Subscribe now

Google recently held its annual developer conference, Google I/O. During the event, the company announced several new AI features and products. Importantly, Google showcased its ChatGPT competitor, Bard, and how it plans to incorporate generative AI into search.

Throughout his keynote, CEO Sundar Pichai hammered home one key point: Google’s products always had AI, and Generative AI is simply the next iteration. In each of the feature announcements, Pichai highlighted how AI was already incorporated into Google’s products (e.g., Magic Eraser).

The company was keen to make the point that they have been doing AI & ML for a while; in fact, many of the key technical breakthroughs that have enabled LLMs came from Google. This slide should assuage the concerns of some investors that were fearful that Google would have to ramp up opex to “catch up” to OpenAI. 

The party line is that Google is just realizing those investments now, with Generative AI being rolled out throughout all of the company’s apps. Since the majority of the features were largely expected and similar to things that Microsoft announced around the Office Suite, the focus of this post will largely be on the announcements related to Bard and Search.

Bard as a Digital Personal Assistant

Bard is Google’s response to ChatGPT (and BingChat, to an extent). Just like ChatGPT, Bard allows users to ask questions and prompt Bard to do tasks with natural language. And unlike the company’s Paris demo, the responses seem to have been error-free1.

Bard already integrates with other Google services and apps, which allows users to prompt actions from Bard and export responses to those apps, e.g., Docs. Similar to Microsoft, Google is trying to weave its suite services together around a common knowledge graph facilitated by generative AI. 

Importantly, Google also announced a number of 3rd party plugins that Bard will be able to interact with directly. Users can prompt the plug-ins to do things directly within Bard.

Bard will be able to tap into all kinds of services across the web with extensions from incredible partners like Instacart, Indeed, Khan Academy, and many more.

For example, users can prompt Adobe Firefly to generate images. 

It’s kind of like giving every human the ability to call APIs directly from the Bard and stitch together their personal workflows. With the plug-ins and integration with Google apps, Google is positioning Bard to be a super-powered “personal assistant” for users on the web. The strategy is similar to ChatGPT’s plug-ins with one key difference: Google is enabling users to interact with Bard directly through Search. The integration is important because the “personal assistant” ties directly with how Google is incorporating generative AI into search. 

What’s Next for Search

One of the most awaited discussions was around Search. How was Google going to incorporate generative AI into search without hurting core search profits? There are two components to this question: 

• What will be the revenue impact?

• What’s the cost of incorporating that type of search?

Revenue Implications

In its demo, Google provides a few examples to showcase how it plans to incorporate Generative AI into search.

In the first example, the user searches for an answer to the question, “What’s better for a family with kids under 3 and a dog, Bryce Canyon or Arches?”. The new search results include a response created by generative AI and traditional search results.

The Generative AI response takes up nearly ½ of the initial page:

Real estate in search results (i.e., ad slots) is incredibly valuable digital real estate. And based on the implementation, it would seem that the implementation is expensive as the company is ceding ~1 ⁄ 2 of the real estate to generative AI. However, if you run the same query in Google today…

…you will notice a few things: (1) the top result is the same, (2) ¼ of the page is already being used for Q&A recommendations (“People also ask”), and (3) most importantly, there are no ads. This type of query falls into the category of non-monetized Google queries.

Per Google, this is what the majority of search queries are:

Nearly all of the ads you see are on searches with commercial intent, such as searches for “sneakers,” “t-shirt” or “plumber.” We’ve long said that we don’t show ads--or make money--on the vast majority of searches. In fact, on average over the past four years, 80 percent of searches on Google haven’t had any ads at the top of search results. Further, only a small fraction of searches--less than 5 percent--currently have four top text ads.

So the net revenue impact on this set of queries is likely nil as there were no ads to cannibalize.

The second example Google demoed falls into the “commercial intent” category. In this example, the user searches for “good bike for 5 mile commute with hills”. The results include a generative AI response but, importantly, preserve the “four top” ad slots by placing them above the generative AI responses. 

Below the sponsored search is a summary of the AI-generated search results that the users can use to research and view more results. From a display perspective, the AI-generated results do not take any more digital real estate than Google's current iteration of “featured snippets.” The generative AI response goes a step further because it incorporates multiple web pages instead of a single source. 

So really, the implementation of “Chat into Search” does not appear to be as negative as people had feared (or at least the stock price suggested). Not much ad space is cannibalized, and the personal assistant opportunity provides Google with some additional optionality.

All of the noise Microsoft made around Search will likely turn out to be for naught. As I’ve written about before, I think we’ll look back at Satya’s infamous “make them dance” as simply a way to grab developer and user attention to say that Microsoft, too, is an AI company. 

Microsoft’s seemingly hell-bent focus on taking share from Google Search will likely turn out to be nothing more than a head fake.

For one thing, Microsoft knows its problem in search is rooted in distribution, not product quality or features. And of all companies, Microsoft is the least likely to be heedless of the fact that competing with an incumbent with a massive distribution advantage is unlikely to bear fruit.

Since Microsoft’s declaration of war on Google Search…

• …Bing’s market share actually went down.

• …Microsoft made some headlines when Samsung was said to have “consider” replacing Google with Bing as the default provider. Samsung abandoned those discussions last week.

But the biggest indictment of Microsoft’s lack of real focus on search is that you can still only use Bing AI on Microsoft Edge. For most Chrome and Safari users, that means forgoing all Chrome Extensions, saved passwords, payment information, bookmarks, etcetera2. On the other hand, Google has made Bard available across all browsers. In many ways, this episode demonstrated the strength of Google’s moat around search. Yes, there’s a technology component to it, but the reach, distribution, and ecosystem around Search is an insurmountable lead.

Furthermore, enabling the chat interface in Google search may actually create ways for Google to better monetize the previously unmonetized “non-commercial” queries.

Google’s AI-powered search results has four components:

1. The AI-generated responses.

2. Ability to ask “follow-ups,” which is powered by Bard.

3. Ability to view the citations that generated the AI response.

4. Traditional search results.

The “follow-up” interface takes users to a full-page view of Bard.

The natural evolution of the Bard seems to be to turn it into a “personal assistant” that combines Search with plugins and generative AI.

With a personal assistant flow, Google could suggest follow-ups to queries that do not immediately have “commercial intent.”  For a user researching travel plans to Bryce Canyon, the natural next step would be to book flights or hotels around Bryce Canyon. Sundar alluded to this on the most recent earnings call:

The main area maybe I'm excited by is we do know from experience that users come back to Search. They follow on. They're engaging back on stuff they already did. And so for us, to use LLMs in a way we can serve those use cases better, I think it's a real opportunity.

For example, Bard’s plug-ins with TripAdvisor and Kayak will likely allow users to do some component of that within Bard. And maybe Google will enable direct flight bookings via Bard in the future? Or the ability to buy directly within Bard? 

Turning Bard into a personal assistant presents a new way for Google to extend Search. Also, Bard, by virtue of being native to Search, will have billions of end-users. And a large portion of those users will have their payment and personal info linked to their Google profiles. All of which incentivize developers to build and invest in plug-ins for Bard. 

How exactly Bard will be monetized is speculation at this point. Maybe it's simply API fees, take rates on payments facilitated via Bard, or “sponsored” follow-ups. 

But once rolled out, Bard will have a number of advantages over ChatGPT. For one, Bard will be subsidized by Google’s ad model, and “premium” features will be free. ChatGPT does not yet have that luxury and charges $20 per user per month to use “premium” features. As a result of that and Google’s already massive reach, Bard will have significantly more users touching the product, which means it's more attractive for developers. 

No doubt, there will be more competition to become the AI personal assistant. There’s Siri and Alexa waiting in the background. It may be that each assistant owns a certain channel, i.e., Siri for iOS. But there are going to be tensions. It will be interesting to see if and how Apple cedes default Search to Google in a world where a percentage of the responses will be “Bard” results. 

Net-net, Google’s approach appears to preserve the monetization ability of the key “commercial intent” queries and increase optionality to monetize the other 80% of queries it does not already monetize with Bard. 

Cost of Implications of Incorporating Gen AI in Search

On the cost of things, investors were concerned that incorporating generative AI into search would increase the cost of queries. But quantifying the cost-to-query is challenging, to say the least. Most of the attempts to do so have assumed fixed cost per query and applied that to some “uptake” of LLM-based queries.

In practice, it will not be a linear relationship for a few reasons:

(1) The majority of search queries are simply not phrased in a manner that requires a generative AI response. You can probably proxy that by cross-referencing the top 1k searches vs. the % that show suggested questions. 

(2) Cost of the query is decreasing as improvements are made to the models. 

(3) Most analyses ignore the impact of caching. When someone searches “Youtube” in Google (one of the most popular queries), Google does not re-run that search query. It will show cached results, which helps reduce latency and compute costs. The historical cache hit rate for Google is between 30-60%. Google does not provide much public data on its search queries, but an analysis done on AOL’s search queries found that caching the top 1M queries captures ~60% of all queries.

Additionally, based on the demo, it seems that Google is using a 2-step retrieval process for incorporating generative AI into Search. When someone searches for a question, Google will find the top search results, and the LLM analyzes the top search results and provides an answer based on the top results with references. The goal of this “grounding” approach is to:

• …avoid incorrect answers by grounding the data on a handful of sources. LLMs are known to hallucinate and provide answers to questions that are incorrect.

• …provide timely information. The data that LLMs are trained on quickly becomes stale. For example, ChatGPT only has information up until Sep. 2021. This two-step approach allows Bard to use the LLMs in conjunction with more timely information.

• …prevent two different people from getting two different answers to “What’s better for a family with kids under 3 and a dog, Bryce Canyon or Arches?”.

Google has not (yet) disclosed if they cache queries, but it seems likely, and Bard seems to agree. Here’s Bard’s response when I ask, “Do you cache queries?”

Yes, Bard caches queries. This means that when you ask me a question, I will first check to see if I have already answered a similar question. If I have, I will return the cached answer. This can help to improve the speed and accuracy of my responses.

The cache is stored in memory, so it is only available for a limited time. If you ask me a question that I have not answered before, or if the cache has expired, I will need to query the underlying data sources to get an answer.

Caching queries can be a helpful way to improve the performance of a large language model like Bard. It can help to reduce the amount of time it takes to respond to queries, and it can also help to improve the accuracy of responses by reducing the number of times I need to query the underlying data sources.

In contrast, here’s ChatGPT’s response

As an AI language model, I don't have direct access to user information or the ability to cache queries. I don't have memory of past interactions or queries unless they are provided to me within the context of the current conversation. Each user interaction is treated as a separate and independent session. Consequently, I can't recall or remember any information from previous sessions or queries. My purpose is to provide information and assist with inquiries to the best of my abilities based on the input I receive in the current conversation.

It’s only logical. LLMs are probabilistic and can produce different results each go around. But you do not want to present different users with different responses for the same search query. And you definitely do not want to provide the same user with a different response to a previously asked question, which means some form of caching will be done. Assuming even a low end of the 30-60% cache hit rate range meaningfully decreases the cost estimates that have been floating around.

Net-net, the negative revenue and cost impact that investors had feared has largely dissipated. The stock price move recently kind of reflects that, but if anything, Google has more optionality around AI than anyone thought or expected.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. This is not investment advice, so do your own due diligence.

Subscribe now

Last week GitLab held a Fireside discussion where they launched and showcased a number of their AI and ML features. Much to the chagrin of its investors, GitLab has seemingly lagged both its direct (i.e., GitHub) and peripheral competitors in launching their own generative AI products. 

But, admittedly, this was a bit unfair for GitLab as the company had been investing in AI and ML functionality. It just has not been in the primary form factor that most investors and customers have become familiar with (i.e., GitHub CoPilot). For example, in 2021 GitLab acquired UnReview, which uses AI to suggest potential merge request reviewers. At that time, the company explicitly highlighted its ongoing development of AI-assisted features and emphasized its significant investments in MLOps tooling, which represented another area of investment for the company.

“Integrating UnReview’s technology into the GitLab platform marks our first step in building GitLab’s AI Assisted features for DevOps,” said Eric Johnson, CTO of GitLab. “By continuing to incorporate machine learning into GitLab’s DevOps Platform, we are improving the user experience by automating workflows and compressing cycle times across all stages of the DevSecOps lifecycle. We’re also building new MLOps features to empower data scientists.” (Source)

But UnReview and GitLab’s other AI-related investments fall into the category of “under the hood” projects i.e., they improve the product but people cannot see them. While these investments make GitLab’s product incrementally better they do not have the visceral effect that products like CoPilot do on both customers and investors. So the underlying reason for GitLab’s AI Strategy call was in reaction to GitHub CoPilot (and its many copycats). The company wanted to demonstrate to investors that they did indeed have the ability, wherewithal, and intention to invest and compete with CoPilot et al1. 

GitLab is not alone in this regard. In the face of the AI tidal wave and democratization of LLMs, most software companies will have to kick off or accelerate an investment cycle with a focus on embedding AI functionality into their product(s). These investments and the associated marketing are targeted at both prospective customers and investors. 

GitLab’s AI & ML Bets and Product Announcements:

The company’s near-term focus is to use AI to drive automation within the DevSecOps workflows and the features/products GitLab launched were primarily focused on that. Here are a few of the features the company announced:

Code Suggestions: AI-assisted code recommendations and suggestions, which help developers write code more efficiently with fewer errors. The functionality is analogous to GitHub CoPilot and is based on Google’s foundation models. 

Suggested Reviewers: Once a developer finished writing the code, GitLab can recommend potential reviewers of the code or merge requests. This feature is an extension of the aforementioned UnReview acquisition. 

Merge Request and Review Summaries: When submitting code for review, developers can use LLM to generate a summary of the change. And on the flip side, reviewers can summarize their review when sending the code back to the developers. 

Explain Vulnerability: The LLMs can analyze new code and existing code for security vulnerabilities and explain the vulnerability to developers/SecOps. 

The company announced a few other features but the overarching theme is that the company is leveraging ML and AI to automate different aspects of DevOps workflow.

GitLab is investing quite aggressively in the opportunity, in part due to competition. Fortunately, GitLab naturally has an edge given its positioning as the core repo and the CI/CD platform for its existing customer base. The impact on the company’s P&L is unclear, primarily because the features are new and the company has not disclosed how they are going to monetize. 

Some of the features will just be included as part of the product and may be available to everyone. Some of it will be tiered and some of it will be charged separately, some on a user basis like per user per month, and some on a consumption basis, where you pay for the computer specifically for that feature in the form of tokens. (Source)

I would expect that the most likely outcome is that GitLab goes with a hybrid monetization model. The company will give the features that are not developer-centric (e.g., issue summaries) away for free and charge for the developer-specific features (e.g., code suggestions) on a per-seat basis. Companies that are charging for the CoPilot-esque features have priced their products at ~$10-20 per month. Some companies like Sourcegraph, which offers universal code search, are monetizing the features by bundling them into their core offering because it’s adjacent to their existing value proposition. 

Assuming a $20 per seat price and 10-20% penetration across the current user base in the first ~24 months, these features would theoretically add 2.5-5% incremental revenue and 350-650 bps to the company’s growth rate in FY ‘25. The impact would not be immaterial but is also unlikely to move the needle for investors, some of which are more concerned about headwinds on the company’s seat expansion2.

More importantly, that revenue uplift is hard to square with the (stated) aggressiveness of GitLab’s investment in generative AI:

Yes, I think a lot of what we do, we generative AI is pretty recent. We really changed how we allocate our resources. And I think now as we look at our development, about 1/4 of our efforts are in generative AI, but I think it's a pretty recent development. So we still have to see a big impact from that.

Now “¼ of our efforts” does not necessarily equate to GitLab spending ¼ of its budget on generative AI. But for argument’s sake, let’s assume that ¼ of the executive team’s time and the company’s net new products are going to be focused on generative AI3. Effectively, GitLab is meaningfully pulling forward its R&D spend. The company did have an AI and ML vision, it just happened to be years in the future and management was taking an incremental approach.

GitLab’s AI and ML Bets

There are a few reasons why GitLab is focusing to such an extent on AI, ML, and generative AI:

Expansion Beyond DevOps. The company sees a long-term opportunity to expand beyond DevSecOps to serve the needs of the engineers and data scientists working on AI and ML projects, which would allow GitLab to serve more personas i.e., seat and TAM expansion.

It’s the current thing. Frankly, AI is the current thing. Customers are interested in learning more about how they can use AI to improve productivity at their company. And investors are looking to discern companies which companies are aligned with or against the current thing.

Competition: GitLab’s competitors in each of their core DevSecOps workflows have announced competitive AI products. For the sake of its marketing message and its IR message GitLab kind of has to have a response.

All the points are interrelated. As previously mentioned, GitLab has been investing in AI and ML for a few years4.

GitLab’s investments can be categorized into three buckets: 

• Applied AI: features and products that use AI and ML to improve the DevOps workflows e.g., AI code generation, suggested reviewers, etc. Generally, most generative AI products fall into this category. 

• MLOps: MLOps is the process by which developers, data scientists, and engineers work together to build, train and maintain ML models.

• DataOps: help companies process data workloads i.e., ETL, which is a prerequisite for building models. 

The near-term focus is using applied AI to improve the existing DevOps5 experience for developers. Because for one, ModelOps and DataOps landscape was crowded before AI took over the zeitgeist and has become even more crowded since. In addition, GitLab’s DevSecOps roadmap is quite expansive as it is and arguably the company has undelivered on their announced features6.

With Applied AI features such as code generation and suggested reviewers, GitLab can serve their existing persona, capture more of their workflows, and make their product stickier. The idea is that over time, the company captures more of company’s DevOps budget and raises prices to realize an ROI on those investments. That’s kind of always been the GitLab thesis and applied AI can simplistically viewed as an accelerant to that thesis.

It’s not meaningfully different than what Microsoft is doing with their CoPilot products. Or what Salesforce is doing with their suite, or what Atlassian is doing with their suite, etcetera. 

The DataOps and MLOps bets were designed to help GitLab expand beyond DevSecOps. Both DataOps and ModelOps were (and are) designed to be long-pole investments to position the company to become an “AllOps” platform: 

We see it with data and operations teams creating DataOps; we see it with machine learning and ops teams creating MLOps. As more companies – and more teams within a company – rely on our platform, we are positioned to become the AllOps platform – a single application for all R&D. (Source)

Given that these bets are likely going to be immaterial for GitLab in the near term, the “why” for GitLab’s investments in those relatively crowded categories is more interesting:

We see ModelOps as a big opportunity for GitLab. ModelOps is a combination of data ops and MLOps. It's everything you need to do to add AI to your application. In GitLab, we're adding AI to our application to GitLab, but with GitLab, customers are creating applications and they now need to AI to that. We want to help them on that road. We already started building this functionality in 2021 in collaboration with the wider community. That allowed GitLab's CICD runner support for NVDI GPUs. More recently, we added the capability to link MLFlow experiment with GitLab experiments.

Later this year, we plan to introduce a model registry allowing our customers to store version, deploy and track the health of their AI and ML models, natively within GitLab. We want to help our customers be more productive every significant application is going to have both code and AI, and we want both managed with GitLab. (Source)

Gitlab (like others) believes that AI is going “eat” software. The idea is that every major software application will have some Applied AI features. Or said another way (more cynical way), Applied AI functionality is going to be table stakes. In this world, there are not going to be bright lines between MLOps and DevOps. Developers will need to work with data scientists, ML engineers, etc. to build their own models and/or fine-tune and implement 3rd-party models. GitLab’s vision to become the one platform for all software developers means they serve the needs of data scientists and data engineers. 

Vision: Identify opportunities in our portfolio to explore ways where GitLab can provide a better user experience for Data Science and Machine Learning across the entire Machine Learning life cycle (model creation, testing, deployment, monitoring, and iteration). (Source)

AI is Eating Software

You’re already seeing “AI-eating software” play out in GitLab’s core DevOps market. One of the main reasons that GitLab even needed to have an “AI Strategy” call is because the company’s direct and peripheral competitors have launched applied AI products. The first to do so was GitHub CoPilot, and recently Atlassian announced Atlassian Intelligence which provides LLM technology across Jira, Confluence, and Bitbucket. More recently, HuggingFace (a future MLOps competitor for GitLab) announced a partnership with ServiceNow to offer code assistance. 

Importantly, it’s the current thing with clear value. Both investors and customers are trying to understand who is well-positioned to offer those features. Investors are also interested in understanding who is not well positioned. In light of Microsoft’s aggressive move with CoPilot and their traction, there’s definitely been a perception that GitLab has fallen behind the competition.

GitLab (like Google) was caught off guard by Microsoft. Unfortunately, the threat to GitLab is more pronounced because GitHub is a far more formidable competitor to GitLab than Bing can ever be to Google.

So in reaction, GitLab has pulled forward its AI and ML investments. 

And when everyone's super…

The imperative to pull forward “AI investments” is not limited to GitLab. For example, most of Google’s recent keynote was focused on highlighting all the ways the company already has AI embedded in its products and the investments they are making. Some of this is just calling products “AI” that were previously not labeled as such.

And over the last few months, almost every B2B software company with more than a handful of engineers has announced features that will incorporate generative AI into their product. A few examples are below:

• Glean

• Outreach

• Zoom

• Five9

• Thoughtspot

For these companies, and other software companies, OpenAI and Microsoft’s moves have presented a Catch-22.

One of the reasons that GitLab could only make incremental AI and ML investments is that the cost to train LLMs was previously prohibitive for any one company to incur. Fortunately and unfortunately, Sam Altman and OpenAI have effectively played the role of Syndrome from The Incredibles. 

OpenAI brought commercial-grade generative AI to the masses. And by doing so, they also forced other vendors (e.g., Google) to open up / commercialize their models, which made it feasible for all companies without deep pockets to incorporate generative AI. 

But to quote Syndrome:

I'll give them heroics. I'll give them the most spectacular heroics anyone's ever seen! And when I'm old and I've had my fun, I'll sell my inventions so that everyone can be superheroes. Everyone can be super. And when everyone's super no one will be. 

By “opening” up OpenAI, OpenAI gave access to the technology to everyone7. The challenge for incumbent software companies is going to be that if everyone has AI features, no one really has AI features. Not literally, of course. It just means the bar for AI features is rapidly increasing and if you are offering similar features as your competitors based on the same out-of-the-box model, there is going to be a cap on how much you are able to charge for those features8. 

The other implication has been that companies that have previously built their businesses around NLP and generative AI are under siege. Or said another way, are no longer super.

But the rapidly advancing technology threatens other companies that have spent years crafting software to automate tasks or to build machine-learning models, including richly valued startups such as Databricks9, Snyk, Zapier, and Talkdesk….

”If companies like Grammarly don’t figure out what is their unique competitive mode pretty soon, they’re going to be quickly replaced by every other text-based interface that integrates LLMs” (Source)

An area where this thesis has been effectively A/B tested is AI copywriting marketing. Some companies had early access to GPT and were able to quickly scale up. But once GPT-3.5/4 was released GA, their early technical advantages faded and competition increased. The scaled players – the ones that got early access – might still be the long-term winners but it’s also non-controversial to say that competition has meaningfully increased for them. 

The pattern extends to more mature categories as well. Generative AI is creating an arms race among incumbents and effectively increasing competition and the R&D burden.

Take the example of Box and Dropbox. Box recently announced AI functionality that can be used across their content storage products, called Box AI:

Box AI will make it easier than ever to uncover and share insights, find timely answers to critical questions, and effortlessly create content based on an organization’s data in Box. The company also announced that it will integrate OpenAI’s most advanced AI models with the Box Content Cloud, further enabling new ways to understand and create content on Box… 

Box AI will bring foundational AI models to where their content is already securely stored, making the files inside of an organization more useful and valuable than ever before. (Source)

And a few days prior to Box’s announcement, Dropbox announced they were laying off 16% of the company’s workforce because of an AI imperative:

The opportunity in front of us is greater than ever, but so is our need to act with urgency to seize it. Over the last few months, AI has captured the world’s collective imagination, expanding the potential market for our next generation of AI-powered products more rapidly than any of us could have anticipated. However, this momentum has also alerted our competitors to many of the same opportunities. (Source)

The company would articulate the AI imperative more explicitly on the next earnings call

The first objective is around building AI-powered product experiences centered on organizing cloud content. We believe for many years and the potential for AI to completely transform and knowledge work and we've been investing in automation and machine intelligence features to help our users organize their cloud content and search and discover more easily.

Did Dropbox lay off 16% of its workforce because Box was going to announce an AI feature set? No.

But both companies understand there’s a strategic imperative to incorporate AI into their platform soon rather than later especially given its relevance for enterprise search and the Microsoft/Google threat. After all, no one wants to be a “Buddy” in a post-GPT world. 

Applied AI features without question make the products better but the actual economic payoff is less certain. In some instances, companies have been able to charge for the tool (e.g., GitHub CoPilot) but in other instances, the functionality is free for customers (e.g., Atlassian Intelligence). 

In the previous investment cycle (cloud), investing in the cloud transition made economic sense. The trade-off in gross margins was worth it and over time, that became the standard. But with AI, it may just lead to incremental R&D spend for some categories unless they find a specific way to realize the ROI.

Atlassian is a good example because I think that’s the model that incumbent application software companies will follow. Atlassian is offering Atlassian Intelligence as a cloud-only product and using it as an incentive for the company to speed up its cloud migration. My model is that mature categories (e.g., CRM) are segments where AI will primarily serve as a “carrot” for something (typically bundling or capturing more usage/workflows). This will inevitably require accelerated investment across the P&L forcing an investment cycle for incumbents (see: Box vs. Dropbox), which was not previously planned nor expected by either investors or management10. 

On the other hand, the parts of the market/companies that I think are going to buck this trend are (1) less mature categories or (2) companies that take some tech risk (e.g., build their own domain-specific model). The second one is tempting because it has the allure of “being proprietary” and “having data moats” but it will be a more capital-intensive model (and more likely to fail). But leaning into this type of uncertainty is kind of required if you’re trying to pull away from the crowd or compete with incumbents. From Jerry Neuman

To be successful a founder must seek out uncertainty and then must manage the company through it and emerge from the other end with a moat. Uncertainty creates a difficult trade-off for entrepreneurs. Without uncertainty they will immediately face competition from many others, including some who are better resourced.

Theoretically, GitLab could have come out and said “we’re going to build our proprietary code-gen model that will be better than GitHub CoPilot”. But investors would’ve probably had a conniption. Because for one it’s clearly going to require a lot of money and two the payoff is entirely uncertain. And investors like the certainty of outcomes, unless the management has a great deal of credibility for taking those types of risks.

For those reasons, GitLab and most other companies have decided to invest in “certainty”11 (i.e., take existing foundation models and fit them to their use case). By definition, this means the returns on increased development spend are going to move towards an average with little room for outlier returns12. On the flipside, the threat for incumbents will likely come from companies that do take the innovation or technical risk such as betting on the fact that models are getting smaller and easier to train.

Fin. But as with every AI and ML post, I’d like to end with: well anyway, it’s early and we’ll see.

Next week I’ll be going deeper into Google’s I/O Conference and its developer Keynote announcements.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. This is not investment advice, so do your own due diligence.

Proceeds from the conference support MSK Kids, the pediatric program at Memorial Sloan Kettering Cancer Center (MSK). You can (should) donate to the cause through the link below:

Donate to MSK Kids

My notes and key takeaways from the conference are below. If you have any questions or suggestions about the notes, please comment below or email me at ardacapital01@gmail.com. And per usual, this is not investment advice.

Investment Pitches

Divya Nettimi – Avala Global

Long Sony — the key thesis points are:

(1) Market underappreciates the quality of Sony’s assets and the recurring revenue/profit mix shift over the next few years. Sony is a misunderstood conglomerate with three key assets: Gaming, Image Sensors, and Music, all of which are number one in their respective segments.

• The company’s gaming business has transitioned from a cyclical hardware-driven business to a subscription and software business.

• By FY ‘25 ~80% of the gaming business’ revenue will come from subscriptions and software so the profit mix of the business will be higher quality and higher margin.

• The image sensor business is the largest sensor manufacturer in the world e.g., used in every iPhone. Business is riding tailwinds from the increased content in cars and increased demand for high-quality sensors in phones.

• Sony Music is less of a focus for the thesis but is also a market leader and is taking share. Should benefit from upcoming streaming price increases.

(2) Operating profits will inflect over the next few years. Both the Gaming and Image Sensor segments faced headwinds over the last two years driven by supply chain issues and lockdowns. Those headwinds are expected to ease over the next two years, which leads to an inflection in operating profits.

• Sales of PS5s were supply constrained during the last two years. As those headwinds fade, PS5 sales and installed base should grow 80% vs. FY ‘22. In addition, Avala forecasts console gross profits will go from negative to positive as input costs come down. The net impact is that gaming profits will CAGR at 27% over the next three years.

• The Image Sensor business has also had headwinds on both the supply and demand side driven by lockdowns in China. As Android volumes normalize, and the company’s supply chain and manufacturing issues ease, margins should increase from ~15% in ‘22 to ~18% by FY ‘25. Avala expects this segment’s profits to CAGR at 22% over the next three years.

• Gaming and sensors are the primary aspects of the thesis but the other segments are attractive as well. Sony Music continues to take share in the market and as the streaming platforms raise prices, the company should benefit.

(3) Attractive Valuation and Risk/Reward. Sony currently trades at 11x Avala’s FY ‘24 estimates. Based on Avala’s forecast for FY ‘25E, they think the stock has a ~58% upside (~30% IRR) with some call options from the company’s equity investments and other investments.

Mala Gaonkar — SurgoCap

Long London Stock Exchange (LSE) — the key thesis points are:

London Stock Exchange has transformed into a data analytics business. Through the company’s acquisition of Refinitiv, LSE has transitioned from a volatile exchange business to a data and analytics (“D&A”) business.

• The acquisition was controversial and dilutive. Has led to the company underperforming since the transaction. 

• But 70% of the company’s revenues now come from the data analytics business. Surgo’s view is that this is a good trade because the D&A revenue stream has low churn and high pricing power.

Refinitiv was a mismanaged asset that LSE has turned around. After LSE acquired Refinitiv, the company made major products and salesforce investments.

• LSE replaced the legacy Eikon platform with the more modern Refinitiv Workspace and has transitioned ~50% of the customers over. 

• The company has also invested in the salesforce and better-aligned sales incentives to incentivize cross-sell.

• The investments are starting to pay off. Refinitiv has taken price for the first in line with their peers with limited pushback and has led to a 2.5x increase in subscription revenue growth.

• Management has increased revenue synergies from 225M to 400M.

Re-rating from business transformation. Surgo argues that the transition should allow LSE to re-rate from ~22x (in-line with exchange peers) to ~25x P/E, which is more in line with the data peers, which trade at ~30x. Gaonkar points to FICO, S&P, and Transunion as case studies for the re-rating. If the company does re-rate to 25x, this represents a 166 pounds PT (~28% IRR from current prices).

David Rosen — Rubric Capital

Long Talen Energy

• Post-reorg equity that will start trading OTC at the end of the month. Will start trading on the NYSE or Nasdaq by EoY.

• Talen’s assets include Susquehanna (leading nuclear power plant), 10GW of non-nuclear energy production facilities, and some data center and crypto-related assets.

• Surprisingly, the company filed for bankruptcy during a period when electricity prices were rising. Talen had hedges that required the company to post collateral as prices rose, liquidity they did not have so they had to file for bankruptcy. So the capital structure was not the core issue.

• The shareholders (incl. Rubric) put in place a new management team.

• The current value for the equity is $2.5bn but the company’s SOTP value is $4.5-6.5bn ($4-5.3bn without the non-generation assets).

Aaron Davis — Boxer Capital

Long Mirati

• Mirati is a targeted oncology company. The company develops drugs to treat the genetic drivers of cancer.

• The stock is down 80% from ATHs; “generational drawdown in biotech. The current market cap is ~$3bn with ~$1bn cash on the balance sheet.

• The lead drug received approval in Dec '22. The annual revenue opportunity for the drug is ~$600M and the second drug in the company’s pipeline has a 3x addressable population. The second drug is currently still in the pipeline.

• Call options on the rest of the pipeline.

Scott Goodwin — Diameter Capital Partners

Long Level3 Bonds

• The pitch is to buy secured (83c) & unsecured bonds (58c).

• Bonds have traded down because of business and revenue declines with the corporate parent (Lumen) but Level3 has an independent cap stack.

• Core Level3 business is doing fine (relative to Lumen) and has both asset and profit coverage for the bonds.

• The bonds are fully covered by the assets and present a total return opportunity of ~100%

David Einhorn — Greenlight

Long Vitesco

• Euro auto parts supplier, spin-out of Continental AG. European mid-cap and is effectively an orphaned equity.

• The business is transitioning from an ICE supplier to an EV supplier. The spin-out was to designed to help the business more aggressively make the transition.

• Current EBIT margins are depressed because of EV business’ margins are negative but will become profitable in the next two years. As a result, the total company EBIT will 3x by 2026.

• The total company valuation is 2.6bn EUR, which per Greenlight, means you get the EV business for free.

Andrew Weiss — Weiss Asset Management

• Long Korean HoldCo's SK Square, Samsung C&T, LG Corp.

• All trading at a 70-80% discount to their net asset value.

• Unclear what's driving the discount other than management and governance concerns, which Weiss thinks are overblown.

Welcome to Tidal Wave, a research newsletter about software, internet, and media businesses. If you have not yet, please subscribe below, and would appreciate it if you could follow me on LinkedIn (because you know, Elon).

Subscribe now

Later this week I will be covering GitLab’s AI announcements and the competitive dynamics in AI as well as any relevant news from the Google I/O presentation.

Macro Discussions

Takeaways from Stan Druckenmiller’s Discussion

• Every time there have been interest rates below 2%, it's been followed by difficult economic times.

• “Biggest maxim in my business is don’t fight the Fed”

• Much of the anecdotal evidence they look at (e.g., trucking, retail, banking) is all pointing negatively.

• Hard to predict in this environment. The most important thing is to keep an open mind and react to policymakers.

• Next year is going to be an election year, which puts the Federal Reserve in a precarious position if there’s a hard landing.

• AI plays for Stan have been MSFT and NVDA. But questions if NVDA even experiences a downturn if there's a recession given chip/AI demand.

• Copper is in the tighest position he's remembered but it’s hard to invest in that going into a hard landing. But coming out of a contraction, it'll be a huge beneficiary given the underlying demand drivers.

Patrick Collison’s Interview of Sam Altman

• Large models vs. open-source (OSS) models:

  • OSS models will get benefits from the community.

  • But large models will always be on the frontier and years ahead of open source. The large models will solve the "harder" problems first (e.g., solving medical problems).

• Sam's main personal use case for ChatGPT is summarization. Doesn’t really use other AI tools. RFS: CoPilot for everything on his computer.

• Facebook’s strategy around AI has been “confusing” but they will be a surprising AI player soon.

• Models are not a threat to search. It’s only a threat to Google if they don't do anything. So far the company is moving with intensity.

• OpenAI will be a platform and consumer-facing co (e.g., ChatGPT). The platform will be used to improve consumer apps.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author holds positions in the securities discussed.

This is a follow-up to my last post on Atlassian, which covered the company’s early beginnings, flywheel, and the company’s Act 1.

Today the company has three growth vectors: (1) cloud transition, (2) enterprise adoption, and (3) ITSM. All three initiatives are bets to lengthen the runway of the company’s growth. If they can deliver on these initiatives, the company likely has the ability to grow its top line in the mid-20s for the next 5-7 years. However, there are a number of open questions that make the stock challenging to chase even after the post-earnings pullback. 

Cloud Transition

For those that have not been following Atlassian, the company is in the middle of a cloud transition. 

As a company founded in 2001, Atlassian started as an on-premise product. Even as the cloud increased in popularity, existing customers of Jira and Confluence were content and preferred to keep their deployments on-premise. At the company’s IPO in 2015, cloud deployments represented only 25% of revenues. 

Jira and Confluence are mission-critical products and store a lot of a company’s sensitive data. And for a long time, the founders were agnostic between cloud and on-premise deployments. 

But by 2015, it was becoming clear that the cloud was the future. For one, ¾ of Atlassian’s net new customers were coming via cloud deployments. Atlassian, unlike other SaaS around that time, did not push customers onto the cloud, primarily because their cloud platform was not ready for it. Back then, if a customer wanted to use Atlassian’s cloud product, the company would create a private container for that customer in the cloud. This meant that the cloud instance could not be easily scaled up or down and importantly it was very slow. One of the common refrains of Atlassian’s cloud customers from this time is that it was “slow”. Even with all the improvements the company has made in the platform, scar tissue remains with this cohort of users as many will still say Jira is quite slow.

Additionally, Atlassian’s cloud product could only support 2K concurrent users, which made it a non-starter for enterprise customers. These shortcomings for the company’s cloud product left the door open for cloud-native entrants to chip away at the company’s TAM, especially for the non-developer use cases. So in 2016, the company began a multi-year cloud transition and investment. The project was “completed” in 2018 but scaling it to serve the enterprise customers that were still on-premise would take another two years. This type of investment to scale cloud products for enterprise reliability is not uncommon, Workday had undergone a similar investment cycle after acquiring Adaptive Insights.

And over the course of the next four years, the company transitioned its cloud platform onto AWS and built it to be scalable for the enterprise. This involved forking the code and engineering organization in two: cloud and on-premise. As a result, this led to the company’s R&D burden increasing for a period.

As the company’s investments pared down, the R&D burden stabilized around ~34%. During the transition Atlassian, naturally, did not push customers onto the cloud platforms and customers also did not have an incentive to make the transition. 

One of the primary motivations for Atlassian to push customers off the server products and onto the cloud platform is that it’s harder to cross-sell customers when they’re on-prem. 

‘20 Investor Day: In my experience with the server and data center business for such a long time, cross-sell, you just got to think of the barriers a customer has to go through on a server deployment to add another product. First off, we'll e-mail, we'll market, we'll do ads. We'll do everything we can to surround a customer with potential new products that they could go use. Then someone, preferably with administration access, needs to go to our website, download those products, install them, get an evaluation or a trial, get everyone else that might be interested, get them on the product to go through the trial, go through the purchasing experience and then integrate that with whatever product they had on-prem before. Like the amount of time, energy and complexity of that is -- and how much Atlassian has visibility into those different pieces is very gapped, right?

After the project was completed, Atlassian began pushing customers off the server and data center products with some carrots and sticks…

• …announcing the end of life for the server products (stick)

• …increasing the prices of the data center and servers products by 15% in ‘21 and ‘23 (stick)

• …offering discounts for the cloud product for early adopters (carrot)

• …most importantly, only offering new features for the cloud products (carrot)

Customers’ reception to the news has been mixed. On one hand, migrating to the cloud decreases the total cost of ownership and there’s no meaningful business value in maintaining Jira and Confluence instances yourself. Jira and Confluence are mission-critical in the same way that an email server is mission-critical for organizations - you cannot work without them but there are few strategic benefits of maintaining an email server. 

On the other hand, customers, especially larger and/or customers in regulated industries like to keep control of their deployments and data1. Most importantly, Confluence and Jira going down can be crippling for an organization. People will not be able to get their work done and there’s the risk of losing critical data. Unfortunately for some customers that did make the transition, the worst nightmare came true when Atlassian’s cloud product had a nine-day outage. This no doubt, slowed the transition of Atlassian’s customers to the cloud. 

In spite of that hiccup, Atlassian has transitioned ~60% of its revenue base to cloud deployments. Some customers have chosen (or forced) to treat the data center as a middle step before fully moving to the cloud… 

…so ~40% of the revenue base remains on either the server or data center. The pace of migrations to the cloud from the data center is choppy but growing. My estimate is that ~8% and ~14% of the data center business migrated to the cloud in FY 2022 and FY 2023, respectively. I assume that trend continues to be driven by cloud-only products and double-digit price increases. 

Benefits of the Cloud Deployments (for Atlassian)

For Atlassian, pushing all the customers to the cloud has a few benefits:

(1) Improved cross-sell. With customers on cloud deployments, Atlassian is able to more programmatically drive upsell and cross-sell of new products. Expansion within the company’s customer base was primarily driven by seat expansion on Confluence and Jira.

​​’19 Analyst Day: So an average customer should add their second product sometime in their second year and on average, over 6 years, most customers have 3 products or more.

But most of those cohorts were customers that landed with Atlassian’s on-premise product. The company’s cloud customers expand faster. 

‘23 MS Conference: So we've really put in a lot of work and that work shows in terms of customer reaction. From a business angle, there is clear evidence that customers in cloud, especially larger customers in cloud, cross-sell faster, expand faster.

Estimated cloud expansion rates seem to be trending higher than the overall company’s, at least the cohorts disclosed at the IPOs.

There are a few confounding factors that make the early cohorts seem better than they likely will be longer term: (1) there was discounting on the initial cloud deals to incent customers to switch and (2) early on-premise customers were on legacy pricing tiers that were simply lower. 

Yet, even if you conservatively assume that Atlassian can transition the remaining cohorts and get cohorts that are adjusted by 1 year, that implies a meaningful tailwind to support at least low-20 growth for the next 3-5 years.

(2) Reduced R&D Burden. Since Atlassian forked the code base and effectively created two different technology stacks, the R&D burden today is higher than in a world where they were only supporting cloud customers. As the company stops supporting the server customers and transitions Data Center customers to the cloud, the R&D burden especially on the infrastructure side should decrease. Long-term margins should be slightly higher than what consensus implies but the puts/takes of data center instances on R&D is not clear. 

(3) Faster deployments and sales cycles. Cloud deployments and sales cycles should be faster. Atlassian has been investing in “enterprise advocates” over the last few years, and the friction in the sales (or trial) process is much lower than with the on-premise products. As customers complete their cloud transitions, the barrier for enterprise advocates to sell ITSM should decrease. 

To further drive adoption, the company is exclusively releasing its new products for cloud customers. Over the last few months, Atlassian has announced the following products:

• Jira Product Discovery - helps product managers ideate and do product discovery, somewhat analogous to Productboard

• Confluence Whiteboards - virtual whiteboards for teams to collaborate on the cloud a la Miro. 

• Atlassian Intelligence - CoPilot for Jira, Confluence, and the rest of the Atlassian cloud products. 

All these products are only available for the cloud customers. By continuing to release new and in the case of AI, trendy, features only on the cloud, Atlassian is forcing the hand of IT and developer teams to move to the cloud. I could make the argument here that the AI product will be a catalyst for re-bundling around Atlassian’s knowledge graph and may accelerate cloud migrations. But I won’t because I’ve made effectively the same argument for Microsoft and the productivity suite. I think that logic broadly applies here. 

Other Bets

In addition to driving more cloud migrations, the other focus areas for Atlassian are (1) ITSM with their Jira Service Management product and (2) Enterprise. Both initiatives seem great on paper but the company has not shown that they’ve de-risked these bets.

45k of the company’s ~260k customers (~17% of total) are currently using Jira Service Management (“JSM”). The company’s motivation for pursuing the strategy is simple, there are many customers that currently use Jira’s ticketing platform “service” workflows including IT requests.

We talk to our customers every chance we get, and nearly 40 percent of them have already been using Jira for service desk operations. After hearing their stories, we set out to give them a powerful new solution to help manage requests even faster than ever. 

Unfortunately, the quote is from 2013. Atlassian has been at this for ~10 years and would probably still argue that they are in the third or fourth inning of this bet, as exemplified by the product penetration. This has been one of the shortcomings of Atlassian. While the company has diversified a bit from Confluence and Jira, they have not been able to organically create another meaningful growth driver.

One of the reasons JSM is not further along is because of the company’s relatively late transition to AWS and slow cloud products. This is something that Freshworks has historically used as a differentiator.

The strategy is both offensive (vs. ServiceNow) and somewhat defensive (vs. Freshworks et al). 

The company’s positioning against the enterprise vendors is that it’s a simpler, easier-to-use, and much cheaper product. ServiceNow’s product is entrenched and built for the Fortune 500 and Atlassian is really targeting the rest of the “Fortune 500,000”. In effect, the company believes there’s a fat middle of the market for which ServiceNow and BMC are too expensive and hard to manage. 

Atlassian’s positioning vs. the Freshworks of the world is that it has feature parity, is ½ of the cost, and sits on the same platform that the company’s developers and IT team use for their software products and internal tooling. Freshworks et al are separate platforms. The benefits of Jira are that there’s no learning curve for new users and the knowledge sharing between the developers and IT is more cohesive on the same platform.  

The management team has “doubled down” on ITSM in an effort to capture the installed base opportunity. And the messaging and timing might just be right. We’re going into a downturn, Jira has a good product at a fraction of the cost of its competitors. And the company’s peers may be forced to pull back because investors focus on efficiency. This is where Atlassian’s business model may serve as a competitive advantage and allow the company to consolidate the ITSM market. 

Traditionally, you would expect the market leader to offer sales incentives such as buyouts to incentivize customers to rip-and-replace but that it’s a bit too anti-Atlassian. It remains to be seen how impactful ITSM and the enterprise push will be on the company’s financials. The results when looking at headline metrics (e.g., incremental ACVs and NDR) would imply the bets are not meaningfully moving the needle. 

Valuation

The lack of conviction on ITSM and the pace of cloud migrations make it hard to lean into Atlassian even after the recent pullback. The company reported earnings and it was fine, broadly in line with consensus and slight miss on street expectations for Q4 guidance. The stock traded down ~10% and is currently at ~$130. The stock is not yet glaringly cheap at ~9.5x NTM revenues and >40x NTM FCF.

The risk/reward at the current price is good on an IRR basis. My base case is that you can see a path for the company to generate ~$7.50 FCF per share in four years. Assuming a three-year hold and a ~30x FCF multiple at exit, you can get to a low 20% IRR. But the multiple assumes the company can demonstrate to the market they sustain mid-20% top-line growth and for that to work, you have to assume that the other bets outside cloud transition payoff. 

The downside case is that the company’s cloud migrations get pushed out and ITSM/Enterprise progress does not meaningfully change. In this scenario, the company trades at ~20-25x FCF with <$7 FCF per share, which implies a mid-single-digit IRR. 

The skew is such that there isn’t meaningful downside risk but conviction in the base case is not particularly high. In other words, the growth model for the next few years does not seem particularly de-risked so for now I’m on the sidelines.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author holds positions in the securities discussed.

In his canonical book, The Mythical Man-Month, Fred Brooks argues that there are three types of tasks:

• Unpartionable tasks: tasks where no matter how many people you add, the time to complete (“ToC”) does not decrease e.g., giving birth

• Perfectly partitionable tasks: tasks where adding more people perfectly decreases the ToC e.g., wheat picking

• Partitionable tasks with intercommunication: tasks that require two-way communication. In these tasks adding people beyond a certain point actually increases the ToC (pictured below).

Software development, and most knowledge work, generally falls into the fourth category. As more people are added to the process, at some point, the time it takes to complete the job grows rather than decreasing.

Collaboration tools are designed to solve this coordination problem by improving communication and knowledge diffusion across the organization1. Atlassian is one such company, the company’s mission “is to unleash the potential of every team”. With their beachhead products, Jira and Confluence, the company has built the leading collaboration software platform. 

Early Pivot

But Atlassian was not always a software company. The founders Scott Farquhar and Mike Cannon-Brookes started working together in 2001. Originally, their business was selling one-off support and debugging services for customers of a Swedish software company (i.e., consulting). After ~1 year, they decided this was not a great business to be in.

Fortunately, by that time the founders had built a few internal tools to help them manage their consulting business: 

• Atlassian Support System - customers of software companies would use this tool to file support tickets that they wanted help fixing. 

• Wiki Tool - to help manage the documentation of the software they were supporting. 

• Email archiving - used to “archive” all the emails they were getting. 

So they made a classic startup pivot - take a tool they’ve built for themselves and sell it to people that look like them, developers. The duo decided to commercialize their service desk tool, Atlassian Support System, and sell it to developers who would use the product to help manage their projects, file bugs, and track progress on projects. Oh, and they fortunately2 decided to rename the product, Jira. 

And it was off to the races!

Theory of Constraints

Well not exactly. Unfortunately for the Farquhar and Brookes, they had two problems: (1) they were building a software startup after the dot-com bust and (2) they were in Australia and the venture ecosystem had just blown up and not in a good way. 

But for us, VC wasn't an option. This is 2001, 2002. It’s the nuclear winter for tech. The world has just blown up. Australia had a couple of VC funds that emerged in 1999-2000 and invested in dStore and all these crazy businesses at absurd prices. All those funds completely blew up. They raised $100 million and returned $5 million — those kinds of numbers…

And so, we had to figure out another way. We used to go around and if we needed to sound smart in Silicon Valley circles we’d say, “Oh, we didn't really go down the VC funding route. We were customer-funded.” People would reply, “Okay, oh, wow. That's an interesting choice.”

And I'm like, “Okay, that worked.” (Source)

So the company, which at this point was just Mike and Scott, had a problem. They were building Jira but they did not have the capital to hire sales reps to sell their products. This funding constraint forced the founders to effectively innovate on the distribution model for their software. 

We also didn't have any money, so we couldn't hire anybody.  So the Atlassian business model came because we knew we needed to sell software somehow online, because we didn't have any salespeople, we couldn't afford to hire them, we didn't have any money, so it kind of had to sell itself. And that required us to do a lot of this tracking and modern things a long time ago….

And it really was a process on constraints. I really think if we had grown up in Silicon Valley and had venture capital around us, that we would have built a very different company that wouldn’t have been as disruptive as ours has been because we didn’t have venture capital…

And almost the thing we built in that model is actually almost less important than the fact that we changed the business model of how people adopted software to be much more consumer-like. (Source)

Atlassian’s founders were at the forefront of self-serve software purchasing. Prospective customers could go to Atlassian’s website, put down a credit card, download the product, and install it themselves within a few hours or a day.

The product had to be cheap. If it was too expensive, they would need sales reps to sell the product. Sales reps they did not have nor could afford to hire. If the product was going to be cheap (the first license was sold for $600) and you want to build a big business, you have to sell a lot of it. If you have to sell a lot of it, especially out of Australia in 2002, you have to sell the product globally. Global sales means 24/7 availability so naturally, the business model had to be internet-centric and self-serve. 

So that’s what the founders did, they built a product that could be purchased on the internet with the swipe of credit and could be installed without external consultants/resellers. The strategy combined with early guerrilla marketing tactics worked.

They woke one morning in 2003 to find an order from American Airlines lying on the fax machine. According to company lore, they looked at each other and said, “Did you talk to them?”

“No, did you?”

“Holy shit! American Airlines saw Jira on our website and bought it just like that!”

The story is analogous to Salesforce, which used the cloud delivery model to go direct to their buyers at a time when most enterprise software companies had to rely on resellers and partners. The companies could not be more different but innovating on the delivery model and go-to-market (relative to the industry) is a theme we often see with technology upstarts. It’s a theme you see often with the “disruptors” in a given market, in many cases, it’s less about product innovation but rather taking advantage of an underused channel or distribution model. Another example is Dell, which used the direct-to-consumer model to disrupt the PC industry.  

For Atlassian, beyond innovation, allowed the company to scale efficiently and helped build one of the most efficient GTM motions. Atlassian likes to emphasize this in their investor materials:

It’s pretty easy to find qualms in the chart and those quals are fair; the companies have different GTM motions, are at different stages of their life cycles, serve different types of customers, etc. 

The better way to see the power of Atlassian is to compare the efficiency of its sales vs. its peers and at similar stages, which the table below shows. 

Atlassian not only has the most efficient sales motion but maintains that efficiency as it scales. Additionally, the company is able to maintain that efficiency while maintaining the strongest growth durability among its peers. Arguably all of the companies in the table would fall into the category of “product-led growth” or “self-serve” and should have low CACs. The Gladwellian narrative would be that Atlassian’s constraints and lack of venture funding3 ultimately led the company to have a different DNA than their peers that were founded much later in better funding environments4.

The company’s sales strategy and efficiency are the foundation of the company’s flywheel5.

Like other great business models, the savings from an area of the business or line item can be used to fund excess investments in other areas, at least relative to peers.

The GTM is what enables Atlassian to invest in R&D aggressively and offer high-value products for low prices. The company was able to use the savings from their efficient S&M, the company could (1) re-invest in product development and (2) charge a lower price than their competitors.

Product Journey

Atlassian’s product investments (organic and inorganic) have allowed the company to consistently expand its product portfolio and the personas the company serves. 

Atlassian started with Jira, which is a project management tool used for agile software development, allowing teams to track tasks, customize workflows, and monitor progress. A year later, the company launched Confluence. Confluence is a collaboration and document management tool that allows teams to create, organize, and share knowledge, information, and ideas in a centralized location.

Because of the inherently cross-functional nature of software development (and maintenance), both Jira and Confluence expanded beyond software teams to serve non-developers. Both Jira and Confluence are used by product managers, IT, designers, marketers, etc. And because IT had already implemented Jira and Confluence for the technical teams, the product was re-purposed to help non-technical teams (e.g., HR, finance, and legal) manage their workflows. 

The combination of the two products would serve as the basis for the company’s growth for the two decades. 

After Confluence, Atlassian launched or acquired a number of other products including Bitbucket (code deployment), Jira Service Management (ITSM), and Hipchat (messaging). The best way to frame the company’s portfolio is that Atlassian built different tools to address different parts of the DevOps workflow, which include maintenance of the software. I will not go into the different components of DevOps here but would suggest reading Gitlab’s overview. 

There are a few unique aspects of Atlassian’s product strategy:

(1) Early to multi-product. The founders went multi-product in Year 2 of the company’s history, which was controversial at the time. The conventional (VC) wisdom is to focus on the core product until you have tens of millions of ARR and then expand to the second product. Most companies never make the leap.

(2) Served multiple personas. Really 1(b) but Atlassian was able to serve multiple personas early in the lifecycle of the company. The critical decision and why it worked was that they did not meaningfully change the product to serve the needs of the other personas. Persona drift is a slippery slope for startups, takes up a lot time, and more importantly, burns capital. Atlassian did not have to do that because their product was infinitely flexible and their customers (e.g., IT and devs) would customize the software to serve the needs of their internal teams. This would later become an area of weakness for Atlassian.

(3) Tuck-in acquisitions. Atlassian employed acquisitions effectively. The playbook seemingly was to acquire a leading developer tool to serve as a beachhead product into different parts of the DevOps workflow or in the case of Trello, the cloud deployment model. 

The combination of self-serve GTM and product strategy has been the foundation of the company’s ability to expand consistently within its customer base. 

What’s Eating Jira and Confluence Users?

But the company has not been infallible. For example, Atlassian acquired Hipchat in 2012. This was literally Slack before Slack. Companies like Uber were using Hipchat but defected to Slack as they scaled. The founders chalk the miss up to the fact they did not focus enough on the opportunity, in fact, they were practically giving Hipchat away. Also, the product wasn’t great – probably a relevant fact. 

Arguably, the more concerning thing for Atlassian is the qualms modern companies have with their core collaboration suite (e.g., Jira and Confluence), which represents ~60% of the business. Jira and Confluence are flexible products with 20 years of features incorporated into the product. The product is built to be everything for everyone, everywhere all at once. An infinitely flexible tool. 

There are advantages to this, such as allowing the company to serve other personas and (once implemented perfectly) can serve more types of companies. This has allowed the company to expand Jira into an “ITSM” platform to serve mid-market customers and compete with ServiceNow. It’s become a platform that other productivity tools companies build into.

There are also disadvantages to this. Jira and Confluence are very complex. If a company wants to implement Jira properly they typically need a trained Jira administrator that can scope the implementation perfectly, execute and handle the change management (i.e., make sure everyone follows the process). For many companies, it’s not worth it, or worse, they do it but other shadow tools pop up within the organization. This creates a steep learning curve for the people implementing the product but also for the organizations.

People have to be trained on how to implement and effectively enforce Jira/Confluence norms. This is not too dissimilar to the dynamic we see today with Salesforce’s Trailblazer Community6, there’s a generation and a half of IT managers, PMs, etc. that have grown up using and managing around Jira and Confluence.

The solution, ironically given the company’s self-serve roots, has been to leverage the agency/partner network to help companies implement the product for larger customers.

There’s even a vibrant community of Jira users that will tell you that they hate using Jira and Confluence. 

Over time, that has created the (perception and) opportunity for companies to effectively unbundle or build more end-user-friendly versions of Atlassian. If you’re driving on the 101 towards SFO, you’ve no doubt seen this. The Marina Trench could likely be filled with the Billboards that have been used to advertise the latest venture-funded project management tool. This is exemplified by the growth in the number of listed competitors in the company’s 10-K. 

The main way that these upstarts gained traction is by counter-positioning and marketing to “non-technical” teams. Today roughly 50% of Atlassian's seats are with “non-technical” teams i.e., marketing, finance, sales, H.R. For Asana et al, while expensive, the strategy has been paying off. 

Recently, the most notable entrant is Notion, which unlike Asana et al is attacking the Confluence use case i.e., the company knowledge base. Today, most early-stage startups and founders default to Notion as the go-to Wiki. Finally, even though Atlassian lists Asana et al in the “Business Teams” category both Asana and Monday.com have made in-roads in serving the broader developer community (albeit limited) and startups like Linear are becoming increasingly popular with developers. 

And of course, while excluded in the chart above, it’s hard to forget about Microsoft looming in the background. 

Competition is clearly increasing in the market, especially for the non-developer use cases. How defensible Atlassian’s business and its long-term market share will be determined by a few things:

• Their positioning as a “value” product. Compared to its newer peers, Atlassian’s offering is more feature rich and typically cheaper, which especially during downturns should allow the company to take share7.

• There’s a bit “Trailblazer” aspect to Jira, because of the company’s relatively steep learning curve, Jira users and administrators are typically reluctant to learn a tool or system of project management simply for the UI/UX benefits.

• Their ability to transition to the cloud effectively. One of the shortcomings with Atlassian’s product over the last ~10 years has been that their cloud product was simply not very user-friendly, in part due to some architectural differences between their platform and their peers8.

In software, the impact of new competition is usually felt in terms of a company’s ability to acquire new customers. Unless the software is not mission critical or a SOR, churn tends to be a very lagging indicator. In some ways, stickiness is table stakes. More often, the impact of competition is felt in two places: (1) new customer additions and the (2) cost of acquiring those customers. So far Atlassian’s CAC paybacks have been holding up well even with the move up-market. 

New customer adds are a bit noisy for Atlassian given the number of products they have but a near-term headwind has been that their free customers are converting at a lower rate than historically. That’s likely due to the economic conditions as companies have pared back headcount growth.

Fortunately for Atlassian, its competitors are also faced with one major headwind: they have to show improving economics because of investors’ increasing focus on efficiency. The strategy that its collaboration software peers have leveraged over the last few years may not be available in a non-ZIRP environment because, well no one will want to fund that level of capital burn (with the exception of a billionaire ex-founder of the company formerly known as Facebook).

That’s one of the reasons Atlassian continues to roll out more features/products at prices meaningfully below the competitors. The founders said as much recently:

​​Turbulent markets provide an opportunity to shake up the leaderboards. We'll focus our investments to take share and strengthen our market position in this environment. We'll, of course, balance our continued investments with the overall growth of our business and be responsive to the macroeconomic conditions.

The flip side is that the company has fallen prey before to better UI/UX products like Notion that were able to improve their product (see: Hipchat).

What’s next for Atlassian? 

To set up the company’s stage of growth and improve its defensibility, Atlassian has made a few bets: (1) transition fully to the cloud, (2) double down on ITSM (again) and (3) push into the enterprise.  

The most important near-term goal for the company is to transition to the cloud delivery model and more importantly, reap the product expansion benefits of the cloud. The other two bets are really contingent on the cloud execution.

Today, only ~60% of the company’s revenues are delivered in the cloud model (up from ~40% two-and-a-half years ago and 25% at IPO). As the company continues to push (literally push in some cases) more of its customers to the cloud and those customer cohorts mature, the benefits should begin to show up in the durability of the company’s revenue growth, which is required for a business trading at ~10x NTM revenues.

In the next post, I’ll cover this cloud transition, its importance, the company’s ITSM bet, the recent product announcements (yes, there’s some AI), and of course, valuation. 

Recommended Reading

If you are interested in learning more about software businesses, especially from an operator perspective, check out OnlyCFO - a newsletter about software, finance, and startups.

Subscribe

writes incredibly thorough and entertaining pieces such as "How to Trick Investors & VCs - Understanding finance tricks in metrics and reporting” and “How to Read Balance Sheets - Software Edition”

Subscribe now

Interviewer: So when you put your investing hat on, what's at the top of your mind?

Druckenmiller: I would guess the top of my mind is just how uncertain for me trying to analyze the environment is going forward. I've been doing this for 45 years. I've studied a lot of economic history, but I've never had a situation where you had free money for 11 years, a very broad asset bubble followed by jacking up rates 500 basis points in 12 months.

So for someone like me who likes to look at history and come up with potential scenarios. This is a particularly difficult period. Two years ago wasn't so difficult when you had two years at 15 basis points and money's probably growing at 30. It didn't take a genius to figure out that was a good risk-reward.

These days things are a lot more complicated, but you just mentioned about everything on our mind; generative AI, inflation, asset bubbles, currencies, all the stuff we usually think about.

Interviewer: So, so how do you navigate it then?

Druckenmiller: Well, luckily I don't have any clients, so I don't have, uh,

Interviewer: You see, I have one client and it's pretty big...

Druckenmiller: It's really big. I don't have to, I don't feel performance pressure. So historically I deal in five or six asset buckets, which tends to keep me out of trouble in terms of playing in an area where I shouldn't be playing at a particular time.

And I think it's one of the most important things to do is not to play when you don't see a fat pitch. I don't see a fat pitch. In fixed income is extremely complicated because I'm in the hard landing camp. Probably be sometime later this year.

But again, this is so complicated. I'm not willing to place a big bet and even if I believe in a hard landing, which I do, what do you do with two-year Treasuries at sub-4% with Fed Funds at five and a quarter, I better be right on my hard landing if I want to own fixed income. And then in the longer end, historically this is easy if I believe we're hard landing, I'm supposed to own bonds, but they're not exactly a screaming bargain. [The] 10-year is at three and a half in the U.S., particularly with a Fed that has certainly shown some mettle in the last year. But, historically, I wouldn't say Jerome Powell is a profile in courage.

So if we get into a hard landing and he moves aggressively, I could see bonds and inflation, coming back with vigor from what I expect to be a lower level than right now.

I will say that, the response to Silicon Valley unnerved me a little because in four days, they printed enough money, [that] they basically wiped out the entire reduction of the balance sheet they had done, for five or six months.

So if I'm trying to look ahead and anticipate I don't have a lot of faith in these guys, should we get into a hard landing that they're gonna hold the line and not do something maybe worse than Arthur Burns?

Interviewer: And, and where does it leave equities?

Druckenmiller: I think equities are really complicated. I think within the equity market, if put a gun to my head, I'd be short the economy to the extent I should [short] something pure like Russell 2000.

Obviously, I don't wanna go into individual short names, but names, like that old economy, economically sensitive stuff. But, let's just say we're gonna have a hard landing and a bad recession in the U.S. What does that mean for Nvidia? I don't know.

I mean, oils and chemicals went up in '73 and '74. Staples have gone up in bad recessions in the U.S. historically.

What do I do if with a company, if you have a bad recession in the U.S. but it's growing wildly throughout that period because we have an arms race going on in its space? It's not clear to me, it goes down.

So I think the equities are complicated. I mean, I'd say the one area, the way that I, I feel reasonably comfortable in is, I'm short the United States dollar. Currency trends tend to run for at least two or three years.

We had a long one here, over 10 trillion - something like 13 trillion - came into the U.S. dollar during the previous decade. I will say full disclosure, I missed the dollar, probably the biggest miss in my career in, in currency trade. I missed the last nine months’ run-up in the dollar. I just couldn't bring myself to own Joe Biden and Jerome Powell.

But, I think now that on a relative basis, the tightening in the U.S. going forward will not be as much, as the foreigners now that we've weaponized the dollar. And you've got people like Lula, running around asking why we need to be trading in U.S. dollars. By the way, it's not a bad question.

Historically, we could be trusted. We had a rule of law, a lot of things. So the only space I have any risk on right now is in the US dollar. Don't run out in short dollars, I could change my mind in a week. But, that's where I am right now. And I'm also long gold obviously for the same reasons.

Interviewer: So you say you haven't seen it as problematic for 45 years and it's really complicated and so on. So what do you do then? Do you sit in cash?

Druckenmiller: Oh no, I'm too, I'd like to, I'd like to say I had, I sit in cash, but I'm too much of a junkie. I'm always doing something. I'd say, our equities were about 3% net short, which is nothing.

But again, I've got, there's always equities we like versus other equities. Our fixed income position is minimal except in JGBs where I don't know whether I'm gonna get paid or not. But I think the risk award is, is ridiculous. It, reminds me a little, just a little of the two-year, two years ago. They have an inflation problem, but again, I'm dealing with government action.

So it's not cash. We got stuff going on, but, I'd say my P&L doesn't move generally more than 30 or 40 basis points a day. That's how buttoned down I have what I'll call our matrix here. Our matrix is the buckets I talked about in the investment within those buckets.

Interviewer: You'd be pleased to hear that I'm, I'm joined by roughly 200 junkies, here in Oslow, who also think it's difficult to not do much.

Now, tomorrow we are releasing a podcast with Steven Schwartzman from Blackstone, and he has one of his 25 rules " Not to lose money". And so I mentioned to him, I said, he's not a bit of a lame advice. It's a bit obvious, but you are kind of in the same camp, aren't you? I mean, I think you never had a down year.

And you are super keen on not losing money. Why is that so important? I mean, again, again, it's a lame question, right? But it's actually not that lame. But, let's listen to you.

Druckenmiller: It's just mathematics. If you go down 50, you gotta go back a hundred to get it back to even. And I've always thought the way to build a long-term track record is when you really see the ball, swing really big. And when you don't see the ball, don't swing.

And if you can build a record and when your terrible years, you're up zero to five, and then throw a couple of fifties and sixties in, the numbers look pretty good over time. If you make a bunch of 30s and then you lose 55 or 60%, you got a long, long way back. It's just the way the numbers work.

And also I'm a sore loser. I don't like to lose so that kind of helps out too.

Interviewer: Just remind us of your compounding.

Druckenmiller: When I had clients, which is the only audited record, it was just a little bit over 30% net a year for 30 years.

Interviewer: Well, that's pretty good.

Druckenmiller: But again, the top year was 99%. I had a lot of five and seven in there. I didn't like, like 20, 30% every year. It was a matter of never losing and then throwing some big numbers in there, maybe 10 times.

Interviewer: Well that sounds like a plan for the fund here as well now.

When you invest, how important is the analysis and how much is now gut feel or pattern recognition?

Druckenmiller: Yeah. I don't know who, but I heard a saying with analysis comes paralysis or Soros used to say, "Invest and then investigate", which I was already doing before I met him.

But, it's more important now, even than it was then. We're in such a fast-moving world with all the new communications that if I get an idea and I think it's attractive and for whatever reason, that security price will be higher in a year or two, I generally go ahead and buy it, and then tell the analyst to look into it.

And if it turns out I was wrong, after they analyze it, I get out. I don't like to wait around.

I've a lot of my best ideas, I'm not that smart. So if I see it, whatever's going on to cause that idea to happen, someone else might see it. And by the time we get done analyzing, I will miss 30 or 40% of the move, and then I'm paralyzed because it just went up 30 or 40% and I don't have the guts to buy it, even if I think it's going higher.

So we're more in the camp of if we got a strong feeling, we'll cut the analysis short and then, by all means, do our analysis thoroughly and then just unload it if it turns out my thesis was wrong.

Interviewer: Do you depend more on this type of pattern recognition than you did in your early days?

Druckenmiller: No. I started in the seventies and my mentor. Turned out to be a chartist. And, I still use that (a) to find things others may not be looking at because we use rate of change, stuff that tends to lead. And (b) it's, if I do take a big position, it's a great check on me not falling in love with security or falling in love with my idea.

You can always find 20 [stocks] that have a good chart and a good fundamental story. So if either one didn't fit, I wouldn't do it.

So, the technical provides a discipline on the fundamental and the fundamental provides a discipline on the technical.

Interviewer: Now, I believe, that one of the hallmarks of a great investor is to have to be stubborn, but at the same time, agile and being able to change your mind, and you have that incredible combination.

Now, how do you question your own beliefs?

Druckenmiller: Price action's important. I try to have young people around me who are not afraid to speak their mind to me and argue with me. If somebody's here too long and they agree with everything I say, they're not here that much longer because, you know, I need healthy debate. But you know, we're constantly following the news. We're following the fundamentals and also the price action.

I will say since the algos and everybody got involved in factor investing and whatever's going on these days, price action is not what it used to be 20 or 30 years ago. 20 or 30 years ago, if a company reported a horrible result and it opened down, say 10%, and ended up on the day, you could almost guarantee that stock was gonna be higher in three months.

It's not the case anymore. Everybody seems to know these tricks. I've seen that action and then a month later the, the stock's right back down again. So a big part of my process, which is price action versus news, I obviously still watch it.

If I've got a thesis and it's really bullish and it's playing out and the stock's not going anywhere, makes me go back and check the thesis over and over.

But price action versus news is not what it used to be 20 or 30 years ago probably because the competition has learned the same game. Whereas when I got in the business, you didn't have that many people that had it as that big a part of their process.

Interviewer: Is that the passive capital going into the ETFs, or is it actually the algo traders and so on?

Druckenmiller: I don't think it's so much the passive, I think it's very active hedge funds that learned the same stuff doing it. And obviously these, these factor investors, algos and quants, they'll really mess up, what used to be historical price action versus news and its message.

A lot of people in my position have complained about them. I, you know, to me it's just a new world I gotta deal with. People gotta deal with me. I might as well have to deal with them.

Interviewer: Lastly, Stan, if you were to leave us with some good advice here, how are we going to make the most money for the next two years?

Druckenmiller: Oh, boy.

I said something that the press ran with eight months ago, maybe it was a year ago, that, I wouldn't be surprised if the stock market was not higher in 10 years. I still believe that, but I do think like the '68 to ' 82 period, we'll have some big swings. So I think the way to make money in the next two years in the equity space is to be patient because I do think we have possibly some rough roads ahead. And I do think the central bank will respond in some crazy way that will give you a period like '70 to '72 where you can make money, or '76 to '78 when you could have made a lot of money. And also for me at least, I think the currency markets are very interesting, but I think this is a movie I've never seen anything.

So I'm gonna be very careful not to dig myself in a hole when I don't have a strong belief to come out because I think the opportunities are gonna be amazing as this movie unfolds in the next year in macro and in equities.

Yesterday, I shared on Twitter the resource page I published for understanding and investing in internet, marketplace, and media businesses. Since not everyone lives on Twitter, I wanted to quickly share the link and list here as well.

If people have other suggestions, feel free to DM me on Twitter or email me at ardacapital01@gmail.com. I’ll continue to update the page over time.

Subscribe now

Table of Contents:

1. Foundational Materials

2. Blogs, Newsletters, and Trade Publications

3. Other Internet Books

4. Media and Telecom Books

Foundational Materials

Strategy:

• Ben Thompson’s Platforms vs. Aggregators, Aggregation Theory, Defining Aggregators, and Spotify, Netflix, and Aggregation are critical to understanding Internet businesses.

• Increasing Returns to Scale by Brian Arthur (book and HBR article)

• Information Rules

Analysis and Trends:

• Mary Meeker’s Internet Trends Reports

• Mauboussin’s The Economics of Customer Businesses

• Tribe Capital’s Unit Economics and The Pursuit of Scale Invariance

• Tren Griffin’s The “Free Parking” Business Model

• Bill Gurley’s Above the Crowd (ATC). It’s worth reading through all the articles at some point but would highlight The Most Powerful Internet Metric of All, The Dangerous Seduction of the Lifetime Value (LTV) Formula, and How to Miss By a Mile: An Alternative Look at Uber’s Potential Market Size, in addition to the ones listed in the marketplace section.

Understanding Marketplaces:

• ATC: Money Out of Nowhere: How Internet Marketplaces Unlock Economic Wealth and All Markets Are Not Created Equal: 10 Factors To Consider When Evaluating Digital Marketplaces

• Perfect Store, The: Inside eBay

History of the Internet, FANG, and Others

• The Highwaymen

• How the Internet Happened

• Amazon: The Everything Store, Amazon Unbound, and The History of Sears Predicts Nearly Everything Amazon Is Doing

• Google: In the Plex

• Facebook: The Inside Story

• Uber, Airbnb, et al: The Upstarts

• Skift’s The History of Booking

Memos and Docs

The memos provide insight into the companies think about their businesses and strategy

• Zuckerberg’s Memo on Owning VR/AR Platform and Acquiring Unity

• Zuckerberg on Monetizing the Messenger Ecosystem

• Gates’ Tidal Wave Memo

• Amazon’s Deal Memo for Ring and Blink

Blogs, Newsletters, and Trade Publications

• Stratechery by Ben Thompson - covers/writes about the internet (understatement)

• Mobile Dev Memo - covers digital advertising, gaming, and the app economy

• Above Avalon - exclusively covers Apple

• Sports Business Journal - everything sports media related

• Entertainment Strategy Guy - covers streaming and media.

• Puck - Matthew Belioni’s Hollywood newsletter is a must-follow.

• Ben Evan’s Newsletter and Annual Presentation

• Skift - covers the travel industry

• AdWeek - covers agencies and advertising

• Octahedron’s Quarterly Review

• Prosus Capital Markets Days - Prosus’ investor presentations are interesting ways to learn about different business models (e.g., food delivery) across geographies.

Other Internet Books

• Engines That Move Markets

• How the Internet Happened

• Platform Revolution: How Networked Markets Are Transforming the Economy―and How to Make Them Work for You

Media and Telecom Books

Hard to separate the internet from media these days so including those recs here.

• Netflixed

• Cable Cowboy

• The Curse of the Mogul: What's Wrong with the World's Leading Media Companies

• Cutthroat: High Stakes & Killer Moves on the Electronic Frontier

Government Documents

DOJ and CMA documents have a lot of detail about the big tech internet businesses

• DOJ’s 2023 Investigation on Google’s Advertising Business

• 2022 - Investigation of Competition in Digital Markets

• 2019 - CMA’s Online Platforms and digital advertising market study

• 2018 - The Role of MVNOs in Cable’s Entry into Wireless

And ICYMI, here’s a link to the Software list.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Later this week I’ll be publishing a two-part piece on Atlassian, subscribe to receive it. Quick disclaimer: this is not investment advice, and the author may hold positions in the securities discussed

Generative AI models and their ability to generate high-quality drawings, portraits, and images are testing the boundaries of U.S. Copyright Law.

Artists and owners of IP have sued the creators of models for infringing on their copyright. The most prominent cases are Andersen v. Stability AI, Midjourney, DeviantArt, and Getty Images v. Stability AI. In both cases, the owners of IP are suing the model developers for copyright infringement. More recently UMG, the music label, asked platforms like Spotify to disallow the “unauthorized” use of their songs to train the models. 

On a related note, UMG has issued DMCA notices and asked the creators of AI-generated Drake and Eminem songs to take down the content from their platforms. 

The original YouTube link is taken down with a message reading, “This video is no longer available due to a copyright claim by Universal Music Group” left in its place. Copies of the song are still all over YouTube, though. Ghostwriter uploads another version today; it’s still up as of this writing.

Clearly, copyright owners are not happy and they’re attacking generative AI from two angles:

• Inputs: They are suing the model developers to prevent them from training their models on copyrighted IP. 

• Outputs: They are issuing takedown notices for AI-generated work and using the platforms (e.g., YouTube and Spotify) to police/enforce.

Copyright law and fair use, which is often used as a defense when sued for copyright infringement, in the U.S. is a bit like the Pirates Code; they are rules, but also more like guidelines and are designed to be applied on a case-by-case basis. Courts do not apply the guidelines consistently1:

As with any multi-factor test, courts have applied the fair use test inconsistently. Indeed, one of the most frequent complaints about the fair use doctrine is that it is incoherent and unpredictable. As one scholar famously put it, the fair use defense may boil down to the “right to hire a lawyer” because its unpredictability means that parties cannot actually rely on it ex ante, instead having to resort to ex post judicial adjudication before they know their rights.

The inherent uncertainty of copyright law creates FUD around using generative AI in business or commercial contexts for fear of creating something that violates the copyright. Incumbents such as Adobe and Shutterstock are starting to take advantage of the FUD to counter-positioning their offerings with the promise of compliance. Moreover, for Shutterstock, generative AI represents a more existential crisis so is using this moment to shift its business away from the quickly commoditizing stock images business towards a potentially more lucrative data business2. 

A copyright case (Andy Warhol Foundation v. Goldsmith) recently heard before the Supreme Court demonstrates some of the nuances in U.S. copyright law and illustrates the problem that Adobe and Shutterstock are trying to “solve” with their generative AI products and initiatives. 

Subscribe now

Copyright 101

Before jumping into the case, it’s helpful to lay some groundwork for the discussion around copyright law and fair use. In addition to the discussion below, I recommend reading Foundation Models and Fair Use, which researchers from Stanford recently released and is a text that I heavily reference in this post.

In the United States, as soon as someone creates something, they get copyright protection for it3. The “fair use” doctrine gives others the right to use copyrighted work under certain conditions without having to get a license from the copyright owner. Courts determine something is considered “fair use” is determined based on four factors:  

• “the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes” 

• “the nature of the copyrighted work” 

• “the amount and substantiality of the portion used in relation to the copyrighted work as a whole” 

• “the effect of the use upon the potential market for or value of the copyrighted work”4

The standards were meant to be intentionally flexible and this flexibility has allowed the Courts to apply copyright law on a case-by-case basis over the last 30 years. Over time, the Judicial system has introduced a number of “subfactors” to each of the factors listed above through Court Opinions. In 1994 Campbell v. Acuff-Rose Music, Inc. introduced one of the most important subfactors, “transformative use”, which the court defined as something that “adds something new, with a further purpose or different character, altering the first with new expression, meaning or message”. 

Since it was introduced, establishing if something is transformative has become a critical component of establishing if something is fair use or not. In “Is Transformative Work Eating the World?” researchers from BYU found that ~80% of Fair Use cases since 1994 have used the Transformative Use Doctrine in their opinions.

The researchers found that when transformative use is cited in opinions, fair use is established 50% of the time. Additionally, since many generative AI models were trained on copyrighted data, determining whether if the model outputs meet the “transformative use” bar will be important in determining whether or not they fall into “fair use”.

But the challenge has been that defining whether or not something is “transformative use” is highly contextual and critics (esp. copyright holders) have argued that it has been applied inconsistently and too broadly, effectively giving rise to unlicensed use5. 

The aforementioned, Goldsmith v. Andy Warhol demonstrates how different Judges’ Courts can apply transformative use. The images and photographs in question have many parallels to the images that users of generative AI are currently making. 

Goldsmith v. Andy Warhol Foundation

In 1981, photographer Lynn Goldsmith was commissioned by Newsweek to photograph an upcoming star, Prince (or The Artist Formerly Known as Prince). After being commissioned for the photographs, Goldsmith took photos of Prince at his concert and during a private studio session. Newsweek ended up using a photograph from the concert but none of the other photographs, which Goldsmith retained for future publication and licensing.

Three years later, Vanity Fair (“VF”) licensed one of the photographs from Goldsmith’s for $400 to be used in their article, “Purple Fame”, which covered Prince’s rise to stardom and hit album “Purple Rain”. The license allowed VF to commission another artist, Andy Warhol, to create a work based on Goldsmith’s photograph. Importantly the license that VF received had limitations:

Vanity Fair further agreed to run only one full-page and one quarter-page version of the illustration, which could appear only in the November 1984 issue. The license specified: “NO OTHER USAGE RIGHTS GRANTED.”

However, Warhol did not create just the one image used in the “Purple Fame” cover, he created 15 other images, now referred to as the “Prince Series”. When Warhol died, the images were passed onto the Andy Warhol Foundation (“AWF”), which over the years has sold, donated, or licensed parts of the series. 

Goldsmith discovered the Series in 2016 when Conde Naste used one of the prints for their Prince tribute cover story. Goldsmith sued6 and the case went to the Southern District of New York (“SDNY”). The SDNY determined that Warhol’s Prince Series did not infringe copyright. In his ruling, the Judge determined the Prince Series constituted transformative use because:

• Warhol represented Prince as “iconic” as opposed to “vulnerable” as Goldsmith originally depicted him.

• the image was cropped out to the torso and neck7, and 

• there was no market damage because the markets were different; the licensing market for Warhol’s, because of his distinct style, is different from the licensing market for Goldsmith’s photographs. 

Goldsmith appealed to the United States Court of Appeals, Second Circuit (“Court of Appeals”), which determined that the Prince Series did not meet the bar for transformative use and dismissed the lower courts’ arguments that the depiction of Prince, use of Warhol’s style and cropping of the torso as the basis of fair use:

[Regarding the first factor] the secondary work's transformative purpose and character must, at a bare minimum, comprise something more than the imposition of another artist's style on the primary work such that the secondary work remains both recognizably deriving from, and retaining the essential elements of, its source material. With this clarification, viewing the works side-by-side, we conclude that the Prince Series is not "transformative" within the meaning of the first factor…

[Regarding the second factor] While Warhol did indeed crop and flatten the Goldsmith Photograph, the end product is not merely a screenprint identifiably based on a photograph of Prince. Rather it is a screenprint readily identifiable as deriving from a specific photograph of Prince.

The Court of Appeals also made important clarifications on the market impact of the derivative work. 

There currently exists a market to license photographs of musicians, such as the Goldsmith Photograph, to serve as the basis of a stylized derivative image; permitting this use would effectively destroy that broader market, as, if artists "could use such images for free, there would be little or no reason to pay for [them].

The AWF appealed to the Supreme Court, which heard the case in August 2022 and the decision is still pending. The case is heavily followed by the legal community because it has far-reaching implications for fair use but also presents an interesting lens through which to look at copyright in the context of generative AI.

Style Transfer

One of the most common prompts and uses of generative AI is “style transfer” (“Draw [Image X] in the style of [Artist Y]”8), which is analogous to how Warhol imposed his style onto the 1994 VF painting. Generative models are very capable of doing style transfer. Here’s an adapted image from a Google blog post that touts their model’s ability to accurately do style transfer.

The Warhol case is important because the ruling will likely decide on the legality of style transfer: 

It is not yet clear how the Supreme Court will rule on this case, but its outcome will likely directly impact the scenario of style transfer in generative images. If the Supreme Court rules that Andy Warhol’s painting was fair use, then style transfer is more likely to be fair use9.

Taken to the extreme, in this scenario, the Court would effectively legalize the ability to superimpose an existing artist's style on existing work and monetize that work. However, styles themselves have historically not been copyrightable, which means the ruling would probably be net-negative for both owners of IP and artists, but net-positive for generative AI technology.  

We find that the most common named entity type used in prompts are people’s names, including the names of artists like Greg Rutokowski, who is referenced 1.2M times. This suggests that users in this community often try to generate images in particular artist styles, which is more likely to be fair use as long as the content itself is sufficiently transformative. However, there are other queries which specifically look to generate known works of art, which would tend towards more risk of losing a fair use defense if the model complies10

In this case, style transfer would meet the “sufficiently transformative” requirement. If you are Greg Rutkowski, you are probably not happy about this outcome and usage. A quick Google Search would confirm that indeed he is not.

Copyright Liability

The second potential outcome is that the Supreme Court rules in favor of Goldsmith and determines that style transfer does not represent significant transformation. And if style transfer does not represent fair use, there is still the question of who bears the liability for the infringement. In the context of generative AI, there are four actors:

• Model developers/creators e.g., OpenAI and StabilityAI

• Model deployer e.g., Drayk.it, Playground AI, Jasper

• End-user i.e., the person prompting

• Hosting platforms i.e., YouTube, Spotify

Who bears, liability and responsibility depends on the scenario:

If, however, the court rules that this is not fair use it is less likely that style transfer will be fair use without significant transformation. There are nuances here, however. If the user provides the original image to be style-transferred, the model deployer may be less liable (since this behaves more like a photo editing software). If the model deployer only takes text and it generates copyrighted Image X in a different style, then the model deployer is more akin to Andy Warhol, rather than a photo editing software.

The first scenario is relatively straightforward. If a user of Midjourney uses the product to create work that infringes on the copyright, the user is ultimately responsible, which is the argument that Stability AI and OpenAI have made in the past.

However, in the scenario that the model itself was designed for specific style transfer, the model deployer is acting less like a photo editing software and may be liable for copyright infringement. The second example part brings to mind, Drayk.it.

Drayk.it, a music generator created by Mayk.it, recently gained steam among fans of the 6 God, who have shared their creations after making them on the site and uploading them to social media. The software allows users to select a topic of their choice, which its GPT-3 will create a song about in the style of the rap star’s own hits. Each generated song will be performed in the voice of Drake and will span one-minute in length. For those stumped or indecisive about which song theme they should choose, the site includes an option to create random topics for you by clicking on its dice icon.

Unsurprisingly, Drayk.it was taken down after a few months.

In the Drayk.it example, assuming there was some commercial aspect to the usage, there are three parties that are potentially liable: 

• the model deployer (Drayk.it)

• the user, if they produce and distribute the Drake-stylized songs

• any hosting platform that hosts content produced by Drayk.it, which can include the website host, the music platform, or eCommerce platforms that are sold MP3   

Under The Digital Millennium Copyright Act (“DMCA”), online service providers are not liable if they take steps to police illegal content and importantly take down the content when they get a notice from a copyright holder. In order to mitigate legal responsibility, media and website hosting platforms implement DMCA policies that they adhere to in order to prevent lawsuits and the ire of intellectual property owners.

Copyright owners are already leveraging these platforms to take down AI-generated content. 

The underlying issue and driver of the uncertainty is the fact that these generative AI models were trained mostly on data that falls under copyright protection. 

The breadth of copyright protection means that most of the data that is used for training the current generation of foundation models is copyrighted material…most foundation models are trained on data obtained from webcrawls like C4 or OpenWebText. Since most online content has copyright protections attached at creation, using them for certain purposes could be considered infringement.11

If the models were trained only on licensed, publicly licensed, and/or domain content, we would not be having this conversation.

Enter Adobe and Shutterstock. 

Adobe and Shutterstock Generative AI Bets

Adobe and Shutterstock, owners of large libraries of stock images, video, and audio, have recently announced Generative AI products and initiatives.

The bets are different but the core positioning is similar; as the legal owners of the licenses to these stock images, the products and data they provide their customers are legal whereas the alternatives are not. Gracefully, they do not explicitly say the last part out loud.

Just kidding, they do. 

Shantanu Narayen Adobe Inc. – Chairman & CEO:

And there are a lot of questions that you can do, associated with that. I think we've been very transparent about the data that we've used. We've used Adobe Stock data. A lot of the other companies are actually using data that potentially -- they're scraping the Internet, they're accessing data that they may or may not have rights and license too. 

Adobe’s bet is to make their own models trained on their stock images and openly licensed images. From their Firefly announcement:

Adobe Firefly will be made up of multiple models, tailored to serve customers with a wide array of skillsets and technical backgrounds, working across a variety of different use cases. Adobe’s first model, trained on Adobe Stock images, openly licensed content and public domain content where copyright has expired, will focus on images and text effects and is designed to generate content safe for commercial use. Adobe Stock’s hundreds of millions of professional-grade, licensed images are among the highest quality in the market and help ensure Adobe Firefly won’t generate content based on other people’s or brands’ IP. Future Adobe Firefly models will leverage a variety of assets, technology and training data from Adobe and others.

Similarly, Getty Images, after suing Stability AI, recently announced that they are working with NVIDIA to generate models on their library of stock images. In February, I wrote about the challenges Getty Images faces in this new era and why is not as well positioned as Shutterstock to make the transition to a new business model.

Shutterstock on the other hand will leverage OpenAI’s model and will license its stock images to companies developing AI models such as Meta and LG. 

This content has created significant data feed and computer vision opportunities, opening up new monetization channels. In addition, we're able to commercially license this content by providing customers with the legal guarantees that they need. In 2022, we saw engagement with technology companies such as Meta, OpenAI and LG. But as we turn the corner into 2023, the size and shape of our customers has evolved to a diverse array of companies, representing technology, financial services, consultancies and other small businesses. Our most immediate use cases are generative AI, content moderation, public safety and interior design use cases.

Adobe and Shutterstock, are both marketing the same thing to their respective customers: certainty around copyright and license. In Adobe’s case, its customers remain artists/creatives whereas ShutterStock is adding a B2B channel where they license their data to the developers of generative AI models. In a hypothetical scenario where customers or developer is choosing between using Dall-E (developed by OpenAI and trained on Shutterstock’s data set) and another model (say one being sued by Getty Images), most companies will probably pick Dall-E to build their business around12.

Each company is playing to its respective strengths, Adobe’s core competency is software so they have decided to build their own models.

Shutterstock, which has historically built around content and its contributor network, is adding a spoke onto the flywheel to sell the metadata they are already collecting to model developers.

Shutterstock seems willing to accept the fact its content business (licensing stock images, video, audio) will be less important moving forward.

Paul J. Hennessy Shutterstock, Inc. – CEO & Director:

I mentioned the content engine [selling stock images, photos, etc]. It was the majority of our business and has been for quite some time, but I'm here to tell you that in short order in the coming years, we believe that the content side of our business will be less than half of our revenue compared to then the creative engine [self-serve editing tools and custom design] and our data engine. 

The company’s “data engine” is effectively selling the metadata (tags, descriptions) on their content (e.g., photos, audio) to companies that want to use both the content and the metadata to train their models. It’s a smart strategy for a company with a large contributor network that is already uploading new content and tagging the data13.

Moreover, it would be a gross understatement to say that the threat that generative AI poses is a much larger threat to Shutterstock than Adobe. Stock images are a relatively small part of Adobe's business, they are the entirety of Shutterstock’s business. So if generative AI does disrupt the stock images market, Adobe might be okay but Shutterstock this is potentially a terminal risk they have to navigate.

The company’s focus on “data” and goal of transitioning the business away from content is a manifestation of the situation. So in some ways, ShutterStock seems to have realized that their days of selling stock images are likely numbered. In other ways, they have not. 

Jarrod Yahes Shutterstock, Inc. – CFO:

Generative AI, metadata, data partnerships. Generative AI is getting really, really good, so good that it can even generate a replica of a tie from the mid-90s in purple Brooks Brothers on a CFO on a stage, incredible stuff this generative AI is capable of, fantastic. But I know 2 things that are even cooler than generative AI and cooler than metadata, and that's revenue and EBITDA.14

Additionally, Shutterstock is still pricing on a per-image basis (~$3 per image) for both stock and AI-generated images; this pricing model is unsustainable, especially for AI-generated images. Even stock image pricing is at risk in a world where the cost of generating a high-quality image is rapidly approaching zero. That being said, it’s also unwise to blow up your business model overnight, especially if you think the transition is more gradual. 

For Shutterstock, an interim saving grace may be that the majority of their business involves serving eCommerce use cases. In this case, the rational thing to do would be further the uncertainty around the legality of generative AI outputs and convince the users of your platform to only rely on your creator tools: 

Paul J. Hennessy Shutterstock, Inc. – CEO & Director:

…customers have peace of mind when it comes to licensing content that's created on our generative AI platform by virtue of being powered by properly, commercially licensed training data.

Also, our customers know that we're paying creators and have uniquely structured our contributor fund to compensate artists for the revenue generated by Shutterstock in connection with the licensing of generative AI content and of our training data…

All while, transitioning your business away from stock photos towards providing metadata to ML model developers. Shutterstock’s bet implicitly is that the duration of this change will give them enough runway to build a business that offsets the declines in the stock image business.

How interesting this opportunity is unclear because it’s an open question whether Shutterstock will be able to scale the data business with their current “marketplace” or if the business will effectively evolve into a professional services business. For reference, Appen is an Australian company that provides tagging services and metadata for ML use cases, the company generates ~$400M growing top-line mid-single digits and trades at ~1x revenues.

The challenge for Appen is that it is effectively a professional services business with very thin margins. However, if Shutterstock can repurpose its contributor network to tag data at scale, the business could have substantial margins and the transition story is more interesting. 

Realistically, the transition will take years. The more near-term opportunity around Shutterstock will likely come from overreactions to news surrounding generative AI. As discussed above, copyright law even without generative AI is fraught with judgments and decisions that take years.

Fading the bearishness (not hype, because that’s hard) that will inevitably take over the stock is potentially an interesting trading strategy but not necessarily investing strategy. 

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

About once every two weeks, as I am doom-scrolling Twitter, a thought crosses my mind:

Why would I, not short a semiconductor stock that is trading at 23x NTM Sales and is already regarded as the bellwether name for the newly forming AI bubble? Well, precisely because it might end up being a trendy or bubble stock. 

I am not the only one to have this thought. It's only been six months since ChatGPT was released, and there are already two distinct camps forming, we'll call them, Team Bubble and Team AI. For our purposes, let's loosely define a “bubble” as an environment where an asset class is priced such that no reasonable investor would expect an adequate return for the risk they're taking, and this behavior persists for a sustained period of time. 

Team Bubble is proactively cautioning that a bubble is forming and points to the wild price action of stocks like C3.ai, Buzzfeed, and NVIDIA (the 800-lb gorilla). Team AI will point to the rapid adoption of ChatGPT, the demand for AI chips, and a few potentially game-changing use cases as reasons to be bullish on the trend and hold stocks like NVIDIA. 

With this setup, an investor has a few options:

• Option A: Go long the AI stocks (e.g., NVIDIA) because they don't think there's a bubble and believe the fundamentals will catch up to the price.

• Option B: Short the asset because, well, you think there's a bubble.

• Option C: Sit on the sidelines and avoid the asset class entirely either because you do not want/cannot short. 

A narrative of bubbles is that unsophisticated investors (especially retail) spurred by pundits and FOMO will choose Option A. Institutional funds will choose Option B or Option C. At some point, the bubble will run out of steam, usually because of interest rates, at which point retail investors will be stuck holding the bag, and hedge funds will profit from their shorts. 

However, hedge fund trading behavior during the dot-com bubble (and other periods of euphoria) presents a slightly different narrative and provides a different model for how investors, especially institutional investors, may behave in our theoretical AI bubble. 

Hedge Funds During the Dot-Com Bubble

It turns out that during the Dot-Com Bubble, hedge funds chose Option D: go long the bubble asset class knowing that it is very likely a bubble1. From Hedge Funds and the Technology Bubble:

Over our sample period 1998 to 2000, hedge fund portfolios were heavily tilted toward highly priced technology stocks. The proportion of their overall stock holdings devoted to this segment was higher than the corresponding weight of technology stocks in the market portfolio. Relative to market portfolio weights, the technology exposure of hedge funds peaked in September 1999, about 6 months before the peak of the bubble. Hedge fund returns data reveal that this exposure on the long side was not offset by short positions or derivatives...

Compared to the overall market, hedge funds were overweight High P/S stocks (our proxy for bubble stocks). At the peak of the Nasdaq, 30% of the “Hedge Fund Portfolio” was invested in High P/S stocks vs. 20% for an overall market portfolio. 

The data contradicts one of the tenets of efficient markets – rational investors, which hedge funds are assumed to be, should trade against mispricing.

However, during periods of sustained mispricings, institutional investors do not necessarily have Options B and C available, which limits their ability to directly attack the mispricing2. Furthermore, because rational investors cannot attack mispricing, the asset bubble persists, and prices continue to increase. 

Literature on limits to arbitrage points out that various factors, such as noise trader risk3, agency problems, and synchronization risk may constrain arbitrageurs and allow mispricing to persist. 

For our purposes, let's (incorrectly) say the arbitrage in the context of equity bubbles is the valuation gap between the market and the bubble stocks4. The natural way to take advantage of the arbitrage is to choose Option B and short the bubble stocks and go long the market or some other stocks you think are undervalued. But the longer the bubble persists, the riskier this strategy becomes. You run into the risk of margin calls and/or redemptions because of underperformance driven by the short positions and your longs that definitely will not be doing as well, because, you know, they're not bubble stocks.

So to go short during a bubble or these story stocks, it's not enough for you to believe the fundamentals do not support valuation, you need other market participants to agree and act in unison to drive down the price. But coordinating this behavior across investors is challenging, creating what academics refer to as synchronization risk.

Synchronization risk arises from an arbitrageur's uncertainty about when other arbitrageurs will start exploiting a common arbitrage opportunity. The arbitrage opportunity appears when prices move away from fundamental values. In a market populated by rational investors this arbitrage opportunity will be immediately explored. However, the recent literature provides evidence that small noise traders can drive prices significantly further from fundamentals...If arbitrager cannot predict investors’ preferences towards the stock, or for how long the market is going to be optimistic about the stock, the arbitrage becomes impossible. Synchronization risk, therefore, leads to market timing by arbitrageurs and delays arbitrage. (Blackburn et al)

Activist short sellers, like Muddy Waters, deal with this issue by publicly issuing well-researched and pointed reports as to why a company is overvalued or fraudulent. More recently, Kerrisdale Capital attacked C3 AI on issues related to accounting, causing the stock to drop ~25%. But the activist short playbook is usually limited to one stock. No one investor can move the whole market; unless it's Powell5. Because of this synchronization risk and the risk of being liquidated for underperformance, hedge funds are reluctant to short bubbles when they see them unless there’s a clear catalyst6. 

What about Option C - just avoid the technology stocks altogether? 

Some hedge funds decided to pursue this route, at least for the time, but the overall industry did not. The risk funds face is that by not owning well-performing stocks, they underperform the benchmarks the limited partners (“LPs”)  set for them. LPs will tolerate the underperformance for a period, but the longer the bubble persists, the higher the risk of redemptions grows. The risk is more pronounced for mutual funds, which tend to have a more retail-oriented investor base that is inherently more flighty. Naturally, the way to avoid this risk is to have some exposure to well-performing stocks, regardless of valuation. This introduces some agency risk, which can be more pronounced during bubbles. 

This dynamic was one of the reasons Julian Robertson’s Tiger Management closed shop in 2000. At its peak in 1998, Tiger had more than $20B under management. But for the final two years, Robertson shunned the overvalued tech stocks and decided to invest in value names. Unfortunately, the value stocks did not do well, and worse, technology stocks continued to outperform. 

More Money Than God:

Rather than fight a bubble he viewed as absurd, Robertson decided to ignore it....

Roberston concentrated his bets on tradition value stocks such as Federal-Mogul Corporation, an auto parts supplier, and Niagara Mohawk, a power company -- the very essence of the old economy...

By the summer of 1999, Robertson's decision to ignore the tech boom was causing a crisis on Park Avenue. Tiger had lost 7.3 percent in the first half of the year; meanwhile, technology-heavy mutual funds were up a quarter or more, and day traders operating from kitchen tables were outperforming Robertson's special-forces unit. Coming on top of the losses on the yen in the fall of 1998, the latest setbacks strained investors' patience: Having withdrawn a net $3 billion from Tiger in the six months to March, Robertson's partners withdrew another $760 million at the end of the second quarter.

Unable to fight back on the withdrawals, among other factors, Robertson announced on March 30, 2000, that Tiger would be closing its doors. 

To avoid a similar underperformance because of not having exposure to the well-performing stocks, funds elect to choose Option D, which is to go long the technology stocks and ride the momentum. We know that hedge funds broadly recognized that this was a bubble because many said as much (though some disagreed), and in early 1999 there was even a failed “attack” on the bubble. 

Following the buildup of an overweighted position in late 1998, hedge funds subsequently reduced their exposure. This is at least somewhat consistent with the remark of Soros Fund Managements’ then-chief investment officer Stanley Druckenmiller that they were “calling the bursting of the internet bubble” in spring 1999. As it turned out, this call was too early. The bubble did not burst yet. Then, within just one quarter, hedge funds increased the weight of technology stocks from 16% to 29% in September 1999. The market portfolio weights only changed from 14% to 17%

The failed attack on the bubble was costly for Druckenmiller, who was running Soros’ Quantum fund. 

At the start of 1999, Stan Druckenmiller, Quantum's supremo, had shared Roberton's conviction that tech stocks were too high; but he acted differently. Undeterred by the market's momentum, Druckenmiller had placed an unhedged outright bet against the tech bubble, picking a dozen particularly overvalued start-ups and shorting $200 million worth of them. Immediately, all of them shot up with a violence that made it impossible to escape: "They'd close one day at $100 and open at $140," Druckenmiller remembered with a shudder. Within a few weeks the position had cost Quantum $600 million. By May 1999, Druckenmiller found himself 18 percent down.

To climb out of the hole, Druckenmiller pivoted and, like others, decided to ride the wave. 

Whereas Robertson had no patience for investing on the basis of momentum, Druckenmiller was fully capable of following the fundamentals in one period and surfing the trend in the next one. In May 1999, Druckenmiller allocated some of Quantum's capital to a new hire named Caron Levit, who loaded up on dot-com stocks. The skeptic who had shorted the bubble now claimed aboard the bandwagon…

When Druckenmiller returned from Sun Valley that summer, he had the zeal of the convert. He allocated more of Quantum's capital to Caron Levit and hired a second technology enthusiast named Diane Hawala…

Levit and Hakala piled into the same sorts of stock that Druckenmiller had shorted in the first part of the year. They were "in all this radioactive shit that I don't know how to spell," Druckenmiller said later. The radioactivity did wonders for Quantum's performance…

In the last months of 1999, Druckenmiller made more from surfing tech stocks than he had made from shorting sterling eight years earlier. Quantum went from down 18 percent in the first five months of the year to up 35 percent by the end of it. Druckenmiller had pulled off one of the great comebacks in the history of hedge funds, and it had nothing whatsoever to do with pushing markets to their efficient levels. 

You can see the differences between Tiger and Quantum's approaches during the last two years of the dot-com bubble in their exposure to high-priced tech stocks. The flow of funds was tightly correlated to this decision. 

The Fall

Even though Druckenmiller was able to masterfully guide Quantum through 1999, he too would soon follow Tiger’s fate. In February 2000, Druckenmiller pivoted out of the bubble but, a month later, would pivoted back in, leading to disastrous results. 

Then on March 10 the NASDAQ turned, and many of the stocks that Quantum held fell faster even than the market. Druckenmiller himself had been a huge buyer of a firm called VeriSign, which lost almost half its value in a month…

“I knew I was dead,” Druckenmiller said later; and by the end of March Quantum had lost about one tenth of its capital. 

Quantum would end up closing in April 2000, only a month after Tiger Management. At the press conference announcing the closure, Druckenmiller summarized his view:

NYT:

Stanley Druckenmiller knew technology stocks were overvalued, but he didn't think the party was going to end so rapidly. "We thought it was the eighth inning, and it was the ninth," he said, explaining how the $8.2 billion Quantum Fund, which he managed for Soros Fund Management, wound up down 22 percent this year before he announced yesterday that he was calling it quits after a phenomenal record at Soros over the last 12 years. "I overplayed my hand."

March 10 would mark the peak of the dot-com bubble. 

And in lock-step, hedge funds would begin to exit technology stocks (see “1” in chart below). The bear market rally in June 2000, would bring some investors back in for a short period. For institutional investors, price action seems to be the best catalyst synchronization.  

Unfortunately for retailer investors, they kept buying. Both directly and indirectly through mutual funds and investment advisors, retail investors would end up increasing their exposure to technology stocks 2-3 quarters after the peak. 

Who Drove and Burst the Tech Bubble?

Interestingly, independent investment advisors buy large amounts during the two quarters following the March 2000 peak and account for most of the increase in institutional holdings after the peak…

Individuals, in contrast, purchase large amounts following individual stock peaks and during the year following the market peak in March 2000…

Our results directly contrast a world where sophisticated investors consistently move against mispricing, a central building block of market efficiency. Nor does the evidence support bubble models where individuals move prices while smart money (institutions) passively stands aside…

Individual investors actively bought during both the run-up and particularly the collapse, highlighting their relatively unsophisticated behavior in the stock market.

Recent Hedge Fund Trading

Everyone following the markets over the last few years could likely draw a lot of parallels. Institutional investors were riding unprofitable technology stocks for the last ~5 years, for good reason.  

Hedge funds increased their exposure in lock-step with the rise in unprofitable tech. Once the Federal Reserve began to raise rates, there was a big unwind specifically around Q4 2021 (A)7.  

And similar to the dot-com bubble, retail investors kept buying for another few months. 

Fading the Hype

If we overextend the lessons from the dot-com bubble and even recent trading behavior, the easy money on the momentum trade has probably already been made in NVIDIA (i.e., going long when ChatGPT came out). As evidenced by hedge fund trading behavior between 1998-2001, the incremental investor will sit on the sidelines until the direction of price action is more clear and, more importantly, the direction of the relative performance is clearer. 

The risk that investors run into if they go short in the face of this type of sentiment is that at some point in the near future, NVIDIA and related stocks8 may end up being the only stocks “that work.” Funds that are underperforming but do not have exposure to these stocks may be pressured to buy, further driving up the price of the complex. Companies seeing price action in the public markets will be empowered to announce more investments in AI, further increasing investors’ expectations for underlying demand. The underlying demand for these companies may even exceed expectations! 

This might sound like an endorsement to go long to ride a momentum trade and get long bubble stock, but it is absolutely not. Riding the momentum and exiting at the exact right time is a cursed strategy, which ultimately led to the fall of one of the best traders of all time, Druckenmiller. 

Enthusiasm for AI and supercycle in chips may simply sputter out, in which case, it’ll be easy for investors to ride NVIDIA’s stock down to a reasonable headline valuation.

But shorting on the valuation thesis alone right now is a daunting proposition.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author may hold positions in the securities discussed.

Databricks has been rumored to go public in 2023 and is one of the few software companies that can likely justify a premium to their last round valuation ($38B) even with the current valuation climate. The company provides a platform that helps companies manage their data, machine learning (“ML”), and artificial intelligence (“AI”) applications.

Given the backdrop and investor enthusiasm (obsession) with AI and ML, the timing probably could not be more fortuitous. A lot of organizations and companies will glom onto the AI and ML theme this year, but Databricks was a company that was purpose-built for managing big data and actualizing the promise of ML. The company’s recent efforts to evangelize a new data paradigm (lakehouse) have, in some ways, put it on a collision course for one of the darlings of the investment community, Snowflake. 

But the beginnings were much more humble.

The Birth of Spark

Databricks’ story begins in 2006 with the "The Netflix Prize" Netflix, then primarily a DVD rental company, announced a $1M prize for anyone that could improve the accuracy of the company’s recommendation algorithm by 10%1. Lester Mackey, then an undergrad at Princeton, wanted to compete for the prize, as did 50k+ contestants from all over the world. Mackey would form a team with classmates, David Lin and David Weiss called “The Ensemble,” which would eventually grow to 30 people. The contest would end up lasting three years, and mega teams would form over the globe. 

By 2009, people had made improvements to Netflix’s baseline model, but no one had matched the 10% requirement. One of the reasons that the challenge was so difficult was because the tooling at the time to work with Big Data and Machine Learning projects was not robust. And the dataset that Netflix provided was massive; the training dataset had 100M+ ratings on 18K movies from 480K+ users. At this point, Mackey had graduated from Princeton and started his Ph.D. at Berkeley.

Reynold Xin, Databricks Co-Founder: Lester was one of the students at a UC Berkeley AMPLab and he was a very competitive person. And plus, as a PhD student back in Berkeley, he was making $2,000 a month. And the competition has a word of mouth was actually a million dollars.

So Lester really wanted to compete, but he ran into a big problem, which is, it was a pretty sizable dataset. He didn't have the tools to transform them and he didn't have the tools to express the machine-learning algorithms on them.

One of Mackey's Ph.D. classmates was Matei Zahari (Co-Founder of Databricks). Zahari started his Ph.D. at Berkeley in 2007 and was doing work on distributed computing (using a network of computers to solve a computational problem or task). To help Mackey, Zahari built an early version of Spark based on his existing work. 

Matei Zahari, Databricks Co-Founder: By seeing the kind of applications that he [Mackey] wanted to run, I started to design a programming model that would make it possible for people like Leicester to develop these applications, and I started working on the Spark Engine, 2009.

Even though only 600 lines of code in the early days, Spark would evolve into a full-fledged processing framework designed for fast and efficient data processing and analytics on large-scale data sets. The early iteration of Spark allowed Mackey and his team members to work with their data more efficiently and iterate faster (more on this later). Mackey and his team were able to use Spark to iterate their solution to Netflix's problem for the final few months of the competition. 

Final, because the contest was a race, on June 26, 2009, a competing team out of AT&T, Bellkor, submitted their solution that met Netflix's criteria of improving the baseline model by 10% on a qualifying data set. Per the rules, Mackey's team had 30 days to submit a solution that could beat Bellkor’s. And they did (and just by an inch)! This meant the two teams’ algorithms would be tested against another data set. Both teams submitted the final versions of their algorithms in September, and they tied. Unfortunately for Mackey and Team Ensemble, their submission was 20 minutes past the deadline. 

From NYT: The losing team, as it turned out, precisely matched the performance of the winner, but submitted its entry 20 minutes later, just before the final deadline expired. Under contest rules, in the event of a tie, the first team past the post was the winner. “That 20 minutes was worth a million dollars,” Reed Hastings, chief executive of Netflix, said at a news conference in New York.

Though the BerkeleyAMP Lab could not celebrate helping win the Netflix Prize, the development of Spark continued. 

From Hadoop to Spark

At the time, Mackey was not the only one struggling with working with large datasets, nor was this the first time Zahari had come across the challenges of working with Big Data.

Zahari: I actually worked pretty early on in 2007 with early users of Hadoop such as Facebook and Yahoo! And so I saw basically across the users there were some common trends. First of all it enabled a new type of application working with these large data volumes on clusters, which was very valuable. But then second what happened is when as soon as people got started they ran into things that the engine couldn't do and the most common ones were actually machine learning which are these iterative computations that make many passes over the same data and interactive queries.

It's hard to talk about Spark without talking about its predecessor, Hadoop.

Hadoop was open-sourced in 2006 and took the (enterprise) world by storm. The pitch was simple, Google, Yahoo, and other tech companies were using data and machine learning to disrupt the world. And with Hadoop, every enterprise could do the same. The messaging was perfectly timed. Because of the proliferation of the internet and mobile, companies had troves of new data about their customers and users. However, companies were constrained by the physical capacity of discs and servers. Hadoop created a file system that could efficiently (by that era’s standards) distribute storage and compute across multiple devices and provided a framework to manage everything. The concept of “Big Data” was effectively enabled by Hadoop.   

An ecosystem developed around Hadoop. Companies like Cloudera, HortonWorks, and MapR emerged to commercialize and distribute Hadoop; they would help enterprises deploy Hadoop clusters on-premise and provide support/services for those deployments. The whole ecosystem got massive adoption and attracted a lot of venture funding. 

But Hadoop and the ecosystem ran into a few issues that inevitably led to its decline2. The fall was driven by a few things: coupling of storage and compute, programmability, and performance.

• Coupling of Storage and Compute: Hadoop was developed during a time when, because of network constraints, the philosophy was to take "compute to the data" As a result, Hadoop tightly coupled storage and compute, and each Hadoop cluster was also a server, which made it expensive for companies to scale their data lakes. 

• Programmability: it was hard for data scientists to work with Hadoop. The product was based in Java and would require copious lines of code. 

• Performance: Hadoop was not able to run queries efficiently and, more importantly, for ML use cases, could not do iterative calculations.

Hadoop made it easy for companies to store data, but it was expensive to scale the storage. And companies using Hadoop found that, like other "Big Data" projects, it was one thing to store all the data, it's another thing to actually use the data. 

Spark was built to address these shortcomings. Zahari would team up with a few others from the Berkley AMPLab (Ali Ghodsi, Andy Konwinski, Arsalan Tavakoli-Shiraji, Ion Stoica, Patrick Wendell, and Reynold Xin) to build the Spark project, which was open-sourced in 2010. The group would also become the co-founders of Databricks, which would commercialize the open-source Spark project. Ghodsi was the VP of Product in the early days and would become CEO in 2016. 

The Spark team made different architectural choices than Hadoop, the technical nuances of which are better explained by others3, but the key things were:

• Separation of compute and storage (i.e., cheaper to-scale data storage)

• Processed data faster and allowed for faster iterations over the data (i.e., better time to value and could be used for ML application)

• Was easier for data scientists to work with

Spark made managing Big Data faster, cheaper, and easier for companies and data scientists. The project was open-sourced in 2010 and took off among the data science and machine learning community. By 2016, Spark had surpassed Hadoop in terms of search queries4.

The post-script on Hadoop isn't that it died. The ecosystem consolidated; Cloudera and Hortonworks merged, and MapR would be acquired by HPE. These vendors' business model was challenged in a world where the cloud was commoditizing Hadoop, and the technology they were providing support was declining. The companies were caught on the wrong end of a technology shift.

And probably did not help because the companies were somewhat captive to the on-premise business models and were late to transition to the cloud. The HortonWorks’ S-1 gives some insight into the company’s thinking at the time:

HortonWorks S-1: Cloud approaches are constrained by performance and compliance limitations. Recent innovations in cloud computing have helped increase business agility and workload efficiency. While big data processing operates optimally when compute and storage are co-located, some cloud computing architectures encourage a decoupling of compute and storage in order to optimize delivery of computing power while minimizing the cost of storage. This means that big data processing must retrieve data from cloud storage separately for any computational analysis, leading to latency, throughput and other scalability challenges. Moreover, corporate and government regulations can inhibit the kinds of datasets enterprises are willing or able to store in the cloud. As a result of these limitations, enterprises are seeking new technologies to collect, store and access higher volumes, velocity and variety of data in a cost-effective manner, and to gain more actionable insight from their increasingly complex and growing data stores. Enterprises need to upgrade their data center architectures to enable them to bring large volumes of data under management and to process and analyze multiple types of data in innovative ways.

It's rumored that Databricks had a sliding door moment similar to the Blockbuster and Netflix story. At one point, the Spark team offered Cloudera and HortonWorks licenses to the project, but they were turned down. Databricks would have to commercialize Spark themselves. 

Databricks and the Worst Pitch Deck Ever

In 2013, the group got the opportunity to pitch Databricks to Ben Horowitz and a16z5. Ben Horowitz, recently recounted of the pitch and deck:

From Fortune:

“The graphics were terrible,” Horowitz recalls of the deck they showed him. “The ideas were somewhere between patronizing and insane,” he says. “It was a very unprofessional pitch deck compared to what we were used to, for sure.”

Ali Ghodsi, a Databricks cofounder who became CEO in 2016, laughed at the memory of it. “I think our first pitch was really, really bad,” he says.

Fortunately for the team, it wouldn’t matter very much. One of the team members, Berkeley professor Scott Shenker, was friends with Horowitz and had told him he thought cofounder Matei Zaharia was one of the “best distributed systems people out of academia in the last 10 years,” Horowitz recalls.

The pitch deck was indeed very bad by today's standards, from the deck:

But the core ideas of the pitch was simple: (1) Hadoop sucks, and Spark is 100x faster, (2) ML and Big Data will be a rapidly growing market that can be unlocked with Spark, and (3) We built Spark, and we're the best team to commercialize it.

Horowitz (unsurprisingly) was able to see through the poor pitch deck and led a $14M Series A in the company. 

Monetizing Spark

Unlike Hadoop, the Apache Spark collaborators commercialized the software themselves, and they had a big advantage compared to its predecessors - the cloud. Around the time Spark took off, companies were beginning to store all of their data and files (unstructured and structured) in the cloud (e.g., AWS S3). 

Databricks monetized Apache Spark by offering its customers a managed version of Spark in the cloud (effectively a SaaS offering). The managed offering makes it easy for customers to get started with Spark deployments and not worry about the maintenance and optimization of the clusters. From Day 1, customers can use the Databricks-managed Spark clusters to query, transform and process the data in the data lake (i.e., storage layer). In addition, Databricks provides other SaaS products on top of the Spark engine, including data visualization and data science notebooks. 

The cloud delivery model allowed Databricks to avoid some of the pitfalls that Cloudera and HortonWorks ran into. Most importantly, the business model was not constrained to offering professional services around someone else’s IP.

Databricks does not, however, charge a subscription fee for its offering. Every time a data scientist runs a query, transforms data, or trains an ML model, some processing power (compute) is required, and Databricks charges for the compute. Similar to Snowflake, Databricks uses cloud infrastructure, so the compute that Databricks charges include a surcharge on top of the cloud vendor’s pricing. 

Companies have the option of managing their own Spark cluster to avoid the surcharge, but it comes down to the classic build vs. buy analysis. Managing your own Spark cluster would require some engineering overhead, and more importantly, there’s usually a meaningful performance drop (2-3x) because Databricks, being the core developers of Spark, are naturally the ones that are able to best manage and optimize it. 

Databricks, the Data Science and ML Company

Given the focus on the data lake and the team's background/early experiences, the personas Databricks catered to (until recently) were data scientists and ML engineers who were typically working with these large datasets. The primary use cases were (1) large-scale data analysis and (2) machine learning. 

A few examples of ML and AI use cases:

• Detecting credit card fraud

• Healthcare companies using electronic medical records to discover new treatments

• And because time is a flat circle, the streaming service formerly known as HBO, uses Databricks to create its recommendation and personalization algorithms. 

To further support the ML engineers, Databricks developed another open-source project, MLFLow. MLFlow is used by data scientists and ML engineers to manage the full machine learning lifecycle. The ML lifecycle is analogous to the software development lifecycle. And just like software engineers have DevOps, ML engineers use MLOps to manage the lifecycle of an ML model from data analysis all the way to model deployment. 

MLFlow provides an interface for engineers to do all of this. There are a host of companies that now specialize in each part of the stack. For Databricks, that’s fine as long as those other vendors work with the MLFlow and Databricks ecosystem, which nearly all do. The goal with MLFlow is for Databricks to make it easier for companies to start, test, and iterate on ML projects on Databricks. Databricks does this to be as cheap and efficient as possible because that means more spend/budget available for compute, the core monetization for the company. 

Machine learning applications are compute-intensive (relatively), and Databricks' goal is to commoditize everything that's not directly tied to compute by open-sourcing it and offering a free alternative. Joel Spolsky wrote about this strategy about ~20 years ago:

But something is still going on which very few people in the open source world really understand: a lot of very large public companies, with responsibilities to maximize shareholder value, are investing a lot of money in supporting open source software, usually by paying large teams of programmers to work on it. And that’s what the principle of complements explains.

Once again: demand for a product increases when the price of its complements decreases. In general, a company’s strategic interest is going to be to get the price of their complements as low as possible. The lowest theoretically sustainable price would be the “commodity price” — the price that arises when you have a bunch of competitors offering indistinguishable goods. So:

Smart companies try to commoditize their products’ complements.

If you can do this, demand for your product will increase and you will be able to charge more and make more.

So it is not surprising that Databricks jumped into the LLM discussion recently, promoting the ability to create a ChatGPT clone using their platform for $30. 

Two weeks ago, we released Dolly, a large language model (LLM) trained for less than $30 to exhibit ChatGPT-like human interactivity (aka instruction-following). Today, we’re releasing Dolly 2.0, the first open source, instruction-following LLM, fine-tuned on a human-generated instruction dataset licensed for research and commercial use.

Even if it seems somewhat gimmicky, Databricks likely achieved its goal of capturing the attention of CIOs at Fortune 2000 companies that want to cheaply experiment with their own version of the ChatGPT. If enterprises can very cheaply train their own LLMs on Databricks, they can also run all future compute required to train, maintain and improve their own LLMs on Databricks.  

While the opportunity to enable companies to do ML and become the de-facto platform for AI/ML is a large opportunity in and of itself, Databricks’ ambitions don't stop there.

Stacking S-Curves, Delta Lake, and Lake House

Databricks’ bread-and-butter persona is the data scientist and ML engineer, but the company wants to expand beyond that, which means they need to evolve their product to serve traditional data warehouse use cases.

To understand why it's helpful to understand the difference between a data warehouse and a data lake. The data lake is a superset of all of an organization’s data, and the data warehouse contains a curated subset of that data designed for BI and visualizations. If an organization decides that certain data from the data lake needs to be in the data warehouse, it'll get cleaned, transformed, and loaded into the data warehouse, where business analysts can query and analyze the data. 

But there's a lot of other data that is better left in an unstructured format (text files, video, audio). For example, a healthcare company likely stores patient information (name, age, sex, address, and pre-existing conditions) in a well-defined table in the data warehouse. All of the doctor’s text and audio notes (which are very valuable) are likely stored as files in the data lake.  

Analysts at the company can easily query the patient data via SQL or create dashboards to help with questions such as “What is the distribution of ages of our patients?” or “What percent of our patients come to us with diabetes?”. At some point, a data science or ML team may get tasked with creating a predictive model (e.g., the likelihood of diabetes) based on pre-existing conditions and sessions with the physician. The team can now leverage Spark, and the unstructured data sits in the data lake to create the model. 

So both data sets and architecture are valuable for the organization. And while big data and AI/ML workloads are rapidly growing, the main way companies and analysts consume and analyze data is via the data warehouse. 

Subscribe now

Large enterprises (again, think Fortune 2000 and large tech companies) typically have both a data lake and a data warehouse. This creates a fragmented data tech stack, duplication of data, and spending at companies. Here’s Ghodsi talking about this fracturing of data and infrastructure created by the two tech stacks:

As you [companies] get started, you store your data, and you start asking some basic questions. The simplest way to do that and the best stack in the past has been to put your data in a data warehouse and then plug in a BI-tool like Tableau or Power BI. And then you get your dashboards.

As you want to move to the right-hand side and you want to ask about the future then suddenly that stack [data warehouse stack] doesn't work. So you have to completely redo it. Use a different technology stack you store your data in a data lake and then you start using AI technologies on top of it. And you hire different type of people. You hire data scientists for that. You hire analytics engineers or analytics folks and there's this big divide between the two [stacks]. 

To bridge the gap between the two tech stacks, Databricks, starting in the mid-2010s, began to develop Delta, which brought data warehouse concepts (e.g., ACID transactions, schema enforcement) to the data lake. The idea is to basically combine the best of the data warehouse with the flexibility of the data lake. In 2019, the company started to open-source the project. And in 2020, Databricks announced Delta Lake, offered SQL support, and began to evangelize the concept of a “Lakehouse Platform”. 

The Lakehouse Platform was designed to provide the best of both the data lake and the data warehouse. Of course, the natural implication of this is that Databricks is going after the incumbent Data Warehouse vendors (e.g., Snowflake). The company ignited a flamewar last when they published benchmarks that claim that Databricks' lakehouse architecture performs better than Snowflake and other cloud data warehouses.

Their latest research benchmarked Databricks and Snowflake, and found that Databricks was 2.7x faster and 12x better in terms of price performance. This result validated the thesis that data warehouses such as Snowflake become prohibitively expensive as data size increases in production.

Databricks has been rapidly developing full blown data warehousing capabilities directly on data lakes, bringing the best of both worlds in one data architecture dubbed the data lakehouse. We announced our full suite of data warehousing capabilities as Databricks SQL in November 2020. The open question since then has been whether an open architecture based on a lakehouse can provide the performance, speed, and cost of the classic data warehouses. This result proves beyond any doubt that this is possible and achievable by the lakehouse architecture.

Snowflake fired back, and there was a benchmarking war for a few weeks. Snowflake has not exactly sat pretty over the last few years, either. The company released Snowpark, which allows companies to work with data in Snowflake with Python, R, and Scala, which is to appeal to data scientists. 

In 2022, Snowflake announced that they have adopted Apache Iceberg, an open-source alternative to Delta Lake, effectively supporting the lakehouse paradigm (with some hedging). Just like Databricks, Snowflake makes money on compute so the company wants to become a platform for data scientists to create and manage ML models as well. Snowflake knows the AI/ML opportunity is massive, from the company’s investor presentation: 

Basically, Databricks started with data scientists and ML engineers and is moving towards serving business analysts/users, and Snowflake is moving in the other direction. This is effectively a reflection of the company’s respective backgrounds, Databricks’ team started with machine learning at Berkeley, and Snowflake’s team came from Oracle and the relational database world.

Will Databricks beat out Snowflake for the cloud warehouse mantle? Maybe. But in reality, the market is large enough that it does not devolve into a zero-sum game like what happened with Cloudera and Hadoop. Most likely, in the near term, Databricks and Snowflake will dominate their respective workloads, and the overlap will be at the edges. But there is some risk for Snowflake.

There is the tail-risk that the lakehouse architecture that Databricks is evangelizing ends up being substantial, and Snowflake is completely unable (or refuses) to adapt to the new paradigm. If that happens, it would be the second paradigm that the Databricks’ team kickstarted and used to bludgeon an incumbent. 

Financials, Valuation, and Beyond

Databricks is smaller than Snowflake but is (reportedly) growing meaningfully faster than Snowflake. If Databricks can (1) sustain its growth given the demand backdrop for AI and ML workloads and (2) get a slight premium to Snowflake’s current multiple (~15x), it might be one of the few companies that raised last year that can actually IPO at a premium (albeit somewhat modest and excl. IPO dilution). 

An IPO like Databricks is probably required to really thaw the late-stage IPO backlog. Regardless of the inevitable timing of an IPO and the outcome of the lakehouse war, the impressive thing about Databricks has been their ability to consistently stack S-Curves. That’s mostly a reflection of the team’s deep bench of technical talent and probably why Horowitz was able to look past the disasterclass of an initial pitch. 

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author may hold positions in the securities discussed.

Over investment cycles, the attractiveness of industries and asset classes waxes and wanes, largely driven by the flow of capital into said industries and asset classes.

Edward Chancellor describes this dynamic eloquently in Capital Returns:

Typically, capital is attracted into high-return businesses and leaves when returns fall below the cost of capital. This process is not static, but cyclical – there is constant flux. The inflow of capital leads to new investment, which over time increases capacity in the sector and eventually pushes down returns. Conversely, when returns are low, capital exits and capacity is reduced; over time, then, profitability recovers. From the perspective of the wider economy, this cycle resembles Schumpeter’s process of “creative destruction” – as the function of the bust, which follows the boom, is to clear away the misallocation of capital that has occurred during the upswing.

The software sector has ridden this cycle over the last ~10-12 years. For a while, subscription software was the darling. The subscription software business model has all the hallmarks of high-quality businesses: annuity-like revenue streams, low churn, high gross margins, and low incremental costs. All of which should translate to high-profit margins at scale.

For those reasons, investors have historically encouraged all software companies to invest aggressively ahead of those profits. The underlying assumption is as growth slows, the companies' margins should inflect. Ultimately this dynamic attracted capital, valuations increased, and competition for the key inputs (e.g., engineers). As a result, investment returns have suffered and will continue to suffer until companies rationalize their investments and there’s enough “creative destruction.”

Companies that have been able to prove their path to profitability in line with investors’ expectations for the business models will be OK, as will their investors. Those that haven’t are effectively stuck in a messy middle between growth and profit, at least for the next 2-3 years.

Growth over profits (at least for a while)

Mature software businesses, such as Veeva, Autodesk, ServiceNow, and even the much-derided Salesforce, have demonstrated that. The margin expansion/potential is most apparent when the companies begin to climb the "top" of their S-Curve. 

Companies earlier in their S-Curves should prioritize growth over profits (within reason). The underlying logic is simply that companies that prioritize growth will get to scale faster and, as a result, generate enterprise value faster.

Take this overly simplistic example, Company A and Company B. Both companies start with $100M of revenues in year 1. Company A will grow consistently at 20% and maintain 20% EBIT margins. Company B starts growing 50% y/y in Year 1 but with negative profit margins. Company B’s growth with decay in line with the rule of 80, but its margins will improve in line with the Rule of 40.

By Year 5, Company B is ~50% more valuable than Company A. 

This also implies that in Year 1, investors are willing to pay a higher revenue multiple for Company B than Company A, which historically they have. The premium that Company B gets vs. Company A is effectively contingent on what the growth decay investors expect for the company1.

But, investors do not want to be rug pulled in Year 5 by Company B. You don't want to get to Year 5 and find that company is far from being able to get to the 20% margins that you were hoping for. So investors used the Rule of 40, which said that revenue growth plus profit2 should be greater than or equal to 40%. And if you chart the profit margins vs. revenue growth for a broad set of software companies over time, you'll more or less see that profit curve. 

Companies that met or exceeded the rule of 40 were rewarded with a premium valuation – and still are. The premium is ~30-50% depending on the company’s growth rate. 

Losing the plot

However, over the last ~3 years, the weight (or attention) investors paid for the profit component of a company's growth diminished. And the market writ large focused more on revenue growth. In effect, every SaaS company was given credit for the long-term margin potential that they showed at their IPO and, more importantly, their ability to graciously achieve those margins. 

As a result, software, and tech companies more broadly, were empowered to prioritize growth over short-term profits. Since headcount, specifically AEs, is the primary bottleneck for growth for software companies, hiring accelerated. Furthermore, demand pull-forward from COVID made it seem that if you do not hire, you’re leaving growth on the table.

Funding these “growth investments” was cheap because (1) companies could issue equity if they were burning cash, and (2) employees could be recruited with stock comp packages. Employees are willing to take this equity because (1) the headline value is attractive and (2) they can sell the shares after vesting to public market investors, so it was effectively deferred cash payments.

For companies, this was kind of the rational thing to do? If investors are granting you access to cheap capital and your employees are willing to take the capital as payment, then you should, of course, issue as much of it as you can to grow your business. 

Incremental opex burden as a leading indicator

The major caveats to this line of thinking are:

1. Revenue/customers being acquired are accretive.

2. The investments you’re making are actually driving efficient growth.

3. Once the period of cheap capital ends, you are able to pare down or unwind the investments.

I’ll quickly touch on the third point first. For both traditional and software companies, the last point is almost 10x harder than what it looks like in a spreadsheet. Unwinding investments are hard for organizations when we’re talking about people, and it takes longer than investors would like it to. 

Analyzing (1) is relatively straightforward for software companies, you can look at LTV and CAC paybacks to get a sense. Unless there's an incredibly abnormal activity by the company or the customers, you can probably assume that the ARR acquired is value accretive to the company. 

The second point can be monitored with the Rule of 40, but for hypergrowth companies, the absolute growth rate can mask inefficiencies. And Rule of 40 is somewhat of a backward-looking metric. It takes 2-3 years for software companies to recoup the cost of acquiring customers (S&M and implementation), and margin expansion in any given year is heavily influenced by the rate cohorts are maturing. So looking at the incremental Opex spending to acquire new revenue/bookings can be a useful leading metric. 

This is effectively a derivative of the rule of 40. If a company (generally) follows the Rule of 40, the incremental Opex should decline as growth slows. So for most software companies, the incremental opex should drop precipitously as growth slows. 

And by the time the company reaches the “steady state” of growth, incremental opex should be < 50%, and in effect, incremental profit should be 50%+. But the reality has been more distributed. The table below shows the incremental opex for a sample of application software companies.

Given that all of these businesses have some degree of growth decay, the numbers in each row should be getting smaller. But that’s hardly the case, even if you exclude 2022, which was a period when everyone missed their plans.

You can bucket the companies into a few categories:

1. Well-managed, defensible, and high-quality software businesses (e.g., Veeva)

2. Companies that manage to the rule of 40 but are pursuing a durable growth strategy (i.e., very little growth decay), which means the incremental opex does not decline much. Historically both ServiceNow and Salesforce fell into this camp.

3. Companies that have approached the very top of their S-Curve but are either reaching for growth or facing increasing levels of competition – are the companies in the middle.

4. Companies that are either not managed well or do not have the levers to get to the Rule of 40 as quickly as investors expect them to (e.g., Coupa and Anaplan).

Companies in the first category are well understood and appreciated. 

Companies that fall into the fourth bucket can go one of two ways: (1) they become a zombie company that muddles around as a public company for a while, or (2) they get acquired by a private equity firm. The second is what has happened to Momentive, Coupa, Anaplan, and Zendesk. The low end of the acquisitions has been ~5x NTM Revenues for enterprise companies and ~3x for Momentive. So finding potentially take-out targets given the amount of PE dry powder is still an opportunity. 

The Messy Middle

The 2nd and 3rd buckets have the most potential for dispersion over the next two years because most of these companies seem to have “found religion”3 when you look at the incremental opex implied by their guides. That is to say, most of the companies are guiding opex that is in line with what you would expect if the company was following the Rule of 40. 

Unfortunately, that’s not enough! Even though most of the companies in the sample above are guiding to incremental opex, which are commensurate with their growth rates, they are far from the Rule of 40. That’s either because the absolute level of opex going into the growth slowdown was too high or the growth deceleration was too much.

To get back to the Rule of 40, companies need to either (1) re-accelerate growth, (2) maintain the level of “austerity” for 2-3 more years, and (3) and/or both. 

If the companies in the middle bucket that are trading at 4-6x4 can get to the Rule of 40, there’s a good chance their multiple re-rate 1-2 turns. How long will that take is the key question. Consensus estimates imply that for almost all of them, it’ll take ~3 years. Based on current estimates, that’s a ~1.5-2x return in three years with some dilution. Not a bad outcome, but carries a lot of execution risk and requires underwriting a few turns of multiple expansions. 

The other risk for companies that fall into the middle bucket, specifically the mid-cap companies, is that someone opportunistically acquires them before they can execute this strategy in the public markets. The Coupa and HMI back-and-forth is more or less indicative of this.

And as I’ve written about previously, it’s cleaner to make this transition in the private markets if you are the CEO. It’s even more attractive if private equity is promising CEOs that they’ll make money hand over fist. 

In summary, SaaS had its own capital cycle of overexpansion, and these companies are facing another 18-24 months of restructuring across the industry. It’s a different backdrop than what SaaS investors have experienced over the last ~10 years, and some of these set-ups (e.g., Smartsheet) might just go to everyone’s too-hard to figure out pile for a while.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author holds positions in the securities discussed.

One of the follow-ups to my recent post on Workday (from Portsea and DMs) was to the tune of: “Why has Salesforce not seen similar levels of operating efficiency as Workday?” Salesforce’s S&M as a % of Revenues is ~39% vs. ~25% for Workday. It’s a fair and pertinent question.

Beyond the simple (not simplistic) “Salesforce wastes,” there are two potential reasons for this:

• Salesforce’s growth has not decelerated to the same extent as Workday’s. 

• Workday’s growth is more efficient than Salesforce’s1.

For both points, there’s a “management” component and a “structural” component. The management components are the near-term investment choices made by the company. The structural components are a function of the end market and products, which are, of course, a reflection of choices made by management years ago. 

There’s definitely a “waste” component to this as well, but from the outside in, that’s obviously hard to analyze.

Salesforce’s Growth Persistence

From a financial model perspective, companies that are consistently growing Net New ARR cannot show meaningful operating margin improvements. Salesforce has been beating this drum for a while and has gone to great lengths to explain that to investors. 

“But what's the growth philosophy of this business [Trailblazer]? The growth philosophy is, we are going to sustain growth, and we are going to grow that incremental amount by 20% per year…There is no structural leverage in this company because the revenue growth is the same as the new business growth. It's not to say this company can't generate leverage. It's to say there's no structural model leverage in this company.”2

And if you were to model this out on a spreadsheet, this largely checks out.

Given Salesforce’s view of the market size and opportunity, management has elected to take the “Trailblazer” approach. The company historically set a goal of growing 20%+ organically, which meant there was not much structural S&M margin expansion over the last ~5 years.

In contrast, Workday’s growth curve has been more traditional, and during the same period, the company’s growth has decelerated.

Per Salesforce’s view of the world, Workday probably falls in the Explorer camp, and Salesforce falls somewhere in between the Explorer and Trailblazer categories. Still, given the scale of the company and the portfolio of products at the company’s disposal, it is still a bit of a head-scratcher as to why Salesforce has not been able to drive down Cost to Book (CTB), which is defined as S&M/Incremental ARR. 

The company offers a few explanations, namely that the shift in the business towards more international bookings hides some of the improvements in the core business. And that’s likely true. But also, given how at length Salesforce’s management has historically talked about the fact they’re unlikely to show meaningful sales and marketing leverage while the company is growing 20%+, it’s unlikely there was even a mandate internally to improve CTB meaningfully.

The net result is that Workday has been able to show slightly more S&M margin improvements than Salesforce. But not by that much! During this period, Salesforce’s S&M as a % of revenues decreased by ~10 percentage points vs. ~11 percentage points by Workday.

So how does Workday have that much lower S&M burden than Salesforce? The main reason boils down to the differences in churn rates.

The Cost of Gross Retention

Workday’s revenue retention is ~98% vs. ~93% for Core Salesforce (~90% for Core Salesforce plus Tableau, Mulesoft, and Slack)3. That difference in revenue churn can be very costly, especially when the company, like Salesforce, has specific growth targets. 

Take this simple ceteris paribus example to illustrate.

In this example, Company A and Company B are both trying to grow 20% YoY. Because of the differences in churn, Company A, which in this case is representative of Salesforce, needs to add 22% more Gross New ARR to meet that goal than Company B. Assuming the same “Cost to Book” for both companies, the S&M for Company A will also be 22% higher. And the Net Cost to Book (or Blended CAC as most public investors calculate it) is also 22% higher for Company A.

You can further extend this analysis by looking at the S&M costs of these two companies at different growth rates. 

The implication is that the company with a lower churn rate will structurally have higher long-term margins, and ergo, the value of their revenue stream and installed base is higher. The interesting thing about this analysis is that the gap between the two company’s S&M costs slightly increases as the growth rate decreases.

So even though Salesforce’s sales and marketing spend will begin to show leverage as growth decelerates unless the company can (1) improve CTB and/or (2) lower attrition, Workday will always have a lower S&M burden compared to Salesforce. 

This dynamic is consistent across a broader sample of SaaS companies. The chart below compares two sets of public companies: ones with revenue churn of less than 5% and ones with churn greater than or equal to 5%45.

Companies with less than 5% churn have an ARR CAC / Net Cost to Book that is, on average, 30% more than their higher churn counterparts. Additionally, these companies, on average, have a higher revenue multiple. This intuitively makes sense. Companies with lower churn will have structurally higher margins at steady state growth (~10%), and as a result, their revenue streams are more valuable67.

In effect, both the Cost to Book and NTM Rev Multiple are functions of a company’s gross churn, which in turn is a function of the company’s product and customer segment. 

All this leads back to the classic aphorism in tech, retention is the key to growth. And in an increasingly constraint-oriented market environment, retention is the key to efficient and durable growth. In 2023 and (maybe 2024 as well), the most rational thing for a SaaS company (both public and private) is to focus most of their energy and efforts on how to increase gross retention above new sales. There are three reasons for this:

• A weak demand environment likely means sales cycles will be long and conversion will be low → CACs will be higher.

• Investors will be more forgiving about weak sales in ‘23

• It sets the company up to re-accelerate growth on a healthier base once the buying environment improves.

Said another way, the LTV/CAC of customers acquired in 2023 will be crappy, so it’s better to invest in increasing the LTV of existing customers and future customers. Like all commentary from investors, this is easier said than done. If it were easy for a company to reduce churn, they would have likely found it and done it.

But for the few companies that may have levers to pull beyond contracting — this might be an opportune year to focus product investments on that (I know Benioff said he wished he’d invested more in the GFC).

Salesforce Gross Retention Opportunity

In my first post about Salesforce’s near-term opportunities, I highlighted churn reduction is likely going to be one of the major drivers for the business moving forward:

At this stage of the company’s lifecycle, the reduction in churn associated with the increase in multi-cloud adoption is going to be very impactful. Today, Salesforce has ~93% gross revenue retention, which is good but not exactly best-in-class, especially for a tool that’s often considered to be mission-critical for companies.

And the company has done that pretty well over the last few years…

…in part due to the company’s ability to drive continued multi-cloud expansion.

However, the lack of improvements in CTB for Salesforce probably lies somewhere here. My hypothesis is that to reduce churn, Salesforce has been acquiring and pushing GTM teams to drive multi-cloud expansion. To help drive multi-cloud adoption, reps have been giving customers meaningful/steep discounts to become “multi-cloud.” In effect, the benefits of cross-selling are being captured by the customers in the form of price.  

This was probably the right strategy, in some instances, Marketing Cloud. And probably the wrong in other instances, Tableau. Given recent comments around Genie / Data Cloud, we probably will see similar behavior there as well. 

The other implication of Salesforce’s go-forward focus on churn reduction is that the company will increasingly encroach into product areas that the company has historically allowed 3rd party vendors to own. Two reasons for this:

(1) Salesforce can find shorter-cycle “non-cloud” products to increase retention

(2) Microsoft’s bet around Dynamics, Viva, Teams, and the Knowledge Graph is pushing the Microsoft ecosystem into a direction where no one outside of the Salesforce administrator has to ever actually interact with Salesforce. This is obviously not a good thing for Salesforce, so the most natural decision/strategy is for Salesforce to increase the surface area of its UI and engagement. 

And given the Zeitgeist, the most natural outcome is for Salesforce to launch more LLM-driven products around their existing product suite for sales reps and marketers.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence.

Subscribe now

Given this is an investment-focused post, a quick disclaimer: this is not investment advice, and the author holds positions in the securities discussed.

As covered in a recent thread, Workday was founded by Dave Duffield and Aneel Bhsuri in 2005 after their previous company, PeopleSoft, was acquired by Oracle in a hostile takeover. Since then, Workday, now led by Bhsuri, has effectively been re-creating a cloud-native version of PeopleSoft. The company's ability to start with a clean slate afforded them a number of product and architectural advantages that have enabled the company to build a best-of-breed ERP for the services industry. 

Their Act 1 was to leverage the cloud advantage to dominate the HCM market. Act 2 was to expand beyond HCM to Financials and build the leading ERP for the services industry. While the company is still in the middle innings of Act 1 and early innings of Act 2, Workday’s Act 3 (Platform/Extend) is beginning to take shape. The combination of those three levers and the company’s investments in the channel provides a meaningful runway for 20%+ top-line growth for the next 5+ years. 

This post is a deep dive into the company’s product story and my investment case for the company, which current levels ($190/share), I believe, represent an attractive risk/reward for a 3-5 year hold.

Best of Breed Platform and a SOR Case Study

Workday has 1.5 systems of records (“SORs”): (a) Human Capital Management (“HCM”) and (b) Financial Planning and Management. The company started with HCM, payroll, and a basic financials suite. Overtime, via product development and some M&A, the company has expanded to build the ERP for the services industry. The core platform can be broken down into three components:

• Human Capital Management: products help companies and HR leaders manage their workforce through their entire lifecycle. Lifecycle management means employers can use Workday to manage recruiting (e.g., Job Board), hiring (e.g., Applicant Tracking), onboarding, payroll, learning, and development. These are all extensions of Workday’s employee system of record.  

• Financials: products that help companies and CFOs manage all of their financial and account processes, which includes Planning (Adaptive Insights), close automation controls, AR/AP, and expense management. The core value proposition for CFOs is that it enables them to get their teams out of spreadsheets and centralize all the relevant financial data in one place. 

• Platform Services: These are services and modules that support things like reporting, data management, and app building across both systems of record. 

ERP for Services Industries

Workday is primarily focused on services industries, which include (but are not limited to) tech, media, financial services, retail and hospitality, education, and healthcare verticals. This focus is out of both necessity and choice. 

Legacy ERP vendors (e.g., Oracle and SAP) are very strong in industries with high-degree of manufacturing complexity and direct costs1. While not technically impossible to build, capturing all the edge cases would have been expensive and requires a zero-defect approach. Additionally, the services industries were faster to move to the cloud, which of course, is the target market for Workday.

Finally and most importantly, in the services industries, human capital management is often the most acute pain point for the C-Suite. For these companies, the ability to effectively recruit, onboard, manage, and develop their employees is a strategic priority. Oracle and SAP have been slower to develop modern and cloud-native HCM suites. And even though the companies have built HCM suites, they have failed to compete effectively with Workday’s more modern platform and modules2. 

The net result is that Workday has consistently taken share across the board. 

• Workday has 20% of the Cloud ERP Market (+15 percentage points since FY ‘16)

• 50%+ penetration in Fortune 500 (+20 percentage points since FY ‘17)

• 25%+ penetration in Global 2000 (19 percentage points since FY ‘15) 

In addition to taking market share and winning new customers, Workday has consistently demonstrated its ability to launch new products that neatly fit the needs of the services industry.

Sustained Product Innovation Around Core SOR

Workday’s ability to continuously expand its product suite and consistently attach new products demonstrates both the company’s R&D capabilities and the power of owning the core system of record. The 1.5 SORs that Workday owns are HCM and Financials. 

I consider Financials to be “0.5” of the “1.5 system of record” mentioned above because control over the financial system of record is a bit murky. Workday sits on top of the accounting system of record — general ledger (e.g., Netsuite). The general ledgers are typically controlled by traditional ERP vendors such as Oracle, SAP, Sage, etc. Many of the traditional ERP vendors, especially Oracle and SAP, have comparable solutions that compete directly with Workday’s Financials Modules. Moreover, because Workday does not own the SOR, the company has to compete directly with other 3rd vendors on equal footing. 

You can see the impact of this competitive dynamic on the product attach rate numbers below. Modules such as expense and procurement have effectively hit a wall with regard to attach rate.

As I mentioned in a previous post, this can be a treacherous place to be if you’re an application software vendor. This is why Workday is investing in solutions such as Close Automation, Accounting Center, and Prism. Close Automation and Accounting Center are designed to be systems of engagement around the general ledger. And, Prism allows Workday customers to aggregate all the ancillary data that does not make it into the GL (e.g., “How many SKUs of xyz products did we sell?”).

These modules are designed to make Workday the core system of engagement for the “Office of the CFO” and effectively turn the GL into just a database for accounting.

There are potentially some upside drivers for Workday if the company can execute its efforts to commoditize the general ledger into simply a database – the main one being the ability to drive better attach rate on expense management and potentially treasury services down the line. 

Promising early execution of “expansion” motion

Putting aside the nuances of SOR competitive dynamics, Workday has consistently demonstrated the ability to drive product penetration and upsell, especially at the “land” phase of a deal. 

And Workday is just beginning to execute the “expansion” motion of its sales strategy. Historically only ~20% of net new ACVs were from existing customers, but that’s shifting… 

…which is leading to ACV growth among the existing customer base of ~45%.

Workday’s evolution is natural and mirrors the other best-of-breed SaaS software companies – Salesforce and ServiceNow. 

And Workday is relatively early in their journey in terms of driving the upsell motion. 

Co-CEO Carl Eschenbach at MS Conference

But I actually think there are some things we can do, and we are in the process right now of looking at our go-to-market strategy, and we're going to modify some things. We're going to sell differently.

An example is something you talked about, typically, at Workday, when a renewal would come up, that's when we'd add new products. Now when a customer wants to add new products, even if it's midstream of a contract, we'll open it up, restructure it, add to new products, and then we'll co-term their existing footprint. So we're doing things like that.

Over time, I expected that ~80% of Workday’s bookings come from the company’s installed base. Again, the impact of this should be that the company’s sales and marketing burden comes down. ServiceNow was able to take out 800 bps of sales and marketing costs. Obviously, Salesforce’s S&M leverage during the same period leaves much to be desired, but from a model perspective, as Workday transitions to more “expand,” the company should see meaningful leverage in the business. That “excess margin” will either be re-invested in the business to drive more growth or fall to the bottom line – either way, the pro forma FCF/share should continue to compound 25-30%+. 

The “expansion” motion appears to be one of the main focus areas for the new Co-CEO Carl Eschenbach. In addition to driving more sales within the customer base, Eschenbach is focused on investing in the channel, which Workday historically has not prioritized.

Channel & Vertical Investments

One of the major gripes companies have with enterprise software, especially ERP software, is that the implementation cycles are long, and change management post-implementation can be painful. Software companies do not want to be in the business of implementation and customization, so usually, those teams are under-resourced or expensive. In addition, given the importance of ERP investments, many companies want advice from someone who’s seen multiple implementations before.

Channel partners (e.g., Accenture) and systems integrators help fill this void in the market. These companies often help the company choose the right vendor for their needs and guide them through the deployment and integration process.

Historically, Workday has underinvested in its channel relationships. Early on, for a good reason, they had to compete with Oracle, which has its own in-house consulting team. But now that the company is going to be approaching $10B in revenues, the company needs to engage with these channel partners for two reasons: (1) it’ll help them reach more customers, especially in the mid-market, and (2) it’ll reduce the burden of having Workday manage these implementations.

Similarly, Workday has invested in vertical-specific SI relationships and products to help drive the adoption of the company’s financials products.

The benefit for Workday of the channel working are: (1) they’ll have to support fewer professional services mandates, which increases gross margins, and (2) higher deal velocity in the channel will lead to better S&M efficiency. For reference, here’s the evolution of ServiceNow’s S&M spend and Professional Services burden vs. the % of the company’s deals sold or influenced by channel partners.

We should expect a similar dynamic to play out with Workday, albeit with small margin uplift, given Workday’s S&M is only 27% of revenues vs. ServiceNow’s >40%. These go-to-market investments are effectively optimizations around the company’s core business, but Working is looking to expand itself even further with its Act 3, the platform services.

Transition to becoming a “Platform” and Developer Flywheel

Workday can be abstracted into three components: 

• Systems of record

• Platform Services: all the APIs, UI/UX components, reporting/querying functionality, etc. 

• Applications: the actual application with the business logic that end-users interact with. 

Workday Extend is the company’s app development platform that allows customers and 3rd parties to use Workday’s platform services to create their own 3rd party apps, workflows, and automation. Over the last five years, Workday has slowly been exposing these different components to customers and its partners.

These platforms are long-cycle investments because (1) exposing a large set of platform services in a controlled manner takes a long time, and (2) customers and channel partners have to invest a meaningful amount of time into learning the proprietary platform and tools. However, the investments are maturing. Workday hosted its 2nd DevCon last year. And there are now >3k Workday developers globally, and that number continues to grow.

It was interesting to note the following exchange at the most recent analyst day (which seems weird but also definitely scripted):

Workday 2022 Financial Analyst Day

Mark Murphy – JPMorgan Chase & Co

Great show. Thanks for having us here. So I wanted to ask you the first customer I bumped into today is with the Fortune 500 company, and I walked them to registration and then found out he works in the IT department of his company. So not HR, not financing. The company hadn't sent anyone from HR or finance to this conference.

So I'm just wondering how common is that trend where you see IT getting involved? And is that opening up a different type of -- kind of greasing the skids for broader and deeper discussions with some of these companies?

Aneel Bhusri Workday, Inc. – Co-Founder, Co-CEO & Chairman of the Board

Well, I'd say that since we started the company, we wanted to have a deeper relationship with IT. In many cases, some of our early customers were driven by IT, like Dave Smoley, Flextronics. But in the last 2 years, we made a concerted effort to really engage with IT, not just because they're involved with the projects, but because Extend is really an opportunity for us to work with them and have them put their resources into the projects with the extensions that they can build together. And so IT is coming, and now there's something that's really in it for them other than being somebody that watches out over the project.

Selling/marketing to IT is important because internal tooling and apps like the ones that Extend enable usually fall into their domain. The app development environment that Workday offers is relatively similar to the apps/workflows that the IT teams create using ServiceNow and Microsoft Power Automate.

While I do not think Workday intends to compete with ServiceNow, there are likely some ServiceNow HCM workflows that are likely subsumed by Extend.

Driving the adoption of Extend is important/beneficial to Workday for a few reasons:

Accretive to gross and (potentially) net dollar retention. The impact on gross retention is clear – companies building apps on Workday are going to make the platform stickier. Extend could also help drive net dollar retention (beyond Extend upsell) by driving wide usage across the enterprise. 

Channel partner and industry verticals. Channel partners and SIs can leverage Extend to create solutions specific to their customer’s use cases and vertical-specific requirements. For the partners, this provides additional service revenues. For Workday, it helps the company address customer/vertical use cases not directly covered by their core platform. For customers, these solutions should help get to value faster and help with change management. 

Workday Certified and Developer Community. Extend enables traditional developers and traditionally non-technical users to build their own apps and workflows. In effect, these users are electing to build skills and their careers around Workday’s platform. At scale, Workday’s developer ecosystem should have an impact similar to Salesforce’s Trailblazer program3.

Platform Flywheel. The combined impact of the aforementioned benefits is that at scale, Workday’s platform strategy should yield flywheel effects. 

• More Workday developers at companies, channel partners, and SIs leads to more apps built to address use cases and customer requirements that Workday does not already address. Over time, this should lead to More Workday Customers.

• More 3rd Party Workday Apps also means that Workday’s R&D teams see more customer use cases that are not addressed. Once there are enough proof points, Workday can re-package those customer apps into out-of-box 1st party apps that it can sell to other customers. This is something that ServiceNow has done exceptionally well over the last five years. 

• More Workday Developers means that deployments and implementations are faster. Faster deployments mean (1) more customers because the time to value is lower and (2) higher velocity of deals. 

Extend, combined with the investments that Workday is making on the channel partnership and developer community side, can be a powerful driver of the company’s sustained growth. The impact of this will show up in a few places: (1) the R&D burden relative to new product launches should decrease, (2) the S&M burden decreases as the velocity of deals increases, and/or (3) the savings from one and two are reinvested to help sustain the company’s revenue growth. The most likely outcome is somewhere between 1, 2, and 3.

Summary Investment Thesis

Investment Merits:

• Workday is building the best-of-breed ERP for the services industry, and the company will continue taking share in the core market.

• Workday is in the early innings of its expansion story. The company’s investments in the channel partnerships and customer base motion – will change the new bookings mix to 80/20 expansion/land over the next five years, which will provide meaningful S&M leverage.

• Investments in the company’s platform services and ecosystem (developer, channel, and SIs) represent an Act 3 for Workday and will be a tailwind for the company to sustain 20%+ subscription revenue growth. 

Near-Term Risks:

• 2023 is shaping to be a dismal year for software purchasing, and long-cycle investments such as Workday will likely be pushed out. Mid-market, which now represents ~50% of new ACVs (+30 percentage points from FY ‘20), will help mitigate the pullback in the enterprise. Additionally, the company has meaningful exposure to markets more resilient to cutbacks (e.g., hospitality, retail, and education).

• Push into the channel will help with the company’s effective sales capacity and deal velocity. However, there is likely going to be some pressure on the headline billings figure. Additionally, Workday (like another enterprise SaaS co’s) has historically been willing to comp/give on pricing and billing terms during slower quarters to land longer-term deals, which will also pressure billings.

While the near-term risks create volatility, I’m bullish on the long-term revenue and margin targets. High-level model and valuation framework below:

Given the near-term risks, this is positioned as a 3-4% of the portfolio with room to double down if/when there’s a drawdown in the stock.

Fin. I’ll continue to track this one and post periodic updates about the business and any updates to the thesis.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about. Again, this is not investment advice, so do your own due diligence. I make no representation, warranty, or undertaking, express or implied, as to the accuracy, reliability, completeness, or reasonableness of the information contained in this report.

Now the market is transitioning to the rationalization phase of the market cycle, and software companies will be required to show meaningful operating leverage. But not all businesses or management teams will be able to.

• For some, it’s a lack of experience. Not that many CEOs or CFOs have experienced a bear market.

• For some, it’ll be a lack of appetite. After a decade of building a company, some CEOs might be tired, and going through this part of the company’s lifecycle is not that interesting.

• For some, it’s hope. Hope that if they invest enough right now, they’ll re-accelerate growth and be the market leader on the otherside. Some might be right, but founders are naturally optimistic (they have to be) and may drive their business into a brick wall.

And it’s not to say that the changes will actually lead to a great financial outcome. The transition from a growth story to a GARP-story (for some, it’ll be a value story) is ugly, and very few companies actually accomplish this. This is typically where private equity comes in and helps reset the story.

Regardless, everyone has one thing in their favor: if managed properly, this transition should not be that challenging from a financial model perspective. The SaaS business model has significant (dare I say, magical) operating leverage embedded in it.

Sources of operating leverage in software businesses

Software businesses have a few sources of leverage:

Sales and Marketing: One of the major drivers of leverage for software businesses and is primarily realized as the company’s net new ARR growth decreases. This impact is largely mechanical since the portion of the company’s cost of acquiring new ARR should effectively decrease (i.e., less sales commission, less required marketing spend to generate leads). The rate at which the company’s net new ARR growth decelerates impacts how quickly the operating margins should increase.

Here’s a simple example that illustrates this point.

The key assumptions are: (a) that the cost of acquiring a $1 net new ARR is $1.50, (b) the cost to support the existing customer base is 20% of ARR, and (c) both costs remain relatively flat1. This is the mental model, but more often than not, software businesses are not managed to follow this framework exactly or unfavorable competitive dynamics limit their ability to do so.

But, even if the business is not managed perfectly, there is still some natural leverage as the company ends up (1) paying less in commissions, (2) renewal fees, and (3) there’s a natural attrition in sales reps. Points (1) and (2) are functions of comp plans2, and (3) ends up being a talent issue (i.e., overall quota shrinks, sales reps don’t meet quota and leave).

However, marketing spending is the biggest opportunity for software companies and the lever that is not flexed enough. Software companies tend to increase their absolute marketing expense irrespective of decelerating growth3. This is in part because of the way companies run budgeting cycles4. But psychologically, reducing overall marketing budget is an admission that the market for their products is saturated, and the company has to transition into the harvest mode of their company’s S-Curve.

So what does the company typically do? Find a new S-Curve, which can take multiple form factors (new regions, market segments, or new products). So the company enters another investment phase to make the “bet,” which sometimes works and most time does not. Unfortunately, some management teams get on this hedonic treadmill and never get off.

There are, of course, exceptions. ServiceNow, Veeva, Hubspot, DataDog, Box, and MongoDB have added multiple S-Curves as public companies. There’s a high correlation between which companies can effectively (and efficiently) create multiple S-Curves and their returns as a public company.

R&D Leverage: This should be a source of leverage for software businesses if the company has a strong market position, but the operating leverage is muted as:

• Competition in the core market increases, and the business has to invest to remain competitive / fight-off new entrants.

• Companies enter a new investment phase to add S-Curves (“make new bets”).
  → The highest ROI investment cycle for public software businesses in the last decade has been transitioning to the cloud or moving up-market.

The truly special companies in software show incredibly high R&D leverage because they are incredibly smart about extending their existing platform into new products (e.g., ServiceNow). The truly terrible companies are those that maintain high R&D spending but also spend significantly on M&A.

Overhead / G&A Costs: self-explanatory (for most management teams).

What does the trade-off between growth and margin look like?

The operating leverage is illustrated by the software profit frontier. The chart below is a scatterplot of the revenue growth of various software businesses and the EBITDA margin over their respective history. To estimate the impact of decelerating revenue on EBITDA margins, we use a polynomial extrapolation to create a “profit curve” for software businesses.

The curve implies that a zero-growth software business should have enough ~40% EBITDA margins with an R^2 of ~48% — and that generally lines up with the SaaS rule of 40. Of course, the data also shows that companies can meaningfully diverge from the profit curve. While easy to put the blame on management, there are, of course, other factors that drive the divergence.

The model businesses that have executed on the profit frontier have been ServiceNow and Veeva, which have diligently balanced growth and profit. Some companies have seen their margins compress as growth decelerates because their end markets became more competitive (e.g., Concur).

The private equity hard “reset”

Companies with strong fundamentals (e.g., low customer churn, favorable competitive environment, and low R&D intensity) but meaningfully divergent from the profit curve will likely have to deal with the prospect of paying the piper (i.e., Orlando Bravo) and a hard private equity reset.

Private equity owners know they can get these businesses to healthy profitability and have the mandate to effect the changes required to balance margins and growth. Again there are many sources of expense reduction that PE owners can implement:

• Rationalization of marketing spending to be in line with the current growth profile

• Pare back sales rep headcount and investments in

• Cut R&D spend in (potentially) low ROI future bets

• Raise prices. Boring, but many software companies do not increase prices as much as possible.

In other words, there are a lot of low-hanging levers companies can pull to increase margins as long as the organization can accept growing at the top line (and for managers, headcount) at a lower rate. This is why the (current5) master of the (software) universe6 is circling.

Subscribe now

In my last post, I wrote that AI-enabled workflows represent one of the core ways that Salesforce and other SORs can drive re-bundling across their product suite over the next decade. The post was focused on Salesforce and its Data Cloud, but Hubspot seems to be similarly embracing ChatGPT with its ChatSpot.ai product1. 

For Microsoft, who has been beating the drum the loudest around enterprise applications of AI, the opportunity is not constrained to simply suites around the SOR (e.g., Dynamics 365). The company’s ambitions are much broader because of its Collaboration and Productivity Suite (e.g., Office 365). 

And on March 16th, Microsoft “showcased” the next iteration of its broader ambitions for Office 365 at its Future of Work event. And two days prior, on March 14th, Google showcased its own AI-enabled productivity suite. I use the word “showcased,” but neither product suite is in GA. 

Microsoft’s Announcement:

In the months ahead, we’re bringing Copilot to all our productivity apps—Word, Excel, PowerPoint, Outlook, Teams, Viva, Power Platform, and more. We’ll share more on pricing and licensing soon.

Google’s Announcement:

As we embark on this next journey, we will be bringing these new generative-AI experiences to trusted testers on a rolling basis throughout the year, before making them available publicly.

Both companies are slightly more focused on beating each other to the punch above all else. Jokes aside, the implication of both “launches,” especially Microsoft’s, have far-reaching implications for how the software stack looks for the next decade.

While both Salesforce and Hubspot are driving towards re-bundling around the customer record, Microsoft is pushing towards a re-bundling around their knowledge graph with their Copilot and Business Chat. In doing so, the company seems to have either created or adopted a new way of reliably working with the LLMs and searching within a business context.

The Copilot System and Information Network Effects

There were three aspects of Microsoft’s recent launch: Generative AI for Microsoft Apps, Business Chat, and the Copilot System.

The Layup — Generative AI for Apps

Let’s get the easy one out of the way first – Generative AI for Microsoft Apps. Microsoft launched their broadly expected adoption of Generative AI features within the Office Suite. The Generative AI features allow users to automate tasks, summarize information, create content, generate analysis, and even create automations across workflows.

Copilot Announcement

Copilot is integrated into Microsoft 365 in two ways. It works alongside you, embedded in the Microsoft 365 apps you use every day — Word, Excel, PowerPoint, Outlook, Teams and more — to unleash creativity, unlock productivity and uplevel skills….

With Copilot in Word, you can jump-start the creative process so you never start with a blank slate again. Copilot gives you a first draft to edit and iterate on — saving hours in writing, sourcing, and editing time.

Copilot in PowerPoint helps you create beautiful presentations with a simple prompt, adding relevant content from a document you made last week or last year. Copilot in Excel, you can analyze trends and create professional-looking data visualizations in seconds. Copilot in Outlook helps you clear your inbox in minutes, not hours. And every meeting is a productive meeting with Copilot in Teams. It can summarize key discussion points — including who said what and where people are aligned and where they disagree — and suggest action items. Copilot in Power Platform, anyone can automate repetitive tasks, create chatbots and go from idea to working app in minutes.

Again, this was a largely expected launch. Beyond making the core Microsoft applications potentially more valuable to its users, the launch has near-term implications for two segments of the market:  

• Copywriting startups built on top of OpenAI or other commercial models → unsure2 what happens to these companies. Maybe they move towards becoming hyper-specialized for specific verticals. 

• High-volume, low-value freelance and agency markets → these companies may have an arbitrage for a short period of time where they can actually use AI to make themselves more efficient and maintain price. But over time, they’ll need to (or get to) focus on higher-order projects/tasks.

Business Chat — Powered by the CoPilot System

The second announcement was Business Chat, which is the company’s Chat Interface that allows users to query/ask for information on the company’s corpus of knowledge (documents, emails, spreadsheets, etc.). 

More or less, it's ChatGPT that has access to enterprise files and applications. But if two months ago, after the error-prone demos of BingChat and Google’s Bard (and the awakening of Sydney), you had pitched someone “ChatGPT for the Workplace,” there would have been two responses: 

• (1) “Oh god”

• (2) the training costs may be prohibitive.

So the more interesting and important aspect of the announcement is what’s under the hood of Business Chat, the CoPilot System. Here’s the company’s visual representation of the Copilot System:

The CoPilot System has three components:

• Microsoft Application Suite – Teams, Outlook, Powerpoint, Word, and Excel. 

• Microsoft Graph3 – the data layer that sits across the Microsoft apps that have structured information such as email, contacts, text within Word docs, etc.

• Large Language Model – OpenAI’s LLM, which is what users query effectively. 

When a user enters a query into Copilot, the following things happen:

1. The user’s prompt is grounded on Microsoft Graph data, and a modified prompt is created (pre-processing) 

2. The modified prompt is then sent to the LLM for a response or app command (processing)

3. The response is then grounded again against the Microsoft Graph (post-processing)

4. Finally, the response is sent back to the user. 

At this point, you are probably wondering why I keep bolding grounded.

If we rewind the tape a month, BingChat launched, everyone clapped, Google launched Bard, everyone mocked its errors, and then it turned out that BingChat had the same problem. The issue is that LLMs have a “hallucination” problem, they come up with an answer even if they don’t know the answer. For certain queries, this is probably fine – but for some contexts (e.g., business setting), it can actually be quite harmful. 

Imagine that in the scenario above, there was no meeting with “Fabrikam.” Because LLMs are designed to be predictive, they “predict” responses, so an unconstrained LLM may simply predict that “Yesterday we shared the following quote” and reference a made-up sales quote. The VP of Sales may reference that made-up quote in a message to the CEO, and we’re off to the chaos races. 

Grounding and LLM-Augmenter

Grounding is designed to obviate this risk, and what seems to be happening under the hood is that Microsoft is using a version of the LLM-Augmenter they wrote about two weeks ago, which is designed to prevent this hallucination issue. 

From Microsoft’s Paper:

Large language models (LLMs), such as Chat-GPT, are able to generate human-like, fluent responses for many downstream tasks, e.g., task-oriented dialog and question answering. However, applying LLMs to real-world, mission-critical applications remains challenging mainly due to their tendency to generate hallucinations and their inability to use external knowledge. 

This paper proposes a LLM-AUGMENTER system, which augments a black-box LLM with a set of plug-and-play modules. Our system makes the LLM generate responses grounded in external knowl- edge, e.g., stored in task-specific databases. It also iteratively revises LLM prompts to improve model responses using feedback generated by utility functions, e.g., the factuality score of a LLM-generated response…

What the LLM-Augmenter system is doing is:

1. Based on the user query/prompt, the LLM-Augmenter retrieves evidence (i.e., information) from a defined data source (e.g., database or Wiki) using an NLP-enabled search. For CoPilot, the data source is the Microsoft Graph, and only information in the graph can be used as “evidence.” 

2. LLM-Augmenter then takes the “evidence” and the prompt (modified prompt) and queries the fixed LLM (e.g., ChatGPT). The LLM can then generate a response to the prompt grounded in the evidence from the Graph. 

3. The LLM-Augmenter then checks the response against the Graph again to confirm that the response has information that can be directly attributed to data and documents in the Graph. 

4. The verified and cited response is sent back to the user.

You can actually see these steps happen one by one in the demo. And the annotated steps below:

In addition to reducing the risk of LLM hallucinating, there are other benefits to this approach:

• Microsoft can use a “fixed” or static LLM, so the users do not have to constantly re-train models to incorporate new corpus of data → cheaper, flexible, and more scalable approach. 

• Only the relevant or required data is passed to the LLM → more secure and compute-friendly. 

Information Network Effects

Driven by the CoPilot system, Business Chat provides users to effectively and reliably search, query, extract, summarize, and build top of all their company’s documents, files, and knowledge in one place. Strategically, the CoPilot System allows Microsoft to foster intercompany network effects. In other words, the summaries and content generated by CoPilot are more powerful if every node (i.e., user) is using tools that feed into the Microsoft Graph because that’s the corpus of data Copilot uses as “evidence.”

In an end-state where knowledge workers are trained to use Business Chat to find information at their company, there’s meaningful value leakage if everyone at the company is not using the same tools. This highlights the primary issue of the last ~10 years of software buying. For the last decade, SaaS applications and productivity tools have thrived on bottoms-up adoption and the departmentalization of software spending, which has inevitably led to the fragmentation of information and tools.

At many companies, especially ones with multiple business units, and a history of acquisitions, it's not uncommon for different teams to use different filesystems (e.g., Sharepoint, Box, Dropbox), collaboration tools (e.g., Teams, Slack, Confluence), and project management tools (e.g., Smartsheet, Asana). For investors, that has been the beauty of software this last decade, bottoms-up GTM and adoption of the cloud have allowed teams to select the tools that best fit their needs. But the corollary to that has been information and knowledge have increasingly fragmented across the enterprise. 

The cost has shown up in two areas: 

• The unquantifiable productivity loss of having information so disaggregated

• The very quantifiable seat bloat (seatware) at organizations. For example, the finance team may not directly need access to Notion4, but that is where the Product and Sales prefer to do their planning/collaboration so everyone in Finance needs a seat because you need the transparency and information for EoY planning. 

Historically this has been bearable for a few reasons: (1) everyone had unlimited SaaS budgets so getting access to the information in another system was not an issue and (2) there was no information or business cost of not standardizing around one system.

Microsoft Copilot and Business Chat will be effectively highlighting the previously cost of unquantifiable cost of productivity loss from searching for information across the company’s various systems.

More recently, there have been entire companies built to address this problem, such as Glean. From the company’s press release:

• Knowledge workers waste 20 percent of their time — more than one day per week — looking for the information to do their jobs (McKinsey).

• They average over 6 hours per week duplicating work that has already been done by a current or former coworker (Asana, 2020).

• On top of that, context switching between a dozen apps 30 times per day and dealing with interruptions every 3-11 minutes (UC Irvine's Donald Bren School of Information and Computer Science) leads to burnout. It is fueling the "Great Resignation" trend.

Standardizing Around the Microsoft Graph

If Microsoft is successful in convincing enterprise users and, more importantly, managers and executives of the value of the Business Chat, there’s going to be a push to re-bundle/standardize on the Microsoft Graph. The cost of not standardizing will become apparent when users query information and get incomplete or incorrect responses because the was pertinent information that sits outside of the graph.

There are two forms of standardizations: (1) all SaaS applications need to build very tight integrations with the Microsoft Graph so their data can be queried by Copilot, and (2) any SaaS tool that is not frictionlessly compatible with the Graph is replaced by a Microsoft equivalent if one exists. 

The Microsoft Graph is not new per se. It’s been around since 2016, and many popular SaaS applications, such as Confluence, integrate directly with the Graph, but there are limitations:

The Confluence Cloud connector has the following known limitations in its latest release:

• Confluence Cloud connector does not index attachment files and comments.

• Indexing Server and Data Center deployments will be released as a separate connector.

Even companies, like Box, have clunky implementations and limitations when it comes to integrating with the Graph. And none of these are the fault of Box or Confluence, this is Microsoft developer APIs, we’re talking about.

Microsoft 365 users, however, do not have to worry about this issue, they have end-to-end integrations that will allow frictionless movement of data and information across their apps and into the LLMs. But nevertheless, this is a problem that other companies need to solve at some point if they want to sell into the Microsoft installed base. Otherwise, here’s how most conversations will go:

VP of Sales: Hey Tim, did we end up sharing the quote with XYZ Client after the meeting? 

Sales Rep: Yeah, it’s saved here in IODT (Indiscriminate Other Document Tool). 

VP of Sales: Ah okay. Can you save it to Sharepoint or Presentation next time? It’s not showing up when the CEO asks for a meeting summary in Teams. 

Sales Rep: Sounds good. Will do. 

VP of Sales: Also, can you stop saving client materials in IODT moving forward. Third time this month this happened. 

[6 months later]

CFO: Hey is anyone using this IODT?

VP of Sales: No, we stopped using that. 

CFO: Great! Will remove it from the budget. 

Obviously, this is a bit cartoonish, but for Microsoft365-centric companies, this is bound to happen in the world that Microsoft is pushing towards. And the sale to managers is very simple because Business Chat is probably one of the greatest enablers of micromanagers the world has ever seen. Managers and executives can now, with a simple query, know anything they want regarding meetings and content5.

What Microsoft is doing with the combination of CoPilot, Business Chat, and Microsoft 365 is creating information network effects within companies. Any tool that does not augment those network effects or, worse, adds friction to those network effects, will likely be jettisoned. 

The Non-Microsoft World

Not every company, especially in the tech world, is a Microsoft company. Most companies in Silicon Valley are G-Suite users.

Fortunately for the productivity tool companies around the Bay Area, Google has not been as aggressive as Microsoft with its productivity suite. Microsoft has been, over the last few years with regards to co-opting very semi-successful productivity suite ideas or features. Google does every now and then co-opt features, but they’re worse at shamelessly copying and distributing when compared to Microsoft.

And importantly, G-Suite’s APIs have been more developer-friendly, specifically Calendar and Gmail. Unlike Microsoft, which has Dynamics, Google does not have a large ERP and CRM suite to upsell its customers onto after it pushes them to standardize everything around G-Suite, so the juice may not be worth the squeeze. 

If G-Suite decides to forgo the aggressive re-bundling opportunity that Microsoft keeps going after, there’s probably an opportunity for a 3rd party vendor to create the “CoPilot System” for the non-Microsoft ecosystem.

And there are probably a few wedges for the 3rd party CoPilot System: the collaboration layer, search, or neutral middleware platform. 

• Notion is the most likely candidate at the collaboration layer and has acquired three companies or teams adjacent to collaboration (Calendar, Workflow, and Search). 

• Glean is a well-funded enterprise search startup that is more or less built to solve the disaggregated knowledge problem, and they have the integration footprint required. 

• Middleware platforms are tougher to identify because they would have to become a standard across companies to really work. 

  • Algolia has domain knowledge but lacks an integration footprint. 

  • Companies like Merge have the integrations but are a bit removed from the domain problem. 

There’s also just the very real possibility that Google goes after this opportunity as well, but their investments around G-Suite leave much to be desired.

Either way — any tool that does not fall neatly into the information graph (Microsoft or G-Suite specific) will be at risk of falling out of the enterprise stack and/or have its TAM severely constrained.

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about.

And in the immortal words of @dalibali: “Not investment advice, obviously”. Do your own due diligence.

Historically, building a software business capable of becoming a unicorn or going public required some insight or innovation (technical, GTM, data). But because of the availability of capital over the last decade, many founders and management teams have underinvested and/or ignored this requirement as they grew their businesses. In an era of cheap capital, it became too easy to hire a few additional reps, spend a little more on SEO and marketing to maintain healthy growth rates, and fend off investor concerns. As a result, many recent software unicorns have built businesses with thin or non-existent durability of growth or moats1. Many software businesses today have unit economics and capital efficiency that are meaningfully worse than what investors have come to expect from the business model.

As the fundraising and capital environment normalizes over the next few years, we should expect a new generation of hopefully more resilient software with better unit economics. Mostly, because the founders or investors will be required to because of capital constraints.

Lower barriers to entry and cost of capital

The SaaS market has grown meaningfully over the past 15 years. At the time of the GFC, the total value of SaaS companies was ~$50-100bn market cap, and now it is >$2tn. Part of that is a transition of legacy software businesses to the SaaS models (e.g., Adobe and Microsoft).

In addition to the broader software tailwinds, a few other factors have helped accelerate the growth of SaaS:

• The cost to launch a SaaS startup has exponentially declined → driven by the growth of cloud vendors.

• Reduced friction in selling and implementing software → driven by departmentalization of IT spending and willingness of companies to use cloud implementations. 

• Increase in venture and growth funding for SaaS co’s → driven by low-interest rates and public market demand/interest in software businesses. 

The net impact is that barriers to launching a software business have decreased precipitously, and the cost of capital to finance these endeavors has decreased. As such, competition in many application software markets has increased, and as a result, the cost to scale a SaaS startup has meaningfully increased over the past decade. The two most prevalent areas are the cost of acquiring a customer and the cost of skilled tech employees (incl. developers).

Customer Acquisition Costs: Public longitudinal data is hard to find, and analyzing public company data is a bit noisy, but Profitwell’s data looks directional correct. Jason Lemkin has recently written about this from the perspective of the impact of excess VC funding on sales rep compensation, which I thought was poigant.

Cost to Hire: This topic has been covered ad-nauseam by many, but the cost to hire and retain high-skilled tech employees, which includes developers, sales leaders, and other execs, has gotten out of whack. Cash-rich tech companies (i.e., FANG) have been a cover driver of inflation. Startups and smaller tech companies have broadly tried to compete by paying more on equity. The impact is noticeable when you look at the SBC burden at software companies over time.

Declining returns on “vanilla” SaaS

So things have happened, the cost to scale has gone up, barriers to entry have gone down, and finally, these companies are increasingly priced to perfection. Investors have broadly fallen in love with the SaaS business model because these companies check a lot of the boxes that investors care about (recurring revenue, high-gross margins, sticky customer base, and high incremental margins2).

Naturally, this dynamic has led to excess capital, more competition, poor management discipline, and increasing capital intensity in the industry. The average lifetime burn multiple for SaaS businesses at IPO has steadily increased over the past few years34.

The businesses that have bucked the trend and will likely continue to buck the trend have been businesses that have typically innovated along some vector (product, GTM, or vertical focus).

Businesses that innovate on the product are rare in application software. And maintaining or extending that advantage is even more rare5. More often than not, product advantages are typically co-opted by an incumbent or other emergent vendors. So most software businesses fall into a bucket of "vanilla SaaS." These software segments are filled early on with multiple competitors with similar feature sets, rough product parity, and relatively low customer switching costs.

While it may not be an enduring framework, these vanilla SaaS businesses are typically not systems of record but are systems of intelligence, engagement, or productivity that sit around a system of record. Martech, productivity software, project management, and sales intelligence tools typically fall into this bucket of vanilla SaaS.

As a result of thin moats and the low cost of financing, vanilla SaaS markets have consumed a lot of capital. These businesses are hard for a variety of reasons (not exhaustive):

• Pricing power is low, or price competition is high, given the competitive environment. In some instances, the alternatives might be free (i.e., spreadsheets), so the willingness to pay is even lower.

• Companies need to consistently invest in expensive R&D to win new customers but not necessarily be able to flex on price to recoup R&D costs.

• CACs are highly relative to the ACV/price, and GTM has to spend an inordinate amount of time educating the customer about the value of their product vs. alternatives.

Generally, the enterprise segment tends to be an area where this pattern bucks the trends. Often this is because the bar for establishing yourself in the enterprise is very high, but once a company does, the willingness to pay is high. The customer segments tend to gravitate towards an industry standard (i.e., copy each other).

So should we call it quits on vanilla software?

No. There are incredibly high-quality software businesses, but companies and investors need to understand what game they’re playing. If the business is going to sell vanilla SaaS, the business typically needs to innovate on the business or GTM model6. Businesses that have done this well have been Bill.com, Hubspot, and Atlassian.

For example, Bill.com’s differentiation has consistently been its ability to acquire new customers with brand awareness and its ecosystem/accounting channels. Emergent vendors (e.g., Ramp) may have better/flashier UI and marketing campaigns, but they struggle to replicate Bill.com’s ecosystem advantages.

And Hubspot has built a business around inbound/organic demand that converts and expands based on product excellence. The founders have historically been disciplined in ensuring the acquisition funnel was incredibly efficient because their business demanded it, i.e., it’s impossible to scale if you’re selling low ACV software with high CAC.

Interestingly, both companies were scaled during a time of higher interest rates, so they had to build businesses that could scale efficiently. The capital was too expensive for them not to. The counter-example to Hubspot is Freshworks, which is serving a similar end-market, but it’s hard to imagine that they’re going to be able to scale with their current sales efficiency metrics.

Bull-case is that the next generation of software business (i.e., ones being built today) and some companies willing to quickly course correct will hopefully have similar dynamics. However, the businesses that were built in the era of low-interest rates will have to massively re-adjust their economic models and go back to first principles to ensure that their unit economics actually makes sense.

There are a few sources of defensibility in application software:

• Mission-critical system of record, e.g., payroll

• Great product with strong organic demand and efficient GTM, e.g., Atlassian

• Becoming an industry standard and how a swath of people do their job, e.g., Salesforce

• Strong ecosystem dynamics, e.g., Bill.com

• Discover an underserved segment or market early and quickly, e.g., Veeva7

The list is not meant to be exhaustive, but those have been the common patterns we’ve observed. For the next-generation software businesses, focusing on building some lasting advantage beyond “better product” that looks like one of these traits is going to be the determinant for if this business will win the market for the long-term because if the opportunity is attractive enough, competition will come.

Subscribe now

Larry Ellison’s Gospel

Software companies can be generalized into three categories: systems of record (“SOR”), workflow systems, or middleware tools that help bring these systems together. Few examples of each:

• Systems of Record – CRM (Salesforce), ERP (Oracle), HCM (Workday), IT (ServiceNow)

• Middleware – Mulesoft, Segment, Workdato

• Application / Workflow – Outreach, Gainsight, Blackline, Coupa

The core systems of record have been CRM (Salesforce), ERP (Oracle), HCM (Workday), and ITSM and the center of gravity in application software has historically been around these SORs. Workflow tools, like those highlighted above, tend to revolve around these SORs. While the SORs have been great investments for investors and extremely sticky with customers, they have a major drawback with regard to end-user UI/UX. Fortunately for the SORs, UI/UX is not always the core buying decision. Workflow software fills the gap in the market by creating software that enables the end-users (e.g., sales reps, customer service agents) to be more productive when it comes to working with or around the SOR.

The workflow tools exist because the user experience built around the SOR was not always optimized for the automation of tasks and end-user productivity. Many times the core problem the SORs are solving is a database problem. And even when the SORs try to serve the workflow use case there was always a tension between the workflow features the customers want and the SOR’s inherent bias toward their own data models. When SAP builds workflows around their SORs, those SORs are designed with only SAP’s SOR and data model in mind.

Many of 3rd party workflow tools became effectively “Switzerlands” in the software ecosystem, integrate with multiple SORs and help employees complete tasks. The primary task that these workflow tools focus on is the movement and retrieval of data across the systems more efficiently. 

As these systems of record mature in their core markets, they typically have to integrate vertically deeper into workflows and expand horizontally into other SORs and sub-SORs1. Both these levers, allow the SORs to persist for a long period of time.

For SORs, vertical expansion can be easier than horizontal expansion. When expanding vertically, the SOR is effectively extending its existing data model. For example, when SAP acquired Concur, they were expending their ERP SOR to subsume a downstream workflow (approvals) but also extending their data model to T&E.

Being caught on the wrong side of the decision by one of these SORs can be treacherous. The workflow vendors that supported fragmented ecosystems with multiple leading SORs tend to be more defensible. But even so, their core TAMs slowly shrink over time as the SORs co-opt some of the features and competition increases from next-generation “best of breed” tools around the incumbent SORs. 

In markets where the end customer is using multiple vendors for the core workflow and drawing data from multiple SORs and sub-SORs (e.g., CRM and Martech), building a vertical integration can be borderline impossible or expensive. 

Larry Ellison in Softwar:

Software from different vendors will have different semantics – even something as simple as defining what a “customer” is may differ – different underlying data schemas, which have to be coordinated but will scarcely ever be fully reconciled, and different user interfaces with conflicting design conventions and display elements. Even if the consultants have a prove integration method to link two pieces of software, APIs still have to be specially constructed to pass messages among distinct database schemas, limiting the amount of information that can be extracted as well as duplicating storage requirements.

The natural progression of this train of thought is that SORs should over time leverage their data model to vertically integrate. But the more controversial Ellison conclusion is that these SORs across the business units should be integrated and combined together with the same data model. 

The answer to the problems that enterprise application customers were facing, Ellison concluded, lay in doing two things, which he calculated that Oracle alone was capable of. The first was for all of Oracle’s applications to be tightly integrated both with one another and with the database through architecture based on a single shared-data model. The second was to produce the first-ever complete suite of applications that would be capable of automating every aspect of a company’s business, from the traditional ERP, or back-office tasks, such as financials, manufacturing and HR, to the fancy new stuff that included customer relationship management (sales, marketing, and customer support), the supply chain, procurement, and Web stores. 

This conclusion (or hypothesis) was somewhat out of self-interest. Ellison’s view was that the database should be the center of this suite and the application suite should revolve around the database. Oracle, the leader in database technology would be the natural vendor for all of an enterprise’s software needs.

Even though Oracle had a variety of challenges executing this strategy, Ellison was broadly right and most software suites have evolved in a similar manner – enterprise software companies start as a core SOR and then expand vertically (via workflows) and horizontally (to other SORs). By the way, he had this vision pre-2000!

Marc Benioff, an Ellison disciple, has been the most public with regard to pursuing this strategy. And like Ellison, Benioff has taken an acquisitive approach to expand his SOR footprint. And like Ellison, Benioff has had challenges executing the consummate with investor expectations. 

How Salesforce got here

Salesforce started with a sales-focused cloud (CRM) and expanded vertically and horizontally via R&D and acquisitions. The singular goal for Salesforce with these acquisitions was to: (1) create a unified view of the customer and (2) drive automation across the sales and marketing process.

Building the suite has been expensive, both from an M&A perspective and also from the R&D cost to integrate all the different clouds. Acquisitions as the core driver of the strategy has presented a few challenges for Salesforce: 

• Integrating these different suites and their backends takes a long time. 

• Customers do not realize the full value of the suite until the integrations are complete because of the disaggregated data models. 

• Middleservices and data warehouses have evolved to the point where customers could effectively re-create the full sales and marketing stack to be close to Salesforce’s bundle without having to buy the bundle.

As a result of these shortcomings, Salesforce’s primary lever to drive adoption was to focus exclusively on CIO/CDOs and/or concede on price. But more often customers would just stitch a variety of different best-of-breed tools together and create their own data model or framework in the data warehouse. 

As a result, for most customers, there has not been a compelling reason to standardize around Salesforce. This is further exacerbated by the fact that Salesforce’s core products are purchased by different teams (sales, marketing, and customer success). The standardization pitch is most effective in the “enterprise” segment, where companies have CIOs, CDOs, etc. that have the mandate to drive the standardization across teams or business units otherwise teams typically choose best of breed tools for their functional group.

Salesforce is not intransigent about this issue and the company has been investing in the backend infrastructure, APIs, and platform to bring these tools together. Arguably, this has been a major drag on the company’s margins and one of the reasons that even as growth has slowed in core Sales and Service Cloud, margins have not inflected as investors have hoped. 

Yes, the acquisitions have made the story noisy, but more importantly, the acquisitions consistently moved the ball for the platform, which meant more backend investments. 

Those investments have all effectively coalesced around Salesforce’s Genie and Data Cloud. Salesforce brought those products and terms2 to market last 2019/2020.

Importance of the Data Cloud 

Salesforce Data Cloud is just a fancy word for the company’s customer data platform. And customer data platform is a fancy word that has all of the relevant customer information from all of a company’s systems in one place. While this sounds relatively simple, this ends up being a very complicated database and orchestration problem3. Companies interact with new and existing customers across multiple teams (sales, marketing, customer service), channels (in-store, online, in-app), devices (mobile, web), and tools. 

The information collected at those various points can be different as well. The goal of a CDP is to basically identify Joe Smith across datasets and unify his “profiles” so that when companies interact with him (sales, marketing, customer services), they understand the full context of his interactions with the company.  

Because of data privacy concerns and regulations, you cannot hamfist every single data point related to a customer into one database. For example, a cable company likely does not want to include the social security number of all their customers in the database that is used to power the mobile app used by the field service or contractors. To solve this, companies have historically either stitched together these CDPs within in-house solutions and/or used external vendors. These centralized databases have all sorts of compliance support and role-based access controls to prevent any data privacy issues. 

On top of doing all of this ingestion and data orchestration, companies have to invest in building dashboards and reporting functionality with a team of data analysts/engineers. All for the “simple” goal of being able to create one unified view of the customer. 

Salesforce with its investments in the Data Cloud is attempting to cut out a significant portion of the intermediary layer that goes into creating a unified customer view. 

From TrailblazerDx

Salesforce Data Cloud lets you bring in data whether it’s streaming data or batch data at scale, into the platform. And then it lets you create the single source of truth so you know exactly what the customer has done with the business. And then you can actually use that data. 

Now the key thing is Data Cloud is not just another data lake or warehouse. It’s an active platform. So that means you can use that insight back into your 360. So that means you can actually call your salesperson, you can create a new service case, you can create personalized ex

What would you do in a normal situation [like this]? You have to take all this data out, put into another lake, run some other program on it, to do AI insights, etc. Not with Data Cloud because we have Einstein built right in. It’s zero ETL and zero copy. 

Again, if you have to do this by yourself, what do you have to do? Copy the data into another data warehouse, leave Salesforce, do it all over again, integrate with Tableau. But with the power of the data cloud, with Zero ETL and Zero Copy, Tableau is running live. 

The broader objective is that the Data Cloud finally unifies Salesforce’s acquisitions into one data model and platform with support from real-time streaming pipelines. When companies rely on the orchestration or aggregation of data based on APIs or data warehouses, there’s always a lag between data being captured by one system, processed, stored, and retrieved by another system. If you’ve ever stood in front of a Bonobos sales rep waiting awkwardly for their email confirmation for the new account you just created to come through, you know what I’m referring to.

Genie and the Data Cloud basically are creating a real-time single source of truth for companies by leveraging their data streaming services and singular data model. 

For those familiar with Salesforce’s story, this is not particularly new. Benioff has been pitching this vision for the past decade and that’s been core to the company’s thesis behind these acquisitions. But the lack of a cohesive data model on the backend has allowed the company to fall short of its potential and more importantly has allowed non-Salesforce tools to co-opt parts of the sales tech stack. 

The unified data model should help Salesforce drive more multi-cloud adoption. 

At the stage of the company’s lifecycle, the reduction in churn associated with the increase in multi-cloud adoption is going to be very impactful. Today, Salesforce has ~93% gross revenue retention, which is good but not exactly best-in-class, especially for a tool that’s often considered to be mission-critical for companies.

But the unified data view alone is not enough to consolidate and upsell the sales and marketing data stack. The bias continues to be toward best-in-breed tools for each part of the stack. And companies are used to the notion of using a hodgepodge of tools to aggregate and process the data in a cloud data warehouse. 

The main carrot Salesforce has to drive this upsell/re-bundling, and where this notion of single customer view really matters, is their automation and AI products. 

Automation and Re-Bundling of Software

Salesforce has been on its AI journey for a few years. The simple goal is to automate tasks and provide predictive insights for sales reps and marketers. Per Benioff tradition, it had a very unassuming and easy to understand name, Einstein.

Originally, Einstein primarily helped companies extract and tag information in their CRM and adjacent systems. It would do things like:

• Capture customers and deal data from emails, which reduces manual entry for reps. 

• Turn interaction data into insights i.e., competitors are mentioned. 

• Score leads and sales opportunities i.e., the likelihood of closing

• Predict account churn

The functionality was largely constrained to data retrieval. Einstein saves reps time from route tasks (to be fair, some of which are a function of Salesforce’s design choices). The features largely looked like “out of the box” data science and data engineering, which while impressive did not meaningfully drive value for the average Salesforce customer because the functionality was limited by a few things:

• Data silos across the sales and marketing tech stack → improving with Salesforce Data Cloud and companies’ investments in cloud data warehouses and data lakes

• Cost of computing and query for ML models → improving with maturation and investment in AI/ML

Salesforce launched the next iteration of this technology at their developer conference, TrailblazerDx. The main focus of the conference was focused on “Einstein GPT”. The product is an extension of the “Einstein AI” that Salesforce launched a few years ago. 

Einstein GPT is the world’s first generative AI for CRM. This is important because it is learning from your trusted customer data. You can bring your models and bring your algorithms. 

It starts with your customer data. Einstein GPT taps into all the data that you have been building as your single source of truth. And then it uses this data to train using security principals to maintain that secure data access to your data because it’s within the Salesforce trust boundary. This protects the privacy of your PII and it’s purpose-built for Customer360. 

Einstein GPT is a gateway where you can use our large language models, or you can bring your LLMs to the process, or you can use the ecosystem for any model that you have already been working with. 

The idea is simple, Salesforce can automate a significant portion of the sales reps’ tasks and more specific tasks associated with data extraction and processing. In this example, the processing means taking information that’s in the CRM and creating an output. Automating this aspect of a sales rep’s job is what an LLM is really good at. 

Critical for companies is that AI/ML finally gives a tangible output/result for their investments in creating a “single source of truth” that the end business user can see. While the idea of a single source of truth makes a lot of sense, in practice many of Salesforce’s end users have not seen the meaningful benefit of having a tightly integrated system. To date, the benefits have primarily been constrained to reporting and human-driven insights i.e., data analysts. 

Beyond reporting and insights, it was hard to justify the business value of aggregating workflows and SORs around one vendor. Yes, there are non-trivial economic vendor consolidation benefits but those benefits are largely the focus of the centralized teams (e.g., finance, data, etc) as opposed to the functional teams using the tools (e.g., sales and customer service). Said more simply, the improvements in AI and LLMs, will finally It finally give companies the incentive to bundle around one SOR, which in this case is Salesforce Data Cloud. 

With AI capabilities like LLMs, a sales rep can: 

• Automate things like customer outreach, research, and service 

• Create new content

• Generate insights without having to leverage a data analyst

Implications for Salesforce and the Ecosystem

For Salesforce, the Data Cloud represents the end of a 10-year investment cycle and the beginning of the company’s re-bundling phase. Tools like Einstein GPT finally provide the tangible business benefits that companies can use to drive vendor consolidation across their teams beyond “it’s less work for the data eng” team. 

The likely losers of this are the application and workflow companies that have existed around the Salesforce ecosystem for the last tens of years. Any tool that leveraged Salesforce’s data “plus one” to drive rep productivity is at threat. Plus one refers to the host of companies that created workflow solutions that would combine Salesforce data and data from one other tool (e.g., Outreach, Scratchpad). 

Salesforce will likely subsume or make some of these ancillary tools redundant for two reasons: (1) AI-powered workflows will obviate the data coordination problem these companies solved and (2) Salesforce knows that the more workflow and data lives on its platform, the better it can make its models. Over time as Salesforce adopts other AI models (e.g., OpenAI Whisper) and modalities, tools that we likely thought were immune from the threat of Salesforce (e.g., Gong) will likely be under threat.

All these factors, likely lead to Salesforce flexing more pricing power and margin over time than we currently expect4. 

Salesforce and Beyond

AI and ML-powered workflows will be automation engines for companies. They will remove a significant percentage of data retrieval and processing that humans currently do when completing tasks at organizations. Most of these tasks and workflows typically have clear constraints and rules, which make them perfect for AI-driven automation. And as per Ellison’s Gospel, having a single system with the same data model that persists through the stack is the best way to drive this automation. 

Because of the importance of training data when it comes to creating and improving these models, having a single interface and database is critical for two reasons:

• Feedback from humans is critical to fine-tuning LLM models. When a sales rep corrects the output of an email, it is important for the ML model to capture that feedback to help re-train the model. 

• Not all of a SOR’s data is exposed to 3rd parties. When Salesforce and other software companies build APIs and data access points, they add layers of abstraction, which can obfuscate or loses some of the metadata along the way5.

The result is that Salesforce is best equipped to drive end-to-end workflow automation. This is not dissimilar to the automation that Oracle and Ellison talk about to this day, ~20 years after Softwar. 

By using the AI and processes in Fusion ERP cloud, we have been able to eliminate more than 30% of our manual accounting activities. And enabled by Fusion HCM cloud, we have seen employee satisfaction levels soar with all-time high rates for things like hiring and onboarding new employees. We've also made it easier for our managers and employees as Fusion HCM reduces the time needed to complete the talent review process by more than 70%. And separately, we're saving more than 20,000 hours of manager time each year with our accelerated job offer process.

In sales, we're using our front-office cloud platform, augmented with machine learning and our own data cloud, to help our salespeople sell more and sell more quickly. With Marketing Cloud, campaign planning now takes days rather than weeks. And with built-in machine learning, we've seen a doubling in lead conversion. We automatically capture millions of activities in sales cloud each year. And with CPQ cloud, ordering is much faster and easier with over 70% of our transactions fully automated. We needed that to handle the increased volumes of transactions as a result of our customers' move to the cloud. In addition, we've adopted the Gen2 Infrastructure, including Autonomous Database, for our custom apps. Our internal IT costs to run these systems are down by millions while, at the same time, we are adopting more than 100 new features each quarter.

Here at Oracle, we are going to continue using our own cloud technology as an intelligent automation engine and continue to simplify our business model and processes.

In fact, every well-built multi-product software company will have similar benefits.

The story here is not about a mythical “data is the new oil” vision. But rather that having the data in one schema/structure helps drive meaningful automation. The maturation and increasingly lower cost of AI are simply enabling automation. When we talk about automation, it is simply a push towards faster and better data retrieval and process, which is what these new models do really well. 

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about.

This is not investment advice. Do your own due diligence.

Historically, the primary drivers of returns for private equity returns have been:

1. Value creation via margin expansion, new revenue opportunities, roll-ups, etc. 

2. Leveraging and then de-leveraging an underleveraged business

3. Multiple expansion

PE firms emphasize the first lever in most of their marketing materials – Thoma Bravo claims they can drive 30+ percentage points of EBITDA expansion once they acquire a company.

But in terms of impact, multiple expansion has historically accounted for ~50% of all private equity returns. Revenue growth and margin expansion have accounted for the remainder. Margin expansion has accounted for less than 10% of the “value creation”

The impact of multiple expansion of returns is not lost on anyone, but I wanted to dig into the other lever – the ability of private equity firms to drive revenue growth/margin expansion for software companies. 

Where’s the value creation?

To analyze the impact of private equity ownership on software businesses, I analyzed companies’ growth and margin structure pre- and post-PE ownership. The sample size of companies that have been public, gone private, and returned to the public markets is small, so I expanded the sample set a bit. 

There are three categories1 of companies that I analyzed:

• The Rounder Trippers: companies that were once public, bought out by PE firms, and then re-IPO’d. Analyzing these companies provide the cleanest view of what the impact of PE ownership has been. Companies in this group include Informatica, SolarWinds, and Instructure.
  

• Recent PE-Backed IPOs: companies that recently came into the public markets after years of PE ownership. Companies in this group include PowerSchool, Ping Identity, JAMF, and Datto2.
  

• Management Plans of Recent PE Buyouts: companies that were recently taken private. Analyzing these companies’ management plans as disclosed in their proxy filings provides an interesting baseline for what public companies think they can accomplish on a standalone basis. Companies in this group include Anaplan, Avalara, Coupa, Medallia, and Sailpoint3.

We used data from these companies’ public filings can be used to create an aggregated view of growth and margin profiles of software companies pre and post-buyout. 

Subscribe now

To our surprise, the margin profiles, growth rates, and combination of the two is not materially different between the two samples. Private equity ownership does not seem to have a trajectory-changing impact on the business. In some cases, such as Informatica, the actual revenue and EBIT performance under PE ownership meaningfully underperformed Management’s Plan at the time of the buyout. 

So what gives?

If they cannot realize the cost savings, how has private equity made money on these deals, and how will they make them in the future? 

Turns out the chart above does not tell the full story because it excludes stock-based compensation (“SBC”). We can debate ad-nauseam how SBC should be treated, but we can agree that it has an economic impact on the business because it is adding to the share count.

Over the past ten years, the SBC burden at software companies has continuously trickled up, driven by various factors (competition for talent, falling interest rates, low cost of equity). Whether this was warranted or not is a separate discussion, but it has happened. 

Private Equity’s SBC Opportunity:

Private equity seems to have also identified SBC as a shadow source of margin compression in software companies. And more importantly, it’s margin compression that they can reverse (at least for a period). PE firms are using public companies’ inability to manage their dilution as an opportunity to acquire depressed software assets and meaningfully improve the economics for equity holders overnight. Since PE sponsors are the new shareholders post-buyout, they are primary benefactors. 

The impact is quite clear when you compare the SBC burden at pre-PE ownership software companies with post-PE ownership software companies. 

So private equity is creating value! Just not on the Non-GAAP profit metrics that software companies like us focus on.

For Thoma Bravo, which recently acquired Coupa, normalizing the company’s (rather absurd) SBC burden would reduce the annual dilution to equity holders by ~80% and increase the economic margin (EBITDA - SBC) by 22 percentage points. And in 2-3 years, Coupa could theoretically re-IPO with a reduced SBC burden. And no matter how one decides to treat SBC, the impact of normalizing SBC expenses should meaningfully benefit the equity.

One way to think about it would be that a higher % of all of the other value creation (revenue growth, margin expansion, deleveraging) accrues to the equity holders. 

Informatica is a great longitudinal case study because we can overlay management’s planned SBC burden at the time of the buyout with actuals under PE ownership4. 

Informatica’s management team planned to run the business with SBC representing 6% of revenues for a business growing 10-15% year-over-year. The PE sponsors ran the business at one-sixth that amount.

To put a finer point on how impactful the difference is: if you assume a similar grant price for both scenarios, the PE sponsor ends up owning 20-30 percentage points more of the equity than a public markets investor would have under management’s plan. Here’s a simple example that illustrates this based on Informatica’s stock-based expenses (plan and actuals)5:

You can argue that Informatica might have performed better as a public company. It might not have missed the cloud transition as much as it did, which would have justified the SBC burden, but it’s not clear that companies’ CFOs are flexing or can flex SBC as a public company. 

To be clear, employees getting shares in the company is a good thing. But private equity has demonstrated that the current levels of SBC at public companies are not a requirement to run a software business. 

SBC Burden on Valuation

Most sophisticated investors price this dilutive impact of SBC into the stock price in some form or another. Some burden the EBITDA/FCF with SBC costs, some just model the dilution, and some do both6. But the net impact is that the stock price is depressed by SBC, and companies with higher SBC burdens are penalized more. 

Investors know that companies like Coupa et al., they are going to have to give management and employees 10-15% of the company every year in SBC. And they are okay with it as long those equity grants are attracting strong talent that can help the company grow top-line 30%+ for a long period of time. But when it does not, the return math does not work – and this is the opportunity that private equity has identified. 

The common themes for the deals that private equity has done over the last 12-24 months have been:

• Decelerating top-line for companies that were at one point growing 25-30%+

• Some recent price dislocation likely driven by a miss on the top-line

• Cash flow positive with little/no leverage

• And a high SBC burden (>15% of revenues)

For software companies, decelerating the top line typically causes some degree of shareholder rotation. And the investors that are rotating into the stock typically care about per-share metrics, which means dilution matters quite a bit. 

For private equity sponsors, this dynamic enables a few things:

• They can acquire these software companies at a lower price because public markets are pricing in that they’ll get diluted 10-15% per year. 

• The intrinsic value to the sponsor is higher because they know they can remove 50-80% of the SBC/dilution overnight.

• Sponsors can add 12-15 percentage points of economic margin by changing the compensation philosophy and re-IPO a company with a right-sized P&L. 

• Add debt (because why not).

How long will this opportunity and dynamic last for PE sponsors? Probably not long, given the current public investor focus on SBC.

But one thing is clear: public software companies’ inability to control their SBC expenses is creating headaches for public investors on both sides. On one side, public investors are getting meaningfully diluted every year. And on the other side, private equity is using this dynamic to rob public investors blind and steal these companies in the public markets.

Private equity can use their change mandate to right size SBC burden overnight. Public investors cannot, and more importantly, the management teams at public companies do not necessarily want to undergo the pain of fixing SBC, right-sizing P&Ls, and losing key people in the eyes of the public markets. Private equity gives them a safe port to land the company and use the PE mandate to fix things.

Other musings on PE…

This is not to say PE firms are out of the water on some of these recent deals even if they remove all the SBC in the world. Most of their return to date was during a period when software multiples 2-3x’d, and some of the recent deals were completed at nosebleed prices. 

And there’s a whole other question of whether some companies can even return to the public markets (I think not, but that’s for a future post). 

Subscribe now

Appendix:

This is a sampling of deals that I was readily able to screen and get data for. I will proactively agree that there are likely deals missing here besides the ones I’ve decided to footnote.

Subscribe now

Spotify recently hosted their event for creators, Stream On. The company debuted a number of new features with a focus on the following:

• Discovery: TikTok-type feed, Previews, Shuffle

• Video: long form and short video for podcasts

• Creator Tools: Tools to engage with fans with video, polls, and Q&A

These are the biggest product upgrades the company has made since its IPO.

The changes could not have come at a better (or worse) time for Spotify. Spotify and Ek find themselves in a precarious position. The company’s growth has slowed, and prospective growth opportunities look questionable. And as growth decelerates, investors are increasingly scrutinizing the company’s inability to expand margins. 

The core KPIs suggest the company is beginning to hit a growth wall for the top line.

• The company is saturated in most markets, and subscriber growth is slowing.

• Pricing power has not been flexed (likely because of competition).

And the company’s bet on podcasts seems to have been a dud in terms of their ability to monetize and improve gross margins. In fact, podcasts have arguably been a distraction for management and continue to be a drag on gross margins.

Spotify’s Crossroads

Looking at the company’s financial model, the challenges/issues become readily apparent. The main challenge Spotify has is that they do not have many levers in its control anymore. Previously Spotify could focus on growing subscribers and driving ARPU. 

However, subscriber growth is/will continue to decelerate because Spotify is saturated in core markets. The company has not flexed pricing meaningfully, but it's unclear how much pricing power they have. 

The only thing that could move the needle for Spotify (outside of a favorable label negotiation) is driving more non-music activity and revenue on the platform. Today, non-music revenue primarily means advertising revenue on top of non-music content.

Podcasting was the right bet for Spotify. However, the company over-rotated into owned & exclusive (O&E) podcasts. Spotify became over-reliant on original content as the strategy to funnel podcast demand to their platform. 

There are a few reasons why Spotify’s O&E bet has not worked out – but the main ones, in my opinion, are:

• The success of new podcasts is not predictable (Archetypes, anyone?)

• Discovery and promotion for podcasts are challenging. And this is made much harder when the podcasts are exclusive to one platform that has less than 50% market share.

• Unlike TV shows, podcasts are not re-watchable, so Spotify found itself on the content treadmill, excluding the top 1-2 creators (e.g., Rogan).

Spotify effectively went too far in the direction of becoming the New York Times as opposed to what Substack is building. What Ek should have focused on was driving a better experience for listeners that naturally funneled all podcast demand Spotify, exclusive or not. But the company lacked the mechanism to effectively do that.

This brings us to this week’s announcements. 

The Spotify Podcast Standard

Most of the press coverage focused on the company’s decision to build TikTok-like features into the app. While I’m sure those features will help with the discovery and recommendation of new content for creators, transitioning user patterns to treat Spotify’s feed the same way one treats TikTok or Reels seems challenging. For one thing, Spotify lacks the diversity of content and creators to take attention away from either Reels or TikTok. 

In practice, what the company seems to be positioning itself for is the re-aggregation of podcast consumption around Spotify.

Over the last few years, podcast consumption has unbundled and disaggregated:

The discovery and video features that Spotify launched are designed to position Spotify as the central hub for podcasters to create and publish content for all three formats. This also has the benefit of allowing Spotify to transition itself from an audio app to a general entertainment platform that includes audio, long-form, and short-form videos.

By supporting all three formats and having artists build their videos and experiences around Spotify’s features, Spotify hopes to condition consumers to consume podcasts and related video content on Spotify’s apps. What the company is trying to (re) capture is the time consumers are spending watching podcasts on YouTube or clips of podcasts on YouTube.

Here’s Julie McNamara, Head Of U.S. Studios & Video, at the event:

Many of our top hosts are also pioneers in video podcasting, and it’s no surprise that adding video has really helped them connect with fans on a deeper level.

….

We’re seeing more and more creators starting to embrace video. In fact, it’s one of the fastest-growing areas of podcasting, and we expect that growth to continue. Right now we have over 70,000 video creators on the platform, and we have a lot more on the way.

Forbidden Fruits with Julia Fox and Niki Takesh will return for season two with an exclusive video edition – I mean who wouldn’t want to see what these two fashion icons are wearing every episode.

And we’re bringing Drew Afualo’s video podcast, The Comment Section, exclusively to Spotify starting on April 5th. We’re also rolling out limited video series from other top podcasters like Mindset Mentor with Rob Dial as well as innovative partnerships with Collab, Creative Juice, Get Engaged Media, Golden Child and Karat.

There are a few important implications for Spotify of bet on video:

• Spotify can re-capture impressions/users that were going to YouTube to watch podcasts.

• Allows the company to experiment with video ads, which have higher CPMs.

• The company can offer a unique in-app experience around video that can help funnel users from other podcast platforms to Spotify. 

On the last point, this comment from Bill Simmons caught my eye:

At The Ringer, we love to experiment with different formats. And Video is one that we’ve come back to over and over again. By doing side-by-side interview panels and instant reaction podcasts, we have mastered companion podcasting. We’re always launching fresh new shows that stand out in their own ways, and most recently we’ve gotten excited about an idea of developing a brand-new game show, which we’ll be sharing more info about very soon.

A game show! I would wager that the quiz show will include audience participation.

What Spotify is trying to do here is add functionality that allows podcasters to engage with consumers in more ways. From Spotify’s Head of Podcast Product, Maya Prohovnik

The new Spotify for Podcasters is available to everyone, in every Spotify market, for free—beginning now.

This platform is the foundation that finally enables us to build the future of podcasting—together with you, the creators. It makes it easier than ever to create and grow your podcast. And it also enables us to provide new, innovative features that help you better engage with your listeners.

That includes things like Polls, Q&As, video—and starting today, we’re also adding Podcast Chapters. This feature lets you create thematic sections in your episodes so listeners can easily browse through the topics you cover. Chapters are available now for all podcasts, all you have to do is include timestamps in your episode descriptions and they’ll show up on Spotify automatically.

And that's just the beginning of the features we'll be rolling out this year for all podcasters.

Not only are these features meant to benefit creators, but they are also meant to benefit Spotify. By creating experiences that are only available within (or optimized for) Spotify’s app, the company is subtly making the listening experience worse everywhere else. Imagine listening to a podcast on Apple Podcasts where the podcast host is pointing to Q&A or discussions that are in Spotify. That seems like a crappy user experience for the Apple user.

Alex Cooper (Call Her Daddy) effectively alluded to this scenario/dynamic:

I can create or upload episodes to publish instantly, I can add video, which is a crucial element of my show and helps me take full advantage of that new home feed Gustav shared. And I can add audience Q&As and Polls to my episodes, so I can hear directly from my real fans, and trust me, the Daddy Gang has a lot to say.

Now to be fair, the reason people make content for YouTube or post to Apple Podcasts is that consumers are watching content on those platforms, and creators want to be where the audience is. Spotify’s hope is that it can add functionality and listening/viewing experiences on Spotify that are not replicable on other platforms. A user could start consuming a video or piece of content elsewhere, but Spotify’s app would provide small “hooks” that drive users to Spotify to continue the experience. 

Another pertinent example is shows like All-In podcasts that actively choose to show charts/graphs during their episodes. If you’re on the Apple Podcast app and you hear someone on the show talk about a chart, it’s a pretty terrible listener experience – because you cannot quickly open up the app to see the video version of the podcast with the visual. 

Spotify’s most natural rival for this end-to-end consumption of audio is YouTube, but they’ve disaggregated the platform into YouTube Premium and YouTube Music. 

Podcast and Beyond

If Spotify executes the strategy well, I think we end up in a place where Spotify becomes an end-to-end entertainment app centered around podcast creators. But there are many natural extensions for these podcasters, such as live and narrative content that fits well into video format:

Here’s Alex Cooper again from the event:

I want to share how some of these features have made a big impact for me and my show. Video episodes of Call Her Daddy not only get more engagement than audio-only episodes: They’ve also helped me expand my artistic capabilities to formats like documentary-style reporting.

This strategy is obviously not a slam dunk, and with Spotify, it never is. But it is the first time in a long-time that the company is relying on product innovation and updates to drive the business forward as opposed to throwing money at podcast creators. 

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about.

This is not investment advice. Do your own due diligence.

Subscribe now

Microsoft’s AI Head Fake

The discussion around Microsoft’s recent endeavors in AI and investment in OpenAI has largely been focused on their self-declared war on Google Search. It’s an engrossing narrative. A sleepy incumbent (Google) is being outflanked by one of its arch-rivals. And you have Satya dropping fire quotes like this in interviews:

From Verge:

“I hope with our innovation they will definitely want to come out and show that they can dance. I want people to know that we made them dance.” — Satya Nadella, Microsoft CEO

Here’s the thing. If Search was really the thing Microsoft was going after, the last thing a prudent CEO would do is go on every news outlet and scream, “We’re going after them. And just as heads up, we’re going to put in an aggressive bid to become the default search engine in iOS”. 

Microsoft’s seemingly hell-bent focus on taking share from Google Search will likely turn out to be nothing more than a head fake.

For one thing, Microsoft knows its problem in search is rooted in distribution, not product quality or features. And of all companies, Microsoft is the least likely to be heedless of the fact that competing with an incumbent with a massive distribution advantage is unlikely to bear fruit.

Being the default position in search matters – and Microsoft is very disadvantaged in that category. Google dominates the desktop and mobile markets with its default position.

Adding a new feature on Q&A to address a small percentage of queries is unlikely to meaningfully move the needle in terms of getting Microsoft the distribution it needs. 

Even more puzzling would be the fact that search advertising is not even an attractive end-market for Microsoft anymore. The company has other more important growth vectors (Azure, Github, LinkedIn, Office365), all of which Microsoft is far better positioned to win than in Search. 

Focus and Making Google Dance

Don’t get me wrong, Nadella is earnest when he says he wants to make Google dance. But making “them dance” is primarily a mechanism to capture the attention of the two constituencies that have underpinned Microsoft's strategy for the last few decades: developers and businesses. 

And if, along the way, that means creating chaos at Google, that’s a cherry on the top. Who doesn’t like it when their old boss returns and starts shipping code?

Microsoft's main priorities are really elsewhere. Over the next 10 years, Microsoft’s revenue growth and equity story will predominantly be focused on just two things:

• Growing and taking share with Azure. As it relates to AI, that means ensuring Microsoft is the leading platform for developers to build, train, and deploy their AI models and applications. 

• Expanding within the enterprise and business customers. As it relates to AI, that means up-leveling their suite of applications with features (e.g., generative AI, a summary of meeting notes).

Over the next decade, ~90% of Microsoft's revenue growth will come from these two areas. Nothing else really moves the needle for the company. Microsoft knows this; coincidentally, 80-90% of Nadella’s performance stock awards in FY ‘22 were tied to those two things. 

Bing could 2-3x revenue scale, and it would only moderately move the needle for Microsoft, especially when you take into account the revenue share the company would need to cede to become the default search engine on Apple et al12.

All this is to say that the marketing blitz that Microsoft orchestrated over the last few months was to:

• Co-opt the leading AI company’s brand (OpenAI) 

• Position Azure to being the leading AI and ML platform for developers

• Highlight (coming) AI capabilities of Microsoft’s applications to their business customers

If Google should be worried about anything, it's really that Microsoft is outflanking them with regard to developer mindshare3 and the next-generation application suite.

Microsoft and Developers:

Developers have always been core to Microsoft’s strategy, especially around platform shifts. During the PC era, that meant convincing developers to build applications for the Windows OS. And during the browser wars, Gates was intensely focused4 on ensuring developers built to the Internet Explorer standard.

From Microsoft’s 1997 Planning Memo

Windows won the desktop OS battle because it had more applications than any other platform. We must make sure that the best Web applications and content become available for IE users first…

There are 6 million developers worldwide who use Microsoft development tools and technology. This is one of our key assets against Netscape. We must help them (ISVs and corporate developers) write to the ActiveX platform, so they develop the rich base of Web applications and controls that establishes the value of the platform.

The strategy fell flat, and the company more or less missed the internet wave. And a decade later, would miss the mobile wave. After missing both platform shifts, Microsoft was able to catch the cloud computing wave. 

For all intents and purposes, we're undergoing another platform5 shift driven by the maturation of AI and ML technology. For this platform shift, Microsoft's primary focus is to drive AI and compute workloads to Azure with an assist from OpenAI. 

And Microsoft’s pitch to developers is simple: Azure will be the best platform to build, train, and fine-tune AI models and applications.

From Nadella's 2022 Ignite Keynote Presentation

We have built the next-generation supercomputers in Azure that are being used by us, OpenAI, as well as customers like Meta, to train some of the largest and most powerful AI models.

Azure provides almost 2x higher compute throughput per GPU and near-linear scaling to thousands of GPUs, thanks to world-class networking and system software optimization. And for inferencing, Azure is more cost-effective than other clouds, delivering up to 2x the performance per dollar.

In Azure, we also offer the best tools across the machine learning lifecycle from data preparation to model management. Data scientists and machine learning engineers can use Azure Machine Learning to build, train, deploy, and operate large-scale AI models at scale — Nadella, Microsoft CEO

Microsoft's primary forms of value capture for this platform shift are Azure and the developer tooling around Github/VSCode (e.g., Github Co-Pilot).

How much value is going to be captured in this next wave is still to be determined. But if we were to use the distribution of cloud spend as a proxy, it would imply that the infrastructure layer captures at least 40% of the total value, and the model layer will capture some % of the value that previously accrued the application layer.

The “supply chain” or stack for AI applications will likely look something like this:

For Microsoft, the goal is quite clear: incentivize and evangelize Azure, OpenAI, Github, and VSCode as the standard for developers building AI applications.

Standards in New Technology Markets

In many ways, this strategy is working. Has any company in recent memory received as much earned media as Microsoft has?6 More importantly, ChatGPT seems to have captured the minds of developers – at least for the time being.  

Microsoft is fully aware of the importance of blitzing the developer market early. In new technology markets, companies and developers tend to centralize around “standards” quickly. 

From The Gorilla Game by Geoffrey A. Moore

Hypergrowth markets, in order to scale up quickly, will often spontaneously standardize on the products from a single vendor. This simplifies the issue of creating new standards, building compatible systems, and getting a whole new set of product and service providers up to speed quickly on the new solution set…Everyone wants its products because they are setting the new standard. Its competitors by contrast, must fight an uphill battle just to get considered. It makes for a huge competitive advantage.

For companies and developers, it’s a simple resource and time allocation decision. When implementing new technology, there are “economies of learning,” and once you learn/familiarize yourself with one vendor and tech stack, the bar to learn and implement another is very high, and a poor allocation of time during a period of rapid growth.

Microsoft’s strategy to help turn Azure and OpenAI into a standard is two-fold:

• Create a tight integration between Azure and OpenAI that drives superior performance relative to the other vendors.

• Offer the lowest price for the core model tokens. After the company’s most recent price cut, OpenAI’s API cost 1/10 of its competitors7.  

The main drawback of Azure’s current approach is that while OpenAI’s APIs are marketed as “General Availability,” getting access to the API is quite restrictive (by developer standards). Microsoft says that’s primarily because of their responsible AI policies. Open-source first platforms (e.g., Hugging Face) allied with the other cloud vendors may inevitably outflank Azure on this8. 

Applications, Enterprise, and Beyond

The other focus area for Microsoft is using OpenAI's “primitives” to up-level their current application suite to expand with its enterprise customers and appeal to a new generation of businesses.

OpenAI’s primitives (so far) are: 

• ChatGPT – LLM that can understand and generate natural language. 

• Codex – the language to code model, which powers Github Copilot. 

• Dall-E – the text-to-image generation model. 

• Whisper – the speech recognition and transcription model. 

Microsoft is integrating each of these primitives into the company’s many applications (i.e., Office, Teams, Dynamics365) with the goal of driving increased adoption of the company’s application suite. In classic Microsoft fashion, the company will likely raise the bundle price in a few years to capture the value of the new features.

From Nadella's 2022 Ignite Keynote Presentation

We envision a world where everyone, no matter their profession, can have an experience like this for everything they do. That includes sales. We are applying conversational intelligence to Viva Sales to transcribe customer calls, and automatically identify highlights and critical follow-ups to ensure better customer service. And Dynamics 365 is using the same capabilities to help sellers identify new opportunities and suggest next steps to close deals

If you missed it, Microsoft used Whisper and ChatGPT to launch a Gong competitor. Is building a Gong competitor really going to move the needle for Microsoft? No. 

But it’s a step in moving towards owning the productivity and collaboration stack for businesses – enterprise, SMB, old and new. Historically, Microsoft has dominated the legacy and enterprise IT/productivity market. But while Microsoft was transitioning its Suite to the cloud in the mid-2010s, G-Suite and a new generation of “core office” applications took hold, especially on the low end of the market. The company's investments in the Microsoft Graph, Office adjacent tools (e.g., Power Automate), and now AI features may make the product more appealing to more types of companies where AI features might have the most immediate impact. 

Overtime, we should expect Microsoft to (1) expand the primitives, (2) continue to integrate the primitives into its application suite, and (3) more aggressively market its application suite.

For companies adjacent to the productivity suite, Microsoft’s intense focus on up-leveling their application stack with OpenAI primitives should be a Code Red. 

If you’re finding this newsletter interesting, share it with a friend, and consider subscribing if you haven’t already.

Share

Always feel free to drop me a line at ardacapital01@gmail.com if there’s anything you’d like to share or have questions about.