BUT the last quarter has been one of the worst quarters in the history of software stocks (note: I didn’t say businesses). So it feels like there’s no other place to start.

1. What’s going on with software?

Software just had one of its worst months ever. Let me show you a graph with the stock performance of the canonical “system of record” software companies over the last two years:

But the fundamentals (revenue growth, margins, retention) haven’t materially changed in these companies.

So to be clear: this “software armageddon” is about the resetting of expectations on the future of software business models, not about the “death of software”. This resetting of expectations has led to software’s lowest multiples of the last decade:

This multiple compression has come primarily from two factors:

1. The narrative of “software as the perfect business model” led to very high multiples

The high gross margins, operating leverage, switching costs, and growth of software businesses led many to the idea that “enterprise software is the perfect business model.” Which, in turn, led to very high revenue multiples on these stocks. But this narrative ultimately overestimated the market size and competitive differentiation for many software businesses.

2. The perceived competitive advantage of software has decreased dramatically with the rise of AI coding tools

I wrote an essay last year on where I thought software business models were trending over the next decade. In short, as software becomes easier to build, software companies would have a lower competitive advantage, causing an increasing amount of value to accrue elsewhere in the value chain.

And I want to reiterate this point now: software is the hardest it will ever be to build today. If you play out the improvement of these models over a 5-10 year period, the cost to build software will continue to decrease dramatically.

This combination of high initial expectations and a clear catalyst to change those expectations has led to the “collapse” of software stocks.

2. So is software dead?

But in that same essay, I also tried to make the point that doesn’t mean “software is dead.”

Let me elaborate on that now. The beauty of enterprise software business models is twofold:

1. High operative leverage: write code once, sell it infinite times (with incremental costs for customer service, R&D, maintenance, sales costs, etc)

2. High switching costs: It is a hassle to move off of software, especially “systems of record” for large enterprises

On the latter point, I’ve been in many, many meetings with customers of large software products (Microsoft, ServiceNow, Databricks, Workday, SAP, etc), and it takes A LOT to get them interested in moving off these platforms. These can be multi-year migrations, which tend to be very painful. And more importantly, if they fail, it’s the project lead’s career on the line.

There’s a popular adage that enterprise software is like an annuity (once you sell it once, you get to collect that revenue for decades). As far as I can tell, for the large system of records, that’s still intact.

The challenge is where does growth come from going forward?

What the rise of AI coding tools means for incumbent software companies is that each new dollar of revenue has to beat out two competitors:

1. A product providing the same service (albeit likely with less functionality) but is much cheaper (given it was much cheaper to build with AI)

2. An AI product that provides a fundamentally different service than the incumbent software solution (who will try to build an AI product as well)

That’s a tough battle to fight! And there’s historical precedent here. Consistent with the history of technology, it’s often not that the revenue of the old technology shrivels up, but that its growth is eaten up by the new technology.

So in summary: High initial expectations + the fear of AI coding tools killing software’s competitive advantage led to the collapse of software multiples. But the switching costs of large enterprise software tools mean most of the existing revenue base will be safe, at least for a while. The question is where growth will come from going forward, and the answer is not clear.

But given the topic of this article, what does this mean for the hyperscalers?

3. What does this mean for the hyperscalers?

The AI infrastructure buildout and the hyperscaler CapEx funding it have been the biggest news in tech over the last few years.

This “death of software” question helps explain why hyperscalers are spending so much to fund their AI buildouts. The logic goes like this:

1. Risk of disruption (stick)

   1. Many of the hyperscalers’ largest customers (outside of the AI labs) are large software companies.

   2. The hyperscalers themselves are among the largest software companies in the world.

   3. This means that if AI does indeed start to capture revenue that was previously going to enterprise software applications, the hyperscalers must capture a piece of that revenue to maintain their exposure to the future of enterprise technology.

2. The amount of compute necessary to fund this expansion is nearly unfathomable (carrot)

   1. Outside of the semiconductor ecosystem, the hyperscalers are the clearest and most immediate beneficiaries of AI.

   2. Microsoft reported that it has roughly $300B in RPO (contracted revenue) from OpenAI.

   3. Presumably (though not confirmed), the resurgence in growth for AWS and GCP is coming from contracts with large AI labs.

3. Combined, these companies have ~$400B in cash and marketable securities on their balance sheets and generate $100B in net income per quarter (no reasonable alternative investment).

This risk of disruption (stick) + immediate revenue opportunity (carrot) + no reasonable alternative for investment continues to lead to staggering investment numbers!

4. This dynamic is leading to a staggering amount of CapEx in 2026 (and over the short term, will be the largest driver of market share changes)

If we look at announced numbers and analyst projections, these four companies are expected to invest roughly $630B in Capital Expenditures next year:

Given the immediate revenue opportunity from funding model training and application inference, this will be the biggest driver of market share changes for the foreseeable future. See the growth rates of these cloud companies below:

For companies with this much revenue (a combined $320B in revenue run rate), these growth rates are just incredible. We can see the updated estimated market share here between the three large cloud providers:

The more interesting metric, though, is market share “momentum”, or net new revenue generated Y/Y. In short, what % of NEW revenue generated over the last four quarters did each cloud company generate?

Again, AWS and GCP had especially strong quarters.

5. One Final Note on Competitive Advantage Periods

In times of technology booms, it’s easy to get caught in the short-term results.

One of my favorite investing papers is Competitive Advantage Period: The Neglected Value Driver by Michael Mauboussin. In summary, moats are discussed frequently in investing. But the durability of moats is equally important.

The question at the crux of this moment in software is about durability. The software multiple collapse is really just investors saying: these moats won’t last as long as we thought. For hyperscalers, the strategy right now is to extend the duration of their competitive advantage as long as possible.

I’m not foolish enough to predict how this plays out, but it’ll be fascinating to see how much clarity (or lack thereof) we have by next quarter. Til then!

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

So my goal for this publication in 2026 is to make everything I publish here worth reading. Attention is scarcer than ever, and I see it as my responsibility to respect all of yours.

With that in mind, why write an annual letter in the first place (by definition, a time-constrained, not quality-constrained artifact)?

A few reasons:

1. I have a special affinity for the annual letter. I’ve gotten tremendous value from reading the greats’ musings on investing, and my goal is to provide a bit of value in mine

2. It’s a selfish endeavor. As people outsource more thinking and writing to AI, the value of doing it well increases dramatically. While I make no claims to either, structuring my thoughts on common investing mistakes has been quite valuable for me

3. In a world where more knowledge work is being automated, good questions and ideas become more valuable than good answers

Onto the topic of the annual letter (what I believe is a good question): why do investors so often underestimate great companies, and how can we avoid those traps?

An Estimation of Underestimations

Investing, by definition, is a game of underestimations: seeing something that others haven’t, OR acting in a way that others don’t. Our job is to generate value by taking advantage of that underestimation.

Plus, as a midwesterner from a town of 500 people in Indiana, I can’t help but have a soft spot for the underestimated.

I’m summarizing my investment thesis these days very simply: I’m looking for the underestimated. Underestimated people. Understand businesses. Underestimated opportunities.

So what more valuable topic to discuss than an observation of the most common underestimations I’ve seen in the investment industry, an estimation of underestimations if you will!

I’ll be framing much of my investment philosophy today through the following underestimations1:

1. The Influence of Consensus

2. Incentives

3. How much History Rhymes (and how little it repeats!)

4. Probabilities & a Focus on the Inputs

5. The Long-Term: Destination Analysis, Compounding, and Concentration

6. The Dynamic Range of Humans

By definition, an underestimation can only exist if others aren’t properly estimating it. Which makes the influence of consensus the foundation for the rest of this essay2.

One: Underestimating the Influence of Consensus

From my observation, the most common investment mistakes come from overweighing others’ opinions on an investment. In the words of Neil Mehta, “the laws of great businesses are the laws of great businesses.”

This is especially prevalent in private markets where information is less available; so when there’s uncertainty, investors tend to look elsewhere for confidence. If they see it from other investors, a fear of missing out takes care of the rest. And if they don’t, they hesitate, as it’s unnatural to make decisions under uncertainty.

But I’ll argue it’s the lack of analysis, not the lack of information, that’s the most frequent bottleneck. From Nick Sleep’s 2012 letter:

In today’s information-soaked world there may be stock market professionals who would argue that constant data collection is the job. Indeed, it could be tempting to conclude that today there is so much data to collect and so much change to observe that we hardly have time to think at all. Some market practitioners may even concur with John Kearon, CEO of Brainjuicer (a market research firm), who makes the serious point, “we think far less than we think we think” - so don’t fool yourself!

If the world was information-soaked in 2012, it’s drowning in 2025!

The analysis (i.e., the independent thought) is the hard part; so when there’s “consensus” from other investors, choosing that company becomes “the easy path.”

At the same time, most investment firms are not structured well for investors to make non-consensus decisions. Personal “mistakes”3 are costly in the investment industry, so herd behavior is also “the safe path.”

That said, there’s reason for optimism! The benefit of the investment industry’s enduring herd behavior is that there’s plenty of opportunity to think for ourselves. And as people outsource more of their thinking to LLMs, I suspect that opportunity will become even more enduring.

Two: Underestimating Incentives

People tend to act in their own best interests; ergo, incentives are the driving force in our world. Incentives are like gravity, and fighting against either should be viewed similarly (it’s going to be tough!).4

The epitome of this is the US healthcare system, based on a fee-for-service model. This incentivized providers to provide as much care as possible, not the right amount. And it incentivized payers (insurance companies) to reimburse only the care they think is required.

This led to an incredible burden of documentation on providers to prove they delivered “the right amount of care”; and as a result, a ton of admin waste throughout the whole system (more than a trillion dollars a year!)

On a more optimistic note, I spend much of my time with healthcare startups, as LLMs have created the ability to automate much of this admin work (due to the fact it’s based on nuanced text-based information). This cuts costs for providers AND payers, and it provides a much better experience for clinicians and patients.

It aligns incentives for everyone involved!

When a company aligns incentives for all parties involved, and even better has a compounding competitive advantage with it, underestimate it at your own risk!

In summary, some wisdom from Charlie: “Never, ever, think about something else when you should be thinking about the power of incentives.”

Three: Underestimating how much history rhymes (and how little it repeats!)

It’s no surprise to anyone reading this how much I value history. That said, it’s surprising to me how often it’s misrepresented. It’s human nature to look at the immediate past and extrapolate. But it seems to me this leads to far more mistakes than not.

To get a bit practical for VCs, the two most common analysis mistakes I see:

1. Underestimating market size because the current market is “small”

The lack of a large existing market today leads investors to underestimate a startup’s ultimate market size, especially when there’s a catalyst to drive market expansion.

My favorite example is Aswath Damodaran’s now-infamous estimate of Uber’s market size and Bill Gurley’s response:

A decade ago, Aswath Damodaran (The Dean of Valuation) estimated Uber’s value to be approximately $6B based on the total TAM and market share. Bill Gurley argued that this number was off by a factor of 25x based on an underestimation of both variables. Uber is now worth $200B (33x Damodaran’s estimate)

In short: latent demand for a service, and a new technology that unlocks the ability to service that demand, is a wonderful thing. Regardless of “existing market size”.

2. Underestimating a company’s value due to “low multiples” or “public comps”5

A common valuation practice is to take today’s revenue or projected revenue over X years into the future, then apply a “reasonable” multiple to that revenue to arrive at a potential valuation. Investors typically get that multiple from “comparable” companies in the same industry.

But this leads to comparing a new company, potentially with a completely different cost structure and competitive advantages, to an incumbent.

The classic example here is Tesla being compared to existing auto companies (and their multiples), leading many investors to underestimate what Tesla could create with new technologies and cost structures.

And as a reminder, multiples are a short-form proxy for the cash flows a company will generate going forward: a second-order effect of business quality!

So companies with higher multiples tend to be much stronger than companies with lower multiples. Which means, all else equal, you’d rather invest in a startup with “low multiple public comps” than high multiple ones. The exact opposite of what much of the investment industry does!

All in all, both of these ways of thinking typically lead to quite a lot of underestimations for companies without a direct historical precedent.

Four: Underestimating Probability-Based Thinking

More often than not, when an investor is discussing investing in a company (especially startups), it’s a matter of “Do I think this company will be successful or not?”

The error here is that our world works in probabilities. Each decision is based on a tree of possible outcomes, with probabilities attached to each outcome. And by using absolutes, you increase the odds you’ll underestimate an outlier company.

Let’s look at Nvidia’s recent $20B pseudo-acquisition of Groq as an example. When the company was founded, if you asked the question, “Will this company work?”; Most investors would’ve said no, saying Groq will fail because the technology is complex, the costs to develop semiconductors are high, go-to-market is hard for semiconductor startups, and Nvidia is the Hulk.

They probably would’ve been right, and they would’ve missed the investment!

So you’d get much closer to making the investment if you asked the question, “What’s the probability Groq will be successful and/or disruptive to Nvidia AND what will happen if so?”

Now, a general equation for investing in a company goes like this (and this just means the sum of all the probabilities times the reward for each probability):

So the question “Will this investment work?” isn’t the right question at all! The question is “what’s the probability this investment will work and what outcome will come of it if so?”

In VC, in particular, an interesting situation arises in which a small number of companies drive the vast majority of returns. So in the above equation, the only probabilities that matter are the ones attached to very large outcomes.

So, in the words of Bill Gurley, “What can go right?” becomes the most important question to ask about a startup.

A note on inputs over outputs:

Implicit in probabilities is a focus on the inputs, not the outputs.

To continue the Groq example from above: let’s say Groq does fail to raise their last funding round and goes out of business. People would’ve seen that investment as a failure instead of a huge success!

So the key point is this: it’s the same investment regardless of the outcome.

Focus on the inputs you can control, let the outputs take care of themselves. Any mistakes along the way should be viewed as learning opportunities. As Nick Sleep said in his 2008 letter:

We would counsel you to think about the inputs to investing rather than the outputs. It is in times like these that the hard psychological and analytical work is done and the partnership is filled with future capital gains: this is our input. The output will come in time.”

Five: Underestimating the Long-Term

I continue to be convinced that a longer time horizon is a meaningful competitive advantage in today’s investing world. There are a whole host of reasons to not invest in a great company in the short-term (and often lots of good reasons to invest in not-so-great companies in the short term!)

To combat this, I operate under a few assumptions:

1. A small number of companies create the vast majority of value (this is true in both public and private markets!)

2. Most of that value is typically created “later” in the company’s life (meaning durability is just as important as growth, in the words of Peter Thiel)

3. Those companies tend to have a compounding advantage that makes them better with time, and that compounding advantage is identifiable earlier than most realize

Some might call that last point “destination analysis.” Said differently, it can be easier to predict the long-term than the short-term.

And when you have figured out the long term (or a decent probabilistic view of it and a willingness to update those probabilities with time), let compounding do its magic!

Now it’s human nature to focus on the short-term, but it’s also human nature to underestimate the value of compounding; for some reason, we just struggle to internalize exponentials. Even with the risk of suffering from hindsight bias, I’ll share a few data points:

1. From Warren Buffett’s Biography: Several times every year a weighty and serious investor looks long and with profound respect at Coca-Cola’s record, but comes regretfully to the conclusion that he is looking too late. [For the record, when that book was published in 1995, Coca Cola’s market cap was roughly $45B. Today it’s $300B.]

2. Shopify, at nine years old, IPO’d at a roughly $1.3B valuation in 2015. It’s now worth $210B (a 161x for anyone counting).

3. Nvidia was worth roughly $15B in 2015, an incredible outcome for a company only 22 years old. It’s now up ~293x that, in years 22-32 in founding.

One more: take a look at the price you could pay for Wal-mart and still get a 10% annual return!

It’s easy to underestimate how long it takes for a great business to emerge, and it’s easy to underestimate how large the truly great ones will become.

A note on underestimating concentration:

Once you’ve found one of these incredible businesses, it’s worth concentrating deeply in those businesses. They’re even rarer than we think.

To abstract the point a bit: in a world where distractions are greater than ever, it’s never been more valuable to concentrate on fewer ideas.

When there’s so much information, adverse selection is a very hard problem to solve, and can only be accounted for with deliberate focus.

The most readily available information may not be the most valuable. The most available opportunity often is not the most important. The company you can invest in may not be the one you should invest in.

Six: Underestimating the Dynamic Range of Humans

Finally, it’s easy to miss how truly special people can influence the world around them.

I’ll take this idea a step further, take this quote from Warren Buffett:

“It was an agrarian economy a couple hundred years ago…Very hard, you know, to get 20 times the wealth of the next guy because you were a little bit better farmer. But if you’re better at some skills now, you can become incredibly wealthy at a very young age … You get to capitalize [the] value of an idea. And so the wealth moves big time, even on an anticipatory basis.”

I’ve made it this far without mentioning AI too much, but I’m afraid I’ll have to break that celibacy here. From my observation, the discovery of transformer-based models are new, but artificial intelligence isn’t. It’s the continued evolution of 50 years of computing, slowly automating knowledge work to make us more efficient.

In summary, giving us leverage.

As I think about AI’s impact in 50 years, I don’t think AI will automate all of human knowledge work (consider me a humanist!), but I do think the top percentile of people will become even more leveraged.

With that, the dynamic range of humans will expand even further. Outside of the societal implications of this furthered divide, it’s led to one operating assumption I’m moving forward with:

Find the people who will be in the long tail of the dynamic range of humans, and figure out how to work with them for the next 50 years.

I feel grateful to have started that journey (thank you to the incredible folks who have invested time in me), and I hope to continue it for decades to come.

If there’s anyone left reading (surely there’s someone!), I’ll leave you with one point: these ideas are easy to say, and hard to do. There must be a conscious effort every day to pursue them.

I’ve no doubt that as I go back and read this letter in a year, it’ll seem obvious and rather rudimentary. I hope so, at least. Perhaps it means I learned something.

In the words of Nick Sleep, ”When I finish writing these letters, I am happy with them, at least happy enough to send them out. I make a habit of re-reading them sometime later and frequently they disappoint. It is as if the letters have a short shelf life.”

I’m going in eyes wide open!

I feel compelled to end this letter as I did my last one, with a thank you to those willing to invest their time into me (including those who read this newsletter regularly).

I don’t take the opportunity lightly to do what I love for a living. In large part, that’s because of the support from you all and those who have heard my thoughts and deemed them worthy of a bet. Thank you.

Onwards and Upwards. To 2026.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

If we take a step back, just over two years ago, OpenAI had built an incredible research team and consumer product, but it felt like they still had a long way to go to be viewed in the same light as the other tech giants.

Now, that’s changed. OpenAI and Anthropic have helped catalyze the largest computing infrastructure buildout in history. Evidenced by the scale of the deals they’ve signed with the hyperscalers in the last quarter, they’ve become the two most important cloud customers in the world.

In summary, they’ve created their own segment of the tech value chain, and put themselves at the center of the largest players in tech.

Below are some thoughts on what that shift means, along with updated data (market share, revenue growth, CapEx) from the hyperscalers.

1. The Value Chain is Shifting

All else equal, it’s best to be the consolidated piece in the value chain, and sell to a fragmented set of customers (less competition, more pricing power).

Before OpenAI and Anthropic’s rise, the cloud providers were selling to a massively fragmented buyer base. Even the largest customers made up just a fraction of cloud revenue.

That’s changing. Let’s take a look at the slew of announcements over the last quarter or two:

1. OpenAI’s $300B deal with Oracle

2. OpenAI’s $250B deal with MIcrosoft

3. OpenAI’s $38B deal with Amazon

4. OpenAI’s deal with Google

5. OpenAI’s $100B deal with Nvidia

6. OpenAI’s equity deal with AMD

7. Anthropic’s deal with Google

8. Amazon’s data center investments for Anthropic

OpenAI and Anthropic have now positioned themselves squarely in the middle of the tech giants, and aligned incentives for everyone to benefit from their success. I’ll just leave you with this Paul Graham quote about Sam Altman (back in 2008):

“When we predict good outcomes for startups, the qualities that come up in the supporting arguments are toughness, adaptability, determination… Sam Altman has it. You could parachute him into an island full of cannibals and come back in 5 years and he’d be the king.”

It certainly feels like OpenAI is king right now.

Along with Anthropic, they’ve created the largest greenfield (net new revenue opportunity) creation the hyperscalers have ever seen. Take a look at AWS’ re-acceleration here, driven by the growing partnership with Anthropic:

It may not look that drastic, but for AWS, that’s at a $132B revenue run rate! According to Semianalysis:

AWS has well over a gigawatt of datacenter capacity in final stages of construction for its anchor customer. AWS is building datacenters faster than it ever has in its entire history. And there’s much more on the horizon.

I can’t confirm this, but it’s likely Google’s acceleration is driven by Anthropic as well.

In summary, the foundation model providers have created the largest greenfield revenue opportunity ever for the hyperscalers. And the question becomes: what are the hyperscalers willing to do to capture it?

2. The hyperscalers are all capacity constrained, but how much CapEx are they willing to invest, and at what economic profile?

Hyperscaler CapEx nearly hit $100B last quarter. If we include data center leases ($11B+ from Microsoft last quarter), it well exceeds $100B.

For all the talk of a bubble, the hyperscalers are still capacity-constrained today. And the hyperscalers are investing to ensure they maximize the amount of durable, recurring revenue they can capture.

Now I’m still of the belief there will be a glut at some point (where we overbuild for a period of time, leading to decreased investments). For what it’s worth, so is Sam Altman:

“There will come a glut for sure and whether that’s in two to three years or five to six I can’t tell you, but it’s going to happen at some point, probably several points along the way. There’s something deep about human psychology here and bubbles.”

But demand is simply greater today than many were expecting (including myself, and shame on me, I spend my days talking to early-stage startups growing like wildfire).

Take a look at the RPO numbers:

• Oracle: Q1 Remaining Performance Obligations $455 billion, up 359% in both USD and constant currency

• Microsoft: Microsoft Corporation Cloud revenue surpassed $49 billion, up 26% year over year, and our commercial RPO grew over 50% to nearly $400 billion with a weighted average duration of only two years.

• Google: Google Cloud’s backlog increased 46% sequentially and 82% year over year, reaching $155 billion at the end of the third quarter.

• Amazon: Backlog grew to $200 billion by Q3 quarter end and doesn’t include several unannounced new deals in October, which together or more than our total deal volume for all of Q3.

$1.2T in backlog for these companies! The hyperscalers will continue to invest until they see signs we’re close to an equilibrium of supply and demand. Quite simply, this will be the largest driver of market share changes for the foreseeable future.

3. For the time being, the largest driver of market share changes will be foundation model revenue.

You could argue the biggest driver of market share changes over the next few years will be who can capture the majority of this greenfield opportunity from OpenAI and Anthropic. We’re already seeing that start to take hold with Microsoft’s market share gains from OpenAI’s computing needs, already generating tens of billions in revenue for Azure.

A better way to view market share changes is through Y/Y net new revenue: of the net new cloud revenue generated over the last four quarters, how much did each capture?

Through that lens, it becomes clearer how much Azure benefited from the OpenAI spend and how important the Anthropic partnership is for Amazon:

It should go without saying, but it’s also important to call out that the profitability of that revenue is important!

4. The economics of that revenue matter.

It’s easy to get caught up in the short-term market share changes from CapEx investments, but it’s also critical to remember that the revenue needs to be durable and/or profitable (should be both) for it to be high-quality revenue for the cloud vendors.

This leads to the risk-reward equation for the hyperscalers: what are they willing to do to capture this revenue opportunity?

Microsoft has been the most vocal about being selective with the types of workloads they take on, and they continue to drop hints of how they’re handling the risk of these CapEx investments.

1. Derisking through investing in short-lived assets and leases, reducing the long-term balance sheet exposure as much as possible.

The way to think about that, and you saw it this quarter in particular and as we talked about twenty-six, the remainder, number one, we’re pivoting toward increasingly, we talked about this short-lived assets, both GPUs and CPUs. Again, we talk about all these workloads are burning both. - Satya, Microsoft Earnings Call

2. It’s okay to be early, as long as investments are within the ballpark of reason, compute demands will eventually catch up. They’re okay with overbuilding by a few quarters.

So we spent a lot of time building out that infrastructure. Now we’re continuing to do that also using leases. Those are very long-lived assets as we’ve talked about, fifteen to twenty years. Over that period of time, do I have confidence that we’ll need to use all of that? It is very high. - Satya, Microsoft Earnings Call

3. They’ve underestimated compute demand so far, and don’t want to underinvest.

[Demand] is very high. When I think about sort of balancing those things, seeing the pivot to GPU/CPU short-lived, seeing the pivot in terms of how those are being utilized. We are, and I said this now, we’ve been short now for many quarters. I thought we were going to catch up. We are not. Demand is increasing. It is not increasing in just one place.- Satya, Microsoft Earnings Call

In summary, Microsoft is doing everything they can to capture ~durable~ greenfield opportunity without putting their balance sheet at risk to do so. Amazon and Google have talked less about this publicly, but we have to assume they’re having the same conversations in their boardrooms.

Satya alluded to this in the BG2 podcast:

Sometimes you may have to say no to some of the demand including some of the OpenAI demand. Because sometimes Sam may say ‘hey, build me a dedicated multi-gigawatt data center in one location for training.’ Makes sense from an OpenAI perspective, doesn’t make sense from a long-term infrastructure buildout for Azure.

It seems he was alluding to Oracle in this comment, which brings me to my final point.

5. Expected value vs expected utility is the final piece of the puzzle to understand these CapEx investments.

So the lens of huge greenfield opportunity + how much investment risk are the hyperscalers willing to take on to capture that opportunity explains much of the changes happening in the cloud market.

But it doesn’t fully explain the last update I haven’t discussed: Oracle’s MASSIVE RPO backlog and commitment from OpenAI.

And they’re dipping into the honeypot of leverage to get into the AI CapEx game. From Doug O’Laughlin:

There is no way for Oracle to pay for this with cash flow. They must raise equity or debt to fund their ambitions. Until now, the AI infrastructure boom has been almost entirely self-funded by the cash flows of a select few hyperscalers. Oracle has broken the pattern. It is willing to leverage up to hundreds of billions to seize a share.

So why are they willing to do this when others aren’t?

My interpretation is this: the other hyperscalers are making their best guess expected value calculation while Oracle has an expected utility calculation they’re making.

Warren Buffett explained expected value simply here:

“Take the probability of loss times the amount of possible loss from the probability of gain times the amount of possible gain. That is what we’re trying to do. It’s imperfect, but that’s what it’s all about.”

BUT let me introduce the idea of expected utility. From Michael Mauboussin:

Economists typically translate expected value into expected utility, an idea that Daniel Bernoulli, a mathematician, introduced in 1738. Utility is a measure of satisfaction and varies from person to person based on individual preferences.

In uncertain conditions, individuals will optimize for expected utility rather than solely expected value. For example, if a favorable casino offered a gambler a bet that they have positive expected value on, BUT the catch is they have to bet their home on it. A person with one home should not make that bet, but someone with ten homes should. (I have no doubt this is deeply flawed to any economist reading this article.)

Whether or not Amazon, Google, and Microsoft lever up to invest even more in this boom, there’s no additional utility for them. They’ll still be one of the big three cloud vendors.

But Oracle, by taking the additional risk with leverage, DOES get expected utility from this. They can move from being a distant fourth-place cloud vendor, to becoming one of the “big four” cloud vendors.

So when you ask Larry Ellison if he’d rather play it safe and stay where they’re at, or make the bet others aren’t if the result is a state change in how Oracle is viewed, he’ll take that bet every time.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

Admittedly, I wanted this article to be a follow-up to my last piece on Free Software. MY GOAL was to learn lessons from the original “free software” companies (i.e., those monetizing open source).

However, after studying the most successful open-source businesses of the last few decades, I want to be honest that the takeaways aren’t as neat and tidy as “playbooks on building open source companies.”

In reality, open source is a wonderful strategy to solve the cold start problem in infrastructure companies: getting people to try it. But getting people to pay for it is very hard. And then protecting that value from competitors is even harder.

So, I’m sharing my notes on the big question in open source: Open source creates so much value for the world, but how can companies best capture that value?

Two disclaimers: I’m biased towards open source businesses, I think they’re good for technology, and I think they’re a great zero-to-one mechanism for building a tech business. Secondly, I’ll be discussing generalized observations in this article. Anomalous outcomes from anomalous companies, so the best companies likely defy the generalized observations I’ll make here.

I. How Have Open Source Companies Historically Captured Value?

Open source software, software developed and given to the world for free, has risen in importance and adoption for the last several decades. The godfather Bill Gurley started writing about it over 25 years ago.

Ever since IBM unbundled hardware and software back in 1970, there’s been an ideological movement around open source (see “Why Software Should be Free” from one of the pioneers of open source). Proponents of “free software” had the belief that distributing it as widely as possible moves technology forward as fast as possible.

But there’s a misalignment with this belief: the best way to make change in a capitalist society is to align financial incentives. For forty years now, companies have been trying to figure out how to align those financial incentives.

Many have done so very successfully (none more so than Databricks, who just raised at a $100B valuation).

The core premise of modern open source businesses is this:

1. Developers lead software adoption in the enterprise, which creates a cold start problem:

   1. You need developers to bring software into the enterprise, but you need the enterprise to buy software for the developers.

   2. Open source solves that by getting developers to use free software, and creating a wonderful market mechanism in the process.

   3. This is why open source is so dominant in databases and developer tooling.

2. But by giving away open source for “free”, it makes it much harder to eventually charge for features. “Why should I pay for a tool I can get for free?”

3. But the reality is that (1) enterprises have unique needs because of the complexity of their deployment and the security necessary for a successful deployment, so (2) the cost of open source is not free.

4. To solve for that, most open source companies give away a core product that can be used by anyone, but will charge for additional enterprise features (security, advanced features, cloud hosting) and/or services (helping customers deploy software).

5. However, the core problem is building sustainable technical differentiation, so eventually open source companies keep more and more of their features behind a “paywall.”

And honestly, this has been the monetization method for the majority of open source companies!

Charge for either services, additional features, compute costs & hosting, or a platform. And to spare you 1500 words on the history of open source (and yes, that was draft #1), I’ll just highlight a few of the companies who helped pioneer those business models:

1. Redhat and Linux: Offer services (customer support, manual updates) to enterprises who had complex needs

2. GitHub and Git: Offer hosting services and use network effects as a moat

3. MongoDB and…MongoDB: Use a free database as a marketing funnel to close enterprise clients who need additional features (Security & Authentication, Operations & Management, Performance & Scaling, High Availability & Disaster Recovery, Compliance & Governance). Confluent, Elastic, and others would follow a very similar playbook.

4. Databricks and Spark: Offer a platform of open source technologies (Spark, Unity Catalog, Iceberg, MLFlow, Delta Lake), and charge for the convenience and performance of the enterprise offering. Databricks would develop more and more features in the Databricks platform to create differentiation.

Here are many of the others that have been successful:

Examples of Most Successful Open Source Companies and How they Monetized

All of them solved some core problem FIRST and monetized after (and most said it was brutally challenging to do so). Then they added on additional features, and the most successful created network effects or platforms in addition to those features. But it’s never been easy.

The challenges of monetizing open source:

Ali Ghodsi summarized the challenge of open source best:

“I sort of liken it to playing baseball and having a strategy that you’re going to hit two home runs in a row after each other consecutively. The first home run is going to be your open source home run. You’re going to give away the software, and it better be a home run. “Home run” here is defined as the whole world loves it and downloads it, otherwise you’ve accomplished nothing, you’ve just given away your intellectual property for free and accomplished absolutely zero. So that’s home run number one…Then you need to hit another home run in Peter Thiel’s parlance another Zero to One that’s 10x better than the open source project.”

So this tells the abridged history of open source business models:

1. Open source is almost a necessity in developer tooling and data infrastructure.

2. It provides both wonderful marketing and a great feedback loop.

3. But it requires an incredible technical team in order to release the amount of quality features to solve customer problems and marketing team in order to get people to pay for the tool

The challenges of protecting open source value:

And finally, protecting that tool from competitors is not easy. We’ve seen every major open source company put out some defense against the hyperscalers. Many have done so through licensing and adding additional features.

MongoDB is a great example. It was originally released under the GNU Affero General Public License (AGPL). But this didn’t prevent companies like AWS from taking that software and releasing their own version of software. So in 2018, MongoDB re-released their database under the SSPL license, which prevented other companies from re-distributing for commercial purposes.

This was largely in response to the hyperscalers offering competitive products using MongoDB’s technology. And it’s brutal to compete with the hyperscalers because they can offer that product (in theory) for free and just make their profits on the compute costs.

The second path, the one that Databricks has most successfully pursued, is just building more and more features. Ali Ghodsi talked about their initial pivot into this here:

So at Databricks we had Spark and the cloud vendors were just giving it away and beating us to it, so we had to come up with something that’s 10x better, so we came up with something called the Photon engine, which is a complete rewrite of the engine under the hood.

So then the value proposition would be, “Hey, you already love Spark, you’re already using Spark, I’m getting the cost of it down to a tenth, and we can split the difference”. The initial reaction of people would be, “Well, wait a minute, that’s not open and you’re not giving away for free anymore”, and it’s like, “Yeah, but we need to make money somehow”, and they said, “Okay, if you can prove to me that you really can lower the cost by that much, I’ll split the difference with you”. That became the strategy, but we only came up with it in hindsight.

Important note: this requires an incredible technical team that can continue to innovate faster than anyone else in the market.

The ability for the hyperscalers to release a competitive product at a lower cost and profit from compute costs points to another observation: many of the most successful applications of open source have come from their strategic value, NOT from charging for an open source product.

II. Open Source as a Strategic Tool

So open source is hard to monetize, but why did Databricks acquire Tabular for $1B+? Why is Meta open-sourcing Llama? Why is OpenAI open-sourcing their trailing-edge models?

Because FAR more value has been created from using open source as a strategic weapon than an offensive. This ‘commoditize your complement’ strategy is straightforward: if you can commoditize one piece of the value chain, value will flow to another. And open-source is a wonderful commoditizer (it’s literally giving away software for free).

In many of these cases, a big tech company has a dominant core business that no other business line can compete with. So it makes far more sense to open-source a product that protects the core business than trying to charge for it.

Google’s been the mastermind behind many of these:

1. Acquiring Android to ensure they don’t lose a key access point to search

2. Open-sourcing Chromium to ensure they don’t lose a key access point to search

3. Open-sourcing Kubernetes to make it easier to switch off of AWS

And this ‘commoditize your complement’ strategy is the primary one we’re seeing today with Open Source AI.

AI & Open Source Today

So the lessons from open source are: open source is an amazing marketing tool, it's a great way to get customer feedback, getting people to pay for it is really hard, and it's a great way to commoditize your complement.

Just a few examples:

1. Why would Meta spend billions on Llama without charging for it?

Meta just wants control over a leading-edge model. They have the monetization engine and the distribution to profit from leading-edge models. However, if they are completely reliant on another model, then value could accrue to that vendor instead of Meta’s distribution.

By open-sourcing Llama, they're hoping no single competitor can establish a monopoly over AI models while protecting their core social media and advertising businesses.

2. Why is Nvidia investing so much in software?

Clearly CUDA gets a lot of attention. But Nvidia continues to develop open-source AI software infrastructure from models (Isaac Groot, Cosmos) to Agent Frameworks (Agent Intelligence toolkit, NeMo). Again, the logic is that Nvidia will never earn as much on software as they do on hardware, but by giving away software for free, they can lock in developers to their hardware.

Commoditize your complement.

3. Why did OpenAI open source their model?

Two-fold. First, open source is great for marketing, and OpenAI is seeing around corners for the unrest that closed-source AI could bring. By open-sourcing their trailing-edge models, they get ahead of that.

Secondly though, OpenAI is monetizing primarily through the application layer, the clearest complement to foundation models. So, in theory, they can survive and thrive without charging for their APIs (this won’t happen in the near term, just a theoretical point).

III. Value Creation vs Value Capture

This debate is not a new one to anyone in the open source community. Open source is a strategy that’s been successful at what it’s good for: getting developer-first businesses from zero to one, gaining mindshare, and strategically commoditizing pieces of a value chain.

We continue to see new successful companies built around open source (take Felicis investments n8n and Supabase for example). And we continue to see successful attempts at commoditizing complements.

While all that is true, far more value has been created from open source than captured. Many open source projects didn’t benefit anyone; they benefited everyone.

For the open source community, there’s no good answer here, which is why this article ended up being more of a “notes” vs “analysis” article. So I’ll make a few predictions:

1. As the barrier to entry to software decreases, open source will continue to become more prevalent, but it will be harder to charge for. Queue the need for innovative business models.

2. The strategic value of open source companies will continue to outpace the discount of future cash flows.

   1. Many open source companies are worth far more strategically than the value they’ll create from cash flows, because they’re such effective complements to a platform. See Databricks’ $1B+ acquisition of Tabular, for example.

3. We will see an increased rate of open source acquisitions for two reasons:

   1. First, the strategic value of open source I talked about above

   2. Second, the bottleneck for so many top tech companies today is AI talent. And there are many incredible technologists building in open source.

And to end on a positive analogy, I write this newsletter because (1) I love investing and (2) I love the process of creating something useful.

There’s a reason I don’t charge for the newsletter. And it’s the exact same reasons why developers open source a lot of their code, they just like doing it.

This was the ideological roots of free software: plant a seed today, so there's shade for people tomorrow. And it's worked incredibly well. Technology today is built on open source.

And the reason I love business models is that if you can figure out how to align incentives, you can figure out how to make the changes to make things better.

Incentives run the world, and I'd like nothing more than more great open source businesses to be built.

If you’re building one of them and agree/disagree with what I said, feel free to reach out.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

This rise has been well-publicized, but it feels to me like people leave the argument at “software is dead.” I want to lay out the second piece of the argument: “software is dead, long live software.”

So I’m writing this post to answer the question: What business models benefit from this trend, and how can companies build business models that take advantage of AI-generated software?

I’ll start with historical precedent.

Free Software

January 1st, 1970: IBM created the modern software industry by unbundling hardware, software, and services. For the first time, a major computing company sold software as a standalone product, creating the model the industry would follow for the next 50 years.

20 years prior, GE “sold” early versions of commercial software. More precisely, they bundled them and gave them away for “free” in order to sell their CNC machines.

It was a means of differentiation, a sales pitch. “Buy our hardware, and you get our software and services for free.”

Software was free, and it was a differentiator.

With the rise of AI coding tools like Cursor, Claude Code, Lovable, we’re entering a world where software business models look more like those of the 1950s then the 2010s.

The TLDR of my argument today is this:

1. Software is getting cheaper and easier to build, and that will only continue.

2. Because of that, technical differentiation will be harder to create and maintain.

3. That doesn’t mean software isn’t valuable; it’s highly valuable, but customers will be less willing to pay for software alone.

4. So, software development looks increasingly commoditized, meaning value will need to be created elsewhere.

5. Because of this, “free software” will be more common, and that will necessitate sustainable business models built around free software.

Those business models include:

1. Hardware: Use software to sell hardware (or vice versa)

2. Vertical Integration: Offer vertically integrated hardware and software

3. Services: Charge for the work itself (accounting, legal) or offer services to integrate software into complex, custom deployments

4. Platforms: Customers will pay for the convenience of platforms, not the functionality of point solutions

5. Payments: Give away software, charge for interchange fees

6. Software as Advertising: Software essentially becomes “interactive content.” (∞) bonus points if network effects are involved

7. Infrastructure / Compute: The platforms that enable “free software” will collect their tax on each piece of it

Disclaimers: Not all of these examples are “free software”, but all are examples of differentiated business models. Software switching costs are incredible; in an alternate universe, this article would be called “searching for switching costs.” Once software’s in place, it tends to be very hard to displace. If there’s an opportunity to rapidly scale software in a new market (like new markets created by AI), it’s wise to cement switching costs as fast as possible. Also, there are many moats to be created outside of these business models. Fault-tolerance matters. Vertical software and regulatory requirements create moats. This is not an all-encompassing article, rather a trend that will play out over the next decade, and my thoughts on how to position yourself so it’s a tailwind rather than a headwind.

Let’s get to it.

1. Use Software to Differentiate Hardware (or Vice Versa)

The clearest beneficiary of free software today is hardware companies. In the past, it was tough to be great at both hardware and software. That barrier to entry is changing.

The core idea of commoditizing your complement is that as prices decrease in one part of the value chain, they increase in another.

If we take a simplistic view of the computing industry, the two primary components of the value chain are hardware and software. If software accrues less value, value flows to hardware.

To take advantage of this, companies can either give away software for free to differentiate their hardware (for example, Google open-sourced Kubernetes to make it easier to switch off of AWS). OR use hardware to differentiate software; if you can sell hardware upfront, you can collect the joys of switching costs.

It helps this argument that the most valuable “free software” in the world, managed by an individual company, is Nvidia’s CUDA.

Since the early 2000s, they’ve been developing the toolkits to “accelerate computing” via CUDA’s programming language, libraries, and drivers.

Today, it is the biggest differentiator for Nvidia’s hardware platform. See this quote from Semianalysis’ analysis on AMD’s MI300X vs Nvidia’s H100/H200:

In short, when comparing Nvidia’s GPUs to AMD’s MI300X, we found that the potential on paper advantage of the MI300X was not realized due to a lack within AMD public release software stack and the lack of testing from AMD.

AMD’s software experience is riddled with bugs rendering out of the box training with AMD is impossible. We were hopeful that AMD could emerge as a strong competitor to NVIDIA in training workloads, but, as of today, this is unfortunately not the case. The CUDA moat has yet to be crossed by AMD due to AMD’s weaker-than-expected software Quality Assurance (QA) culture and its challenging out of the box experience. As fast as AMD tries to fill in the CUDA moat, NVIDIA engineers are working overtime to deepen said moat with new features, libraries, and performance updates.

Nvidia has 1000s of employees dedicated to CUDA, for free software! And the ROI on those investments is in the tens (or hundreds) of billions of dollars.

Alternatively, you can use hardware to sell software. Take our friends at Vast Data for example (who’re reportedly raising a round valuing them at $30B (!)). They sell software, that runs on storage boxes in data centers, to manage the data for AI workloads.

They initially sold integrated hardware and software. But in 2021, they stopped building hardware, opting to let partners do that for them. They still have heavy involvement in the configuration and design of the hardware, but don’t manufacture it themselves.

Now, customers buy the hardware from their partners, powered by Vast’s software. The beauty of this is that Vast gets the upside of the data center boom, but their revenue is recurring from software, so they don’t get the downside when the hardware investment cycle slows.

Use hardware to sell software!

The takeaway: Hardware is hard, but difficulty creates differentiation. If software gets commoditized, more value will flow to hardware. And this article is about finding differentiation in software. If a company can use hardware to get performance gains, it’s a wonderful way to build moats.

For next-gen hardware companies (semiconductor companies, for example), the decreasing cost to build software creates a new way to build moats and distribute your hardware.

2. Offer Vertically-Integrated Solutions

The other way to benefit from hardware is to offer vertically integrated software and hardware products. A company sells one product, at one price, that integrates hardware and software to provide the best customer experience possible.

We’ve seen many of the largest companies of the last three decades leverage this strategy to provide MUCH better customer experiences than incumbents: Apple, Tesla, SpaceX, The Hyperscalers, Anduril.

The basic premise is this: by controlling every piece of the user experience, you can more easily create magical experiences. This is why Steve Jobs meticulously reviewed every single piece of Apple’s products. As software gets easier to build, it becomes easier to own every part of the value chain.

We’re also seeing new players leverage this strategy like Radiant Nuclear, Base Power, and many others.

Packy McCormick has written a wonderful series on vertical integrator startups (I hope I don’t too grossly overcite his articles, but do go read them if you haven’t). The startups are integrating sets of proven technical components, and then innovating at the systems level: The integration is the innovation.

I’d argue Anduril is the perfect example of the modern vertical integrator (and what a rise it’s been, they’re up to $30.5B since this chart was made):

Pre-Anduril, the majority of contracts won by defense contractors were cost-plus contracts. By the way, can someone tell me a worse way to align incentives?

Combined with the consolidation and oligopoly of the defense contractors, all incentives were aligned AGAINST innovation.

Until companies like Anduril (and Palantir as its predecessor, but we’ll discuss them in a bit) came along. To win government contracts, they not only had to beat incumbents, but beat incumbents’ solutions so badly that they overcame the decades of inertia and relationships those contractors had built up. And the best way to do that was vertical integration.

They compiled one of the best hardware AND software teams in the world, and started selling integrated solutions, from watchtowers, to drones, to submarines. A lot of hardware, enabled by software. And it resulted in a better, faster, cheaper products that built a $30B company.

If you ask me one area I think will generate the highest “slugging ratio” of $10B companies over the next two decades, it’s this theory applied to healthcare, robotics, manufacturing, and more!

3. Offer Services

For as valuable as software has been, its promise has always exceeded its impact.

The problem with selling software is that you can sell it, but you can’t guarantee the customer uses it. And you DEFINITELY can’t guarantee they use it optimally. This statement gets put on steroids with AI.

Services, meaning including human labor with the software, help to prevent this. There are two applications of services in this context:

1. Forward Deployed Engineers: Help the customer implement the software; the more complex the implementation, the more valuable the services.

2. Sell services, not software: If an AI tool actually provides performance improvements, why not build a business around that tool and outcompete competitors with better prices and a better experience? Instead of taking a software-sized bite out of the value chain, you can own the whole thing.

For the Forward-Deployed-Engineering model, the logic goes like this:

Software is only valuable if it can be adopted. The barriers to adopting software are often the technical know-how of the adopter. For complex environments, like:

1. In financial services, integrating with mainframes and COBOL code

2. In healthcare, deploying software to systems with hundreds of clinics, large hospitals, and thousands of patients, with life-or-death consequences if there are mistakes

3. In government, well…you can imagine what it’s like to implement software in the government

In these examples, it’s not an option, but a necessity to provide implementation services. For a simple framework, the complexity of the industry directly scales with the success of the FDE model. Growth is slower, but it's got more durability!

Palantir has (now) famously pioneered the FDE model, which explains their success in selling government software. It also explains why they’re so successful in healthcare deployments. For projects, their teams go on site, leverage the Palantir platform, and co-build solutions to their customers’ specific problems.

On to selling services themselves. For this, I’d rather explain through a thought experiment: Why haven’t there been any large companies built selling AI to hedge funds?

The software is immensely valuable, very smart people have worked on these problems, and could raise money to sell the software. But the answer is if the algorithms actually provide alpha, it’s far more profitable to create a hedge fund, and leverage the software to generate alpha.

If you try to sell the software, the logic breaks down in two ways:

• If the software doesn’t work, no one will buy it.

• If the software does work, and you sell to many customers, the competitive advantage will be eaten away, making it less valuable and easier to copy.

If we now apply that logic to the legal world, a common AI use case:

• You can sell AI software to legal firms, which are notoriously difficult buyers, in large part because of the misaligned incentives of hourly billing.

• Or if the AI actually works, you can vertically integrate as a legal firm, get the 50% productivity improvement of AI, and outcompete the law firms who wouldn’t adopt the AI software. Over time, as AI automates more of the work, the firm collects more of the value, looking more and more like a software company.

So the basic idea of selling services, is that it’s far more valuable to sell “the work” than the software. Instead of taking a high-margin software-sized bite out of the value chain, you take a low-margin approach, but gobble up the whole thing.

Similar to how Steve Ballmer described Microsoft’s shift to the cloud,”You should expect us to have lower gross margins going forward, but we’ll make it up in volume.”

The two attempts to do this are in starting an AI-native startup or rolling up (acquiring) existing businesses and using their customers as a distribution network to deliver the value of AI. Both are logical strategies for creating a differentiated business model. (Whether or not this makes sense as a VC investment is worth its own article.)

Three recent examples of the rollup strategy, led by Thrive Holdings, are Crete (Accounting), Shield Technology Partners (IT Consulting), and Savvy Wealth (Wealth Management).

4. Process Payments

As the world stands today, companies still need to pay interchange fees on payments, regardless of who they choose. If companies can integrate payments into their platform, it provides another way to profit off of software without having to charge for it.

I was recently arguing with a friend on what Ramp’s valuation should be. His argument was that “it’s not real software revenue.” If we take a step back, the benefit of enterprise software is that it’s:

1. Recurring

2. Predictable

3. Hard to switch off of

In Ramp’s example, they give software away for free, then collect a % of expenses processed by Ramp’s expense cards, through Ramp’s travel platform, or through Ramp’s procurement tools. If a customer is using all three of those offerings, Ramp’s business is:

1. Recurring (and aligned with the success of the business)

2. Predictable (although less so than traditional SaaS contracts)

3. Hard to switch off of (same software switching costs as traditional SaaS)

Ramp’s not alone. I talked about this in my Toast Deep Dive: getting restaurants to spend a lot of money on software is really hard. But getting them to pay for a payment processor is easier, they’re paying interchange fees either way!

In a world where enterprise software is more and more competitive, people will pay less for it. But if you give away the software for free, and can offer payments through the platform, it provides a different value prop for the customer:

You already have to pay interchange fees, you might as well pay them to us, and we’ll give you all of this software for free!

5. Sell a Platform

When I was working at Microsoft, one competitive advantage stood out over all others: the strength of their platform, and the ensuing ability for them to upsell through their distribution network.

Microsoft seemingly had a relationship with every single enterprise in the world. When they released a new product, they could give it away for nearly free, upsell through their distribution network, embed switching costs, and then raise prices.

Even better, a company’s already got data stored with Microsoft, oftentimes very high-value data, and so they trust Microsoft. Plus IT doesn’t need to manage another vendor, just add it on to the contract!

They beat point solutions over and over (and over) again with this strategy.

As industries consolidate, the variable for competition changes: from functionality, to performance, to convenience, to price (commodity):

My argument for building platforms is that we’re in the convenience phase of software. There’s not enough differentiation for customers to pay for point solutions, but they will pay for the convenience of not having to manage many different pieces of software.

Take the success of Rippling (now valued at $16.8B) as an example. Look at their launch announcement:

Today, we’re launching a best-in-class, all-in-one HRIS, payroll, and benefits system. But Rippling itself isn’t just an HR system. It’s a level deeper in the stack—a system for employee information that connects to all your key business systems and automates manual processes across the entire company, not just HR.

If there were ever a time to go to market with a platform, it’s now.

But even better, if there ever a time TO BUILD A PLATFORM, it’s now.

With the decreasing barrier to entry of software, it’s now possible to build a platform MUCH faster than ever before.

6. Advertising

The shortest and simplest of these micro-essays is centered on one fact: The two best business models of all time were built on “free software.”

They did so by aggregating attention, via network-effect-driven marketplaces. Obviously, if you can build network effects into your business, it’s a wonderful thing. But that’s very hard to do.

It also points to another fact: if you can own attention with software, you can monetize it.

In an age where attention is scarcer than ever, you can monetize it even more.

Consider this a prediction that we’ll see more and more “software as content”.

Brands will take unique data they have from their business, build out applications, and use it as another way to garner attention for their business.

For a mini example, look at Ramp’s AI Index. Ramp then gets shared thousands of times across articles for leveraging their data, building out lightweight software to visualize it, and telling the world about it.

7. Infrastructure, GTM, Inertia, and All the Reasons Theories are Brittle

Finally, what about the infrastructure that enables this shift?

I’d argue that open source infrastructure has been pioneering these business models from the beginning. Starting with Red Hat in 1993, offering services and compute built around Linux.

So I’ll create one more category: the infrastructure that enables “free software.” From cloud services, to hosting, to AI coding, to databases. The companies enabling this software to be built will collect a tax on each piece of it.

On top of that, distribution/marketing/sales, deep data integrations & switching costs, and any moats built through vertical-specific regulations are more important than ever.

The barrier to software development has been dropped to near zero. And with it, decreased the barrier of turning creativity into creation.

I remember the first time I used GitHub Copliot in 2022, then when I used Cursor in 2023, and then Lovable at the end of 2024.

Each time, I remembered feeling the same thing: that the barrier for creation is lower than every before. Anyone can build software, and anyone can build businesses.

It’s the best time in history for those who do want to build businesses. Hopefully, these playbooks help them do it.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

The hyperscalers provide the best insight into that. When I was working at Microsoft, I was constantly reminded just how broad its reach was. I mean, every company in the world seemingly uses Microsoft (that’s actually not true, its something like one million enterprises, but the point stands).

The tech world is built on the hyperscalers (which in turn, are built on the data center and semiconductor ecosystem they source from).

These are the most important businesses in technology. They're the largest demand generator for the semiconductor ecosystem. They're the largest vendor to software companies. They're the utilities of the computing age.

Now, they’re driving the majority of the spending on AI CapEx and supplying much of the compute to the AI labs and startups building the next supercycle of tech.

It’s why the one dedicated quarterly article I write on Generative Value is an update on their businesses:

1. Revenue & Performance

2. Market Share

3. Five Observations on the State of the Cloud

So while I spend 90+% of my time thinking about what the largest tech businesses will be in a decade, the hyperscalers justify a dedicated article.

On to the update.

I. Market Share & Statistics

For those that prefer cold hard numbers over commentary, feel free to look at these graphs and move on with your day! A reminder that these are my personal estimates based on 10Qs, but some assumptions have to be made around Azure’s historical numbers and Google Cloud, including GSuite in their numbers.

AWS generated revenue of $30.9B, up 17.5% Y/Y. Azure generated roughly $21B in revenue, up 39% Y/Y. Google Cloud generated $13.6B in revenue, up 32% Y/Y.

Combined, the three cloud giants hit a $262B revenue run rate, growing 27% Y/Y.

Oracle is also rapidly ramping up CapEx, up to a projected $27B in 2025 according to Semianalysis. They also project nearly $40B in GPU cloud revenue run rate by 2027, which would give them meaningful market share. Perhaps I’ll include them in next quarter’s update.

Based on those revenue numbers, that translates to the following market share numbers:

A more interesting view of market share is by net new revenue generated over the last four quarters. Of all the revenue generated in the last year by the hyperscalers, what percentage of share did each take?

Consider this a gauge of “market momentum”, or where market share is headed.

So on to my observations from the quarter, starting with the one most on everyone’s mind: how in the world is Azure growing 39% Y/Y at a $75B run rate?

II. Five Observations on the Quarter

1. AI is driving Azure’s outperformance.

Let’s revisit the chart above. In three years, Azure has moved from generating ~30% of annual net new cloud revenue to ~43%. They’ve accelerated growth from 27% two years ago to 39% this quarter.

To start, they will write case studies on Satya’s execution with AI. It’s led to the biggest differentiator in their cloud business today. First, because of their partnership with OpenAI, they can offer the OpenAI API directly through Azure. To my knowledge, OpenAI is the only other API provider of OpenAI’s models.

Second, they’re the primary cloud provider to OpenAI, which could grow into the world’s largest cloud customer in the coming years. Note that Azure lost exclusive provider status after not being able to deliver enough capacity, but they’re likely still the largest OpenAI cloud provider by far.

A large % of Azure’s growth has been driven by these AI workloads over the last few quarters:

If you include Microsoft’s share of OpenAI today (reportedly around ~33%), Microsoft’s investments in OpenAI will result in hundreds of billions in value creation.

When you add on the power of Microsoft’s distribution network (their relationship with nearly every enterprise in the Western World), it’s an incredibly powerful combination that’s challenging to compete with.

So yes, Azure’s taking share, but all three providers continue to do very well, especially at their size. Growing these businesses is priority number one for the hyperscalers; a large part of that is growing the AI business, on to point number two.

2. AI applications continue to grow ~rapidly~ and exceed (their already lofty) expectations.

The hyperscalers are in the unenviable position of trying to predict how much supply the AI ecosystem will need over the next decade, and matching their investments to meet that demand. When in doubt, they’ll err on the side of overinvesting to ensure they don’t miss the current land grab moment.

So they’re closely watching demand signals to see if they need to decrease, maintain, or increase investments. And those demand signals are a screaming buy right now.

Take a look at some of these recent (reported) stats from AI model and application companies:

• OpenAI roughly doubled its revenue in the first seven months of the year, reaching $12 billion in annualized revenue. They expect that number to hit $20B by the end of the year.

• Anthropic hit a roughly $4B run rate at the beginning of July…by the end of July, that number had grown to $5B. They expect that number to grow to $9B by the end of the year.

• Cursor hit a $500M revenue run rate, in roughly two years of selling.

• Lovable grew from $1M-$100M in 8 months.

• ServiceNow expects to hit $1B in AI revenue in 2026.

We can see some of those numbers visualized from The Information here:

For all the questions around AI today, there’s no question about the growth rates of these companies!

3. Those demand signals are driving more and more CapEx.

So the hyperscalers are basing their investment decisions off those demand signals, which are telling them to keep spending. Because there's a land grab to be the cloud provider to support these AI applications, the hyperscalers will invest whatever they need to win those early AI application workloads.

If we look at the early history of AWS, much of their early dominance came from the fact that they built a near monopoly on digital native & SaaS business building on the cloud. While Microsoft focused on the enterprise, AWS built mindshare in this “small market.”

But as I keep bringing up: an early monopoly on a rapidly growing market is a dangerous recipe!

The hyperscalers want to do everything they can to support the fastest-growing AI application startups. This is reflected in the continued "foot on the pedal" capital expenditures:

Despite these investments, the hyperscalers remain capacity-constrained, as noted by all three providers. Interestingly, they also called out the changing “shape” of CapEx investments.

4. These CapEx investments are changing shape, shifting towards inference and more scalable workloads.

Both Meta and Microsoft made comments along the lines of "We also expect a greater mix of our CapEx to be in shorter-lived assets in 2025 and '26 than it has been in prior years."

From my understanding, this means they're spending less on physical data center buildouts and more on shorter-term assets like GPUs, networking, and storage.

This tells us two things: (1) they've made progress in physical buildouts, (2) they’re derisking their CapEx investments as it becomes increasingly risky to invest in 15-year assets at the latter part of an investment cycle, and (3) they want their CapEx to be catered towards inference and as scalable as possible. I called this out in last quarter's update, and we're seeing it come to fruition:

Inference, as the enabler of applications, resembles the recurring SaaS revenue model of the last decade. The hyperscalers will do everything they can to power those applications.

Training, on the other hand, requires larger, upfront investments. As Zuck described, “I’d guess that in the next couple of years, the training runs are going to be on gigawatt clusters, and I just think that there will be consolidation.” These large-scale expenses present a less attractive risk-reward profile.

The last two years have come with a lack of clarity on the future of AI. In the absence of clarity, the biggest risk was underinvesting.

However, as the economics of AI come into focus and these infrastructure investments unlock AI applications, the hyperscalers can focus on investing in a future that provides them the best ROI per dollar deployed with the least risk of overinvestment. Risk-adjusted AI, if you will!

If it’s not clear, I’m overwhelmingly positive on the short-term and long-term economic characteristics of these businesses. However, I will share the biggest question for their businesses (in my opinion) over the next decade.

5. How different does the value chain in an AI world look from the value chain in a SaaS world?

The biggest question for the hyperscalers is not that they lose market share, but that the value chain shifts at the expense of the hyperscalers, in favor of the foundation model and app companies.

If we look at the rise of the SaaS wave in the 2010s, the largest early driver of AWS's success was digital natives like enterprise SaaS companies and consumer tech companies.

These SaaS products were essentially a set of "database and compute" wrappers. As the primary provider of database and compute, AWS became the infrastructure that captured the majority of value.

Today, though, the primary infrastructure for AI applications are the models themselves, i.e. the largest applications being “LLM Wrappers.” The AI labs are assuming the position of both infrastructure and application providers, sitting on top of the hyperscalers.

So while OpenAI and Anthropic are a fraction of the hyperscalers' size today, they could grow to match their significance over the next decade. As I said at the beginning, the hyperscalers give us the best insight into both supply and demand of AI. And those dynamics between the hyperscalers and its AI suppliers (Nvidia) and its AI demand generators (OpenAI, Anthropic, etc) will be fascinating to watch.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

Translated: we have to believe in AI’s ability to solve real problems and build companies around solving those problems.

So I’ll be doing a series of deep dives on AI applications starting with one of the biggest problems our country faces: the state of US healthcare. It’s becoming prohibitively expensive, weighed down by complexity, bureaucracy, and the ensuing administrative burden on providers (and in turn, patients).

It’s a problem that touches all of us, and it’s a problem technology is so well suited to solve.

This article is my best attempt to explain (1) why I’m so excited about AI in healthcare, (2) the state of healthcare software, (3) the state of healthcare & AI investments, (4) the areas where AI can help most, and (5) my thoughts on what it will take to make the impact I’m so excited about.

By no means is it perfect, but I do hope it starts a conversation. If you’re interested (as a builder or buyer of these tools) to be part of that conversation, feel free to shoot me a message.

Let’s get to it.

I. The US healthcare system is ailing.

In Part I, we covered why the US healthcare system is so expensive. Over the last 50 years, incentives have led to a system of powerful middlemen, each taking their cut from the $5T spent on healthcare in the US annually. The complex relationships between the players involved (patients, providers, insurance companies, PBMs, pharma, and the US gov’t) lead to intense scrutiny over everyone getting their “fair share.” Translated, lots of paperwork, lots of bureaucracy, lots of administrative waste:

To quantify it, over $1T of admin spend annually…and while we’re much less healthy overall. Full blog from myself and the team at Felicis here:

But when we dig into the details of those administrative tasks, they are text-heavy, voice-heavy, and often straightforward tasks being brute-forced by teams of administrative workers.

They’re calling, emailing, and writing letters to patients, PBMs, insurance companies; many of them spend hours a week faxing documents. We’re talking about AGI, and our largest industry is doing business via fax!

As succinctly as possible, here’s why I’m so excited about the opportunity for healthcare in AI:

1. This is a critically important problem for our country to solve. No one is satisfied with the system we have today (not patients, not providers, and not payers).

2. LLMs are perfectly suited to automate MANY of these administrative tasks in healthcare. Great companies resemble the technology waves they ride in on.

3. If a company can break in, the moats in healthcare are incredible. Of Epic’s over 600 customers, they’ve never lost a single one.

4. Finally, any meaningful improvement in efficiency creates a massive business opportunity. When you’re tackling a (literal) trillion-dollar problem (seriously, it’s literally a trillion dollars of spend, this is not VC talk for the ‘next trillion dollar company’), any meaningful % improvement can create a very large opportunity.

This article focuses primarily on the administrative applications of AI in healthcare, there are many others that are outside the scope of this article today. To be clear: this will not be easy. AI will not “solve” everything. There are two paths to fix this problem: reform the system or play within the bounds of the system today. As a technologist and investor, I choose the latter.

II. The State of Healthcare Software Today

Before we can even START to talk about AI’s impact, we have to understand two topics:

1. The healthcare software landscape today

2. Why its impact has been limited

The healthcare software landscape today

Given our focus on AI in healthcare today, I don’t want to bog you down with an exhaustive list of companies. I’ll point out the largest healthcare software companies:

The biggest categories are the healthcare systems of record:

1. EHRs (electronic health records): storage of patient medical records

2. RCMs (revenue cycle managment): manages the billing process between payers, providers, and patients

3. PMSs (practice management software): manages scheduling, billing, AND EHRs for smaller clinics

But you may notice: wait, these companies are kind of small relative to healthcare, a $5T industry? Yes! Epic, the healthcare giant, only has $5B in revenue?

It points to an unfortunate fact that those in healthcare know well: healthcare software’s impact has been relatively limited in both dollars and automation.

Why has healthcare software’s impact been limited?

I want to lay out very clearly why its impact has been limited AND why it changes now:

1. Data in healthcare is nuanced, and can’t be forced into an objective system

Remember how I was talking about all parties wanting their “fair share” and the ensuing documentation requirements because of that?

Well, the big problem with that is it forces nuanced, unstructured healthcare data into a structured format. Every patient interaction has to be translated to one of thousands of medical codes, turned into a bill, and justified to insurance on why the final diagnosis/prescription was made.

Ask any physician, and they’ll tell you: it’s not so simple.

All this did was force a burden on clinicians to take copious notes, put them into the EHR, and then be forced to justify their notes/prescriptions for any questions/issues from insurance (and there are a lot of them).

This work couldn’t be effectively documented and certainly not automated before AI. LLMs change that.

LLMs are wonderful for understanding nuanced, unstructured data, and they create the ability to start to automate these workflows.

2. Healthcare providers have (rightfully) been slow adopters of technology

Providers have been slow to adopt software for three more reasons:

1. Low profits = low budgets

2. Low fault-tolerance (i.e., mistakes are costly) = high barrier to adoption

3. Integration complexity = higher costs, lower adoption.

Today, low budgets are actually a tailwind. Providers told me they have to cut costs to survive, and AI is a really good way to cut costs. Companies are taking spend from labor budgets, not just software budgets, to pay for AI tools. This expands the software market a lot.

Second, fault tolerance is a real concern, but is much higher in administrative workflows where mistakes are already made all the time.

Finally, integration complexity is still an issue. Because healthcare systems are complex, they have complex technology stacks, which further cements switching costs, making adopting new technology more expensive.

Healthcare AI companies and startups must solve this problem themselves. They must take a page out of Palantir’s playbook with the Forward-Deployed-Engineer model where their team goes in to healthcare systems and builds out the integrations required to successfully adopt this software.

In summary, LLMs ability to actually automate workflows, deliver ROI in the form of cost savings, and a new model for delivering software create a new age of healthcare software opportunities.

As one source told me,”Providers were basically impossible to sell into two years ago, and that’s completely changed.”

III. Where AI is Helping Healthcare’s Administrative Burden

I’m not kidding when I say there’s another task that can be automated around every corner in healthcare. From my research, I came across five key categories of workflows that take up the most time and effort from healthcare providers. See the startups taking on those categories (with a few others) below:

To illustrate AI’s impact today, we can essentially walk through every single step of the healthcare process:

Scribing: Starting with the patient visit, AI scribes take notes for clinicians so that they can focus on engaging with the patient. Abridge is now valued at $5B, and two other unicorns have been created in the scribing market. When I talk about AI being adopted way faster in healthcare than any previous software, this is what I’m talking about.

Some estimates have the % of hospitals using or testing scribes at over 80%:

Medical Search: Once the doctor’s seen the patient, they need to do research to come up with a diagnosis. Previously, this was using antiquated medical search tools or Google. Now, many are using AI search tools, led by OpenEvidence, who just raised at a $3.5B valuation.

Over 1/3rd of US physicians are using OpenEvidence 12 months after launch:

Prior Authorizations: After a diagnosis has been made, many medications or procedures require prior authorizations. Essentially, if insurance doesn’t explicitly cover a prescription or procedure, the provider needs to get pre-approved for insurance. This can be done through prior auth portals, phone calls, and god forbid fax machines.

There are entire PA teams in hospitals just doing prior auths all day long, mistakes are frequently made, and doctors get pulled in to help fix them. It’s the quintessential highly painful + very common workflow that AI is automating.

Revenue Cycle Operations: Once a procedure is complete, that clinical note needs to be translated to medical codes (of which, there are about 70k). That code is translated to a bill/claim, and sent to insurance, who approves or denies it.

Insurance companies are automating claims processing, or reviewing of the claim. If it’s denied, the provider appeals it, which is likely overturned. (And many providers are using AI to write appeals letters now).

After all that, a bill is sent to the patient.

Customer Service: Finally, if the patient wants to ask questions, schedule a follow-up, or pay a bill, they’ve got to call the provider. The majority of basic questions can now be automated with AI. Essentially, immediately cutting spend and improving patient satisfaction with AI.

All five of these categories of administrative spend are centered around voice and text, that AI can now automate, and it’s why we’ve seen such rapid adoption of these tools.

All in all, we’ve already seen large, well-funded companies emerge in all of these categories:

To be clear, there are MANY other companies in each of these categories, this is just illustrative of the excitement from both investors and customers for these solutions.

IV. The Path Ahead: How to Maximize Impact on the Healthcare System

So, early demand is there, the impact in several categories is already being seen. How do we ensure this is a sustainable trend in both (1) building enduring companies and (2) ensure those companies maximize their impact?

I see three paths to do so:

1. Early point solutions need to expand as platforms.

In my Toast Deep Dive, I laid out the path to go from owning a workflow -> being a system of record. When I wrote that, I had Abridge in the back of my mind. Once you own the clinical note, you can now automate the key workflows around that clinical note, build a platform, and expand to become a system of record.

In Abridge’s case (and other AI scribes), you can take the clinical note, automatically translate that to a medical code and claim, send that to insurance, and expand to own the workflows around revenue cycle management.

For the investors at the $5B valuation, this is likely the vision they see.

2. Companies need to scale GTM as quickly as possible

To maximize their impact, companies to maximize their reach. Time for capitalism to shine!

I’ve exhaustively covered the Power Law on this publication. If there’s one applicable lesson from it, you have to start your power law compounding as quickly as possible.

This means whoever moves fastest often wins the mindshare of customers, the mindshare of investors, and creates a flywheel of increasing returns. The goal is to get that flywheel turning as quickly as possible, either through:

• Early traction (either sheer volume or top-tier customers)

• Fundraising (again through amount or top-tier customers)

• Marketing (owning mindshare)

All of these variables then compound on each other.

3. Leverage this newly automated data to improve more than just efficiency, but make a real impact on patient outcomes

Many clinicians I talked to about this problem were completely uninterested in the admin side of workflows. They were obsessed with creating better outcomes for patients.

The fact that OpenEvidence already supports 1/3rd of US physicians less than 12 months after launching is incredible.

But as we aggregate more and more data, the ability to positive affect patient outcomes improves as well.

In a perfect world, we’ll have the ability to aggregate much of the (now automated) data from AI scribes into one database so that every single symptom -> diagnosis is tracked in the US. As the dataset grows and models get better, we’ll get faster and faster diagnoses.

Hopefully, we’ll leave a world behind someone has to bounce around between ten specialists before getting a diagnosis for a rare disease.

One final thought on the exciting thing about boring things.

Prior auths, medical coding, scribing…these processes themselves are boring. But the impact of automating these tasks is not.

As we think about the economic impact of AI, much of it will come down to tasks like scribing and prior authorizations.

Nobody wants to do them, it’s routine work, and so we’ll hand it off to AI. It makes the economy more productive, very few will notice; and in aggregate, it will make us meaningfully more productive.

In the process, it will help fix one of our most important and most broken systems. And one that MUST be solved for the US to thrive moving forward. And that’s something to get very excited about.

As always, thanks for reading!

Disclaimer #2: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

It feels like an obvious statement, but we can’t help ourselves from falling into pattern matching. Looking for the next Google/Meta/Amazon, or the Uber for X, the Airbnb for Y, the AI Agent for *literally any industry*.

Well, to not get swept up in the new trend every week, we’ve got to look back. Something akin to Churchill telling us,“The farther back you can look, the farther forward you are likely to see.”

So I wanted to do an analysis of the largest companies founded in recent history. To not get caught up in narrative, but in what the data tell us, taking an outside view in the words of Mauboussin!

Causal thinking is a form of storytelling that comes naturally. It is a compelling way to anticipate the future and a convincing way to explain the past. Our minds are great at creating facile narratives to explain what happens in the world around us.

The second method is to think statistically, commonly referred to as the outside view. Rather than weaving a story based on causal links, the statistical approach examines what happened to an appropriate reference class of cases in the past. The results of the reference class are called base rates.

So this will be a deep dive on:

1. Data on value accretion in technology over the last 30 years

2. Lessons from that value accretion

3. What those lessons might tell us about investing in technology today

The TLDR:

1. The next $100B company will not look like the last

2. Know the game you’re playing: home runs, grand slams, or something more like Space Jam for Baseball (aiming for outer space)

3. Software is like chicken, 80% of it tastes the same

4. “Market size” may be the single greatest reason for great investors missing great companies

5. Companies resemble the technology waves they ride in on

6. Just because it would be irresponsible of me not to mention this, don’t underestimate the Power Law, ever.

Also, we published our Felicis Call for Startups last week. Do go check it out for the areas we’re excited to invest in.

On Methodology: The vast majority of technology value has accrued to the largest companies, so I pulled all Pitchbook, IT-labeled companies worth $5B+, founded since 1995. (Note: this does exclude Amazon, Nvidia, Microsoft, and Apple.) I had Claude help categorize the companies, so do consider the exact data quite directionally accurate, and precisely inaccurate.

Let’s get to it.

Data on 30 Years of Technology Returns

The dataset includes $13T worth of value creation, 300+ companies, across 65 (!) categories. See the highlights of the most successful of those companies here:

Now I won’t belabor the Power Law anymore than I have to, but the top seven companies accounted for nearly 50% of this dataset.

Which leads to the first and most important takeaway:

1. The next $100B company will not look like the last.

First and foremost, the value of technology is driven by anomalous companies, founded by anomalous people. The very fact that they’re anomalous means pattern matching is more likely to make us miss the great companies than to catch them.

It’s hard to visualize a company’s path if it’s never happened before. How can you estimate the size of Google’s market in 1998? Or Meta’s in 2004? It’s impossible.

Take the current most anomalous AI company: OpenAI. It was founded as a non-profit research lab, without a clear technical vision, lost its cofounder, with a convoluted governance structure, and it’s on a path to be one of the most important companies in history. As anomalous as it gets.

There is no “public comp” for the most successful companies. They’re unique. The largest companies create categories, and that’s what makes them so hard to see.

Neil Mehta put it as looking for the “the very small number of the world’s founders who are going to produce a significant proportion of the value that humans enjoy.”

To start to get into the data, see the largest companies founded since ‘95:

The majority of these companies either created their industry or hugely expanded it in a way that they basically created their industry (Tesla for example).

Looking at the data by category leads us to the next takeaway:

2. Know the game you’re playing: home runs, grand slams, or something more like Space Jam for Baseball (aiming for outer space).

If we go back to Mauboussin’s idea of base rates, I’d argue we need a different mental model for investing in these various categories.

The majority of value has been created by consumer companies (dominated by the Power Law). However, nearly twice as many enterprise software companies were founded.

To visualize that, I added a column for ‘Slugging Ratio’, Value/Company, to get a sense of how Power Lawed different industries were:

Over the last 30 years, consumer companies have tended to be network-driven marketplaces with genuine winner-take-all dynamics. If you happen to invest in one of those giants, the only mistake tends to be underestimating how big they will become. Cue Yuri Milner’s investment in Facebook at $10B.

If a company can genuinely architect network effects into their business model, the upside is immediately increased.

Hardtech companies, any company building hardware, have the second highest ‘slugging ratio’, mostly because it’s harder to survive as a hardtech company. Typically, it requires more capital, takes longer to get to scale, is harder to build products, is potentially more exposed to fundraising challenges, and is harder to disrupt incumbents.

However, if they can get past that speed run, the market opportunities tend to be massive.

But there’s only room for a handful of consumer and HardTech companies. Because of that, enterprise software has become a perfect investment vehicle for an expanding venture capital landscape.

Businesses that scale rapidly, in non-winner-take-all markets, with strong moats, and low operating costs. In an environment with many venture funds, there are more winners to go after, more proven markets, and all-in-all much less risk. But plenty of upside if it does go right. A good recipe to derisk an inherently very risky industry.

3. Software is like chicken, 80% of it tastes the same.

I stole this line from Vista Equity Partners founder Robert Smith who said, “Software companies taste like chicken…They’re selling different products, but 80% of what they do is pretty much the same.”

If we look at the majority of the largest enterprise software companies, they’re either:

1. An application built on databases with unique workflows

2. Infrastructure to build those applications

3. Security to protect those applications

This isn’t to say these companies aren’t differentiated, but that their differentiation is much more nuanced than it seems on the surface. Sales, marketing, and building mindshare are all just as, if not more important, than technical differentiation.

In a world where software is getting easier and easier to build, features are replicated in days, and AI coding tools are getting better, technical moats in software may be confined to unique data or integrations.

The point is that technical differentiation tends not to be the defining variable of enterprise software companies.

In this context, I find the “GPT Wrapper” argument fascinating, stating that AI application companies are just rebundling LLMs. The majority of enterprise software companies are SQL (or NoSQL) databases, with unique workflows built for a specific set of customers.

If we look at the largest AI enterprise application companies recently, they’re “LLM wrappers.” But this is the exact same thing we saw over the last decade with the largest enterprise software companies, and they became $100B+ companies!

As I mentioned before, enterprise software is less risky and more predictable than other categories. Outside of horizontal enterprise software though, market sizing seems to be less of a constraint than it seems on the surface. “How big can this company become?” and “How big is this market?” are two very different questions.

4. “Market size” may be the single greatest reason for investors missing great companies.

If there’s one thing humans don’t handle well, it’s uncertainty. And that’s exactly what new markets introduce.

Palantir, Shopify, Uber, and many others, all (more or less) created new markets that didn’t exist before.

Even trying to add certainty to an inherently uncertain problem can lead to foolish behavior.

Take Aswath Damodaran and Bill Gurley’s infamous back and forth on valuing Uber. Gurley’s conclusion: Uber’s potential market could be 25 times higher than Damadoran’s original estimate.

I took a look at companies founded since 2010, with the highest multiple returns in public markets, an underestimation metric if you will.

A few patterns emerge:

1. Investors underestimated market size, especially for market expanders or vertical markets: Shopify, Guidewire, Zillow, AppFolio were underestimated. Similarly, in private markets, investors underestimated Toast, ServiceTitan, and other vertical software companies.

2. Companies had multiple-expansion tailwinds as new business models surpassed old: Tesla (the most extreme example) and (to a lesser extent) all the software companies listed saw a reframing of multiples compared to the incumbents they were competing against. Today, Tesla alone is worth nearly $1 trillion, more than twice what all of the big automakers were worth, combined, when it entered the market.

3. Investors underestimated the power of platforms: ServiceNow, Palo Alto, Crowdstrike, Workday, Atlassian, and Datadog all expanded their markets by expanding their product lines. As software became easier to build, it reduced the technical differentiation between platforms, making customers prefer platforms over point solutions. In times of consolidation, it’s good to be a platform!

This isn’t to say market size doesn’t matter, it’s to say: it’s really easy to get it wrong.

5. Companies resemble the technology waves they ride in on

If the last section was a “market sizing” section, this is a “why now?” section. The proverbial “why now?” question in venture asks: Why hasn’t this company been created before? What new revelation now allows this company to exist?

Most of the time, it’s a new technology wave that unlocks the ability for the business to exist. Today, it’s AI. Before, it was the internet, then mobile, then the convergence of internet and mobile, then the cloud.

We can see below when $5B+ companies were created by sector:

The internet connected the world, allowing the aggregator business model to emerge.

Mobile took it a step further, putting the internet in everyone’s hand, unlocking consumer marketplaces.

Fintech was a rare example of regulation unlocking a new technology industry, and saw a boon post-2010s Durbin Amendment.

The cloud was the most disruptive wave in the history of technology, unlocking the ability to build software with a credit card instead of a data center.

Now, with AI, what companies are unlocked and what will they look like?

1. AI coding tools take the cloud a step further, unlocking the ability for anyone to create software, not just developers. This will lead to a similar explosion of software that the cloud introduced.

2. It unlocks the ability to automate voice and text-based workflows. We’re already seeing this across coding, customer service, and AI scribing. But this will expand to many, many more applications.

This expands the market for software in ways we haven’t seen before. Just one example: in this dataset, there’s not a single legal software company valued at $5B+. Harvey, founded just three years ago, is already valued at $5B.

A good thought experiment on what we’re seeing with AI, from Rex Woodbury:

I liked Alfred Lin’s parallel between mobile and cloud. Back during mobile, a worthwhile exercise was to break apart features of the iPhone, then predict which companies each feature could enable. His example: the GPS allowed couriers to drive around with Google Maps, delivering food. This led to DoorDash.

There’s a narrow window for new companies unlocked by waves of technology, and we’re seeing that window of companies emerge right now.

6. So what comes next?

I was reading Will and Ariel Durant’s History last week, and came across this quote: History smiles at all attempts to force its flow into theoretical patterns or logical grooves; it plays havoc with our generalizations, breaks all our rules; history is baroque.

Perhaps this article is foolish to even attempt to fit the most anomaly-based industry out there into logical grooves!

What doesn’t change is human nature. If we solve by inversion, humans struggle to picture exponential growth, they struggle with anomalies, and they struggle with uncertainty.

To handle that uncertainty, our best bet is to:

1. Understand the “base rates” of company categories (what’s likely to happen)

2. Know where differentiation comes from (in software, sometimes it’s mostly sales and marketing)

3. View market size as a first-principles exercise on the problems to be solved instead of a pattern-matching activity

4. Understand that each wave of companies will be unique, it will be hard to predict, and that’s a good thing

Steve Jobs said of computers,”I think one of the things that really separates us from the high primates is that we’re tool builders…What a computer is to me is it’s the most remarkable tool that we’ve ever come up with, and it’s the equivalent of a bicycle for our minds.”

Steve was right. Computers unlocked a wave of creativity never before seen.

And now, we’re seeing the greatest bicycle for the mind ever created. What a time to be alive.

As always, thanks for reading!

Disclaimer #2: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

If reports are true, Meta is hiring SSI CEO Daniel Gross, his partner (former GitHub CEO) Nat Friedman, and a stake of their venture firm. Meta also considered acquiring Ilya Sutskever’s SSI (last valued at $32B), Perplexity (last valued at $14B), and is offering 9-figure pay packages for AI researchers to join Meta’s superintelligence team (total spend could be up to $5B).

So, Meta was prepared to spend at least $30B, err let me rephrase, whatever it takes to be competitive at the leading edge of AI.

This is apparently in response to the underperformance of Meta’s model developments, from Stratechery:

The proximate cause for this reevaluation is the apparent five alarm fire that is happening at Meta: the company’s latest Llama 4 release was disappointing — and in at least one case, deceptive — pushing founder and CEO Mark Zuckerberg to go on a major spending spree for talent.

This will not be an analysis of whether or not these investments will work, but rather why they’re making these investments. Similar to my arguments on big tech’s CapEx investments, this is all logical. In summary, Meta has:

1. Personal motivations: Meta’s history of conflict with the platform they sit on top of (Apple) has led to a “do anything necessary” mentality to own the next big platform.

2. Profitable applications of AI: Meta has one of the most profitable applications of AI in the world (advertising systems) to justify ROI on these investments.

3. An unsatisfactory rate of progress: Meta’s not making the progress Zuck wants to in AI. Of the three largest paths towards model improvements (compute, data, and talent/algorithms), talent is the most differentiated today.

4. Incentives + no reasonable alternative: Meta has the applications to profit from AI (carrot) + a risk of relying on someone else for the next big platform shift (stick) + no reasonable investment alternative ($20B in profit last quarter!) = play the AI arms race.

If we squint, we can see the big vision for AI take form: a personal assistant at our fingertips, with context on our lives, that gets better as it becomes more powerful and more personalized. As it gets more integrated into our lives, it gets harder to replace.

The window to become that personal assistant is closing. This is the prize Meta is chasing, and the risk of underinvesting towards that vision is far greater than the risk of overinvesting.

This will be a rundown of Meta’s AI strategy, their motivations from past scars, the current state of their tech, and their vision for the future.

I. Some Historical Context

Three pieces of important historical context to understand Meta’s motivations:

1. Meta initially viewed themselves, and has always had the aspirations to be, a platform company.

This started with the Facebook platform in the early 2010s, enabling companies to build apps on top of Facebook’s ecosystem (you guys remember Farmville?). In the early days, the platform was just as important as advertising in the eyes of Zuck and the team.

However, as mobile took off, it became clear Facebook was a mobile application, but the mobile platform would be Apple’s. And that fact has caused Zuck as much stress as any other:

"There are all these things that, over time, Google—but especially Apple—have prevented us from doing, even though I think they would be good for consumers. At some point, we ran the numbers and realized that if we were able to do all the things we thought were beneficial—considering the additional engagement, increased app usage, and avoiding various fees—I think we might be twice as profitable as a company as we are now." - Zuck in a Stripe Sessions Interview (lightly edited for clarity)

2. To avoid that pain again, Meta made two opinionated bets on the platforms of the future in the early 2010s: AR/VR and AI.

For AR/VR, Meta acquired Oculus for $2B in 2014. As has been well-documented, they’ve spent well over $60 billion in pursuit of owning that platform.

For AI, they hired Yann LeCun in 2013, as well as other top AI researchers, and developed a near duopoly on top AI talent with Google. FAIR (now Meta AI) became one of the most respected AI research labs in the world.

3. Meta created one of the most profitable applications of AI of any company in the world.

Meta has profitably spent tens (perhaps hundreds) of billions of dollars on GPUs. Digital advertising is one of the most profitable businesses in history, and there are only three ways to increase revenue: increase users, increase screen time, or increase the likelihood that an ad converts to a purchase.

Meta’s recommendation systems are the best way to capitalize on the third variable, taking all the data they have and personalizing ads based on that data. Before spending GPUs on training Llama, this is what all those GPUs Meta purchased were for (and many of them still are):

In summary, Zuck’s personally and financially motivated to own the next platform, and he doesn’t care what price it takes to do so.

II. Meta’s Current AI Offerings

For a quick summary of Meta’s models (description from Devansh’s breakdown of Llama 4):

• Scout: an efficient multimodal LLM with a 10 Million Token Context Window that can run on one GPU.

• Maverick: a more powerful model (still reasonably efficient) that stomps out the “intelligent models”- GPT 4.5, 4o, Claude 3.7, DeepSeek R1, Gemini 2.0 Flash etc.

• Behemoth: A 2 Trillion model that is being called best in class, but is not released yet b/c it’s still mid-training. This is used as the teacher model for the lighter ones.

Much of their CapEx is going towards data centers and GPUs to fund model training:

These large investments are about ensuring they have control over a leading-edge model. They don’t need to profit from it because of their other profitable applications, ergo the open-source strategy. Open source provides a differentiated strategy and the open-source flywheel of community development and data to help them achieve it.

So, how are their models performing?

According to benchmarks, those models are struggling to keep up with the competition:

In LMArena, Llama 4 Maverick clocks in at 39th place on text models. According to Artificial Analysis, it’s fairly uncompetitive in terms of intelligence (logical, given it’s not a reasoning model). It does perform reasonably well on speed and price, however isn’t leading on any metric:

I want to be very clear: benchmarks can be misleading. As a friend recently said, “I never met a benchmark I didn’t like.”

But a litany of reports have suggested the same:

• From the Platformer, “The company lost last year to the Llama 4 dysfunction.”

• From Stratechery: Llama 4 was widely viewed as a disappointing model, and a big portion of the original Llama team has long since left Meta.

• From the Information: Meta delayed Llama 4, the newest generation of its AI model, because it wasn’t performing as well on technical benchmarks as the company had hoped.

• From the WSJ: Company engineers are struggling to significantly improve the capabilities of its “Behemoth” large-language model, leading to staff questions about whether improvements over prior versions are significant enough to justify public release, the people said.

This is not meant to be an indictment of Meta, I keep this newsletter optimistic! But based on reports and actions, it feels fair to conclude that Meta’s models are underperforming expectations. So why should they keep chasing the prize?

III. Meta’s Vision for AI

Zuck has described a few ~big~ opportunities for Meta in AI:

1. Creating “The Ultimate Business Results Machine”

Meta’s advertising engine has become so good that they know business’ customers far better than they do in most cases. This advertising black box is the clearest path to AI ROI for Meta.

In a Stratechery interview, Zuck described what this system would look like:

“Improve recommendations, make it so that any business that basically wants to achieve some business outcome can just come to us, not have to produce any content, not have to know anything about their customers. Can just say, “Here’s the business outcome that I want, here’s what I’m willing to pay, I’m going to connect you to my bank account, I will pay you for as many business outcomes as you can achieve.” (lightly edited for clarity)

2. AI-Generated Personalized Content

Facebook and Instagram have transformed from social media to media, as people mostly engage with content with people they’re not connected to:

Now, we may see the next evolution: personalized content generated by AI.

From Stratechery: “You can think about our products as there have been two major epochs so far. Well, the third epoch is I think that there’s going to be all this AI-generated content and you’re not going to lose the others, you’re still going to have all the creator content, you’re still going to have some of the friend content.” (lightly edited for clarity)

However, as referenced earlier, the big vision Zuck’s fighting for with these AI investments is a personalized and ever-present AI. Again, the ROI is increased engagement on the platform, leading to more advertising dollars.

3. The General Purpose AI Assistant

This is the biggest prize on the AI field right now. Viewed through this lens, OpenAI’s Jony Ive acquisition makes sense. The importance of memory becomes clear. The future is a world where you have your assistant with you at all times, and whoever owns this, has pole position on being the AI company. It’s what OpenAI, Meta, and Google are fighting for.

"There are two primary values we're aiming to deliver. On the AR and mixed reality side, our main goal is to create a strong sense of presence. The other major focus is personalized AI, which is where Llama, Meta AI, and related efforts come in. There's a lot of ongoing development to make these models increasingly intelligent, but I think things will become truly compelling when the AI is personalized specifically for you." - Zuck on Huge if True (lightly edited for clarity)

Meta’s got user data, the lead in AR/VR/AI glasses, and they’ve got a model that could be in the race for the leading edge. They’re well-positioned to compete for this prize, and it’s the one that’s got to be top of mind for Meta when making these investments.

IV. How do they achieve that vision (hint: people still matter)?

We’ve now come full circle. Meta and Zuck have the personal scars, the financial motivation, and the vision for their AI future. But how do they get there?

There are three primary vectors of improvement for AI: compute (GPUs/data centers/energy), data, and algorithms (as a proxy for human talent). The compute bottleneck has eased, and Meta’s over-acquired GPUs compared to rivals since 2023.

With the Scale investment and all the talent investments referenced above, they’re making a bet on data and on talent. But mostly on talent.

For all the talk of AI automating human labor, AI advancement is being driven by humans.

Steve Jobs told us to “Never forget the dynamic range of humans.” Well, Zuck isn’t. And it’s his best shot at achieving his vision and avoiding the same platform pains that have played out over the last decade.

As always, thanks for reading!

Disclaimer #2: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

“So if you really believe as we do that healthcare is about conversations—that it's one of the first original signals in healthcare—then you start to see that any number of different workflows are beyond it.”

Implicit in this statement is the idea that you can own a key workflow, then backwards integrate into becoming the system of record (the “single source of truth” system that stores the most important data for companies).

Contrast this approach to the systems of record today (Salesforce, ServiceNow, Workday, SAP, etc.), which started as databases and expanded into applications.

But companies resemble the technology wave they ride in on. The last wave of software (SaaS) was built on the structured database. This wave is built on the unstructured workflow.

Charlie gave us the advice,”Pick an extreme example and ask what the hell happened here?”

This brings us the extreme example at hand: Toast.

They took what may be the hardest customer set of any industry (small restaurants), owned the most important workflow (orders), expanded as a platform (became the system of record), and became a $25B company in the process.

Their journey lays a playbook for companies, especially in vertical industries, to build a new system of record.

See a summary of that journey here:

Also I'm working on a project documenting the stories of founders of companies like Toast, and compiling a booklet of their lessons. If you happen to have an intro to one of those folks, it'd be much appreciated. You guys are the best.

Onward.

The Story of Toast

Toast’s starts at Endeca, a database search engine founded by Steve Papa that Oracle acquired for over $1B in 2011. Three leaders (Steve Fredette, Aman Narang, and Johnathan Grimm) had spent years talking about starting a company together.

After the acquisition, they were, as Aman put it, “shamed into starting a company” because they’d talked about it for so long.

They get the idea that ordering food takes too long, so they’re going to build an app that allows you to order from your phone (scanning a QR code at restaurants). They go to raise money, and VCs do not want to invest. Restaurants are terrible customers, contract values are small, and the founders are arguing with each other in meetings. The only person willing to invest is Steve Papa, a $500k check at a ~$3M valuation. That stake would eventually be worth over $3B.

In 2012, they launched the app. Nobody really wants it, it solves a problem people don’t really care about. But what they do realize is that once they get the order, they can’t really send it anywhere because the Point of Sale systems (where you swipe your card and waiters put in your order at restaurants) are all on-prem and closed off from integrations.

So, for restaurants to be able to move into this new wave of mobile ordering, they were going to need new point-of-sale systems. Those systems happened to manage the most important data in the restaurant: the order.

1. Own the key workflow

So, in 2013, Toast launched their cloud-based Point of Sale system. They made two key decisions early on in their development:

1. Building on Android, not iOS, for maximum flexibility to accommodate the complexity of restaurants

2. Integrating payments into the point of sale, allowing them to create another revenue stream (and aligning their incentives with the restaurants). They essentially offer the same rate as existing payment processors and got little to no pushback.

They had immediate and violent product-market fit. As Arman described it, "We would go in and pitch our app, and people wouldn’t really care. But when we asked about their point of sale, they would talk for hours about how much they hated it."

By the end of 2015, they had 1000 customers. The complexity of restaurants (orders, inventory management, table management, payroll, high employee turnover, marketing) became their opportunity, and they immediately started building out features around the PoS.

2. Once you have the key workflow, expand to the most painful applications now unlocked by the unique data

After they had the PoS data, they could solve problems that were previously impossible to address: kitchen display systems and order routing, replacing paper tickets. They also offered features that solved the biggest problem all restaurants face (generating profits): digital gift cards, points-based loyalty programs, and online ordering.

There’s no replacement for good old-fashioned product velocity.

In Bessemer’s 2015 memo, they shared, “First off, the sheer amount of software the team has built in a short span is impressive – feature for feature they are already much more in the class of the >20 year old enterprise systems than the next gen “Bistro” players, and so for restaurants with any level of sophisticated feature requirements they win easily.”

Then, in 2015 and 2016, they showed the path they’d execute on for the next decade:

1. Build product functionality to move up market (multi-location management)

2. Expand the horizontal functionality to become the system of record for restaurants

In 2016, they released inventory management and food cost tracking, connecting front-of-house data with back-of-house data. At this point, they had solved the most important problems in the restaurant. The next step was to become a platform for others to build on.

3. Build an ecosystem for other companies to solve the problems you can’t

Due to the complexity of restaurants, there were still a lot of problems to be solved. And Toast couldn’t solve all of them, so they started to create an ecosystem.

They offered an API to allow developers to build with Toast’s data. This kept restaurants happy, allowed Toast to generate another source of revenue, and, most importantly, kept everyone in the Toast ecosystem.

“Basically several other startups we’ve seen in the restaurant space are simply modules on Toast. We think an app store model over time (a la Shopify) is possible and could contribute to Toast’s defensibility along with integration into several ecosystem players.” - BVP Memo

4. Expand to everything the customer needs

Their product velocity then continued until their 2021 $20B IPO. By then, they offered (among other features): payroll, labor management, handheld PoSs, marketing automation, and loans.

Even when COVID hit, and they thought they might lose it all, they released Toast Delivery Services, Toast Order & Pay, Toast Go 2. As the VCs say,“JuSt KeEp ShiPpInG.”

After their IPO, they’d offer products for: invoice processing (accounting), international products (UK, Ireland, Canada), hotels, retail (food & beverage), and catering.

Toast had become the Microsoft of restaurants, owning the most important workflows in the restaurant, starting with the one that mattered most: the order itself.

Which leads us to the final step: profit.

II. The Current State of Toast

Since IPO, Toast has done nothing but execute:

• Location growth: 31% CAGR

• Revenue growth: 36% CAGR

• Gross payment volume: 33% CAGR

• Improved operational efficiency: 20% gross profit CAGR

• Net income: $56M last quarter

Today, Toast processes payments for roughly 25% of the 520,000 small and medium restaurants in the US.

Their revenue comes from two primary sources: 80% comes from payments (taking a cut of every transaction) and loans, while 20% comes from software subscriptions (60% and 40% of gross profit, respectively).

See their metrics over time here:

With a quarter of the SMB market already captured, the easy wins are gone. The question now is how much growth remains, which breaks down into three areas: market expansion, market share gains, and revenue per location.

1. Market Expansion: How Much Runway Remains?

Market expansion will come from three areas:

1. Upmarket expansion

2. Expansion to other verticals

3. International expansion

Upmarket expansion is the clearest opportunity. Toast already offers most enterprise functionality; the gap is primarily in customer support and service levels. Recent wins with Applebee's and Topgolf demonstrate early traction, though competing against entrenched enterprise vendors like NCR and Oracle requires different sales motions and longer cycles.

Vertical expansion beyond restaurants is possible, but significantly more competitive. Toast's core advantage has always been deep restaurant expertise.

International expansion is the final growth vector. As of May 2024, Toast had 2000 live international locations, so they’ve shown early success. Especially in markets where there are similar workflows to the US (Canada, UK, and Ireland). Expansion beyond these markets requires adapting to different payment systems, regulations, and operational practices.

Management shared the goal of 10,000 locations across these three segments by end of year, and it will be the most important metric to watch moving forward. If they can execute, they still have only touched a fraction of the global market, see a directionally accurate market analysis here:

Based on these numbers, there’s still $13B+ worth of addressable ARR and $7B within the US. This doesn’t include the potential to expand to more geographies and industries.

2. Competition: Can They Keep Winning?

Across all segments, Toast is the fastest-growing company in the space. Qualitatively, they win by being the best product for restaurants: the most features, the easiest setup, the best user experience. However, that varies by the needs of individual customer segments.

In micro-businesses, Square does well through its simplicity and cost. Will Schreiber from Business Breakdowns shared that customers below $50K tend to go with Square because they don’t need the additional functionality of Toast and can get started for less than $50 on Square with no monthly fee.

According to that same interview, over 60% of Square sellers process $125,000 a year or less, compared to an average of over $1M a year for Toast. Toast, on the other hand, is $70-$160/month with a little more expensive hardware. Toast becomes cost-effective ~$43k in annual revenue due to lower payment processing rates (about 30 basis points cheaper).

In larger chains (15+ locations), customers historically chose enterprise solutions like Revel (owned by Shift4, serving 200,000+ customers), Oracle, or NCR. However, as Toast expands functionality, this becomes much more addressable.

Finally, there are strong switching costs around point of sale systems and its surrounding software, and Oracle and NCR still have the largest market share in US restaurants. In addition, there are other competitors like Clover from Fiserv (targeting $3.5B in revenue this year), Lightspeed ($1.1B in revenue).

The market’s huge and the market’s competitive.

See an overview of the competitive landscape here:

3. Can They Increase Revenue/Location?

We’ve seen revenue growth outpace location growth since IPO (ARR/location CAGR of 5.7%). If we go back to 2015, they’ve grown ARR/location at a CAGR of 10%.

So, they’re likely to continue expanding that number, albeit at a much slower pace than their location and revenue growth.

Finally, AI offers a new and clear expansion path: customer service, phone answering, and drive-thru ordering are all in play and have been mentioned by Toast.

In summary, the market Toast is taking on provides plenty of room to grow in my opinion. But the execution risks increase with each new market they enter. After twelve years of execution, they’ve earned the benefit of the doubt.

III. Some Takeaways from Toast’s Journey

As I think about the takeaways, the big lessons, the playbooks to take away from Toast, I’ve narrowed that down to three thoughts (and one application):

1. Toast Laid out the Ideal Playbook to Move from Workflow -> Platform

1. Figure out the most important piece of data to a given customer set (the order in restaurants, the clinical note in healthcare).

2. Own and automate the workflow that controls that data (the Point of Sale System in restaurants, AI scribes in healthcare)

3. With that new data, solve all the most immediate problems that were previously unsolvable.

4. Build a platform for others to build on to cement your status as the central force of the workflow.

5. Expand over time to own more and more of that platform.

2. Complexity is a Moat

Toast succeeded because restaurants are the worst customer set I can think of (again, all due respect). Complex, low margin, low annual spend.

And that difficulty is what made it hard to execute and limited the amount of competition that would go after it. Who wanted to build hardware, software, and take on the GTM challenge of selling and marketing to 100k(!) locations?

Well, the one who did became a $25B company.

At the same time, the constant turnover of restaurants creates a consistently new market of customers, perfect for upstarts.

I’m reminded of Steve Schwarzman’s (Co-founder of Blackstone) quote: “The harder the problem, the more limited the competition, and the greater the reward for whoever can solve it.”

3. The existence of Toast precludes the existence of the next Toast.

Toast’s playbook is brilliant, but it’s already been played out in restaurants. There can’t be another toast because Toast already exists. They also benefited from weak or non-existent incumbent System of Records. The weaker the competition, the bigger the opportunity.

4. The Current AI Opportunity

Dare I end on one more Charlie Munger quote (from Founders podcast, so loosely translated): “If you're actually paying attention, right, you can spot opportunities in parallel industries.”

A new age of workflows is now automatable with AI, many of which own the most important data to a given customer set (customer meetings in sales, clinical notes in healthcare, legal documents in law, and much of the “agent landscape”).

The path for those companies to expand into systems of records is clear: use the new data that was previously inaccessible, automate the most painful customer workflows with that newfound data, expand as a platform, and gradually build the system of record.

While it took time for Toast (5+ years to expand to a platform), and it will take time with AI companies, this is the clearest path I see to building the “next Toasts.”

That combination of Toast’s playbook + a complex problem in a niche industry + AI-automatable workflows creates a new opportunity to build billion-dollar vertical software companies. An opportunity Mark Leonard (and his littany of hyper-niche-vertical software companies) could get excited about!

As always, thanks for reading!

Disclaimer #2: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

I’ve come away believing the opportunity for AI’s impact on healthcare is one of the most exciting applications in the world. It’s a rare opportunity for a (very) important problem to meet a new technology perfectly suited to solve it. But to lay out that case, we have to first understand the problem.

The US Healthcare System has become a $5T web of suppliers, providers, payers, and middlemen accounting for nearly 20% of US GDP:

This has created an administrative burden that costs $1 trillion annually. Hospitals have teams of people filling out forms and documentation, in addition to the clinicians themselves. I’d include some more stats, but just Google physician burnout and you’ll get the picture.

My goal of this article is to explain (as best I can) the structure of the US healthcare industry today, the incentives that made it that way, and how technology can be a part of the solution.

In summary:

1. There’s no one problem that’s created this system; many variables have contributed to it.

2. While reform is important, it’s much harder than automating processes within today’s system.

3. That system connects the US gov’t, insurance companies, pharmaceutical companies, healthcare providers, a few extra middlemen, and patients; all requiring documentation between each.

4. That administrative burden, while well-intentioned, decreases the well-being of both providers and patients.

5. AI provides a perfect solution to this burden.

To explain the final point, we first have to understand the system today, how we got here, and the incentives for each party involved.

Thank you to the thirty or so people I talked to in preparation for this article. If it takes that many people to understand something, it’s a real indictment of either the complexity of the system or the intellect of the one trying to understand it. I’m afraid, in my case, it may be both.

1. An Overview of the US Healthcare System

I had a friend recently tell me, ‘You love those charts with boxes and arrows…I hate them.’ Well, buckle up!

There are eight primary players in the US healthcare system:

1. Suppliers like (1) pharma companies and (2) medical device manufacturers who sell to providers either directly or via (3) distributors

2. Providers like (4) pharmacies, (5) hospitals, clinics, dental offices, etc, who have (6) group purchasing organizations (GPOs) to purchase on their behalf

3. Payers like (7) insurance companies and the (8) US Gov’t

We can visualize those parties involved here:

The Suppliers

Pharma companies and medical device manufacturers are the two most profitable segments of the value chain.

The vast majority of the most valuable US healthcare companies fall into these categories:

Pharma companies essentially act as the venture capitalists of the healthcare industry: developing drug pipelines in the hope that they find a breakthrough compound that can become a blockbuster drug.

About 10% of drugs make it to market from the clinical trial pipeline, and 1% provide half the profits to pharma companies (from the Acquired episode on Novo Nordisk).

So, pharma companies get a lot of criticism for being profitable, and they do have strong gross margins (70-80% for the large pharma companies). However, much of that is funneled back into R&D to fund further drug development.

Drugs have 20-year patents from filing, but since patents are filed early in development, most drugs have only 7-12 years of protection remaining when they reach the market. After which, drugs become much less profitable due to generic competition. As they say in VC, you’re only as good as your last deal (drug)!

They then sell through distributors or directly to hospital organizations. For context, 92% of prescription drugs are sold through distributors. It’s a lower percentage for medical devices, as they often need to provide service directly to hospitals.

The three largest distributors are McKesson, Cardinal Health, and Cencora, who run very low-margin businesses but have 90+% market share:

Providers

Providers include pharmacies, hospital organizations, clinics, and any other healthcare provider directly serving patients. The CMS publishes annual data on National Health Expenditures, and breaks down spending by category:

Payers

Finally, payers are the insurance providers. The US Gov’t (Medicare + Medicaid) makes up $1.9T of that spending and private insurers make up $1.5T.

According to a 2022 report from the US Census, 54% of the insurance-covered population is on employer-sponsored insurance, 36% on gov’t-sponsored plans, and the remaining 10% covered through marketplace coverage. About 92% of the population has insurance.

Now, if this were the whole system, it would be relatively straightforward. Suppliers make products, providers deliver care, and payers cover costs.

But if there’s one thing Charlie taught us, incentives rule the world. And incentives created middlemen, and middlemen created complexity.

2. An Abridged History of How We Got Here

Three big historical forces created today's system:

First, employers and the government became the primary payers. This started in WWII when wage controls led employers to offer health insurance as benefits, then became permanent with tax exemptions. Medicare and Medicaid in 1965 made the government a major healthcare payer, eventually becoming the largest.

Second, as the industry grew more complex, middlemen emerged to facilitate transactions. PBMs formed to manage pharmacy benefits and negotiate drug prices on behalf of insurers. GPOs formed to give hospitals collective bargaining power with suppliers. This made sense initially when the industry was fragmented; however, as both sides consolidated, these middlemen retained their power despite being less necessary.

Third, over the following decades, the industry consolidated. Pharma companies merged because drug development became increasingly expensive and risky. The "VC model" of drug discovery required firms with enough capital to withstand multiple failures while hunting for blockbuster hits. Insurance companies consolidated for increased bargaining power and the benefits of vertical integration. Now, providers are consolidating too, as hospitals merge into massive health systems to become more efficient and increase negotiating power.

This created a system of oligopolies, where every transaction requires multiple intermediaries, each taking their cut and requiring documentation to pay for their services.

3. Insurance & Healthcare Incentives: Why is Everything so Expensive?

Insurance seems simple enough: you get a service, the provider submits a claim, insurance approves or denies it, then the provider appeals or not, and you get a bill. On the back end though, everyone wants to ensure they get their piece of the pie.

Abridge's founder said it well: "We're not compensated as doctors for the care that we deliver; we're compensated for the care that we documented that we deliver."

Once that documentation is complete, it's translated to medical codes, which are then sent to the insurance provider.

Meanwhile, insurers scrutinize claims closely, and frequently deny them due to poor documentation. This creates an adversarial relationship where providers must extensively document everything to defend against denials.

PBMs & How Drugs are Priced:

Perhaps the most contentious and complex piece of the value chain are pharmacy benefit managers (PBMs). They negotiate rebates for the insurance companies in exchange for the insurance companies putting their drugs on a formulary (list of drugs covered by insurance).

Because the pharma companies know they'll have to pay a cut of their fees back, they raise prices. On average, about 30% of gross drug costs come from those discounts. However, they aren’t paid out as discounts, they’re paid out as rebates.

So pharma sets high list prices (known as the wholesale acquisition cost (WAC)) -> PBMs negotiate rebates -> insurance pays high list prices -> pharma pays rebates back to insurers -> PBMs take a cut of those rebates.

The fact that brings up the most contention is that two of the three largest PBMs are owned by the insurers they’re representing (Cigna/ExpressScripts and UNH/OptumRx) and the third is owned by a pharmacy submitting claims to that PBM (CVS/Caremark - CVS also owns a health insurance company):

This vertical integration means PBMs are parts of the very same organizations they’re representing, and still taking a cut of prices. See operating margin of UNH and Optum:

Hospital & Procedure Pricing

This same pricing process plays out with medical procedures, forcing list prices higher. Every procedure in a hospital has an agreed-upon price with insurers. Providers will set artificially high list or “chargemaster” prices because they know it will be negotiated down by insurers.

This creates two problems:

1. It locks everybody into the insurance ecosystem. Because if patients don't have insurance, then they don't get all of those discounts, making services very expensive.

2. All this negotiation creates more administrative overhead to manage. The exact billing process varies by patient and their insurance, which can include a combination of deductibles, copays, and coinsurance across multiple plans. Thereby creating even more administrative overhead for both providers and payers.

The Administrative Costs of All This Bureaucracy

Now, all of this complexity and bureaucracy creates this massive administrative burden I’ve been alluding to. And what a burden it is!

• Admin costs make up ~20-25% of the $5T US healthcare industry.

• Insurance administration costs are ~7% of all US healthcare costs.

• Physicians spend roughly between 25% and 50% of their time on admin work (studies vary widely).

This clinical admin work starts with documentation. Every clinician has to document interactions with patients to ensure they can be billed for those services. Some services require prior authorization, which forces the provider to call/fax/submit a PA to the payer to get approval before providing a drug/procedure. Drug pricing programs like 340B help keep expensive drugs affordable, but require a another process of documentation and compliance.

“So much of the story on the clinician side is about agency... You're always paying debt on work, you're never able to get ahead, and you don't have any control over your time. That's why they call this 'pajama time'—this affliction where doctors are writing notes after dinner or after their kids are in bed.” - Shiv Rao, founder of Abridge

This is where AI comes in. For the first time, we have a technology perfectly suited to handle these text and voice-based administrative processes.

4. Healthcare Software, AI, and the Case for Market Expansion

If you made it this far, (1) I apologize for the headache, and (2) time for some optimism!

The last two decades were about digitizing healthcare, getting everything electronically stored. Now, the industry's mostly digitized. However, that digitization came at the expense of providers.

Clinicians had to document patient interactions in free-form notes, then admin teams had to restructure that information for billing and claims. Any mistakes led to disputes with the insurance company.

Because of this complexity, healthcare software's efficacy was limited, and it placed a huge burden on providers. This meant value primarily accrued to EHRs (documentation) and revenue cycle management (billing).

With LLMs, we now have the best way to process unstructured information in the history of technology.

The starting point is AI scribes, which immediately reduce the admin burden on clinicians. Instead of clinicians spending appointments or evenings taking notes, they can spend them with patients.

Abridge, the $5B AI scribing company, has been at the lead of recent scribing startups (after their competitor, Nuance, got acquired by Microsoft for $20B).

Their CEO laid out the path to expanding past scribes (lightly edited for clarity)

"If you believe as we do that healthcare is about conversations—that it's one of the first original signals in healthcare—then you start to see that any number of different workflows are beyond it. It's not just clinical notes, it's also orders... After orders is a claim, is a code, is a bill that goes to the insurance company—so all things revenue cycle."

This initial automation creates a flywheel of:

1. Automate admin work

2. Improve documentation and increase the value of EHRs/RCMs

3. Automate more admin work (while models get better)

Given over $1T is spent annually on healthcare admin, that becomes one heck of a flywheel. I think it lays out the case for pretty massive market expansion across the healthcare software/AI landscape.

Flare Capital Partners made an excellent map of the various companies across this landscape:

Now, as basically every healthcare professional will tell you, it’s not perfect: models are still improving, regulations make adoption slower in healthcare, mistakes can be high consequence, and data privacy is a real concern. So, adoption cycles are long, and it may take longer than expected to provide the scale of impact that’s possible.

But like I said at the beginning, this is a rare opportunity where a massive problem meets a new technology perfectly suited to solve it.

That’s something to get excited about.

If you’re working on this problem or with these tools and have thoughts on the future, feel free to reach out. As always, thanks for reading!

Disclaimer #1: While my goal is to simplify, this is not a simple issue. I’m focusing on one problem: the administrative burden, how it came to be, and how it could potentially be fixed. There are many others, and this impacts real lives. I could make many more disclaimers but I’ll leave it at that.

Disclaimer #2: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

TSMC introduced the foundry business model to the semiconductor industry in 1987, allowing companies to outsource chip manufacturing, taking advantage of economies of scale and Asia’s cheap labor. This would unlock a new wave of semiconductor companies and make the idea of a “semiconductor startup” much more feasible.

Over the next 30 years, TSMC would reshape the technology industry around it. Now, it’s the 11th highest valued company in the world, it manufactures the vast majority of the world’s most advanced chips, and it’s become an index on the world’s most advanced technologies.

This is my (humble) attempt at covering TSMC in one article, a fool’s errand perhaps! We’ll cover:

1. TSMC’s Past: How TSMC reshaped the semiconductor industry

2. TSMC’s Present: The moats of TSMC

3. TSMC’s Future: The variables likely to determine its future

See a summary of that journey here:

Onward.

1. How TSMC Reshaped The Semiconductor Industry

Two ideas laid the groundwork for TSMC’s founding.

1. As semiconductor manufacturing got more complex, the barrier to entry for new companies got higher.

2. Morris Chang believed more semiconductor companies would be founded if startups didn’t need to build a fab to start producing chips.

In 1987, when TSMC was founded, the world’s largest semiconductor companies were vertically integrated: designing their chips (sometimes with their own software and machines), building fabs, and manufacturing chips.

We had just started to see the rise of the modern EDA providers (Cadence Design Systems was doing ~$500M in revenue) and the modern semicap companies (Applied Materials was also doing ~$500M in revenue). But the vast majority of the industry’s revenue went to vertically integrated chip manufacturers:

Per Moore’s Law, semiconductors were getting more complex, which leads to one important observation on the evolution of industries: complexity necessitates specialization.

Morris Chang, through his experience at Texas Instruments, believed more people would start semiconductor companies if they didn’t need to build fabs. Through a winding path, he ended up at Taiwan’s Industrial Technology Research Institute (ITRI), where he’s effectively forced to resign. In his own words, he wonders if he’s risen to the level of his incompetence.

Now, he did have one successful accomplishment at ITRI: the founding of Taiwan Semiconductor Manufacturing Company.

TSMC’s Rise

TSMC’s early customers follow the disruption playbook exactly: large companies like Intel contract with TSMC to manufacture their lagging-edge chips to free up space for their high-margin leading-edge chips.

Because the margins are worse at the lagging edge, TSMC was forced to innovate. They built new fabs, pioneered new manufacturing techniques like SMIF wafer-carrying pods, and took a chance on a new lithography company called ASML.

After a couple of years, Morris’ thesis on fabless companies comes to fruition. Qualcomm, Nvidia, Broadcom, and Marvell (founded 1985, 1993, 1991, and 1995) all pursued the fabless business model.

The value prop of the fabless model was this: With TSMC, fabless companies innovate faster, have lower CapEx, and better operating margins. TSMC, in turn, gets manufacturing volume, which slowly compounds in efficiency to draw in more customers, keeping the flywheel going.

As modularization unlocks specialization, this new ecosystem of fabless companies outperforms the old. TSMC continues to compound their learnings in manufacturing, slowly encroaching on the leading-edge.

By 2000, 60% of TSMC’s revenue comes from fabless companies. Over the next decade, the fabless model would emerge as the clearly superior business model, with AMD spinning off their fabs in 2009. Intel would be the last leading-edge logic integrated manufacturer.

However, TSMC would see increased competition as well, prompting Chang’s 2009 return (from a retirement in 2005). He emphasized one point on his return: focus. The core foundry business is the company’s priority, and that’s what they’d focus on.

They created a wave of fabless chip companies and it was Morris’ intention to ride that wave.

The Apple Partnership and TSMC’s 2010s

In the early 2010s, TSMC started manufacturing chips for Apple. For TSMC, it meant an initial investment of $9 billion dollars and devoting 6000 employees to building a dedicated plant for Apple in just 11 months. For Apple, it meant “no backup plan.”

Over the next decade, Apple would grow to ~25% of TSMC’s revenue, or about $18B worth. In return, TSMC gave Apple preferred access to the leading-edge nodes.

In parallel, the cloud was becoming the preferred model for computing, further expanding the world’s computing demands. Leading up to the AI moment, which, as we all know, further expanded the world’s computing demands.

While these trends massively expanded the semiconductor market, TSMC consolidated the foundry market:

In twenty years, TSMC went from one of 26, to one of two. It was the perfect combination of growing pie and TSMC’s slice getting bigger.

To summarize TSMC’s journey:

1. They introduced the foundry business model to the industry.

2. That unlocked a wave of fabless chip companies.

3. Three tailwinds led to TSMC’s meteoric rise: a growing semiconductor industry, increasing importance of fabless chip companies, and increasing market share of TSMC.

That brings us to the current state of TSMC.

2. The Moats of TSMC

Instead of exhaustively covering TSMC’s technology and revenue, I’ll link to two great pieces covering them: Construction Physics’s “How to Build a $20B Semiconductor Fab” and Fabricated Knowledge’s latest TSMC update.

For a quick summary on financials, TSMC surpassed $25.5B in quarterly revenue last quarter, up 35% Y/Y, with a gross margin of 58.8% and a net income margin of 43.1%:

Instead, I want to focus on what I see as the clearest competitive advantages of TSMC (outside of just monopoly from sheer technical complexity):

1. TSMC’s Economies of Scale

Manufacturing ultimately comes down to constant, subtle improvements that compound over time. This makes volume especially important, and once those advantages compound, it becomes increasingly challenging for competitors to break in.

For TSMC, the process goes like this:

• Invest CapEx, primarily into fab equipment like ASML’s machines, into new fabs at the leading edge.

• Depreciate those costs over 5 years on the P&L statement. At the same time, those fabs get increasingly profitable on unit economics as the process gets better.

• After the fab is fully depreciated, use those profits to fund R&D and CapEx into new process nodes.

This results in leading edge nodes being unprofitable for the first few years, followed by strong profitability to fund more leading edge nodes:

All the while competitors have less volume at trailing edges, worse yields at the leading edge, and fight to maintain the ability to manufacture at the leading edge.

2. The Collaboration of the Semiconductor Industry

Secondly, as semiconductor complexity gets harder to scale, TSMC has to partner with both suppliers and customers to ensure chips can be manufactured.

The industry’s gains are coming from (1) improvements in EDA tools, (2) improvements in design, (3) improvements in semicap tooling to further increase transistor density, (4) improvements in packaging these components together more efficiently, and (5) the foundry’s ability to integrate all of these improvements together.

It’s not easy for new entrants to break into any piece of that workflow, nonetheless the foundries, who are at the center of the process:

With customers, an engagement with TSMC is typically a multi-year process. TSMC will offer a Process Design Kit (PDK) that’s like a detailed instruction manual for how/what features will be manufactured. Customers will pay an upfront cost for how much work and customization is required. They’ll then pay per wafer produced.

For a company to switch fabs, they’ll need to switch PDKs, re-design chips with the new fab, test to ensure there aren’t defects, and lose out on the sunk costs of the prior relationship.

And that’s if there’s even another viable alternative on the market!

3. Position in the Value Chain

Austin Lyons wrote an exceptional article called ”Monopsony and TSMC”, describing what happens in a market when there’s only one buyer. And while we aren’t quite there with TSMC, that’s the logical extreme we’re approaching.

TSMC’s importance in driving forward semiconductor progress combined with their increasing market share has allowed them to take an increasing share of revenue and profits from the industry:

Because they’ve become the bottleneck of the industry, they have extreme pricing power. Now, they have very healthy margins and no incentive to squeeze their customers or suppliers on costs. Regardless, all else equal, every industry participant would prefer a fragmented supplier and customer base. And both sides of the value chain have to come to TSMC.

4. Those competitive advantages have allowed TSMC to take an increasing amount of market share:

While the numbers aren’t public, this number is much higher at the leading edge.

Depending on who you ask, TSMC has two to three potentially credible competitors at the leading edge in the near future: Intel, Samsung, and SMIC:

Today, TSMC is the clear dominant player in foundries.

There are two primary risks to TSMC that I’m monitoring:

1. A disruptive technology doesn’t emerge to make semiconductor manufacturing cheaper and easier.

2. Semiconductor manufacturing continues to drive computing performance improvements.

These two variables ensure TSMC is still leading and driving forward semiconductor scaling, which means they can maintain or increase (1) their share of foundry revenue and (2) their share of customer expenses (i.e. what % of spend of their COGs go to TSMC).

For example, chiplets could fit the bill for a disruptive technology. They could push a higher % of spend towards advanced packaging, which has lower competitive advantages compared to wafer manufacturing. Perhaps a primer on chiplets is due in the future.

3. The AI Opportunity and the Path Ahead

I want to simplify the variables that matter to TSMC as much as possible. First and foremost, TSMC is a customer service company. The success of TSMC is the success of their customers. Those customers broadly fall into two categories:

1. Edge devices via Apple, Qualcomm, MediaTek

2. AI & Data Centers via Nvidia, AMD, Broadcom, Intel, and the Hyperscalers

Apple and Nvidia made up 22% and 12% respectively of TSMC’s revenue in 2024. From 2022, we have estimates of who the other largest customers are:

The most immediate driver for computing growth is AI workloads, primarily through Nvidia:

And the great Sravan Kundojjala modeled out what that growth could look like over the next 5 years, with AI approaching 40% of TSMC’s revenue in 2029:

With exposure to AI, edge devices, and seemingly every other major technology on the horizon (robotics, the metaverse, etc), TSMC has become an index on the future of technology.

As I said last year, and as I’ll be saying for many years: it’s never a good idea to bet against technology, and one I won’t be making any time soon.

As always, thanks for reading!

[You’ll notice I didn’t discuss geopolitical concerns. I have nothing unique or valuable to add to that conversation.]

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

While I make no personal claims of first-rate intelligence, I’ve observed two opposing ideas from two people who undoubtedly do have first-rate intelligence:

1. “Show me the incentive and I will show you the outcome." - Charlie Munger

2. “I’m not saying OpenAI would never consider ads, but I don’t like them in general, and I think that ads-plus-AI is sort of uniquely unsettling to me.” - Sam Altman

In The Future of Foundation Model Business Models, I made the case for how foundation model companies would pursue profitability.

At the time, I laid out the case for vertical integration, profiting from applications, and using models as a differentiator. However, that argument was incomplete.

It focused on the 4% of paying ChatGPT customers, and not the 96% of free users. Advertising could be the answer to monetizing those users, which I’ve started to lay out in my previous two articles:

1. The Evolution of Digital Advertising

2. The Current State of Digital Advertising

This article will be a bookend on that series, covering:

1. The Incentives of Advertising

2. The Challenges with Advertising

3. The Timing of AI Advertising

4. How will ads be implemented?

5. The impacts of AI & Advertising

Let’s get to it.

Disclaimer: ‘The Siren Song’ refers to the allure of advertising, but not necessarily the detrimental nature of the Sirens in classical Greek Mythology. Analogies can’t be perfect!

1. The Incentives of Advertising

In consumer technology, no business model better incentivizes all parties involved than advertising.

As we saw with the development of digital advertising, advertisers paid to use the platform because that’s where attention flowed, giving them a chance to influence consumers. Secondly, advertisers pay so much for advertisements because the user data and algorithms to target those users allows them to pay for improved odds of purchasing.

Over the last three years, we’ve seen attention flow to LLMs. With the development of memory, LLMs are getting to know their users. Secondly, and more importantly, we’re seeing the glimmer of a world where those LLMs take actions on your behalf. In that world, consumers no longer search for what toothbrush to buy; they ask their assistant to buy it for them.

How much would advertisers pay for hyper-targeted attention AND vastly improved odds of purchasing?

Put eloquently, a lot.

If ChatGPT were indeed able to aggregate purchasing into its app, that “Everything Engine” would create, in the words of Ben Thompson, “the most compelling ad opportunity since web search results pages.”

The Dual-Benefit of Advertising

The challenge with advertising has always been the perception that “the user is the product.” The challenge has never been how profitable the model is. Advertising helps monetize products in two ways:

1. It helps monetize the 96% of free users today (and in turn, helps build a sustainable business model that keeps those services free; i.e., advertising isn’t all bad).

2. It encourages free users who don’t like ads to upgrade to the paid version.

For OpenAI, the targeting of those free users is coming:

2. The Resistance to Advertising

Despite all of those incentives, there’s a remarkable consistency of companies resisting and then pursuing ads:

Google: "We expect that advertising-funded search engines will be inherently biased towards the advertisers and away from the needs of the consumers." - Page & Brin, The Anatomy of a Large-Scale Hypertextual Web Search Engine (1998)

Amazon: “Advertising is the price you pay for having an unremarkable product or service.” - Jeff Bezos

Perplexity: “Perplexity was founded on the belief that searching for information should be a straightforward, efficient experience, free from the influence of advertising-driven models.”

Netflix: “We want to be the safe respite where you can explore, you can get stimulated, have fun and enjoy - and have none of the controversy around exploiting users with advertising.” - Reed Hastings

Meta: I couldn’t find a direct quote, but I can quote The Acquired Podcast here: “The other thing that’s happening here is Mark still thought ads were gross at this point in history.”

This isn’t to disparage these companies at all, it’s to highlight the strength of the advertising business model!

Which leads us a few questions:

1. Why is the distaste for ads so consistent?

2. When do companies eventually make the shift to ads?

3. How can you introduce ads in a way that’s not detrimental to the user experience?

To keep the first answer short, the ethics of advertising have always been debated. Nobody likes the idea of “being the product.” Additionally, since LLMs give direct answers and not a suite of answers, influenced answers are particularly concerning. Finally, people are rightfully nervous about AI and its societal impacts. The idea of it being funded by advertising makes it even more concerning.

Because of this, figuring out the right way and time to introduce ads is especially important.

3. The Timing of AI Advertising

Different companies in the ecosystem have different incentives on how immediately they need to figure out advertising.

Google, for example, will be the first company to see lost revenue from traditional Search (given its market dominance). So they’re heavily incentivized to figure out the advertising model.

Perplexity, as well, is incentivized to figure out advertising, given their direct competition with Google. In their own words, “To fully deliver on our mission to spark the world’s curiosity, we need to invest in building not just a beloved product, but a robust and self-sustaining business.”

However, a company like OpenAI is less incentivized to pursue advertising in the short-term for two reasons:

1. Monetization follows attention, so the timing of monetization matters less.

For aggregators, the asset is the network of users. The monetization is a second-order effect. For example, take a look at Facebook’s Daily Active Users (DAUs) vs revenue over time:

Facebook DAUs over time have compounded at approximately 12% since IPO, while it’s revenue has compounded at 35% since IPO. The Instagram acquisition does change those numbers, but the point stands: focus on the user base, monetize later.

2. There’s no immediate need for profitability (i.e. they have plenty of funding.)

The parallels between search and LLMs are incredible, take a look at this article on Google’s early years:

For the first four years of Google’s life, they resisted advertising, despite the urging of the two venture capitalists, John Doerr and Michael Moritz, who had invested in Google and were upset that it was bleeding money. Brin and Page insisted that traditional banner ads would offend their users. They believed that if they built search traffic they would eventually engineer a way to make money; in the meantime, they hoped to improve people’s lives through information.

Eventually, Google pursued advertising because of pressure to find a sustainable business model. Investors like Mike Moritz started to get “antsy”, saying: “We really couldn’t figure out the business model…there was a period where things were looking pretty bleak.”

The pressure for a sustainable business model has historically been the impetus to pursue ads. Given the parallels between search and LLMs, I think ads will come, but they’ll come slowly.

4. How could ads be implemented?

From my observations, the incentives of advertising are too powerful to ignore. However, companies like OpenAI have the opportunity to be some of the largest companies in the world twenty and thirty years from now. So their brand is too important to damage with the idea of selling user data for personalized ads.

So, the question for these AI companies is how can you profit from aggregation without damaging the user experience?

1. Be hyperselective with the ads you show

2. Be hypersensitive about user data, don’t personalize ads at first.

3. Influence responses as little as possible

This lowers the maximum monetization potential of ads, and that’s okay. It’s much more important to do it ethically and correctly, than in a profit-maximizing way. Monetization follows aggregation.

Today, the initial attempts around ads are basic banner ads, suggested questions, and Pay-Per-Click Ads. Take Microsoft Copilot or Perplexity for example:

However, people come to LLMs for unbiased information. Because of this, it’s unlikely the AI providers provide personalized responses influenced by ads.

Why the affiliate strategy makes so much sense

With rumors of a Shopify and OpenAI partnership, an alternative path for AI providers is to surface personalized recommendations for e-commerce, centralize purchasing, and take a commission on any purchases through their platform.

Sam Altman hinted at this in his Stratechery Interview:

“The kind of thing I’d be much more excited to try than traditional ads is…e-commerce, for example, and is there a way that we could come up with some sort of new model…if you buy something through Deep Research that you found, we’re going to charge like a 2% affiliate fee or something.”

LLMs can combine the information search of Google, the personalized user data of social media, and (in the future) the proximity of purchase of Amazon.

E-commerce via LLMs saves users time, increases platform stickiness, and increases revenue for the merchant. It also allows platforms to very clearly demarcate ads vs non-ads; information is not influenced, but recommended actions are.

More interestingly, it gets users used to the action of buying through ChatGPT, which lays the path towards the absolute best case for ChatGPT.

For this to be really successful, ChatGPT would need to become more than an LLM. It’d look a lot more like super apps in other parts of the world. It could become the default go-to for shopping, search, and (maybe even) social media.

[To be clear, affiliate fees are not directly advertising; however, they are a means of monetizing free users with a clear path towards advertising.]

5. Who Does This Impact?

We’ve only seen a glimmer of the potential of new AI companies; in the early days of industries, it’s easier to see the losers than the winners.

As WB said,

“I’ve always said the easier thing to do is figure out who loses. And what you really should have done in 1905 or so, when you saw what was going to happen with the auto is you should have gone short horses. There were 20 million horses in 1900 and there’s about 4 million horses now. So it’s easy to figure out the losers, you know the loser is the horse.”

Traditional, ad-saturated ten-blue-links search will decline as users switch to LLMs. As we’ve seen with the last waves of digital advertising, monetization lags adoption. So while it will happen slowly, as search moves towards LLMs, eventually advertising dollars will try to move as well.

If this thesis turns out to be true, it provides not only a clear path for AI companies to generate sustainable business models, but it hints at the best-case scenario for today’s AI companies becoming the super apps of tomorrow.

Finally, it opens up a host of questions about the potential for new companies to be built around this ecosystem. I have to imagine those answers are stored in the minds of many first-rate intelligence people reading this, so feel free to reach out if you have thoughts.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

Because of this, they give us a perspective into AI markets that no one else can. It’s why I’m so fascinated by the game theory of their decision making, and why I write a quarterly update on their businesses: investment strategy, market share updates, and growth rates. Let’s get to it.

The Game Theory of Hyperscalers’ AI Investments

Given that AI seems to be the driving force for accelerated computing needs over the next decade, the hyperscalers are as well-positioned as nearly anyone to benefit from AI. As AI makes humans more productive, they’ll “outsource” their labor to computers.

As AI coding tools further decrease the barrier to entry to software, more and more software will be built. We’ll only get more compute-hungry, and the hyperscalers couldn’t be happier.

Because of this, they’ve fully pursued AI, pouring hundreds of billions into infrastructure investments.

As I laid out last year, this was logical:

The equation for hyperscalers is relatively simple: data centers are at least 15-year investments. They are betting compute demand will be higher in 15 years (not a crazy assumption).

If they do not secure these natural resources, three things could happen:

1. They will lose business to competitors who have capacity.

2. They’ll have to secure sub-optimal land with worse price/performance profiles.

3. Challengers will buy this land and power capacity and attempt to encroach on the hyperscalers.

If they secure this land and don’t need the computing power right away, then they’ll wait to build out the “kit” inside the data center until that demand is ready. They might’ve spent tens of billions of dollars a few years too early, and that’s not ideal. However, the three big cloud providers have a combined revenue run rate of $225B+.

As Google’s CEO Sundar said, “The risk of under-investing is higher than the risk of over-investing.”

However, for the first time, we’re hearing reports of these investments slowing down.

In part, this is simple downside limitation. If demand constricts, they won’t be exposed to tens of billions worth of idle compute.

However, the more interesting part is that they’re adjusting investments based on “the shape of workloads” moving forward. As Satya said:

“The reality is we've always been making adjustments to build, lease, what pace we build all through the last, whatever, 10, 15 years, it's just that you all pay a lot more attention to what we do quarter-over-quarter nowadays.

Having said that, the key thing for us is to have our builds and lease be positioned for what is the workload growth of the future, right?

You don't want to be upside down when the shape of demand changes because, after all, with essentially pre-training plus test time compute, that's a big change in terms of how you think about even what is training, right.”

I'll lay out a revised equation. The cloud providers still want to maximize upside from three areas:

1. AI inference (APIs, GPUaaS, etc.)

2. AI companies building on their platforms (just like AWS rose by supporting the early cloud software winners)

3. Their own AI products (The AI pie is still up for grabs)

Inference, as the enabler of applications, resembles the recurring SaaS revenue model of the last decade. The hyperscalers will do everything they can to power those applications.

Training, on the other hand, requires larger, upfront investments. As Zuck described, “I’d guess that in the next couple of years, the training runs are going to be on gigawatt clusters, and I just think that there will be consolidation.” These large-scale expenses present a less attractive risk-reward profile.

The last two years have come with a lack of clarity on the future of AI. In the absence of clarity, the biggest risk was underinvesting.

However, as the economics of AI come into focus and these infrastructure investments unlock AI applications, the hyperscalers can focus on investing in a future that provides them the best ROI per dollar deployed with the least risk of overinvestment. Risk-adjusted AI, if you will!

Updated Market Statistics

What does this mean? The hyperscalers are getting cautious, not in the short term, but in the long term. They haven’t changed their short-term spending meaningfully; if they do, it won’t be reflected immediately.

The good news is that they’ve got the money to invest. Combined, the three cloud providers are at a $247B run rate, growing 23.9% Y/Y.

A reminder on all of these numbers: the three cloud providers report different product lines in their “cloud” units. Azure just restated their numbers a few quarters ago, GCP includes Google Workspace. The numbers are all directionally correct, and precisely incorrect. Azure includes estimates of revenue based on public commentary.

We can see updated (estimated) market share here:

And the more interesting way to view market share is what % of net new revenue generated over the last 12 months did each company capture:

Based on this, Azure has had a few very strong quarters, perhaps because of the OpenAI relationship.

I’ll end with one final stat (and one that Warren Buffett might find amusing). For all of the noise of the last 6 months, the average stock performance for Google, Amazon, and Microsoft is just about 0%.

There’s a lesson in that. Unfortunately for me, it may be to read fewer quarterly updates, and fewer updates in general. As Nick Sleep said, “We keep our discussions to as high a level as we can manage in the belief that, in the long run, the high level is all that matters.”

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

In it, Arthur observes that for a small handful of companies, they don’t see the decreasing returns that had been observed for several centuries, but increasing returns. The strong get stronger. The rich get richer.

If you had invested behind this one theory for the last 30 years, it’s been a very fruitful thirty years for you (outside of any catastrophic behavioral or leverage mistakes, of course).

For the last two decades, the digital advertising market has epitomized those increasing returns. Consolidating to a few, massive, power-lawed, network-driven, attention marketplaces. In that time, not only did they produce two of the largest IPOs ever, but they also achieved wonderful public returns.

Translated: Early on, people saw how incredible Google and Meta would be. The only thing they missed was just how incredible.

Now, however, the market’s at its most interesting point in a decade:

1. Attention is shifting as information retrieval fundamentally changes.

2. Regulators are looking to remove the advantages built up by digital ad networks over the last two decades.

3. Consumers are more immune to advertising, leading advertisers to pay most for the ability to influence buying decisions.

4. One of the harbingers of digital ads, social media, has fundamentally shifted from social media to media.

I can’t help but be deeply interested when there’s a massive industry facing the shifting sands of disruption.

With that, let’s get to it. This will be a deep dive on the digital advertising industry: its value chain, market statistics, and trends shaping the industry.

A Quick Primer on Digital Advertising

Given the profile of these businesses, I’ll keep the primer section short (and perhaps best explained in visual format below).

Broadly, we have a few categories that make up the advertising industry:

1. Walled Gardens like Meta, Google, Amazon, Tiktok that require you to go through their platform if you want to advertise with them.

2. Non-Walled Gardens include the rest of the channels that showing digital ads such as publishers like the Wall Street Journal, music providers like Spotify, mobile games, and streaming providers like Netflix.

   1. Demand-Side Providers (DSPs): Platforms like the Trade Desk, that aggregate advertiser demand, to connect to ad exchanges/SSPs.

   2. Supply-Side Providers (SSPs): Platforms like Google’s AdX (now integrated into Ad Managers) that aggregate advertising inventory from multiple publishers, connecting to various Demand-Side Platforms (DSPs).

   3. Channel-Specific infrastructure: Platforms that provide an ad network to a specific vertical; AppLovin, for mobile gaming, is the best example of this.

3. Ad Agencies like WPP and IPG that build campaigns and manage advertisements for advertisers.

4. Publishers like the WSJ that sell ad inventory.

5. Supporting tooling like tooling to design ads (Canva, Adobe), analytics to track ROI (typically rolled up into an ad platform or CRM), and ad servers (store and place ads based on the ad auctions).

The Walled Gardens make up over 50% of the US digital ads market:

Three primary types of digital ads drive the walled gardens: search (consumers come for information), “social” media (consumers come for entertainment), and e-commerce (consumers come to buy).

The nature of these systems is built around accurately matching large amounts of supply and demand. What the supply and demand are varies by use case, but they could be content, ads, advertisers, or user preferences.

Regardless, these matching problems are a perfect use case for AI. Matching also needs to happen in real-time, meaning the hardware requirements for latency are quite high. Because of this, digital advertising essentially bankrolled the AI industry for the last decade. It was the most valuable application of AI, and funded the developments that have taken the world by storm over the last two years.

Outside of those walled gardens, the majority of spend goes through the “open web.” With that context, I see four trends driving the industry:

1. Regulation & Google Ad Network’s Decline

2. The Rise of E-Commerce Advertising: Increasing Consumer Ad Immunity and the Pursuit of ROI

3. The Shift from Social Media to Media

4. The Shifting of Attention in Search

Trend #1: Regulation & Google Ad Network’s Decline

Last week, Google was deemed a monopoly, specifically on the supply side of the “open web.”

This starts with Google Search. For advertisers to place ads on search, they need to go through Google Ads (formerly known as AdWords). Those open-web ads get placed through Google Ad Exchange.

Publishers, on the other hand, need access to that vast number of advertisers through Google Ad Platform. This includes ad servers, which store and place ads on ad inventory in real-time.

The argument from the DoJ was that Google unfairly tied these services together, specifically the ad server and the ad exchange. In doing so, they created a self-reinforcing loop for advertisers and publishers:

1. Advertisers needed to access publishers, which is done through Ad Exchange.

2. Publishers needed to access advertisers, which is done through Ad Exchange.

3. By tying together the ad server and the ad exchange, publishers were locked into the Ad Exchange. In turn, that locked both parties into the Ad Exchange, which locked them into the whole network.

I won’t comment on the unfair practices, but regardless, it led to a dominant flywheel of an attention marketplace. As publisher supply increased, so did advertiser supply, and vice versa. Google then leveraged those data advantages, and their AI prowess, to further cement their advantages.

That led to Google having the largest market share in North America in DSPs (see above), Ad Servers (the DOJ estimated Google had a 90% market share in ad servers), and SSPs.

On top of that, they have both data and existing relationships from advertisers needing access to both Search and Youtube.

With all those advantages, Google should have continued to get more dominant. And they did, for a while. However, over the last three years, that dominance has flattened:

Whether voluntarily or involuntarily, Google’s ad network has ceded share to competitors like the Trade Desk. This means the $30B+ open web ad market may be up for grabs, and Google’s ad network may get spun off from the parent company.

However, the more interesting implication is this: if Google's Ad Network gets broken up, could it create a Standard Oil-like effect on the digital ad network? Will Google’s Ad Network get a cool new name and be unlocked to fully leverage those advantages listed above?

Trend #2: The Rise of E-Commerce Advertising: Increasing Consumer Ad Immunity and the Pursuit of ROI

One other interesting ruling from the Google case was that advertisers are actually “agnostic” to the platform they advertise on, they just want ROI:

Substantial trial evidence showed that advertisers reallocate resources among different digital advertising channels based on perceived return on advertising expenditures (“Ultimately, advertisers are following eyeballs. They are trying to reach people wherever they are.”).

There are two ways to pursue this ROI:

1. More engaging content

2. More influence over buying decisions

The rise of e-commerce advertising epitomizes the latter. Amazon, for example, is the closest platform to consumer buying decisions. By owning the distribution network of buying, they can auction off the most valuable attention in the world: those who are willing to buy.

One of the beauties of Google’s business model is that users come with intent to the platform, they know what they’re looking for. Amazon supercharges that by its proximity to the purchase.

For as long as Amazon has the ability to influence buying decisions, their advertising business will do well.

As we consider the potential for advertising in LLMs, they can combine the information search of Google, user data like social media, and (in the future) purchase proximity data from Amazon.

I’ll call out that OpenAI is reportedly working on a native integration to Shopify to allow you to check out from ChatGPT. Take that for what you will.

Trend #3: The Shift from Social Media to Media

The other avenue for improved ROI is more engaging content, and this has led to the decline of “social” media.

Perhaps the most interesting data point of the last few months came from Meta’s slide on the shifting of user behavior on social media:

In 2025, only 17% of user time on Facebook is spent engaging with friends’ content and 7% of time on Instagram.

This transition was pioneered by Tiktok, which has benefited massively from it. According to Sacra, ByteDance generated $146B in 2024 revenue, of which $34B came from Tiktok.

And at the expense of Meta. See a few quotes here from the recent trials:

1. Facebook’s and Instagram’s share of the amount of time people spend on social media apps has “gone down meaningfully,” Meta Platforms CEO Mark Zuckerberg testified on Wednesday,

2. “Friends and family sharing” was “less and less” a core way that people use Facebook.

3. He testified that TikTok is “bigger than either Facebook or Instagram,” while “people spend more time on YouTube than Facebook and Instagram combined.”

4. “TikTok’s growth so far is worrying and their trend-line projections have them overtaking Instagram in the US in terms of total [time spent] … TikTok in the US is a much bigger threat to our entire family of apps.”

On one hand, this shift from social media to media created an incredibly profitable opportunity for Tiktok and Meta. On the other hand, it creates an Innovator’s dilemma of sorts. The core job to be done of connecting people online is now unfilled, as the social media companies have pursued the most profitable activities.

It could be why Sam Altman’s serious about creating a social media component of OpenAI.

Perhaps this opens the door for another technology that can offer equivalent engagement and a social component. (This sounds a lot like Meta’s vision for the metaverse). *Hey Siri, remind me to write an article about the metaverse and technology realization times.*

Trend #4: The Shifting of Attention in Search

Conveniently, these trends all align with a few observations:

1. Advertising is harder than ever, leading to demand for new advertising channels.

2. Advertisers will pay most for the actual ability to influence buying decisions.

3. At the same time, the way we search for information is fundamentally shifting.

The opportunity for LLMs and advertisements creates both an incredible opportunity and an incredible conundrum. And one that creates perhaps the most interesting debate in the history of this industry. We’ll see that debate play out over the coming years, and it’s worth understanding deeply.

Attention -> valuable data -> influence -> revenue.

In other words, attention is all you need, and attention is shifting.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

As I’ve been thinking about those business models for AI companies, I’m afraid I sorely overlooked the most powerful economic force that drove the last wave of consumer giants: advertising.

Nearly every large consumer technology company founded over the last two decades was built on the business model of marketplaces, advertising, or both.

The business models that ~really work~ ultimately come down to aligning incentives.

On the internet, advertising was the only way to align incentives for all parties involved. Consumers wanted free products. Creators wanted to get paid for their services. The aggregators wanted to make money. Advertisers wanted to sell to consumers. Advertising via aggregated networks aligned incentives for everyone.

For the most obvious example, Google and Meta capitalized on aggregating and auctioning off attention. They did so with a dominant consumer product, user data, and training models on that data. In doing so, they created arguably the best business models of all time.

For a less obvious example, AppLovin helped incentivize the mobile gaming ecosystem by creating an advertising network for mobile gaming. It provided a means for creators to profit from their games. In a world of thousands of mobile apps, advertising created economic incentives for creators to build better apps and profit from them.

Let me play those two examples back:

1. Dominant consumer products, user data, and training models on that data.

2. In a world of thousands of mobile apps, advertising creates economic incentives for creators to build better apps and profit from that.

The parallels we’re seeing in LLMs and AI app development platforms are remarkably parallel to those two developments in recent waves of technology.

This brings us to a very clear question: Can advertising be the answer to align incentives for all parties in the AI ecosystem?

This will be the first in a three-part series covering advertising business models in the age of AI.

This article will explore the evolution of digital ads and lessons from the last generation of advertising giants.

Let’s get to it.

The Evolution of Digital Advertising

As we look back at the history of digital advertising, we see a few very clear trends:

1. Google captured and monetized interest, by targeting ads when a user came to them.

2. Facebook captured and monetized attention, by targeting ads based on users’ interests.

3. AI supercharged both models, by taking user data and improving targeted advertising.

4. Tiktok catalyzed the shift from social media to media, by supplying content based on interest instead of relationships.

5. Amazon captured and monetized intent, by targeting ads when a user came to buy.

6. Data privacy measures were implemented, forcing ad networks and advertisers to rethink how to advertise.

Throughout all of this, there were a few key lessons:

1. If you can capture attention, you can monetize through advertising.

2. The more you can influence consumer behavior, the more you can charge per ad.

3. The more engaging and personalized ads can be, the more you can charge per ad.

The final lesson is that while advertising is an incredible business model, basically everyone has some level of discomfort with it. Google resisted ads, Facebook resisted ads, Amazon resisted ads, (more recently) Perplexity resisted ads. But it’s just an incredibly powerful way to provide services for “free” on the internet, create incentives for creators, provide value to advertisers, and create value for the ad networks.

This article calls out several digital advertising articles on Stratechery; I highly recommend the decade-plus of articles he has on these companies.

The Early History of Digital Advertising

For the better part of 100 years, advertising meant newspapers. They had monopolies on distribution and monopolies on attention via that distribution. A big brand would partner with an ad agency, who would maintain relationships with many publishers, negotiate bulk rates, and place those ads. Over time, they developed ads for clients as well.

Newspapers had a few problems: they had little “air time” by only running once a day. And they weren’t very immersive. Radio and television ads changed both of those:

1. Radio could run 24/7, which meant ads had much more air time.

2. TV had much more immersive advertising, with video and images.

All three had major limitations, including the number of ads to be shown, ad engagement, and tracking the success of those ads.

The First Wave of Digital Advertising

On all of those metrics, digital advertising was incredible for the industry. Advertisers could now directly target interested customers, track the success of their ads, and create a positive feedback cycle of data, improving ads, driving more sales, and gathering better data.

Google, founded in 1998, would go on to capture the VAST majority of value created by search. In Ben’s framework below, Google was the perfect model for “conversion.”

A buyer came to them, looking for information, and Google auctioned off the ability for companies to “provide” that information.

By monopolizing search, aggregating intent, and auctioning that intent off to the highest bidder, Google created one of the best business models of all time. But while Google captured intent, internet ads weren’t particularly good at capturing attention.

The Second Wave of Digital Advertising

Facebook changed that, starting with its founding in 2004. Facebook created this new paradigm of the “attention economy.” Users didn’t come to Facebook with intent; it was Facebook’s job to plant seeds of intent via advertising.

Again, they auctioned off that attention and created one of the best business models of all time in doing so.

That said, Zuck did not like ads for a long time and wanted Facebook to be a platform company. In Facebook’s early years, they essentially outsourced advertising to Microsoft.

Despite the distaste for ads, in hindsight, it’s clear just how powerful Facebook’s ads were. Five years after founding, they passed $1B in revenue, more than doubling Y/Y.

Then the smartphone flipped all previous notions of advertising on its head; Facebook had to rearchitect its platform. Now, with technology completely immersed in our lives, it garnered even more attention. With that attention, Facebook’s ad business got even better.

However, there was a huge lag time between the introduction of the iPhone and the adoption of mobile advertising. Facebook wouldn’t fully commit to mobile for years, at which point they had to go convince advertisers of the value of mobile ads. The attention had shifted, but the monetization hadn’t.

At this point, we knew a few things:

1. The more data you have on consumers, the more targeted the advertising can be.

2. The more ads you have on your platform, the more targeted it can be.

3. The more attention you can capture, the more ads you can sell.

Both Google and Meta had been capturing huge amounts of consumer data, but hadn’t maximally profited from leveraging that data.

The Introduction of AI to Advertising

Right around the same time Facebook started pushing mobile ads, the AlexNet project was released, kicking off the modern era of deep learning.

Google and Meta essentially created a duopoly on top AI researchers in 2013, 2014, and 2015. They offered (1) pay, (2) flexibility for AI research, (3) open-source support, and (4) the resources you want to do your research.

They realized that they could profitably spend billions on AI researchers and hardware, because the returns on targeted advertising were massive.

The equation of data + GPUs + attention + AI researchers became one of the most profitable equations in the world.

Just the year before, though, another monumental shift in advertising was kicked off with the founding of Musical.ly (which would go on to be acquired by ByteDance and become TikTok).

This marked the turning point for the social media industry. From this point on, the industry would trend towards becoming just media, and not social media. It matched users with content they interacted with, and became essentially an AI company.

Somewhat incredibly, ByteDance revenue nearly passed that of Meta’s last year.

Amazon Enters the Fray

In terms of both AI and advertising, Google and Meta were creating near duopolies. The network flywheels, the talent, and the innovation seemed impossible.

Amazon entered the advertising business in 2012, and it really took off later in the 2010s.

It took off for a few reasons

1. First and foremost, it could influence buying decisions more than any other platform in the world. It was the closest to the buyer, and the closest to the bottom of the funnel. People go to Amazon to buy.

2. Secondly, it had a unique dataset and distribution network. Amazon’s e-commerce market share is 40%+ in the US. They’re not the only game in town, but they’re definitely the biggest game in town.

Put simply, Amazon provided advertisers the most direct way to influence the buying decisions of the most number of consumers.

And boy, has it worked. Amazon did $17.2B in advertising revenue last quarter, and that’s likely at a VERY profitable margin. Analysts estimate EBIT margin is greater than 50% for Amazon ads.

The last variable that made Amazon such an ad juggernaut was that it mostly avoided the incoming data privacy regulations.

Apple App Tracking Transparency & Cookie Deprecation

In July 2021, Apple released their App Tracking Transparency policy, essentially restricting iOS user data from being tracked by apps like Facebook. Pre-ATT, data flowed freely; post-ATT, data did not flow freely.

Shortly after, Google would announce the planned deprecation of cookies as well.

So this had a few effects.

First, it opened up ~some~ competition as the data moats weren’t quite as the aggregators’ data moats weren’t quite as strong anymore. So companies like Uber, Instacart, and DoorDash were more incentivized to offer advertising.

Second, it forced apps like Meta to start leveraging probabilistic models instead of deterministic models, a fun precursor to AI’s models today.

Third, though, Meta still needed to increase ad revenue (more ads or better ads). Which conveniently leads us to the right framework for thinking about digital advertising.

A Framework for Thinking about Digital Advertising

This brings us mostly to the current state of digital ads, dominated by three players: Google, Meta, and Amazon. With Tiktok as an emerging fourth player.

We’ve got an increasing trend of consumer aggregators (Uber, Instacart, DoorDash) leveraging ads for essentially free margin.

Finally, to be discussed more in the next article, we have infrastructure providers (like The Trade Desk and AppLovin) who match advertisers with ad slots across all platforms.

Now, summarizing everything, we have a fairly simple mental model for digital advertising.

1. The more attention you capture, the more ads you can place.

2. You can either increase attention by increasing the number of users (more apps or more users) or increasing engagement of the ads (Tiktok and Reels).

3. The more you can influence buying decisions, the more you can profit (Amazon > Google > Meta).

4. The better data you have, and the better AI you have, the better you can influence buying decisions.

We’re seeing the history of digital ads play out again: attention is shifting, wide-scale monetization has yet to follow, but a new generation of consumer giants is being built in front of our eyes. These companies have plenty of attention, incredible amounts of user data, and…pretty good AI.

Lest we forget Charlie’s wisdom, “Show me the incentive and I will show you the outcome."

The next couple of articles in this series will explore the combination of this mental model for advertising, applied to today’s AI companies, and see if the incentives can indeed show us the outcome for advertising in the age of AI.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

One of those important questions I’ve been thinking about is: How do established software companies’ advantages play out in the age of AI?

Given the rise of AI, agents, and Salesforce’s focus on both, Salesforce feels like a good case study to help answer that question.

To start, Salesforce’s AI strategy thus far has been excellent.

Leverage their data, distribution network, and brand; go all-in on marketing Agentforce.

At the same time, Salesforce’s software is notoriously complex. They’ve got more competition now than ever before.

So this will be a deep dive into Salesforce: its history, what made it successful, the state of the business, and the AI opportunity (or threat) today.

Ironically, the Salesforce story largely lays out the strategy to take on large incumbents, so we’ll start there (with a preview of the story below):

Without further ado.

The Story of Salesforce

As we go through Salesforce’s story, I want to highlight four central themes to keep in mind:

1. Getting the right idea at the right time is really, really important.

2. There’s no such thing as too much marketing, as long as it’s got a clear message.

3. The bigger the platform, the higher the switching costs.

4. Once you have the platform, expand it to as many adjacent markets as possible.

Much of this section came from Marc Benioff’s books on Salesforce.

The Founding Story (Lesson #1: The Right Idea at the Right Time)

Just like Phil Knight and Sam Altman, a sabbatical helped plant the seeds of Marc Benioff’s success. Marc had spent thirteen years working under Larry Ellison at Oracle, where he had risen up through the ranks. He decided to take several months off to think about his career.

At that time, Amazon.com was exploding, and Siebel Systems (on-prem CRM) had IPO’d. Marc was an investor in Siebel and knew the insider flaws with the product.

Combining the two ideas, he came up with the idea for an online CRM: “Amazon.com meets Siebel Systems.” Companies would no longer need to buy on-prem software, set up servers and networking, install the software, and have expensive contracts to fix the software when it went down.

With Salesforce’s launch in 1999, it became one of the first widely available cloud software products.

The Early Years (Lesson #2: There’s No Such Thing as Too Much Marketing)

Salesforce has always been an innovative sales and marketing company. This has been true from the very beginning.

Salesforce pioneered the cloud software model, but customers didn’t really get it. They didn’t see the point in it and didn’t trust its reliability or security.

From the beginning, Marc and Salesforce understood the importance of marketing: “Always ask yourself, "What's my message?" Position yourself either as the leader or against the leader in your industry.”

Salesforce did just that, launching their initial campaign of “The End of Software.” Companies would no longer need to buy software; they could now “rent” their software and remove the upfront cost and headaches of on-prem software.

Most of the Salesforce team hated the idea. But, in Marc’s words,”I felt that their arguments were overruled by the most important rule in marketing—the necessity to differentiate your brand.”

Famously, Salesforce hired fake protesters to set up shop outside of a Siebel Systems conference. They then hired journalists to go interview those protesters. Marc’s motto was “always, always go after Goliath.”

In 2003, Salesforce would host their first Dreamforce event as well, which would go on to become the most successful annual software conference in the world.

And this all caught on like wildfire. Once customers saw the value of not managing on-prem software, Salesforce took off. When people talk about PMF, this is what it looks like. By the end of 2003, Salesforce would do ~$100m in revenue, just four years after founding. Those are AI-like numbers!

Becoming a Platform (Lesson #3: In Software, the bigger the platform, the higher the switching costs.)

The bigger the platform, the higher the switching costs. The root of Salesforce’s (and enterprise software’s) moats are switching costs. Even if they don’t offer the best product, nobody else has been able to offer a better enough product to convince users to switch. This started with advice from Steve Jobs himself.

Planted by advice from Steve Jobs, Marc began focusing on building an ecosystem around Salesforce. Steve told Marc that he needed to build an application ecosystem for Salesforce. So, Marc came up with the name “The App Store”, but his team disapproved, so they changed the name to “AppExchange.” When Steve was ready to launch the App Store, Marc gifted the domain name back to him.

AppExchange would become one of the first early product expansions for Salesforce in 2005. Developers could now build applications on top of Salesforce, and it even had its own programming language.

For an idea of the legitimacy of this platform, Veeva ($37B public company) built its platform on Salesforce. They didn’t officially decide to move off of it until last year.

Inorganic Expansion (Lesson #4: Acquisitions should have clear strategic value leveraging the existing strengths of the acquirer.)

In 2008, Salesforce passed $1B in revenue, approximately a decade after being founded. At this point, they were becoming the market leader in the CRM market. The cloud is taking off. But, Salesforce needed to expand to continue this growth.

So they started a series of acquisitions to expand into adjacent markets.

Many of these acquisitions are good lessons on real acquisition synergy: leverage existing advantages (data and distribution network), expand into a new market, and upsell that new product with your advantages.

In 2006, Salesforce would make its first of mobile startup Sendia. This would start its tradition of entering a market via acquisitions. Up to today, it would do the same with:

• Heroku for PaaS services

• ExactTarget for Marketing Automation

• Demandware for E-commerce

• Tableau for data visualization

• Slack for enterprise messaging

Most of these clearly add a new vertical with new customer data or a new way to communicate with customers that fits naturally into Salesforce’s portfolio.

Those acquisitions mostly catch us up to speed and tell the root of Salesforce’s story.

Salesforce has continued to expand market share in their core markets, growing to a market cap of $230B and playing out the natural evolution of an enterprise software company: grow revenue, become more profitable, and create shareholder value!

Salesforce: A Platform of Platforms

I won’t spend much time on their technology, but I will provide a quick overview for the purposes of our discussion.

The basic idea everyone has of Salesforce is that it is a database with pre-organized dashboards and customizable business logic for sales teams. This is true and a testament to the clarity of branding mentioned earlier.

But actually, less than 25% of Salesforce’s revenue comes from their “sales cloud.”

So the right way to think about them is a database for all of a company’s customer-facing organizations: sales, customer service, marketing, and eCommerce.

• Salesforce sales cloud: account data

• Customer service cloud: incident management data

• Marketing cloud: marketing data

• Commerce Cloud: e-commerce data

Then, there’s a platform layer, made up of data services and an application builder. Data includes a warehouse-like service for unifying data storage, Mulesoft for integrating all of those various data sources together, and Tableau for visualizing data from those sources. Slack is also in there, but it doesn’t really fit so cleanly in the portfolio.

Agentforce

Finally, Salesforce’s current area of focus is Agentforce, their AI agent offering. They offer agents with pre-defined workflows for customer service, sales, and marketing, as well as an agent builder product.

As Salesforce’s head of AI describes:

Agentforce basically has pre-built automations in the form of AI agents for their various workloads. So Agentforce and what we built is a mixture of effectively out-of-the-box agents that mirror kind of our lines of business or clouds, right? So service, sales, commerce, marketing, et cetera. And so we think about having agents doing all kinds of like long tail type background work. Really unstructured data is -- I don't want to say gold mine, but it's something now that we can do so much more with all of the data, transcripts.

The most popular use cases today are in customer service and sales:

So I think service is maybe the most obvious as you're pointing out, kind of almost like it's low hanging fruit, but it comes up a lot. Another really kind of big interesting use case is really about more on the employee side, like internally facing agents that are doing basically pre and post like meeting prep. We have an SDR product, sales development rep product that takes leads. Your sales plan does research on the leads and it does outbound e-mails to them. So we have 5,000 Agentforce deals in Q4. About 3,000 of those were paid.

Bringing it all together:

CRM: A Market of Giants

One other important point to highlight is the top-heavy nature of the market. Salesforce is undoubtedly the CRM leader. But, second, third, fourth, and fifth place are all massive, incumbent enterprise software companies.

Salesforce’s biggest competitors are Microsoft, Oracle, SAP, and Adobe.

Okay, forgive me for skipping over some of the details here. But as I said, Salesforce is the canonical example of an enterprise software company. Low churn, consistently decreasing revenue growth (as expected), and increasing profitability:

CRM epitomizes big, mission-critical software that nobody wants to switch off of.

Given that this article is asking the question of how incumbent advantages play out in AI, the switching costs of CRM provide the perfect case study on it.

What’s the Outlook for Salesforce?

The bull case for Salesforce is that Salesforce’s switching costs, data advantages, and distribution network are strong enough that Salesforce adds another multi-billion dollar product line and creates an additional layer of value with their AI offerings.

But, the more interesting question is, what if that doesn’t happen?

I think the greatest risk to Salesforce is not head-head competition but a shifting of value in the value chain. In this scenario, value shifts from the “CRUD databases with business logic” in Satya’s words to new data storage and new workflows.

The current model of CRM is simple: store customer data and have applications (sales, customer service, marketing, eCommerce) built on top of that.

A very interesting data point came up from Packy McCormick’s deep dive on Rox: data has been shifting from the CRM to the data warehouse over the last five years. Up to 40% of the data stored in data warehouses is customer data. What that means is that the CRM database may be losing some of its position in the value chain.

If this is true, then the value from data storage may shift to lakes & warehouses. With the rise of agents, application value may shift to agentic workflows:

Despite this being several years out (if it happens at all), I think it’s the threat that Salesforce is most concerned about (and I think it’s informing their data strategy with Zero Copy Partner Network and Agentforce).

This is a conversation not just about Salesforce, but about the future of enterprise software. Software is meant to automate workflows and make us more productive. If agentic workflows can do that more efficiently, then value is likely to shift there.

With that shift, the new value chain may look much different from the old one.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

I had also seen a statistic that Waymo now has a 25% market share of ridesharing in San Francisco. Naturally, it made me think about Waymo’s effect on Uber.

That leads to a more interesting question: What’s the future of aggregators?

Many of the largest companies founded in the last 30 years have been aggregators. They aggregated information as supply, aggregated attention as demand, and provided a service to match the two. What happens to the aggregators in a world where attention is shifting to AI?

To answer that, we need to understand what went so right over the last two decades and what led to the success of those companies. To start, we’ll focus on a specific example of an aggregator model: the marketplace.

Now’s a particularly interesting time to study marketplaces, we’ve got:

1. Questions on demand: Who owns attention in an agentic world? (i.e. ChatGPT booking hotels)

2. Question on supply: What happens if supply can be automated? (i.e. Waymo vs Uber)

3. Questions on economics: Will marketplaces achieve the long-expected profits of market dominance?

This article will act as a primer on marketplaces and a framework for thinking about their future, covering:

1. The History of Marketplaces

2. The Economics & Business Models of Marketplaces

3. The “Late-Stage Economics” of Marketplaces like Uber

4. The Potential Future of Marketplaces

Without further ado.

1. The History of Marketplaces & Timing Technology Waves

Physical marketplaces developed thousands of years ago, as a means of trading resources. This started with agriculture (i.e. I’ll trade you a pound of grain for a pound of meat). Over time, marketplaces grew larger, increasing the number of buyers and sellers.

This led to an important observation: The larger the market, the more specialized the labor could be. This means the bigger the market, the better. Until it hits operational limits like space, safety, and geographical constraints.

Adam Smith would go on to observe this via the Division of Labour. He observed that when people specialize their skills and trade with others, the productivity of the society goes up!

The First Digital Marketplaces

The internet unlocked the first wave of digital marketplaces. It networked the world’s information, and unlocked the ability to aggregate that information.

As we dive into digital marketplaces, we should keep one idea in mind: business innovation is a mirror of the underlying technology that unlocked it.

In 1995, Pierre Omidyar launched AuctionWeb, which would go on to become eBay, a site "dedicated to bringing together buyers and sellers in an honest and open marketplace."

In just two years, the site would sell $500M…of Beanie Babies alone. Over the next decade, the marketplace would sell cars, jets, and even yachts. They’d acquire PayPal, Skype, and StubHub. eBay epitomized what internet marketplaces could become.

By 2000, Amazon would launch its 3rd party Marketplace as well.

The other large category of companies that rose was OTAs, or online travel agencies. Expedia was founded in 1996. Priceline was founded in 1997, on the back of a simple idea: “Airlines regularly flew only two-thirds full, with millions of seats empty. A million hotel rooms might go unused every night. We thought, what if we harnessed the Internet to drive demand, filling those planes and hotels?”

This was the beauty of marketplaces: they took unused supply and matched it with demand. As Bill Gurley put it, creating “money out of nowhere.”

Travel marketplaces also consolidated to just a few players over the decades since their founding.

If we recall earlier: the bigger the marketplace the better, up to certain physical constraints. The internet didn’t have those constraints. Which means that it lends itself towards winner-take-most market share.

The Second Wave of Digital Marketplaces

The next wave came in tandem with a wave of disruption (the cloud) and a wave of distribution (mobile).

It’s important to point out the importance of timing in these businesses. As we laid out earlier, marketplaces tend to be winner-take-most markets. The winner is the marketplace that turns the flywheel of supply and demand the fastest. Being early almost always helps in technology, but it especially matters in marketplaces.

Over a few years, we saw the following subsequent innovations:

1. Google releases tools to integrate Google Maps into websites and applications in 2005.

2. Amazon debuts AWS in 2006.

3. Apple releases the iPhone in 2007 and announces it would run mobile applications.

If we’re observant at this time, we’d see there’s a new way to build applications, a new way to distribute them, and new functionality we can build out within them.

The first of our main characters was founded in 2008, with Airbedandbreakfast. For a bit of cloud lore, Jeff Bezos gave a talk to a group of Y Combinator founders in April of 2008 about how easy it is to build on AWS. It’s cheap, it’s easy, and you can iterate faster than ever before. That talk convinced the Airbnb founders to build on AWS.

Similar to Priceline a decade before, Airbnb connected unused supply and created demand for that supply. Prior to Airbnb, vacation rentals were a small market and a black box. It was risky to stay in someone else’s home! Airbnb aggregated a fragmented market, brought trust to the booking process, and created value for all sides involved.

In 2009, Uber was founded, putting instant transportation in the hands of everyone. Over time, they’d expand the “taxi market” several times over. DoorDash, the third largest marketplace of the cloud era, was founded in 2013.

To emphasize the importance of timing, the Acquired team described of DoorDash:

“DoorDash had one of the best "why now" answers of all-time: mass-market smartphone adoption (not just high-end) made three things true that were never possible before: 1. average consumers could order online conveniently, 2. couriers could plug into the network via their own devices, and 3. restaurants (most of whom didn't have wifi or desk staff) could accept orders online. Before 2013 a business like this would simply have been impossible to build.”

The promise of these marketplaces was that they’d compound via the network effects shown below:

Whoever got the marketplace flywheel turning fastest was going to win the market. The easiest way to turn the flywheel faster? Pour money into the business.

Airbnb raised over $6B as a private company. Uber raised over $20B. Lyft raised ~$5B. DoorDash ~$3B. This led to a battle of attrition. Whoever could subsidize their business long enough would eventually be entitled to the majority of the market’s profits.

Summarizing the key lessons of the marketplace:

1. The larger the market, the more specialized the labor could be. This means the bigger the market, the better.

2. With a lack of constraints in the digital world, marketplaces trend towards winner-take-most market share.

3. Whoever got the marketplace flywheel turning fastest was going to win the market. The easiest way to turn the flywheel faster? Pour money into the business.

2. The Economics of Marketplaces

The basic economics of marketplaces are very simple, so I won’t spend much time on them: connect fragmented buyers to fragmented sellers, take a cut off the top.

The criticism of marketplaces is that they’re commodity businesses with cost as the determining factor for victory. So the question is:

1. Why are they good businesses in the first place?

2. What variables make some marketplaces more attractive than others?

3. If a marketplace is a good fit for a problem set, how do they scale?

Why are marketplaces attractive business models?

To that, I’ll let Nick Sleep explain.

In Nick Sleep’s 2004 annual letter, he described what would make the best business in the world. He concludes it’s a firm that would have a huge marketplace (offering size), high barriers to entry (offering longevity), and very low levels of capital employed (offering free cash flow).

For those that know Nick Sleep’s story, Amazon would go on to become a huge % of their holdings. So he’s got to be describing Amazon here, right? Wrong! He’s describing eBay.

He goes on to say: In our opinion a business such as eBay could be the most valuable in the world. It has a huge marketplace, the biggest, an auction marketplace naturally aggregates to one player, offering high market share and high barriers to entry to the winner. Better still eBay makes the customer pay for a high proportion of the assets used in the transaction such as PCs, modems, phone lines and so forth. But best of all, the incremental assets required to grow are so small.

Some of the largest of those marketplaces can be seen here:

So marketplaces are attractive with sufficient moats, how do you build those moats?

A marketplace’s revenue is the total value that flows through an ecosystem * the take rate of the marketplace.

The “Net Revenue” divided by the “Gross Revenue” is the take rate. The long-term economics of marketplaces are determined by what % of the transaction the marketplace can take.

The take rate is ultimately determined by the competitive advantages of the marketplace. This is determined by a few variables:

Concentration of suppliers and buyers on the marketplace:

The more fragmentation, the less negotiating power from individual customers or suppliers. The marketplace’s ability to drive incremental revenue for suppliers plays into this as well: the more fragmented the industry, the more likely the marketplace drives meaningful revenue for its suppliers.

The higher the transaction amount, the lower the take rate:

Suppliers are willing to accept less of a fee for higher-priced items. For example, it’s one reason why real estate transactions don’t go through credit cards. Sellers aren’t willing to pay 3% of their transaction to the payments value chain.

Reasonable alternatives:

The less competition, the more pricing power. This competition comes in the form of both alternative marketplaces and alternatives to using the marketplace.

Value created in value-added services:

Airbnb expanded its market by adding trust into its process. Once they established that trust, no other competitor could enter the market; it’s too risky for customers. Other services include payments, quality control and authentication, insurance, and fraud prevention. All of these services are hard to replicate and build in switching costs to marketplaces.

Given that a marketplace is a good fit for the problem, how do they scale?

In large part, the competitive advantages of marketplaces are driven by scale. So how do companies build scale?

Lenny Rachitsky did some excellent work on how marketplaces scale several years back. It comes down to solving the chicken-and-egg problem. I.e., you need supply to bring in customers, but you need customers to incentivize suppliers to come to the platform.

This typically starts with constraining the marketplace to a subset of users and scaling up supply in that marketplace, like Uber starting in San Francisco.

Then it’s a constant game of scaling supply and scaling demand.

There’s no one cheat code to making this work; it’s just a constant grind for years and years until the flywheel is sustainably turning faster than competitors. If there is a cheat code, it's in raising far more capital than competitors.

3. The Late-Stage Economics of Marketplaces

At this point, we’ve laid out the economics of marketplaces, when they are a good fit, and how to scale them. All of this capital flowing into marketplaces was under one assumption: at a certain point, they would generate profits from their consolidation.

Uber and Lyft give us the model of what this looks like. It starts with the compounding network effects:

This is despite Lyft’s higher take rate (note Uber Eats has a much lower take rate, which biases this down, but Lyft’s take rate last quarter was 37% compared to Uber’s 30.3% take rate on mobility).

As market share increases, so does pricing power:

Finally, valuation reflects that:

Now, this does not mean all marketplaces will follow this path. Actually, very few have. This brings us to our final section: what does the future of these marketplaces look like?

4. What Happens to the Aggregators?

The goal of this article was to provide a primer on marketplaces (above) and a framework for thinking about their future (below).

I’ll also point out that I’m speaking in general terms, but the answer to these questions ultimately comes down to the unique supply and demand of each marketplace. Of which, I’ll leave to you all to have an opinion.

Outside of the widespread adoption of decentralized software, there are three potential outcomes for today’s large marketplaces: continued compounding, trending towards commoditization, and being out-competed by closed-loop systems.

Outcome #1: Inertia & Continued Compounding

This path is straightforward: the moats of these network-driven marketplaces lead to the predicted profitability of marketplaces. These companies will grow until they saturate the market; at which point, they’ll grow at the pace of the market. They’ll continue to become more profitable over time.

Airbnb’s brand, for example, is likely so strong that it will prevent competitors from encroaching. Amazon’s distribution network makes it brutal to encroach on their marketplace.

In these situations, the marketplace has economies of scale and pricing power. Their economics should continue to improve, just as has been predicted over the last decade.

Outcome #2: Commoditized Marketplaces

This has been the critique of the marketplace business model. If a company is primarily a demand aggregator without unique supply, then it becomes much harder to protect profits.

This highlights how hard it is to have a unique supply like Airbnb. However, all is not lost.

Take Booking and Expedia, for example, both arguably without unique supply. They are two of Google’s biggest clients, spending 31% and 50% of revenue on advertising last year. And yet, both are quite profitable businesses.

Outcome #3: Competition with Closed Loop Systems

Now, we get to the interesting part of the conversation! What if the supply of the marketplace can be automated? Closed loop systems are those where a company provides the supply directly to customers.

The most interesting example is the debate between Uber and Waymo.

For homogenous marketplaces, consumers pay for one thing: a consistent customer experience at the lowest price possible. If you’ve ridden in a Waymo, you know the experience is excellent. It’s no coincidence that Waymo has risen to a 25% market share in San Francisco:

Does that mean Uber is done for? I’d argue not. The beauty of marketplaces is their scalability. The beauty of closed-loop systems is their control over user experience.

These are fundamentally at odds with each other. Should Waymo build to peak or average load?

Waymo can’t build to peak, just like Taxis couldn’t build to peak! There were only 13,000 taxi cabs in NYC for a long time, compared to 100,000 rideshare vehicles in NYC.

Research suggests the optimal model may be a hybrid one like Uber and Waymo are experimenting with:

We could see a core base of closed-loop providers with a flexible marketplace network on top. This combines the benefits of top-tier user experience most of the time, with the extensibility of marketplaces on top of that.

Now, in this world, the open question is about economics. Where does value accrue? Who owns the most defensible piece of the stack, and is their business model set up to collect profits off the value chain?

I liken this to the Google vs ChatGPT debate. I’m not here to say whether one model will be dominant or the other, but it certainly doesn’t make Google’s business model better.

Finally, there’s one important question we need to ask: Who owns the attention?

Ben Thompson wrote an article back in 2015 called Airbnb and the Internet Revolution, which described how Airbnb and the sharing economy have commoditized trust, enabling a new business model based on aggregating resources and managing the customer relationship.

But what happens to that business model if you no longer manage the customer relationship?

Let’s use this commonly discussed example of ChatGPT booking you a hotel room based on your preferences.

In all likelihood, this would require a native integration between ChatGPT and Expedia or Booking.com. You ask ChatGPT to book you a hotel, it navigates Expedia for you and books the hotel. Over time, your preferences get stored in ChatGPT; you go to the OTA less.

How does that affect the business model of the OTA? Again, hard to say, but it certainly doesn’t make it better.

And by the way, I didn’t conjure this up in my head; this section was inspired by an executive conversation between DoorDash and OpenAI.

This is one of the most important second-order effects of AI, and one I don’t have a good answer to today. But hey, a mediocre answer to a great question is more important than a great answer to a mediocre question.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

Somehow, the team at Databricks got a hold of it, and Ali Ghodsi shared it on LinkedIn. The internet’s a funny place. (Ali/Ali’s team, do let me know if you want to come on for an interview.)

Now, the opportunity to write about Databricks has come up again. The last three articles of mine have formed a semi-intentional series on the lifecycle of industries. As WB says,”The boat you’re in matters much more than how hard you’re rowing.”

As a bookend on this series, I wanted to write a deep dive on one of my favorite companies that exemplifies riding the SaaS -> AI S curve. Databricks is as good of an example as there is.

Let’s start at the beginning.

The History of Databricks

I can’t recommend this incredible article about Databricks founding story enough, much of this section is inspired by it.

The Netflix Prize

Sometimes stories write themselves, and Databricks has this incredible founding story that does so.

So in 2006, Netflix announces a $1M competition called the “The Netflix Prize” for anyone who improves their recommendation algorithm by 10%.

For some context at this time, the rise of the internet led to the proliferation of big data and unstructured data. Companies were exploring how to best leverage this data. To do so, companies needed to store data, write algorithms, and run those algorithms on that data (which requires huge amounts of computing power). The predominant technology at the time was Hadoop, released in 2006. Hadoop, however, was too inefficient and difficult to use to be effective in this new age of big data:

So, Netflix needs a more efficient way to run this process. By 2009, 50k participants had entered the contest, but no one had quite met the 10% improvement requirement. The dataset included all of these movies (18k+), all of these ratings (100M+), and all of this user data (450k users). Millions

A Berkeley Ph.D. student named Lester Mackey had decided to compete in the contest. He faced a clear problem: there was no tool to efficiently run algorithms on that large of a data set.

Matei Zaharia, a classmate (and now a legend in data), built a tool to more efficiently process algorithms. That tool would go on to become Spark.

Mackey uses this tool from Zaharia to iterate over the dataset and get the 10% improvement Netflix had been hunting for. Unfortunately, their submission was 20 minutes past the deadline, and a team out of AT&T won the prize (feels timely to add in some old adage about the turtle and the hare).

The Rise of Spark and the Founding of Databricks

At this point, Zahari had seen issues with big data processing while at Yahoo and knew the shortcomings of Hadoop as an early contributor. He decides to team up with fellow academics at the Berkeley AMPLab to build out Spark and commercialize it via Databricks.

The early vision for Databricks is articulated remarkably similar to today: one unified platform for analytics with the goal of enabling artifical intelligence.

Let’s take a step back: why does Databricks need to exist in the first place? That comes down to two questions:

1. Why does Spark need to exist?

Processing big data is really hard at this time and it’s going to keep getting harder. If companies want to be able to glean information from large amounts of data, a tool like Spark needs to exist. Tailwind #1: Big Data.

2. Why does Databricks need to exist?

Secondly, the Databricks team knew managing big data was only going to get harder. The cloud fundamentally unlocked the ability to scale compute and storage separately, AND HAVE VIRTUALLY UNLIMITED AMOUNTS OF BOTH. Open-source can be really hard to manage, so Databricks offers the power of Spark and the ease of use of managed cloud software. Tailwind #2: The Cloud.

Databricks even turned down a $20M offer from a financial services firm to offer Databricks on-prem…when it had only $10M in revenue.

They articulate Databricks as a tool to “make simple things easy” and “make complex things possible.” Put simply, Spark allows large compute workloads to be run faster and more cost-effectively than Hadoop. A Spark driver takes the program written by the user and distributes the tasks that need to be executed. The Cluster Manager allocates the necessary resources to accomplish the task, and the executors process the tasks accordingly.

Because machine learning workloads are so compute-intensive, Databricks’ initial customer base was data scientists and machine learning engineers, typically at large companies. Contrast this to Snowflake, whose workloads were much more focused on business intelligence.

It’s important to call out that Spark is a processing framework, not a database. So, it’s a way to distribute compute workloads rather than store data. It sits on top of raw object storage like S3, typically running workloads on structured and semi-structured data. This is a fundamentally different value proposition than a tool like Snowflake, which started out as a data warehouse designed for structured data.

Sparks fly…eventually so does Databricks

From 2013-2015, the team continued to build out the Spark project, looking for the proverbial “product market fit.”

In 2015, in Ghodsi’s words, “Spark exploded…at that point we felt like we had succeeded, but we just didn't have any monetization. So we went through a difficult time in 15 and we had to figure out, okay, if we can't monetize, maybe there is nothing here. Uh, and the company kind of went through a little bit of introspection.”

Ali Ghodsi then steps in as CEO in 2016, and the leadership team changed, as Databricks brought in a team that had experience selling enterprise software.

Most of the original founding team stayed at the company (and is still involved today), but they led technology while the new team focused on commercializing that technology.

Part of this change required making Databricks a little bit less open-source. Ghodsi said, “So we need proprietary secret sauce, sauce that we're not just giving away to the hyperscalers. So that was critical also in changing, and that's difficult to do…but it was the only way to succeed. We wouldn't have been around if we hadn't done that pivot.”

Tailwind #3: The Rise of Open Source

The monetization of open-source tech remains a problem for startups today, and one without a one-size-fits-all solution. But Databricks decides to start keeping more of Spark’s features proprietary so they have a value proposition to sell to customers.

How does Databricks make money? Since Databricks doesn’t store data, they made their revenue from charging a hefty margin on their compute costs. That was built on the cloud infrastructure providers, so Databricks takes an additional margin on their costs. Because of the nature of big data, Databricks sold to very large enterprises or startups with large amounts of data.

From this point on, the business really takes off:

In 2017, Databricks makes a really smart strategic move with a partnership with Microsoft Azure. This was a jointly developed product where Microsoft customers could purchase Azure Databricks, integrate with other Azure services, and both Databricks and Microsoft recognize revenue. They leverage the distribution network and sales teams of Microsoft while establishing their own technology within Microsoft’s customer base.

Perhaps lessons from this helped inform Microsoft’s OpenAI partnership.

Databricks goes Mainstream

Over the next five years, Databricks would raise five rounds of venture funding and take them from a $520M valuation to a $28B valuation:

1. Series D of $140M in 2018

2. Series E of $250M in 2019

3. Series F of $400M in 2019

4. Series G of $1B in 2021

5. Series H of $1.6B in 2021, valuing them at $28B

Over that time, they’d pursue a key piece of their strategy: becoming a platform. They’d release two more pillars of the open-source data/AI community: Delta Lake and MLflow.

MLflow provided an open-source framework to build out ML models. Delta Lake provided an “open table” format which provided a way to store metadata about data in data lakes, which allowed more efficient organization and processing of that data.

Perhaps, most importantly, this laid the groundwork for Databricks “data lakehouse” vision, which brings us full circle to Databricks “Big Picture” in 2013: a unified platform for analytics.

Over the 2010s, data management had diverged in two paths: data warehouses for SQL-based analytics and data lakes for storing bulk, unstructured data. Both had problems:

1. Data warehouses were expensive, closed-source, and didn’t handle unstructured data well.

2. Data lakes were difficult to manage, leading to data disorganization.

Databricks went full steam ahead on the data lakehouse vision to finally bring their vision to life. The lakehouse merged the data lake and the data warehouse, bringing three primary benefits: centralizing data management onto one platform, opening the data platforms to open-source tools, and doing so at a theoretically cheaper price. (I wrote a full article the lakehouse last year.)

Databricks did a wonderful job marketing the lakehouse, bringing us back to the key point: one platform for all your data, and easily get insights from that data.

It also opened the door for Databricks to release Databricks SQL, a competitive data warehouse to Snowflake that now makes up 20% of their revenue. If you’re going to store all your data on one platform, you need to store both your structured and unstructured data!

The Current State of Databricks

Databricks had spent a decade building its reputation in the AI/ML community. By the time the ChatGPT moment hit, they were wonderfully positioned.

They’ve made acquisitions and product releases over the last two years to enable the entire data and AI workflow. This includes ingestion, orchestration, transformation, processing, SQL database, and a host of AI tools (agent frameworks, agent evaluation, vector search, AI governance, model fine-tuning services, and managed MLflow).

I want to take a step back because the challenge with understanding platforms is the thousands of features they offer. Databricks can be broken down simply: they aim to be the unifying layer for data and enable companies to get insights from that data.

1. Data science: Data science jobs, running on managed Spark in the backend, writing code in Databricks notebooks.

2. Data warehousing: Store all your structured data and query it.

3. Data engineering: Pipelines, orchestration, transformation.

4. Build your own AI: Traditional ML (MLflow, feature stores, model serving, training models, fine-tuning models) and Agents (Agent frameworks, vector stores)

5. Governance & security: Index all of your data assets (know where they are, what’s in it, and who should have access to it)

6. AI on top of that: I speculate Databricks envisions a world where you have one application for your data, one agent if you will. You ask that data agent any question you want and it can answer it for you. That’s been the vision from the beginning: don’t ask what happened last week, ask what’s going to happen next week.

The Business Today

This brings us to the current state of the business today. In December of 2024, Databricks raised a $10B round valuing the company at $62B. With that announcement, we got a look at their financials as well:

To be clear, the moats in enterprise data are exceptional. Databricks will be around a decade from now and they’ll be important a decade from now.

From an investment perspective, the uncertainty comes from valuations and growth expectations. Databricks had a $3B annual revenue run rate, up 60% from ~$1.9B run rate at the end of 2023. That values them at a Price/Run Rate of ~21, and a price/run rate/growth of ~.35. If we compare that to comps of other data companies, that seems like a reasonable price to pay:

What’s particularly interesting is that while the rest of the industry’s growth slows, Databricks’ growth is accelerating. They accelerated growth by 10% Y/Y, while the other three companies shown declined ~7% Y/Y on average. There may be something to this whole AI business after all!

At a macro level, the questions around Databrick's valuation come down to industry growth as a whole, how long Databricks can sustain its outperformance, and wide-scale multiple contraction or expansion in data companies.

Now, I alluded to the big question of this article at the beginning: how do you effectively ride S curves?

Databricks & The Path Ahead

At an industry-wide level, we’re simultaneously seeing the consolidation of software and the rise of AI. I suspect how Databricks navigates these trends will determine their next decade.

As I see it, this leads to three key elements of their strategy:

1. Offering a Platform in Times of Consolidation

After a decade of rapid growth, enterprise SaaS is consolidating. As Redpoint put it, “Market Leaders Have Continued to Benefit Disproportionately.”

What that means is that you want to be as broad of a platform as possible for your specific customer base. And that’s exactly what Databricks has done: releasing more and more features to support the entire data ecosystem.

Now, this is powerful alone, but it’s even more powerful with open-source. The power isn’t just being a platform, it’s in being an open-source distribution network.

2. Becoming the Open-Source Data Distribution Network

From the beginning, Databricks has been an open-source-first company. Initially, this was the value prop: open-source is wonderful but it’s hard to manage. Let us handle the hard stuff.

Now that open-source has exploded, the value prop is this: let us handle all of your open-source, we’ll offer a managed version and you just build great products.

This is so brilliant because the better open-source does, the better Databricks does. This creates an open-source flywheel: The more open-source products you use on Databricks, the harder it is to switch.

This was the same dynamic that Databricks had to fight off from the hyperscalers in the early years!

Finally, these first two strategies are positioning for industry consolidation. The last point is positioning for industry growth.

3. Becoming the AI Database

Ghodsi described the future of Databricks, ”Basically Databricks [is] an AI database. We don't call it that. We don't market it that way, but essentially it's an AI database under the hood. And you will need that for every application, every company in the planet will need that. And, you know, I hope that we can be the company that powers all of these applications in the future.”

To succeed in this, Databricks needs to do two things: convince startups to build on their platform and build as many native integrations as possible with AI applications.

The first allows Databricks to profit from the coming wave of AI applications, and the second protects Databricks core data business.

Like I said prior, Databricks will be around a decade from now and they’ll be important a decade from now.

The question comes down to how much bigger they’ll be. The answer to that comes down to execution.

Two years ago, I speculated Databricks wanted to be the biggest tech IPO ever. That sight is still within reach. Meta IPO’d at a market cap of $81B in 2012 according to Pitchbook. Could Databricks IPO at a $100B market cap in the coming years?

There are a lot of questions between now and then, but I don’t see why not.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.

It brings up a more important point altogether: how should we think about investing throughout the lifecycle of an industry? How should we think about investing in companies in new industries vs old ones? For a fun example of the game we play: if OpenAI were to IPO at the $300B reported valuation of their next round, they’d be worth just less than Bank of America (roots trace back to 1784!) and right above Wells Fargo (founded 1852!).

Summarized, the question I’m asking is: What’s a good theory on the lifecycle of industries? As Charlie said:

What is elementary, worldly wisdom? Well, the first rule is that you can’t really know anything if you just remember isolated facts and try and bang ’em back. If the facts don’t hang together on a latticework of theory, you don’t have them in a usable form.

Consider this my humble attempt at a theory on the lifecycle of industries.

I think it’s particularly relevant today as we’re simultaneously witnessing the rapid growth of AI and the slow descent of enterprise software growth rates:

My goal of this article is to answer a few questions:

1. What is a good general theory on the evolution of industries?

2. How do companies differentiate and succeed in the early stages of an industry lifecycle?

3. Who benefits as industries mature and start to consolidate?

A spoiler: it’s a lot like my middle school basketball coaches telling me how to defend big guys: Do your work early.

A Brief Intro to the Lifecycle of Industries

One theory I particularly like on the evolution of industries, which comes from Steven Klepper and Michael Gort, lays out five stages of the industry lifecycle:

1. Birth - Introduction of a new product by its first producer.

2. Expansion - Sharp increase in the number of producers.

3. Peak - Number of entrants approximately equals number of exits.

4. Consolidation - Negative net entries into the industry.

5. Equilibrium - Small number of producers continue until an industry dies or is reinvented.

Those early days of an industry is where I’ll start.

Industry Birth = Problems to Be Solved

Michael Mauboussin, in “New Business Boom and Bust,” laid out a wonderful analogy comparing the development of the human brain to the development of new industries.

As a child’s brain develops, it creates far more neurons than are actually necessary, laying the pathways for all the potential neural pathways that may be used. As the brain determines what pathways will actually be used, it “prunes” the unused pathways:

As new industries develop, products flood the market, providing many potential solutions to customer problems.

The “market,” or the aggregate demand from all customers, then selects the products or services that best solve their problems.

For a new industry like AI, we’re just starting to see the market select its preferred solutions.

Now, the question becomes, how do you differentiate in these early cycles?

Differentiation in the Early Stages of Industry Lifecycles

The earlier you compound feedback cycles, the faster you compound. This is why first-mover advantage matters so much.

The earliest producers get to pick the lowest-hanging fruit, the easiest problems to solve. They get the “king-of-the-hill” position on all those customers, which puts the burden on competitors to provide enough performance improvement to convince customers to switch.

It also creates positive feedback loops: get data on customer needs, develop products for customers, build relationships, and get more data on customer needs. At the same time, a company’s reputation compounds, revenue grows, and companies can raise more capital, which further allows companies to solve customer problems.

This is particularly important in network effect businesses. It’s not a guarantee you’ll win the market; you might take the wrong strategy and get beaten by a superior approach later on (i.e. Google showing up late to the search party). But you’d rather be in pole position.

How else can companies differentiate?

1. Verticalization & Domain Expertise: The better a company understands the unique problems of an industry, the better they can solve those problems.

2. Distribution Networks: Companies with existing distribution networks can release products and upsell them to existing customers rapidly (also adding to the speed-to-market point).

3. Data Advantages: Proprietary data leads to proprietary insights, whether accessed through partnerships, integrations, or experiments.

4. The X factor of finding ways to solve customers problems: At an event last year, I heard David Senra tell a story about Fastenal, who sells an entirely commoditized product: quite literally nuts and bolts. Somehow, they had 4x the margins as competitors and were still gaining market share. They did that by having the best customer service on the planet. Their founders were obsessed with customer experience; they went out of their way to show customers that, and they dominated a market by doing it.

And a reminder from Jensen: "My will to survive exceeds almost everybody else's will to kill me."

Depending on these variables, the market starts to “select winners” and those winners start their Power Law compounding. The rate of competitor entry slows, until eventually hitting its peak.

Industry “Peak” & Maturation

The peak of an industry lifecycle is defined as the point at which the number of companies exiting (acquisition, bankruptcy, etc.) equals the number of companies being founded.

This is particularly hard to measure for software-based industries because the capital requirements to start and run a software business are so low. This leads to the constant influx of companies and few exits of companies.

So rather than definitively saying “SaaS is consolidating,” I’ll say that it’s clear the growth of the industry is slowing.

One of the side effects of this growth decline is the rise of “zombies” in private markets: companies that aren’t large enough to draw interest from investors but aren’t profitable enough to draw interest of acquirers.

As described by Jamin Ball, “A zombie is a company that (in it’s current form) is not attractive to public markets or acquirers. There are so many companies sitting in the $50m - $300m ARR range, growing <20%, and burning money or barely breakeven.”

For context, see the total unicorns created and their ensuing “down rounds” (raising money at lower valuations) below:

Who benefits when industries consolidate?

Oftentimes, industry consolidation creates something like a Golden Age for industry leaders. The number of competitors dwindles as the industry grows.

The perfect scenario is a company that has strong competitive advantages, barriers to entry, a growing industry, and competitors that are “donating share.”

Now, at this point, your hackles might be raised, and you say, “Eric, there are no free lunches in investing; when have we seen that happen?”

Well, look, it is rare! But this is the story of many of the largest tech companies today:

1. Spot a company benefitting from the Power Law in an industry with winner-take-most dynamics.

2. Spot an industry that’s going to grow larger and for longer than everyone thinks.

An early monopoly on a rapidly growing industry creates a recipe for winners in technology.

Let’s take the semiconductor industry for example:

In the semiconductor industry, there have only been a few public semiconductor companies founded since 2000. A few others have been spun off or created out of mergers. But the vast majority of companies were founded between 1970 and 2000:

Yet the industry has nearly 3xed revenue in that period, with almost no new companies, from $222B in 2000 to $626B in 2024.

Now, the barriers to entry in semiconductors are massive which has made this more pronounced than most industries. But the point stands.

While it’s common in technology, not all industries trend towards consolidation. Even if they do, that doesn’t mean profits accrue there. Take airlines for example; the industry is relatively consolidated but those businesses are infamously challenging to run.

One final point: “broken economics” may start to sort themselves out, such as in railroads or even recently in ride-sharing. Natural monopolies lend themselves towards these outcomes, requiring huge upfront capital expenditures on the front end but monopolistic profits on the backend.

See Uber’s economics over the last five years here:

Then, as the industry stagnates, competition decreases, innovation slows, and more problems arise that need to be solved. And so goes the industry lifecycle.

Why this time is (a little bit) different:

If you’re looking for me to predict the next 12 months in AI or “where we’re at in the cycle,” I'm sorry to disappoint!

As I said in my annual letter, “I’m afraid I don’t have the foresight or the inclination to predict what will happen in technology or markets over the next 12 months (and I can’t imagine anyone else wants that!).”

What I will say is this: one fascinating observation is that industry lifecycles seem to be compressing, meaning industries rise and consolidate faster than before.

In software-dominated fields, this is only reasonable. First, software products iterate much faster. Second, they’re iterating faster and faster via the advancements in AI coding tools. Third, product market fit can be found faster because of the faster dispersion of information. Social media allows good companies to find their customers and hear their feedback faster than ever!

Additionally, and specifically to AI, enterprises are clamoring to adopt AI. Every organization I talk to is interested in adopting AI.

Combined, we have faster product development times (more neurons created to explore potential paths of “product-market fit”) + faster positive feedback loops (neurons are used earlier and paths are hardened faster) + more eager customers = compressing of early industry lifecycles.

The implication of that is this: Positive feedback loops are faster. Power Law winners emerge faster. Strike while the iron’s hot, but chasing others’ successes may be more dangerous now than ever.

As always, thanks for reading!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of Felicis Ventures Management Company, LLC.