1The bloodbath was scheduled
In the third week of March 2024 two of the most admired AI companies of the previous year came apart in the same news cycle. Inflection, which had raised more than a billion dollars to build a personal chatbot, saw its founders and most of its staff walk into Microsoft, which paid the investors back through a licence and kept the people; three days later the chief executive of Stability, whose image models had made the company famous, resigned and told the world he was leaving to pursue decentralised AI.1 The commentary called it a bloodbath. It was a schedule.
Anyone who was working during the last war of the titans had seen the shape before. Microsoft won the first browser war by crushing Netscape, and then lost the long one to open source, to Firefox, to Chromium and to WebKit, until the browser became a thing nobody could charge for; and Facebook, which arrived too late to own any of the surfaces where people met the internet, search, the browser or the phone, spent the next fifteen years paying rent to the companies that did.2 In 2024 Microsoft was all around OpenAI, and open source was all around Microsoft, because the company that had learned the rent lesson hardest was releasing its models for free to make sure nobody could charge for them either. Open source can deny a win to a competitor. It has rarely been a way to win, and the moats in this war were never going to be built from model technology at all.
The economics of a foundation model start-up did not make sense in 2024 and they do not make sense now, for a reason that has nothing to do with the quality of the research. The technology is an order of magnitude more expensive to deliver than any previous generation of knowledge technology, and the buyers on both sides said the same thing about it in different words: consumers thought it did too little for what it cost, and companies thought it did too little for what they needed. A start-up with a few hundred million dollars and a good model was trying to solve a supply chain and a market, not a technology, and supply chains can stay rational a great deal longer than a venture-funded company can stay solvent.
2Four inputs, and the cheap one is the only one that scales the product
Strip any AI business to its inputs and there are four: people, who carry the intellectual property; compute, which is energy with a delay; data; and distribution, which is the market. A company needs an edge in at least one, and in 2024 it was already possible to say which edges would hold.
People and their ideas are the input a start-up can most easily afford, and the one it can least easily keep, because the academic and open-source community publishes every breakthrough within weeks of its being made and the companies releasing their weights make sure the frontier is a public good. There is no lock-in on model IP at the foundation layer, and a business built on it is a business built on a lead measured in months.3
Compute is the interesting one, because it is different this time, and the difference explains why the previous generation of the industry was so much cheaper. A decade and a half ago falling chip prices and broadband gave us the cloud, and the cloud gave us new business models, but it did not make applications smarter; compute scaled with the number of users, and your software did not become twice as bright with twice the servers unless somebody wrote the code and somebody else paid for it. With these models the compute is the product. More of it means a better model, in training and in use, and a better model means a better application on every axis at once, independent of anything else the company does. That turns compute into a competitive edge and turns the edge, one step down, into the cost of energy, which is the only real bottleneck underneath it, and the labs began saying in public that compute and energy would be the currency of the future.4 The prediction was correct and useless to almost everyone, since unless you own a fusion company the currency is not yours to mint, and even if you do the grid is a long game with many bottlenecks and the connection queue does not care about your roadmap.
3Distribution belongs to three companies, and one start-up became a fourth
The right market can solve most technology problems, and the consumer market had a distribution problem that looked terminal. ChatGPT had reached a hundred million users faster than any application in history, and then the curve bent: a year later it had around a hundred and eighty million users, half of what the exponential curves of the previous decade would have delivered. Without distribution there was not enough consumer demand for AI to feed one lab, let alone ten, and the places where demand could be created, an assistant on the phone or a browser that acts, belonged to Apple, Google and Microsoft, with a billion, five billion and a billion users respectively.5
That paragraph needs its correction, because the correction is the most interesting thing that has happened since. OpenAI did what the argument said could not be done: it became the distribution. By late 2025 the application had eight hundred million weekly users, it had become the default place people go to ask a question, and the incumbents’ assistants either slipped, as Apple’s did, or arrived as features inside the one thing people already used.6 For every other foundation model start-up the prediction held to the letter. Inflection, Adept and Character were absorbed into Microsoft, Amazon and Google within six months of each other through the same licence-and-hire manoeuvre, and Stability became a different company under new owners. The consumer market did not choose the labs; it chose one lab and the three incumbents, and left the rest without a surface to live on.7
4Enterprise said meh
For companies that cannot make it as a consumer product there is always the enterprise, and by 2024 most of the labs had discovered it. The question was what they would offer that a company could not get elsewhere, because the cloud providers were already delivering capable open models at a low price, and the price was falling by an order of magnitude a year.8 Worse, the companies that had been interested at first had now met the copilots and the bolted-on chatbots and had said, in the word every chief information officer used to me that year, meh. They agreed they would pay more for AI that did more; what they were being sold looked like an overpriced gimmick, and the only categories that had made commercial sense were chatbots with a retention problem and robotic process automation with a new name.9
The meh has since been measured. A widely read study in 2025 found that the great majority of enterprise generative AI pilots produced no measurable effect on profit and loss, and the reason it gave was not the models but the gap between what a general tool can do and what a company’s work actually requires: the context, the workflow, the data the model was never shown.10 The lab that pivots to the enterprise has to build the whole of that, and none of it is model-specific; it is software engineering and product design, the least glamorous inputs in the industry and the ones the labs are worst at.
5Data is the expensive input, and the cheap kind needs a scorer
The last input is the one that decides the rest. The only significant variance between the large models in 2024 was the data they had been trained on, and since they had all used the same public internet they were mostly the same, which is why most consumers could not tell them apart. From a supply-chain view data is also the dearest component: there is a finite stock of public text, it took decades and trillions of dollars of infrastructure and human labour to create, and the labs had already consumed most of it. Scaling a transformer to anything like general intelligence on the same recipe would need ten to fifty times more, and the stock was not there; the researchers who counted it put the exhaustion of public human text somewhere in the second half of this decade.11 In the enterprise the problem inverts, because the data that matters there is proprietary, and the large models lose the one edge they had.
The cheaper alternative is synthetic data, which costs a fraction of the real kind to generate, and it has a condition: you need to know what winning looks like. AlphaGo Zero beat the version that had learned from human games using no human data at all, and it could do that only because Go has rules and a way of knowing who won, so every self-played game could be scored without a person in the loop.12 That was the prediction in the March 2024 post that landed hardest. Within a year the frontier moved to models trained by reinforcement learning on tasks with verifiable answers, mathematics and code first, where the scorer is cheap and unarguable, and the lab that did it most visibly did it from China with a fraction of the compute and wiped half a trillion dollars off the chipmaker’s market value in a day.13 Clear winning criteria and easy validation turned out to be worth more than a bit more scraped text, and the same is true a fortiori in the enterprise, where nobody has the text anyway: define what counts as done and a smaller model in a scored world beats a larger one without one, at a fraction of the cost.
The other frontier is time. Nearly every model is trained on a snapshot, and a model trained on sequences of events, with enough variance in them, can predict the next event or action rather than the next word. That part of the prediction arrived under a different name, as agents, and it has been slower than the scoring, because an event has consequences a token does not.
6Most foundation model start-ups are screwed, and the field is wide open
The conclusion of the 2024 post stands, and two years of evidence have made it less controversial than it was when it was written. The consumer market runs on distribution, and it was won by one lab that became the distribution and by three incumbents who already owned it; the foundation model companies that could not tell themselves apart were absorbed for their people, as the browser makers were, and the models that survived have converged to the point where they are rented by the token from several suppliers and switched between without anyone noticing. Model IP did not become a moat. Compute became energy, and energy belongs to whoever got into the grid queue first. The labs’ pivot to the enterprise required an entire infrastructure that has nothing to do with models and everything to do with the kind of engineering and product work that turns a capable model into something a company will pay for.
That is the wide-open field, and it is the same field it was in 2024 with the fences moved. The winners in it will not be the companies with the best model, because the best model is available to everyone by the token; they will be the ones who hold the proprietary data, who can say what winning looks like in a given piece of work so that the cheap synthetic data can be scored, and who own the surface where the work gets done. The oil was never the business. The pipeline to the coast was.14