Recently, Dwarkesh posted a series of questions. One of them was:
What’s the most plausible story where foundation model companies actually start making money? If you consider each individual model as a company, then its profits may be able to pay back the training cost. But of course, if you don’t train a bigger, more expensive model immediately, then you stop making money after 3 months. So when does the profit start? Maybe at some point scaling will plateau, but if progress at the frontier has slowed down, then the combination of distillation and low switching costs (cloud margins result from high switching costs) makes it really easy for open source to catch up to the labs, eating into their margins. So how do the labs actually start making money?
I thought this was an interesting question, and it will be interesting to see in a couple of years how it shakes out. For the sake of posterity, my take is written down here.
Consumer and Enterprise AI Markets
The two most obvious AI markets are consumer and enterprise.
Consumer
I think the consumer market is actually somewhat underappreciated. Right now, a common way to look at this market is to look at the chatbot market, which has about a billion users with a 5% conversion rate to a $20/month subscription. That implies about a $10–15 billion annual business. That is clearly attractive, but not extraordinary relative to the broader AI opportunity.
But I do think the market could be much larger if personal agents meaningfully increase willingness to pay. If agents become reliable enough to perform useful tasks, users may want their personal agent running for multiple hours a day. In that world, it seems plausible that over the next 2–5 years, a much larger share of chat users would want a premium subscription.
For example, if we assume 60–70% penetration due to the usefulness of personal agents, that would imply roughly $150 billion in annual revenue.
The attractive feature of the consumer market is stickiness. A personal agent will accumulate context, connect to many services, and may eventually be trusted with money. Once that happens, switching costs become very high. Whoever gets there first could capture a large share of the market.
Right now, there has not been any breakthrough personal-agent product; the closest so far has been open claw. It seems like there is still a bit of a capability gap, as well as a product gap, to create a personal-agent experience that is compelling for everyday users. I think frontier model companies are very well positioned to close this capability and product gap, since they can co-design the model and the product. Once a model-product pair can handle most personal tasks, extra capabilities matter little, so investment in new models for this product category can be limited.
Enterprise
Enterprise is the larger and harder-to-size market. In theory the TAM could approach the total wages of all work done on a computer, which is about $20 trillion a year. I would divide this market into two segments.
1. Computer work highly levered by frontier intelligence
Some computer work is unusually levered by frontier intelligence. In these domains, even a small capability edge can be enormously valuable. If a model helps discover a drug, improve chip design, discover better optimization algorithms, generate trading alpha, or optimize industrial processes, the customer may pay far above inference cost because the model contributes to high-value outcomes where small improvements can make a huge difference.
The reason this market is especially attractive for frontier labs is that frontier capabilities appear to require frontier scale. We can see this today: the strongest math, science, and coding models are generally large, general-purpose frontier models, and the next leap in capability is likely to come from even larger models like Mythos or Spud.
To illustrate this point a bit more, suppose you are designing a state-of-the-art gas turbine. In that case, you want your model to be world class at engineering, coding, research, physics, and broad enough in its general knowledge to come up with novel insights and connections. A scaled model that has all of these capabilities will be most effective at improving your turbine design, and you will almost certainly be willing to pay a high premium even if the result is only slightly better.
I do think intelligence is like “roundedness,” which would imply that at some point there will be diminishing returns. On the other hand, the human benchmark gives us a useful lower bound on how far models can go for sure, and current models are still far from matching the best human on every task. My guess is that this leaves room for steady progress over at least the next 3–5 years.
Right now, adoption of frontier intelligence is growing at an incredible rate, which justifies almost any training spend. But over the next couple of years, I expect the market to discover how much work truly relies on frontier capability and what average premium customers are willing to pay. My personal prediction is that the market for frontier capability is substantial, and that the capability gap will remain large enough to matter, easily supporting multi-billion-dollar spend for the next model generation, at least for the next 3–5 years.
This is also why I am skeptical that open-source models will remain competitive here. Frontier development requires continuous large investments, while this market rewards only the very best models. If a model is excellent but still behind the top proprietary systems, its value will be disproportionately lower, so the economics do not seem to pencil out for open source.
2. Fixed-intelligence computer work
Most computer work does not require frontier-level intelligence. We can already see in coding that, for many tasks, cheaper open-source models are becoming good enough. I expect this trend to continue.
For these tasks, the required level of intelligence is relatively fixed. Once models clear that threshold, the market should increasingly optimize around cost. As a result, this segment will likely fragment and specialize over time.
That does not mean no one will make money here. Companies that deliver cheap, reliable models for specific verticals can still build good businesses, I think. But the economics are structurally limited: by construction, the TAM of any one vertical is capped, and once a vertical becomes large enough in revenue terms, cheaper and more specialized models will emerge to compete for it.
This is probably also a market where open-source models will put even more pressure on margins. For popular verticals, like average coding tasks, it makes a lot of sense for companies, philanthropists, etc. to create open-source models to support the broader ecosystem.