The Three Types of AI Company
There are roughly three categories of company in the current AI market: the labs that build frontier models, the startups that wrap them in products, and the infrastructure companies that rent compute.
None of them are making real money. But each category is not making money in a different way, and understanding the difference matters for figuring out which ones eventually will.
Infrastructure: The Commoditisation Problem
AI infrastructure , data centres, GPU rentals , is one of the most commoditised businesses in the history of technology. CoreWeave, Nebius, Iron, Lambda: all backed by Nvidia, all selling the same hardware at similar prices, all competing on a product they cannot meaningfully differentiate.
None of the major Nvidia GPU generations have been clearly profitable for the operators renting them. The economics are approximately: rent GPU time from a company that built a data centre to do so. Charge customers for that GPU time. Hope the spread is positive. For most operators, it is not, or barely.
Nvidia is extraordinarily profitable. The companies buying Nvidia hardware to resell as compute are not. The machine is enriching the component supplier and running thin margins everywhere downstream.
The Wrappers: Margin Math That Doesn't Work
AI startups that build on top of foundation models face a specific economic problem. Their cost structure is the API pricing of the underlying model. Their revenue is whatever their customers will pay for the product built on top of it.
The margin between those two numbers has to cover engineering, sales, marketing, and customer support. For most AI startups, that margin is negative or close to it. The pricing pressure from users , who can often access the underlying model directly , keeps product prices low. The API costs, especially for the most capable models, keep cost of goods high.
There is no sign that inference costs are coming down in a way that fixes this. OpenAI and Anthropic are not particularly concerned with bringing those costs down , their revenue comes from the API pricing itself. The wrapper startups are trapped in a margin problem that the underlying economics are not moving to solve.
The Labs: Spending What Doesn't Exist Yet
The frontier labs are spending as if future revenue is certain. It may not be. Every major AI lab is running on investor capital, not operating profit. The revenue is real , OpenAI and Anthropic are generating billions. The gap between that revenue and what would be needed to justify current spending is also real, and it is large.
The honest comparison is the dot-com bubble. Pets.com was spending $250 per customer acquired. Webvan was building logistics infrastructure for grocery delivery that the market wasn't ready to pay for. The economics didn't make sense and everyone in the industry privately knew it.
But here is the other half of that comparison: Pets.com became Chewy. Webvan became Instacart. The business models were right. The timing and capitalisation were wrong. Most of the companies failed. The idea survived and eventually created enormous value for the survivors.
The Thing Everyone Is Betting On
The AI market is operating on the assumption that costs will fall, revenue will rise, and the economics that currently make no sense will eventually make a great deal of sense.
That assumption could be correct. The inference cost curve does tend to follow hardware improvement curves over long periods. Revenue from AI applications is growing. The window between "burning cash" and "profitable at scale" might be shorter than the dot-com era.
Or it might not be. The current market is not making bets on which companies will win. It is making a bet that the underlying assumption holds , that this is a timing problem, not an economics problem. That bet has not yet been proven right. The receipt is still being written.