Why AI companies lose money

A quick honesty note before the argument, because it is the sort of thing this site should do: not literally every AI company is losing money. Plenty of profitable businesses use AI, the chip makers are enormously profitable, and some model providers are likely to be making money on the inference itself. The companies losing spectacular amounts are the frontier labs and the startups competing with them. The headline is a shorthand, and shorthand is where bad understanding comes from, so let us be exact instead.

The thing AI broke

Software was the best business ever invented for one reason: the second copy is free.

Write a program once, and serving one more user costs approximately nothing. Costs are fixed, revenue grows with users, and margins climb towards the ceiling as you scale. That single property is why software companies were valued the way they were.

AI does not have that property. Every answer costs real compute on real hardware, so cost grows with usage. That is not a software business. It is closer in shape to an airline: heavy fixed costs, a real marginal cost per unit sold, and a permanent fight over utilisation.

Once you see that, the strange behaviour of the whole industry becomes legible.

Where the money actually goes

Inference, which is usage. Every query costs. A free tier is not a marketing expense that stops when you hit scale, it is a subsidy that grows precisely as fast as the product succeeds. In software, a user who loves your product costs you nothing extra. Here, your most enthusiastic free users are your largest single expense.

Training, which is a bet. Building a frontier model means an enormous concentrated spend on compute before you have any idea whether the result is competitive. Published figures for frontier training runs are estimates rather than audited numbers, because the labs mostly do not disclose them, and I would rather tell you the figures are contested than quote one with false precision. Stanford HAI's AI Index collects the public estimates and is honest about their limits.

Depreciation that is brutally fast. A model is a capital asset that a competitor can obsolete in a few months. A factory depreciates over twenty years. A frontier model can lose most of its commercial value in one, and there is no way to hedge that.

A price war on top of all of it. Per-token prices have fallen steeply and repeatedly, driven by competition and genuine efficiency gains. Falling prices are wonderful for you and me and hard on anyone trying to earn back a training run.

The distinction nobody explains

Here is the thing worth knowing, because it is where most public commentary goes wrong.

Gross margin is whether you make money on the thing you sold. Net loss is whether you make money overall after research, salaries and training runs.

It is entirely possible, and is probably the actual situation at the larger labs, to have healthy gross margins on inference and enormous net losses, because the money is going into the next model rather than into serving the current one.

That matters because the two situations have completely different futures. A company losing money on every query has a broken business. A company earning a margin on every query while spending it all on the next model is making a bet. It might be a bad bet. It is not the same disease, and "AI companies lose money on every query" is usually stated as though it were established fact. When you see that claim, ask which of the two numbers it refers to.

Why raise so much, then

If usage costs money and prices keep falling, why pour in capital?

Because the plausible prize is a general-purpose technology layer, and the people funding it think landing that is worth almost any present loss. Whether it turns out to be a platform like the smartphone, a utility like electricity, or a commodity like bandwidth is genuinely unsettled, and those three outcomes have wildly different returns. Anybody telling you confidently which one it is, in either direction, is guessing with more conviction than the evidence supports.

What would actually change the picture

Rather than a prediction, here are the things to watch, which is more useful:

• Inference cost per unit of capability. It has been falling fast. If it keeps falling faster than prices, margins repair themselves without anything else changing.
• Whether model quality plateaus. If everyone's model is roughly as good, nobody can charge a premium and it becomes a commodity, priced like bandwidth.
• Whether anything becomes load-bearing. The most valuable position is not the smartest model, it is the one businesses have wired so deeply into their operations that leaving is painful.
• Power and hardware. Compute is the constraint, and it is increasingly an electricity and supply-chain story rather than a software one.

The summary worth repeating

AI took the best property of the software business, near-zero marginal cost, and removed it. Everything strange about the industry, the free tiers that get more expensive as they succeed, the enormous raises, the relentless price cuts, the obsession with efficiency, follows from that one change.

The losses are structural rather than a sign that everyone is being stupid. Whether the bet pays is a separate question, and one I would not trust anybody's confident answer to, including mine.