Anthropic now makes its money running models, not building them. It will not say what running them costs.
The company briefed prospective investors at its San Francisco headquarters this week on revenue of $16.8 billion for the three months to the end of September, up 45%, after $4.7 billion in the first quarter and $11.6 billion in the second, as the BBC reported. Read as a market number, that is a growth curve, and the share sale the briefing is meant to prepare — expected by the end of the year — will be priced off it. Read as an engineering number, it says something more specific. A business that adds $5.2 billion in a single quarter while the price of a unit of model capability keeps sliding is a business whose volume is moving considerably faster than its revenue. Anthropic’s chief executive has said publicly that the company spends several times more on inference — running the models for users — than on training them. At that ratio, this is a quarter of serving, and the growth is coming from models wired into production systems: code assistants, support queues, document pipelines. That is what consumption looks like, as distinct from pilots that end.
What the curve cannot tell you is whether it pays, because the deciding number was not on the agenda. Revenue is price times volume, and the price term is not a constant. Tokens — the chunks of text a model is priced in — get cheaper to serve at a rate that has nothing to do with how fast they sell. Epoch AI, which tracks what a fixed level of model capability costs over time, has found that price falling by roughly an order of magnitude a year. If per-token pricing keeps sliding anywhere near that rate, then 45% revenue growth implies token volume roughly doubling in a quarter: an order of magnitude more inference in six months. Doubling tokens is not a marketing achievement. It is a serving problem, and the work is unglamorous — batching, meaning many requests run together; quantization, meaning the model’s numbers stored with less precision; distillation, meaning a smaller model trained to imitate a bigger one so the traffic can be carried more cheaply. Margin at Anthropic is decided by how fast that internal engineering work outruns the price curve, and the briefing did not include that number. An Anthropic spokesperson declined to comment. That is standard practice, and it is also the whole question.
Which is where the pressure on hardware names belongs, rather than being waved off. Chip and data-centre stocks have come under pressure as investors do the arithmetic on infrastructure spending, and the nervousness is not superstition. Two of the risks are technical rather than sentimental. The first is yield on training compute: a frontier training run costs hundreds of millions of dollars and can be obsoleted inside a year by somebody else’s run, which means the asset depreciates on a competitor’s schedule rather than an accounting one. The second is the mismatch between when capacity is paid for and what it earns: a data centre is bought with dollars today against revenue whose unit price is falling by roughly an order of magnitude a year, and capacity provisioned for peak agent-style workloads — features that act rather than answer — sits idle between spikes. Anthropic’s sequential progression is the strongest evidence available that this particular demand is real consumption rather than procurement theater; you cannot double token volume with a press release. But revenue is not margin, and the labs are now asking public investors to fund a build-out whose unit economics have never been shown to anyone outside the companies.
OpenAI, whose eventual listing is expected to value it above $1 trillion, said in September that it will not go public this year while it addresses concerns about the technology’s safety, the BBC reported. I will grant Sam Altman the strongest version of that. A safety case is an engineering artifact: a documented threshold, an evaluation battery with defined failure conditions, and an authority that can stop a shipment after the demo works. If the delay bought any of that, it is worth more than a quarter of listing proceeds. The trouble is the public record on the gate. Jan Leike, who ran OpenAI’s long-term safety research, left in 2024 writing that “safety culture and processes have taken a backseat to shiny products.” The deployment risk has since stopped being hypothetical: OpenAI’s own published analysis with outside researchers found roughly one in seven hundred weekly users showing conversation patterns consistent with psychosis-like states — a small number, a real one, and precisely the failure mode that memory-persistent, always-on products aimed at heavy users make worse. In October 2025 the Federal Trade Commission opened a study under its Section 6(b) authority — its power to examine an industry rather than wait for a complaint — into how OpenAI’s products affect young people and how reliable its safety commitments are. That is the safety-versus-ship tradeoff at the engineering level. Persistent and agentic features multiply the blast radius of a model failure: a hallucination in a chat window costs the user an evening, while a hallucination routed into a booking, a payment, or a prescription costs a third party who never consented to the experiment. The evaluation regimes that gate a chat feature do not gate those deployments at the same thresholds. And a trillion-dollar expectation makes the tradeoff structural, because every quarter the listing waits, the incentive to demonstrate shipping velocity compounds, and safety becomes the sentence in the filing that explains the timing.
The lever here is disclosure, and it is not complicated. If the public markets are going to fund the capacity, they are entitled to the two numbers the private markets never demanded: what a million served tokens costs at the operator’s scale, and the name of the gate with documented authority to stop a deployment. My father’s bar mill posted the shift’s throughput and the cost per tonne beside it, every shift, because the cost per tonne was the only number that decided whether the shift had been worth running. The labs can manage the same arithmetic. They are asking the public to fund the mill.