Does a $200 AI subscription really buy $14,000 of compute?

You have probably seen the screenshot. A $200 a month ChatGPT Pro subscription delivers about $14,000 a month of API usage. Claude Max, at the same price, about $8,000. It circulates with great confidence, usually with the conclusion attached: the labs lose a fortune on every heavy user, and this cannot last.

The source is better than most viral numbers. SemiAnalysis bought one of each OpenAI and Anthropic plan, ran long-horizon coding tasks until the weekly caps ran out, and priced the resulting token volume at API list rates. The earlier rule of thumb was about $2,000 per plan, so this is a real finding. A credible shop, real work, with the detail behind a paywall. I buy and use these tools every day for cloud and software work, so I followed the number to its source. The research holds up. The reading does not.

API list prices carry large margins. Epoch AI estimates that GPT-class models are served at around 50% gross margin. Hosted API businesses generally run at 60 to 80%, and Anthropic's API margin has been reported above 80%. Price $14,000 of retail API usage at cost and you get perhaps $3,000 to $7,000 of compute.

The figure also assumes a user who exhausts the weekly caps every week. SemiAnalysis's own companion number is that OpenAI crosses into loss at roughly 11% of a base plan's limits, so most subscribers sit well inside the line and a minority of heavy users far past it. The $14,000 is what the heaviest possible use would have cost you at retail, not what it cost them.

The honest version is that OpenAI loses money on its heaviest users, by an amount nobody outside the company knows, on a plan most subscribers use lightly. Less quotable, and true.

What interests me more is how readily the figure was adopted, because nobody had to lie.

For the labs, "we are losing money on you" is a useful sentence. It pre-justifies every price rise for three years, flatters the buyer as a power user rather than a customer, and signals ambition to investors. When Sam Altman says it out loud, it is humblebrag and expectation management in one breath.

The amplification layer has its own incentives. Much of the coverage is content farms rewriting one tweet. Some comes from crypto and decentralised-AI outlets, for whom the finding is a sales argument for their own product. No coordination is required. A lot of people independently found one number useful, and that is how durable received wisdom usually forms.

Subsidise your way in, own the market, raise prices later.

It is a decent argument and deserves a real answer rather than a dismissal. Uber burned roughly $30bn cumulatively over about fourteen years before 2023, its first full year of GAAP profit. It won. Loss-making entry into a new market is an ordinary strategy with a respectable track record, and if that were the whole story the $14,000 screenshot would be a curiosity.

In three places, and one of them cuts the other way.

AWS is the misremembered case. It is widely assumed to have lost money for years, and mostly it did not. When Amazon first broke it out in 2015, AWS was already running at roughly 19% operating margin on about $8bn of revenue. Capital hungry, not loss making: the investment sat in capex, depreciated over a decade and earned throughout. On this axis OpenAI looks more like AWS than Uber. Serving is gross-margin positive, at roughly 43% on the 2025 figures, and the loss sits in R&D and sales.

The moat mechanism is weaker. Uber's subsidy bought liquidity in each city, which reinforces itself and has to be outspent city by city. AWS bought data gravity and skills lock-in. Model labs buy distribution and habit, which are real, and workflow embedding, which is rising. But there is no network effect between users, and at the API layer switching costs are close to nil. Uber never faced a rival giving rides away at cost with an open-source dispatcher. The labs face that roughly once a year.

The cost curve is better than Uber's and worse than it looks. Uber's marginal cost was a human driver and never fell, which is why its route to profit was taking more of the fare. Inference genuinely gets cheaper, around 95% per query for a given level of capability, which is why code is getting cheap. But the cost of running the actual frontier rises steeply, and demand migrates to the frontier as fast as it arrives. Uber never had riders demanding journeys forty times longer every year.

Uber built a dispatch network once. AWS built data centres that depreciate over ten years and earn the whole time. OpenAI's 2025 R&D spend of $19.18bn, on the leaked statements Ed Zitron obtained and the Financial Times verified, buys a model that is commercially stale within twelve to eighteen months and matched by rivals, including free-weight ones, before the build is amortised. Uber and AWS paid once to get in. OpenAI pays again every year, and more each time, to stay in. The nearest comparison is not Uber or AWS but a pharmaceutical company with a two-year patent.

Brand, distribution and agent-level lock-in do accumulate even while the models decay, and whether they accumulate faster than the models decay is the whole question. Nobody knows yet.

Survivorship bias, first. We reach for Uber and AWS because they worked. The same playbook produced WeWork, MoviePass and a graveyard of delivery firms. The analogy shows the strategy is viable. It does not show it is likely.

Anthropic, second, as the live counterexample. Roughly 60% gross margin in 2026, up from deeply negative in 2024, tracking towards positive operating profit, enterprise-led, and without subsidising consumer power users nearly as hard. If that route works, the land grab is a choice rather than a law.

On the leaked 2025 statements, OpenAI had revenue of $13.07bn, cost of revenue of $7.5bn, R&D of $19.18bn, sales and marketing of $5.73bn and an operating loss of $20.92bn. Uber's entire fourteen-year cumulative loss is smaller than OpenAI's 2025 operating loss alone, and that loss sits against reported compute commitments in the region of $600bn through 2030.

The strategy is ordinary. The size of the bet, relative to any market that currently exists, is not.