opinion 7 min de lecture

Someone Else Is Paying for Your AI, Just Not the Part You Think

The comforting story about your 20 euro AI subscription is that it is sold below cost and the real bill is coming. In 2026 that stopped being true. The subsidy is real, but it sits somewhere else, and it reaches you as rationing rather than as a price rise.

CourionAI

13 août 2026

Cet article n'est pas encore disponible en français. Voici la version anglaise.

Twenty euro leaves your account every month, plus VAT, for a chat window that answers at three in the morning and never gets tired. In the same week you read that the company sending you that invoice expects to lose about fourteen billion dollars this year. The two facts sit next to each other and produce an obvious conclusion: you are being subsidised, your subscription is sold below what it costs to serve, and one day the grown-up price arrives.

It is a tidy story. In 2023 it was mostly true. In 2026 it has quietly stopped being true, and the version that replaced it is more useful to know.

Your monthly subscription is probably no longer the part that loses money. Investor cash now pays for the free tier and for the next model, and that bill reaches you as rationing rather than as a price rise.

Serving you got cheap, embarrassingly fast

The cost that matters here is inference, which is simply the cost of running a finished model to answer your question, as opposed to the much larger cost of building it in the first place.

Epoch AI tracks what it costs to buy a fixed level of ability over time. Across tasks, the price to reach a given performance level has fallen somewhere between 9 times and 900 times per year, and the spread is the interesting part. PhD level science questions got cheaper at roughly 40 times per year. General chat quality moved far more slowly. Epoch’s own caveat is worth repeating: the steepest drops happened in the most recent year, so nobody should assume they continue.

The company numbers point the same way. OpenAI has reportedly cut its inference cost per unit of work by around 95 percent since GPT-4 launched in early 2023. Its gross margin, meaning what is left after paying to serve the product, ran at about 33 percent across 2025 and roughly 39 percent in the first quarter of 2026, according to leaked financials reported by the Financial Times and others. Anthropic’s inference margin reportedly climbed from 38 percent to above 70 percent during 2026, carried by heavy Claude Code use.

A fair caveat before you build anything on those figures: neither company publishes them. Every margin number in this article is reported or leaked, not audited in public.

The money burns one floor up

If serving is profitable, where do fourteen billion dollars go?

Look at 2025. OpenAI booked 13.07 billion dollars of revenue and posted an operating loss of 20.92 billion. In the same year it spent 17.2 billion dollars with Microsoft Azure alone, more than its entire revenue. Anthropic’s 2025 inference bill reportedly passed 2.7 billion dollars, more than triple the year before, while its overall gross margin landed near 40 percent against a 50 percent target.

None of that gap lives in your chat window. It lives in training runs, data centre commitments, and salaries for a few thousand people. Your twenty euro is not covering a loss on your conversation. It is buying a seat while somebody else finances the option on the next model.

You are paying for the person who pays nothing

Here is where the working title of this piece was right, just aimed at the wrong target.

Roughly 95 percent of ChatGPT’s weekly users pay nothing, and serving them reportedly costs about 3.9 billion dollars a year in inference. Spread revenue across every weekly user and you get somewhere near 28 dollars per user per year. A Plus subscription is 240. The cross-subsidy is real, it simply runs sideways: from paying users and investors to free users, and from light payers to heavy ones.

Watch what happened next, because it is the clearest signal in the whole story. In January 2026 OpenAI did not raise the price of Plus. It launched an advertising-supported ChatGPT at roughly 5 to 8 dollars a month in global markets. When the bill arrives, it does not land on your plan. It lands on the free tier.

The strongest case against this

The heavy user genuinely is served below cost, and nobody serious disputes it. Somebody running coding agents eight hours a day consumes far more compute than twenty euro buys. That is exactly why the ladder exists: ChatGPT Pro at 100 dollars, a 200 dollar tier above it, Claude Max at 100 and 200. It is why Claude Pro runs a five hour session limit plus a weekly cap across all models, and why ChatGPT Plus quotes its allowance in messages per three hours rather than per month.

That is a strong objection, and it is correct about the tail. But notice what it predicts. If the twenty euro plan were structurally underwater, the fix would be to charge more for it. Instead the fix has been to fence the usage and sell a bigger fence. That is a business protecting a profitable light user from a costly heavy one. The caps are the mechanism, not the counterexample.

The honest hole in my own argument is different: none of this is auditable from outside. If the real figures are meaningfully worse than the leaked ones, the naive story wins and I am wrong.

What is different in Europe

Two things. First, the headline prices you read are American and pre-tax. In Germany 19 percent VAT rides on top of the number in the blog post you were reading, so the amount on your statement is the checkout screen, not the conversion.

Second, Europe currently has the cheapest credible option on the table. Mistral’s Le Chat Pro sits at 14.99 a month against 20 for Plus and Claude Pro, with a free tier of roughly 25 messages a day. Whether that price is durable is a separate question, because Mistral is spending investor money too, and the second cheque problem means it is doing so with less of it. But the European reader is the one who can leave without giving up a working tool.

The ad-supported tier, meanwhile, launched in global markets, which in practice tends to mean Europe later and with more consent friction, courtesy of GDPR. Cheaper access here will probably arrive as a smaller free tier before it arrives as ads.

What this means for you. If you are curious and not paying, nothing changes this week, but watch the free tier rather than the paid prices. That is where cost discipline lands first, and it is already moving. If you pay twenty euro and use it a few times a day, you are the profitable customer and you are fine. If you are a power user, read the usage limits page before the pricing page. Annual prepay is a bet that the caps stay where they are, and caps are the one thing that can change without an announcement. Keep your own exports either way.

What would change my mind

A published, audited gross margin from OpenAI or Anthropic showing consumer subscriptions serve below cost. That would mean the simple story was right all along.
Headline price rises on the 20 dollar tier at two major providers within twelve months, without a matching loosening of usage caps. Rationing and price rises together would mean the flat fee really is underwater.
Epoch AI’s inference price series flattening or turning upward for two consecutive quarters. The whole argument rests on serving costs falling faster than usage grows, and that is an assumption, not a law.

Sources

Article suivant

Stop Asking Whether AI Is a Bubble. Start Reading the Lease Agreements.

The bubble debate has two defensible answers and no useful ones. Splitting the market into layers helps a little. What actually gives you an early warning is duller and far more specific: four year leases on buildings meant to last decades, and the guarantees the tenants signed to get them.

Risograph illustration of a desk with a thick lease document, a pen resting on the signature line, and a small data centre building sketched in the margin