Tokens are not money

A chart Bloomberg shared in July showed that more than half the tokens used by American businesses now run through AI models built in China. The line written underneath it says Chinese AI models are taking over American business and beating OpenAI and Anthropic in their own backyard.

The numbers are accurate. The line underneath them is not. The chart in effect measures price competition at the cheap end of the AI market, but it is being reported as competition across the whole market. The first is real and matters a great deal. The second isn't happening, and the same data shows that it isn't.

Where the number comes from

CNBCChinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surgeRecent model releases from Chinese companies including DeepSeek and Z.ai are seen by many as highly competitive compared to leading U.S. frontier systems.

The data comes from a business called OpenRouter, which sits between software developers in businesses and some four hundred models from dozens of companies. A developer sends a request, OpenRouter forwards it to whichever model they have chosen, and handles the billing. It's a real business: handling roughly 25 trillion tokens a week for about eight million developers. In August Stripe agreed to buy it for more than $7bn. Consequently, the data is worth taking seriously.

This is what makes the data useful, but it is also the case that its customers are unrepresentative. People who buy this service through a router are, by definition, the people most willing to move provider the moment a cheaper model appears. The traffic mix on OpenRouter is a picture of that group, not of the entire market.

Bloomberg and CNBC both reported the numbers over the summer. On OpenRouter, in June 2025, models built in America handled about 70% of the tokens used but that number was down at 30% a year later. Among accounts billed in the United States, the Chinese share has been above 30% in every week since February and was reported as high as 58% in July. Geography on the platform is inferred from the billing address, so "US firms" means accounts paying with an American card.

What the traffic actually is

OpenRouter's own published study is candid about what the numbers cover. More than half of the usage is outside the United States. And “roleplay” and creative fiction account for about 52% of all open-source model usage on the platform.

Roleplay, by the way, is exactly what it sounds like. Character chat apps: you invent a character, or pick one somebody else has made, and talk to it. The largest by volume is Janitor AI, which has an adult-content toggle (so I think it's fair to assume that a decent chunk of the category is adult). That goes a long way to explaining why Chinese models dominate it. This is not about capability. This is about price, and furthermore, OpenAI's and Anthropic's usage policies rule most of that traffic out. Part of what the chart shows is Chinese models winning a business the American labs have chosen not to be in, and not a Fortune 500 procurement decision, which is what the headline invites you to picture.

What's not in the chart

In fact, only a tiny proportion of the money in the AI market moves through routers. The vast majority moves through three other channels.

The first is subscriptions. More than 900 million people use ChatGPT every week, over 50 million of them pay for it, and there are more than seven million workplace seats. Claude, Gemini and Microsoft Copilot sell the same way. Not one of those subscriptions appears in the chart, because traffic from these apps goes straight to the model provider instead of being channelled through OpenRouter.

The second is direct enterprise contracts, signed either with the labs or through Azure OpenAI, Amazon Bedrock and Google Vertex. None of that appears either. Enterprise contracts often explicitly disallow retention or model training against their data.

The third is direct API traffic: developers who point their code at the model provider’s infrastructure directly, rather than going through a router.

The scale of what is missing from the OpenRouter data is easy enough to establish, because the frontier companies publish it. At its developer day in October 2025 OpenAI said its API alone was processing more than six billion tokens a minute; by March this year that figure had risen to around fifteen billion per minute. That's roughly 650 trillion tokens a month from the API channel alone, before a single ChatGPT subscription or enterprise client is counted. For the record, OpenRouter does about 100 trillion a month.

Anthropic tells the same story in its financials. Its annualised revenue run-rate went from about $9bn at the end of 2025 to $65bn as of July, with roughly 80% of it from API customers and enterprise contracts. Almost none of that business is visible on OpenRouter either.

Tokens are a volume measure

Putting all these caveats to one side, token share is still the wrong yardstick.

A token is a unit of output, not a unit of work done.

Prices in this market also vary by more than a hundredfold. DeepSeek's V4 Flash lists at $0.28 per million output tokens. GPT-5.5 lists at $30 and Claude Fable 5.1 at $50. Justin Summerville, who runs data and analytics at OpenRouter, puts the savings from Chinese open-source models at between 60% and 90%.

On benchmarks, the gap between the best American and the best Chinese models is not enormous – something like 60 against 52 on one widely used index. Fifty-two compared with sixty for a 180th of the price sounds like an easy decision. But for serious users, it isn't quite so straightforward, because the value of being right is more important.

Think about what you actually do with an answer. A model that is right 80 times in 100 and a model that is right 95 times in 100 are not 15% apart in usefulness. They sit on either side of an important line. Below it you haven't automated anything, you have just created a proofreading job. Above it you can let the thing run. Almost all of the money in this market is being paid for the last few points before that line, and those are the hardest and most expensive points to buy.

It gets sharper the longer the job runs. A coding agent does not answer one question. It takes hundreds of steps in sequence, and each one has to be right for the next to make sense. Take a modest twenty-step task. A model that is right 95% of the time at each step finishes it cleanly on about a third of attempts. At 99% it finishes on four in five.

Set against that, the token bill is usually trivial. Take a development task that a good engineer would spend a day on. Run it through a cheap model and it costs 50 cents; through a frontier model, $50. If the cheap one produces partially-working code that has to be picked through and fixed, you have spent 50 cents and most of an engineer's day. If the expensive one produces code that is reliable, you have spent $50 and nothing else. That is not a hundredfold difference in price. It is the difference between having bought the work and having bought a not-very-good draft of it.

Price only decides the matter when the task is worth almost nothing if it fails: sorting documents into piles, translating a catalogue, running a (dodgy) character in a chat app. Which is precisely the work the cheap models are winning, and precisely why they are winning it.

OpenRouter's figures show it. Anthropic runs somewhere between 12% and 15% of the tokens and accounts for about half the money spent. OpenRouter's own conclusion is that the closed models get the work worth paying for, and the open models get the work done in bulk.

What the money says

Menlo Ventures2025: The State of Generative AI in the EnterpriseFor all the fears of over-investment, AI is spreading across enterprises at a pace with no precedent in modern software history.

Not surprisingly, establishing who is actually winning involves counting the money. Menlo Ventures has surveyed enterprise spending on model APIs since 2023. In that first year OpenAI had half of it and Anthropic 12%. By the end of 2025 Anthropic had 40%, OpenAI 27% and Google 21%. DeepSeek had 1%. Over the same year the open-source share of enterprise usage fell from 19% to somewhere between 11% and 13%.

Ramp is a large US financial technology company that provides corporate credit cards and financial automation software to businesses. It monitors corporate card spending across more than 70,000 American businesses. In July, 6.1% of them were paying a model-serving platform, the closest available proxy for open-weight use, and 96.4% of the same group were also paying OpenAI or Anthropic.

Meanwhile the American labs have started cutting prices at the frontier. OpenAI took GPT-5.6 Sol from $5 and $30 per million tokens (in and out, respectively) to $4 and $20 in August. Anthropic has held Claude Opus 5 at $5 and $25, the same rate as the two generations before it, while charging double that for Fable 5 at the very top.

That looks like a price war, and it is. But look at who it is with. Even after the cut, the output price of OpenAI's Sol is around 70 times DeepSeek's. Nobody cuts from $30 to $20 to compete with 28 cents. OpenAI cut to $20 because Anthropic was already at $25. The frontier is competing with the frontier, and the cheap end is competing with the cheap end. Those are two different fights, and the chart only sees one of them.

Where China really has won

The Chinese model win is genuine. When a software developer wants a model they can download and run themselves rather than rent from a lab, they now mostly reach for Chinese ones. Chinese models account for 41% of everything downloaded from Hugging Face, the main public library for this sort of thing, and Alibaba's Qwen displaced Meta's Llama as the most-used family in September 2025. Martin Casado of Andreessen Horowitz has said that of the startups pitching his firm which already use open models, roughly 80% are on a Chinese one.

And it isn't just developers experimenting. Airbnb runs Alibaba's Qwen in customer service, and Brian Chesky told Bloomberg why:

Airbnb's support bot runs on thirteen different models and has cut resolution time from about three hours to six seconds. Cursor, the coding tool, built its own Composer model on Moonshot's Kimi (another Chinese model). Interestingly, both companies received letters from Congress in April asking why American businesses were building on Chinese models.

The capability gap is down to months now, not years. Epoch AI puts the average lag behind the American frontier at seven months. And the Chinese models have closed it on a fraction of the money spent: American private investment in AI is running at roughly 23 times China's.

What this means for the valuations

So, the headline here isn't a simple conclusion about winners and losers. If Chinese models were displacing American ones in paid enterprise work, it would show up in the spending data and it isn't.

The real problem is somewhere else, and interestingly it isn't about China. It's that the middle of the range of AI competency is on its way to being worthless. Very few companies download these models and run them on their own hardware, and I don't expect that to change soon. Customers no longer have to buy them from whoever built them. A dozen firms will host Qwen or DeepSeek for you, and since they are all serving identical weights, price is the only thing they can compete on. When anyone can supply the need, ordinary competence stops being something you can charge much for.

What's still worth paying for is being genuinely at the frontier, owning the place where the work actually happens (the seats, the tools people already have open in front of them), and carrying the customer's legal and regulatory risk. That last one is one of the most important when a large company makes a buying decision in this space. Anthropic's October IPO will be the first proper test of whether public markets agree.

A word of caution on the infrastructure argument, which runs that cheaper tokens mean more tokens, more tokens mean more computers, so falling prices are good for the picks and shovels suppliers to this industrial revolution. OpenRouter found that cutting the price by 10% bought only 0.5% to 0.7% more usage. Volume is rising because people keep finding new jobs to hand these systems, not because the price fell. That's still a reason to own the infrastructure. It's just a different one.

What would change my mind

There are a couple of issues that I will be keeping a close eye on that could indicate that this analysis was wrong. They are:

1. If the open-weight share of enterprise API spend in Menlo's survey started to increase. It fell in 2025 as I reported earlier in this piece. If it turns however, the substitution story might start to have legs.
2. If a frontier price started moving towards the Chinese level rather than towards the price of a competing American lab. Sol or Opus at low single-digit dollars per million output tokens would be that signal. $20 against $25 is certainly not.
3. If Ramp's overlap figure started falling. If the businesses using open models started to stop paying OpenAI or Anthropic as well, that would be a sign that they are replacing the frontier models rather than using them for different tasks.

In summary

1. Token share rewards whichever model is cheapest, and Chinese models are 60% to 90% cheaper.
2. Anthropic has 40% of enterprise spending on model APIs against DeepSeek's 1%. Of the American businesses using an open model, 96.4% also pay OpenAI or Anthropic. And when the labs cut prices at the frontier, they cut them to match each other, not to match China.
3. China has won the open weight contest. That costs the American frontier labs the revenue from the most basic tasks, not their customers.

Going right back to the start of this note, my original objection was to the commentary that accompanied the OpenRouter chart. My conclusion is that if you want to know who is winning in the AI industrial revolution, attention should be focused on where the money goes, and what it is buying.

On that measure the American frontier labs are further ahead than they were a year ago, which might surprise those commentators who appear to be too quick to declare Chinese open-model hegemony in the fast-moving AI race.