Half the AI models you can buy now come from Chinese labs
97 of the 190 commercially priced models we track are Chinese, and Alibaba alone ships 49 of them — more than OpenAI, Google and Meta put together. They are not, however, the cheap ones.
TL;DR — the short version
97 of the 190 commercially priced AI models we track come from Chinese labs. That is 51 %. US labs account for 42, or 22 %.
Alibaba alone ships 49 models — more than a quarter of the entire market, and more than OpenAI, Google and Meta put together.
They are not the cheap option. The median Chinese model costs $1.13 per million output tokens against $0.30 for the median American one. The “cheap Chinese model” story does not survive contact with the price list.
Where they do dominate is memory: 24 of the 33 models offering a million tokens of context are Chinese.
Count the models a developer can actually call today — published price, public endpoint, available now — and the map of who supplies the world’s AI looks different from the one in the headlines.
Half the shelf
Of 190 commercially priced models, 97 come from Chinese labs: Alibaba, Z.AI, DeepSeek, MiniMax, Moonshot, Tencent, Xiaomi and others. Forty-two come from American ones. Fifty-one come from everywhere else, a group Mistral dominates with eighteen models on its own.
The single most striking bar is the first one. Alibaba ships 49 models — 25.8 % of everything on the market. The Qwen family has become the Linux distribution of language models: a version for every size, every budget and every context length, released faster than anyone can benchmark them.
OpenAI appears with four. That is not a measure of importance — it is a measure of strategy. OpenAI sells a handful of models to an enormous number of people; Alibaba floods the shelf. Both work. They just produce very different-looking charts.
The part that contradicts the story
The received wisdom is that Chinese labs compete on price. Our price list says otherwise.
| Origin | Models priced | Median output price | Range |
|---|---|---|---|
| Chinese labs | 97 | $1.13 | $0.13 – $15.00 |
| US labs | 29 | $0.30 | $0.08 – $10.00 |
| Everyone else | 51 | $0.65 | $0.03 – $8.00 |
The median Chinese model costs nearly four times the median American one, and the most expensive model on the entire market is Kimi K3, from Moonshot AI, at $15.00.
There is a real reason for this, and it is not that anyone got more expensive. The American median is pulled down by a large number of small open-weight models — NVIDIA’s Nemotron line, Google’s Gemma, Meta’s Llama — published cheaply or free as ecosystem plays. The Chinese median is pulled up by Alibaba’s flagship Qwen Max tiers. Compare like with like and the gap mostly closes: in the under-fifty-cents band, there are 18 Chinese models against 22 American ones.
Where they genuinely lead: memory
Thirty-three models on the market offer a context window of a million tokens or more. Twenty-four of them are Chinese — six are American, three are from elsewhere.
The median context window follows the same line: 262,144 tokens for Chinese models, 193,536 for American ones, 131,072 for the rest. If your problem is “this whole codebase has to fit in the prompt”, the long-context shelf is where the concentration is real.
Why a market count is not a capability ranking
We are counting models available with a published price. That is a genuine measure of supply — of who is filling the shelf a developer picks from — and it is not a measure of who has the best model, who serves the most tokens, or who makes the most money.
A lab that ships one model used by four hundred million people and a lab that ships forty-nine models used by specialists appear here as 1 and 49. Both facts are true and they answer different questions. This article answers: when you open the model picker, whose work is in it? Increasingly, the answer is a Chinese lab.
Method
190 models with published commercial pricing, collected via OpenRouter from provider price pages and refreshed daily; the snapshot is 7 August 2026. Origin is assigned by the lab that trained and publishes the model, not by hosting location: Alibaba, DeepSeek, Moonshot AI, Z.AI, MiniMax, Tencent, Baidu, ByteDance, Xiaomi, StepFun and InternLM count as Chinese; OpenAI, Anthropic, Google, Meta, NVIDIA, Microsoft, Cohere, xAI, Amazon, Perplexity, Poolside and AI21 as American; everything else, including Mistral, in the third group.
One gap worth naming: a lab appears here only if its models are sold through the price source we read. Anthropic, for instance, has no entry in this count — not because it ships nothing, but because its models are not listed there. Read this as a census of one very large shelf, not of the whole world.
Medians are computed over models with a non-zero output price, which is why the counts in the price table are smaller than the counts in the chart — free tiers are excluded from price statistics but included in the market count. Open-weight models that anyone can self-host appear here only when a provider also sells them as an API.