A third of American AI already runs on models its builders can't name
Something quietly shifted this year, and most of the Western AI feed missed it.
Since February, Chinese-origin models have handled more than 30 percent of the tokens routed by US companies through developer gateways, every single week. At the peak it touched 46 percent. A year ago that number averaged 11 percent. In the first half of 2025 it was under 5. This isn't a slow drift. It's a re-plumbing of who actually powers the apps people ship.
The names doing the work aren't household ones in San Francisco. DeepSeek, Qwen, Kimi, GLM, MiniMax. When Zhipu shipped GLM-5.2 in June, it became the fastest-growing model anyone tracked all year: roughly 27 times the daily token volume and 80 times the customer count in its first full week. It lands within a hair of the frontier on agentic tasks at about a fifth of the cost. Vercel's own team put it plainly: price is doing the work. When a task doesn't need the best model, teams route it to the cheapest one that's good enough.
Here's the part worth sitting with. The teams building on these models often can't tell you which lab made them. The model got abstracted behind a gateway, picked by a dropdown and a price. The best model and the known model stopped being the same thing.
That gap is the whole opportunity, and it runs both directions.
Western AI founders are drowning in models and starving for distribution. Forty capable options, no clear way to be the one a developer actually tries. Chinese labs have the opposite problem. They shipped genuinely good work the rest of the world is now running without knowing their names, and almost none of it is legible outside China. One side needs to be seen. The other needs to be understood.
We spent two years watching this exact dynamic in crypto. The projects that won were rarely the ones with the best tech. They were the ones that showed up, got found, and got the first hundred people to actually try the thing. The rooms are different now. The gap is identical.
That bridge, between the labs the West already runs and the West that can't name them, is where we're spending our attention. Not another agent demo. The unglamorous work of making good builders legible to the people who'd use them.
More field notes soon.
English
