China's open models now dominate two developer gateways as Washington opens inquiries

On 26 September, CNBC published numbers that give the clearest measure yet of how far the model market moved this year. OpenRouter routes developer traffic to many providers. There, Chinese models went from a low single digit share in February to a clear majority by mid September. On Vercel, the share rose to 55% in August from 11% in January.

OpenRouter's figures cover companies in the U.S., Europe and what it calls the Global South, 82 countries across Central and South America, Africa and Asia. Vercel did not say how its data breaks down geographically.

The gap that closed

Two years of benchmark releases have narrowed the distance between the top of the U.S. field and the best open weights from China. Most benchmarks still favour the most advanced American models, CNBC reported. But DeepSeek, Z.ai and Alibaba all shipped models this year with large gains on coding and other agentic work. Peter Walker, head of insights at OpenRouter, put the change in plain terms. Chinese open source models released this year can credibly perform in advanced agentic use cases, especially coding, in a way that was not true in late 2025. That is a statement about task completion, not about leaderboard position.

OpenAI and Anthropic both announced cheaper models in the same week CNBC's piece ran. Dianne Penn, who leads product management for research and labs at Anthropic, told the network the company was trying to make answers more efficient, so that they use fewer tokens depending on an effort setting. The timing is not a coincidence.

Cost is not the only variable, but it is the one buyers can measure. A model that completes the same pull request for a tenth of the token spend wins the default slot in an agent pipeline, even if it loses on a hard reasoning benchmark.

Who is actually using them

The geography matters more than the headline share. Businesses in the countries OpenRouter groups as the Global South have been the heaviest users of Chinese models on its system in recent weeks, with 67% of their tokens going to Chinese providers. About half of all tokens on OpenRouter come from U.S. companies, a group that still spends more overall on frontier American models. Daniel Remler, a senior fellow in the technology and national security program at the Center for a New American Security, told CNBC that the integration of Chinese AI models risks pulling countries into a Chinese technology sphere of influence that hardens into geopolitical alignment. He called the trend a real economic and security risk for the United States.

Remler also predicted where the growth goes next. He pointed to Southeast Asia's economic and cultural links with China and its expanding digital infrastructure, and to entrepreneurs and governments from Lagos to São Paulo to Jakarta who want cheap, open models.

"Anywhere from Lagos to São Paulo to Jakarta where entrepreneurs and governments are looking for cheap, open models, will look first to Chinese AI."

That quote, from Remler, is the argument Washington is now acting on. Two U.S. House committees are investigating the impact of rising adoption of Chinese models. The stated concerns are technology competition, security and Beijing's global influence.

Export controls remain the main lever. The U.S. has restricted Chinese AI companies from buying the most advanced chips, and Washington is worried about workarounds: access to Nvidia hardware through overseas data centres, and distillation, where a new model is trained to mimic an older and more established one. Neither tool addresses the price gap directly. A developer choosing a model on OpenRouter is not buying a chip, and a model downloaded from a public repository cannot be un-downloaded by a licensing rule.

There are also domestic pressures pulling the other way. U.S. labs have spent the year defending premium pricing, and the cheap-model announcements from OpenAI and Anthropic in the same week suggest they read the same usage charts.

AI was a major focus when U.S. President Donald Trump and Chinese President Xi Jinping met this week, CNBC reported, which puts the model market inside a diplomatic frame as well as a commercial one.

What the numbers do not show

Token share is not revenue share, and the two diverge sharply. U.S. frontier models still attract more overall spending even where they lose on volume, because the hardest tasks and the largest enterprise contracts stay with them. Nor does a majority token share on two gateways describe the whole market. OpenRouter and Vercel are developer-facing routers, not the full universe of enterprise deployments, and Vercel declined to break its data down by region. The 57% to 67% range CNBC reported for OpenRouter covers a single week in September, not a settled trend line.

What the data does establish is that the default choice for a large and growing set of developers changed within about seven months, and that the change was driven by capability clearing a bar and price falling below it. Benchmarks move slowly. Procurement decisions do not.