Chinese AI models now process most of the traffic on two widely used developer gateways, OpenRouter and Vercel, according to CNBC data cited by PhoneWorld Saturday. The rise, led by open models from DeepSeek, Z.ai and Alibaba, comes as U.S. lawmakers look more closely at Chinese AI.

What Happened

On OpenRouter, a service that lets developers send requests to models from many providers through a single interface, Chinese models accounted for 57% to 67% of tokens in the week of Sept. 14, 2026. In February, the figure was 6% to 13%. On Vercel's AI gateway, their share reached 55% in August, up from 11% in January.

OpenRouter also breaks usage down by region. In what it calls the Global South, a group of 82 countries across Asia, Africa and Central and South America, 67% of the tokens businesses use go to Chinese models. The United States still accounts for about half of all tokens on the platform, so the shift is not only happening overseas.

RelatedNew York City Comptroller Urges Anthropic to Delay IPO Over Safety and Governance Concerns

The gains are concentrated in coding and agentic tasks, where a model plans and carries out multi-step work. Peter Walker, OpenRouter's head of insights, said Chinese open models can now handle advanced agentic work, particularly coding, in ways they could not in late 2025. Harpreet Arora, head of agentic infrastructure at Vercel, said businesses are more likely to pick a cheaper model once it reaches the quality a task requires.

Why It Matters

Gateway data shows what developers actually pay to use, not just which model tops a benchmark. Tokens are the unit that providers bill by, so a majority share means Chinese models are now doing much of the routine paid work, such as code generation and agent runs, that passes through these services.

U.S. labs still lead many frontier benchmarks, but Arora's comment points to how buyers behave in practice: for a large class of jobs, the top model is not required. Open weights add to the appeal. Companies can run models such as DeepSeek's or Alibaba's Qwen family on their own hardware or through third-party hosts, which lowers cost and reduces dependence on a single vendor.

The data has limits. It covers only traffic that flows through OpenRouter and Vercel. Enterprises that buy directly from OpenAI, Anthropic or Google, or that host models privately, are not counted, so the figures describe a developer-heavy slice of the market rather than all AI spending.

Concerns in Washington

Daniel Remler, a senior fellow at the Center for a New American Security, said Chinese AI could create economic and security challenges for the United States, and raised concern about China's growing technological influence abroad.

Two House committees are investigating the rise in adoption of Chinese models. U.S. officials are also examining two ways Chinese labs may be keeping pace despite export controls: gaining access to Nvidia chips through data centers located outside China, and distillation, in which a model is trained on the outputs of a more capable one. Washington has restricted sales of Nvidia's most advanced accelerators to China since 2022, which is why remote access through overseas facilities has become a policy target.

What to Watch

Watch for hearings or findings from the House committees, and for any federal move to limit U.S. companies' use of Chinese models or to close off remote chip access. The other open question is whether the share holds as U.S. labs ship new models and cut prices, or whether the cost gap keeps developers on Chinese open weights for coding and agent workloads.