Chinese AI models take majority token share on two gateways, but reach has limits

Chinese AI models moved from a small slice of developer traffic to a majority of tokens on two major gateways in 2026, according to usage data reported by CNBC. Across the company groups CNBC examined on OpenRouter, Chinese models accounted for 57% to 67% of tokens in the week of Sept. 14, up from 6% to 13% in February. Those are platform-supplied figures reported by CNBC, not an independent census of all model use. On Vercel’s AI Gateway, CNBC reported a 55% Chinese-model token share in August, versus 11% in January.

That is a meaningful shift in production usage, but it is not the same as saying Chinese models have “won” the market. These gateways sit between applications and model providers, so they reveal what teams are choosing when cost, latency and quality matter. They do not measure every enterprise or every region, and CNBC also reported that U.S. frontier models still attract more overall spending.

Why the traffic moved

The reported explanation is a combination of relatively low prices and improving performance on some tasks; the sources do not establish a universal price or quality trend across Chinese models. CNBC reported that lower prices and stronger performance are helping drive adoption, especially for coding and other agentic workflows. Peter Walker, OpenRouter’s head of insights, told CNBC that Chinese open-source models released this year can credibly handle advanced agentic use cases, especially coding, in a way that was not true in late 2025.

For teams routing many requests, the cost of a model that meets their own quality threshold can change the economics of a workload. Vercel’s data on falling average token prices is consistent with that incentive, though it cannot prove why any one customer switched.

Vercel’s own AI Gateway Production Index supports the broader pricing story. In its September report, Vercel said open-weight models ran 56% of gateway tokens in August, the first time that class reached a majority, and that the average token price fell 23.2% that month. The report is not China-specific, but it shows the same pattern: production users are shifting high-volume work toward cheaper models while reserving frontier systems for harder tasks.

Editorial inference: the two gateway reports together suggest Chinese models are not just being evaluated in labs or hobby projects; they are being used for enough routine production work to matter operationally.

Who should care

AI product teams, platform engineers and procurement leads have the clearest reason to pay attention. If a coding assistant, workflow agent or internal tool can meet quality thresholds on a lower-cost model, the savings can be immediate. That changes routing policy from a purely technical choice into a spending decision.

The data also implies segmentation rather than wholesale replacement. CNBC reported that U.S. frontier models still attract more overall spending, while Vercel said Anthropic accounted for 64% of spend on its gateway in August. The token and spending measures therefore describe different parts of the market; they do not identify the exact tasks assigned to each model class.

Geography appears to matter too. CNBC said OpenRouter’s data covered companies in the U.S., Europe and what it defines as the Global South — 82 countries across Central and South America, Africa and Asia — and that businesses in that category were the biggest users of Chinese models on the platform in recent weeks. Vercel did not disclose a geographic breakdown. That leaves an important boundary in the evidence: the data does not show a complete global map or why each customer chose a model.

Policy is becoming part of the product decision

The commercial shift is running into security scrutiny in Washington. In a July 31 statement, the House Committee on Homeland Security and the House Select Committee on China said they were continuing a joint investigation into the national security, cybersecurity and economic implications of U.S. companies integrating PRC-developed open-weight models into consumer platforms, enterprise software and systems with access to sensitive information. The committees said they were examining model provenance, training methods and whether capabilities were advanced through unauthorized distillation from U.S. systems.

That does not prove a problem in every deployment. But it does mean model choice is no longer only a benchmark or price conversation. For companies handling customer data, internal code or regulated workflows, the decision now includes governance questions: where the model came from, what data it may have absorbed, and how much control the company has over downstream exposure.

The central limitation remains important. OpenRouter and Vercel are valuable windows into production usage, but they are not the entire market. CNBC’s platform-supplied figures put Chinese models above half of token volume on two gateways, while U.S. frontier labs still capture more spend. That combination is the real story: substantial observed gateway use alongside a large premium-model market, with incomplete visibility into geography and task mix. Watch whether Chinese-model share stays above half while U.S. spend remains dominant; that split shows when cheaper models are winning routine production work and when frontier systems still justify their premium.