Chinese AI models dominate corporate platforms, raising alarm in Washington

Rapid expansion on major platforms

By 2026, Chinese artificial intelligence models had captured at least 55% of enterprise usage on the OpenRouter and Vercel platforms, according to CNBC data. This surge was driven by prices up to 12.5 times lower and more competitive performance, prompting U.S. House committees to investigate potential economic, technological, and geopolitical risks.

On OpenRouter, a platform aggregating models from various vendors, Chinese models’ token share jumped from 6–13% in February to 57–67% during the week of September 14. By mid-year, these models accounted for roughly 61% of all tokens processed on the platform.

Vercel did not disclose the geographic breakdown of its data, but growth on the platform was also significant.

Price as a key driver of adoption

The cost difference between Chinese and American models varies by product. DeepSeek V3.1 charges $0.25 per million input tokens, while Anthropic’s Claude 3.5 Sonnet costs $3—a 12-fold difference. In April, DeepSeek made permanent a 75% price cut for its V4-Pro model.

Alibaba’s Qwen Plus starts at $0.40 per million tokens, compared to $5 for Claude Opus, representing a reduction of approximately 12.5 times.

Harpreet Arora, Vercel’s head of agentic infrastructure, says price is the primary factor behind growing adoption. “Chinese models are becoming capable enough for more tasks at a much lower cost. Once a model meets the required quality standard, that price difference becomes irresistible,” he said.

Performance nears Western standards

In May, NIST assessed that DeepSeek V4 Pro was about eight months behind the U.S. technological frontier. In July, the institute found the GLM-5.2 model comparable to a U.S. system released just six months earlier, indicating a narrowing gap.

In interface programming tests, Alibaba’s Qwen3.8-Max ranked fourth in the Frontend Code Arena, behind only Claude Opus 5 at maximum effort and tied with the same model at high effort. This shows that in some specific tasks, Chinese models already outperform Western competitors.

Peter Walker, OpenRouter’s head of insights, noted that open Chinese models released in 2026 can fulfill advanced agentic use cases, especially in programming, in ways that were not possible at the end of 2025. He also emphasized that these models are “incredibly competitive in cost” compared to most U.S. lab systems.

U.S. response and geopolitical restrictions

In response to China’s advances, OpenAI and Anthropic announced new, cheaper models. Anthropic launched Claude Opus 5.5, with a 40% cost reduction compared to the previous model. Dianne Penn, Anthropic’s head of product management, explained that the company aims to make its models’ responses “more efficient, so they use fewer tokens depending on the effort setting chosen.”

The U.S. government imposed restrictions through the 2026 fiscal year defense budget law, requiring the Department of Defense to exclude any AI from DeepSeek or its parent company, High-Flyer, from its systems and prohibiting government contractors from using these tools. Zhipu AI, owner of the GLM model, has been on the U.S. list of entities with presumed restricted access to the country’s technology since January 2025.

Washington also fears that Chinese companies may remotely access advanced Nvidia chips in foreign data centers and use “distillation” techniques to mimic established systems. Two U.S. House committees are investigating the impact of the growing adoption of these models.

The geopolitical concern was discussed at a meeting between Presidents Donald Trump and Xi Jinping, who addressed the balance between technological competition and cooperation in artificial intelligence.

The role of the Global South and the popularity of open models

Companies from Global South countries are the largest users of Chinese models on OpenRouter: 67% of tokens consumed by these companies run on Chinese systems. However, about half of all token volume on the platform still comes from U.S. companies.

Daniel Remler, a senior researcher at the Center for a New American Security (CNAS), notes that Southeast Asia may see significant adoption due to economic and cultural ties with China and the growth of digital infrastructure. “From Lagos to São Paulo, through Jakarta, where entrepreneurs and governments are looking for open and cheap models, they will first look at Chinese AI,” he said.

The Qwen family has already surpassed 1 billion downloads and serves as the basis for about 40% of new derivative models published on Hugging Face, an open-source model-sharing platform.