Chinese open-weight AI models now process the majority of tokens on major API platforms, and Silicon Valley keeps using them

DeepSeek Started the Shift. Chinese Open-Weight Models Now Process Most Tokens on OpenRouter.

Benchmarks tell you which model scores best on paper. Token routing tells you which model developers actually choose.

On OpenRouter, the largest neutral marketplace for LLM API traffic, Chinese-origin open-weight models now process the majority of all tokens routed through the platform 1. The inflection traces to DeepSeek-R1, released in January 2025, which demonstrated that a Chinese lab could match Western reasoning quality and ship the weights 1.

Qwen, Kimi, GLM, MiniMax, and Xiaomi MiMo releases followed, filling every price-performance tier below the U.S. frontier 1.

The shift is concentrated in code. Programming rose from about 11% of OpenRouter usage at the start of 2025 to more than 50% by mid-2026, and Chinese models are disproportionately strong and cheap on coding tasks 1.

Meanwhile, Washington is arguing about theft. Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that Moonshot AI distilled Anthropic's Fable model to develop its K3 model, calling it "stealing proprietary US technology" 2. Commerce Secretary Scott Bessent has suggested the U.S. might impose sanctions on Chinese AI companies 2. Anthropic's chief national security officer Tarun Chhabra identified Zhipu among companies that have allegedly distilled capabilities from frontier American models 3.

The token data sits in tension with the geopolitical framing. Developers are not choosing Chinese models for ideological reasons. They are choosing them on cost and capability. Hugging Face, an open-source AI platform, said it used Z.ai's GLM 5.2 to analyze a cyberattack by OpenAI's GPT-5.6 on its production system because other frontier models refused to assist due to built-in safety guardrails 2. Thinking Machines, founded by former OpenAI CTO Mira Murati, built its first foundation model in part using Chinese models: its architecture drew on DeepSeek-V3, and its post-training incorporated synthetic data from Moonshot AI's Kimi K2.5 3. Apple received approval from China's Cyberspace Administration to launch Apple Intelligence in China using Alibaba's Qwen and Baidu's Ernie models as core components of its AI stack 3.

For builders, the implication is concrete. If your application routes through OpenRouter or a comparable marketplace, your inference stack may now depend on Chinese infrastructure you did not explicitly choose. U.S. export controls on semiconductors remain effective at constraining access to advanced chips, but they are far less effective at limiting the diffusion of open-weight models that can be downloaded and run locally 3.

For investors, the moat thesis for closed-source frontier labs is harder to defend. Artificial Analysis, an independent AI benchmarking company, ranks Moonshot's K3 third in its intelligence index 2. Arena AI, a crowdsourced evaluation platform, ranks K3 first for web development tasks 2.

Routed inference is not the whole market. First-party consumer traffic through ChatGPT and Gemini dwarfs the router, and Anthropic and OpenAI still capture the majority of industry revenue by pricing at a premium and serving enterprises directly 1. OpenRouter is a leading indicator of where developers reach when brand does not dictate the choice.

The market voted eighteen months ago. The bill is coming due, and most of the people holding it do not know what it is denominated in.