Chinese AI models went from a sliver to nearly half of token consumption on OpenRouter in about a year. "American startups are switching to Chinese models" is still overstated.
First the honest caveat, because it matters and most people sharing this chart skip it. This is OpenRouter data only. OpenRouter is a platform developers use to test and route between many models, and it's heavily self-selected. If you're using a US frontier model seriously, you usually go direct: Anthropic's API, OpenAI's API, AWS Bedrock, Google Vertex. You don't route it through OpenRouter. Meanwhile, a lot of people who want to use Chinese models specifically avoid the Chinese labs' own servers for privacy reasons, so they access them through OpenRouter instead. The result: this particular platform systematically undercounts US model usage and overcounts the Chinese share.
So the chart is not "half of all AI usage is now Chinese." It's "on the platform where people shop around between models, Chinese models have become a huge share."
A year ago, Chinese open-weight models barely registered even on OpenRouter. Now they're consumed at the scale of trillions of tokens per week. Developers are running real workloads (aka a lot of it coding) on DeepSeek, Qwen, Kimi, GLM, and the rest. The models got good enough, and cheap enough, that for many tasks the price-performance is impossible to ignore.
The US frontier labs mostly went closed and premium: best models, accessed through paid APIs, priced for the value they deliver. The Chinese labs largely went open-weight and cheap: release the weights, let anyone run them, compete on price. Open plus cheap drives adoption the same way a $200 LiDAR unit or a $299 home battery drives adoption. You commoditize the thing the incumbent is charging a premium for.
The core question stands: which layer captures the value? If frontier intelligence becomes a cheap, open, interchangeable commodity that developers route between based on price, that's wonderful for everyone building on top and brutal for anyone trying to earn a margin on the model itself.
It's the same dynamic as solar panels. The panels got so cheap that the money moved to installation, integration, and the systems built around them. If AI models follow that curve, the value moves up the stack to the products, the orchestration, the proprietary data and workflows.
One more note on the spike: part of the March 2026 surge was a Chinese model offered free for a period, which inflates the numbers. Free is the most aggressive price point on any cost curve, and it's a familiar tactic for buying adoption fast.
The takeaway isn't "China won AI." The data is too biased to say that. The takeaway is that Chinese labs have made frontier-adjacent intelligence cheap and open at a speed that mirrors every other technology they've commoditized, and the open-weight cost curve is now bending in AI the way it bent in solar a decade ago. Whether that ends with intelligence as a cheap commodity and the value somewhere else is the trillion-token question