Chinese models drive OpenRouter growth

This shows OpenRouter is becoming less a storefront for premium U.S. models and more a price discovery market for interchangeable intelligence. When Chinese open-source models jump from the margins to more than half of usage, it means developers are using routing to buy enough quality for each task, not the most prestigious model by default. That lifts token volume fast, but because OpenRouter takes a roughly 5% cut on cheaper inference, revenue per token falls as mix shifts downmarket.

• The usage shift is concrete, not abstract. Chinese labs now account for 50%+ of OpenRouter usage, up from about 2% in mid-2025, led by DeepSeek at 3T tokens per day, Tencent at 1.8T, Xiaomi at about 760B, and Z.ai at about 570B.
  1 sacra
• OpenRouter works because a developer can plug into one API and swap models without rewriting their app. That matters more when open models are good enough. DeepSeek in particular is easy to drop in because its API is compatible with existing OpenAI style workflows and supports tool calling for agents.
  2 sacra 5 sacra
• This is the same market split seen across AI gateways more broadly. OpenRouter acts like a wholesale token marketplace taking about a 5% brokerage fee, while companies like Kong sell internal routing, caching, and policy controls to enterprises that want to steer work to the cheapest acceptable model behind the firewall.
  1 sacra 3 sacra 4 sacra

The next step is a larger share of AI traffic moving to brokers and gateways that choose models dynamically in the background. If Chinese open-source labs keep improving on reasoning and agent workflows while staying cheap, OpenRouter can keep compounding volume growth, but the prize shifts from owning the best model to owning the routing layer where model choice, usage metering, and margin control happen.