RedMonk: Chinese models hit 63% of OpenRouter enterprise tokens

RedMonk's analysis traces the current situation back to DeepSeek's release of DeepSeek-R1 at the end of December 2024, which it says began serious competition between closed and open-weight models. Since then, the newest open-weight models have closed much of the gap with closed frontier systems.

The economics behind open-weight models remain unsettled. AWS CEO Matt Garman said at 2025's re:Invent that developing open-weight models is expensive and often lacks a clear economic model. RedMonk lays out the variations: IBM sells consulting and software around Granite, Nvidia uses Nemotron to sell more hardware, and Alibaba and Google funnel users toward their clouds. Google and Meta can fund model work from advertising revenue, while Anthropic and OpenAI depend on venture funding and near-trillion-dollar valuations. China backs open-weight models as a route to AI leadership by 2030, including through an $8 billion national fund.

Divergence also shows up in release practice. Chinese model weights typically appear within 10 to 14 days of a model's announcement, while US weights often take over 100 days, if they are released at all. Distillation runs through all of these issues: RedMonk says substantive distillation claims almost always originate in the US and target Chinese competitors, and that concerns about distillation are at least part of why US providers are increasingly reluctant to publish weights.