OpenRouter data: lower-cost Chinese AI models made by companies such as DeepSeek and MiniMax have overtaken their US rivals in token consumption since February
Context & Ripple Effects
Chinese model makers had already been pursuing lower training costs under export controls, including through smaller training data sets. DeepSeek then positioned its open-source V3 as competitive while using fewer chips to train in an earlier account of its training approach.
This usage shift gives observable marketplace weight to a strategy that Chinese firms had broadened through low-cost services and models after DeepSeek-R1's debut in the earlier price-and-release wave. It matters because OpenRouter usage reflects what model users actually route, not only benchmark claims.
First-order effects
- DeepSeek, MiniMax and other lower-cost Chinese providers gain a concrete distribution signal: their models are now consuming more tokens on OpenRouter than US rivals.
- US model providers face immediate pressure to justify higher-priced usage with clearer capability, reliability or product-integration advantages; OpenRouter users gain stronger evidence to test and switch among suppliers.
Second-order effects
- Competition shifts toward price-performance at inference: providers that cannot match the Chinese models' cost per token will need to cut prices, improve efficiency, or target workloads where their models retain an advantage.
- OpenRouter's role as a routing layer becomes more strategically valuable as buyers use it to compare and allocate demand across a broader cross-border model supply base.
Third-order effects
- If usage persists, frontier-model competition may become less concentrated around a small set of US providers and more dependent on efficient, open or readily accessible model alternatives.
- The result would strengthen buyer leverage over model vendors, while making access, distribution and inference economics—not training scale alone—more consequential competitive bottlenecks.
The trend: This is one data point in the shift from a frontier-model race defined by training scale to a buyer-driven market organized around cost-effective inference and multi-model routing.