China's DeepSeek, MiniMax, and Xiaomi Are Crushing US Rivals on This Surprising AI Metric
Chinese AI labs now serve over 12 trillion tokens combined, outpacing US models.
Chinese AI models have overtaken US counterparts in actual usage, measured by token consumption on OpenRouter. DeepSeek V4 Flash, MiniMax M3, and Xiaomi's MiMo-V2.5 collectively account for over 12 trillion tokens since early 2024. DeepSeek leads with 4.63 trillion, MiniMax at 4.13 trillion, and Xiaomi at 3.8 trillion. This shift is driven by lower energy costs and more efficient model architectures, allowing Chinese labs to offer significantly cheaper per-token pricing.
The Financial Times analysis highlights that this cost advantage is structural, not temporary. Chinese AI labs benefit from government subsidies on electricity and highly optimized inference pipelines. While US models like GPT-4 and Claude still lead in raw benchmark performance, the market is voting with usage: developers and enterprises are choosing the cheaper alternatives for production workloads. This trend could erode US dominance in AI deployment and shift the center of gravity for applied AI to China.
- DeepSeek V4 Flash leads with 4.63 trillion tokens consumed since early 2024.
- MiniMax M3 and Xiaomi's MiMo-V2.5 follow with 4.13T and 3.8T tokens respectively.
- Cheaper energy and efficient models give Chinese labs a decisive cost-per-token edge over US rivals.
Why It Matters
Cost-efficient Chinese AI models are winning production usage, threatening US leadership in applied AI deployment.