chinese models topped us models in token usage on openrouter in march and have been up only since. how real are the chinese labs, and what do they mean for the frontier moats? this and more in delphi research's latest report from @accelxr on evaluating the china labs. x.com/ceterispar1bus/status/20…
China’s AI hardware disadvantage is turning efficiency into a competitive advantage. Chinese labs have spent the past several years trying to keep pace with the frontier under tighter restrictions on Nvidia’s most advanced chips. Restricted access made it harder to improve models by simply adding more compute. Labs had to look for gains in how models were trained and run. Much of that work focuses on improving chip utilization during training and designing models that require less compute and memory. Post-training methods are also extracting more reasoning capability from existing base models. Chinese AI progress cannot be reduced to distillation. These systems change how models are built and run while lowering training and serving costs. That cost advantage changes how models compete. Many workloads do not require the strongest model available. A lower-cost option can win once it performs well enough for the task. Several Chinese models already combine competitive performance with substantially lower prices. China is much closer to running competitive models on domestic chips than it is to training them from scratch. Frontier training still depends on advanced hardware China cannot yet fully replace. The next generation of Chinese models will show how far better software and more efficient architectures can narrow that gap.
· 310K Views
