I could be completely missing something, but how is a Chinese lab training an incredible 3T model (probably same order of magnitude as Fable or GPT 5.6) when they have orders of magnitude less compute than American frontier labs?? For context, the best Nvidia chips that could be exported to China (up until a few days ago) were the H100, H20, and A800. Those are between 66-92% *fewer* FLOPs per watt than the B200/B300s that OAI/Anthropic have had for ~1.5 years. On top of that, sanctions restricted them from buying up nearly the magnitude of compute that the hyperscalars here have Either Huawei has caught up BIG TIME with training chips, Chinese labs smuggled in Blackwells, they have way bigger power budgets in their datacenters, or Kimi has really figured out how to push training efficiency in meaningful ways Looking forward to hearing more, but this is a huge hit to our perceived edge in compute scaling

@teortaxesTex They may never catch up even if US stops since China is in demographic collapse and this was the last time they trained 3T model! 2 more weeks untill 2 year gap! TRUST THE PLAN!