Moonshot AI taps Alibaba's 20,000 Nvidia GPUs to train Kimi K3 model

Chinese AI startup Moonshot AI used a computing cluster supplied by Alibaba that contains roughly 20,000 Nvidia GPUs to train its new Kimi K3 model. The hardware is reported to be Hopper‑generation accelerators, most likely the H200 chips, and the model has shown performance advantages over Alibaba’s own Qwen series despite comparable compute resources.

U.S. officials have accused Moonshot of circumventing export‑control rules by accessing the latest Blackwell‑generation GPUs through cloud services routed via Southeast Asia, although the Alibaba cluster itself is said to use the older Hopper architecture. Alibaba has denied providing H200 chips directly to Moonshot but confirms the existence of a large GPU cluster. The episode underscores ongoing tensions between the United States and China over AI hardware and the limits of current export‑control regimes.