Moonshot trained Kimi K3 on 20,000 Nvidia chips from Alibaba’s cloud

Moonshot trained Kimi K3 on a 20,000-chip Nvidia cluster supplied by Alibaba, Bloomberg reported on Thursday. The chips form a substantial portion of Moonshot’s total compute capacity. Multiple sources said the agreement is specifically for H200 processors, Nvidia’s most powerful Hopper-generation accelerators. Alibaba denied providing H200s but did not deny the 20,000-chip arrangement and declined to say which chips are involved instead.

The dynamic is extraordinary. Alibaba is one of Moonshot’s largest investors and expects backed companies to use its cloud. But Kimi K3, trained on Alibaba’s infrastructure, now outperforms Alibaba’s own Qwen models on key benchmarks. Bloomberg reported that some people inside Alibaba are disappointed that a rival team with access to similar resources produced a better model than their own. Kimi K3 launched as the world’s largest open-weight model at 2.8 trillion parameters, matching Anthropic’s Fable 5 and OpenAI’s GPT-5.6 on several metrics.

The chip sourcing question is politically explosive. White House official Michael Kratsios accused Moonshot of illegally acquiring Nvidia’s most advanced Blackwell chips. A person familiar with Moonshot confirmed the company has a channel for accessing Blackwell processors via Southeast Asia, without specifying whether it involves rental (generally legal) or direct purchase (a breach of US regulations). Moonshot is reportedly seeking additional Blackwells for its next model. The 20,000 Hopper chips from Alibaba are a separate arrangement, but the physical location of that compute is unclear. Chinese cloud companies have sometimes accessed restricted chips owned by foreign firms in overseas locations and re-leased that capacity domestically.

Kratsios also claimed, without providing evidence, that Moonshot used distillation against Anthropic’s Fable model to train K3. Anthropic accused Alibaba’s Qwen lab of running the largest distillation campaign ever against Claude using 25,000 fake accounts. Now Moonshot, Alibaba’s investee using Alibaba’s compute, faces the same accusation. The export controls were designed to prevent Chinese labs from accessing the chips needed to build frontier models. Kimi K3 demonstrates that between Hopper chips from cloud agreements, Blackwell access via Southeast Asian intermediaries, and alleged distillation of American models, the controls have not worked as intended.