Moonshot's Kimi Runs on 20,000-Chip Nvidia Cluster From Alibaba

The most interesting line in Bloomberg's report on Chinese AI startup Moonshot isn't the size of the cluster, though 20,000 Nvidia GPUs is a lot; it's who owns the cluster and who owns Moonshot. Both answers point to the same company. Alibaba took a 36% stake in the startup during a February 2024 funding round and also provides the cloud infrastructure the Kimi models run on.

Sources told Bloomberg the accelerators in that cluster are Hopper-generation H200s, though an Alibaba spokesperson denied providing H200-based computing power while not disputing the 20,000-chip agreement itself. That distinction matters because which Hopper variant Chinese firms can legitimately access is exactly what US export policy has been trying to nail down.

Why this matters beyond one company: Moonshot released its open-source Kimi K3 model earlier this month to real attention, going toe-to-toe with OpenAI and Anthropic on some metrics. If a Chinese lab can front-run US frontier labs on open weights while training on Western silicon rented from its own cloud investor, the export-control story becomes less about smuggling and more about which structures policy actually catches. The Bloomberg piece lands days after White House OSTP Director Michael Kratsios publicly alleged Moonshot had accessed banned Nvidia GB300 servers via Thailand, and after separate reporting that Moonshot is now pushing for more Blackwell chips for the next model, Kimi K4.

The honest caveat is that most of the specific hardware claims here are single-sourced to unnamed people, the H200 identification is actively contested by Alibaba, and the reporting does not put a dollar value on the compute contract or a date on when the chips were provisioned. What the reporting doesn't give you is the split between Hopper variants, whether the arrangement predates or postdates the tightening of US rules, or how much of Alibaba Cloud's own capacity this cluster represents. What it does give you is the shape: a top-tier Chinese open model built on Western GPUs, provisioned through a hyperscaler that is also an equity holder, at the exact moment Washington is trying to make that pattern harder.