Moonshot’s Kimi Relied on 20,000 Nvidia Hopper Chips
Bloomberg’s Peter Elstrom reports that Moonshot built its Kimi K3 model partly with compute supplied by investor Alibaba, using roughly 20,000 Nvidia Hopper-generation chips, according to Bloomberg’s sources. Alibaba acknowledges supplying the chips but denies that its compute uses Nvidia’s H200s, while the sources identify the cluster as H200-based. Elstrom argues that the arrangement exposes a limit of US export controls: restricting chip sales into China does not necessarily prevent Chinese AI companies from accessing advanced Nvidia compute.
Alibaba-supplied Nvidia compute is part of Kimi’s training picture
According to Peter Elstrom, Moonshot’s Kimi K3 was built in part on computing capacity supplied by Alibaba, using roughly 20,000 Nvidia chips from the Hopper generation. That arrangement matters because Kimi has recently drawn attention for performance Elstrom described as near the leading edge of models from OpenAI and Anthropic.
Chinese AI companies have demonstrated significant engineering progress before—Elstrom pointed to DeepSeek as the prior year’s standout example—but Kimi’s performance raised a parallel question: what infrastructure enabled Moonshot to build it? Alibaba is both an investor in Moonshot and, Bloomberg reports, a provider of compute used for the model.
The reported number and the hardware designation should be separated. Alibaba did not deny supplying Moonshot with 20,000 Nvidia chips. It denied providing compute powered by Nvidia’s H200 chips and declined to identify the generation of hardware it does provide. Elstrom said Bloomberg’s sources identified the chips as H200s, an advanced product in Nvidia’s Hopper line.
We know that Alibaba has an investment in Moonshot, and what we're reporting in this story is that Alibaba is also supplying a bunch of the computing power behind that model.
The central dispute, then, is not whether Alibaba has supplied substantial Nvidia capacity to Moonshot. It is whether that capacity includes H200s. Elstrom also said Michael Kratsios at the White House had stated that the administration believed Moonshot was accessing Nvidia Grace Blackwell chips, which Elstrom characterized as more advanced than the Hopper hardware at issue in Alibaba’s response.
A July 22 post by Kratsios shown on screen made a separate allegation about K3’s development: that Moonshot had distilled Anthropic’s Fable and developed an internal platform for large-scale distillation against U.S. models. That assertion concerns the model-development process, rather than Alibaba’s reported role in supplying compute, but it places Kimi’s capabilities within a broader U.S. concern about how Chinese firms acquire access to frontier AI inputs.
Export controls restrict sales into China, not all access to Nvidia compute
The policy significance of the Alibaba arrangement lies in the distinction between chip sales, domestic use, and access to computing infrastructure elsewhere. Peter Elstrom described the export-control regime as “complicated” and shifting. Nvidia is not allowed to sell its most advanced Blackwell chips into China, he said. It had previously been able to sell certain Hopper-generation and older chips there, while the U.S. more recently approved sales of H200s into China.
Chinese companies operating in China face a second constraint: Beijing’s own preference for domestic semiconductor capacity. Elstrom said Beijing has pushed back on companies using H200s, even though their use may be possible under the U.S. position, because it wants to support China’s chip industry—particularly companies including Huawei and Cambricon.
That makes the unresolved H200 question politically meaningful. If Moonshot were using H200-powered capacity inside China, it would sit in a space where U.S. rules permit access but Beijing would prefer firms to rely on domestic alternatives. Alibaba’s public statement leaves the precise configuration of its supplied cluster unspecified.
It is still entirely possible, entirely legal for Chinese companies to access all of Nvidia's chips outside of the country.
Elstrom said Chinese companies can obtain access to Nvidia chips and compute in Southeast Asia, including H200s. Current restrictions, in other words, do not necessarily prevent a Chinese firm from using advanced Nvidia infrastructure merely because comparable chips are restricted, unavailable, or politically disfavored inside China.
That overseas-access route is a loophole U.S. China hawks have sought to close, Elstrom said. For now, it remains available. The reported Alibaba–Moonshot relationship therefore illustrates a broader limitation of focusing only on direct shipments into China: frontier-model development can depend on a mix of domestic cloud relationships, permitted hardware, and overseas compute access.