VECTOR WIREAI INTELLIGENCE
NVDA$1,847+3.2%MSFT$512+1.1%GOOGL$199-0.4%META$728+2.7%AMD$184-1.2%TSM$212+0.6%PLTR$98+4.1%AI IDX4,821+1.9%
PKT
SEEDRefresh Models Deals Regulatory Sources

Moonshot Built Kimi K3 on 20,000 Nvidia Chip Cluster Supplied by Alibaba

Bloomberg reports Moonshot used a 20,000 Nvidia Hopper-generation chip cluster supplied by investor Alibaba to build its frontier-class Kimi K3 model.

Vector Wire — AI-assisted editorial illustration

Moonshot built its Kimi K3 model using a cluster of 20,000 Nvidia chips supplied by Alibaba, according to a Bloomberg report1,2. The chips are Nvidia's Hopper generation. Bloomberg's sources said Alibaba is providing the H200 variant specifically, though Alibaba denied that the chips are the leading-edge H200s.

Alibaba holds an investment in Moonshot and is supplying the computing power behind the model. Moonshot introduced itself to a Western audience a couple of weeks ago with the Kimi K3 model, which Bloomberg described as performing near the leading edge of frontier labs such as OpenAI and Anthropic.

The report surfaces several regulatory threads around the compute supply chain. Nvidia is not currently allowed to sell its most advanced Blackwell chips into China. The U.S. recently signed off on Nvidia selling H200 chips into China, but Beijing has pushed back on domestic companies using H200 chips. Chinese companies can legally access Nvidia chips outside China, including in Southeast Asia. The White House has said it believes Moonshot is also accessing Grace Blackwell chips.

ANALYSIS The Bloomberg report fills in a key question that followed the Kimi K3 launch: what compute infrastructure underpins Moonshot's frontier-class results. The answer — a 20,000-GPU Hopper cluster routed through an investor-supplier relationship with Alibaba — illustrates how Chinese AI startups are assembling large-scale training runs despite U.S. export restrictions on the most advanced chips.

The conflicting accounts over whether the cluster uses H200s specifically — Bloomberg's sources saying yes, Alibaba saying no — leaves the exact chip variant unresolved. The distinction matters because H200s carry more memory bandwidth than earlier Hopper SKUs, which directly affects training throughput for large language models.

The regulatory picture is layered: U.S. approval for H200 sales into China, Chinese government resistance to domestic use of those same chips, and the availability of offshore access routes create a complex compliance environment for both Nvidia and its Chinese customers.

CORRECTIONS: none for this article · this piece updates automatically as the story develops · corrections policy & trail →