Kimi Trained on Alibaba's Cluster of 20,000 Nvidia Chips
Claire Weston
Moonshot AI has access to roughly 20,000 Nvidia chips through a confidential deal with backer Alibaba, Bloomberg reports — but the chip model and legality of the arrangement are drawing scrutiny from both Washington and Beijing.
Where do the chips come from, and how many?
People familiar with the matter say Moonshot AI (月之暗面) signed a confidential computing deal with Alibaba, gaining access to about 20,000 Nvidia chips.
This means → Moonshot did not build its own cluster. It trained Kimi on its investor's cloud infrastructure — a common but rarely confirmed arrangement among Chinese AI startups.
Alibaba is a major Moonshot investor and typically requires portfolio companies to use its cloud services, tying compute access to the equity relationship.
Why has the chip model become a flashpoint?
Several people close to Moonshot say the deal involves Nvidia H200s — the most powerful chip in the Hopper family.
Alibaba denied supplying H200s but did not deny providing about 20,000 Nvidia chips, and did not specify the actual model.
In plain terms = H200s have always required a U.S. export license for sale to China. Washington has issued some licenses, but a senior U.S. official said actual deliveries have been "minuscule." The model dispute is really a compliance dispute.
What is the Blackwell angle?
White House science-and-technology adviser Michael Kratsios publicly accused Moonshot of illegally obtaining Nvidia's latest-generation Blackwell chips.
He said Moonshot accessed Blackwell chips through an unnamed entity in Thailand via a compute-rental arrangement, possibly for K3 training. He also claimed Moonshot used distillation — training a model with the outputs of a stronger model — drawing on Anthropic's top model, but offered no specific evidence.
A person familiar with Moonshot's procurement confirmed a Southeast Asian channel exists but did not say whether it was a lawful rental or a prohibited purchase.
How should we read the investor-vs-competitor tension?
Kimi K3 has outperformed Alibaba's own Qwen series on multiple benchmarks.
A person aware of internal discussions said the result disappointed some inside Alibaba — both teams had access to comparable training resources, yet the portfolio company came out ahead.
This reflects a structural tension: Alibaba is simultaneously Moonshot's funder, compute supplier, and direct competitor. The more resources they share, the more awkward the rivalry becomes.
What does this mean for the U.S.–China AI race?
Since its early-month launch, Kimi K3 has approached the top models from OpenAI and Anthropic on several benchmarks, fueling debate across Washington, Wall Street, and Silicon Valley about a narrowing U.S.–China AI gap.
This means → the controversy over Moonshot's chip sources is effectively a stress test of U.S. export controls — the rules exist, but chips are still flowing into Chinese AI training through multiple channels.
Beijing is simultaneously pressuring domestic firms to prioritize Huawei and other homegrown chips, keeping the policy tug-of-war alive on both the supply and demand sides.
Content is for reference only, not financial advice.