Moonshot’s Kimi Built on 20,000 Nvidia Chip Cluster From Alibaba

Chinese AI champion Moonshot has a computing power agreement with Alibaba Group Holding Ltd. for the use of around 20,000 Nvidia Corp. chips, underscoring China’s continued reliance on Western semiconductors to fuel its AI development, according to people with knowledge of the companies’ operations.

That hardware forms a substantial chunk of the overall computing capacity Moonshot uses for its Kimi models, said the people, who requested anonymity to describe a confidential deal. The Moonshot team burst onto the global stage earlier this month with the debut of Kimi K3, an open-source AI model that goes nearly toe-to-toe with the leading models from the likes of OpenAI and Anthropic on some metrics.

Moonshot, Alibaba and Nvidia declined to comment on the arrangement.

Alibaba is one of Moonshot’s largest investors and expects the companies it backs to use its cloud, the people said. But the two Chinese firms are also competitors — and now Kimi, using infrastructure provided by Alibaba, has surpassed Alibaba’s own Qwen products in important performance metrics. That’s been a source of disappointment for some people at Alibaba, according to one person familiar with internal talks, given that these rival AI teams have access to similar training resources.

Kimi K3 sparked concerns from Washington to Wall Street to Silicon Valley that Chinese AI companies are closing the gap with US frontier labs, a moment with echoes of Deepseek’s breakthrough last year. Much of the worry centers on the question of hardware.

Read More: China AI Sensation Moonshot’s Gamble on Big Models Pays Off

The US has for years restricted China’s access to Nvidia chips — and to the tools needed to make its own AI processors — all while American firms are plowing hundreds of billions of dollars into data centers filled with these coveted accelerators. And yet Moonshot managed to build K3 anyway. That ignited a debate among investors about the economics of the Western AI infrastructure buildout, and among policymakers about the effectiveness of the US government’s approach.

But according to a White House official, Moonshot skirted the rules. Michael Kratsios, who directs the office of science and technology policy, last week accused the company of illegally acquiring Nvidia’s most advanced Blackwell chips, which are the most powerful AI accelerators on the market. It’s unclear how many of these processors the US government believes Moonshot has.

Sales of these chips to Chinese firms have always required US permission, and President Donald Trump has explicitly said he doesn’t want China buying accelerators from Nvidia’s Blackwell lineup. US officials also believe Deepseek has Blackwell chips, Bloomberg has reported.

Read More: US Official Says Moonshot Accessed Banned Nvidia Chips

Yet the US rules generally allow Chinese firms to access these very same chips via compute rental agreements. Kratsios said Moonshot did this, too, via an unnamed party in Thailand — and that the company likely used these accelerators for training K3. He also claimed, without providing concrete evidence, that Moonshot used a technique called distillation to train K3 using outputs generated by Anthropic’s top Fable model. That would help compensate for more limited computing resources.

A person familiar with Moonshot’s procurement strategy confirmed that the company does indeed have a channel for accessing Blackwell processors via Southeast Asia. They didn’t specify whether this was a rental channel — which is legal, in most cases — or direct purchases, which are a breach of US regulations. The Information reported this week that Moonshot is seeking additional Blackwell processors to train its next model.

The 20,000 chips Moonshot accesses via Alibaba, meanwhile, are from Nvidia’s earlier generation of Hopper products, the people familiar with the agreement said. Hoppers were the industrial standard for training and running AI models until Nvidia launched Blackwells.

The top processors from that lineup have always required a US export license to ship to China, while less advanced models — the H800 and H20 — were at one point legal to ship freely but now need a green light from Washington.

Multiple people close to Moonshot said the agreement is specifically for H200 chips, the most powerful Hopper accelerators. An Alibaba spokesperson denied that the company provides H200-powered compute to Moonshot. They did not deny that the company supplies 20,000 Nvidia chips of compute capacity to Moonshot, and declined to comment on what chips — if not H200s — form the basis of the agreement.

The US has issued licenses for some direct H200 sales to China, a major policy reversal under Trump. But a senior US official recently said only a “trivial” number have actually been shipped with those permits. Beijing has been hesitant to grant approval for domestic firms to purchase the chips and has pressured them to use Chinese-made chips, like those from Huawei Technologies Co.

Chinese cloud companies have sometimes worked around the US controls by accessing restricted chips owned by foreign firms in overseas locations — which is generally allowed — and then re-leasing that compute as part of their own cloud. The physical location of the compute that Alibaba provides to Moonshot is unclear.