China‘s AI startup Moonshot AI built its latest Kimi K3 model using a cluster of around 20,000 Nvidia chips provided through Alibaba, according to a Bloomberg report. The arrangement has given a rare look at the computing power behind one of China’s biggest AI breakthroughs and shows how Chinese companies are finding ways to compete with US rivals despite export restrictions.
Moonshot launched Kimi K3 on July 17, calling it the world’s largest open-source AI model with 2.8 trillion parameters. The company says the model performs on par with leading systems from OpenAI and Anthropic. Its rapid progress has also driven a sharp rise in its valuation, with estimates climbing from about $4.8 billion earlier this year to between $20 billion and $35 billion. The company is also reportedly considering a Hong Kong IPO.
Alibaba played a bigger role than just investing
Alibaba is one of Moonshot AI’s biggest backers, owning about 36% of the company after participating in its February 2024 funding round. It is also the cloud provider supplying the computing infrastructure used to train and run Kimi.
According to people familiar with the companies’ operations cited by Bloomberg, the 20,000 Nvidia chips provided through Alibaba make up a significant portion of Moonshot’s computing capacity for its Kimi models. Alibaba has traditionally encouraged companies it invests in to use its cloud platform.
The partnership has also created an unusual situation inside Alibaba. Bloomberg reported that Kimi has recently outperformed Alibaba’s own Qwen AI models on several benchmarks, even though both teams have access to similar computing infrastructure. That has reportedly disappointed some groups within Alibaba.
Questions remain over the Nvidia chips
The exact Nvidia chips powering Moonshot’s AI models have become the focus of growing attention. People familiar with the matter told Bloomberg that the Alibaba cluster consists of Nvidia’s Hopper-generation H200 chips. These accelerators were used for advanced AI training before Nvidia introduced its latest Blackwell architecture.
The issue has become more sensitive after Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that Moonshot illegally acquired Nvidia’s newer Blackwell chips to train Kimi K3. According to Bloomberg, US officials believe the chips may have been accessed through an unnamed third party in Thailand.
It is also reported that a person familiar with Moonshot’s procurement strategy said the startup has a channel to access Blackwell processors through Southeast Asia. The source did not say whether the chips were rented legally through cloud services or obtained through direct purchases that would violate US export controls.
The report shows that building cutting-edge AI models still depends on enormous computing power, even as companies develop more efficient software and model architectures.
It also shows the growing challenges of enforcing US export restrictions on advanced AI chips. While Chinese companies continue to rely on domestic cloud providers such as Alibaba, reports of access to newer Nvidia hardware through Southeast Asia suggest global supply chains remain difficult to control.