AI Lab Moonshot Leases 20,000 Nvidia GPUs from Alibaba Cloud in Major Compute Deal

Chinese artificial intelligence lab Moonshot AI has entered into a significant compute capacity agreement with Alibaba Cloud, securing access to approximately 20,000 Nvidia graphics processing units (GPUs) to support the training and operation of its Kimi family of models. The arrangement, reported by Bloomberg and confirmed by sources familiar with both companies' operations, underscores the intensifying demand for high-performance computing resources among Chinese AI developers amid tightening export controls.

The leased GPUs are said to represent a substantial portion of Moonshot's total compute capacity, highlighting the scale at which the company is operating to develop and deploy its large language models. According to the sources, the chips involved are from Nvidia's Hopper generation, specifically the H200 model. However, Alibaba has pushed back on that characterization, telling Bloomberg that claims the GPUs were H200s are "completely groundless," while declining to specify the exact chips provided under the agreement.

The location of the compute clusters supporting Moonshot's workloads has not been disclosed. Alibaba Cloud operates data centers across numerous regions in China and has been rapidly expanding its international footprint in recent months, giving it broad capacity to host large-scale AI infrastructure. The deal reflects a growing trend of Chinese AI labs relying on cloud providers to secure access to advanced chips, particularly as direct purchases of cutting-edge Nvidia hardware become more complex due to U.S. export restrictions.

Moonshot recently made headlines in July with the release of its Kimi K3 open-source AI model, which boasts 2.8 trillion parameters. The company has claimed the model's capabilities approach those of frontier systems developed by leading U.S. labs such as OpenAI and Anthropic. The new compute agreement is expected to bolster Moonshot's ability to scale its training runs and improve model performance, though the company declined to comment on the Bloomberg report, as did Nvidia.

The deal also carries geopolitical undertones. Michael Kratsios, the White House science advisor, alleged on social media platform X that Moonshot had distilled Anthropic's Fable model to train Kimi K3, and further claimed the lab had "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models." Alibaba operates two data centers in Thailand, a fact that has drawn attention to the region as a potential conduit for advanced chip access. Bloomberg's sources confirmed that Moonshot has a channel to access Blackwell-generation GPUs in Southeast Asia, though they did not clarify whether this was arranged through rental agreements or direct purchases—the latter of which would violate U.S. restrictions on sales of the latest-generation chips to Chinese entities.

Alibaba's role as both a cloud provider and a major investor in Moonshot adds another layer to the relationship, given that Alibaba also develops its own large language models under the Qwen brand. This dual position allows the tech giant to benefit from Moonshot's AI advancements while simultaneously strengthening its cloud business. For the broader industry, the agreement illustrates how cloud platforms are becoming critical intermediaries in the global AI compute supply chain, especially in markets where hardware access is constrained by policy and export controls.