Moonshot AI trained its Kimi K3 large-language model using a cluster of approximately 20,000 Nvidia GPUs supplied via Alibaba Group Holding Ltd. [1]
This partnership allows the Beijing-based firm to access the massive compute power necessary to compete with U.S. AI models. By utilizing Alibaba's cloud data-center infrastructure, Moonshot can scale its operations while sidestepping U.S. export restrictions that limit the direct purchase of high-end Nvidia chips by Chinese firms [1, 5].
The deal, reported on July 31, 2026, highlights the critical role of cloud providers in the global AI arms race [1]. While some reports suggested the lab built a data center without Nvidia hardware, primary reporting from Bloomberg and Reuters confirms the Kimi K3 model relied on the Alibaba-hosted cluster [1, 4].
The scale of the compute provided by Alibaba has enabled Kimi K3 to outperform some of Alibaba's own internal efforts. Specifically, the Kimi K3 model has shown the ability to outperform Alibaba's Qwen models, despite those models having access to the same underlying compute resources [3, 5].
This arrangement underscores a growing trend where Chinese AI startups leverage the existing infrastructure of tech giants to accelerate development. The use of roughly 20,000 chips [1] places Moonshot in a higher tier of computational capacity, narrowing the gap between domestic Chinese models and those developed in the U.S. [5].
“Moonshot AI trained its Kimi K3 large-language model using a cluster of approximately 20,000 Nvidia GPUs”
The collaboration between Moonshot and Alibaba illustrates a strategic shift in the Chinese AI ecosystem. By using 'compute-as-a-service,' startups can bypass hardware sanctions and achieve state-of-the-art performance without owning the physical chips. This creates a paradoxical situation where cloud providers may inadvertently fuel the growth of competitors who eventually outperform their own proprietary models.
