Moonshot AI has secured a computing-power agreement with Alibaba Group Holding Ltd. to access a cluster of approximately 20,000 Nvidia chips [1].
The deal underscores the critical dependence of Chinese AI champions on Western-made semiconductors to train next-generation models. Despite ongoing U.S. export curbs on high-end hardware, the ability to leverage cloud infrastructure allows companies to maintain a competitive pace in the global AI race.
Moonshot AI, based in Beijing, is utilizing the hardware to train its Kimi K3 model [2]. The specific hardware involved in the cluster consists of Nvidia H200 chips [4], which provide the massive computational power required for the complex training processes of large language models.
Alibaba provides this access through its cloud infrastructure, acting as the intermediary that hosts the chip cluster [1]. This arrangement allows Moonshot AI to scale its operations without needing to own and maintain the physical hardware directly.
The scale of the cluster, roughly 20,000 chips [1], reflects the immense resource requirements for the Kimi K3 project. By partnering with Alibaba, Moonshot AI can bypass some of the logistical hurdles associated with procuring and deploying high-end GPUs in the current regulatory environment.
Industry observers said such partnerships are becoming a primary strategy for Chinese firms to circumvent hardware shortages. The use of H200 chips [4] specifically suggests a need for high-memory bandwidth to handle the vast datasets associated with the K3 model's development.
“Moonshot AI has secured a computing-power agreement with Alibaba Group Holding Ltd.”
This agreement demonstrates that cloud-based access remains a viable pathway for Chinese AI developers to utilize restricted US technology. By leveraging Alibaba's existing infrastructure, Moonshot AI can deploy thousands of high-end GPUs that would be difficult to acquire as standalone units under current export restrictions, signaling that the 'compute gap' may be narrower than hardware shipping data suggests.
