Moonshot AI trained its flagship Kimi K3 large-language model using a cluster of roughly 20,000 Nvidia graphics chips leased from Alibaba Group Holding Ltd. [1].

The deal underscores the immense computational requirements of next-generation AI and highlights the continued reliance of Chinese developers on Western semiconductor hardware to remain competitive.

Moonshot, headquartered in Beijing, secured the hardware through a computing-power agreement with Alibaba [1]. The startup utilized Alibaba's cloud data-center facilities in China to house the cluster [1]. This infrastructure provided the necessary scale to develop Kimi K3, positioning the model as a direct competitor to leading AI systems developed in the U.S. [2].

The disclosure of the arrangement came on July 31, 2026 [1]. By leveraging a cluster of 20,000 chips [1], Moonshot was able to bypass the prohibitive cost and logistical challenge of purchasing and maintaining such a vast amount of hardware independently.

Industry analysts said that the partnership reflects a broader trend in the Chinese tech sector. While the government encourages domestic chip production, the current performance gap between local hardware and Nvidia's GPUs often necessitates these types of leasing agreements. The scale of the Kimi K3 project demonstrates that high-end AI training still requires massive concentrations of GPU power, often in the tens of thousands, to achieve state-of-the-art results [2].

Alibaba's role as a provider of this compute power further cements its position as a critical infrastructure layer for the Chinese AI ecosystem. By offering high-density chip clusters to startups like Moonshot, Alibaba enables the rapid iteration of models that would otherwise be impossible for smaller firms to build [1].

“Moonshot AI trained its flagship Kimi K3 large-language model using a cluster of roughly 20,000 Nvidia graphics chips”

This partnership illustrates the critical bottleneck in the global AI race: the availability of high-end compute. Despite efforts to develop sovereign silicon, the use of 20,000 Nvidia chips by a top Chinese startup confirms that U.S.-designed hardware remains the gold standard for training frontier models. It also signals a shift toward 'AI-as-a-Service' infrastructure, where a few cloud giants control the physical hardware that determines which startups can successfully scale their models.