Moonshot Is Using 20,000 Nvidia Chips From Alibaba to Train Its Kimi Model

Moonshot, the Chinese AI startup behind the Kimi language model, has struck a deal with Alibaba Group Holding Ltd. to use a cluster of about 20,000 Nvidia Corp. chips. Bloomberg reported this on July 31, 2026, citing sources familiar with both companies' operations. The deal underscores how heavily leading Chinese AI labs still depend on Western semiconductors, despite years of US sanctions pressure.

What We Know About the Moonshot–Alibaba Deal

According to Bloomberg, the two sides agreed on access to computing capacity: Alibaba provides Moonshot with GPU infrastructure built on chips from Nvidia Corp., the largest maker of AI accelerators. A cluster of about 20,000 chips is a serious compute base. Large training runs for modern language models require thousands of accelerators working in parallel, and a cluster of this size puts Moonshot alongside first-tier organizations.

• Parties: Moonshot (flagship model: Kimi) and Alibaba Group Holding Ltd.
• Cluster size: about 20,000 Nvidia chips
• Format: computing power agreement, i.e. an arrangement for access to computing capacity
• Details (specific chip models, terms, financial conditions): not disclosed by Bloomberg
• Sources: people familiar with both companies' operations

Moonshot was founded in 2023 and is known for Kimi, which stood out for its ultra-long context support, one of the key competitive advantages at launch. The startup quickly became one of the most talked-about players in China's AI industry and raised substantial venture funding. Kimi competes with Alibaba's Tongyi Qianwen, Baidu's Wenxin and the internationally known DeepSeek, and steady growth in compute is essential to staying in that race.

Why China Still Depends on Nvidia Chips

Moonshot is far from the only Chinese AI company building infrastructure on Western accelerators through cloud platforms. According to Bloomberg, leading local labs continue to rely on American GPUs, drawing on capacity from providers that stockpiled chips before the latest rounds of export restrictions.

Since 2022, the US has steadily tightened export controls. The A100 and H100 accelerators were banned first, followed by their China-market versions, the A800 and H800. As Bloomberg's report makes clear, ways to work with Western chips in China remain, including agreements with cloud providers that hold existing stockpiles. Domestic alternatives, above all Huawei's Ascend accelerator, are gradually scaling up. But a gap remains in the software ecosystem and in real-world performance on large training runs.

"The situation highlights China's continued reliance on Western semiconductors for the development of artificial intelligence,"

Bloomberg writes.

For Alibaba Group, the Moonshot deal pays off on two fronts: the tech giant keeps a promising AI startup inside its ecosystem and monetizes its accumulated GPU fleet. Competition for AI startups among China's leading platforms (Alibaba, Tencent, Baidu, ByteDance) has long been a strategic battleground for future positioning in the technology race.

What This Means

A single startup running a cluster of 20,000 Nvidia chips is clear evidence that the US sanctions strategy has not yet stopped China's AI race. It has only raised the race's cost and organizational complexity. Moonshot holds its place in the first tier, and the Alibaba deal demonstrates a working model of adaptation: partnering with a platform instead of buying banned components directly. The infrastructure gap with domestic alternatives is not yet closed, and the pace of the race is not slowing.