DgcityEven under the world’s most scrutinized technology embargo, Nvidia’s most powerful artificial intelligence chips are finding their way to China’s leading AI labs. Moonshot AI, the Beijing company behind one of the country’s most capable language models, leased a cache of top-end processors this summer, regulatory documents reveal.
The contract, signed in August, covers 32 Super Micro servers built around Nvidia B300 GPUs, the cutting-edge accelerators at the heart of the current generation of large model training. Financing filings made with Beijing regulators locate the machines inside China. The servers are owned by Lingmeixin, a Beijing technology company, and rented to Moonshot rather than sold outright.
The arrangement cuts straight to the contradiction of America’s China chip strategy. Rules prohibit B300-class processors from being shipped to China without explicit American approval, designed to keep the most powerful hardware out of Chinese labs. Yet the filings show that a company closely watched as a national champion was able to access precisely that equipment domestically.
How the servers arrived remains the critical question, and one the documents do not answer. Possibilities include re-routing through third countries, purchases made before the latest restrictions tightened, or licensing decisions taken quietly in Washington. Whatever the route, the disclosure will intensify scrutiny of both export-control enforcement and the booming grey market for high-end GPUs.
The rental model itself is revealing. Instead of buying chips outright, Chinese AI firms increasingly lease computing power from intermediaries that own the hardware. That structure spreads financing costs, obscures supply chains and lets young model-makers present lower capital expenditure to investors even as they train on the best silicon available.
For Moonshot, the stakes are domestic as well as international. It is locked in a fierce race with DeepSeek, MiniMax, Alibaba’s Qwen team and Baidu, where model quality depends heavily on the compute available for training. Falling behind on chips can mean falling behind in products that matter for everything from chatbots to enterprise software.
Beijing is pouring money into domestic alternatives, and Chinese accelerators are improving. But the B300 lease is a quiet admission: for now, even China’s most ambitious AI company would rather rent American silicon than wait for a homegrown replacement.
