Moonshot AI's 32 B300 Server Lease Exposes U.S. Export Control Gaps

By Grace Kim

A financing contract for 32 servers carrying Nvidia's Blackwell Ultra B300 graphics processing units (GPUs), valued at 432 million yuan, names Moonshot AI as lessee in a deal that likely breached U.S. export controls on advanced AI chips.

Documents obtained by Bloomberg on the 10th show Moonshot AI, one of China's leading AI model developers, signing the contract in August to lease the servers from Super Micro Computer. The filing lists their storage location as Qingyang in Gansu province and their combined value at 432 million yuan, about 86.5 billion won.

The B300 is the highest-performing AI accelerator currently available. Its per-server valuation runs 350% above the prices at which comparable units trade in the United States. Bloomberg read that gap as evidence that export controls have produced a black market rather than closed one.

The disclosure came as Nvidia CEO Jensen Huang acknowledged an extramarital affair with a woman suspected of spying for China.

Why does the B300 trigger U.S. export controls?

The H200 remains the highest-performing accelerator currently permitted into China, two generations behind the B300. The B300 can ship with explicit U.S. government approval, but in principle cannot be shipped. Export controls (U.S. restrictions on specified technology bound for China) make the lease highly likely to have violated U.S. rules.

How was the purchase financed?

Lingmeixin owned the servers, and Beijing Zhongguancun Bank supplied the funds to buy them. The contract was registered in the database of a credit information agency under the People's Bank of China.

That registration places the purchase in a formal credit channel, not an informal payment arrangement. It also leaves a paper trail, and one that survived only until Bloomberg began reporting: references to Moonshot AI were abruptly deleted from the document.

Moonshot AI denied the allegations, saying it has "no business relationship whatsoever with Lingmeixin." Bloomberg did not say who removed the company's name from the filed contract.

What does this change for China's AI buildout?

It changes the price of restricted compute more than the availability of it. Bloomberg's read: Chinese AI companies remain dependent on American semiconductors despite U.S. export controls. The sourcing trail in the documents runs through a financing contract, while related chip use has surfaced through data centers in Southeast Asia and the Middle East and on China Telecom's network.

• Alibaba and ByteDance have leased Nvidia chips through data centers in Southeast Asia and the Middle East.
• U.S. suspicions persist that Moonshot AI trained Kimi K3, the high-performance model it unveiled in July, using Blackwell-equipped servers secured in Thailand.
• The Information reported that 32 B300 GPUs were detected last month on the network of China Telecom, a state-owned telecommunications carrier.
• SemiAnalysis, after comparing supply-chain research with production and installation data, estimates that hundreds of thousands of AI chips have flowed into China.

The scale matters because domestic substitutes have not closed the gap. "Chinese-made AI chips are struggling to compete with Nvidia GPUs in training AI models," said Su Jianjun, senior analyst at Omdia. "The software performance of Chinese AI chip makers is also still lacking."

For buyers, the practical effect is a two-tier market: legal channels deliver the H200, while Blackwell-class capacity trades at a 350% premium through intermediaries. Export controls have changed who profits from advanced compute. They have not reduced how much of it reaches Chinese model developers.