Moonshot AI Seeks 8,000-GPU European Clusters to Reach Nvidia Blackwell

Moonshot AI is negotiating to lease European data center clusters of 8,000 or more Nvidia Blackwell GPUs, CNBC reported. The Chinese lab, best known for the Kimi chatbot, wants the capacity primarily for inferencing. It has also asked providers about compute for training runs.

The talks show how Chinese AI companies now route around US export controls. They cannot buy Blackwell-class GPUs. They can rent them.

How does the loophole work?

US rules ban Chinese firms from acquiring Nvidia Blackwell and other AI GPUs that match or exceed its performance when those products contain American technology. But the rules have not stopped Chinese companies from renting the same silicon inside data centers abroad. Moonshot is exploiting exactly that gap.

Washington is moving to close it. The US House has passed the Remote Access Security Act. The bill still requires Senate approval and a presidential signature before it becomes law. Until then, Chinese AI labs continue to train models on foreign data centers that legally bought Nvidia chips.

Why is Moonshot in a hurry?

Time pressure comes from two directions. First, the pending legislation could ban the rental arrangement outright. Reports like the CNBC disclosure may increase political pressure to advance the bipartisan bill sooner, which would cut off the option entirely.

Second, Nvidia itself has set a limit. The company has reportedly told its data center clients to restrict their business and avoid serving more than 49% of their capacity to Chinese customers. If enforced broadly, that cap constrains how much compute Moonshot and its peers can secure even before any law passes.

What about chips sold into China?

The policy picture has shifted twice. Washington initially tightened export bans to make it harder for Chinese AI labs to acquire hardware built on American technology. President Donald Trump then reversed course and permitted Nvidia H200 sales to China. Beijing responded by blocking H200 imports at the border to protect China's domestic AI chip industry.

The result leaves Chinese labs with no legal route to top-end Nvidia silicon at home. Foreign rented capacity is the remaining path.

Does the strategy change the AI race?

Chinese open-weight models now trail US frontier models by only months, according to the report. Models such as Kimi K3 are also significantly easier to run and could be much cheaper to use than their US counterparts. That efficiency matters for inferencing economics, which may explain why Moonshot prioritized inference clusters in its European negotiations.

These gains arrived despite successive US export restrictions. Access to rented Blackwell capacity abroad, however limited by the 49% rule, helps sustain the trajectory.

What happens next?

Moonshot's window depends on three actors. The Senate and the White House will decide whether the Remote Access Security Act becomes law. Nvidia will decide how strictly data centers apply its 49% capacity guidance. And Beijing's border blocks on the H200 show that Chinese policy, not only US policy, shapes the company's options. If the bill advances, expect European data center operators to renegotiate or cancel Chinese leases rather than risk US penalties — and expect Moonshot to accelerate its search for compute before that happens.