Moonshot AI, the Beijing lab behind the Kimi models, has been trying to rent Nvidia Blackwell capacity in Europe, according to four people familiar with the discussions who spoke to CNBC. The company cannot import the chips, which U.S. export controls effectively bar from China. It can, under current law, rent them from a data center that already owns them.

Talks this year have covered capacity from neoclouds hosting Nvidia GPUs at European facilities, with one source describing discussions about clusters of 8,000 or more Blackwell GPUs. Two of the sources said Moonshot was interested in inference capacity, meaning servers that answer user requests rather than train new models, and one said training clusters were also discussed.

The White House said in July that Moonshot had accessed Blackwell chips through a facility in Thailand. The European conversations have taken place since then, according to one of the sources, who described the lab as "very aggressive" in its pursuit of compute. Moonshot, Nvidia, the White House, the Commerce Department and Chinese officials did not respond to CNBC's requests for comment.

Renting what cannot be imported

The gap in the rules is structural rather than accidental. Export controls restrict shipments of advanced chips into China. They do not give U.S. regulators broad authority over whether a Chinese company can buy time on those chips through a cloud provider hosted somewhere else. A lab in Beijing with a credit card and a contract can reach the same silicon a lab in California reaches, as long as the servers sit outside the restricted territory.

That is why the deals are structured as leases. A Chinese customer typically rents from a non-Chinese neocloud or operator that can buy the hardware directly, then pays for capacity over a multi-year term. Customers usually pay an upfront fee to secure the contract, calculated as a percentage of its lifetime value.

Moonshot's offer has been richer than the norm. One source said the lab proposed paying 30 to 50 percent more upfront than a non-Chinese offtaker would. For the seller, that premium is compensation for risk: lenders and neoclouds financing these deals are pricing in the possibility that remote-access agreements get banned and the customer disappears mid-contract.

Why the numbers are so large

Moonshot released a model in July that it said closed the gap with systems from Anthropic and OpenAI, and demand ran past its capacity. The company paused new subscriptions to protect compute for existing members and said it would reopen spots in batches. A lab in that position has one problem, and it is not model quality.

Chinese models have gained adoption through 2026 as cheaper alternatives, and every additional user is an inference bill. Training the next model is a second, separate bill paid in the same currency. Renting 8,000 Blackwell GPUs is the kind of order that shows up in a neocloud's quarterly revenue and in Nvidia's guidance, which is why the arrangement matters beyond one company.

The demand is also visible in supply chains that sit outside the data centre. The same squeeze has pushed memory makers to reallocate capacity, and it has made tool licensing a bargaining chip, as with SK Hynix seeking permission to keep upgrading its China fab.

The rule that would close the door

Washington has a bill for this. The Remote Access Security Act would extend export controls to cloud-based access of critical hardware and software. It passed the House in January and has not passed the Senate, which means the loophole is currently legal and politically contested at the same time.

Nvidia has been tightening its own rules in parallel. The Financial Times reported on Tuesday that the company told cloud providers selling access to its chips that they must sell at least 51 percent of those services to non-Chinese customers. Asked about the reports, an Nvidia spokesperson said America cannot win the AI race unless its infrastructure wins worldwide, and warned that clouds relying on foreign chips would produce foreign models.

There is a domestic complication on the Chinese side as well. Nvidia said in August that Washington had licensed some sales of its older H200 chips, but that the Chinese government restricted those purchases while it pushes domestic suppliers. The result is a market where the buyer wants the chips, the seller wants the revenue, and both governments are pulling in opposite directions.

For neocloud operators, the calculus is a premium today against a ban tomorrow. For Moonshot, it is capacity now against a contract that may not survive the next Congress. The lease is the only instrument that satisfies both, which is why the deals keep being discussed even when nobody will confirm them. Similar pressure has already reshaped the talent side of the industry, where a Samsung engineer went to prison for moving to a Chinese memory maker.