Moonshot is shopping for 8,000 Nvidia Blackwell GPUs in Europe

Moonshot AI has spent this year trying to rent its way onto the silicon Washington will not let it buy. The Beijing lab behind Kimi talked to neocloud operators about leasing Nvidia Blackwell capacity hosted in European data centres, according to four people familiar with the matter who spoke to CNBC — clusters of 8,000 or more GPUs, three of them said. Two described the interest as inference capacity, the compute that serves finished models to users; one also knew of interest in training. None of the four said a deal had been signed. One called Moonshot "very aggressive" in its hunt for compute, CNBC reported on Thursday.

That last detail is the story. A lab with a 2.8-trillion-parameter model and a capacity problem is not shopping for a discount. It is shopping for a jurisdiction.

The loophole is now the business model

Export controls stop a box crossing a border. They do not stop someone logging into it. Current US law gives regulators no broad authority over remote access to controlled chips hosted abroad, which is why renting from a non-Chinese operator in a third country has become the standard workaround — and why the hardware route is already a criminal business, as we wrote when Taiwan's AI-server indictment exposed the gap in chip export controls in August.

Congress has a bill for exactly this. The Remote Access Security Act would extend the Export Control Reform Act to cover access through a network connection, including a cloud service, from outside the item's physical location. It passed the House 369 to 22 on January 12 and was referred to the Senate Banking Committee the next day, where it has sat ever since. Bipartisan, two-thirds majority, nine months of nothing. That gap is the space Moonshot is operating in.

The White House has already named the practice once. In July, OSTP director Michael Kratsios said on X that Moonshot "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models," CNBC reported at the time. One of the four sources says the European conversations came after that accusation. The Thailand disclosure did not end the strategy; it moved it.

Nvidia has started policing the channel itself. The Financial Times reported on November 3 that the chipmaker told cloud and data-centre customers worldwide that at least 51% of their services under new contracts must go to non-Chinese companies — delivered verbally through sales teams rather than in writing, existing leases exempt, with no stated metric for how the share is measured. Nvidia did not confirm or deny it, according to the FT; its statement mentioned neither the number nor China.

Europe was the wrong shelf to shop

The real question is not whether Moonshot wants 8,000 Blackwell GPUs. It is whether Europe has them to spare, and largely it does not.

Europe's Blackwell capacity is being built against contracts signed years in advance, mostly by American labs. Nscale is deploying 12,600 GB300s at Start Campus in Sines, Portugal — the continent's first and largest Blackwell Ultra deployment — for Microsoft, with more than 66,000 Rubin GPUs planned there from late 2027, per DataCenterDynamics. Its Loughton campus in Essex is due to house roughly 23,000 GB300s for Azure from the first quarter of 2027. The European fleet is not idle; it is pre-sold.

That is what the reported premium is for. One source told CNBC it had been proposed that Moonshot pay 30% to 50% more upfront than a non-Chinese customer would — not a spot price for spare capacity, but a fee to jump a queue and to compensate the operator if remote-access rules change mid-contract. The neocloud and its lenders take more cash at signing because the back half of the deal is the part that can be legislated away. The premium is an insurance policy, priced by people who think the loophole is closing.

The arithmetic still works, which is the part that should worry Washington. A GB300 NVL72 rack holds 72 GPUs, so 8,000 chips is about 111 racks. One rack of B300s sells for roughly $600,000 internationally and traded in China last month at 18.5 million yuan — about $2.8 million — the Seoul Economic Daily reported, citing the FT. Renting in Europe at a 40% premium is cheap next to a 4.7x scarcity markup, and it carries none of the smuggling risk.

Who wins, who loses

Nvidia wins either way, which is why its rule is a customer-mix condition rather than a refusal. It keeps the revenue, keeps the political heat off, and pushes the compliance burden onto operators who say finding enough non-Chinese customers outside the US is hard. A Singapore provider told the FT that strict enforcement "will clearly hurt our business." That is the sound of an unauditable rule being absorbed rather than obeyed.

Europe still buys Moonshot two things Thailand cannot: latency for European customers, and a jurisdiction that is not the subject of a White House statement. It is just harder political ground than Southeast Asia. A neocloud selling frontier capacity to a Beijing lab hands a gift to every EU policymaker who wants a bloc-wide rule of their own. The 30–50% premium is what that risk costs today; it will cost more after the first story naming the counterparty. Beijing constrains Moonshot too — it decides which Chinese firms may import even older Nvidia parts, as we covered when ByteDance and Tencent got an H200 quota. Washington will not sell it the good chips; Beijing will not simply let it buy them.

The skeptic's case

Strip out the sourcing and the story is thinner than it looks. Four anonymous people, no completed deal, no named counterparty, and no comment from Moonshot, Nvidia, the White House, Commerce or the Chinese embassy. "Deals have been discussed" is doing a lot of work. A lab exploring options is not a lab signing.

Then there is the 51% rule's shape: unwritten, applied only to future contracts, with no definition of whether the share means revenue, capacity or headcount, and no stated enforcement mechanism. A condition you cannot audit is a nudge, not a control. Nobody quoted in the FT's account describes how compliance would be verified.

Worth separating what is policed from what is not. Training is legible — a cluster runs hot for months at a shape hard to disguise, which is how Thailand got noticed. Inference is not: a token request from a European endpoint looks like any other, and the operator has every commercial reason not to look closely at who is on the other end. If Moonshot is mainly renting inference, as two sources said, it has picked the half of the workload regulators are worst equipped to see.

What to watch

Whether a named European operator confirms a Chinese offtake — that is the moment the politics changes. Whether the Senate Banking Committee moves on a bill that has sat since January; a 369–22 House vote is a strange thing to leave on a desk. Nvidia's next quarterly results, where the customer-geography question now has a number attached. And Moonshot's Hong Kong listing, filed in September at a reported $50 billion: a public company has to tell its underwriters where the compute comes from, and that disclosure will be more specific than four anonymous sources.

The tell will not be an announcement. It will be a neocloud's capacity that quietly stops being available, and a contract nobody will describe.

If you ran a European neocloud, would you take the 30–50% premium or the political risk? Write to us via our contact page.