Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Moonshot AI Reported Seeking European Nvidia Blackwell Capacity, Offering Up to 50% More Upfront

Moonshot AI has been negotiating to lease clusters of more than 8,000 Nvidia Blackwell GPUs through European data centers, and is reportedly willing to pay 30% to 50% more upfront than non-Chinese customers. The arrangement would give the Kimi maker access to hardware Washington has effectively banned from being shipped to China, exposing how cloud access can route around export controls.

Published

Moonshot AI is exploring deals to lease access to clusters of more than 8,000 Nvidia Blackwell GPUs through European data centers, according to three people familiar with the matter cited by CNBC. The Beijing-based developer of the Kimi models has been talking to neocloud operators that host Nvidia hardware in Europe.

The reported terms are unusually direct: Moonshot would pay 30% to 50% more upfront than a non-Chinese customer for comparable capacity. One source described the company as “very aggressive” in its pursuit of additional computing power. Neoclouds typically lease capacity to offtakers who pay fees plus an upfront amount to lock in a contract.

Two of the sources said Moonshot's interest in European clusters was mainly for inference, the work of running trained models to serve users, while one source said capacity for model training was also under discussion. Together, those two needs point to sustained production capacity rather than a one-off experiment.

The talks have a clear starting point. In July the White House said the Chinese lab had accessed Nvidia Blackwell chips through a facility in Thailand, and discussions about European capacity began after that allegation surfaced. Moonshot, Nvidia, the White House, the US Commerce Department, China's embassy in the UK and its foreign ministry did not respond to requests for comment.

For European operators, the higher upfront payment is less a premium than a risk buffer. If remote-access arrangements are restricted later, someone has to absorb the loss, which is why Chinese customers paying more in advance look attractive — and that is precisely the loophole US lawmakers want closed.

Demand is the other half of the story. After releasing a new model in July, Moonshot said demand had stretched its available capacity and forced it to pause new subscriptions, promising to reopen spots in batches as it adds compute. For a model developer, available capacity sets how many users it can actually serve.

Washington has already moved partway. The House passed the bipartisan Remote Access Security Act in January, which would extend export controls to cloud-based access to critical hardware, but the Senate has not taken it up. The House Select Committee on China urged Congress to finish the job, saying Chinese labs leasing advanced chips from US and European companies circumvents export controls. Nvidia has reportedly told cloud providers that at least 51% of their services must go to non-Chinese customers, according to the Financial Times, while a company spokesperson argued that if the world's clouds run on foreign chips, the model and application layers will be foreign-built too.

Restricting where advanced chips can be shipped may not be enough if companies can still rent the same hardware from abroad. Three things to watch: whether the Senate takes up RASA, whether European clouds and their financiers actually sign such contracts, and how Nvidia's 51% customer-mix requirement is enforced in individual leases.

Why it matters

If such leases become common, the practical boundary of export control shifts from where chips are shipped to who uses the compute, adding pressure on the US Senate to pass the Remote Access Security Act and repricing compute access for Chinese model developers.

Moonshot AINvidiaExport ControlsCompute
Back to realtime news

Nearby Updates

All

11/07, 13:32

Microsoft AI Chief Suleyman Calls Delaware's AI-Agent Company Plan ‘Deeply Concerning’

Microsoft AI chief executive Mustafa Suleyman has publicly criticized a Delaware draft bill that would let autonomous AI agents manage and operate corporate entities, saying the idea demands extreme caution. The proposal, which envisions a 30-month sandbox for AI-run companies, has also drawn a document request from a House oversight chairman warning that it could shield AI companies from liability.

11/07, 13:03

Anthropic's IPO framed as the first real test of AI's financial viability

BankInfoSecurity argues that Anthropic's listing, expected to be the largest IPO in history, is the first genuine test of whether the AI sector's economics can survive public markets. Its S-1 discloses $895.4 billion in commitments and leases against an $80.75 billion net loss, making the offering a benchmark for the entire supply chain.

11/07, 12:49

Perplexity launches Teammates, always-on agents for Max subscribers

Perplexity has introduced Teammates, a set of AI agents built for ongoing, large-scale work, available first on the web to Perplexity Max subscribers. They run around the clock, coordinate with each other and can message coworkers through Slack, free during a promotional period before credit-based billing begins.

11/07, 12:46

Qwen posts FP8 build of its 2.4T-parameter flagship, Qwen3.8-2.4T-A95B-FP8

Qwen has published Qwen3.8-2.4T-A95B-FP8 on its official Hugging Face repository, labelled as a quantized build of Qwen3.8-2.4T-A95B with text-generation tooling and safetensors weights. With 2.4 trillion total parameters, roughly 95 billion active per token and 8-bit weights, the update is about deployability rather than a new architecture.