Moonshot Rents Its Way Around America’s Chip Wall

Human Verified 0

Share to: , and

A European data center building at night with rows of glowing server racks visible through glass

A Chinese AI lab is testing the limits of America’s chip curbs from the other side of the Atlantic. Moonshot AI, the Beijing company behind the Kimi models, has discussed renting access to Nvidia’s most powerful Blackwell chips through data centers in Europe, in talks that would put clusters of 8,000 or more GPUs at its disposal without a single chip crossing into China. The discussions expose a gap in Washington’s controls: they restrict shipments of hardware, not the rental of computing power hosted abroad.

The talks

The discussions took place this year with European neoclouds, the specialist operators that buy Nvidia GPUs and rent them out by the hour, according to four people familiar with the matter. Three of the sources said Moonshot was seeking Blackwell-powered systems, the advanced chips that US export controls effectively bar from import into China.

Two of the sources said the company wanted clusters for inference, the work of serving models to users, while one was also aware of interest in training. One source said talks had covered arrangements for clusters of 8,000 or more Blackwell GPUs hosted in the region, and described Moonshot as very aggressive in its pursuit of more computing power.

The loophole

US export controls restrict shipments of Nvidia’s most advanced chips to China, but remote access through overseas data centers sits in a legal grey zone. Current law gives regulators no authority over a company’s access to those chips through cloud services hosted abroad, so a Chinese lab can rent hardware it cannot buy, provided the servers stay in Europe.

The arrangement comes at a price. One source said a proposal would have Moonshot pay 30% to 50% more upfront than a non-Chinese customer, a premium that reflects the risk neoclouds and their financiers take on. Deals with Chinese customers can be more lucrative upfront, as operators hedge against a future ban on remote-access agreements.

A stylized world map with glowing network connections between Europe and East Asia

Washington’s answer

Lawmakers have already moved to close the gap. The Remote Access Security Act, which would extend US export controls to cloud-based access to critical hardware and software, passed the House of Representatives in January but has not passed the Senate.

The Financial Times reported on Tuesday that Nvidia had told cloud providers selling access to its chips that at least 51% of their services must go to non-Chinese customers. The move would squeeze the same neoclouds Moonshot is courting, by capping how much of their business any Chinese customer can represent.

Moonshot, Nvidia, the White House, the Commerce Department and China’s embassy did not respond to comment requests. Asked separately, an Nvidia spokesperson said America cannot win the AI race unless its AI infrastructure wins worldwide, warning that if the world’s clouds rely on foreign-made chips, the model and application layers will be foreign-built as well.

Our read

The story is not really about a lease. It is about the difference between controlling atoms and controlling bits: Washington can stop a GPU crossing a border, but it cannot easily stop the computation that GPU performs from crossing one.

Moonshot’s interest in inference clusters, not just training, shows the prize is access to serve its models, not a one-off training run. Its Kimi K3 model, released in July, drew demand so heavy that Moonshot paused new subscriptions to protect compute for existing users, and a lab that cannot buy Blackwell chips will pay a premium to rent them abroad.

The premium is the market pricing the regulatory risk. Neoclouds are charging Chinese customers more because the access could be banned tomorrow, and the Remote Access Security Act sitting in the Senate is the reason why. Until that bill moves, the loophole stays open, and every Chinese lab with money to spend knows it.

Why It Matters

Export controls were meant to slow China’s AI progress by denying it the best hardware. If the same hardware can be rented by the hour from a European data center, the controls slow the shipment of chips without slowing the training of models.

That shifts the contest from hardware to law: either Congress extends the rules to remote access, or the curbs become a tax on Chinese labs rather than a barrier. Enforcement already reaches the physical side: Malaysia seized $12.1 million in smuggled AI chips this year.

Background

Moonshot was founded in 2023 and released Kimi K3 in July, a model it said closed the gap with leading systems built by Anthropic and OpenAI. The same month, the White House said the lab had accessed Blackwell chips through a facility in Thailand. Meanwhile, about 600,000 Nvidia H200 chips sit in Hong Kong awaiting shipment to mainland China.

Source: CNBC’s report and Unite.AI’s coverage

Share to: , and

Faheem R. Khattak
Founder & Senior Editor

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

✓ 10+ Years Covering Google Search & AI

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.