Moonshot AI Seeks Access to 8,000 Nvidia Blackwell Chips Through European Cloud Providers

Europe becomes a target in Moonshot's search for computing power

Moonshot AI has discussed renting access to clusters of at least 8,000 Nvidia Blackwell graphics processing units in European data centers, as the Beijing company seeks more computing power for its growing Kimi artificial intelligence business, according to people familiar with the talks.

Four people described discussions this year with specialist cloud providers operating in Europe. Three said Moonshot was seeking Blackwell capacity, while one described a proposed cluster containing 8,000 or more chips. The hardware would remain abroad, with Moonshot accessing its computing power remotely.

The discussions place Moonshot at the center of a dispute over the reach of American export controls. Washington restricts shipments of its most advanced AI chips to China, while current law gives regulators more limited authority over Chinese companies using those chips through overseas cloud services.

Some proposed arrangements would also make Moonshot pay more money before receiving access. One person familiar with the discussions said a proposal called for an upfront payment 30% to 50% higher than a comparable customer outside China would pay, reflecting concern that future American restrictions could interrupt the business.

No completed European agreement, named provider, specific data center location, contract value or starting date has been disclosed. The 8,000 figure describes capacity discussed in negotiations.

Why Kimi needs more capacity

Moonshot, founded in 2023, develops the Kimi family of AI models. Its search for hardware follows the July release of Kimi K3, which the company said narrowed the performance gap with leading systems from OpenAI and Anthropic.

Demand around that launch stretched Moonshot's computing resources enough for the company to pause new subscriptions and prioritize existing members. For customers, that turned a technical constraint into a practical limit: new users could not immediately buy access while the company worked to expand capacity.

Moonshot described its response in a LinkedIn statement about the subscription pause:

"We're adding capacity as fast as we can and will reopen new subscription spots in batches," Moonshot said.

The company has attracted users seeking a cheaper alternative to leading American AI services. Lower prices can help bring in customers, but serving a growing audience still requires computing equipment every time those customers submit requests.

Moonshot's Kimi K3 announcement on July 17, 2026, described a model with 2.8 trillion parameters and a context window of one million tokens. Parameters are numerical values learned during training. A context window is the amount of material a model can work with in a single interaction.

The earlier announcement put K3 ahead of major American competitors in some evaluations, including a comparison of website coding outputs based on human preferences, while placing it behind the strongest systems on a broader intelligence index. Its appeal combines competitive performance, lower advertised prices and the planned availability of model weights for developers.

What Moonshot wants the chips to do

Two people familiar with the discussions said Moonshot was interested in European computing clusters for inference, the process of running a trained AI model to answer users. One also described interest in capacity for training.

Training and inference use similar equipment for different jobs. Training adjusts a model's parameters by processing large amounts of data. Inference uses those learned parameters to produce an answer, write code or complete another task after a customer sends a request.

For Moonshot's subscribers, inference capacity is the more immediate issue. Every additional customer adds potential demand, and long documents or complex tasks can require more computation than short exchanges. A popular model therefore needs a substantial supply of computing power after its development is finished.

A cluster connects many graphics processing units, or GPUs, so they can work together. These chips are particularly useful for AI because they can perform many calculations at once. The proposed scale of at least 8,000 Blackwell GPUs shows that Moonshot is discussing a large pool of capacity rather than occasional access to individual servers.

How overseas cloud access changes the transaction

European cloud providers offer Moonshot a route to computing power without delivering the underlying chips to China. The providers buy and operate the equipment, then rent customers access through network connections.

The companies involved in the discussions were described as neoclouds. These are specialist cloud businesses built around renting powerful GPUs for AI work, rather than providing the broad mix of storage, business software and other services associated with large general cloud platforms.

A Chinese developer using such a service can send work to servers overseas and receive the results remotely. The equipment remains in the foreign data center, while the developer pays for access to its processing capacity.

American export controls have focused on restricting the physical movement of advanced chips to China. Remote access creates a different regulatory problem because the customer can use the hardware without importing it. The European discussions concern that distinction between where a chip sits and who can use its computing power.

For European providers, the opportunity comes with exposure to policy decisions made in Washington. A contract can promise computing capacity over an extended period, while a later rule could change whether a particular customer is allowed to receive it.

The Thailand allegation and the sequence of events

Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged in a July 22 post on X that Moonshot had acquired servers equipped with Nvidia GB300 systems and used GB300 servers in Thailand, likely for model training.

GB300 systems use Nvidia's Blackwell technology. The allegation placed overseas access at the center of Washington's concerns about Moonshot, shortly after the company's new model drew attention to its growing capabilities.

At least one person familiar with the European talks said discussions continued after the White House raised the Thailand allegation. The chronology connects Moonshot's rising demand for computing power with mounting American scrutiny:

• January: The U.S. House of Representatives passed the Remote Access Security Act, according to the legislative status described in the accounts.
• July 17, 2026: Moonshot announced Kimi K3, with full model weights scheduled for release on July 27.
• Around the K3 release: Moonshot paused new subscriptions as demand stretched computing capacity.
• July 22: Kratsios alleged that Moonshot had used Nvidia GB300 servers in Thailand.
• After the July allegation: Discussions over European computing capacity continued.
• August: Nvidia said Washington had licensed some H200 sales to China, while Chinese authorities restricted those sales.

Why providers want more money upfront

Cloud providers and their financiers are weighing the possibility that Washington will restrict remote access more tightly. That risk is influencing the proposed payment terms for Chinese customers.

Customers seeking a large allocation of AI computing power commonly make an advance payment to reserve capacity. For the provider, that money helps support a business requiring expensive chips and data center infrastructure before rental income arrives.

The proposal described for Moonshot would raise that initial payment by 30% to 50% compared with a customer outside China. A simple illustration shows the difference: an upfront payment of $100 for the comparison customer would become $130 to $150 for Moonshot. Those amounts are illustrative, not disclosed contract figures.

The reported premium applies to the advance payment. The rental rate and total cost would depend on the rest of the agreement, including its duration and the capacity reserved.

Providers have a financial reason to accept the business: a large customer can fill substantial computing capacity and bring in revenue early. They also face a commercial risk if new restrictions arrive before the agreement has run its course. Asking for more cash at the start shifts part of that risk to the customer.

Washington considers extending the rules

The Remote Access Security Act, known as RASA, would extend American export controls to certain forms of remote access to controlled technology. Its purpose is to give authorities a stronger legal basis for regulating access through cloud services.

The House passed the measure in January, but it had not passed the Senate at the time described in the accounts. A proposal in Congress must complete the legislative process before its provisions become law.

For Moonshot and potential European providers, the legislation makes the timing and terms of a contract more consequential. A service that can be offered under current rules could become subject to additional controls, permissions or restrictions if Washington changes the law.

The dispute reaches beyond whether chips can enter China. American policymakers are also considering who should be allowed to use American technology once it has been installed elsewhere. That brings overseas cloud businesses into an export control debate previously centered more directly on manufacturers and shipments.

Nvidia faces pressure from both markets

Nvidia has reportedly required cloud providers renting access to its GPUs to sell at least 51% of their services to customers outside China. That condition would require a majority of the affected services to go to other customers, shaping how much business providers could allocate to Chinese demand.

The requirement introduces another commercial constraint alongside government rules. A provider negotiating with Moonshot would need to consider its wider customer mix as well as chip availability, financing and the terms of the proposed contract.

Nvidia has argued that American AI infrastructure needs a strong global presence for the United States to remain competitive. Its position is that allowing rival hardware suppliers to dominate overseas cloud infrastructure would also help competitors build the models and applications running on it.

The company also faces restrictions from Beijing. Nvidia said in August that Washington had licensed some H200 chip sales to China, but that Chinese authorities had restricted those sales as Beijing sought to support domestic chipmakers.

Moonshot's European search sits between these pressures. It needs capacity to serve more customers and develop future models, while providers must decide how much business they can accept under changing American policy and Nvidia's reported customer requirements. The immediate commercial issue is whether enough computing power can be reserved on terms both sides are willing to accept.

Key Points

• Moonshot has discussed European access to clusters containing at least 8,000 Nvidia Blackwell GPUs.
• The intended uses include serving Kimi users through inference, with interest also reported in training capacity.
• Demand around Kimi K3's July release prompted Moonshot to pause new subscriptions while expanding capacity.
• One proposal called for an upfront payment 30% to 50% higher than a comparable customer outside China would pay.
• The White House alleged in July that Moonshot had used Nvidia GB300 servers in Thailand.
• The Remote Access Security Act had passed the House but awaited Senate passage, leaving overseas cloud access at the center of the policy dispute.