Moonshot AI, an AI company based in Beijing, China, has been negotiating with cloud providers to use NVIDIA Blackwell GPUs installed in European data centers. CNBC reported on November 5, 2026 that one anonymous source said the talks involved more than 8,000 GPUs. However, it has not been confirmed whether a deal was reached or whether the company has actually started using the GPUs.
Moonshot AI was reportedly seeking computing resources mainly for "inference," the process of using a trained AI model to respond to user inputs. The negotiations concern remote use of advanced GPUs, which U.S. export controls restrict from being shipped to China, while the hardware remains installed in European facilities. There was also reportedly a proposal to ask Moonshot AI for prepayments 30–50% higher than those required of non-Chinese customers, in anticipation that the U.S. might in the future regulate GPU use through overseas clouds. In securing AI computing resources, regulatory risk, meaning whether access can continue in the future, is starting to affect contract terms alongside hardware performance and pricing.
Moonshot AI's Talks Over 8,000+ Blackwell GPUs in Europe
According to CNBC, Moonshot AI began discussing compute lease agreements in 2026 with several neoclouds that operate GPUs in European data centers. A neocloud is a cloud provider that specializes in supplying GPU computing resources for AI. Of the four anonymous sources familiar with the matter whom CNBC spoke to, three said Moonshot AI was seeking NVIDIA Blackwell GPUs, and one of them referred to a European cluster with more than 8,000 GPUs. The specific figure of 8,000 or more, therefore, was not independently confirmed by multiple sources.
Founded in 2023, Moonshot AI is the Chinese company behind the Kimi series of AI models. In these negotiations, two sources said the company was seeking computing resources for inference, while another said it was also interested in training clusters. However, the names of the counterparties, the countries where the facilities are located, the contract value, and the planned start date have not been disclosed. CNBC said Moonshot AI, NVIDIA, the White House, the U.S. Department of Commerce, and others did not respond to requests for comment, so the report remains at the stage of contract negotiations.
In July 2026, the White House alleged that Moonshot AI had remotely used NVIDIA Blackwell GPUs through a facility in Thailand. One of CNBC's sources said the European talks took place after that allegation. Still, the U.S. government's claim about use in Thailand and the new contract negotiations in Europe are separate matters. There is no information indicating that Moonshot AI moved its computing from Thailand to Europe, nor can it be concluded that the 8,000+ GPUs in Europe are already being used for Kimi services.
Growing Demand for Inference GPUs as Kimi Usage Expands
Behind Moonshot AI's search for large numbers of GPUs is rising inference demand as AI models spread. According to CNBC, when the company released Kimi K3 in July 2026, a surge of users caused a computing shortage. Moonshot AI reportedly temporarily suspended new paid subscriptions to prioritize service for existing members. That was a measure taken at the time and does not mean new sign-ups are still suspended, but it illustrates that developing a high-performance model alone is not enough to keep up with growing usage.
Training and inference place different kinds of demands on GPUs. Training uses large amounts of data to update a model's parameters, whereas inference uses the trained model to process user requests. Inference itself splits into two stages: "prefill," which processes the input text in bulk, and "decode," which generates the response step by step. Because the two stages stress GPU compute power and memory bandwidth differently, simply adding more GPUs does not necessarily improve performance uniformly.
According to NVIDIA's technical explanation, prefill processes the entire input in parallel, so compute performance is typically the main bottleneck. In ordinary decoding, by contrast, units called tokens are generated one at a time, referencing a "KV cache" that stores past computation results. Because the KV cache must be repeatedly read from HBM, a type of high-speed memory, memory bandwidth often limits decode performance. In other words, response speed and the number of requests that can be handled simultaneously depend not only on GPU compute performance but also on how quickly data can be retrieved from memory.
That said, the inference bottleneck varies with model design and processing methods. NVIDIA explains that "speculative decoding," which generates candidate outputs ahead of time, can in some cases make decoding more constrained by compute performance than by memory bandwidth. Using a prefix cache to reuse shared input across multiple requests also changes the characteristics of prefill. These optimizations are general technical descriptions of inference systems; it has not been confirmed that Kimi uses the same mechanisms.
From the scale of 8,000+ mentioned in these talks, it is not possible to estimate how much Kimi's response speed or concurrent user capacity would improve. Actual performance depends on GPU configuration, memory capacity, model size, inference software, and the nature of user requests. Even so, the fact that Moonshot AI sought inference-specific computing resources separately from training suggests that securing infrastructure to deliver services reliably has become as important as developing the models themselves.
Regulatory Risk Behind the 30–50% Higher Prepayment Proposal
One notable aspect of the talks is the proposal to require higher-than-usual prepayments. Citing one anonymous source, CNBC reported that terms were proposed under which Moonshot AI would prepay 30–50% more than non-Chinese customers. However, this is a comparison of prepayment amounts; it does not mean hourly GPU rates would be 30–50% higher. Nor does it mean 30–50% of the total contract value would be paid upfront. The specific contract value, term, and the share of the total that prepayment would represent have not been disclosed.
Behind this condition is the possibility that the U.S. will in the future regulate GPU use through overseas clouds. Even if a cloud provider signs a long-term contract with Moonshot AI, a change in U.S. regulations could force it to stop providing service midway, leaving it unable to recover expected revenue. According to CNBC, the providers of the equipment and the financial institutions backing their financing are wary of such risk. A proposal to receive larger prepayments can be seen as a measure to recover more funds early in case service is cut off during the contract term.
However, what has been reported is only a negotiating proposal; it has not been confirmed that Moonshot AI accepted the terms or actually made any payments. Moreover, shipping chips themselves to China and using chips installed overseas over a network are treated differently under U.S. export control rules. This difference is important for understanding both these contract talks and the push for stricter U.S. regulation.
Based on CNBC's report, along with the U.S. Congress's official legislative record and the text of the bill passed by the House, the situation can be summarized as follows.
| Item | Status as of November 5, 2026 | Meaning |
|---|---|---|
| Shipping Blackwell GPUs to China | CNBC says it is effectively prohibited under U.S. export controls | A restriction on bringing the physical GPUs into China |
| Remote use of GPUs located in Europe | Negotiations to rent access to 8,000+ Blackwell GPUs were reported | Whether a deal was concluded or the GPUs are in operation is unconfirmed |
| Prepayment at contract signing | A proposal emerged for 30–50% more than non-Chinese customers | A comparison of prepayment amounts, not a rise in total contract value or hourly rates |
| H.R.2683 (Remote Access Security Act) | Passed the House, not yet passed by the Senate | A bill expanding authority to regulate remote use from overseas; not an enacted blanket ban |
At present, shipping advanced U.S. GPUs directly to China and Chinese companies remotely using GPUs installed in places like Europe are treated differently under the rules. In the U.S. Congress, however, moves are advancing to explicitly bring such remote use within export controls. If the rules actually change, the impact could reach not only cloud providers that own the equipment but also the business plans of the Chinese companies renting their computing resources.
What Does the Remote Access Security Act Passed by the U.S. House Regulate?
The Remote Access Security Act (H.R.2683), under consideration in the United States, is a bill that would amend the Export Control Reform Act of 2018 to add remote access to items under U.S. jurisdiction as a new regulated category. According to CNBC, under the current U.S. export control system, shipping advanced GPUs to China is strictly restricted, while the authority to broadly regulate use through overseas cloud services is limited. H.R.2683 is positioned as a bill to address this regulatory gap. However, whether Moonshot AI's individual contracts are lawful under the current system depends on the terms of the transactions and the regulations that apply, and this report alone does not allow a conclusion.
Section 2(1) of the bill passed by the House newly defines "remote access" in the Export Control Reform Act. Specifically, it covers foreign persons or foreign companies using items under U.S. jurisdiction from a location different from their physical location, via networks including the internet and cloud services. However, rather than uniformly regulating all remote use, the bill includes a condition that the Secretary of Commerce determine that the use could pose a serious risk to U.S. national security or foreign policy. In addition, Section 2(3) explicitly provides authority to add remote access to covered items by foreign persons to export controls and regulate it.
This bill would not immediately impose a blanket ban on foreign companies' cloud use of U.S.-made GPUs. If enacted, it would expand the legal authority of the U.S. government to regulate remote use, but exactly which items and forms of use would be regulated would also depend on subsequent rulemaking and administrative decisions. Therefore, the bill's enactment should be distinguished from an actual prohibition on companies like Moonshot AI using GPUs in Europe.
According to the official congressional record, H.R.2683 passed the House on January 12, 2026, by a vote of 369 to 22. It was sent to the Senate the next day and referred to the Committee on Banking, Housing, and Urban Affairs. As of CNBC's November 5 report, Senate passage had not been confirmed, and the bill had not become law. Whether Senate deliberations advance will help determine how use of advanced GPUs through overseas clouds is regulated.
For Moonshot AI, securing access to Blackwell GPUs in Europe could be an option for meeting Kimi's inference demand. But making that happen would require not only concluding a contract but also settling the start date of use, the allocation of computing resources, and contract terms that can withstand future regulatory changes. Whether the company will actually be able to use 8,000+ GPUs, and how much of that would be devoted to Kimi inference, remains unclear.
What this report highlights is that competition over AI companies' computing resources is expanding beyond the ability to procure GPUs to securing cross-border usage rights. As U.S. export controls tighten, Chinese companies are exploring ways to use overseas computing resources, but new regulations may extend to that use as well. Going forward, the focus will be on the outcome of Moonshot AI's contract negotiations, as well as on how H.R.2683 progresses and how far restrictions will extend to advanced GPU use through overseas clouds.