Moonshot AI, the Chinese company behind the Kimi family of AI models, reportedly held talks with European cloud infrastructure providers to rent compute capacity built on NVIDIA Blackwell GPUs. According to a CNBC investigation published on November 5, 2026, the negotiations involved clusters of 8,000 or more GPUs — a route that would let the company use chips barred from direct export to China without physically installing them in the country.
Moonshot AI and NVIDIA’s GPUs in 20 seconds
- Moonshot AI reportedly negotiated with European specialized cloud providers, known as neoclouds, to rent capacity built on NVIDIA Blackwell GPUs.
- One of the sources cited by CNBC mentioned clusters of 8,000 GPUs or more.
- The main interest was reportedly in running AI models, though there were also said to be talks about training capacity.
- U.S. restrictions limit the export of these chips to China, but remote access from foreign data centers opens a regulatory gap.
The information comes from four people familiar with the talks who spoke to CNBC on condition of anonymity. It does not confirm that Moonshot has signed a contract, that it already has that infrastructure in place, or that the negotiations ended in a deal. It also does not identify the European providers involved.
The deal would matter to the AI infrastructure market because it places European data centers in the middle of the technology restrictions between the United States and China. Companies that rent out GPUs can offer compute capacity to customers in other countries without the accelerators ever physically crossing those countries’ borders.
European Data Centers as an Alternative Route to NVIDIA Blackwell
According to the sources cited by CNBC, Moonshot had sought compute capacity at European facilities run by providers that can buy and host NVIDIA’s accelerators. These companies are usually called neoclouds, and they specialize in renting out accelerated computing resources for AI workloads.
Three of the four sources said Moonshot’s interest was directed at systems built on the Blackwell architecture. Two people familiar with the talks said the company was interested in clusters for inference, while another source was also aware of possible interest in training infrastructure.
The difference between the two workloads matters. Training requires large amounts of compute to develop or update models, while inference uses already-trained models to generate responses and handle user requests. The latter needs capacity available continuously whenever an AI service sees a high volume of queries.
A cluster of 8,000 GPUs is a large-scale operation, though the figure cited by one of the sources should not be read as contracted or installed capacity. CNBC is reporting on talks about possible deals, not on infrastructure that is already running for Moonshot in Europe.
The company develops the Kimi family of models, which it uses to compete in the AI market against U.S. providers such as OpenAI and Anthropic. Access to more NVIDIA accelerators could help give it additional resources to develop its models and serve its users, though the published information does not reveal how much capacity it currently has or how much additional compute it needs.
U.S. Restrictions Leave an Opening for Remote Rental
The possible deal sits within U.S. limits on exporting NVIDIA’s most advanced accelerators. These measures block certain chips, including the Blackwell systems mentioned in the report, from being shipped directly to China. Access to servers installed in third countries, however, is a different situation: the hardware stays abroad, and the customer pays to use its compute capacity over the cloud.
That is the mechanism Moonshot reportedly pursued. Instead of importing the accelerators, the company could contract resources hosted in European data centers and run its workloads there. The infrastructure would stay under the provider’s management, while the customer would access the resources remotely.
This is not a new controversy. In July 2026, the White House accused Moonshot of having accessed NVIDIA Blackwell GPUs through a facility located in Thailand, an example of the same “chip hole” that lets restricted GPUs reach Chinese users without the hardware ever being imported into the country. The information published by CNBC indicates that the talks to secure capacity in Europe took place after that accusation. Moonshot did not respond at the time to questions about that alleged access, according to the U.S. outlet.
U.S. law had also not generally closed off this possibility. Export controls focus on the physical departure of products and technologies, while renting compute capacity from foreign data centers presents additional regulatory difficulties. The Remote Access Security Act (RASA), a legislative proposal that would extend export-control restrictions to remote access to covered hardware and software, had passed the House of Representatives in January but was still pending in the Senate, according to the published information.
There is also a financial angle. One of the sources cited in the reporting on the talks said Moonshot had reportedly been asked to pay 30% to 50% more upfront than a non-Chinese customer. That premium would compensate providers and lenders for the risk that Washington might restrict remote-access contracts in the future. This is a proposal attributed to a source, not an agreed or confirmed rate.
NVIDIA Is Also Pressuring Cloud Providers
The search for European capacity coincides with reports of new commercial terms NVIDIA is setting for its data center and cloud partners. According to the Financial Times on November 3, the company had told infrastructure providers that at least 51% of their services under new contracts had to go to non-Chinese customers.
The measure, according to that outlet, was reportedly communicated verbally and would not affect existing rental agreements. NVIDIA did not publicly confirm those terms, though it defended the importance of letting data centers and clouds built with U.S. chips compete worldwide. Other outlets, including The Information, also reported that the company was asking some providers to reduce their exposure to Chinese customers and share information about end users — a push that echoes a recent case in which security researchers traced NVIDIA Blackwell Ultra B300 GPUs to IP addresses linked to China Telecom despite export controls.
If applied as described in those reports, the 51% condition would limit the weight of Chinese customers in new contracts, though it would not by itself amount to an outright ban on serving them. It would also require operators to review how they allocate their clusters’ capacity and which companies actually use the accelerators.
For European neoclouds, the situation poses a difficult business decision. Large GPU contracts require heavy investment in hardware, data centers, power, and cooling. A customer looking to reserve thousands of accelerators can represent a major source of revenue, but accepting that contract can also raise regulatory risk if the end user is subject to U.S. restrictions.
The upfront payment premium mentioned in the talks with Moonshot fits that uncertainty: the provider would receive more money at the start of the contract to reduce its exposure if a regulatory change forces it to cut off service. That said, the available sources do not identify the European companies that may have taken part in the negotiations, nor do they reveal whether any of them accepted those terms.
The case shows that control over AI accelerators no longer depends solely on who buys the chips or where they are installed. It also matters who rents the capacity, from which country they connect, and what conditions manufacturers impose on their business partners. For Europe, which is seeking to attract investment in data centers and AI computing services, these deals add a geopolitical dimension to managing its infrastructure.
Frequently Asked Questions
What is Moonshot AI and what models does it develop?
Moonshot AI is a Chinese artificial intelligence company known for developing the Kimi family of models. It competes in the advanced-model market with companies such as OpenAI and Anthropic.
How many NVIDIA Blackwell GPUs does Moonshot AI want to rent?
The talks mentioned by CNBC involved clusters of 8,000 GPUs or more. That figure comes from a source familiar with the negotiations and does not confirm that a signed contract exists.
Why is Moonshot AI looking for GPUs in Europe?
U.S. restrictions limit the direct export to China of certain advanced accelerators. By renting capacity at European data centers, Moonshot could use those chips remotely without physically moving them to Chinese territory.
What restrictions does NVIDIA want to apply to cloud providers?
According to the Financial Times, NVIDIA reportedly asked its partners to direct at least 51% of their services to non-Chinese customers under new contracts. The report does not confirm that this condition has become a public contractual rule.
This article is also available in Spanish: «Moonshot AI busca acceso a miles de GPU Blackwell de NVIDIA desde centros de datos europeos», on revistacloud.com.
