Moonshot in Talks to Lease Capacity From European Data Centers, Including 8,000 or More GPUs
Key Facts
Moonshot AI, the developer of the Kimi model, is in talks to lease clusters of Nvidia Blackwell GPUs from at least one European data center, according to coverage by Tom's Hardware citing CNBC. One of the discussions involved a cluster of 8,000 or more GPUs, the report said. The figure refers to computing capacity under consideration for lease, not an order to buy chips or a completed agreement. The report places the talks within Chinese AI companies' efforts to access computing power outside China. It did not give a timeline for signing an agreement or bringing the GPUs online.
Inference was the main use for the models discussed in the talks, according to the report, with interest also in securing computing capacity for training. The focus on inference means the capacity could be used to run models after they have been developed, while training enables models to be built or updated. The report said that leasing capacity from data centers outside China had allowed Chinese companies to access computing despite restrictions on chip purchases. It did not specify how much capacity would be allocated to inference and training, respectively. It also gave no details about the data center's identity or the lease terms.
The discussions come against the backdrop of U.S. legislative action to regulate remote access to advanced technology through cloud services. The U.S. House of Representatives passed the Remote Access Security Act on January 12, 2026, by 369 votes to 22, according to the House record. But the coverage reporting the talks said the bill still needs to pass the Senate and the White House before becoming law. The House Select Committee on the Chinese Communist Party said the bill expands the government's authority to restrict foreign adversaries' access to technologies, including AI chips, through cloud services. The legislation therefore focuses on access to computing power remotely, alongside restrictions on exports of the chips themselves.
Moonshot's computing needs highlight the operational strain the company faced as demand for its services grew. In July 2026, it temporarily stopped accepting new subscribers to Kimi 3 after demand neared the limits of its operating capacity, saying this happened within 48 hours. The Associated Press described Kimi 3, which Moonshot launched that same month, as an open-weight model with 2.8 trillion parameters. The report quoted an analyst as saying the company did not have enough computing chips to meet the sudden surge in demand. Moonshot said it was prioritizing existing subscribers and adding capacity as quickly as possible, and that it would reopen new subscriptions in phases.
Also in July, a White House official accused Moonshot of accessing servers equipped with Nvidia GB300 chips in Thailand, saying it may have used them to train its models, according to Investing.com, which cited a post by the official. This remains an allegation attributed to a U.S. official, not an independently established finding in the report. In the context of export policy, the U.S. Commerce Department said in January 2026 that export applications for Nvidia H200 and comparable AMD MI325X chips to China would be reviewed case by case if security requirements were met. Nvidia said in its 2026 annual filing that U.S. export restrictions had changed and expanded over the previous three years, and that under the rules and geopolitical conditions in effect at the time of the filing, it could not offer a data center product for China approved by both the U.S. and Chinese sides. These developments show how rules governing access to computing and chips changed during the year, while the report does not say how the European talks will turn out.
Nvidia shares closed at $186.64 on November 6, 2026, according to EL7 data. The price provides a reference point for the stock in the session before the news was published, without the data linking that close to the Moonshot talks or to any specific regulatory development. On the technology, Nvidia says its Blackwell platform was designed to run large-scale AI applications, including inference. The company's announcement describes the GB200 NVL72 system as linking 72 GPUs within a single NVLink domain. The available coverage gives no price for leasing the European capacity or estimate of its financial impact on Moonshot or Nvidia.