Chinese AI developers are evaluating cloud rental of Nvidia Rubin GPUs, though cost, complexity, and regulatory barriers pose deployment challenges.
China-based AI hardware developers are making rapid progress with accelerators of their own design, yet China's most advanced AI developers increasingly acknowledge that domestic hardware is unlikely to catch up with U.S. leaders in the near term, which greatly limits the development of competitive models. To remain competitive with American peers, Chinese AI developers are exploring ways to rent Nvidia's upcoming Rubin GPUs in the cloud. When Nvidia introduced its Rubin datacenter platform for AI in January, it publicly named American customers and omitted Chinese ones, reflecting U.S. export rules, its commitment to comply with them, and its intent not to signal to investors about the opening of the Chinese market. This message prompted Chinese companies to begin exploring ways to obtain leading-edge processors from Nvidia remotely to avoid falling behind their U.S. rivals.
Chinese AI companies have begun negotiating access to NVL144 GR200 and other Nvidia Rubin-based systems hosted in data centers outside China, particularly in Southeast Asia and the Middle East. Up until the middle of this week, these arrangements were generally considered legal, though they come with inherent limitations: compute is rented rather than owned, capacity is shared rather than dedicated, and deployment timelines depend on third-party operators rather than internal schedules.
The difference between renting Rubin in a remote cloud data center and deploying it locally is profound. U.S. hyperscalers can integrate Rubin accelerators at scale, tune their software stacks tightly around the new hardware, and reserve massive GPU clusters for long training runs. Chinese developers renting Rubin capacity, by contrast, will have to cope with limited allocations, cross-border latency, limited freedom to customize systems, and, in some cases, queuing. If they rent enough systems, they may well train their models without much hassle; however, if they cannot find appropriate clouds on time, they will have fewer AI accelerators per project and, in some cases, be unable to run large training jobs, which will directly cap model size, experimentation cadence, and iteration speed.
Chinese developers already have experience with this challenge. Using fleets of different Nvidia GPUs consisting of A100, H100, H800, and H20 to train frontier models, they rented additional capacity in the cloud—an experience that insiders say was costly and operationally awkward. As models scale further, the value of uninterrupted access to large, homogeneous GPU clusters grows increasingly important, yet rented capacity rarely delivers this. Even if deals are secured, they typically leave Chinese developers at a structural disadvantage relative to well-funded American competitors that can deploy tens of thousands of accelerators under one roof.
The financial constraints facing Chinese AI developers compound these challenges. According to UBS estimates, China's hyperscalers spent about $57 billion on capital expenditures last year, roughly one-tenth of U.S. peers—less than Meta's CapEx of over $70 billion alone. Given these financial limitations, it remains to be seen whether Chinese AI developers will be able to stay more or less competitive with their American counterparts.