Sources: Nvidia verbally told cloud compute clients to sell 51%+ of their services to non-Chinese companies, imposing new limits on China's access to AI chips (Financial Times)

Nvidia reportedly gave cloud-computing clients a new customer-mix requirement. They would need to sell more than 51% of their services to non-Chinese companies. The goal is to reduce cloud providers’ exposure to Chinese customers and limit China’s access to Nvidia-powered computing.

For example, a data-centre partner serving both Chinese and overseas businesses would need a majority of its services sold outside China. The key mechanism is commercial access, rather than only direct chip shipments. By changing which customers can use partner capacity, Nvidia could narrow Chinese companies’ access to advanced AI computing.

The requirement was reportedly delivered verbally, so the excerpt does not describe a formal written rule or enforcement process. Still, it marks another restriction involving Nvidia’s data-centre partners. Chinese firms could face fewer cloud options, higher costs, or less access to computing needed for advanced AI.

Selling more than 51% of services to non-Chinese companies means the provider’s customer base must be majority non-Chinese. The threshold is measured by services sold, such as rented computing time or access to cloud resources. It does not necessarily mean that Chinese customers are completely excluded.

For example, if a provider sells 100 units of computing capacity, at least 52 units would need to go to non-Chinese companies. Chinese customers could receive the remaining 48 units or fewer. The central mechanism is a business-allocation rule that limits how much capacity can be directed toward Chinese clients.

This could make Chinese access more difficult even when service remains technically available. Providers might reduce Chinese contracts, ration capacity, or prioritize other customers. The supplied article does not explain whether Nvidia would measure the threshold by revenue, computing hours, capacity, or another metric, so that detail remains unclear.

Nvidia has an outsized role in supplying data-centre chips for advanced artificial intelligence. Its processors are widely used to train and operate modern AI models. That makes Nvidia important not only as a chipmaker, but also as a gatekeeper to computing capacity delivered through cloud providers.

The article focuses on Nvidia’s chips and its data-centre partners, but it gives no percentage, shipment total, or revenue figure describing Nvidia’s exact market share. In general industry reporting, Nvidia is commonly described as the dominant supplier of high-end AI accelerators. That broader point comes from established outside knowledge, not a number stated in the supplied excerpt.

Because so many AI services rely on Nvidia-based systems, restrictions involving its partners can have effects beyond direct chip sales. If cloud companies reduce service for Chinese customers, access to Nvidia hardware may become harder even when the customers do not buy chips themselves.

Advanced AI models require large amounts of computing power during training and operation. If Chinese companies have less access to Nvidia-powered cloud services, they may struggle to obtain enough suitable hardware. Training could take longer, become more expensive, or require smaller models and fewer experiments.

For example, a company that previously rented many Nvidia-equipped servers through a cloud provider might receive less capacity under the new customer rule. It could then wait for access, pay more elsewhere, or divide work across less capable systems. The mechanism is a reduction in available high-performance computing, not necessarily an immediate ban on every Chinese user.

The result would depend on how broadly providers apply the reported requirement and what alternatives remain available. Chinese firms could continue developing AI, but constraints may slow progress or raise costs. Over time, limited access could make it harder to compete in building and deploying advanced models.

The United States and other governments restrict access to the most advanced AI chips because these chips provide enormous computing power. That power can support sophisticated AI systems with potential military, intelligence, cybersecurity, and surveillance uses. Governments therefore treat certain chips and related equipment as strategically sensitive.

A restriction can work through export controls, licensing requirements, or limits on cloud access. For instance, preventing a Chinese company from obtaining high-end processors directly may be less effective if it can simply rent the same hardware abroad. Rules affecting cloud providers attempt to close that route by limiting access to computing services, including those delivered through Nvidia-powered data centres.

The broader policy aim is to slow the transfer of advanced capabilities while protecting national-security interests. Such controls can also reshape global technology supply chains. The supplied excerpt identifies new limits on China’s access, but it does not detail the specific government rules or security assessments behind them.

Chinese companies have several possible ways to obtain AI computing without unrestricted access to the newest Nvidia-powered services. They could use domestically produced chips, older hardware already in the country, overseas cloud providers, or smaller and more efficient AI models. They might also combine many weaker processors to replace some high-end capacity.

Each option has trade-offs. Domestic or older chips may provide less performance, while combining more processors can increase software complexity and electricity use. Overseas clouds may be difficult to access, expensive, or subject to the same customer restrictions. More efficient models reduce computing needs but may not match the capabilities of systems trained on larger resources.

These alternatives could help Chinese companies continue operating, but they may not fully replace the scale and convenience of Nvidia-powered infrastructure. Availability also depends on future export controls and provider policies. The supplied article confirms new limits but does not identify which alternatives Chinese firms are already using.

A GPU, or graphics processing unit, is a chip designed to perform many calculations simultaneously. Although originally developed for graphics, GPUs are well suited to the repeated numerical operations used by machine-learning systems. They can process large batches of data in parallel rather than handling one calculation at a time.

During AI training, a model repeatedly adjusts millions or billions of numerical parameters. GPUs accelerate the matrix and vector calculations behind those adjustments. A cloud provider can link many GPUs in a data centre, giving companies access to far more computing power than a single ordinary computer could provide.

This is why access to high-performance GPUs matters for advanced AI. More suitable processors can shorten training and support larger models, though they also require substantial infrastructure and energy. Nvidia is central to the article because its AI chips are used in data centres that cloud providers make available to customers. The source does not explain GPU technology itself; this definition uses established computing knowledge.