Why America’s New AI Chip Rule Names Five Countries but Has China in Its Sights
Who Gets To Use An AI Chip After It Clears Customs?
A company in China can train an AI model on chips it has never bought, in a building its engineers have never visited. It can rent time on servers overseas, send the training job to the data centre and retrieve the result. The chips stay in their racks for the entire transaction.
On 16 November 2026, the US Commerce Department said it had issued an interim final rule targeting this kind of remote access where it could undermine American technology restrictions. One concrete measure is a requirement for licences to re-export covered advanced AI chips to five destinations: Brazil, Indonesia, Malaysia, Singapore and Thailand.
China is central to the announcement even though it is absent from that list. The five countries are places the hardware could go. Washington’s concern is who could use it once it arrives.
Why a Re-Export Licence Matters
Consider a distributor outside the United States that wants to send a covered processor to a data centre in Singapore. Moving the chip from one foreign country to another is a re-export. Under the announced measure, that movement requires a licence. A requirement to seek permission leaves room for officials to approve a shipment, attach conditions or refuse it.
This gives Washington a point at which to examine a transaction while the equipment is still in transit. It also explains an apparent oddity in the announcement. The government is trying to influence what happens over an internet connection, yet one of its chosen instruments is a licence for moving a physical object.
A shipping document can identify the distributor, the customer buying the equipment and the address taking delivery. All three details could be accurate and still tell an incomplete story. A data centre operator buys servers precisely so that other organisations can use them. Its customer list can change throughout the life of the equipment.
The announced licensing provision concerns covered chip re-exports. The initial statement gives no basis for concluding that everyone opening an AI application in one of the five countries now needs US permission. Precise chip thresholds, exceptions and the treatment of equipment already installed depend on the regulatory text.
What a Customer Actually Rents
A graphics processing unit, or GPU, can perform many calculations in parallel. Training a large AI model involves repeating numerical operations across vast collections of data. For a developer, access to several GPUs that can exchange information quickly can determine whether an experiment finishes in time to be worth running.
Amazon Web Services offers a familiar example of what customers can buy. An individual machine in its P5 cloud range can provide access to as many as eight Nvidia H100 GPUs. A team can also use interconnected machines for a larger training job. Amazon supplies the hardware, networking and physical infrastructure. The customer decides what computation to run. The example illustrates the service model, rather than identifying a product covered by the November rule.
Imagine a small company developing speech-recognition software. It has recordings, training code and engineers, but it needs substantial computing power for only part of the year. Renting machines for a training run spares it the expense of owning a cluster that might later sit idle. At the end of the rental, the provider can give the same machines to somebody else. The speech company can keep the model it trained.
That last detail matters for export policy. Denying a developer the opportunity to own an advanced processor has a limited effect if it can still buy enough time on equivalent hardware to finish the same project.
Why China Was Already Part of the Licensing Decision
US export controls already reached beyond a shipment’s destination before Monday’s announcement. In May 2026, the Bureau of Industry and Security clarified that an existing licensing requirement for certain advanced computing items could apply according to where the recipient, or its ultimate parent company, was headquartered. A foreign subsidiary could therefore fall within the requirement even when its own offices were elsewhere.
Remote rental presents a different case. An independent operator can buy equipment for its own premises and retain ownership while selling computing time to another business. Checking the parent company of the hardware buyer answers one question. It still leaves open who will be directing a training run six months later.
The announcement came as President Donald Trump was leaving for an Asia trip that included a planned meeting with Chinese President Xi Jinping. For Chinese AI developers, the practical issue is how much high-performance computing they can obtain, at what price and for how long. A restriction that makes a large cluster harder to reserve could delay an experiment. Its effect would be smaller for a team that already has sufficient capacity elsewhere. The announcement alone cannot establish how much computing power will become unavailable.
The Two Decisions Facing a Data Centre
For an operator planning a new facility, the first decision comes before the server racks are filled. A business ordering covered equipment for one of the named destinations needs to know whether it can obtain the necessary authorisation in time. A delay could leave it paying for a building, power and cooling before it can earn money from customers.
The second decision comes when somebody asks to reserve those machines. A university booking a cluster for its own researchers presents a relatively straightforward question about who will use it. A computing broker reserving the same cluster on behalf of several clients presents a harder one. The name on the invoice identifies who pays. The operator may need more information to understand who directs the work.
That is where the practical effect of the rule will become visible. A licence decision can determine whether servers reach a facility. The terms attached to that permission, and how operators apply them when taking bookings, will help determine whether a developer in China can still buy a week on those servers. The chip may cross a border only once. Its computing time can be sold again on Monday morning.
Sources
Next Reads
For the physical side of chip controls, follow a case involving processors allegedly shipped to China. Smugglers of Silicon: Inside the Black Market for AI Chips
For a closer look at what customers are renting, explore the processors, networking and software inside a modern AI cluster. Why NVIDIA Dominates AI: The Chips, Customers and OpenAI Deals Behind Its Empire
For the history behind the dispute, trace the decisions that took Nvidia from gaming PCs into the data centre. The History of Nvidia Through Three Machines That Changed Its Future