US Closes the AI Chip Cloud Loophole as Commerce Blocks China From Renting American Computing Power Abroad

In This Article

Washington has moved to close one of the most consequential gaps in its technology controls on China. On Monday, November 16, 2026, the United States Commerce Department released a new interim final rule that stops Chinese companies from renting access to advanced American artificial intelligence chips through data centres in other countries. The chips themselves never enter China. Instead, the computing power they provide is rented remotely, and for years that arrangement sat outside the export controls that govern physical shipments. The new rule ends that distinction. Any data centre anywhere in the world that rents large amounts of computing power to companies based in China or other restricted countries will now need a United States licence before it can receive the advanced chips and servers that make that computing possible. The move is the Trump administration’s first formal policy response to the scandal that forced out Nvidia chief executive Jensen Huang, and it lands three days before the rule is due to take effect in the Federal Register, on the same day President Donald Trump is scheduled to meet Xi Jinping in Shenzhen. China rejected the measure on Tuesday, setting up the latest confrontation in a technology rivalry that has become the defining economic contest of the decade.

1. What the New Rule Does

The rule, issued by the Commerce Department’s Bureau of Industry and Security and titled “Addressing Remote Access to Advanced Semiconductors”, runs to 166 pages and carries the regulatory identifier RIN 0694-AK50. At its core it does one thing: it treats remote access to advanced computing power as an export. Until now, American export controls largely governed the physical movement of chips and servers. A company could not ship a restricted processor to China, but a data centre in a third country could buy that same processor, install it in its own facility, and rent its computing power to a Chinese customer over the internet. The hardware stayed outside China, so the shipment itself was lawful, even though the practical result was the same as if the chip had crossed the border. The Bureau of Industry and Security said the rule “is designed to ensure that existing restrictions on access to sensitive American technology cannot be circumvented through remote computing arrangements involving data centres in third countries.” That single sentence is the clearest statement of the policy’s intent: the loophole is closed, and the rental market is now inside the control regime.

The mechanics of the rule are built around licensing. Companies will need a licence to export, re-export or transfer advanced artificial intelligence chips and servers to any data centre, anywhere in the world, that provides computing power above a defined threshold to companies headquartered in restricted destinations. The licence requirement does not apply to every data centre on earth. It applies only where the facility’s customers include firms based in the countries Washington has designated as restricted. That targeting matters, because it means ordinary commercial data centres serving ordinary customers in ordinary markets should see little direct change, while facilities whose business model depends on serving restricted customers face a new and serious compliance burden.

The rule is an interim final rule, which means it takes effect without waiting for the usual full rulemaking process to finish, although the public will have 60 days to comment and the department can adjust the rule in light of what it hears. The choice of an interim final rule signals urgency. The administration did not want to spend months consulting while the workaround it was trying to close remained open. It also signals confidence: the department believes the legal basis for the rule is settled and that the policy is ready to be enforced. Industry groups, data centre operators and chipmakers will now spend the coming weeks reading the 166 pages closely, because the details of how the rule is written will decide how much of the global cloud market it actually reaches.

1.1 The Cloud Loophole, Explained

The workaround the rule targets is simple to describe and difficult to police. Advanced chips are among the most controlled goods in the world, and the United States has spent years tightening the rules on where they can be sold. But a chip does not need to travel to be useful. A data centre in Singapore, Malaysia, Indonesia or any other country can buy advanced processors, install them in racks, and sell computing time to customers anywhere with an internet connection. The customer never touches the hardware. It sends data to the data centre, the data centre’s machines do the work, and the results come back. For a company training a large artificial intelligence model, that arrangement is nearly as good as owning the chips outright, and in some ways better, because it avoids the capital cost of buying them and the difficulty of maintaining them.

For Chinese companies, the arrangement was especially valuable. Direct purchases of the most advanced American chips were already restricted, and each round of controls made the restrictions tighter. Renting computing power abroad offered a way around those limits without breaking the letter of the law. The chips were legally exported to a lawful destination. The customer was in a different country. The transaction that connected them was a service agreement, not a shipment, and service agreements were not covered by the export control rules as they stood. American officials have been warning about this gap for years, and Congress has been pressing the Commerce Department to close it. The new rule is the answer to that pressure, and it reframes the entire question: what matters is not where the chip sits, but who ultimately uses its power.

The scale of the workaround is difficult to measure precisely, because the rental market is opaque and the customers are not always identified in public filings. But the fact that the Commerce Department chose to act with an interim final rule, and that the rule covers data centres worldwide rather than in a handful of countries, suggests officials believe the practice is widespread. The rule’s drafters also understood that the workaround would simply move if they only covered a few destinations, so they wrote the rule to follow the customer rather than the location. Wherever the data centre sits, if it serves a restricted customer above the threshold, the licence requirement applies.

1.2 The 12 Million TPP Threshold

The rule draws its line at 12 million aggregate total processing performance, a measure known as TPP. Any data centre, or network of linked data centres, capable of providing more than 12 million TPP of computing power to companies headquartered in a restricted destination falls within the rule. That figure works out to roughly the maximum output of about 750 Nvidia H100 or H200 chips, the workhorse processors of the modern artificial intelligence industry. The threshold is deliberately set high enough to catch serious artificial intelligence training workloads while leaving ordinary cloud computing untouched. A data centre running email, websites and ordinary business software will not come close to it. A facility training a large language model will cross it easily.

The threshold is also a statement about what Washington is actually worried about. The rule is explicitly narrowed to high-end machine learning workloads, and the Commerce Department has said standard consumer applications are untouched. The concern is not that a Chinese company might rent a server to run a website. The concern is that it might rent enough computing power to train a frontier artificial intelligence model, the kind of system that can be used for advanced research, for cyber operations, or for military applications. By setting the threshold at the equivalent of about 750 top-end chips, the rule targets the handful of facilities that can offer that scale of power, and it leaves the rest of the cloud market alone.

There is a subtlety in how the threshold is calculated that data centre operators will need to study. The measure is aggregate TPP, counted at a single facility or across linked data centres, and it applies to the computing power provided to companies headquartered in restricted destinations or whose ultimate parent is based there. That means a facility cannot escape the rule by splitting its computing power across several buildings, or by serving a restricted customer through a subsidiary registered in a different country. The rule follows the ultimate parent, not just the name on the contract. That anti-avoidance design is one of the most consequential features of the rule, and it is likely to be one of the most closely examined in the comment period.

2. Why Washington Acted Now

The timing of the rule is not an accident. It arrives at a moment when the politics of artificial intelligence, the leadership of the world’s most important chip company, and the diplomacy of a presidential summit are all colliding. The Commerce Department has been under pressure from Congress for months to close the remote access gap, and the scandal that forced Jensen Huang out of Nvidia gave that pressure a sharper edge. The rule is the administration’s first formal policy response to that scandal, and its release on a Monday, with effect scheduled for Thursday, shows an administration that wants to be seen acting decisively. The fact that the rule takes effect on the same day President Trump meets Xi Jinping in Shenzhen is either a coincidence of scheduling or a deliberate signal. Either way, it means the two leaders will be talking about trade and technology on the very day the new controls become law.

The broader context is a technology rivalry that has been intensifying for years. The United States has used export controls as its main tool to slow China’s progress in advanced computing, and each round of controls has produced a new workaround. Chips were restricted, so companies rented them. Rental was restricted in some places, so the business moved to others. The new rule is an attempt to break that cycle by regulating the use of computing power rather than the location of the hardware. Whether it succeeds will depend on enforcement, on how other countries respond, and on whether China finds yet another route to the computing power it wants. But the direction of travel is unmistakable: Washington is no longer content to control the shipment of chips. It wants to control who can use them, wherever they are.

2.1 The Jensen Huang Scandal

The rule is the Trump administration’s first policy response to the scandal involving former Nvidia chief executive Jensen Huang, and that connection explains both its timing and its framing. Huang, who built Nvidia into the most valuable company in the artificial intelligence boom, resigned amid a scandal that shook the chip industry and the administration’s technology policy at the same time. The details of the scandal have been the subject of intense reporting, but the policy consequence is clear: the administration faced questions about whether its export control regime had failed, and it needed to show that it was tightening the rules rather than letting them drift. The new rule is that demonstration.

The connection between Huang and the remote access question is direct. Nvidia’s chips are the hardware at the centre of the rule. The H100 and H200 processors named in the threshold calculation are Nvidia products, and the company’s data centre business is the largest in the world. Any rule that governs who can rent computing power built on Nvidia chips is, in effect, a rule about Nvidia’s market. The administration’s decision to frame the rule as its first response to the Huang scandal ties the policy to the person who embodied the industry it regulates, and it signals that the administration intends to be more assertive about controlling the technology, not less, in the aftermath of his departure.

It is also worth noting what the rule does not do. It does not punish Nvidia, and it does not restrict the company’s ability to sell its chips in lawful markets. What it does is extend the logic of the existing controls to a new kind of transaction. That is a significant step, but it is a step within the existing framework rather than a break from it. The administration is building on the export control system it inherited and has already been tightening, not replacing it with something entirely new. The scandal provided the political moment; the policy itself is the product of years of debate inside the government about how to control computing power in an age when the hardware and the user can be on opposite sides of the world.

2.2 The Remote Access Security Act Push

The rule is also expected to add momentum to the Remote Access Security Act, known as RASA, a bill in Congress that would write similar restrictions into law. The relationship between the rule and the bill is important. A rule issued by the Commerce Department can be changed by a future administration. A law passed by Congress is far harder to undo. Supporters of the bill have argued for months that the executive branch needed to act, and the administration’s decision to issue an interim final rule gives them a concrete policy to point to when they argue that the approach is workable. It also gives opponents something concrete to criticise, which is the normal course of a major policy change.

The bill’s supporters in Congress have been among the most vocal critics of the remote access gap. They have argued that allowing Chinese companies to rent American computing power makes a mockery of the export controls, and they have pressed the Commerce Department to act faster. The new rule answers that criticism, but it also raises the stakes for the legislative debate. If the rule works well, the case for the bill becomes stronger: the policy will have been tested in practice, and Congress will be asked to make it permanent. If the rule creates problems, whether for American companies or for allies, the bill’s critics will have fresh evidence to cite. Either way, the rule has moved the debate from the abstract to the concrete, and the coming months will show whether the policy survives contact with the real world.

The timing of the rule also matters for the legislative calendar. With the rule taking effect this week and the comment period running for 60 days, Congress will be debating the bill while the department is receiving feedback from industry. That overlap means the legislative debate will be informed by the practical experience of the rule’s first months, which is unusual and potentially valuable. Lawmakers will be able to ask data centre operators and chipmakers what the rule actually costs them, rather than relying on predictions. That evidence could shape the final form of the legislation in ways that no amount of testimony could.

3. The Countries and Companies in the Crosshairs

The rule’s scope is defined by a list, and the list is what gives the policy its bite. The licence requirement applies when a data centre provides computing power above the threshold to companies headquartered in Macau or in a Country Group D:5 destination. That group includes China, Russia, Iran and North Korea, along with 18 other countries. The inclusion of Macau alongside the D:5 countries is significant, because Macau has been used as a routing point for technology shipments in the past, and its inclusion closes off an obvious alternative route. The rule also looks through corporate structures to the ultimate parent, so a company cannot escape the controls by registering a subsidiary in a country that is not on the list.

China is the primary target, and no one involved in the policy pretends otherwise. The rule was written with Chinese artificial intelligence companies in mind, and the public debate about it has focused almost entirely on China. But the rule’s design is deliberately broader, because the administration does not want to create a precedent that applies to only one country and leaves the same loophole open for others. Russia, Iran and North Korea are all subject to extensive American sanctions already, and the new rule adds another layer of restriction on their access to advanced computing. The 18 additional countries in the D:5 group are a reminder that the export control system is a global framework, not a single-country policy, even when one country dominates the discussion.

3.1 The D:5 List and China

Country Group D:5 is a category in the American export control system that covers countries subject to the most restrictive controls, and its members are those Washington considers the highest risk for the diversion of sensitive technology. China is the largest economy in the group by an enormous margin, and it is the only member with a globally competitive artificial intelligence industry that depends on access to advanced computing. That is why the rule matters so much more for China than for any other country on the list. Russian, Iranian and North Korean companies have far less capacity to make use of rented computing power at the scale the rule targets, and their access was already heavily restricted by other measures. China is the country with both the demand and the money to rent computing power at scale, and it is the country whose artificial intelligence ambitions the United States is most determined to slow.

The rule’s treatment of Chinese companies is also notable for what it does not require. It does not ban all computing services to Chinese customers. A data centre can continue to serve Chinese companies below the threshold without a licence, and it can serve them above the threshold if it obtains one. The licence requirement is a control, not a prohibition, and the Commerce Department will have discretion over which licences to grant. That discretion is a powerful tool, because it allows the department to distinguish between different kinds of customers and different kinds of uses. A licence for a university research project is a different proposition from a licence for a company with military links, and the department will be able to make that distinction case by case. How it exercises that discretion will be one of the most closely watched features of the new regime.

For Chinese companies, the practical effect is likely to be higher costs, longer delays and greater uncertainty. Renting computing power abroad was attractive precisely because it was fast and straightforward. Adding a licensing step introduces delay, and the possibility that a licence will be refused introduces risk. Some companies may decide the new process is not worth the trouble and turn to domestic alternatives, which is one of the outcomes Washington may be hoping for, since it would push Chinese firms towards less capable hardware. Others may try to structure their arrangements to stay below the threshold, which is why the rule’s anti-avoidance provisions matter. The cat-and-mouse dynamic that has characterised the export control regime for years is likely to continue, but the mouse now has less room to run.

3.2 Five New Licensing Destinations

Alongside the remote access control, the rule adds a licence requirement for advanced chips shipped to five countries: Indonesia, Malaysia, Singapore, Thailand and Brazil. These are not restricted destinations in the ordinary sense. They are major economies, several of them close American partners, and all of them are home to significant data centre industries. Their inclusion is a recognition that the remote access workaround has been concentrated in precisely these kinds of countries: places with the infrastructure to host large data centres, the connectivity to serve global customers, and the legal systems that made them attractive locations for the rental business. By requiring licences for shipments to these five countries, the rule adds a second layer of control that operates alongside the remote access provision.

The inclusion of Singapore is particularly significant. Singapore is one of the world’s largest data centre hubs, and it has been a focal point of the debate about remote access for years. The city’s combination of reliable power, undersea cable connections and a business-friendly environment made it a natural location for facilities serving the whole of Asia, including China. Requiring licences for advanced chips shipped to Singapore will be felt immediately by the data centre operators there, and it will force them to think carefully about which customers they serve. The same is true, on a smaller scale, for Malaysia, Indonesia and Thailand, all of which have been growing their data centre industries rapidly and competing for the same business. Brazil’s inclusion extends the logic beyond Asia, signalling that the administration sees the workaround as a global phenomenon rather than a regional one.

The five countries will now need to decide how to respond. None of them wants to be seen as a conduit for evading American export controls, but none of them wants to lose the data centre investment that has been flowing into their economies either. The licence requirement does not ban shipments to these countries, and licences for facilities serving lawful customers should be obtainable. But the additional compliance burden is real, and it will change the economics of building data centres in these markets. Governments in the region will be watching closely to see how the Commerce Department handles licence applications, and how quickly it processes them, because the speed of the licensing process will determine whether the new rule is a manageable adjustment or a serious drag on their digital economies.

4. How Data Centres Will Have to Comply

For the data centre operators caught by the rule, the next three months will be a scramble. The rule gives existing operators 90 days from its effective date to obtain a licence if they want to continue providing the affected services, and that clock starts when the rule is published in the Federal Register. Ninety days is not a long time to navigate a licensing process that many operators have never used before, and the industry’s first task will be to work out whether it is even covered. That means auditing customers, mapping which of them are headquartered in restricted destinations or owned by parents based there, and calculating how much computing power each of them receives. For a large facility with hundreds of customers, that is a substantial exercise, and it will need to be completed before any licence application can be filed.

The compliance burden is not only administrative. Data centre operators will need to build systems that can monitor their customers’ usage on an ongoing basis, because the licence requirement is triggered by the amount of computing power provided, not by a one-off transaction. A customer that starts below the threshold and grows above it changes the facility’s obligations. A customer that changes its ownership structure changes them too. Operators will need to know, at any given moment, how much computing power each restricted customer is using, and they will need to be able to demonstrate that knowledge to the Commerce Department if asked. That is a new kind of compliance for an industry that has traditionally thought of itself as a landlord for servers, not a gatekeeper for technology policy.

The rule also creates a difficult commercial question. Operators that serve restricted customers will need to decide whether the business is worth the licensing risk. A licence application can be refused, and a refusal would mean losing the customer. Even a licence that is granted comes with conditions and with the possibility that it will not be renewed. Some operators may decide that the safest course is to stop serving restricted customers altogether, which would achieve Washington’s objective without the department having to refuse a single licence. Others may decide to apply and hope for the best. The choices they make in the coming weeks will shape the structure of the global cloud market for years.

4.1 The 90-Day Window

The 90-day window is the rule’s most immediate deadline, and it is deliberately tight. Existing data centres that are already providing the covered services will have 90 days from the rule’s effective date to obtain a licence. That means they must identify their exposure, prepare their applications, and submit them, all while continuing to serve their customers. The Commerce Department has not said how quickly it will process applications, and the industry will be watching the first decisions closely for signs of how the process will work in practice. A licensing system that processes applications in weeks is a manageable burden. One that takes months would make the 90-day window effectively impossible to meet, and would force operators to suspend services while they wait.

The window also creates a test of the department’s own capacity. The Bureau of Industry and Security is a relatively small agency, and it will now be responsible for licensing a category of activity it has never regulated before. The number of applications it receives in the first 90 days will be a measure of how large the remote access market actually is, and the speed with which it processes them will be a measure of how seriously the administration has invested in enforcement. If the department is overwhelmed, the rule’s credibility will suffer. If it processes applications efficiently, the rule will be seen as a workable model for future controls. The first 90 days will therefore be watched not only by the industry but by other governments considering similar measures.

There is also a question about what happens to operators that miss the deadline. The rule does not spell out a grace period beyond the 90 days, and operators that continue to provide the covered services without a licence after the window closes would be in violation of the export control rules. That exposes them to enforcement action, including fines and the loss of their own ability to receive American technology. The stakes are therefore high, and the incentive to comply quickly is strong. Industry associations are already advising their members to begin the audit process immediately, before the rule even takes effect, so that applications can be filed on the first day the window opens.

4.2 What Counts as Remote Access

The definition of remote access is the heart of the rule, and it is where the most difficult questions will arise. The rule applies when a data centre provides computing power to a company headquartered in a restricted destination, and the computing power is measured in aggregate TPP. But the modern cloud is a complicated place, and the line between providing computing power and providing a service that happens to use computing power is not always clear. A company that rents a virtual machine and runs its own software on it is plainly receiving computing power. A company that buys a finished software service, such as a customer relationship management tool, is receiving the output of computing power, not the power itself. The rule is aimed at the first case, but the boundary between the two will need to be worked out in practice.

The rule’s drafters have tried to make the boundary clear by focusing on high-end machine learning workloads and by setting a threshold that only serious computing users will cross. A company buying an ordinary software service will not be using 12 million TPP of computing power, and the data centre providing that service will not need a licence. But the distinction may become harder to draw as artificial intelligence is built into more and more software. A customer relationship tool that includes an artificial intelligence assistant is using computing power in a way that a traditional tool did not, and the question of whether that counts as remote access to advanced computing will need an answer. The 60-day comment period is likely to produce detailed submissions on exactly this point, and the department may need to issue guidance to clarify the boundary.

The rule also has to deal with the reality that computing power is often shared. A data centre’s machines serve many customers at once, and the same physical hardware can be used by a lawful customer in one moment and a restricted customer in the next. The rule measures the computing power provided to the restricted customer, which means operators will need to be able to attribute usage to individual customers with reasonable accuracy. That is technically possible, since cloud providers already meter usage for billing purposes, but it requires the metering to be accurate enough for regulatory purposes, and it requires the operator to know which of its customers are restricted. Both of those requirements are new, and both will require investment in systems and processes that many operators do not currently have.

5. China Pushes Back

China rejected the new rule on Tuesday, November 17, the day after it was announced, and the speed of that rejection tells its own story. Beijing has been preparing for tighter American controls on artificial intelligence for years, and its response was ready before the rule was even published. The rejection was framed in the familiar language of trade disputes: the measure is discriminatory, it violates the principles of free trade, and it is an attempt to suppress China’s development. That language is now routine in these disputes, but the substance behind it matters. China is the country most affected by the rule, and its government has made the development of domestic artificial intelligence a national priority. A rule that cuts off a major source of computing power is a direct challenge to that priority, and Beijing’s response was never going to be quiet acceptance.

The rejection also matters for the diplomacy of the week. The rule takes effect on November 19, the day President Trump meets Xi Jinping in Shenzhen, and China’s public rejection of the measure on Tuesday sets the tone for that meeting. The two leaders will be discussing trade, technology and the future of the relationship between the world’s two largest economies, and the new rule will now be part of that conversation whether either side wants it to be or not. The meeting was already expected to cover nuclear arms reductions and the war with Iran, with Ukraine off the agenda according to American officials. The addition of the artificial intelligence controls to the agenda makes the meeting even more consequential, and it raises the stakes for both sides.

5.1 Beijing’s Rejection

China’s rejection of the rule follows a pattern that has become familiar in the technology rivalry. Beijing argues that American export controls are an abuse of national security justifications, that they harm the global technology industry, and that they will ultimately damage American companies as much as Chinese ones. There is some truth in the last of those claims: American chipmakers and cloud providers do lose business when controls are tightened, and the industry has repeatedly warned that overly broad restrictions will push customers towards non-American suppliers. But the argument has not persuaded Washington in the past, and there is no sign that it will persuade the current administration now. The rule was issued in full knowledge of how Beijing would respond, and the administration plainly decided that the national security benefit outweighed the commercial cost.

What Beijing does next matters more than what it says. China has several options. It can accelerate its investment in domestic chip production, which it is already doing at enormous scale. It can seek computing power from suppliers outside the American orbit, although the rule’s global scope makes that harder than it used to be. It can retaliate against American companies operating in China, a step it has taken in past disputes. Or it can use the issue as leverage in the negotiations with Washington, offering concessions on other matters in exchange for a softening of the controls. The most likely outcome is a combination of all of these, pursued simultaneously and over a long period. The technology rivalry is not a single dispute to be settled in a single meeting. It is a structural feature of the relationship between the two countries, and it will be managed rather than resolved.

5.2 The Impact on Chinese AI Firms

For Chinese artificial intelligence companies, the rule is a serious practical problem. The most advanced models in the world are trained on hardware that China cannot buy directly, and renting that hardware abroad was one of the few remaining routes to the computing power needed to compete at the frontier. The rule does not close that route entirely, since licences can be granted, but it makes the route slower, more expensive and less certain. For a company racing to train a model before a competitor, delay can be as damaging as a refusal. The companies most affected will be those working on the largest models, the ones that require thousands of advanced chips running for weeks or months. Those are precisely the projects the rule is designed to slow down.

The impact will not be evenly spread. Large Chinese technology companies with their own data centres and their own chip supplies will be less affected than smaller firms that depend on rented computing power. That may accelerate a consolidation in the Chinese artificial intelligence industry, as smaller companies find it harder to compete and larger ones absorb their talent and their projects. It may also accelerate China’s investment in domestic chip production, since the lesson of the rule is that dependence on foreign computing power is a vulnerability. The Chinese government has been subsidising its domestic chip industry heavily for years, and the new rule will be cited as further justification for that spending. In the long run, the rule may make China more self-sufficient, which is the opposite of what Washington intends, but that is a risk the administration has evidently decided to accept.

There is also a human dimension that is easy to overlook in a discussion of export controls. The engineers and researchers who train artificial intelligence models are among the most sought-after people in the world, and they go where the computing power is. If Chinese firms cannot offer their researchers access to the best hardware, some of those researchers may leave, either for companies in other countries or for the large Chinese firms that can still provide the resources they need. The rule therefore affects not only companies but careers, and its long-term effect on the global distribution of artificial intelligence talent may be as significant as its effect on any single model.

6. What It Means for Nvidia and the Chip Industry

No company is more central to this story than Nvidia. Its H100 and H200 chips are the standard against which the rule’s threshold is measured, and its data centre business is the engine of the artificial intelligence boom. The rule does not restrict Nvidia’s ability to sell chips in lawful markets, but it does change the environment in which those chips are used. A data centre that buys Nvidia chips will now need to think about who rents the computing power those chips provide, and that changes the commercial calculation for everyone in the supply chain. Nvidia itself will need to consider whether its customers are complying with the new rules, and whether it has any responsibility for how its products are used after they are sold.

The industry’s response to the rule will be shaped by the experience of the past few years, in which each round of controls has been followed by a period of adjustment and then by a return to growth. Chipmakers have learned to live with export controls, and they have built compliance systems to manage them. But the remote access rule is different in kind from the controls that came before it, because it regulates the use of chips rather than their sale. That is a new frontier for the industry, and it will require new thinking about what a chipmaker’s responsibilities are. The companies that adapt quickly will have an advantage. Those that do not may find themselves caught out by enforcement action.

6.1 Nvidia’s New Leadership Moment

The rule arrives at a moment of transition for Nvidia. Jensen Huang, the chief executive who turned the company from a maker of graphics cards for gamers into the most important supplier of artificial intelligence hardware in the world, has resigned amid a scandal, and the company is in the process of choosing his successor. President Trump has publicly backed Tim Cook, the former Apple chief, to take the role, calling it a case of one legend replacing another. Whether Cook takes the job or not, the leadership change comes at a time when the company’s relationship with Washington has never mattered more. The new chief executive will need to manage not only the company’s technology and its markets but also its position in a geopolitical contest that the company did not choose and cannot avoid.

The leadership transition also raises a question about the company’s strategy. Under Huang, Nvidia built its dominance by selling the most powerful chips to anyone who could pay for them, and by building a software ecosystem that made those chips indispensable. That strategy depended on open markets and on the free movement of technology. The new era of export controls, of which the remote access rule is the latest example, depends on the opposite: on restricting who can use the technology and where. Nvidia’s next leader will need to navigate that contradiction, and to do so while the company’s valuation, its workforce and its customers are all watching closely. It is one of the most consequential leadership transitions in the history of the technology industry, and the rule is a reminder of how much is at stake.

For the moment, the company’s position is strong. Demand for its chips remains enormous, and the rule does not reduce that demand in lawful markets. If anything, by restricting the rental market, the rule may push more customers towards buying chips directly, which would benefit Nvidia’s sales. But the longer-term picture is more complicated. If the controls accelerate the development of competing chips in China, Nvidia could eventually face a rival that it helped to create by forcing Chinese companies to build their own hardware. That is the paradox of export controls: they slow a rival down in the short term, and they may make the rival stronger in the long term. Nvidia’s new leadership will need to think about that paradox every day.

6.2 The Cloud Providers Caught in the Middle

The cloud providers are the companies most directly affected by the rule, because they are the ones who will need to obtain the licences and monitor their customers. The largest providers, including the American hyperscalers and their international competitors, have extensive compliance operations and will be able to adapt. But the rule also affects a long tail of smaller data centre operators, many of which have never dealt with the Bureau of Industry and Security before. For those companies, the compliance burden may be disproportionate to their size, and some may decide that the simplest response is to stop serving restricted customers altogether.

The cloud providers also face a commercial dilemma. The customers affected by the rule are, by definition, among the most demanding users of computing power, and they pay accordingly. Losing them means losing high-value business. Keeping them means accepting the cost and risk of licensing. The providers’ decisions will shape the market: if most providers stop serving restricted customers, the price of computing power for those customers will rise sharply, and the market will split into two tiers, one for lawful customers and one for everyone else. If most providers apply for licences and receive them, the rule will have added a layer of bureaucracy without fundamentally changing who gets access to computing power. The truth is likely to lie somewhere between those two outcomes, and the industry’s behaviour in the first 90 days will be the best guide to which way it is heading.

There is also a question about how the rule affects the competition between American and non-American cloud providers. A provider based outside the United States that does not use American chips is not subject to the rule at all, which creates an incentive for customers to seek out non-American hardware. That incentive already exists, and the rule strengthens it. The long-term effect may be to accelerate the development of a non-American computing ecosystem, built on chips from China or from other countries that are not subject to American controls. That is a slow process, and American chips remain far ahead of the alternatives, but the direction of travel is clear. Every control the United States imposes creates an incentive for someone else to build an alternative, and the remote access rule is no exception.

7. The End of the AI Diffusion Rule

The new rule formally rescinds the Biden administration’s AI Diffusion Rule, a framework introduced in January 2025 that set country-level limits on advanced chip exports. The Trump administration had already said it would not enforce that rule, and the Bureau of Industry and Security now says it was never enforced. Its formal repeal is therefore less a change of policy than a confirmation of a decision already made. But the repeal matters, because it clears the decks for a different approach. The AI Diffusion Rule tried to control the spread of advanced computing by sorting the world into tiers and setting different limits for each. The new rule takes a narrower approach: it targets a specific workaround, the remote rental of computing power, and it leaves the rest of the system largely as it was.

The difference between the two approaches is instructive. The AI Diffusion Rule was broad and ambitious, and it attracted criticism from industry and from allied governments who felt it was too sweeping and too difficult to implement. The new rule is narrow and specific, and it is aimed at a problem that almost everyone agrees exists. That contrast suggests the administration has learned from the earlier debate, and that it prefers targeted controls that can be enforced over broad frameworks that generate resistance. Whether the narrower approach works better remains to be seen, but the shift in method is itself a significant development in American technology policy.

7.1 What the Biden-Era Rule Tried to Do

The AI Diffusion Rule was an attempt to manage the global spread of advanced artificial intelligence by dividing the world into three tiers. A small group of close allies would have essentially unrestricted access to advanced chips. A middle group would face caps on how much computing power they could import. A third group, including China and other restricted countries, would face the tightest limits of all. The logic was that American technology should be shared generously with friends and withheld from adversaries, and that the middle tier needed to be managed carefully so that it did not become a conduit for diversion to the third tier. It was, in concept, a coherent framework, and it reflected a serious attempt to think about the global diffusion of artificial intelligence as a single problem.

In practice, the rule ran into trouble almost immediately. Allied governments objected to being placed in the middle tier and subjected to caps, arguing that they were being treated as untrustworthy. Industry groups warned that the caps would be difficult to administer and would damage American competitiveness. And the incoming Trump administration, which took office days after the rule was issued, made clear that it did not intend to enforce it. The rule was never implemented, and its formal repeal this week is the final step in a process that began almost as soon as it was announced. The episode is a reminder that a policy, however well designed, needs political support to survive, and that a change of administration can undo years of work in a matter of weeks.

7.2 A Sharper, Narrower Approach

The new rule represents a different philosophy. Rather than trying to manage the global diffusion of computing power through a tiered system, it targets a single, specific practice: the remote rental of advanced computing power to restricted customers. That narrowness is its strength. It is easier to explain, easier to enforce, and harder to object to, because it addresses a gap that is widely recognised as a problem. Allied governments are less likely to object to a rule that targets Chinese companies renting computing power than to a rule that caps their own imports. Industry is less likely to object to a rule that affects a small number of facilities than to one that reshapes the entire global market.

The narrower approach also reflects a change in how the administration thinks about the problem. The AI Diffusion Rule was built on the assumption that the United States could control the spread of computing power by controlling where chips were shipped. The new rule is built on the recognition that the shipment of chips is only part of the story, and that the use of computing power matters as much as its location. That is a more sophisticated understanding of the technology, and it suggests that future controls will be similarly targeted at specific uses rather than broad categories. The administration has, in effect, moved from controlling the hardware to controlling the service, and that shift may prove to be the most consequential feature of the new policy.

Whether the narrower approach succeeds will depend on enforcement, and on whether the targeted controls can keep pace with the ingenuity of those trying to evade them. The history of export controls suggests that evasion is a constant, and that each new control produces a new workaround. But the narrower approach has an advantage: it is easier to update. A rule that targets a specific practice can be amended when that practice changes, without reopening a debate about the entire framework. The administration has given itself a tool that is flexible as well as focused, and the coming months will show whether flexibility is enough.

8. What Happens Next

The immediate future of the rule is governed by dates. The rule takes effect when it is published in the Federal Register, which is scheduled for November 19, 2026. From that date, the 60-day comment period begins, and the 90-day window for existing data centres to obtain licences starts to run. That means the industry has until the middle of February to file its licence applications, and the department has until the middle of January to hear what the public thinks of the rule. Those deadlines will structure the debate, and they will determine how quickly the rule’s effects are felt. The first licence decisions, expected in the weeks after the rule takes effect, will be the clearest signal of how the department intends to enforce it.

Beyond those dates, the rule’s future will be shaped by three things: the response of industry, the response of China, and the response of Congress. Industry’s response will be visible in the licence applications and in the comment submissions. China’s response will be visible in its diplomacy and in its investment in domestic chip production. Congress’s response will be visible in the progress of the Remote Access Security Act, which the rule is expected to accelerate. All three will be playing out at the same time, and they will interact in ways that are difficult to predict. The rule is the beginning of a process, not the end of one.

8.1 The Federal Register Clock

The Federal Register publication date of November 19 is more than a bureaucratic detail. It is the date the rule becomes law, and it is the date from which the 90-day and 60-day clocks start to run. The choice of that date, coinciding with the Trump-Xi meeting in Shenzhen, means the rule will be at the centre of the diplomatic conversation from the moment it takes effect. It also means the administration will be able to point to a concrete action when it discusses technology policy with Beijing, rather than a promise of future action. That is a strong negotiating position, and it is likely to be part of the administration’s calculation.

The publication date also matters for the industry, because it sets the deadline for compliance. Data centre operators that have been following the debate will have started their audits already, but those that have not will need to move quickly. The 90-day window is not generous, and the licensing process is unfamiliar to many of the companies it now covers. The department has an interest in making the process work smoothly, because a licensing system that collapses under its own weight would undermine the rule’s credibility. The first weeks after publication will therefore be a test of the department’s administrative capacity as much as of the policy itself.

There is one more detail worth noting. The rule is an interim final rule, which means it takes effect immediately but remains open to comment and to change. The department will read the comments it receives over the 60-day period, and it may amend the rule in response. That means the rule as published is not necessarily the rule as it will exist in six months. Industry groups, allied governments and other interested parties will use the comment period to argue for changes, and the department will need to decide which of those arguments to accept. The rule is therefore a living document, and its final form will be shaped by the debate that follows its publication.

8.2 The Bottom Line for the AI Race

The bottom line is that the United States has taken a significant step to control who can use its most advanced computing technology, and it has done so by extending its reach from the shipment of chips to the rental of computing power. The rule closes a loophole that had become central to how Chinese companies accessed the hardware they need, and it does so with a mechanism that is targeted, enforceable and difficult to evade without substantial cost. Whether it succeeds in slowing China’s progress in artificial intelligence will not be known for years, and the history of export controls suggests that the answer will be complicated. But the direction is clear, and the commitment behind it is serious.

For the wider world, the rule is a reminder that the artificial intelligence race is not only a contest of research and engineering. It is also a contest of policy, of export controls and licensing regimes, of diplomacy and industrial strategy. The companies that build the chips and the data centres that house them are now part of that contest, whether they want to be or not. The rule will reshape the market for computing power, it will change the calculations of cloud providers and chipmakers, and it will add another layer of complexity to a relationship between the United States and China that is already the most consequential in the world. The coming weeks will show how the industry adapts, how Beijing responds, and how the two leaders talk about it when they meet in Shenzhen. The answers to those questions will shape the technology landscape for years to come, and the rule that takes effect on Thursday is where the next chapter begins.

Author

Comments

comments

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top