The ban on selling high-end AI chips directly to China has been in place for years.
But many people are unaware that another route has remained open over the past two years.
Instead of shipping the chips back to China, companies leave them in overseas data centers and lease the computing power remotely.
Plenty of companies have taken this route.
In October 2026, the U.S. think tank IFP estimated that by the end of 2027, roughly one-third of China's effective AI computing power would come from remote access to U.S. AI chips.
Then, on November 12, NVIDIA CEO Jensen Huang, who had consistently opposed shutting China out of the market, abruptly resigned.
Just four days later, the U.S. issued a 132-page rule that brought compute leasing under chip export controls for the first time.
The rule applies regardless of which country hosts the data center.
Many are asking what the new rule actually means. Let's unpack it layer by layer.
What Are the New U.S. Rules?
On November 16, 2026, the U.S. Department of Commerce issued its new rule on Addressing Remote Access to Advanced Semiconductors, which formally took effect on November 19.
This time, Washington is no longer content merely to stop chips from entering China. It is reaching directly into the contracts of overseas data centers.
The rule's core logic is simple: any provider offering computing power above an aggregate threshold of 12 million TPP must obtain U.S. authorization.
What does “12 million aggregate TPP” mean in practice?
Research institutions estimate that it is equivalent to about 750 NVIDIA H100 or H200 GPUs.
Training a large model now routinely requires tens of thousands of GPUs, so the threshold captures almost every training project of meaningful scale.
Even small and midsize AI startups are effectively shut out if they need to lease GPUs overseas.
The harshest provision is a 90-day grace period, which starts a countdown for Chinese companies and overseas cloud providers.
Washington's message is that these overseas cloud providers must apply for U.S. authorization within the next 90 days.
If they fail to secure approval by the end of that period, they must terminate all computing power supplied to Chinese customers.
If they refuse, they will be in violation of the rules and face U.S. sanctions.
In effect, overseas cloud providers must choose between losing their Chinese customers and being sanctioned by the U.S.
The rule has set off a fierce public debate. Some mock Washington for trying to starve its rivals, while others believe the measure will accelerate the replacement of foreign computing power with domestic alternatives.
But chip restrictions have existed for years. Why have Chinese companies still been able to lease so many GPUs overseas?
How This Route Emerged
The old U.S. approach to chip controls was straightforward: investigate the buyer.
The first two rounds of restrictions initially barred advanced chips from being sold into China, then extended the ban to overseas subsidiaries of Chinese companies.
But that framework had a fundamental flaw: the U.S. regulated the physical transfer of chips, not the right to use them remotely.
As long as the chips remained in overseas data centers, Chinese companies could lease their computing power over a network, creating a legal gray area.
That gray area widened dramatically in 2025.
As the U.S. changed administrations, Joe Biden introduced the AI Diffusion Rule on January 15, leaving this back door open.
Donald Trump took office on January 20, and the Commerce Department scrapped the old rule on May 13. But its replacement was slow to arrive.
The resulting regulatory vacuum, which lasted more than a year, gave Chinese companies an opportunity to secure computing power.
With no effective restrictions in place, Chinese companies began stockpiling capacity aggressively.
Citing industry sources, Caixin reported that of the roughly 1.5 GW of data center projects approved in Thailand, 60% to 70% either have Chinese backing or primarily serve major Chinese technology companies.
Companies were prepared to pay almost any price to secure resources:
The Financial Times reported on September 30 that Tencent had signed a five-year lease with Oracle to reserve about 100,000 advanced AI chips in Southeast Asian data centers, in a deal worth roughly $7 billion. But on November 4, the newspaper reported that Oracle had halted the deal under pressure from the U.S. government.
On November 5, CNBC reported that Moonshot AI was seeking to lease a cluster of more than 8,000 Blackwell chips in Europe. The proposed terms required an upfront payment 30% to 50% higher than for non-Chinese customers.
The rush to secure GPUs at a premium put NVIDIA under the most immediate pressure.
NVIDIA is not only the seller but also a direct target of U.S. export-policy oversight. If Chinese companies accumulated too much capacity, the U.S. government was bound to push back.
According to a November 3, 2026 Financial Times report, NVIDIA began conducting internal checks over the summer and verbally instructed cloud providers to limit Chinese customers to no more than 49% of new contracts.
But NVIDIA's concession did not calm Washington.
Congress sought to define remote GPU leasing as an export, but the relevant legislation remained stalled in the Senate.
With the legislative process moving too slowly, the executive branch decided to act without Congress.
Jensen Huang's resignation came at this critical moment.
Caixin reported on November 17 that the new rule was drafted directly by the White House National Security Council and the Treasury Department and forced through despite internal concerns at the Commerce Department.
The prevailing view is that Huang's departure removed a leading pro-business voice, leaving Washington's hardliners firmly in control and determined to cut off Chinese companies' remote access to overseas computing power.
But these data centers operate on Malaysian and Singaporean soil. What gives the U.S. the authority to decide which customers they can serve?
How the U.S. Can Regulate Data Centers Abroad
The answer lies in the supply chain: if these data centers want to keep buying new NVIDIA servers, they must follow U.S. rules.
The force of the new approach lies in shifting the focus from inspecting goods to verifying identities.
The U.S. has effectively turned suppliers such as NVIDIA into intermediaries.
To demonstrate compliance, sellers must require data centers to disclose their full customer lists.
If a data center refuses to cooperate, the U.S. can determine that the seller knowingly enabled the violation and cut off its access to future supplies.
To protect their own business, sellers will inevitably pressure data centers to disclose every detail about their Chinese customers.
With that leverage in place, the U.S. has also moved preemptively to close off every obvious workaround:
First, relocating the data center will not help.
Whether a data center is in Europe or Japan, it must obtain authorization if it exceeds the threshold. Indonesia, Malaysia, Singapore, Thailand and Brazil, all popular destinations for Chinese companies leasing computing power, are specifically named.
Second, using shell entities will not help either.
Even if a company breaks up a data center or leases GPUs across multiple regions, systems connected to perform the same tasks will be treated as affiliated centers, with their computing power aggregated against the threshold.
Third, having a foreign company sign the contract will not work.
The rule requires scrutiny of the underlying ownership structure. If the computing power ultimately serves a Chinese company, or the resulting model ends up in the hands of a Chinese company, the transaction must be declared even if a foreign entity signs the contract.
Fourth, companies cannot exploit loopholes after a deal is signed.
Even if a purchase was compliant at the outset, any later change that provides Chinese customers with computing power above the threshold must be reported immediately.
Earlier rules had more loopholes because Washington was still testing its approach.
For example, an early draft dated November 12 covered only Blackwell chips and did not apply retroactively to existing data centers.
The final version is far tighter. It covers mainstream chips including the H100 and H200, while existing data centers must also file disclosures within 90 days.
Researchers once estimated that the earlier restrictions would affect only 14% of China's computing power.
The final rule not only closes the remote-access route but also targets the existing overseas computing power that Chinese companies had previously obtained in compliance with the rules.
With the route so comprehensively blocked, will Chinese companies be completely unable to lease computing power in the future?
Approval, Not an Outright Ban
The new rule does not seal off the route entirely. Instead, Washington has adopted a more calculated approval system.
The document does not impose a blanket ban. Instead, it stipulates that:
Every transaction involving advanced computing power must receive U.S. government approval.
Because this is an approval regime, enforcement can be tightened or relaxed. The rule leaves three openings:
First, it gives U.S. companies a back door.
For business in five Southeast Asian countries, no new authorization is required if the end customer or parent company is American.
The condition is that sellers must disclose every detail, reporting the parties to each transaction, chip models, computing-power scale and other information to the Commerce Department on the first business day of every month.
In plain terms, U.S. giants can still make money in Southeast Asia, but Washington will scrutinize every transaction.
Second, it gives lower-risk businesses a way to continue operating.
Computing power leased for tasks such as ranking e-commerce products or recommending short videos may remain in use after being declared.
But the red line is rigid: if a system can generate text, images, audio or code, it will be strictly regulated as a large model, even if it also performs recommendation tasks.
Third, the U.S. is abandoning a broad global crackdown to focus specifically on China.
The new rule directly scraps the Biden-era framework that divided countries into different tiers.
Washington's current strategy is clear: stop burdening allies indiscriminately and concentrate its pressure on Chinese customers and high-risk transit hubs such as Southeast Asia.
Since companies can still apply, does that mean approval is simply a matter of filing the paperwork?
Quite the opposite. This approval system is more punishing than an outright ban.
In the document, the Commerce Department estimates that the rule will generate only about 100 additional applications per year, with total new compliance costs of just $1,900.
That sounds trivial, but the market understands that the real weapon is not the $1,900 in fees. It is delay.
Submitting an application does not guarantee approval, and approval does not guarantee a timely decision. Data centers may also refuse orders altogether to avoid liability.
That is why Southeast Asian countries whose economies benefit from data centers were the first to react once Washington claimed control over who could access computing power.
On November 16, two core Singapore government agencies issued urgent statements saying they were negotiating with the U.S. in an effort to secure whitelist status for local companies.
Industry observers see the move as an upgrade from physically cutting off supply to controlling access: U.S. chips can still be sold and U.S. companies can still make money, but Washington will decide, case by case, who gets access to large-scale computing power.
The slower those approvals arrive, who will buckle first?
Who Gets Hurt First?
The first to suffer under the new U.S. rule will not be Chinese companies, but the small and midsize overseas cloud providers that depend on Chinese customers.
One Singapore cloud provider said there simply are not enough non-Chinese customers outside the U.S. and that strict enforcement would inevitably hurt business.
Once the 90-day deadline passes, providers that refuse to hand over customer lists will have no choice but to watch their expensive servers sit idle.
Chinese large-model companies are next in line.
Model training is not like gaming. It depends heavily on continuity.
Once large-model training begins, it is like bringing water to a boil: the computing power must be supplied continuously.
Previously, leasing overseas computing power was a commercial contract. Pay the money and receive the service.
Now it has become a regulatory gamble. Even with a signed contract, training could be forcibly interrupted if the U.S. government delays approval or suddenly demands additional documents over “geopolitical risk.”
If the computing power disappears before the data run is complete, a training program costing hundreds of millions of yuan can collapse, wiping out all prior investment.
This is not simply a price increase. It is political risk that can derail a project at any moment.
As overseas options narrow, where will this demand for computing power go?
Only one direction remains: back to China.
Can Domestic Computing Power Meet Demand?
Chinese AI companies previously allocated computing power globally, using the best chips to train models as quickly as possible.
The U.S. is now forcibly cutting off that route, leaving companies to meet their computing-power needs within China.
The global computing-power market is splitting in two: a U.S.-controlled “authorization zone” and a Chinese market built around domestically produced systems.
NVIDIA is still selling chips, but its guidance for the past several quarters has already assumed zero data center revenue from China.
Domestic suppliers such as Huawei Ascend are rapidly filling the space NVIDIA once occupied.
China's intelligent computing capacity has grown by more than 70% in a year. NVIDIA's market share in China has fallen to 34%, with Huawei close behind at 28%.
That would have been unimaginable two years ago.
More importantly, Huawei has open-sourced its CANN software, offering a way around NVIDIA's CUDA ecosystem, the biggest barrier to adoption. Domestic chips are finally moving from merely functional to practical.
Industry observers say the U.S. effort to constrain Chinese AI through computing-power controls has instead forced the emergence of a self-contained Chinese computing-power ecosystem.
But a domestic ecosystem is no guarantee of security.
Demand that previously went to overseas computing power is now being driven back into China. The only question is whether domestic chips can absorb it.
The answer hinges on HBM.
HBM is the high-bandwidth memory used in AI chips. Each AI chip requires several HBM units and cannot be delivered without them.
HBM is now in short supply globally, and procurement costs are rising, directly pushing up prices for domestic chips.
In the market, prices for domestic chips have risen 20% to 50% in just two months. Demand is spilling over so quickly that supply and demand have reached a critical point.
Even with money in hand, buyers may not receive their chips on schedule.
The decisive issue for domestic computing power has therefore shifted from whether it works to whether suppliers can deliver.
The companies that can secure HBM capacity, reliably connect tens of thousands of GPUs into clusters and maintain stable deliveries without being choked by foreign restrictions will capture the vast market NVIDIA is leaving behind.
Industry observers say U.S. chip controls have steadily moved deeper over the past several years.
First, Washington controlled the goods, barring advanced chips from being sold into China. Then it targeted buyers, preventing overseas subsidiaries of Chinese companies from purchasing them. Now it is regulating usage rights: the GPUs can be located anywhere, but Washington decides who may use them and how much capacity they may access.
Chinese companies are no strangers to restrictions that tighten step by step.
China's difficulties with GPS in the 1990s eventually led to the BeiDou system.
When Huawei lost access to Google services in 2019, it was pushed to develop HarmonyOS. Each time access was cut off, a domestic alternative eventually emerged.
But none of those alternatives was built in a matter of months.
BeiDou took more than two decades to progress from project launch to global coverage.
Large models are trained in successive cycles, and competitors will not stop and wait.
Can China's domestic computing-power suppliers deliver the chips these companies need before the next training cycle begins?

Comments
00No comments yet. Be the first to weigh in.