China spent two years discovering that the most powerful AI chips in the world are available by wire. If Beijing cannot import Nvidia's H100s or H200s through customs, its AI labs can simply rent them from data centers in Singapore, Malaysia, or Thailand, where the chips sit legally under existing U.S. export rules. On Monday, November 16, 2026, the Commerce Department's Bureau of Industry and Security ended that arrangement. The agency issued an interim final rule requiring U.S. export licenses for any advanced semiconductor shipment to an overseas data center that provides compute to companies in arms-embargoed countries, with a 90-day compliance deadline for facilities already in the market.
The Rule That Rewrites the Global AI Cloud Market
The BIS interim rule formally rescinds the Biden administration's AI Diffusion Rule, a January 2025 framework that set country-level limits on advanced chip exports. The Bureau of Industry and Security confirmed that the new rule was written at the direction of the Treasury Department and the National Security Council, an unusual arrangement that one analyst described as a rare rift in Washington's export-control bureaucracy. The Biden framework was never enforced: the Trump administration said in May 2026 it would discard the country-level cap approach, leaving a gap that Chinese AI firms immediately exploited by renting compute from Southeast Asian data centers rather than trying to import restricted hardware.
The new rule fills that gap at the transaction level rather than the country level. TechCrunch first reported that BIS requires a license for any shipment of advanced AI chips or servers to an overseas facility if the exporter knows or has reason to believe that facility provides computing power to companies headquartered in arms-embargoed nations, including China, Russia, Iran, and North Korea. The rule covers not just direct shipments but also transfers through brokers, resellers, and any intermediary. Exporters and data center operators must monitor their customers' ownership structures and activities to ensure restricted entities are not accessing compute through affiliates or shell companies.
BIS also designated five additional countries facing new licensing requirements: Singapore, Malaysia, Thailand, Indonesia, and Brazil, which Commerce says "present a higher-than-normal risk of diversion." Chips destined for U.S.-owned subsidiaries operating in those countries are exempt. The rule is expected to publish in the Federal Register on November 19, 2026, the same day as the scheduled Trump-Xi summit in Shenzhen. Data centers already operating in the market have 90 days from that date, closing around mid-February 2027, to obtain licenses or restructure their customer base.
How 750 Nvidia H100s Became the Legal Threshold
The rule does not capture every cloud provider that sells GPU time. The licensing requirement activates only at facilities supplying more than 12 million total processing performance, the BIS standard unit for measuring AI chip capacity. TechCrunch's reporting calculated that 12 million TPP is roughly equivalent to 750 Nvidia H100 or H200 GPUs, a cluster size well within range of a commercial AI training or inference operation but above the threshold of small research or enterprise deployments. The 12 million TPP floor also exempts commercial recommendation systems, the kind used by e-commerce platforms and social media services, which typically run lighter specialized models rather than general-purpose large language models.
The threshold number is not explained in the rule text, and lawyers who cover export controls immediately flagged a consequential ambiguity: Commerce has not specified whether the 12 million TPP limit applies per individual customer allocation, per physical data center, or per a company's total network of facilities worldwide. That distinction could determine whether medium-sized GPU rental operators in Malaysia and Singapore fall under the rule or not. Commerce is expected to issue clarifying guidance but had released none as of Monday. Data center operators are unlikely to invest in compliance programs until that guidance arrives, which could effectively shorten the usable 90-day window to six or eight weeks of real compliance time.
The rule does not apply retroactively to chips already installed. Commerce confirmed that existing hardware in overseas data centers is not subject to new licensing under the rule. Only future exports, re-exports, and transfers require authorization. That distinction matters: a Singapore facility that already has 2,000 H100s on the floor today does not need a license for those chips. It needs a license before it accepts any additional GPU shipments from U.S. vendors while maintaining Chinese customers above the TPP threshold.
Why Singapore, Malaysia, and Three Others Are in the Crosshairs
The five designated countries were not chosen arbitrarily. U.S. officials have been tracking the flow of advanced AI chips through Southeast Asian data centers for at least two years. CBS News reported that the rule arrives days after Commerce Secretary Lutnick faced public criticism for failing to stop earlier diversion of Nvidia chips to China. Reuters had reported on November 12 that Chinese AI firms were accelerating their efforts to lock in overseas computing capacity in anticipation of exactly this kind of rule, signing longer-term contracts with Southeast Asian data center operators before tighter controls arrived.
Megaspeed, a Singapore-based GPU rental company, said on November 15 that it operates lawfully and has retained outside counsel for an internal review. Commerce has not publicly named Megaspeed as a violator, but the company became the focal point of the controversy surrounding Jensen Huang's departure from Nvidia: Huang reportedly had a relationship with a former Megaspeed executive whom U.S. authorities suspect of connections to Chinese intelligence. Megaspeed has denied that any intelligence activity occurred and says its former executive left the company in 2025. China's government separately warned the United States against tightening chip restrictions, calling the move an overreach.
India is not among the five designated countries, a decision that industry observers say reflects both the depth of the U.S.-India semiconductor partnership and the relatively smaller role Indian cloud providers currently play in the GPU rental market. The Next Web noted that the exclusion of India from the list gives U.S. officials room to expand the partnership as an alternative compute destination for non-Chinese firms displaced from the five designated markets. That gap is also likely to be tested commercially: if operators in designated Southeast Asian countries cannot serve Chinese clients without navigating a lengthy U.S. licensing process, that compute demand will not disappear. It will move to alternative destinations, and Commerce will face pressure to expand the designated list or rely on broader export-control regulations to close the new corridor.
The Nvidia CEO Exit That Pushed the Rule Across the Finish Line
The BIS announcement cannot be separated from the most consequential leadership exit in Silicon Valley this year. Jensen Huang resigned as Nvidia's CEO on November 12 after reports of a relationship with a former executive at a company under Commerce scrutiny, following questions from Treasury Secretary Scott Bessent about whether Nvidia leadership had pushed sales that may have been illicit to boost revenue and earnings. Nvidia CFO Colette Kress became interim CEO. Huang has been asked to testify before the Senate Banking Committee on December 3.
Nvidia had already begun preparing for tighter controls. CNBC reported on November 17 that Nvidia warned data-center customers that the Remote Access Security Act, known as RASA, could expose both the company and its clients to legal liability over Chinese access to Nvidia chips. RASA, which would give the U.S. government broader authority to regulate cloud access to chips in the same way it regulates physical hardware exports, passed the House in January but has not cleared the Senate. Nvidia also told buyers, through verbal instructions reported by the Financial Times, that it would not sell chips to customers renting more than 50% of their GPUs to Chinese companies under future contracts.
Nvidia shares fell as much as 2.9% on Monday following the BIS announcement. AMD dropped 5.6%, Broadcom fell 3.9%, and CoreWeave, which rents out Nvidia chips to AI developers, lost 6.1%. Trump said at the White House on Monday that selling less advanced chips is "a good thing, not a bad thing." He was traveling to Asia for the November 19 Shenzhen summit with Chinese President Xi Jinping. Trump also said he "would not mind" if Nvidia reinstated Huang as CEO, a signal that the administration's position on Nvidia's leadership situation remains unsettled. Bloomberg covered the full market reaction, noting that the rule's scope extends beyond Nvidia to any U.S. semiconductor vendor whose products are used in overseas AI data centers serving restricted countries.
Why Critics Argue the Rule Arrives Too Late to Matter
The bear case for this rule is straightforward and not easily dismissed. Export-control rules have consistently failed to slow China's AI progress because enforcing them requires visibility into transactions that occur entirely on foreign soil. The BIS rule requires exporters to monitor customer ownership structures, but the Commerce Department has limited investigative reach in Singapore or Brazil. Critics argue that Chinese firms will simply migrate compute demand to undesignated countries, create shell companies in jurisdictions not on the five-country list, or accelerate domestic alternatives like Huawei's Ascend AI chip series, which has steadily improved in capability and is designed precisely to replace Nvidia products for Chinese buyers.
Skeptics point out that the rule may also arrive too late for its primary objective. If China's leading AI labs, including Moonshot AI and others, have already secured enough overseas GPU capacity under long-term contracts signed before November 16, the new licensing requirement affects future transactions but not existing compute arrangements. CNBC reported that Moonshot AI sought Nvidia chip access through European data centers, suggesting Chinese firms have already diversified beyond Southeast Asia. A licensing requirement that covers only new transactions may succeed in preventing future arrangements while doing nothing to unwind the capacity already in place.
The Council on Foreign Relations analyst Chris McGuire called the rule "a major step forward," but the five designated countries have pushed back. Southeast Asian governments invested heavily in data center infrastructure expecting to serve regional AI demand, and U.S. licensing requirements put that business model at risk. Thailand and Indonesia have argued that their own AI development programs depend on access to U.S. chip technology. The rule was also issued without the standard public comment period, with BIS citing national security grounds to bypass notice-and-comment rulemaking. That shortcut opens it to legal challenge from affected operators, particularly in Singapore, which has an active trade law community and a government that has been vocal about its sovereignty over data infrastructure decisions.
What to Watch Before Mid-February and Beyond
Three markers will determine whether the BIS rule is an enforcement milestone or a placeholder. First, watch for Commerce guidance on the 12 million TPP threshold calculation, expected within the next two to four weeks. The ambiguity over whether the limit applies per customer, per facility, or per operator network is not academic: an operator in Malaysia with 1,200 total H100s serving dozens of customers, including some Chinese firms, needs to know whether it falls under the rule before it can begin the license application process. Commerce's guidance will narrow or widen the rule's effective scope, and the industry is treating that document as more important than the rule itself.
Second, the December 3 Senate Banking Committee hearing with Jensen Huang will determine whether RASA advances in the next Congress. If Huang's testimony produces new evidence that Nvidia sales practices knowingly facilitated Chinese chip access, Senate leadership will face pressure to pass the bill, which would give the BIS rule statutory backing rather than relying solely on executive authority. That matters for durability: an executive-level rule can be rescinded or modified by any future administration, while a statutory framework is significantly harder to unwind and would bind both Commerce and the courts to treat remote access as a physical-hardware-equivalent export.
Third, watch the November 19 Shenzhen summit outcome. The BIS rule was timed to arrive just before Trump's meeting with Xi Jinping, as was the Biden AI Diffusion Rule before his summit in 2025, suggesting the rule functions in part as a negotiating instrument. If Trump and Xi reach a side agreement on technology access as part of a broader trade package, the BIS rule could be modified, its enforcement delayed, or specific country exemptions granted. The 90-day window closing around mid-February 2027 is the first hard deadline the market will watch to see whether the rule bites or gets bargained away. By that date, the semiconductor industry will know whether Washington is serious about the cloud-access enforcement it has been signaling for two years.
The rule does not stop China from building AI. It stops the United States from supplying the cloud on which China builds it.
Key Takeaways
- 12 million TPP (~750 Nvidia H100s) is the threshold : BIS requires a U.S. export license for advanced AI chip shipments to any overseas data center above this compute level that serves Chinese or other restricted companies.
- Five countries designated: Singapore, Malaysia, Thailand, Indonesia, Brazil : existing operators in these countries have 90 days from November 19 to obtain licenses or restructure their customer base.
- Biden AI Diffusion Rule formally rescinded : the January 2025 country-level cap framework was never enforced by the Trump administration and is now replaced by this transaction-level rule targeting remote access rather than physical shipments.
- Nvidia fell 2.9%, AMD fell 5.6%, CoreWeave fell 6.1% : chip and GPU rental stocks dropped sharply Monday as the market priced in tighter restrictions on overseas AI data center operations.
- Jensen Huang testifies December 3 : the former Nvidia CEO, who resigned November 12 amid a scandal involving a Megaspeed executive suspected of ties to Chinese intelligence, faces the Senate Banking Committee in a hearing that could shape RASA's legislative fate.
Questions Worth Asking
- If Chinese AI labs have already secured long-term GPU rental contracts before November 16, does the BIS rule meaningfully slow their compute access, or does it only raise costs at the margin for future capacity expansion?
- The rule was written at Treasury's direction rather than Commerce's initiative. Does that signal a shift in how Washington frames AI chip controls: less as trade policy and more as national security finance?
- Southeast Asian governments built data center capacity expecting to serve AI demand from across the region. If U.S. licensing now restricts that model, which countries step in to fill the gap, and do they face the same restrictions within the next 12 months?