On November 16, 2026, the US Commerce Department issued an interim export control rule: any data center or compute network with total processing performance exceeding 12 million TPP (roughly equivalent to 750 Nvidia H100 or H200 GPUs) must apply to the US government for a license before providing remote AI compute access to users in restricted countries.
The core of the rule is to close the indirect-access loophole. Directly selling high-end AI chips to China is already banned, but establishing data centers in Singapore, Malaysia, Thailand, Indonesia, or Brazil and then renting remote compute to Chinese companies was previously not explicitly prohibited. Commerce Department officials stated: "This control is a logical extension of existing controls. These restrictions will weaken the ability of companies in restricted countries to use advanced US semiconductors to train advanced AI models."
How Big Was the Loophole: One Company Tripped the Regulatory Trigger
The specific case that triggered the rule was Megaspeed. The US government determined that the company purchased large quantities of Nvidia chips and was essentially acting as a compute proxy for Chinese companies. The Commerce Department has begun adding Megaspeed to the Entity List.
A few days before the rule was released, Nvidia co-founder and CEO Jensen Huang announced his resignation. The trigger was a scandal linked to Megaspeed—individuals involved had connections with suspected Chinese intelligence personnel. Huang had previously been one of the most recognizable business lobbyists in US AI policy toward China.
The rule is a technical patch for a specific loophole and also sends a political signal: regulators' tolerance for lobbying pressure on chip export issues has significantly decreased.
Scope of Controls: Why Brazil and Southeast Asia Are Included
The rule distinguishes two restricted categories. The first targets D:5 embargoed countries, including China, Russia, Iran, North Korea, and 18 other countries. The second targets five intermediate hub countries with "above-normal diversion risk": Singapore, Malaysia, Thailand, Indonesia, and Brazil.
The 12 million TPP threshold corresponds to 750 H100s. The rule is not aimed at small academic clusters or startup test environments, but at compute infrastructure with industrial-scale training capability.
Who Will Be Materially Affected
Operators running large data centers in the five intermediate hub countries face a choice: apply to the Commerce Department for a license and accept review, or restrict remote services to Chinese customers.
For China's AI industry chain, some large-model companies and cloud service providers previously accessed overseas high-end compute through third-party channels. Once the licensing system covers this, the actual probability of approval is close to zero.
The Southeast Asian data center market has expanded rapidly in recent years. The new rule makes the growth logic of cross-border compute leasing legally unstable.
Loophole Logic vs. Technical Reality: Can Such Rules Keep Up?
Remote access controls face enforcement difficulties: how can data center operators verify users' ultimate affiliation. VPNs, multi-hop proxies, and shell company structures can obscure the true destination of compute access. The rule requires "reasonable knowledge" rather than "complete confirmation," which in practice will generate a large number of gray-area compliance disputes.
After each past AI chip export control rule was issued, workarounds were found within months. The H100 export ban spawned the H800; after the H800 was restricted, the shift went to cloud access. This rule targets the cloud access path and is a form of catch-up regulation.
It usually takes months from formulation to implementation for a rule, while new detours in the compute leasing market often iterate within weeks. The rule's effect is to raise costs and friction, not to create physical isolation.
Independent Assessment
The significance of the controls lies in marking a shift in the mindset of US regulators: they no longer tolerate diversion risk in intermediate hub countries out of concern for harming allies' commercial interests. Jensen Huang's resignation provided a political window; at this moment, no company dares openly oppose tightening regulation.
The rule will accelerate two forces: China's investment in self-reliant compute infrastructure will speed up further; Southeast Asian data center operators will separate two business architectures—serving restricted-country users and not serving restricted-country users—with costs shifted to end users.
The loophole is closed, and the search for new loopholes begins simultaneously. The evaluation criterion is whether the friction coefficient is high enough that alternative paths cost more than their commercial value. Judging from the precise design of the 12 million TPP threshold, the Commerce Department's approach is pragmatic.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接