The US Commerce Department is drafting an unprecedented new export control rule, with the core aim of prohibiting Chinese AI companies from remotely leasing high-performance Nvidia GPU computing power through data centers in third countries such as Thailand, Singapore, Malaysia, and Japan. According to The Information, a small team within the Commerce Department's Bureau of Industry and Security (BIS) is leading the drafting effort, with the draft expected to be circulated to AI companies and industry associations for comment as early as September 2026.
This marks a fundamental upgrade in US export control logic. Over the past several years, Washington has built a computing power blockade against China by restricting chip exports, but the current export controls only cover the buyers and actual holders of chips, not the remote use of computing power through overseas data centers. Chinese companies do not need to purchase restricted chips; they only need to open an account on servers in Southeast Asia, pay, and start using the computing power.
Trigger Point: Moonshot AI and Kimi K3
Michael Kratsios, director of the White House Office of Science and Technology Policy (OSTP), publicly accused Moonshot AI of secretly holding servers equipped with Nvidia GB300 chips and training its flagship model Kimi K3 — a model reportedly with 2.8 trillion parameters — by accessing a GB300 computing cluster in Thailand. Kratsios further alleged that Moonshot AI developed a sophisticated internal platform for large-scale distillation of top US models, dynamically switching between different access channels to evade detection.
The GB300 belongs to Nvidia's Blackwell chip family and is explicitly restricted by US export controls from being sold to China. Kratsios did not make public specific evidence such as Kimi K3's training logs or chip procurement routes. Moonshot AI has not yet publicly responded to the above allegations.
According to CNBC, ByteDance, Alibaba, and Tencent have all been accused of using Nvidia computing power through Asian data centers in Thailand, Malaysia, Japan, and elsewhere. Tencent open-sourced Hy4-preview on August 29, 2026, a MoE model with 770 billion parameters and a 1 million token context window.
A Loophole of Their Own Making
In the final days of the Biden administration, the "AI Diffusion Rule" was introduced, which included a "foundry due diligence rule" requiring data centers providing computing services to Chinese customers to conduct strict "Know Your Customer" (KYC) checks.
In May 2025, the Trump administration fully rescinded the aforementioned AI Diffusion framework, citing "over-regulation" and "harm to US competitiveness," and explicitly stated it would not enforce the due diligence requirements within it. This decision removed constraints on Southeast Asian computing intermediaries. The Commerce Department is now attempting to piece together an alternative that is functionally similar but narrower in scope.
Assistant Secretary Jeffrey Kessler, who oversees BIS matters, stated clearly at a House hearing in July 2026 that he had no intention of launching a replacement for the Biden AI Diffusion Rule, and directly criticized the rule as "a bad rule to begin with." Yet weeks later, the agency under his oversight was internally drafting a new rule of a similar nature.
The Hollow Legal Foundation
Export control lawyers at the law firm Baker McKenzie told The Information that the industry generally agrees the Commerce Department does not have statutory authority under the current legal framework to regulate remote computing power access. BIS's enforcement authority derives from the Export Administration Regulations (EAR), which were designed to control the export of tangible goods, not cross-border data service access.
Closing this legal gap requires congressional legislation. The Remote Access Security Act (RASA) would grant BIS explicit authority over remote computing power access. The bill passed the House in January 2026 by a vote of 369 to 22, and is currently pending in the Senate Committee on Banking, Housing, and Urban Affairs.
If the new rule is formally issued before RASA passes, it will almost certainly face immediate legal challenges, and courts are highly likely to halt its enforcement on the grounds of lack of authority.
The Industrial Logic Behind the Escalated Controls
If the rule ultimately takes effect, the impact will be structural. The training computing power sources of China's top AI companies can be divided into three categories: domestic compliant chips (mainly pre-restriction H100/A100 inventory and Huawei's Ascend series), domestic alternatives procured through legitimate channels, and high-end Nvidia computing power accessed remotely through Southeast Asian data centers. The third category, being closest to the frontier in performance, serves as a key supplement for maintaining model competitiveness. Michelle Nye, visiting fellow at the Center for a New American Security (CNAS), noted: "If they cannot access the most powerful computing resources, China's ambitions to catch up with the latest US AI models will be noticeably frustrated."
For hyperscale players like ByteDance and Alibaba that possess substantial data center assets of their own, the impact is more about rising costs and reduced flexibility; for startups like Moonshot AI that rely on external computing resources for rapid iteration, the blow is more direct.
The inference side will also be affected. Compared to training, inference has higher latency requirements, and the viability of providing real-time inference services through cross-border remote access was already limited. However, batch inference and data processing tasks also depend on high-performance computing power. If remote access channels are cut off, Chinese models will face significantly increased pressure on both cost and speed.
Southeast Asia's Dilemma
The rule's spillover effects on third parties are also present. Countries like Thailand and Singapore have been actively courting investment from the Nvidia ecosystem in recent years, and a data center construction boom is in full swing. If the new rule requires these data centers to implement KYC checks when providing computing services to Chinese customers, or even comprehensively prohibits providing services to specific entities, operators will face the dual pressure of soaring compliance costs and the loss of Chinese customers. Nvidia's Southeast Asian computing sales will also be directly hit.
In March 2026, BIS issued another regulatory draft that was withdrawn after just over a week due to strong opposition from the US AI industry.
Assessment
The US export control system is undergoing a logical upgrade: from controlling the physical flow of silicon chips to controlling the bit-level flow of computing power. The direction is right, but the difficulty of enforcement far exceeds that of controlling chips. Chips can be seized; computing power access is an API call that happens in milliseconds and is distributed across multiple jurisdictions worldwide.
The more fundamental dilemma is that this administration abolished the KYC framework that could have constrained such behavior in May 2025, and is now rebuilding a narrower version without supporting legal authority. This policy path is destined to subject the new rule to dual pressure from courts and the industry.
The final outcome is likely not an across-the-board computing blockade, but rather tiered verification requirements after negotiation and compromise — similar to an extension of the financial industry's KYC system into the computing domain. For Chinese AI companies, this means rising compliance costs and the closure of some channels, rather than a complete cutoff of computing power. But for teams that rely on overseas top-tier computing power to maintain their training pace, the window of opportunity is narrowing.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接