U.S. closes remote compute loopholes, tightening grip on Chinese AI firms’ offshore training routes

Photograph shows the U.S. Department of Commerce building in Washington DC

A new BIS rule extends export controls from chips to global compute services in five countries, threatening the overseas computing model used by Chinese AI developers.

By Zhou Yuan 

The U.S. Commerce Department has moved to restrict Chinese companies’ access to advanced AI computing power through overseas data centers, closing off a route that has allowed them to circumvent restrictions on direct chip purchases.

On Nov. 16, the Bureau of Industry and Security (BIS) issued an interim final rule titled Securing Remote Access to Advanced Semiconductors. Drafted and directed by the U.S. Treasury Department and National Security Council, the measure is designed to prevent Chinese AI companies from accessing advanced U.S. computing resources through remote services.

Until now, U.S. export controls have prevented Chinese companies from directly purchasing high-end Nvidia AI chips. But they have retained two ways to use them: renting bare-metal servers in overseas data centers and remotely accessing the machines to train models, or renting computing capacity through overseas cloud services.

This has given rise to a compromise strategy known as “train overseas, deploy domestically.” Chinese AI companies conduct computationally intensive model training using overseas resources, then run less demanding inference operations on computing infrastructure inside China.

Industry sources told Caijing that the new rules appear to cover virtually every way Chinese companies have accessed overseas computing power. The likely consequence, they said, is greater reliance on domestically produced chips.

Five countries brought into the regulatory net

The most significant change is that Washington is extending export controls beyond physical chip shipments to the provision of computing services worldwide, particularly through third-country data centers and remote access arrangements.

The rules introduce four main measures.

First, advanced chips and servers shipped to, or transferred through, Malaysia, Singapore, Thailand, Indonesia, and Brazil will generally require a license. An exemption is available where the ultimate consignee, or its ultimate parent company, is headquartered in the U.S., provided the supplier submits monthly reports.

The controls are aligned with existing U.S. semiconductor export restrictions and cover controlled GPUs including Nvidia’s latest Hopper-generation chips, not just its Blackwell series.

One industry source said Brazil’s inclusion appeared intended to address a route used by Chinese internet companies expanding overseas and renting computing capacity there.

Second, the rule sets a threshold of 12 million total TeraFLOPS Processing Performance (TPP) for remote computing services.

If a provider knows that equipment will be used to supply computing services to an entity headquartered in China, including Hong Kong, or in a country in the U.S. Department of Commerce’s D:5 country group, it must apply for a license if the aggregate capacity exceeds this threshold. Applications will be reviewed individually by BIS.

The requirement applies to data centers worldwide.

TPP is a U.S. measure of chip computing performance, broadly comparable to a chip’s horsepower. Previous export controls focused on the performance ceiling of individual chips. The new rule instead measures aggregate computing capacity.

This means adding together all the chips a data center supplies to a single customer. Where multiple data centers are interconnected, pool their resources and operate as a unified computing cluster, their capacity must also be combined.

Low threshold

The threshold is relatively low. Wu Zihao, chief executive of Ronghe Semiconductor Consulting, estimates that an Nvidia H100 has a TPP rating of approximately 16,000. The 12 million threshold therefore corresponds to around 750 H100s. For the newer Blackwell generation, it would take only a little over 300 chips, equivalent to roughly four or five server racks.

By comparison, pre-training trillion-parameter models typically requires clusters of tens of thousands of chips. Industry estimates put the computing resources needed to train a GPT-4-level model at tens of thousands of high-end GPUs, while next-generation frontier models could require more than 100,000.

The rule also requires data centers to conduct know-your-customer (KYC) checks, examining clients’ backgrounds and identifying direct and indirect ownership or control by restricted entities. This is intended to prevent third parties from providing computing services on behalf of Chinese companies.

Importantly, the restrictions cover inference as well as training. Computing capacity above the threshold is controlled whether used to pre-train models or run advanced AI applications.

Existing chips and limited exemptions

Third, the rules also cover chips already installed in data centers before their publication. If these continue supplying computing services above the relevant threshold to restricted customers, providers must generally apply for a license within 90 days of the rules taking effect. Changes in the end user’s intended use will also require renewed authorization.

Fourth, the rules provide a limited exemption for computing capacity used exclusively for product, service or content recommendations on e-commerce and social media platforms. Providers can apply for an exemption by submitting customer details, computing volumes and evidence of the intended use.

However, systems that also use generative AI to produce text, images or video do not qualify, even if they perform recommendation functions.

Model distillation remains an option

The new restrictions do not prohibit Chinese companies from purchasing AI inference services from U.S. providers overseas, leaving some room for an alternative training method known as model distillation.

This involves using the outputs of an established model as training data for a smaller student model, without directly accessing the original model’s full pre-training computing infrastructure.

But the approach has limitations. Distilled models generally cannot match the capabilities of their source models, while the provider of the original service retains access to the interaction data and earns revenue from the process.

The policy, one industry source said, has two objectives: restricting China’s access to AI computing power while protecting the commercial interests of U.S. AI model developers.

Washington has been considering remote-access controls for some time. The House version of the Remote Access Security Act (H.R. 2683) passed in January 2026 but stalled after being referred to the Senate Banking Committee. It has not become law.

Rather than wait for congressional action, BIS has now moved ahead under its existing export-control authority, introducing the restrictions through an interim final rule.

Source: 
Caijing

Share the story:
,