November 18, 2026

CNAS Insights | Unpacking the Remote Access Rule

On November 16, the Commerce Department’s Bureau of Industry and Security (BIS) released an interim final rule (IFR) to address remote access—the practice of Chinese companies accessing advanced American chips through the cloud that they are otherwise banned from importing. Access to top performing chips is essential to train and deploy advanced AI. According to one estimate, by the end of the year, remote access will account for 23 percent of China’s access to AI chips. Commerce released the IFR amid the Nvidia scandal and ongoing allegations that the department had failed to add Megaspeed, a Nvidia customer suspected of selling remote access to Chinese firms, to the Entity List. Below are key insights to understand what the IFR does, what it gets right, and where the administration and Congress can go further.

What the IFR Does

The IFR takes three main actions. First, it introduces a license requirement for chips exported to data centers abroad serving customers headquartered in China or other countries of concern above a threshold of 12 million total processing performance (TPP), the equivalent to about 760 Nvidia H100 chips. Data centers already serving these customers must obtain a license to continue doing so within 90 days. Second, it introduces a new license requirement for advanced chip exports to Indonesia, Malaysia, Singapore, Thailand, and Brazil. Third, it formally rescinds the AI Diffusion Rule, a Biden-era framework that sorted every country into tiers that determined their access to advanced AI chips from U.S. companies.

What the IFR Gets Right

The IFR represents a meaningful first step toward closing the remote access loophole in chip export controls. The 12 million TPP threshold strikes the right balance. For context, it represents about a third of the computing power DeepSeek reportedly used to train its V3 model. This threshold is high enough to omit small-scale uses of cloud computing but low enough to catch deals that potentially involve training AI models, such as the ones in Figure 1. Smaller-scale chip rentals could still be concerning: An AI developer could stay under the threshold by splitting model training across several cloud providers. But the highest-risk deals involving Chinese AI labs easily exceed the threshold.

Figure 1: The new remote access rule’s threshold covers major commercial deals, while leaving out low-risk cases

Each bar shows a reported remote access arrangement in terms of H100 equivalents. The red line represents the IFR’s threshold of about 760 H100 equivalents; anything to the right of the line now requires a license.

The IFR targets five countries where remote access risk is most concentrated. The IFR introduces a new license requirement for shipping advanced AI chips and related servers, software, and technology to five countries: Indonesia, Malaysia, Singapore, Brazil, and Thailand. Most of these countries have suspected or confirmed cases of remote access by Chinese entities. Malaysia, which has the fourth-most AI compute in the world, counts many Chinese AI labs like ByteDance among its data center customers. The country has also been at the center of high-profile investigations of firms like Megaspeed suspected of remote access and chip smuggling. Moonshot AI reportedly trained its Kimi K3 model using Nvidia Blackwell chips accessed remotely from a data center in Thailand. Although there haven’t been any publicly confirmed cases of remote access in Brazil, ByteDance is currently building a $10 billion data center and is reportedly weighing a second-phase expansion.

The IFR allows U.S. companies to build in the five countries. U.S. companies can import chips to these countries without a license if they report the shipments to BIS. This incentivizes these countries to partner with U.S. firms, complementing the Trump administration’s initiative to export the American AI stack abroad.

The IFR is more inclusive of U.S. allies and partners than the Biden administration’s AI Diffusion Rule. The AI Diffusion Rule invited widespread pushback, notably from the European Union, for leaving 17 EU member states out of Tier 1, which faced no limits on chip imports. The new rule flips the approach, adding no new license requirements for most partners outside the five named countries. This aligns with the Trump administration’s goal to pursue a “bold, inclusive strategy” for exporting AI to trusted foreign partners, while controlling the highest-risk cases.

Where the IFR Falls Short

The IFR omits data centers in the United States. Investigative reports from Reuters and The Information have documented several instances of Chinese state-linked groups procuring compute from AWS, Microsoft, and Google. Although the public reports did not mention the location of the U.S.-owned data centers serving Chinese requests, a more effective policy would cover remote access regardless of where chips sit, even if they’re in the United States.

he IFR overrelies on chip exporters’ “knowledge” of whether a data center serves Chinese firms above the threshold. In practice, this knowledge rests on what data centers report about their own customers, and the IFR does not set a standard for what information they should collect or verify. A stronger approach would be a mandatory know-your-customer (KYC) regime requiring cloud providers serving American chips, at home and abroad, to verify their large compute customers’ identity and report suspected remote access violations to BIS and other relevant government agencies.

Building on the IFR

The Senate is considering complementary legislation, the Remote Access Security Act, to help close gaps in the new IFR. The bill would give BIS jurisdiction to regulate remote access itself—not just new transactions involving remote access customers.

This authority would also enable BIS to establish a rigorous KYC regime to help cloud providers and the government monitor high-risk cases of remote access. But that authority alone may not be enough. BIS weighed a rule mandating KYC in 2024, but it was never finalized amid industry pushback that existing privacy laws prevented them from sharing customer information with the government and that KYC requirements would undermine U.S. cloud competitiveness vis-à-vis Chinese cloud providers. The Cloud Security Act, introduced by Representatives Josh Gottheimer and John Moolenaar, would remove the privacy obstacle. And a well-designed regime could address the competitiveness concern by applying only to customers above a high compute threshold, lightening the regulatory burden.

Enforcing the IFR requires greater BIS capacity. The bureau’s current enforcement capacity remains stretched thin. It only has one export control officer for all of Southeast Asia—an insufficient presence in the region at the center of several remote access and chip smuggling cases. The congressional funding picture offers a path forward. Both chambers’ appropriations committees have proposed major increases to BIS’s $235 million budget, with the House bill supporting hiring more export control officers abroad and the Senate proposing at least an additional $30 million for advanced AI chip export enforcement.

The IFR is a welcome step forward in closing the remote access loophole, but real gaps remain. BIS should pair the IFR with a new KYC regime for cloud providers to ensure rigorous monitoring of remote access incidents. In addition, Congress can help close gaps by passing the Remote Access Security Act and Cloud Security Act and appropriately funding export control enforcement abroad. Together, these measures will help the United States maintain its AI lead over China.

Michelle Nie Duncan is a visiting fellow with the CNAS Technology & National Security Program.

  1. The author’s calculation is based on a SemiAnalysis estimate of an Nvidia H100 at roughly 15,800 TPP, meaning the IFR’s threshold works out to about 760 H100s. ↩
  2. Michelle Nie Duncan. Calculations are based on reporting by SemiAnalysis, The New York Times, Culper Research, The Wall Street Journal, Financial Times, and Reuters. H100 equivalents assume 15,831 TPP per H100. ↩

View All Reports View All Articles & Multimedia