U.S. Clamps Down on AI Chip Exports to Block Remote Compute Access by Chinese Firms

The U.S. government is introducing sharper export controls targeting AI chips and cloud-based compute power, aiming to stop companies in China and other sanctioned nations from leveraging overseas data centers to bypass existing restrictions. The interim rule from the Department of Commerce mandates that exporting, reexporting or even transferring AI semiconductors and servers used in data centers require a special license if those systems enable compute services for Chinese firms. The rule takes effect under current U.S. administration policies and carries forward earlier efforts to tighten national security protections.

What’s New: Remote Compute Under License

The updated regulation prohibits shipping advanced AI chips, servers or related infrastructure capable of giving remote computing access to Chinese entities—no matter where the hardware is located globally. Under the rule, anything that enables compute “services” to companies in China must meet licensing requirements. This includes data centers equipped with high-powered hardware defined in terms of Total Processing Performance (TPP), a metric designed to capture compute capacity. Specifically, networks surpassing 12 million TPP—equivalent to roughly 750 Nvidia H100 or H200 chips—now fall under the new controls.

Additionally, firms delivering such compute must monitor their customers’ ownership and operations to ensure they aren’t being operated by entities from banned countries. The rule also bars using intermediaries or brokers to skirt restrictions. Exports or transfers to locations like Malaysia, Singapore, Indonesia, Thailand and Brazil are newly covered, though exemptions exist if the end-users are U.S.-based.

Policy Revisions and Fallout

One significant change is the rescinding of the 2025 AI Diffusion Rule, which had imposed limitations on exporting powerful chips and moving high-end AI models across national borders. That framework has now been replaced by the current licensing-heavy approach, which puts stronger emphasis on supervising remote access and compute delivery.

Entities that already operate under frameworks impacted by the rule must move to secure licenses within 90 days from its activation. Companies affected include chip manufacturers, cloud providers, and data center operators that facilitate model training or operations that could support foreign military or intelligence-linked applications.

The provisions exclude systems used solely for commercial recommendation engines—like those behind social media or online shopping—but are squarely aimed at advanced AI systems deemed capable of high-risk uses.

Strategic Drivers and Reactions

This policy shift comes amid concerns over China setting up or contracting overseas compute capacity to train large AI models, possibly evading earlier export constraints. U.S. authorities view advanced AI compute as increasingly tied to national security threats. The expanded controls are therefore framed as essential to hampering efforts by sanctioned nations to exploit foreign infrastructure for otherwise restricted technologies.

The rule follows the resignation of a prominent chipmaker CEO amid allegations involving one of the firm’s major customers accused of acting as a Chinese intermediary to acquire thousands of AI chips. This incident reportedly prompted U.S. officials to press for tighter enforcement and to scrutinize entities that could serve as backdoors for restricted semiconductor technologies.

Regions added under the rule as presenting “higher-than-normal risk of diversion” include several Southeast Asian nations and Brazil—reflecting U.S. concerns about redistribution of advanced chips or the technology being used indirectly by firms in sanctioned countries.

The move represents a deliberate shift from earlier frameworks to broader licensing demands hinged on compute capability, not just where hardware is shipped. It tightens rules around who has access to high-end compute remotely and introduces more scrutiny of global computing networks.

Why it matters: This rule marks a turning point in how the U.S. regulates AI hardware exports—moving from prohibiting exports to certain countries toward policing remote access and compute power irrespective of geography. Companies worldwide will need to overhaul compliance protocols, especially cloud providers and semiconductors involved in cross-border AI work. Key issues to monitor include how license reviews are handled, what thresholds become enforcement bottlenecks, and whether this pushes China to develop fully domestic compute infrastructure. Enforcement effectiveness—and how quickly firms adapt—will determine if this truly shifts the AI export landscape.