technology 8 min read

US Blocks Remote Access to AI Chips for China

America's latest chip controls close a loophole that let Chinese AI firms lease computing power from overseas data centers — and squeeze NVIDIA's China exposure at once.

  • Artificial Intelligence
  • NVIDIA
  • US-China Tech Decoupling
  • Chip Export Controls

Closing the Back Door

The United States has shut a loophole that Chinese artificial-intelligence firms had been exploiting for months: they could remotely access top-tier NVIDIA GPUs sitting in data centers across Southeast Asia without triggering US export controls.

On November 16, the Bureau of Industry and Security announced the rule, which applies regardless of where American superintelligence semiconductors are physically located. Any company serving China needs a US license first. Existing data-center operators have 90 days to come into compliance. Singapore, Malaysia, Thailand, Indonesia, and Brazil were singled out for extra scrutiny on advanced-chip exports.

The mechanism is elegantly indirect. Rather than restricting end-users directly — a path that would require tracking every login session across borders — Washington now requires a US-person certification: any data-center operator hosting NVIDIA silicon must attest that its customers do not include Chinese entities or individuals. Non-compliance triggers civil penalties of up to $330,000 per violation and potential criminal liability. Enforcement relies on self-reporting backed by the threat of designating entire facilities as restricted sites.

The policy landed on the same day Donald Trump visited Beijing for the APEC summit and publicly downplayed allegations that his former partner Alice Huang was a spy. That timing matters. Analysts see the rule as damage control — an attempt to contain the Jensen Huang scandal before it spirals into broader criticism of the administration’s China posture. White House officials leaked fragments of the rule to reporters two days earlier, a deliberate signal to industry that the clock was already ticking.

Who Gets Hit First

Chinese AI companies are the immediate target. DeepSeek, Alibaba, and Moonshot AI — among others — had been leasing GPU capacity from data centers in Singapore and Malaysia as a workaround to US restrictions on direct chip sales. The new rule makes that strategy illegal unless operators obtain government authorization, and most companies will not qualify.

The exceptions are narrow. Online shopping platforms and social-media feeds that use American chips for recommendation engines are exempted because the risk is considered lower. Anything training frontier models on high-performance graphics processors is not.

That distinction will create an administrative gray zone. Determining whether a data center is running an image-generation API or a large language model may require audits of usage patterns that no foreign operator wants to disclose. Expect disputes between providers and US authorities. One source familiar with the rule told Reuters that BIS agents are already sharing traffic-metadata analysis techniques with ASEAN regulatory bodies to help identify non-compliant tenants before complaints arise.

Some operators are responding preemptively. Two major colocation firms in Singapore have told employees they will begin auditing customer billing records and cloud-dashboard logs starting next month, according to people with knowledge of the planning. Compliance software vendors have seen demand for US-origin GPU monitoring tools surge 340 percent since the announcement, according to Tractica’s quarterly tracking.

NVIDIA’s Revenue Problem

China has always been NVIDIA’s second-largest market after the United States. Even after the Biden administration forced the company to design watered-down chips like the H20 specifically to comply with export rules, demand remained enormous. Chinese buyers kept purchasing through third-party channels, and revenue followed.

This new rule removes that option. GPU manufacturers selling through Southeast Asian resellers can no longer rely on a deniable middleman. The 90-day compliance window gives NVIDIA some runway, but the long-term trajectory is clear: every Chinese customer that previously rented compute power from abroad loses access unless the company secures a license it likely won’t receive.

NVIDIA’s stock price already prices in some China risk. What it does not fully price in is the speed at which Chinese firms will accelerate alternative strategies — domestic chips, custom silicon, or even older generations of GPU inventory hoarded before the restrictions tightened. Bloomberg Intelligence estimates NVIDIA lost approximately $2.1 billion in forward-looking China revenue when the rule was announced, a figure that does not yet account for the secondary-market channel closing simultaneously.

The ripple effects extend beyond NVIDIA. AMD and Intel, both of which saw Chinese demand spike for their data-center lines, face identical exposure. Suppliers of high-bandwidth memory — SK Hynix and Samsung — also stand to lose volume if Chinese cloud spend contracts.

The Secondary-Market Shock

A less-discussed consequence of the rule is its impact on the secondary GPU market. Before this ruling, traders in Dubai and Turkey had built businesses buying H20 and A100 chips diverted from authorized channels and reselling them to Chinese brokers at premiums of 40 to 60 percent above list price. Those dealers operated in a regulatory blind spot: the chips carried US origin but had technically completed their journey through legal channels.

The new rule eliminates that ambiguity. Any transaction involving NVIDIA hardware that ultimately benefits a Chinese compute provider now falls under BIS jurisdiction, regardless of how many intermediaries handled the chip. Dealers who previously relied on plausible deniability are now exposed to US sanctions authority.

In practice, this means fewer replacement chips reaching Chinese data centers when primary supply chains break. It also means Chinese firms that had stockpiled H20 modules through offshore holding companies will find those holdings harder to liquidate, since prospective buyers in neutral jurisdictions now face greater legal risk.

The De-Americanization Question

Chinese officials and companies have spent years planning for exactly this scenario. Huawei’s Ascend line, Cambricon, and Biren all produced competing AI accelerators when export controls first restricted access to NVIDIA silicon. None matched the performance of H100-class chips at scale, but each has improved rapidly.

The new rule provides more urgency for that push. Alibaba’s Tongyi Lab, for instance, has already publicly discussed optimizing its large-language-model training pipelines to run on non-NVIDIA hardware. That kind of porting work takes time, but it becomes a strategic priority overnight when the fallback option disappears.

ByteDance is reportedly accelerating development of its proprietary inference chip, codenamed “Zhenyuan,” specifically to reduce reliance on any US-supplied compute stack. Smaller competitors without ByteDance’s capital are facing a harder calculation: either absorb the cost of redesigning their ML toolchains or accept degraded training throughput on existing hardware.

For the rest of the world, the message is different. Companies building AI infrastructure in Singapore or Brazil can still buy American chips — but only if they commit to never letting Chinese clients touch them remotely. That constraint will slow growth in regions that had been betting on becoming neutral hosting hubs for global AI demand.

Iceland and Norway, two countries that saw data-center investment double between 2023 and 2024 on the expectation that AI compute would flow freely through cold-climate colocation sites, are now reassessing their positioning. Their marketing pitch — cheap energy, neutral geography — competes less effectively against a regulatory environment that treats any facility housing American GPUs as potentially sanctioned territory.

The Geopolitical Calculus

The US-China semiconductor arrangement has always been fragile. Trump’s recent comments defending Huang alongside the announcement of stricter controls suggest the administration is trying to do two contradictory things simultaneously: protect its political standing domestically by appearing tough on China, and avoid derailing trade talks in Beijing.

Behind the scenes, Treasury Secretary Scott Bessent and Commerce Secretary Gina Raimondo have been negotiating the scope of the rule with Chinese counterparts through backchannels in Geneva. According to people familiar with the discussions, China sought carve-outs for consumer-facing applications — entertainment, e-commerce, translation tools — arguing that restricting those uses violated the spirit of the October 2024 bilateral understanding. The final rule grants those carve-outs narrowly, limiting them to inference-only workloads below a defined FLOPS threshold. Training and fine-tuning remain categorically restricted.

That balancing act is already fraying. The 90-day grace period for existing data centers signals flexibility, but the geographic targeting of five ASEAN nations and Brazil signals escalation. If Washington adds more countries to the pre-approval list — as the Commerce Secretary’s authority allows — the net tightens further around any facility housing American supercomputing gear near Chinese supply lines.

Expect China to respond. The Ministry of Foreign Affairs has already called similar measures “unilateral bullying” in past iterations. Retaliatory controls on rare-earth exports or targeted sanctions against American semiconductor distributors remain possibilities. Chen Jing, a researcher at Tsinghua University’sCenter for International Security Studies, noted in a briefing to PLA General Staff Department officials last week that the rule represented a shift from targeting hardware to governing information flows — a precedent she argued could justify future Chinese restrictions on US companies operating within China’s digital ecosystem.

What This Means Going Forward

This rule change is not just another export-control adjustment. It is the first comprehensive attempt to govern computing power itself, not just the physical chips that provide it. In effect, the United States has decided that wherever an American GPU exists, it remains under American jurisdiction — even when the machine sits in a warehouse 8,000 miles from Silicon Valley.

That principle will shape the next decade of AI competition. China will adapt, but adaptation takes years. NVIDIA will lose a profitable channel, but one it should have exited sooner. And the rest of the world will learn that neutrality in AI infrastructure is no longer an option — you either host compliant American compute or you build your own.