AI Chip Export Controls: Technology as Geopolitics

In October 2022, the United States did something unusual. It deliberately hobbled its own companies' most lucrative export market. The Bureau of Industry and Security issued sweeping export controls blocking China's access to advanced AI chips, chip manufacturing equipment, and even the expertise of American citizens working in Chinese semiconductor facilities.

The controls have been tightened three times since — in October 2023, late 2024, and early 2025. Each round closed loopholes from the previous one. Each round generated new loopholes to close.

This is not just trade policy. It is an attempt to control the global distribution of AI capability through the hardware layer. And it is reshaping who gets to build AI, where, and for what purpose.

What's Actually Restricted

The controls target three layers of the semiconductor stack.

Advanced chips. Nvidia's A100 and H100 GPUs — the workhorses of AI training — cannot be sold to Chinese customers. The restrictions use specific performance thresholds: chips exceeding certain compute densities and interconnect bandwidths are blocked. The October 2023 update lowered these thresholds after it became clear that chips just below the original cutoffs were still highly capable for AI workloads (Bureau of Industry and Security, 2023).

Manufacturing equipment. ASML's extreme ultraviolet (EUV) lithography machines — the $150-million-per-unit tools required to fabricate chips below 7 nanometers — have been restricted since 2019 under US pressure on the Dutch government. The 2023 rules extended restrictions to ASML's older deep ultraviolet (DUV) systems, which China had been using to push the boundaries of less advanced manufacturing processes.

Design tools and expertise. Electronic design automation (EDA) software from companies like Cadence, Synopsys, and Mentor Graphics is restricted. So are US persons — American citizens and permanent residents — who were previously working in Chinese chip fabs. The personnel restrictions forced hundreds of American engineers to choose between their jobs and their citizenship.

The cumulative effect: China cannot buy the best AI chips, cannot buy the machines to make them, and cannot easily hire the people who know how to build the machines.

The Strategic Logic

Jake Sullivan, Biden's National Security Advisor, articulated the framework in a September 2022 speech at the Special Competitive Studies Project. The US would maintain a "small yard, high fence" — narrowly targeted restrictions on technologies with clear national security implications, but enforced rigorously (Sullivan, 2022).

The rationale rests on the dual-use nature of AI. The same chips that train language models also train surveillance systems, autonomous weapons, and signals intelligence platforms. Advanced AI capabilities in Chinese military applications, the argument goes, pose direct national security risks. Maintaining a generational lead in AI hardware — not just an incremental one — became explicit policy.

This was a significant departure. Previous US technology policy generally aimed to stay ahead through innovation, not by actively degrading a competitor's capabilities. As Chris Miller argues in Chip War, the export controls represent a shift from running faster to kneecapping the other runner (Miller, 2022).

Nvidia's Regulatory Arbitrage

Nvidia found itself in an awkward position. China represented roughly 20-25% of its data center revenue. The export controls didn't just affect future sales — they disrupted existing customer relationships.

Nvidia's response was swift and creative. Within months, the company designed China-specific chips that technically complied with the export thresholds. The A800 replaced the A100. The H800 replaced the H100. These variants reduced interconnect bandwidth to fall below regulatory limits while retaining most of the raw compute performance useful for AI training.

This was regulatory arbitrage in silicon. The chips were meaningfully less capable for certain large-scale training configurations — specifically, the multi-chip communication speeds critical for training the very largest models. But for most AI workloads, they were nearly as good as the unrestricted versions.

The October 2023 rules closed this loophole by lowering the performance thresholds and introducing a broader "performance density" metric that caught the workaround chips. Nvidia reportedly lost $5 billion in projected China sales in the year following the tightened restrictions (Reuters, 2024). The company lobbied aggressively against the controls, arguing they simply pushed Chinese customers toward domestic alternatives without meaningfully slowing China's AI progress.

They were not entirely wrong.

China's Response: Forced Innovation

Export controls created an explicit mandate for Chinese semiconductor self-sufficiency. Beijing had been pursuing this goal since the "Made in China 2025" initiative, but the controls transformed a long-term aspiration into an urgent industrial priority.

Huawei's Ascend chips. Huawei, already under severe US sanctions since 2019, developed its Ascend 910B AI accelerator — a chip that, while trailing Nvidia's latest by a generation, proved capable of training large language models. The Ascend ecosystem remains immature compared to Nvidia's CUDA, but it is functional and improving.

SMIC's manufacturing advances. Semiconductor Manufacturing International Corporation, China's leading chipmaker, produced 7-nanometer chips for Huawei's Mate 60 Pro smartphone in 2023 — a feat that surprised Western analysts who assumed SMIC couldn't reach that node without EUV lithography. They did it using older DUV machines with multi-patterning techniques: slower, more expensive, lower yield, but it worked.

DeepSeek's efficiency breakthroughs. Perhaps the most striking consequence has been in software. DeepSeek, a Chinese AI lab, achieved performance competitive with frontier Western models using significantly less compute. Their R1 model, released in January 2025, demonstrated reasoning capabilities rivaling much larger models. Chip constraints didn't just force hardware innovation — they forced algorithmic efficiency. When you can't brute-force with more compute, you get clever with less.

The pattern echoes technology sanctions throughout history. Restricted access to IBM mainframes in the Soviet era didn't prevent the USSR from computing; it produced a different computing tradition. Constraints reshape development paths more often than they halt them. And the results are not always inferior — DeepSeek's efficiency work may benefit the entire field, just as Japan's fuel-efficiency innovations, born from oil scarcity, eventually reshaped the global auto industry.

The Allied Dimension

Export controls are only as effective as their geographic coverage. Advanced chips designed in the US are manufactured in Taiwan and the Netherlands. Manufacturing equipment comes from the Netherlands and Japan. Restricting US exports alone would leave enormous gaps.

The Biden administration pressured the Netherlands and Japan to impose parallel restrictions. In January 2023, the three countries reached an agreement — the details of which were never fully published — to coordinate controls on semiconductor manufacturing equipment. The Netherlands restricted ASML's exports of both EUV and advanced DUV systems. Japan restricted exports from Tokyo Electron and other equipment makers.

This extraterritorial dimension is significant. The US effectively leveraged its position in the semiconductor supply chain — where American technology touches nearly every link — to extend controls beyond its own jurisdiction. Companies that use US-origin technology, software, or equipment in their products can be subject to US export rules regardless of where they are headquartered.

The approach has generated friction. ASML's CEO has publicly questioned whether the restrictions serve long-term European interests. Japanese companies have expressed concern about losing market share to Chinese equipment makers that will eventually emerge. Allied compliance is real but not enthusiastic.

Unintended Consequences

Export controls are a blunt instrument applied to a complex system. The second-order effects are substantial.

Cloud access as a workaround. Chinese companies can still access advanced chips indirectly through cloud computing services. US cloud providers initially continued serving Chinese AI customers through their Asian data centers. The 2024 rule updates attempted to close this by requiring "know your customer" provisions for cloud providers, but enforcement remains difficult. How do you verify what workloads a cloud customer is running?

Stockpiling. In the months before each new round of restrictions took effect, Chinese companies purchased enormous quantities of chips — billions of dollars worth. Nvidia shipped record volumes to China in the gap between announcement and enforcement. These stockpiles provide a buffer, though one that depletes over time as chips age and workloads grow.

The smuggling economy. A grey market for restricted chips has emerged, flowing through intermediary countries. Chips nominally destined for Southeast Asian buyers end up in Chinese data centers. The Commerce Department has pursued enforcement actions, but supply chains are difficult to police at the chip level.

Accelerated decoupling. Chinese technology companies are actively reducing dependence on all American technology, not just restricted chips. The controls accelerated a broader bifurcation of the global technology ecosystem — exactly the opposite of what decades of engagement policy aimed to prevent.

The Redistribution Lens

Export controls are, fundamentally, a redistribution mechanism. They redistribute AI capability — directing it toward some nations and away from others.

Who benefits? US technology companies maintain their premium pricing and technology leadership, at least for now. The US national security establishment gains a lever over Chinese AI development. Allied nations with access to frontier chips — Europe, Japan, South Korea, Taiwan — maintain their position in the AI hierarchy.

Who pays? Chinese companies and researchers face higher costs and capability gaps. Nvidia and other US chipmakers lose significant revenue. But the most underexamined cost falls on developing nations caught in the crossfire. Countries that are neither US allies nor Chinese partners face a two-tier world: those with access to frontier AI hardware and those without. The controls don't just restrict China — they create a framework where AI capability becomes a function of geopolitical alignment.

This is the redistribution question at its sharpest. AI chips are not consumer goods. They are infrastructure — the substrate on which AI capabilities are built. When infrastructure access is determined by geopolitical alliance rather than need or capacity, the resulting distribution mirrors power, not opportunity.

The Deeper Question

Should AI infrastructure be treated as a strategic resource subject to geopolitical control? The precedents are instructive but imperfect.

Oil embargoes in the 1970s demonstrated that resource controls create short-term leverage but accelerate substitution and enmity. Nuclear technology controls through the Non-Proliferation Treaty have been partially successful but created their own inequities — a world of nuclear haves and have-nots that the have-nots have never accepted as legitimate. Internet governance debates in the 2000s showed that controlling information infrastructure generates both security benefits and censorship concerns.

AI chips sit uncomfortably across all these analogies. They are physical goods like oil, dual-use technology like nuclear materials, and enabling infrastructure like the internet. No single governance framework fits.

What is clear is that the current approach — unilateral controls extended through alliance pressure — is unstable. It generates workarounds, accelerates the very self-sufficiency it aims to prevent, and creates a global AI landscape defined by where you sit on a geopolitical map rather than what you need to build.

The alternative is not naivety about military applications. It is recognizing that treating AI hardware purely as a weapon of strategic competition ensures it will never function as shared infrastructure. And shared infrastructure is what a technology this consequential probably needs to be.

It is worth acknowledging that policymakers face a genuinely difficult problem. Dual-use technology creates real security dilemmas, and the people designing these controls are navigating trade-offs with no clean answers. The Nuclear Non-Proliferation Treaty, for all its inequities, did help prevent nuclear war. The question is whether AI chip controls can achieve their security objectives without foreclosing the possibility of broadly shared compute infrastructure in the future — and whether multilateral frameworks, however difficult to negotiate, might serve both goals better than unilateral action extended through alliance pressure.

Related

• Hardware Bottleneck
• Compute Is the New Oil
• Who Controls Supply Chain
• Frontier Lab Oligopoly
• China's AI Strategy
• Sovereign AI Movement
• DeepSeek Price of Open AI
• Magnificent Seven

Sources

• Bureau of Industry and Security. "Export Controls on Advanced Computing and Semiconductor Manufacturing Items to China." Federal Register, October 2023 (updated). Federal Register
• Miller, Chris. Chip War: The Fight for the World's Most Critical Technology. Scribner, 2022. Publisher
• Sullivan, Jake. "Remarks on the Biden-Harris Administration's National Security Strategy." Special Competitive Studies Project, September 2022. White House
• Reuters. "Nvidia's China Sales Hit by Tightened US Export Controls." Reuters, 2024. Reuters
• Allen, Gregory C. "Choking Off China's Access to the Future of AI." Center for Strategic and International Studies, October 2022. CSIS