Why the United States banned selling chips to China

Today we talk about the political decision with the greatest sector-wide impact of the last five years, and the one on which most Spanish-speaking readers haven't read a single piece with context. It's the October 7, 2022 rule from the Bureau of Industry and Security of the U.S. Department of Commerce. A hundred and thirty-nine pages of technical jargon. So boring on the surface that it went almost unnoticed in the general press. Its effect: to stop dead the export to China of the chips that train frontier models. Why does it matter to the European reader? Because that decision is reshaping the planet's technological geopolitics, and because the side effects may have produced the opposite of what was intended. My take is biased because I've followed the rule since the day after it was published. Form your own.

I've kept a notebook since October 2022 where I jot down the amendments to that rule. There are eight, as of this article's cutoff date. Each amendment closes a hole that had opened. Every time a hole closes, suppliers and buyers look for another. The dynamic is that of a dike with so many leaks that patching is the main activity. Let's take it slowly.

What exactly the October 7, 2022 rule bans

The original rule was published in the U.S. Federal Register on October 7, 2022 under regulation code 87 FR 62186. Its technical core is complicated, but the central idea is graspable. The United States bans the export, re-export or transfer to China —and to Macao— of chips that exceed certain combined thresholds of compute capacity and interconnect bandwidth. The thresholds are deliberately designed to capture the Nvidia H100 and A100 GPUs used in training frontier AI models.

The rule isn't limited to chips. It also bans the export of semiconductor manufacturing equipment capable of producing those chips in China, the design software tools (EDA) needed to create them, and services rendered by U.S. citizens that help develop them. It's an extensive rule meant to close every reasonable avenue of access to the critical hardware.

The immediate question is why. The official answer, in the document's own preamble, is that China was using these chips to modernize its military and surveillance capacity, and that this modernization posed a risk to U.S. national security. The real answer, read between the lines, is broader. The United States had concluded that the ability to train frontier AI models was a geostrategic advantage of the first order, and that holding that advantage required denying China access to the necessary hardware.

The Biden administration, which approved the rule, presented it not as general containment but as a surgical measure on military technology. That presentation was rhetorically convenient but technically misleading. The blocked chips are exactly the same ones used for civilian tasks —training ChatGPT, training medical models, simulating proteins, training translation models. The barrier between military and civilian use is, in this technology, nonexistent. The rule assumes as much; the public discourse hides it.

Nvidia's response and the cycle of cut-down variants

Nvidia, which lost roughly 25% of its market in a single week to the rule, responded with a classic sector strategy: produce cut-down variants legally compatible with the thresholds. In November 2022 it launched the H800 and the A800, versions of the H100 and A100 with lower interconnect capacity —but similar compute capacity— designed to fall just below the rule's thresholds. They were technically legal and were exported en masse for a year.

The Biden administration closed that hole on October 17, 2023 with an amendment that lowered the combined threshold, leaving the H800 and the A800 also banned for China. Nvidia responded again with a new, even more cut-down variant, the H20, which met the new requirements. That chain of adaptations —ban, cut-down variant, new ban, new variant— has repeated three times as of this article's cutoff.

The resulting pattern matters. The restrictions don't apply to a specific product; they apply to a technical capacity. When the manufacturer adjusts its product to dodge the restriction, the regulator adjusts the restriction to reach the product. It's a dynamic game where no equilibrium is stable. The consequence is that no Chinese customer, not even legally buying the cut-down variants in force at any given moment, can plan more than twelve months out.

The AI Diffusion Rule of January 2025

On January 13, 2025, a week before the change of administration, the outgoing Biden administration published a new rule called Framework for Artificial Intelligence Diffusion. The rule extended the logic of chip export control to a broader framework, dividing the world's countries into three tiers.

The first tier included the United States and its closest allies —the United Kingdom, Japan, the Netherlands, South Korea, Australia, Canada, Germany, France, Spain, among another eighteen countries. These countries could buy Nvidia chips without quantitative restrictions.

The second tier included most of the rest of the world, including most countries in Asia, Africa and Latin America. These countries could buy Nvidia chips but with an annual numerical cap and under a regime of revisable licenses. The exact size of the cap depended on the country and its political regime.

The third tier included countries under embargo: China, Russia, Iran, North Korea, Cuba, Venezuela, Belarus and Myanmar. These countries could not buy advanced Nvidia chips under any condition.

The rule was immediately criticized by industry —Nvidia issued an official statement calling it "unprecedented and misguided"—, by the foreign ministries of several second-tier countries —Singapore, Israel and others voiced displeasure—, and by international tech-policy organizations. The Trump administration, which took office on January 20, 2025, reviewed it and formally rescinded it in May 2025, replacing it with a more bilateral, less universalist scheme. The China-specific controls, however, have been kept and reinforced under the new administration.

Why what's done is done

The logic of the embargo, read honestly, rests on a clear strategic hypothesis. If the United States can deny China access to the hardware that trains frontier models for a long enough period —five to ten years—, China will fall one or two generations behind. That generational gap, multiplied by the geostrategic importance of AI, translates into years of military and economic advantage won without firing a shot.

The hypothesis has an empirical basis. During the Cold War, Western export controls on technology toward the Soviet Union played an important role in the resulting technological asymmetry. Current doctrine on AI and semiconductors applies the same playbook to a different competitor.

But the hypothesis also has limitations that U.S. agencies themselves recognize internally. The first is that the embargo assumes the technical frontier moves only linearly, with more hardware. If the frontier also moves with algorithmic efficiency —training better models with the same hardware—, the embargo can be evaded by innovation rather than smuggling. The second is that the embargo assumes China won't be able to produce equivalent hardware on its own. If China builds its own internal Nvidia —the equivalent would be Huawei Ascend, which already produces chips used for AI—, the embargo only delays access, it doesn't deny it permanently. The third is that the embargo creates enormous incentives for China to invest massively in alternatives, accelerating an independent ecosystem that without the embargo would have taken much longer to build.

These three limitations materialized in January 2025 with DeepSeek-R1.

The DeepSeek case and why it matters

DeepSeek-R1, the Chinese model that on January 20, 2025 triggered the largest market loss in Nvidia's history, is the operational proof of the embargo's three limitations.

DeepSeek trained R1 using Nvidia H800 chips —the cut-down variants, legal in China at the time of acquisition— combined, according to the company's own public statements, with algorithmic techniques particularly efficient in memory and compute use. The result was a reasoning model that rivaled OpenAI's o1, launched a few months earlier with far superior hardware.

This demonstration had two immediate consequences. The first was to reveal that algorithmic efficiency can compensate for hardware inferiority by an order of magnitude. The second was to show that the Chinese technical ecosystem had reached the capacity to compete, not just to copy, through its own R&D, funded in part by the quantitative fund High-Flyer of DeepSeek's own founder.

These two consequences, together, call into question the embargo's strategic hypothesis. If the frontier can move with efficiency and not only with hardware, and if China can already produce that efficiency without access to the most advanced hardware, the embargo delays but doesn't stop. And, meanwhile, it has produced an important side effect: it has forced the Chinese ecosystem to invest in efficiency, in homegrown hardware (Huawei Ascend, Cambricon, Biren) and in open-weight models published under permissive licenses. Each of those effects, individually moderate, adds up to a strategic alternative to the Western ecosystem that without the embargo would have taken longer to build.

What this means for Europe

Here comes the part worth naming without drama. Europe doesn't take part in this game, except as a spectator and as an enforcer of the restrictions decided in Washington. ASML, as we saw in 0073, has ceded commercial autonomy over a quarter of its market. The large European companies that would like to sell or buy Chinese technology operate under the double regime of U.S. rules (which apply extraterritorially when there's a U.S. component) and European rules (which add layers of their own). The operational consequence is that Europe adopts the United States' technological foreign policy without having designed it and without negotiating any quid pro quo in the process.

There's another indirect consequence. The embargo divides the world into technological blocs. China builds its ecosystem. The United States and its allies build theirs. Europe ends up inside the American bloc by default, without having actively decided to stay. For a region that officially defends multilateralism, strategic autonomy and commercial diplomacy, the situation contradicts its own rhetoric.

The political question

This is personal opinion, but I hold it on the record of the decisions themselves. The chip embargo on China is one of the most consequential and worst-explained geostrategic bets of the century. It's consequential because it's redesigning the global flows of technology, capital, talent and research. It's badly explained because the public conversation —even in the United States, much more so in Europe, infinitely less in Spanish-speaking countries— is reduced to headlines without analysis of the underlying hypothesis and its limitations.

What I would ask of the European public conversation is clarity on three questions. The first: does Europe have its own policy on the embargo, or does it adopt the American one by default? The second: has Europe calculated the industrial costs to itself of adopting it, including the cost to ASML, to European makers of semiconductor equipment and to the European AI labs that depend on the global ecosystem? The third: does Europe have a plan B if the embargo bet doesn't work and China manages to build a self-sufficient ecosystem?

Those three questions aren't being discussed seriously in Brussels or in any European capital. And the consequence will start to show in five to ten years, when we see whether the American bet worked and, depending on the result, whether Europe is on the right side of the wall or not.

The concrete figure that closes it: according to the report Restrictions on Exports of Advanced Semiconductors from the Center for a New American Security, published in November 2024, U.S. exports of AI chips to China fell roughly 60% between the fourth quarter of 2022 and the fourth quarter of 2024, while the number of frontier AI models published by Chinese institutions grew by around 130% in the same period. Those two figures together, belonging to the same period, are the numerical summary of the embargo's current state. What was banned has shrunk. What was being pursued has not.

Definitions

BIS (Bureau of Industry and Security): the office of the U.S. Department of Commerce responsible for export controls on strategic technologies. It publishes the specific rules in the Federal Register.

Federal Register: the official gazette of the U.S. federal government where rules and regulations are published. The BIS rules on semiconductors are published there.

Combined threshold: the technical criterion used by BIS rules to classify a chip as restricted. It combines compute capacity (TFLOPs) and interconnect bandwidth. It allows a product to be included or excluded without naming it explicitly.

Extraterritorial application: the application of a national rule to transactions outside its territory when they contain components or technology subject to the rule. The U.S. rule applies extraterritorially when a product contains American technology, even if the transaction is between non-American companies.

References

Bureau of Industry and Security, Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items (Federal Register, October 7, 2022; subsequent amendments 2023-2025). The original regulatory text and its updates.

Bureau of Industry and Security, Framework for Artificial Intelligence Diffusion (Federal Register, January 13, 2025; rescinded in May 2025). The Biden administration rule, briefly in force.

Center for a New American Security, Restrictions on Exports of Advanced Semiconductors (CNAS, November 2024). Empirical analysis of the effectiveness of the controls.

Center for Strategic and International Studies, Updated Analysis on US Chip Export Controls (CSIS, 2023-2025). Recurring analytical coverage from Washington.

Chris Miller, Chip War (Scribner, 2022). Historical framework for understanding the logic of semiconductor export controls since the Cold War.

Graham Allison, Destined for War: Can America and China Escape Thucydides's Trap? (Houghton Mifflin Harcourt, 2017). Analysis of the strategic dynamic between Washington and Beijing that frames the embargo decision.

Federal Register, public archive of U.S. Department of Commerce regulations on semiconductors and artificial intelligence (2018-2025). The primary source for the full chronology.

Going deeper

Dan Wang, Breakneck: China's Quest to Engineer the Future (W. W. Norton, 2025). A recent analysis from the Chinese side of the file.

Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Houghton Mifflin Harcourt, 2018). Background on the technological competition between the United States and China.

Henry Farrell & Abraham Newman, Underground Empire: How America Weaponized the World Economy (Henry Holt, 2023). Theoretical framework on how the United States uses global technological dependence as a geopolitical tool.

Paul Triolo, regular columns in DigiChina and at the Stanford Cyber Policy Center. Expert, level-headed coverage of the Chinese front of the file.

You may also like

• Why OpenAI wouldn't exist without a factory in Taiwan
• ASML, the Dutch company ChatGPT, Claude and the rest depend on
• DeepSeek and the end of the American monopoly
• Nvidia is worth more than Spain's economy

Elsewhere

• Lessons for Europe from China's quest for semiconductor self-reliance (Bruegel)
• Securing Semiconductor Supply Chains (Center for Security and Emerging Technology, Georgetown)