Chip Export Controls Are the Only Real Brake on the Industry

The Biden administration's export controls on advanced semiconductors, announced in October 2022 and significantly expanded in October 2023, are the most consequential AI policy action any government has taken. They're also the policy that the AI industry talks about most ambivalently, because they represent real constraints on a real lever, and the industry prefers real constraints on nothing.

Here's what the controls actually do, why they're considered the most effective current brake on AI capability development, and what they don't do.

What Gets Controlled

The controls restrict the export of advanced AI chips—primarily NVIDIA's H100, A100, and subsequent high-end GPUs—to China and a list of other countries of concern. They also restrict the export of chip-making equipment from companies like ASML, KLA, and Applied Materials that would allow those countries to manufacture equivalent chips domestically.

Compute is the primary input to frontier AI training. Training GPT-4 required somewhere around 25,000 A100 GPUs running for several months. Training frontier models at the 2026 scale requires significantly more. Without access to chips in those quantities, you cannot train frontier models. Full stop.

The controls are designed to create a persistent compute gap between US-based labs and potential rivals—specifically China—that can't be closed by buying chips on the international market or by domestic manufacturing in the near term.

Why Compute Controls Are Different From Other Governance

Most AI governance proposals—transparency requirements, safety evaluations, registration schemes, liability frameworks—rely on self-reporting, voluntary compliance, or enforcement against distributed actors. They can be evaded, delayed, or worked around. International coordination is required for them to bite at the frontier.

Compute controls are different because the advanced chip supply chain is geographically and commercially concentrated in ways that make enforcement practical. TSMC, the only company capable of manufacturing the most advanced chips, operates in Taiwan and sells to the world through a well-documented supply chain. NVIDIA designs in the US. The equipment manufacturers are in the Netherlands and the US. Controlling exports from these choke points is achievable without international coordination, because the choke points are in the US and among US allies.

Yann LeCun and other critics of AI risk have argued this is excessive. They're wrong that the controls are unnecessary (capability scaling clearly matters for safety), but not wrong that the controls are powerful. They are. That's the point.

The Workarounds and Their Limits

Chinese AI labs have responded in several ways. Huawei's Ascend chips are a domestic alternative, significantly less capable than NVIDIA's best but improving. Chinese companies have purchased chips before the controls and through intermediary countries. Some compute has moved to cloud services in jurisdictions not subject to the same restrictions.

None of these fully close the gap, but they're closing it faster than optimists hoped. Huawei's Ascend 910C, announced in late 2024, benchmarks closer to H100 performance than its predecessors did. The Chinese government has invested substantially in domestic semiconductor development. The gap that exists today is likely larger than the gap that will exist in three years.

This is the essential tension: the controls create a meaningful pause, not a permanent stop. They buy time. How much depends on how fast domestic alternatives develop and how tightly the enforcement can be maintained against evasion routes.

The Compute Overhang

A separate but related concept: compute overhang. The idea is that there are algorithmic improvements and architectural innovations that could dramatically improve AI capability without requiring more raw compute—they're "sitting in potential" waiting to be discovered or deployed. When they are, a given amount of compute becomes much more powerful.

This matters for policy because export controls target compute, not algorithms. A country or organization with limited compute but access to a major algorithmic breakthrough could close a substantial gap quickly. The history of AI is full of examples: attention mechanisms, transformer architectures, mixture-of-experts, chain-of-thought prompting—each dramatically improved what you could do with the same hardware.

The compute overhang argument suggests that the margin of safety from export controls is smaller than the raw compute gap implies. Capability isn't purely a function of chips. It's chips times algorithms. Controls on chips don't control algorithms.

What This Implies for Policy

The current approach—export controls as the primary policy lever, supplemented by voluntary safety commitments from labs—is better than nothing but insufficient as a long-term strategy. Export controls create a gap; they don't solve alignment. They buy time for alignment research and governance to catch up; they're not a substitute for that catch-up.

The honest assessment of where we are: the US government has identified the right lever (compute) and is pulling it, with real effect. The AI safety community has identified what needs to happen with the time that lever buys (alignment research, governance frameworks, international coordination). The pace of the second thing has not been commensurate with the importance of the first.

Chips without alignment is dangerous. Controls that buy time that isn't used well are not a solution. They're an opportunity. So far, the opportunity hasn't been matched with the urgency the situation implies. The competitive dynamics that drive capability investment don't slow down just because one country's export policy creates a temporary compute gap.