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. The AI industry talks about them more ambivalently than anything else, because they put a real constraint on a real lever, and the industry would rather have a real constraint on nothing at all.

What follows is what the controls do, why they count as the most effective current brake on AI capability development, and where they fall short.

The Chips and Tools on the List

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

Compute is the primary input to frontier AI training. Training GPT-4 took somewhere around 25,000 A100 GPUs running for several months, and training frontier models at 2026 scale takes considerably more. Without access to chips in those quantities, you cannot train frontier models at all.

The point is to create a persistent compute gap between US-based labs and potential rivals, China above all, that can't be closed in the near term by buying chips on the open market or by building fabs at home.

Why a Compute Lever Beats the Alternatives

Most AI governance proposals rely on self-reporting, voluntary compliance, or enforcement against distributed actors. Transparency requirements, safety evaluations, registration schemes, liability frameworks: all of them can be evaded, delayed, or worked around, and all of them need international coordination to bite at the frontier.

Compute controls work because the advanced chip supply chain is concentrated, geographically and commercially, 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 makers sit in the Netherlands and the US. Controlling exports from these choke points needs no international coordination, because the choke points already sit in the US and among US allies.

Yann LeCun and other critics of AI risk have argued this is excessive. They are wrong that the controls are unnecessary, since capability scaling clearly matters for safety, but they are right that the controls are powerful. That is exactly why the controls exist.

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 are closing it faster than optimists hoped. Huawei's Ascend 910C, announced in late 2024, benchmarks closer to H100 performance than its predecessors did, and the Chinese government has poured money into domestic semiconductor development. The gap today is probably larger than the gap will be in three years.

So the controls create a meaningful pause rather than a permanent stop. They buy time, and how much depends on how fast domestic alternatives mature and how tightly enforcement holds against the evasion routes.

The Compute Overhang

Here a separate but related idea comes in: the compute overhang. There are algorithmic improvements and architectural innovations that could sharply raise AI capability without any more raw compute, sitting in potential, waiting to be discovered or deployed. Once they are, a given amount of compute becomes far more powerful.

That matters for policy because export controls target compute, not algorithms. A country or organization with limited compute but a major algorithmic breakthrough could close a large gap quickly. AI's history is full of examples: attention mechanisms, transformer architectures, mixture-of-experts, chain-of-thought prompting. Each sharply raised what the same hardware could do.

The compute overhang argument means the margin of safety from export controls is thinner than the raw compute gap suggests. Capability comes from chips multiplied by algorithms, and controls on chips leave algorithms untouched.

Where the Controls Leave Us

The current approach, export controls as the primary lever supplemented by voluntary safety commitments from labs, beats doing nothing but won't hold up as a long-term strategy. Export controls create a gap; they do not solve alignment. They buy time for alignment research and governance to catch up, and they are no substitute for that catch-up actually happening.

The US government has found the right lever, compute, and is pulling it to real effect. The AI safety community has worked out what to do with the time that lever buys: alignment research, governance frameworks, international coordination. The second effort has moved nowhere near as fast as the first one's importance demands.

Chips without alignment are dangerous, and controls that buy time nobody uses well are an opportunity rather than a solution. So far that opportunity hasn't been met with anything like the urgency the situation calls for. The competitive dynamics that drive capability investment don't slow down just because one country's export policy opens a temporary compute gap.