Did the Export Controls Actually Work?

In October 2022, an unnamed senior US official made a remark to a small group of reporters that has become one of the more revealing statements of the entire US-China technology competition. The official was describing the philosophy behind the new semiconductor export controls, and the phrase was this: “We want to as long as possible maintain as large a lead as possible.”

That is not a strategic objective. It is an aspiration. The difference matters enormously when you try to evaluate whether the policy has worked.

Strategic objectives, properly stated, are measurable and time-bound. “Prevent China from training a model competitive with GPT-4 before 2026” is a strategic objective. “Ensure that China cannot deploy AI systems for autonomous drone targeting before 2028” is a strategic objective. “Maintain as large a lead as possible for as long as possible” is a preference, an orientation, a direction of travel — not a goal that can be evaluated against outcomes.

This ambiguity is not accidental. Vague objectives are politically useful because they are unfalsifiable. You cannot demonstrate that the export controls failed to “maintain as large a lead as possible” because no one specified what lead was required, over what time period, measured by what standard. The policy’s supporters can always claim it is working because China has not yet matched US capabilities on whatever capability they choose to invoke. The policy’s critics can always claim it is failing because China continues to develop AI capabilities that, in a world without the restrictions, it would have developed more quickly.

The honest analytical task is to evaluate the controls against the most specific plausible reading of their objectives, then assess what the actual outcomes have been.

The most specific stated objective was preventing China from training models that could rival American frontier capabilities in the near term. By this standard, the export controls have had a measurable effect. The compute available to Chinese AI labs in early 2027 is significantly lower than it would have been without the restrictions. Nvidia estimates — based on the revenue foregone from Chinese customers — that China’s AI data center compute has grown at perhaps 40 percent of the rate it would have without export controls, accounting for both direct chip restrictions and the secondary effects on ecosystem development.

The gap this has created is real but not as large as some proponents hoped. DeepSeek’s architectural innovations, which have been widely documented and replicated by international researchers, demonstrate that Chinese AI teams have partially compensated for compute constraints with algorithmic efficiency. A model trained with half the compute can achieve most of the performance of a model trained with full compute, if the training is done with sufficient architectural creativity. This is not a permanent offset — there are limits to how much engineering ingenuity can substitute for raw compute — but it has narrowed the effective gap relative to the hardware gap.

The second stated objective — protecting American national security by preventing Chinese AI capabilities in surveillance, autonomous weapons, and intelligence analysis — is harder to evaluate because the relevant capabilities are not publicly benchmarked. What is observable is that China’s domestic AI surveillance infrastructure, which predates the export controls and was already highly capable, has continued to develop. The systems deployed in Xinjiang, which attracted international human rights attention beginning in 2018, did not depend on the frontier AI capabilities that the export controls targeted. They depended on computer vision, facial recognition, and pattern matching capabilities that were already commoditized by 2022.

The third objective — which was stated less explicitly in public documents but is visible in the strategic logic of the policy — was signaling. The export controls were, among other things, a demonstration that the US was willing to use its control over the technology stack as a strategic weapon, accept economic costs for strategic ends, and enforce that use against the resistance of affected American companies. As a signaling exercise, the controls were unambiguously successful. China, the Netherlands, South Korea, Japan, and Taiwan all received a clear signal about American willingness to act. So did every other country trying to assess American seriousness about its stated technology security objectives.

The cost of that signal was the revenue foregone by Nvidia, AMD, Applied Materials, and KLA — companies whose Chinese sales were constrained by the restrictions. Nvidia’s CEO Jensen Huang has been publicly and consistently critical of the export controls, arguing that they cost American companies revenue that is now being captured by Chinese alternatives, while the Chinese alternatives develop capability that would have been imported instead. This argument has some validity: Huawei’s AI chip revenue is partly a substitution for Nvidia revenue. Whether the substitution is complete — whether every yuan Huawei earns from Ascend sales is a yuan that Nvidia would have earned from H100 sales — is disputed.

The deeper economic argument is about long-run market structure. Semiconductor markets are characterized by enormous fixed costs and network effects that reward scale. If Chinese chip companies build domestic market share under the protection of forced import substitution, they will develop cost structures and customer relationships that persist after the geopolitical context changes. The export controls may be creating a more competitive Chinese semiconductor sector than would exist in their absence — not despite the restrictions, but because of them. The historical analogy is the Japanese semiconductor industry, which built its global competitiveness in the 1970s and 1980s partly behind domestic procurement preferences that American companies found deeply frustrating.

The question of unintended consequences is where the analysis becomes genuinely uncertain. The export controls were designed to constrain a specific set of Chinese capabilities while minimizing disruption to the global semiconductor supply chain and to US commercial interests. The designers understood that there would be costs, but believed the strategic benefits outweighed them.

What they may not have fully anticipated is the accelerant effect on Chinese industrial policy. China’s domestic semiconductor investment was already substantial before October 2022. After the export controls, it became a stated national priority with essentially unlimited capital support from the state. The companies that received that capital — SMIC, CXMT, Yangtze Memory Technologies, and dozens of smaller firms in the supporting supply chain — have developed capabilities they would have developed more slowly in a market environment. The forced march has costs (inefficiency, misallocation, poor yield on rushed programs), but it has also produced results that are real.

If the export controls produce, over a ten-year period, a China that has a complete domestic semiconductor ecosystem at 80 percent of frontier capability, independent of American supply chains, they will have failed on their own terms while arguably also demonstrating that the policy was correctly targeted. A China that can operate its critical infrastructure on domestically produced chips is a China that has removed a vulnerability that American strategic planners had assumed would persist.

The verdict, one year after the most significant tightening, is that the controls are achieving a version of their objectives — slowing Chinese AI development — while also accelerating a Chinese industrial response that may ultimately produce a more resilient adversary. Whether that tradeoff is worth it depends on how you weight short-term capability gaps against long-term strategic independence.

The “maintain as large a lead as possible for as long as possible” framing is inadequate because it doesn’t engage with this tradeoff. The honest strategic question is: what do we want China’s AI capability to look like in 2035, and does this policy make that more or less likely? That question has not been publicly answered, which suggests either that the answer is classified or that the policy was constructed without a clear terminal objective.

Both possibilities are uncomfortable. For different reasons.