Hardware Restrictions and AI Competition: The Effect of NVIDIA GPU Export Restrictions on AI Benchmark Performance

• In Event: Technology Diffusion, Industry, and Global Competition

Abstract

In October 2022 the United States restricted advanced computing GPU exports to China and in October 2023 tightened the rule and expanded coverage, raising the question of whether these controls slowed observed AI model performance development. This paper contributes by building a developer-level monthly panel from the Hugging Face Open LLM Leaderboard that other researchers can reuse for causal analysis of policy shocks in AI development and extend as restriction rounds and benchmark data continue to evolve.This paper estimates the effect of the two export control waves with diagnostic event studies, staggered difference in differences, and mechanism-specific extensions. Across the main estimators, the paper does not find robust detectable evidence of a net negative effect on Chinese benchmark performance, while the mechanism evidence is more consistent with adaptation than with collapse. That adaptation was concentrated rather than uniform: organizational scale separated developers who absorbed the restriction from those who did not, substitute GPUs offset part of the first-wave pressure during the window they remained available before the second wave eliminated that channel, and China-affiliated research output rose after the first wave in a response that did not extend to later-treated cohorts. The stockpiling hypothesis finds no support in the available data. The net null is therefore better read as the product of direct pressure and offsetting adaptation than as evidence that the controls imposed no constraint.The results suggest that hardware restrictions may have imposed real pressure without producing a clear decline in benchmark scores of state-of-the-art models, so policy evaluation must distinguish between the direct constraint and the adaptation channels that offset it. By placing observable AI benchmark outcomes inside a causal inference design focused on export controls, the paper bridges the international trade and export-control literature with the AI benchmarking literature, connecting the two through direct measurement of how policy shocks propagate to development outcomes. The staggered and dosage-aware difference-in-differences designs, together with the bivariate decomposition that separates ban effects from substitute chip mitigation, provide an empirical framework for evaluating technology restrictions whose impact is absorbed through multiple adaptation channels rather than appearing as a single aggregate decline.

Author

• XINGYUAN ZHAO, University of Texas, Dallas
  Presenting Author