Published July 26, 2026
| Version v1
Publication
Open
Compute, Policy, and Frontier AI: An Empirical Analysis of Export Control Regimes
Description
Advanced-computing export controls are the largest instrument actually deployed in AI policy, and among the least
evaluated. We ask whether they measurably moved where frontier training happens — by developer nationality, by
compute scale, and by reported hardware. We assemble 650 catalogued AI models (353 with a training-compute
estimate) and 404 AI compute clusters (385 with a capacity estimate) from the public Epoch AI records, attach the
five Bureau of Industry and Security tranches from October 2022 to January 2026 at their Federal Register effective
dates, and estimate difference-in-differences around each tranche with randomisation-based inference.
Abstract
Abstract
Our central result is a null that is robust in a specific and informative way. Across five tranches and two outcomes,
no estimate survives scrutiny: two of ten reach nominal p < 0.05 under organisation-permutation, but the strongest
(December 2024, −1.13 dex of training compute) has a placebo-in-time p = 0.06, the largest Benjamini–Hochberg q
value across the grid is 0.19, and the October 2022 estimate changes sign with the estimation window (−0.11 at ±1
year, +0.86 at ±2 years, +1.28 at ±3 years). Cluster-level capacity DiD is null throughout (p ≥ 0.28). We conclude that
the public record cannot support a causal claim that these controls raised or lowered the training compute of
Chinese-developed models.
What did change, sharply and on schedule, is which chips get named. The share of China-based models disclosing an
export-control-specific NVIDIA SKU (A800/H800/H20) went from 0% before October 2022, to 9% between the first
and second tranches, to 54% afterwards; disclosed Chinese domestic accelerators scaled from 512 chips (PanGu-Σ,
March 2023) to 8,192 (Pangu Ultra, April 2025). Meanwhile the geography of disclosed capacity diverged
dramatically: the US share of disclosed H100-equivalents rose from 53% to 81% while China's fell from 28% to 6%,
and the US/China ratio moved from 1.1× in 2021 to 27× in 2025.
We also report three ways this data will mislead an unwary analyst, each of which we hit: Epoch publishes every
Chinese cluster anonymised with chip type withheld in both eras (so a naive disclosure DiD manufactures a large
spurious effect), geocodes all Chinese systems to one national coordinate (so location maps imply false precision),
and its catalogue composition shifts over time (US models with compute estimates fall from 26 in 2022H1 to 6 in
2025H2 as China's rise). Hardware disclosure is collapsing everywhere — China 100%→10% and the US 79%→18%
between 2021H1 and 2025H2 — which is itself the binding constraint on future evaluation of this policy.