Export controls — Gregory Allen on the new architecture

Gregory C. Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies in Washington and previously chief of strategy and communications at the Defense Department's Joint Artificial Intelligence Center, has produced the most rigorous open-source analysis of the post-2022 US semiconductor export-control regime. His CSIS reports from 2022-2024 are the standard reference.

Allen's core analytical move is to read the October 2022 controls — and their October 2023 and December 2024 extensions — as a categorical shift from the previous US export-control posture. Pre-2022 controls were primarily product-based: specific items on specific lists subject to specific licensing rules. The 2022 controls layered onto this a capability-based architecture: restrictions on items designed to limit the entire class of advanced-AI-training compute available to specific entities or to entire geographic regions.

The three operational pillars Allen identifies. First, compute caps on advanced AI accelerators — initially defined by a combination of interconnect bandwidth and floating-point performance thresholds, refined in October 2023 to close loopholes that emerged. Second, foundry equipment controls — ASML EUV equipment is now subject to Dutch and Japanese controls coordinated with the US, with sub-EUV deep-ultraviolet equipment added in 2023-2024. Third, the long-arm 'foreign direct product rule' — items made anywhere in the world using US technology, software, or design tools are subject to US controls when destined for restricted end uses or end users.

Allen's analytical concern, expressed across his 2024 work, is about implementation. Compute-cap controls are technically complex and have been gamed by chip vendors (notably Nvidia with the A800 and H800 China-specific variants prior to the October 2023 tightening). Foreign direct product rule enforcement depends on the willingness of allied governments to support the US extraterritorial reach. The Entity List covers 187 advanced-semiconductor-related entries by end-2024 but Chinese semiconductor firms have demonstrated substantial capacity to restructure around individual designations.

His longer-run argument is that the controls will substantially slow Chinese leading-edge progress but will not stop it, and that the policy success criterion should be measured in years of capability lag rather than in absolute denial. He has been publicly critical of policymakers who frame the controls as denial rather than as time-buying.

­Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies, previously the founding director of the Department of Defense Joint Artificial Intelligence Center's Strategy and Policy office, has produced the most detailed open-source readings of the October 2022 export-control package and its 2023 and 2024 amendments. His running analysis at CSIS, combined with policy-memo work for the Center for a New American Security and other think-tank publications, has become the standard non-classified reference for what the new architecture covers, how it operates in practice, and what its enforcement limitations are.

The October 2022 package introduced two major innovations beyond the existing Commerce Department Entity List framework. First, it imposed performance-based controls — restrictions keyed to specific computational capability thresholds (initially the 4800 total processing performance and 600 GB/s interconnect bandwidth lines), with later amendments introducing performance density measurements as additional criteria — that target advanced AI accelerator chips regardless of which manufacturer produces them. Second, it imposed a foreign direct product rule on advanced semiconductors and on the equipment to make them, extending US export-control jurisdiction to any chip or equipment whose production involves US-origin technology or US-person expertise even if the final manufacturing occurs outside the United States.

The 2023 amendments tightened the original thresholds (in response to NVIDIA's release of the A800 and H800 cut-down variants that fell below the original ceilings) and added additional jurisdictions and entities to the controlled-geography list. The 2024 amendments further refined the performance metrics, addressed advanced-packaging and high-bandwidth-memory considerations, and adjusted the licensing presumptions for specific use cases. The cumulative architecture is the most extensive technology-export-control regime the United States has imposed since the COCOM-era controls on Soviet-bloc technology flows.

The enforcement architecture rests on the Commerce Department's Bureau of Industry and Security and the Department of Justice's national-security division. BIS has been resourced more aggressively since 2022, with the Disruptive Technology Strike Force established in February 2023 as a joint Commerce-Justice initiative for cross-jurisdictional enforcement actions. Allen's tracking of enforcement actions through 2024 and 2025 documents a series of indictments and civil-penalty actions covering transshipment networks (particularly through the UAE, Hong Kong, and Singapore), licensing-violation cases against US-domiciled corporate compliance failures, and a smaller number of cases targeting foreign nationals operating as procurement agents for sanctioned end-users.

The third-country dimension is the part of the architecture with the most uneven coverage. The Netherlands and Japan, which together hold the upstream-equipment chokepoints, have aligned their export-control regimes with the US framework through 2023 and 2024 amendments to their respective national lists. South Korea has aligned more selectively, with exemptions for specific Korean-domiciled fabs operating in China (Samsung and SK Hynix's mainland operations). The EU's broader export-control coordination, operationalised through the Dual-Use Regulation and a series of additional implementing measures, has progressed unevenly across member states. India, Indonesia, Vietnam, and the major Gulf states remain outside the aligned-control framework, which Allen's analysis flags as the largest open question for the architecture's medium-run effectiveness.

The strategic-doctrinal frame Allen has articulated is that the export-control regime is best understood as a moving target rather than a static restriction. The performance thresholds will need to tighten as the controlled technology advances; the third-country alignment will need to widen as the substitution networks adapt; the enforcement architecture will need to scale as the volume of transactions subject to scrutiny grows. The architecture introduced in 2022 has produced measurable effects on the Chinese semiconductor industry's leading-edge progression; whether those effects are sustained depends on the political-economic durability of the alignment with the supplier and intermediate countries, and on the ability of the enforcement bureaucracies to scale with the policy ambition. Both are open questions and both will be decided over the late-2020s window during which the current generation of the architecture is operating.

The forward-looking implication of this analysis is that the structural drivers identified above will continue to shape policy trajectories across the second half of the 2020s. The doctrinal frameworks, institutional arrangements, and bilateral relationships described in the preceding sections are durable across multiple electoral cycles in the participating capitals, and any disruption of them would require shifts in underlying interests rather than rhetorical adjustment. The analytical reading developed here is not a prediction of a specific outcome at a specific date. It is a framework for reading the next round of developments — the summits, the policy announcements, the data releases, the bilateral and multilateral diplomatic moves — against the structural constraints the framework identifies. Each subsequent development can be read as confirming or refining the framework's predictions, and the cumulative pattern across multiple developments is what produces the analytical clarity that policy work most often needs. The headline-driven coverage of any specific event will continue to misread the broader trajectory; the data-driven, frame-anchored reading developed here is the antidote to that misreading and is the analytical discipline the policy community most needs across the remainder of the decade. The arithmetic of the underlying interests does not change quickly. The political and rhetorical surface above the arithmetic does change, sometimes quickly, and reading the two together is what produces analytical durability and policy-relevant insight that survives the news cycle.

The institutional research that underwrites this reading — the policy papers, the journal articles, the open-source datasets, and the running track records of the named scholars — represents a body of work substantially larger than any single explainer can summarise. Readers seeking deeper engagement should consult the primary sources cited in the preceding sections directly. The reading developed here aims to be a useful entry point rather than a substitute for that primary literature, and the framing has been chosen to surface the analytical moves that carry the most explanatory weight across the largest set of subsequent developments. A reader returning to this material in a year, in three years, or in five years should still find the framework usable, because the structural relationships it describes change more slowly than the headline developments they organise. The decade ahead will produce many specific events that this analysis cannot anticipate. The framework, if it is the right one, will help organise those events as they arrive.