U.S. export controls on advanced chips and AI tools may be self-defeating: they can spur China’s push for technological self-reliance while raising legitimate WTO legal challenges, so policymakers should avoid overbroad security claims under GATT Article XXI(b).

Main Finding

Export controls have become a central strategic tool in the U.S.–China AI competition, but their long‑term effectiveness is doubtful. Rather than permanently denying access to capabilities, these controls tend to (1) incentivize supply‑chain diversification and circumvention in the short run and (2) accelerate domestic self‑reliance and independent R&D in the long run. Overbroad or arbitrary export restrictions also risk breaching WTO obligations; the authors argue for a constrained reading of GATT Article XXI(b) that excludes routine commercial or dual‑use AI models and semiconductors from security exceptions.

Key Points

• Recent U.S. actions (BIS rulemaking and proposals)
  • Major regulatory steps: Oct 2022 (initial controls on advanced computing & semiconductors), Oct 2023 (expanded scope and stricter licensing), Apr 2024 (clarifications to prevent circumvention), Dec 2024 (added equipment, software tools, entities), Jan 2025 (Framework for AI Diffusion — added controls on closed‑weight AI model parameters).
  • Legislative proposal: ENFORCE Act (2024) would broaden BIS authority to require licenses for exporting covered AI models to China.
  • Tactical focus: restrict upstream semiconductors, semiconductor manufacturing equipment, and increasingly downstream model weights/closed‑weight models.
• Chinese countermeasures
  • China imposed export restrictions on gallium, germanium, graphite, antimony and certain heavy rare earths (2023–2025 actions).
  • China filed a WTO complaint against the U.S. (WT/DS615) alleging GATT violations.
• Real‑world responses and circumvention
  • Firms adapted product lines (e.g., Nvidia A800/H800, later H20) aimed at complying with or evading controls; allegations of indirect routing (third‑country intermediaries) in enabling Chinese firms like DeepSeek.
  • Both states are investing heavily in domestic capacity (China: national big‑fund investments in chipmaking; U.S.: efforts to develop rare‑earth supply chains and “mine‑to‑magnet” initiatives).
• Legal analysis under WTO
  • The paper examines justiciability of security claims and whether Article XXI(b) covers AI‑related export controls.
  • Authors argue: security exceptions should be narrowly interpreted; commercial/dual‑use AI semiconductors and models generally do not qualify as measures “necessary to protect essential security interests.”
  • Overuse of security exceptions risks WTO liability and complicates dispute settlement.
• Net effect
  • Short term: controls can disrupt flows and raise costs, but are porous and spur circumvention and substitution.
  • Long term: controls accelerate indigenous R&D and domestic ecosystem formation, reducing intended leverage.

Data & Methods

• Nature of study: doctrinal legal and policy analysis combined with descriptive chronology and case exemplars.
• Primary documentary sources:
  • U.S. Federal Register/BIS rulemakings (Oct 2022; Oct 2023; Apr 2024; Dec 2024; Jan 2025 Framework).
  • Chinese customs/export control notices (2023–2025 notices on gallium, germanium, graphite, antimony, heavy REEs).
  • WTO filings (complaint WT/DS615).
  • Congressional materials (ENFORCE Act, committee statements).
  • Public reports, press coverage, and industry statements (e.g., Nvidia statements, reporting on DeepSeek).
• Analytical methods:
  • Doctrinal interpretation of GATT Article XXI(b) and related WTO jurisprudence, including assessment of justiciability, “essential security interests,” and necessity.
  • Policy analysis of bilateral tit‑for‑tat measures and their likely economic/innovation consequences.
  • Use of examples and market events (product substitutions, entity‑list additions, investment programs) as empirical illustrations rather than formal econometric analysis.

Implications for AI Economics

• Trade and diffusion effects
  • Export controls fragment supply chains and raise costs for AI compute and hardware, slowing cross‑border diffusion of capabilities.
  • Short‑run market disruptions can be significant (capacity constraints, price increases), but markets adapt via alternative suppliers, product splits, and third‑country routing.
• Innovation and industrial policy effects
  • Controls create strong incentives for targeted states to invest in indigenous capacity (chip fabs, design, model development), accelerating domestic R&D and industrial maturation.
  • Over time, this reduces the leverage export controls aim to secure and can produce parallel ecosystems (de‑facto decoupling).
• Strategic uncertainty and investment
  • Legal uncertainty (WTO disputes, shifting rulemaking) raises investment risk and may bias firms toward on‑shore investment or splitting operations across jurisdictions.
  • Firms with global footprints may face higher compliance costs, fragmentation of product roadmaps (e.g., split SKUs), and innovation inefficiencies.
• Welfare and global efficiency
  • While controls pursue national security, they can produce global welfare losses: slower global innovation, duplication of capital investment, and less efficient allocation of R&D resources.
• Policy recommendations suggested by the authors (economic reading)
  • Favor restrained, narrowly tailored export controls focused on demonstrable military end‑uses rather than broad commercial or dual‑use goods.
  • Pursue multilateral coordination where possible to reduce circumvention and informational problems while limiting collateral damage to civilian AI diffusion.
  • Clarify international legal standards (WTO interpretation of Article XXI) to reduce uncertainty and avoid weaponization of trade law that undermines global innovation.
• Takeaway for AI economists
  • Export controls are a strategic lever with complex dynamic effects: they can deliver temporary frictions but catalyze long‑term substitution and domestic innovation. Modeling and policy evaluation should therefore account for adaptive responses (supply‑chain reconfiguration, investments in endogenous R&D) and legal/dispute‑settlement risk when assessing the net economic impact.