State provision of curated datasets and compute can plausibly be treated as WTO‑actionable subsidies, yet legal ambiguity and weak evidentiary tools mean subsidy disciplines will be hard to apply consistently; policymakers and adjudicators must navigate doctrinal pathways, specificity tests and opaque digital markets to determine when data‑led industrial policy distorts trade.

Main Finding

The paper argues that many forms of state provision of curated data and computational capacity—already central to modern AI industrial policy—can plausibly be characterized as subsidies under the WTO Agreement on Subsidies and Countervailing Measures (SCM Agreement). Using principles of technological neutrality and existing WTO jurisprudence (notably US – Large Civil Aircraft), Chen maps a doctrinal pathway showing how digital inputs (raw/curated data, compute access, and state data‑curation services) can meet the SCM tests of (1) financial contribution, (2) benefit, and (3) specificity. At the same time, the paper emphasizes substantial practical and evidentiary limits that make enforcement difficult and leave important policy space for states.

Key Points

• Framing and motivation

  • States now treat data and compute as strategic, production inputs for AI; policy tools include compute vouchers, curated data repositories, and national compute/data platforms (e.g., Chinese city computing vouchers, EU data spaces, US NAIRR).
  • This shift raises the question whether SCM disciplines (negotiated in a pre‑digital era) apply to intangible digital inputs.

• Interpretive foundation

  • Technological neutrality (and evolutionary treaty interpretation) supports applying 1994 rules to modern digital inputs rather than freezing meanings to tangible goods only.
  • WTO jurisprudence (US – Large Civil Aircraft) accepts that intangibles with commercial/industrial utility can fall within subsidy disciplines, opening the door to treating data and data‑related rights as subsidizable.

• How data/compute can fit SCM Article 1.1

  • Financial contribution: state provision of goods or services (including non‑tangible assets) that are not “general infrastructure” may count as a financial contribution under Art. 1.1(a)(1)(iii).
  • Categories distinguished: raw data (unprocessed), curated datasets (cleaned/labeled, high industrial value), compute resources (access to HPC), and data‑curation services (state performs costly prep work). Each can convey economic value.
  • Benefit: must be measured against a market benchmark (e.g., price paid vs. market price); overcompensation or privileged access indicates a benefit.
  • Specificity: targeted or limited distribution of the advantage (by firm, sector or region) can satisfy the SCM specificity test.

• Theoretical and policy context

  • Situates discussion in Sykes–Neven economic/legal debate: subsidies may be legitimate policy tools (market failures) but disciplines are focused on cross‑border distortions rather than every public expenditure.
  • Distinguishes technological neutrality (interpretive principle) from competitive neutrality (economic leveling principle).

• Practical and evidentiary constraints

  • The SCM framework can theoretically cover digital inputs, but practical barriers are large:
  • Difficulty defining/valuing datasets and compute in market terms; opaque digital pricing and lack of observable market transactions.
  • High information asymmetries and concentrated private markets make demonstrating injury and benefit challenging.
  • The WTO moratorium on customs duties on electronic transmissions complicates application of traditional countervailing duties to digital flows (enforcement paradox).
  • Ambiguity over “general infrastructure” and government claims of public interest/public good status for data projects.
  • Specificity findings require detailed evidence about target recipients and distribution; governments can design measures broadly to avoid specificity.

Data & Methods

• Methodological approach: legal doctrinal analysis and conceptual mapping rather than empirical econometrics.
  • Textual interpretation of SCM Agreement provisions (Article 1.1, Article 2, general infrastructure exception).
  • Review and application of WTO jurisprudence (Appellate Body and Panel reports, especially US – Large Civil Aircraft).
  • Normative and theoretical framing using economic literature on subsidies (Sykes, Neven) and industrial policy literature.
  • Typology construction for different forms of state data and compute provision (raw data, curated datasets, compute, curation services).
  • Use of contemporary policy examples (compute vouchers, NAIRR, EU data spaces, Chinese initiatives) to illustrate doctrinal points and practical realities.
• Evidence: legal precedent, treaty text, reported government programs and policy documents; the paper highlights the evidentiary burdens a complaining WTO Member would face in practice.

Implications for AI Economics

• Market structure and competition

  • State provision of curated data and subsidized compute can materially lower entry costs and accelerate domestic champions, reinforcing winner‑take‑most dynamics in AI.
  • Where government support is targeted and excludable, it can enhance incumbency advantages and raise global concentration.

• Global industrial policy and trade tensions

  • Recognizing data/compute as subsidizable inputs brings digital industrial policy squarely into trade remedies and dispute risk; however, enforcement frictions (valuation, evidence, moratorium) mean many measures will persist de facto.
  • Unclear WTO coverage increases geo‑economic tensions: states may exploit legal ambiguity to pursue strategic AI objectives while avoiding clear breach risk.

• Policy design recommendations to mitigate trade frictions (and for economists/policymakers to consider)

  • Make publicly funded data/compute broadly accessible and non‑discriminatory (to reduce specificity risk).
  • Price state services at observable market rates or publish price benchmarks and access conditions to reduce evidentiary opacity.
  • Increase transparency and notification of data/compute programs to facilitate international assessment and reduce dispute likelihood.
  • Where public objectives justify support (R&D, public goods), clearly document non‑commercial intent and open access provisions to help distinguish public investment from targeted subsidies.
  • Invest in internationally agreed metrics for valuing data and compute (benchmarks, proxies) to improve the feasibility of legal and economic assessment.

• Research needs for AI economics

  • Empirical work to value curated datasets and computing access (market proxies, transaction datasets, cost‑based valuations).
  • Quantitative analysis of how state data/compute programs change entry costs, innovation rates, and market concentration.
  • Development of metrics and evidence standards usable in trade disputes (e.g., shadow prices, comparable private transactions, usage intensity measures).

Overall, the paper shows that existing WTO subsidy law can be stretched to encompass state‑led data and compute support, but doing so will be analytically complex and practically constrained. For AI economics, this means trade disciplines may shape—but not fully constrain—the conduct and effects of public digital industrial policy unless evidentiary, definitional, and institutional gaps are addressed.