2.1.1 Jurisprudence on Intangible Inputs and Intellectual Property

Can data be characterized as a ‘good’ under the SCM Agreement even though not tangible? While not explicitly defining intangible data as a generic ‘good’, the rulings in US–Large Civil Aircraft paved the way for addressing data in subsidy law, with key implications for the modern data economy. In this case, the Appellate Body addressed the provision of scientific information and rights over data generated through government-funded research by NASA and the US DoD, recognizing the transfer of such data and tech rights as a financial contribution under the SCM Agreement.Footnote ¹⁹ The Panel in that dispute reasoned that the ordinary meaning of the term goods should not be restricted to tangible, possessable items. Instead, the Panel stated that there is no reason to assume that the ordinary meaning of goods necessarily excludes intangibles, particularly when those intangibles possess clear commercial value and are utilized as inputs in an industrial process.Footnote ²⁰

To discuss this issue, it is essential to examine the Sykes–Neven theory. The foundational way trade law looks at subsidies comes from the work of Alan Sykes and Damien Neven. They offer different views on whether subsidies are an economic problem that needs fixing or whether they are just a government’s way of making its own rulesthat should be respected. Neven and Sykes suggest that legal experts should check if a recipient is getting an unfair advantage. They argue that who owns the intellectual property should not be the only aspect that determines if a government action counts as a benefit.Footnote ²¹ Sykes contends that subsidies often serve legitimate public policy objectives and that broad-based trade restrictions on subsidies may be economically inefficient.Footnote ²² His economic analysis emphasizes that subsidies can address market failures (such as externalities in research and development) and that the case for subsidy disciplines rests on contestable assumptions about welfare effects.Footnote ²³

In both Neven and Sykes’ view, a government contract might still be an efficient arrangement even if it confers an advantage, unless the state overpays for services or provides over-compensation through data rights. This perspective emphasizes that the primary job of the law is to identify an economic advantage that is relative to the market, rather than focusing on the intangible nature of the asset itself.

The Sykes–Neven framework provides important context for understanding why state data provision warrants SCM disciplines even when considering Sykes’s skepticism toward broad subsidy regulation. Sykes argues that many subsidies reflect legitimate policy judgments and that disciplines should be calibrated to address only those subsidies that create clear cross-border distortions.Footnote ²⁴ Nevertheless, Sykes’s analysis posits that subsidy benefits are fully utilized upon transfer. Specifically, it assumes that once a financial transfer takes place, the subsidy is ‘expended’ and fails to yield continuous strategic advantages. Data subsidies operate under different economic logic. Because data is non-rivalrous, a government can provide the same high-quality training dataset to multiple domestic champions at minimal marginal cost, thereby creating perpetual, compound competitive advantages that persist and amplify over time as models are retrained and refined. This structural feature of data subsidies distinguishes them from the transitory subsidies Sykes discusses and aligns them with Neven’s concern regarding entrenched market enclosure. Furthermore, the excludability aspect of state-curated data, which is realized via technical gate-keeping and vetting mechanisms, implies that the socialization of artificial intelligence development costs is intentionally concentrated. This concentration differentiates data subsidies from truly public goods (such as widely disseminated basic research) and places them within Neven’s paradigm of selective rent creation.

Furthermore, the non-rivalrous nature of data does not disqualify it from being a ‘good’ in the trade law sense. On the contrary, as the state can provide the same high-value input to multiple domestic champions at near-zero marginal cost, this non-rivalry amplifies the scale of the potential distortion. This concentration of fiscal support and regulatory readiness creates a strategic advantage for firms in high-income economies, helping them to navigate the immense costs of data acquisition and model development.Footnote ²⁵ The high fixed costs associated with data resources and computing power are identified as primary factors contributing to the high concentration of AI capabilities in a handful of leading economies and dominant firms. The main point is not whether data is labeled as a digital product or a data-cleaning service, what really matters is that data is an economic resource from the government that gives a clear financial boost to a specific industry.

2.1.2 Distinguishing between Raw Data and Prepared Datasets

Does the legal character of state support change when data is processed for industrial use? The application of the Large Civil Aircraft jurisprudence to Generative AI requires a distinction between raw, unstructured information and AI-ready training sets. High-quality data is recognized not as a naturally occurring resource but as a foundational asset requiring substantial investment in an institutional and infrastructural capacity. Applying prior findings on scientific information to modern technology requires a distinction between raw, unstructured information and datasets that are ready for training. Besides, the collection, labeling, annotation, and verification of training data are the labor-intensive activities essential to the development and optimization of AI models, often demanding scalable storage solutions and compliance with complex data regulations. Raw data, such as unindexed medical records or unprocessed satellite imagery, has limited utility for training a Large Language Model (LLM). The commercial value is realized only through a labor-intensive lifecycle of data preparation.

How do government cleaning and labeling activities create a financial advantage for domestic firms? When a government helps create a national data pool, it is often providing much more than raw information. These initiatives typically involve the state that is checking the data to remove errors and labeling it to provide necessary context for machine learning. By doing this, the state engages in a state assumption of commercial costs that would otherwise be paid by the firm. The global market, for prepared AI training data, is projected to grow to US$ 17.04 billion by 2032,Footnote ²⁶ which demonstrates the economic value of this prepared assets. Access to resources and the high costs of development are significant barriers to entry. While existing rules in the SCM Agreement apply primarily to hardware, the state provision of these AI-ready assets functions as a production input with a clear market equivalent in the form of data licensing.

2.2.1 Functional Criteria for Public Bodies

The first attribution pathway involves the direct provision of data by a public body. In the landmark case United States–Definitive Anti-Dumping and Countervailing Duties on Certain Products from China, the Appellate Body established a functional test to define what constitutes a public body.Footnote ²⁷ An entity is a public body if it possesses, exercises, or is vested with governmental authority. The formal legal status of the entity is not dispositive: the core of the test is its function and relationship to the state.Footnote ²⁸

In the digital realm, state-owned commercial banks, national research institutes, and government-funded cloud platforms often fulfill this definition. These entities are frequently taskmasters of national AI strategies, and their provision of data resources under delegated governmental authority is a measure to achieve technological sovereignty. For instance, a national data platform managed by a government research council that aggregates public sector records and licenses them to domestic tech giants is acting as a public body. Its actions are direct expressions of state policy, and the economic resources it transfers are directly attributable to the government under Article 1.1(a)(1).

2.2.2 Government Direction of Private Data Resources

The second attribution pathway, provided for in Article 1.1(a)(1)(iv), applies when a government uses a private body as its proxy to make a financial contribution. In US – Countervailing Duty Measures (China), the Appellate Body affirmed that ‘entrust’ refers to giving responsibility to a private body and ‘direct’ means commanding a private body to perform a task. This doctrine provides a key interpretive tool for analyzing modern ‘data realms’, where governments may rely on regulatory mandates to compel private entities to socialize their data assets.Footnote ²⁹ This attribution pathway is crucial for analyzing policies such as the Common European Data Spaces, where private platforms may be commanded or incentivized to share their proprietary data with domestic rivals. A government action need not be a legally binding command to constitute direction, it is sufficient if the government provides an incentive that leaves the private body with no realistic commercial choice but to act in a particular way. For example, if a national AI strategy document provides R&D tax credits to private firms on the condition that they contribute their datasets to a national platform for use by startups, this constitutes a financial contribution through direction. By maintaining this surgical distinction, the analysis ensures that governments cannot bypass subsidy rules by using nominally private consortia or public–private partnerships to bestow data-based advantages.

2.3.1 Standards for Universality and Access

The Panel in EC and Certain Member States–Large Civil Aircraft provided the definition of the infrastructure exception, stating that general infrastructure is that which is available to all or nearly all entities in the economy.Footnote ³⁰ The criterion for this concept lies in its universality and non-excludability. Classic examples of general infrastructure include national road networks or public sanitation systems.

In contrast, state-facilitated data pools for AI are inherently excludable and selective. The legal standard for infrastructure requires that the resource be provided as a public good, yet data pools are frequently designed with technical and legal barriers that limit their use to a select group of technologically advanced firms. If a resource is restricted to a specific sector or to firms meeting high computational thresholds, it cannot be described as general. Besides, control over data infrastructures is a form of power that trade law has traditionally sought to discipline.Footnote ³¹ The strategic provision of AI-ready data to nurture national champions almost certainly constitutes the provision of goods or services other than general infrastructure. The failure to meet the ‘general infrastructure’ exception does not stem merely from the unequal technological capacity of downstream users. Rather, it derives from the state’s active design of technical gatekeeping, vetted user registries, and explicit eligibility criteria that legally and structurally restrict access to a designated group of domestic enterprises, thereby breaching the requirement of general availability across the economy.

2.3.2 Selective Curation for Industrial Use

This selective essence is manifested in three main ways. Firstly, there exists technical exclusivity. Access to petabyte-scale training sets necessitates specialized high-performance computing (HPC) environments that are restricted to pre-approved participants. Second, there is targeted curation: the commercial value of data lies in its preparation for specific tasks. When a state cleans, labels, and annotates data specifically for AI training, it is providing a bespoke production input tailored to a single industry (AI development), not a public good for general use. Third, there are national security and privacy carve-outs: many high-value datasets cannot be made universally public. When a government provides privileged access to these restricted datasets for a select group of domestic firms, it is furnishing a specific commercial asset. This threshold issue found in the Appellate Body Report, US–Softwood Lumber IV, Article 1 is distinct from the Article 2 specificity requirement, a distinction that this analysis strictly maintains.Footnote ³²

2.4.1 Market Benchmarks for Data Licensing

In Canada–Aircraft, the Appellate Body noted that although Article 14(d) primarily concerns the calculation of the amount of a subsidy in CVD investigations, it also provides essential context for the interpretation of benefit in Article 1. Under Article 14(d), a benefit exists if the government provides a good or service for less than adequate remuneration relative to prevailing market conditions.Footnote ³⁵ In the AI sector, the marketplace is no longer merely a site of transactions but a baseline for commercial rationality.Footnote ³⁶

When a government provides domestic AI firms with free or discounted access to curated training sets, it is acting inconsistently with the behavior of a commercial operator. The AI training dataset market alone is expected to experience rapid growth, with projections indicating that it will reach US$ 17.04 billion by 2032.Footnote ³⁷ This expansion reflects the increasing economic importance of data and data licensing as foundational inputs for the development of generative AI systems. This burgeoning market provides a baseline for constructing benchmarks. If a private data broker, recognizing the monopolistic potential of a high-quality dataset, were to demand an equity stake or substantial royalties, the state’s failure to seek such remuneration would constitute a clear economic benefit.

2.4.2 State Assumption of High Entry Costs

A more sophisticated paradigm for the AI treats data as a capitalized intangible asset. In the AI economy, the primary barrier to entry is the staggering cost of data preparation. Training new foundation models is an intensive process that demands massive computational power, vast datasets, and substantial capital investment. Furthermore, the physical infrastructure required, such as hyperscale data centers, involves significant financial outlays, often reaching billions of dollars.Footnote ³⁸ When a state undertakes data cleaning and labeling, it is not simply offering a ‘commodity’; rather, it is socializing the most substantial capital risk within the AI lifecycle. By endowing domestic firms with a cost advantage that is not available to foreign rivals, the state confers a strategic rent that directly improves the recipient’s market capitalization.Footnote ³⁹ This capitalization approach provides a more robust legal foundation compared to standard price-comparison models. This is because it has the structural advantage of being the first to train on exclusive national datasets in a market with winner-takes-all dynamics.Footnote ⁴⁰

2.5.1 De Facto Specificity and the Infrastructural Threshold

While some national data strategies are de jure specific (naming the AI sector in law), most utilize nominally neutral language. In such cases, the analysis must focus on de facto specificity under Article 2.1(c).Footnote ⁴¹ As established in United States–Subsidies on Upland Cotton, de facto specificity exists if the actual distribution of benefits is concentrated in a limited group.Footnote ⁴² This paper argues that state-facilitated data pools are inherently specific because of the infrastructural threshold.Footnote ⁴³

AI adoption tends to cluster in large, urban, and digitally connected firms. These firms, particularly those in highly productive metropolitan areas and innovation hubs, are led by larger enterprises with the resources necessary to manage complex production processes and invest in the digital infrastructure required for AI development and deployment.Footnote ⁴⁴ If a government data platform is theoretically open to every firm, but only a handful of domestic AI champions possess the high-performance computing power required to process petabytes of training data, those firms are the predominant users. Regarding the literature that also paid attention to the risk of such ‘digital inequality’, where infrastructural control generates massive rents for a few dominant firms,Footnote ⁴⁵ the state is making a targeted industrial investment whereby only a specific group of companies can realistically satisfy the specificity requirement.Footnote ⁴⁶

2.5.2 Industry Targeting and Restricted Access

The potential for AI to drive inclusive growth is challenged by the high concentration of development capacity within a limited number of firms and economies. High-income and upper middle-income economies account for most of global subsidy measures targeting AI-related products, which demonstrates a substantial risk of further concentrating AI capabilities and widening the structural gap across income groups. There comes a problem of data inequality, arguing that control over data conveys significant social and economic power. When a government oversight committee exercises discretion to approve data access for ‘trusted’ national champions while denying foreign competitors, it fulfills the Article 2.1(c) test of discretionary specificity.Footnote ⁴⁷

Furthermore, the curation process itself acts as a mechanism of enclosure. When a state cleans and labels data specifically for the requirements of a domestic AI model, it is providing a bespoke industrial input that has no utility for other sectors. This form of ‘algorithmic capture’ ensures that the benefits of state support remain locked within the national AI industry. By maintaining the surgical distinction between the definition of the subsidy and its specificity, this analysis demonstrates that national data policies are precise instruments of digital industrial policy. They are designed to benefit a domestic industry by providing privileged access to the algorithmic commons, thereby distorting the level playing field that the SCM Agreement was intended to protect.

3.3.1 Federal Support and Restricted Access: The First Prong of Article 1 Subsidy Definition

First, NAIRR receives financial support from the federal government in coordination with 10 agencies and 25 private partners, such as Amazon and NVIDIA.Footnote ⁵⁶ Critically, access to NAIRR resources is not universal or automatically available. Rather, it is expressly restricted to vetted research teams and domestic entities physically based in the United States.Footnote ⁵⁷ This gatekeeping structure immediately signals the non-universal nature of the program. While private firms cannot access NAIRR resources directly, they may do so indirectly by partnering with academic institutions designated as official recipients, thereby creating a domestically favoring partnership mechanism.Footnote ⁵⁸

This access restriction is decisive for meeting the requirements of Article 1(1)(a)(1) of the SCM Agreement, which stipulates that a subsidy exists only when there is a ‘financial contribution by a government’ that bestows ‘a benefit’. The SCM framework clearly acknowledges that not all government expenditures qualify as subsidies; instead, a subsidy must both offer financial assistance and confer an advantage to a particular recipient.Footnote ⁵⁹ By limiting access to vetted US-based teams, NAIRR does not act as a neutral public utility open to all economic participants. Instead, it functions as a targeted benefit mechanism, concentrating advantage among a restricted class of beneficiaries.

3.3.2 Technical Gatekeeping and the Rejection of the ‘General Infrastructure Exception’

Second, NAIRR does not qualify for the general infrastructure exception under SCM jurisprudence because it employs significant technical gatekeeping mechanisms combined with a selective vetting process.Footnote ⁶⁰ The concept of ‘digital gatekeeping’ refers to the institutional practice of restricting or mediating access to critical infrastructure based on control mechanisms implemented at multiple points and levels. In the context of NAIRR, technical barriers serve as a de facto form of access control. Specifically, only organizations equipped with adequate computational expertise and institutional capacity can effectively leverage the platform. This gate-keeping function gives rise to what scholars refer to as ‘infrastructural barriers to participation’, which are obstacles not only ingrained in formal policies but also embedded in the technical architecture of the system itself.Footnote ⁶¹

The unequal control over data infrastructure is acknowledged as a fundamental origin of economic power and strategic edge. Thomas Streinz has persuasively contended that ‘control over data infrastructures should not’ be concentrated in the hands of a single state or actor with privileged access.Footnote ⁶² Nevertheless, the architecture of NAIRR precisely does this: it confers control over petabyte-scale data resources and the computational infrastructure to the federal government, which will only allocate them to vetted domestic teams. Given that the resource is restricted to vetted US teams and prioritized for national research objectives, as manifested in the program’s stated mission of ‘advancing the AI industry’,Footnote ⁶³ it does not operate as a universally accessible public utility that can be utilized by both domestic and foreign competitors. Instead, the program’s capacity is restricted to a specific, favored group of users rather than to the entire economy.Footnote ⁶⁴

The WTO jurisprudence on specificity clearly establishes that programs which appear universally available but function with de facto or technical restrictions constitute specific subsidies.Footnote ⁶⁵ NAIRR’s technical design ensures that even if formally ‘open’, the practical barriers to entry and use ensure that AI developers dominate the user base, thereby satisfying the implicit targeting requirement embedded in the SCM framework.Footnote ⁶⁶

3.3.3 Market Benefit: Calculating the Subsidy Value through Commercial Comparators

Third, the benefit conferred by NAIRR is identified by comparing the zero-cost or heavily subsidized access to the market rates charged by commercial cloud providers such as AWS and Microsoft Azure.Footnote ⁶⁷ The SCM Agreement defines benefit as a comparative advantage: i.e., what matters is not the absolute value but the advantage the recipients achieve relative to what they would obtain on the open market.Footnote ⁶⁸

When the US government provides an AI startup with free or near-free access to petabyte-scale terrain mapping, medical imaging, or other specialized datasets, it is supplying a production input that has a clear market equivalent. Commercial cloud providers charge substantial fees for equivalent computational capacity and data access.Footnote ⁶⁹ Moreover, NAIRR offers specialized data curation services, including data cleaning, labeling, and preparation for machine-learning applications. This implies that the government undertakes the commercial costs associated with data preparation, which would otherwise pose a significant financial obstacle for domestic enterprises aiming to develop AI models in a competitive manner.Footnote ⁷⁰ Such cost-absorption by government is paradigmatic of the kind of financial contribution that generates a countervailable benefit under Article 1 of the SCM Agreement.

3.3.4 Specificity: Sector-Specific Targeting and De Facto Concentration of Benefit

Beyond mere access and benefit, the measure satisfies the specificity requirements of Article 2 of the SCM Agreement through multiple mechanisms. First, although NAIRR is nominally ‘open to various research purposes’,Footnote ⁷¹ its design and operational capacity ensure that AI developers become the predominant users. The program was explicitly designed with ‘the advancement of the AI industry as its central goal’. The technological infrastructure, including GPU clusters, specialized AI datasets, and frameworks optimized for deep-learning, is de facto dedicated to AI applications. Researchers aiming to utilize NAIRR in non-AI domains encounter technical obstacles and will find that the infrastructure is less than optimal for their requirements.Footnote ⁷² This implicit sectoral targeting satisfies de facto specificity under Article 2.1(c), which directs adjudicators to consider whether a subsidy programme is predominantly used by a limited number of enterprises.

Second, the limitation to US teams gives rise to explicit specificity based on geography and nationality, as the SCM Agreement stipulates that a subsidy is specific when ‘access to the subsidy is restricted to certain enterprises’ or when it is ‘de facto restricted to certain enterprises’.Footnote ⁷³ By confining the NAIRR to US-based researchers and excluding foreign competitors, the program centralizes its economic advantage within a defined domestic group. Yet from a WTO perspective, such nationality-based restrictions precisely delineate a specific subsidy subject to SCM regulations.

Third, although the National Artificial Intelligence Research Resource (NAIRR) does not explicitly impose export requirements, it indirectly attains comparable outcomes via its selectivity mechanisms. Access is contingent upon research teams meeting vetting criteria that are in line with the research priorities of the US government.Footnote ⁷⁴ This conditionality, namely the requirement for researchers to align their work with national objectives, operates as a form of de facto performance requirement analogous to those prohibited under Article 3.1(b) of the SCM Agreement, which bars subsidies contingent, ‘whether solely or as one of several other conditions, upon the use of domestic over imported goods.’Footnote ⁷⁵ Even though these criteria are phrased in neutral terms, such as ‘research excellence’ and ‘alignment with national interests’, they effectively direct the subsidy towards domestically favored firms whose research agendas are consistent with the priorities of the US government.

3.3.5 Synthesis: Non-Universal Infrastructure, Targeted Benefits, and Subsidy Determination

The National Artificial Intelligence Research Resource (NAIRR) represents a novel form of subsidy in the digital era: government-provided data and computational infrastructure, featuring restricted access, technical control, and de facto sectoral and geographical targeting, meet all the criteria of the definition in the SCM Agreement. The federal government offers free or highly subsidized access to computational and data resources valued at market rates, conducting specialized data preparation services, and sustaining the infrastructure through government funding.Footnote ⁷⁶ Access gained at zero or below-market rates means that recipients establish a comparative advantage over private-sector alternatives;Footnote ⁷⁷ and specificity through multiple means, such as de facto sectoral targeting towards AI developers, explicit geographical and nationality-based targeting towards US-based teams, and conditional access contingent on alignment with government research priorities.

Because NAIRR satisfies all three definitional prongs of a specific subsidy, it falls squarely within the SCM Agreement’s scope. It represents precisely the kind of government intervention that SCM disciplines are designed to regulate: a measure that provides advantage to domestic enterprises at the expense of foreign competitors, thereby distorting the conditions of fair international competition.

3.4.1 The Paradox of Non-Rivalry: Public Good versus Strategic Enclosure

The primary contradiction is the public good versus strategic enclosure paradox. Economically, data is a non-rivalrous resource: its use by one firm does not inherently preclude its use by another. This characteristic suggests that data should be a public good for global innovation. However, the mechanisms used by the ‘Big Three’ (vetted platforms, sectoral data spaces, and factor exchanges) effectively enclose the resource for the benefit of domestic champions. This selective enclosure of a non-rivalrous resource contradicts the principle of non-discrimination that underpins the WTO. States are utilizing the digital nature of data to create ‘club goods’ that shift the competitive relationship with foreign rivals without appearing on traditional balance sheets.Footnote ⁷⁹

3.4.2 The Conflict of Sovereignty: Regulatory Autonomy versus Multilateral Neutrality

Jurisdictions in all three realms justify their data socialization policies as acts of ‘digital sovereignty’. According to the World Trade Report 2025, high-income and upper-middle-income economies provide significantly more financial support for AI-related production. The report confirms that these economies account for the vast majority of the total count of global AI-related subsidy measures, demonstrating a substantial risk of further concentration of capabilities.Footnote ⁸⁰

3.4.3 The Knowledge Subsidy: Infrastructural Control as Factor Endowment

Lastly, states are no longer solely offering financial subsidies; instead, they are establishing the infrastructure conducive to knowledge generation. Thomas Streinz identifies this as a pervasive form of digital inequality, where unequal control over the infrastructures that generate and store data becomes a structural barrier to development.Footnote ⁸¹ The state provision of AI-ready data socializes the massive capital risks of the fourth industrial revolution for a limited group of firms. By doing so, it entrenches the dominance of those with the capacity to exploit these infrastructures, widening the gap between the ‘data haves’ and ‘data have-nots’. Trade can be a powerful enabler for an inclusive AI transformation, but the case studies suggest that without disciplining these data subsidies, the digital divide will only be further institutionalized. To maintain the integrity of the system, WTO Members must move beyond simple applications of law and engage with these structural realities, ensuring that the algorithmic commons are governed by the principles of fair competition and multilateral oversight.

4.1.1 The Linkage: Subsidizing Goods-Inputs for Service-Outputs

Under Article 1.1(a)(1)(iii) of the SCM Agreement, the primary legal test for identifying data provision as a financial contribution is its classification as the ‘provision of goods or services other than general infrastructure’. The value of such a subsidy is conceptualized as follows:

\begin{equation*}{V_{subsidy}} = \sum\limits_{i = 1}^n {\left( {{P_{market}} - {P_{gov}}} \right)} \times {D_{volume}}\end{equation*}

This equation quantifies the benefit conferred by calculating the delta between the prevailing market benchmark ( ${{\text{P}}_{{\text{market}}}}$) – such as commercial licensing fees for ‘AI-ready’ datasets or the market rental rate for high-performance GPU clusters – and the price actually charged by the government or public body ( ${{\text{P}}_{{\text{gov}}}}$). In jurisdictions where ‘compute vouchers’ or ‘data vouchers’ are utilized, ${{\text{P}}_{{\text{gov}}}}$ can represent a price as low as 20% of the market rate, effectively covering up to 80% of a startup’s operational costs in the Large Language Model training phase. This ‘negative cost’ advantage socializes the heavy capital risks of the fourth industrial revolution, creating a competitive distortion that satisfies the ‘benefit’ test established in Canada–Measures Affecting the Export of Civilian Aircraft, where the focus is on the advantage received by the recipient relative to the marketplace rather than the cost to the state.

The intangible nature of data should not serve as a shield against subsidy disciplines. As a foundational resource for training Large Language Models (LLMs), curated data possess distinct commercial value and market benchmarks (e.g., data licensing agreements). While Article 1.1(a)(1)(iii) excludes general infrastructure, platforms such as the NAIRR, the Shanghai Data Exchange, and Common European Data Spaces utilize technical gatekeeping and participant vetting mechanisms. Consequently, these do not meet the legal threshold of being ‘universally available’ and cannot be classified as general infrastructure.

A critical jurisdictional gap exists regarding the ‘input–output’ bridge in the AI industry. Most AI models are delivered as cross-border services via Mode 1 (cross-border supply) or Mode 3 (commercial presence), yet the strategic support they receive is often provided at the input stage. As established in United States–Measures Affecting Trade in Large Civil Aircraft (Second Complaint), the provision of scientific information and rights over data can be characterized as a financial contribution under the SCM Agreement. This suggests that even if the final product is a service, the state mobilization of curated data repositories can be disciplined as a subsidy to the model’s production.

AI is increasingly recognized as a general-purpose technology with the potential to transform the world economy and significantly boost international commerce. Simulations suggest that the use of AI could increase global trade in goods and services by nearly 40% by 2040.Footnote ⁸⁴ However, this growth is mediated by the algorithmic models that act as intermediaries. If a government provides its domestic AI industry with free or discounted access to national healthcare or financial archives, it is subsidizing the foundational resource of the service output. Under the principle of technological neutrality, as discussed in US–Gambling, commitments under the GATS apply regardless of the delivery mechanism.Footnote ⁸⁵ Yet, a fundamental asymmetry remains: while the SCM Agreement provides binding disciplines for goods-based inputs, GATS Article XV remains a mere placeholder for services subsidies. This services loophole creates a structural vulnerability, allowing states to potentially shield their digital champions by framing the entire support mechanism through the lens of services trade, where multilateral rules have reached a state of perpetual stasis.Footnote ⁸⁶ It is important to clarify the legal boundary: while the SCM Agreement explicitly disciplines subsidies to goods, digital AI models often bridge the gap between goods and cross-border services (Mode 1). For the SCM Agreement to directly apply in this hypothetical scenario, the complaining Member must either frame the AI model software itself as a digital good/tangible production input, or assert that the input subsidy directly distorts downstream trade in goods reliant on AI integration. Otherwise, if characterized purely as a service, the dispute reveals the very 'jurisdictional vacuum' under GATS Article XV discussed herein.

4.1.2 Legal Impediments to Applying CVDs to Service Suppliers

The lack of progress in negotiating disciplines under GATS Article XV has led to what scholars like Pietro Poretti describe as the ‘Waiting for Godot’ of trade law.Footnote ⁸⁷ Article XV recognizes that subsidies can have trade-distortive effects and mandates WTO Members to enter into negotiations to develop necessary disciplines, but these negotiations have remained in a rule-making stasis for nearly three decades. As a result, there are no explicit GATS rules for the calculation of benefit or the imposition of countervailing duties (CVDs) for services.

Applying traditional CVDs to AI service providers faces significant legal and practical obstructions. Unlike manufacturing sectors where an investigating authority can identify a ‘like product’ and assess its price in a domestic market, the AI services market is characterized by opaque, tiered, and highly customized pricing models, such as API-based subscriptions. Establishing a ‘normal value’ for a digitally delivered service is exceedingly difficult. Furthermore, traditional trade remedies are designed to offset price advantages at the border, yet services are supplied through a variety of modes that bypass customs checkpoints. Yi-ting Cathy Cheng has proposed adapting the SCM Agreement definition to services, taking into account the four modes of supply. However, without multilateral consensus, such unilateral attempts at ‘service CVDs’ would likely be challenged as violations of National Treatment or Market Access obligations. Regulatory coherence is essential for an inclusive AI transformation.Footnote ⁸⁸ The lack of progress in negotiating disciplines under GATS Article XV has led to uncertainty of trade law. Article XV recognizes that subsidies can have trade-distortive effects and mandates WTO Members to enter into negotiations to develop necessary disciplines, but these negotiations have remained in a rule-making stasis for nearly three decades. As a result, there are no explicit GATS rules for the calculation of benefit or the imposition of CVDs for services. However, the current framework faces challenges in ensuring a level playing field, as existing multilateral disciplines (including the SCM Agreement) apply primarily to goods and do not encompass services or intellectual property.Footnote ⁸⁹ Furthermore, the absence of specific safeguard mechanisms in services trade to address sudden import surges may discourage market opening and hinder the global diffusion of AI-related services.

4.3.1 The Burden of Proof Following the Expiration of Article 6.1

Article 6.1 of the SCM Agreement established a presumption of serious prejudice for certain egregious categories of subsidies, such as those where the total ad valorem subsidization exceeded 5% or those intended to cover operating losses. This provision was designed to simplify the burden of proof for complaining Members. However, Article 6.1 expired on 31 December 1999, after Members failed to reach consensus on its renewal. Consequently, there is no longer a ‘safe harbor’ or a presumption of prejudice for large-scale data subsidies.

Any claim of serious prejudice must now be based on positive evidence of adverse effects as detailed in Article 6.3. In the context of AI, a complaining Member needs to demonstrate that a state-led data sharing policy has resulted in significant price undercutting, price suppression, or a significant loss of market share in the global market for AI services. While the development of foundation models remains concentrated among a few dominant firms, the emergence of cost-efficient model architectures and open-weight alternatives is beginning to shift competitive dynamics within the AI sector. However, proving that a low-cost data advantage in one jurisdiction caused a specific price effect in another is an econometric challenge of the highest order. The non-market nature of state data provision and the opacity of digital pricing models mean that positive evidence is often difficult to obtain. As the Appellate Body noted in US–Upland Cotton, the causal link between the subsidy and the price effect must be clearly demonstrated, a requirement that places a heavy evidentiary burden on smaller economies with limited investigative capacity.Footnote ⁹³

4.3.2 The Dual Role of Injury in Multilateral and Unilateral Tracks

A final systemic consideration is the dual role of injury under the SCM Agreement. Footnote 11 of the Agreement clarifies that the term ‘injury’ applies both to the unilateral track (as a basis for CVDs) and to the multilateral track (as a category of adverse effect). This ensures that Members have multiple pathways to address injurious subsidization. In the global race for AI supremacy, proving that a data subsidy has materially retarded the establishment of a domestic AI industry in another Member country could become a central claim in future WTO litigation.

Trade can be a powerful enabler for an inclusive AI transformation by fostering the diffusion of innovation and unlocking new development pathways. However, the high concentration of targeted government support within a small number of economies suggests a substantial risk of further consolidating AI capabilities. Without proactive policy responses and greater international cooperation, these disparities risk locking in long-term inequalities and widening structural gaps across income groups. If a subsidized firm from a major data realm uses its structural advantage to dominate the algorithmic commons, it may prevent the emergence of localized AI solutions in developing regions. Proving this material retardation is legally possible under Article 5(a), but it requires a level of data transparency that the current notification system fails to provide. Strengthening notification requirements and utilizing the Trade Policy Review Mechanism to document non-monetary inputs like data access are essential steps to ensuring that the dual role of injury remains an effective safeguard against digital protectionism.

4.4.1 Prong 1: The Financial Contribution Test (Article 1.1)

Under Article 1.1(a)(1)(iii) of the SCM Agreement, a complaining Member (assume the United States) must establish that the EU has made a ‘financial contribution’ by ‘providing goods or services other than general infrastructure’.

Firstly, the United States would classify the European Health Data Cloud’s (EHDC) provision of curated health data as a ‘service’ within the framework of SCM jurisprudence, referring to the precedent set in the case of United States – Measures Affecting Trade in Large Civil Aircraft (Second Complaint).Footnote ⁹⁹ The United States would contend that data curation, which involves the labor-intensive procedures of cleansing, anonymizing, standardizing, and annotating medical records for machine-learning applications, amounts to the provision of services.

The government directly made the financial contribution. The direct subsidy and the annual government expenditure on services are financial contributions that confer a benefit to the recipient that meet the requirements of Article 1.1(a)(1)(i).Footnote ¹⁰⁰

First, data does not meet the criteria of ‘general infrastructure’. Unlike public road networks or sanitation systems, the EHDC is not universally accessible. The explicit limitation to entities registered in the European Union and the vetting condition (‘alignment with EU healthcare objectives’) indicate that the resource is selectively offered to a specific group of beneficiaries.Footnote ¹⁰¹ As determined in the case of EC and Certain Member States – Large Civil Aircraft, general infrastructure must be available to ‘all or nearly all entities in the economy’.Footnote ¹⁰² The EHDC’s exclusion of non-EU competitors fails to satisfy this criterion.

Second, data curation services possess commercial value. The United States would introduce economic evidence to illustrate the market value of comparable data-preparation services. Studies indicate that investing in high-quality training datasets in the artificial-intelligence domain implies high fixed costs, and a complete dataset suitable for training machine learning models requires a large scale of observations depending on its complexity.Footnote ¹⁰³ By regulating data access and sharing, the EU government is trying to strike a balance between private incentives and social benefits.

Regarding the attribution to the government, the United States would prove that the EHDC is a public entity according to the functional test established in United States – Definitive Anti-Dumping and Countervailing Duties on Certain Products from China.Footnote ¹⁰⁴ Although nominally managed by a ‘European Health Authority’ (a technically private entity), the authority exercises delegated governmental powers. It is fully funded by the government, its board is appointed by the EU and member-state governments, and its operations directly implement public health policies. Therefore, the financial contributions can be attributed to the EU government.

Even if the EU were to claim that data provision merely represents ‘information sharing’ rather than a ‘service’, the United States would adopt the substance-over-form approach. Regardless of the terminology used, the transfer of valuable, AI-ready data from the government to selected private enterprises constitutes the provision of a good or service for the purposes of Article 1.1.

4.4.2 Prong 2: Establishing Benefit (Article 1.1(b))

Upon establishing the financial contribution, the complaining Member is required to demonstrate that the measure bestows a ‘benefit’ upon recipients, which is an advantage in relation to the market.

The United States would adopt the ‘commercial comparator’ methodology. According to the ruling in Canada – Aircraft, a benefit exists when the government provides a good or service ‘at less than adequate remuneration relative to prevailing market conditions’.

The investigatory authority would formulate the benefit calculation as follows:

\begin{equation*}Benefit{\text{ }} = {\text{ }}\left( {Market{\text{ }}Price} \right){\text{ }} - {\text{ }}\left( {Government{\text{ }}-{\text{ }}Charged{\text{ }}Price} \right)\end{equation*}

where:

\begin{equation*}\begin{array}{*{20}{l}} {Market{\text{ }}Price{\text{ }}\left( {Pm} \right){\text{ }} = {\text{ }}cost{\text{ }}of{\text{ }}comparable{\text{ }}data{\text{ }}acquisition{\text{ }}and{\text{ }}curation{\text{ }}in{\text{ }}commercial{\text{ }}markets} \\ Government{\text{ }}-{\text{ }}Charged{\text{ }}Price{\text{ }}\left( {Pg} \right){\text{ }} = EUR\ 100{,}000\ per\ annum\ per\ firm \end{array}\end{equation*}

The US would present evidence demonstrating that the European Health Data Cloud (EHDC) constitutes a substantial implicit subsidy to domestic firms, as commercial market comparables reveal that medical-grade cloud storage, specialized AI data licensing, and high-quality data annotation for petabyte-scale datasets typically entail annual costs ranging from €50 million to 150 million, conservative independent valuations tailored to the EHDC’s specific scale appraise its integrated bundle of unrestricted health data, AI-ready preparation services, and high-performance computing .Footnote ¹⁰⁵

By charging beneficiary firms only €100,000 , the EU effectively delivers an annual subsidy of €50 million to 150 million per recipient. Since access to these premium datasets is currently restricted to just 15–20 EU-based biotech firms, the program concentrates €750 million to €3 billion in annual benefits among a narrow domestic cohort. This pricing structure creates a severe competitive disadvantage for US AI and biotech companies, which are compelled to either procure equivalent health data at 500 to 1,500 times the cost through commercial brokers or forgo access entirely, thereby restricting the geographic and epidemiological diversity of their training datasets and undermining their global competitiveness (see Table 1).

Table 1.

Benefit calculation summary table

Table 1 Long description

The table compares estimated market costs, government-provided rates, and the resulting implicit subsidy for major data and computing cost components. For data storage of 1,000 petabytes per year, market cost is USD 5 to 10M while the government rate is included, implying a USD 5 to 10M subsidy. Data preparation and curation is the largest component at USD 25 to 300M at market rates, also included by government, implying the same subsidy range. Quality assurance and compliance is USD 2 to 5M at market rates and included by government, implying a USD 2 to 5M subsidy. Computational infrastructure is USD 10 to 20M at market rates and included by government, implying a USD 10 to 20M subsidy. Overall, total annual benefit per firm is shown as market costs of USD 42 to 335M versus a government rate of USD 0.11M, yielding an implicit subsidy of about USD 41.9 to 334.9M. Ranges indicate uncertainty or variability across firms and assumptions, and the government rate being listed as included suggests costs may be covered through provision rather than direct billing.

Note: The Government–Charged Price of €0.1M (€100,000) is not allocated across individual cost components (hence Included in each row) but is netted against the aggregate market value at the total level. Precisely, Total Implicit Subsidy = €49.9M–€149.9M; this is presented rounded to €50–150M for consistency with the whole–million precision used elsewhere in this table.

This quantification demonstrates that the benefit is substantial, identifiable, and confers a decisive competitive advantage to beneficiary firms.

4.4.3 Prong 3: Establishing Specificity (Article 2)

The final criterion necessitates demonstrating that the subsidy is ‘specific’, i.e., concentrated within a particular enterprise or industry. Although nominally open to all EU-registered research organizations, the US would contend that the EHDC exhibits de facto specificity under Article 2.1(c) of the SCM Agreement through three converging mechanisms. First, technical gate keeping creates sectoral specificity: the platform’s infrastructure is heavily optimized for AI and machine learning while imposing structural barriers on traditional epidemiological or clinical researchers, thereby channeling benefits almost exclusively to AI developers and biotech firms. Second, the explicit EU-registration requirement establishes de jure specificity under Article 2.1(a) and, per United States – Subsidies on Upland Cotton, functions as geographic specificity that concentrates advantages domestically by systematically excluding major non-EU competitors such as OpenAI, Anthropic, DeepMind, and Chinese AI developers.

Third, the mandate that applicants demonstrate ‘alignment with EU healthcare policy objectives’ operates as a de facto performance requirement analogous to those prohibited under Article 3.1(b) of the SCM Agreement. This effectively conditions subsidy access on the extent to which AI development prioritizes European disease patterns and public health goals, thereby converting a broadly framed initiative into a targeted industrial policy instrument. This conclusion is reinforced by the WTO precedent in United States – Measures Affecting Trade in Large Civil Aircraft (Second Complaint), where the Appellate Body held that even unmarked government research transfers constitute specific subsidies when confined to a limited beneficiary group, making the EHDC’s explicit eligibility and performance conditions a clear instance of WTO-inconsistent specificity.

4.4.4 Prong 4: Demonstrating Serious Prejudice (Article 5)

After determining that the subsidy satisfies the criteria of Article 1 and Article 2, the United States would subsequently be required to demonstrate ‘serious prejudice’ to prove that the subsidy is actionable.Footnote ¹⁰⁶

Legal Standard: Article 5(c) of the SCM Agreement identifies ‘serious prejudice to the interests of another Member’ as an actionable category of adverse effect. Article 6.3 operationalizes this standard, providing that serious prejudice may arise where the effect of the subsidy is ‘displacement or impedance of the imports of like products of another Member into the market of the subsidizing Member’ (Art. 6.3(a)) or ‘displacement or impedance of exports of a like product of another Member from a third country market’ (Art. 6.3(b)).

The US would establish its evidentiary pathway by first characterizing the relevant ‘like product’ as generative AI models and AI-driven diagnostics. These products serve functionally equivalent roles in global healthcare markets irrespective of their origin in EHDC data or US-sourced commercial datasets. To demonstrate serious prejudice through export displacement, the US would present market data showing substantial long-term growth in the global medical AI sector, with the EU accounting for a significant portion of demand, while US firms’ regional market share has experienced severe erosion from its historically dominant global position within Europe.

Causal evidence from industry surveys and cost–benefit analyses would directly link this decline to the EHDC subsidy, revealing that EU procurers prefer domestic partners due to the platform’s low-cost data access, which drastically reduces AI development expenses and enables EU offerings to be priced substantially below US alternatives, thereby displacing a vast volume of annual US medical AI export sales and satisfying the WTO’s serious prejudice threshold. Furthermore, the US would demonstrate actionable prejudice under Article 5(c) by documenting third-market share erosion, as EHDC-subsidized EU firms leverage their structural cost advantage to undercut American competitors in Asia and Latin America, with the EU market share in international healthcare AI contracts rising sharply, directly mirroring corresponding losses for US exporters.

4.4.5 Prong 5: Remedial Constraints–the Implementation Deficit

Despite establishing a prima facie case of actionable subsidization causing serious prejudice, the US would confront significant systemic enforcement hurdles that severely constrain its remedial options. First, the absence of binding CVD disciplines under GATS Article XV means that even a favorable WTO panel ruling would not authorize unilateral CVDs on EU medical AI services, leaving Washington to choose between protracted DSB litigation spanning two to three years, diplomatically unlikely multilateral negotiations or legally precarious unilateral service taxes that risk violating WTO National Treatment obligations.

Second, any attempt to impose border measures on digitally delivered AI services would directly contravene the WTO e-commerce moratorium, which remains in force until March 2026,Footnote ¹⁰⁷ thereby forcing the US to either pursue the diplomatically fraught delisting of AI services or carefully design internal tax-based remedies that must strictly comply with GATT Article III.

Third, the EU would likely mount robust substantive defenses by characterizing the EHDC as a legitimate public health initiative shielded by public policy exceptions, arguing that the underlying health data remains ‘government-held’ and thus falls outside the SCM Agreement’s scope regarding privately received inputs, or by invoking the GATT Article XXI essential security exception to categorize sensitive health data as beyond the reach of standard trade disciplines.Footnote ¹⁰⁸ Collectively, these procedural, treaty-based, and doctrinal barriers would significantly complicate US enforcement despite a strong substantive claim. Nevertheless, the United States could counter such public-health justifications by demonstrating that the explicit exclusion of foreign entities constitutes an arbitrary or unjustifiable discrimination, or a disguised restriction on international trade under the chapeau of GATT Article XX.

4.4.6 Synthesis: The Hypothetical’s Instructive Value

This worked hypothetical demonstrates that while the SCM Agreement provides sufficient conceptual tools to classify government-integrated health data as a subsidizable input, calculate its financial benefit, establish specificity, and evidence serious prejudice,Footnote ¹⁰⁹ the complaining Member would nonetheless face remedial paralysis due to insurmountable practical obstacles: the absence of binding GATS Article XV disciplines for services countervailing duties, the WTO e-commerce moratorium blocking tariffs on digitally delivered AI, attribution ambiguities that enable defensive public-interest arguments. Without amending the GATS Article XV framework or renegotiating the e-commerce moratorium, WTO Members lack effective mechanisms to discipline highly trade-distortive data subsidies, thereby creating a structural loophole that invites governments to deploy data as an unpoliced industrial policy tool.Footnote ¹¹⁰ To restore the credibility of the multilateral trading system in the AI era, Members must urgently negotiate binding GATS Article XV disciplines targeting AI-related services subsidies, clarify the status of digital transmissions under the e-commerce moratorium to permit targeted countervailing measures, and develop WTO jurisprudence establishing that proven data subsidies trigger mandatory multilateral remediation regardless of current Article XV gaps.