Summary Under the proposed Cloud and AI Development Act (CADA), "AI computing resources" for frontier AI projects are not defined by specific hardware specifications (such as GPU counts or FLOPS) but by access time on existing high-performance computing (HPC) infrastructure, specifically the Union's share of European High Performance Computing (EuroHPC) access time. Article 9 of the proposal mandates that the Union and Member States allocate sufficient resources from their compute capacities to support designated "frontier AI priority projects." Crucially, the Union commits to matching the computing resources contributed by Member States, but only "to the extent that sufficient AI computing capacity is available within the Union's share of European high performance computing access time." This mechanism leverages the existing EuroHPC ecosystem rather than creating a separate hardware pool, prioritizing projects that are pioneering, involve broad Union participation, and focus on scaling frontier technologies.

Detail

The proposed Cloud and AI Development Act (CADA), COM(2026) 502 final, establishes a targeted framework to support the development of "frontier AI," defined in Article 2, point (4) as "AI models or AI systems built upon such models that can perform a wide variety of tasks and that approach, reach or exceed the current state of the art." To enable the development of such advanced models, which require unprecedented computational power, the proposal introduces a specific regime for the allocation of "AI computing resources."

Unlike typical infrastructure regulations that might define resources by physical assets, CADA defines these resources procedurally through access rights to existing supercomputing capacity. The core of this definition is found in Article 9, titled "Computing support for AI projects," and is contextualized by Recital 35 of the explanatory memorandum.

The Legal Definition: Access Time, Not Hardware

Article 9 does not list specific hardware types (e.g., AI-optimized servers, quantum accelerators) as the definition of "AI computing resources" in the context of support allocation. Instead, it anchors the definition in the allocation of capacity from existing high-performance computing (HPC) systems.

The article operates on three distinct pillars:

  1. Mandatory Allocation (Article 9(1)): The Union and Member States are legally obliged to ensure that "sufficient AI computing resources from their compute capacities are allocated to support the development of frontier AI priority projects." This allocation is conditional: it must be within the "limits of available capacity." The projects eligible for this support are those designated as "frontier AI priority projects" under Article 8, which requires them to be pioneering, focused on scaling frontier AI, and involve broad participation (e.g., via European Digital Infrastructure Consortia or entities eligible for Union funding involving at least three Member States).

  2. The Union Matching Mechanism (Article 9(2)): This is the most critical provision for defining the scope of resources. It states: "The Union shall at least match the AI computing resources contributed by Member States to frontier AI priority projects to the extent that sufficient AI computing capacity is available within the Union's share of European high performance computing access time."

    • The "Union Share": This phrase explicitly ties the definition of available resources to the EuroHPC Joint Undertaking (EuroHPC JU) framework. It refers to the portion of supercomputing access time funded by the EU budget, distinct from capacity funded solely by Member States.
    • The Matching Rule: If a Member State contributes a specific amount of compute time (from its national share or co-funded share) to a priority project, the Union is obligated to contribute an equivalent amount from its own share.
    • The Cap: This matching is not unlimited. It is strictly capped by the availability of capacity within the Union's share of EuroHPC access time. If the Union's allocated time is fully utilized by other operations, the matching obligation cannot be fulfilled.

  3. Broader Support (Article 9(3)): Beyond the strict matching for frontier AI, the proposal states that the Union and Member States shall "endeavour to provide sufficient computing resource for AI industrial innovation, physical AI and public sector AI projects." This creates a secondary, non-binding channel for support for other strategic AI domains, though it lacks the mandatory matching mechanism of paragraph 2.

Recital 35: Operational Context and EuroHPC Integration

Recital 35 provides the essential interpretative context for Article 9, clarifying that the proposal does not intend to build a parallel infrastructure but to integrate with the existing EuroHPC ecosystem.

The recital acknowledges the "unprecedented scale of resources required for frontier AI development" and confirms that the Union's matching is "on a proportional basis... within the limits of available European high performance computing (EuroHPC) capacity."

Crucially, Recital 35 addresses the operational friction between new priority projects and existing commitments. It states that the "EuroHPC JU access policy should be accommodated to reflect the allocation of such computing resources in an efficient, transparent and timely manner." However, it includes a vital safeguard: this accommodation must be done "without prejudice to the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources."

This confirms that "AI computing resources" under CADA are a prioritization mechanism within the existing EuroHPC access time pool. They are not new hardware; they are reserved slots on supercomputers that are already part of the Union's HPC infrastructure.

High-Performance Computing for Training and Scaling

The proposal implicitly recognizes that "training and scaling" frontier AI models requires the massive parallel processing capabilities found only in high-performance computing environments. By linking support to EuroHPC access time, CADA ensures that frontier AI projects have access to the specific type of infrastructure required for:

  • Training: The iterative, compute-intensive process of training large-scale models.
  • Scaling: The ability to expand model capabilities and handle complex, multimodal tasks.
  • Validation: Testing models in real-world environments as referenced in Article 4(4) regarding physical AI.

The definition of resources is therefore functional: it is the time required to run these training and scaling workloads on the Union's most powerful supercomputers.

The "Union Share" and Access Time Dynamics

The concept of the "Union share of access time" is central to understanding the scope of CADA's support. Under the EuroHPC framework, supercomputers are often co-funded by the EU and Member States. The "Union share" refers to the access time rights that belong to the EU budget contribution.

  • Leverage: The matching mechanism in Article 9(2) is designed to leverage national contributions. If a Member State commits 100 hours of its national share to a frontier AI project, the Union is obligated to commit 100 hours from its share, effectively doubling the resource pool for that project.
  • Constraint: This leverage is bounded by the physical reality of the "Union share." If the Union's allocated time is already committed to other strategic projects (e.g., climate modeling, drug discovery, or existing AI factories), the matching obligation cannot be met.
  • No New Hardware Mandate: The proposal does not mandate the construction of new supercomputers solely for this purpose. Instead, it relies on the "limits of available capacity" within the current EuroHPC fleet.

What this means for you

For technology leaders, research consortiums, and AI developers, the definition of "AI computing resources" in CADA has profound strategic implications.

1. Consortium-Based Access is Mandatory

You cannot apply for these resources as a standalone entity. Article 8 stipulates that frontier AI priority projects must be undertaken by a "European digital infrastructure consortium established pursuant Decision (EU) 2022/2481 or another legal entity eligible for funding under Union law" and must involve "the participation of at least three Member States."

  • Action: If you are an SME or a research lab, you must partner with a consortium or join an existing EDIC to qualify for the matching compute resources.

2. The "Matching" is Your Leverage

The Union's commitment to match Member State contributions is a powerful tool for project funding.

  • Strategy: When designing your project proposal, explicitly quantify the compute time you need from Member States. A strong national commitment (e.g., from a national HPC center) triggers the Union's obligation to match it. This effectively doubles your access to EuroHPC capacity without additional cost to the project, provided the Union share is available.

3. Priority is Not Guaranteed

Access is subject to the "limits of available capacity." The EuroHPC JU access policy must be "accommodated," but existing rights are protected.

  • Reality Check: Even with a designated priority project status, you may face competition for the Union's share of access time. Early engagement with the EuroHPC JU and national HPC centers is essential to align your project timeline with available slots.

4. Beyond Frontier AI: The "Endeavour" Clause

If your project is "industrial AI," "physical AI," or "public sector AI" but does not meet the strict "frontier" criteria of Article 8, you are not excluded entirely.

  • Opportunity: Article 9(3) states the Union shall "endeavour" to provide resources for these sectors. While not a binding matching obligation, this creates a political and administrative channel to advocate for access to HPC resources for non-frontier but strategically important initiatives.

5. Focus on "Time," Not "Specs"

Do not expect CADA to define resources by chip architecture (e.g., "only EU-made GPUs"). The regulation defines resources by access time.

  • Implication: Your project planning should focus on the duration and intensity of compute time required, rather than specific hardware procurement. The infrastructure is provided; your job is to optimize your workload for the available time slots.

Common misconceptions

Misconception 1: CADA creates a new, separate pool of supercomputers. Correction: No. Article 9 and Recital 35 confirm that the resources are drawn from the existing EuroHPC infrastructure. The proposal creates a prioritization and matching mechanism for access time on these existing machines, not a new hardware fleet.

Misconception 2: Any innovative AI project can access these resources. Correction: The mandatory matching and priority allocation in Article 9(2) are strictly limited to "frontier AI priority projects" that meet the rigorous criteria of Article 8 (pioneering nature, multi-state participation, scaling focus). General AI projects do not automatically qualify for the Union matching mechanism.

Misconception 3: The Union will provide unlimited compute time. Correction: The matching obligation is explicitly capped: "to the extent that sufficient AI computing capacity is available within the Union's share." If the Union's share is fully utilized, the matching cannot occur. Access is constrained by physical availability.

Misconception 4: CADA defines resources by specific hardware metrics like FLOPS. Correction: The proposal defines resources procedurally as access time on HPC systems. While the value of that time depends on the hardware's performance, the legal definition in Article 9 is tied to the allocation of time within the EuroHPC framework, not the specific technical specifications of the chips.

Misconception 5: Existing EuroHPC users will lose their access. Correction: Recital 35 explicitly protects existing users. The allocation for frontier AI must be done "without prejudice to the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources."

Official sources

Related

This is general information about a draft EU regulation, not legal advice.