Summary Under Article 9 of the proposed EU Cloud and AI Development Act (CADA) — COM(2026) 502 final, a Commission proposal that is not in force — the Union and the Member States would "ensure that sufficient AI computing resources from their compute capacities are allocated to support the development of frontier AI priority projects" recognised under Article 8, but only "within the limits of available capacity" (Article 9(1)). On top of national contributions, the Union "shall at least match" the AI computing resources contributed by Member States, "to the extent that sufficient AI computing capacity is available within the Union's share of European high performance computing access time" (Article 9(2)). For a broader set of projects — AI industrial innovation, physical AI and public sector AI — the duty is weaker: the Union and Member States only "shall endeavour to provide sufficient computing resource" (Article 9(3)). Article 9 concerns access to compute, not cash grants, and creates no new fund.

Detail

CADA is a legislative proposal. Nothing in Article 9 is in force, the wording can change during the legislative process, and the obligations described below bind the Union and the Member States rather than private providers directly. Article 9 is also about computing resources — access time on high-performance computing infrastructure — not about money. CADA does not create a new fund, and the matching mechanism in Article 9(2) is a matching of compute, not of grants.

The provision rests on two existing pieces of EU architecture: the frontier AI priority project regime in Article 8, and the European High Performance Computing Joint Undertaking (EuroHPC), established by Council Regulation (EU) 2021/1173.

What Article 9(1) requires

Article 9(1) provides that "the Union and the Member States shall ensure that sufficient AI computing resources from their compute capacities are allocated to support the development of frontier AI priority projects that fulfil the criteria set out in Article 8, within the limits of available capacity."

Two points matter for a provider reading this. First, the obligation runs to "the Union and the Member States" — they allocate resources "from their compute capacities," i.e. from public compute they already hold or control. It is not a duty imposed on commercial cloud or data centre operators. Second, the duty is qualified. "Shall ensure" sounds firm, but it is bounded by "within the limits of available capacity." There is no fixed quantity, no guaranteed minimum, and no commitment to build new hardware to satisfy demand. If public capacity is fully committed, the obligation does not manufacture more.

To benefit, a project must be a recognised frontier AI priority project under Article 8. Recognition is by Commission decision, for projects selected through open calls for expression of interest supporting grand challenge 3 in Annex I. The Article 8 criteria require that the project (a) is a pioneering frontier AI project; (b) is carried out by a European Digital Infrastructure Consortium (EDIC) under Decision (EU) 2022/2481 or another entity eligible for Union funding that brings together at least three Member States; and (c) involves participating Member States that pool computing time and other relevant resources. Compute support under Article 9 therefore flows only to projects that have already cleared this Article 8 gateway.

The Union's matching commitment in Article 9(2)

Article 9(2) is the most concrete of the three paragraphs: "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 mechanism is that the Union draws on its own share of EuroHPC access time and uses it to match what Member States contribute. "Shall at least match" is a stronger formulation than the best-efforts wording in Article 9(3), but it is still conditional. The match only happens "to the extent that sufficient AI computing capacity is available within the Union's share" of EuroHPC access time. If the Union's share is exhausted, the matching duty does not extend beyond it.

Recital 35 adds useful colour without changing the operative text. It describes the allocation as being of strategic importance and explains that the Union should match the resources contributed or committed by Member States "on a proportional basis and within the limits of available EuroHPC capacity." It also states that this is without prejudice to Council Regulation (EU) 2021/1173, and that the EuroHPC access policy should be accommodated to reflect such allocation in an efficient, transparent and timely manner — but without prejudice to the continuity of ongoing operations and to the rights of projects already benefiting from allocated EuroHPC AI computing resources. In other words, existing EuroHPC users are not displaced to make room for CADA frontier projects.

A note of caution on how Article 9(2) is sometimes described: it is "at least match," read together with Recital 35's "proportional basis." That is not the same as a promise to double every euro of national compute, and it should not be presented as an unconditional guarantee. The match is capped by available capacity within the Union's EuroHPC share.

Article 9(3): a weaker, broader duty

Article 9(3) provides that "the Union and the Member States shall endeavour to provide sufficient computing resource for AI industrial innovation, physical AI and public sector AI projects."

Two differences from paragraphs (1) and (2) are important. First, the obligation is weaker. "Shall endeavour" is a best-efforts duty — the parties must try, not guarantee — and it is markedly softer than the "shall ensure" of 9(1) or the "shall at least match" of 9(2). There is no matching mechanism and no link to a defined share of EuroHPC access time. Second, the scope is broader. Paragraph (3) reaches beyond frontier AI priority projects to cover AI industrial innovation, physical AI and public sector AI projects. These categories are named in the text itself; the proposal does not, in Article 9, define them by reference to specific sectors, so they should not be over-specified.

The practical upshot is a layered structure. Frontier AI priority projects get a qualified "shall ensure" allocation (9(1)) plus a qualified Union match (9(2)). The wider field of industrial, physical and public sector AI gets only a best-efforts "shall endeavour" (9(3)). None of the three paragraphs is an unconditional guarantee of compute.

What this means for you

As a provider, the first thing to keep in mind is that Article 9 places duties on the Union and the Member States, not on you. It does not require commercial operators to reserve capacity, and it does not, by itself, give you a right to claim Union compute. Its relevance to you is indirect, and it is contingent on a proposal that has not been adopted.

Frontier work runs through Article 8 and EuroHPC. If you want to support a frontier AI priority project, the route is through a recognised Article 8 project — typically led by an EDIC or by an entity eligible for Union funding spanning at least three Member States, with participating Member States pooling compute. Providers usually engage as infrastructure partners within such a consortium rather than as direct applicants for Union compute.

Treat the matching commitment as conditional. Article 9(2)'s match is bounded by the Union's available EuroHPC share, and Recital 35 frames it on a proportional basis. Do not build a business case on the assumption of guaranteed or doubled compute.

The broader market signal sits in 9(3). Industrial, physical and public sector AI are flagged as areas where the Union and Member States will "endeavour" to provide compute. That is a policy direction, not a funding promise, but it indicates demand beyond frontier model training.

Existing EuroHPC arrangements are protected. Recital 35 preserves the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources. If you operate within or alongside EuroHPC, the proposal does not, on its face, contemplate displacing those rights.

Watch the EuroHPC access policy. Recital 35 contemplates that the EuroHPC access policy may be accommodated to reflect these allocations. Any operational detail relevant to providers is likely to surface there rather than in Article 9 itself.

Common misconceptions

Misconception 1: Article 9 hands out money or creates a CADA fund. Reality: Article 9 concerns AI computing resources — access to compute — not cash grants. The matching in Article 9(2) is a matching of compute drawn from the Union's share of EuroHPC access time, and CADA creates no new fund.

Misconception 2: The Union will double every Member State contribution. Reality: Article 9(2) says the Union "shall at least match," and only "to the extent that sufficient AI computing capacity is available within the Union's share" of EuroHPC access time. Recital 35 frames the match "on a proportional basis and within the limits of available EuroHPC capacity." It is a capped, conditional match, not an automatic doubling.

Misconception 3: All three obligations are equally binding. Reality: The strength differs by paragraph. Article 9(1) is "shall ensure ... within the limits of available capacity"; 9(2) is "shall at least match ... to the extent that sufficient ... capacity is available"; 9(3) is only "shall endeavour to provide sufficient computing resource" — a best-efforts duty. Paragraph (3) also covers a broader set: AI industrial innovation, physical AI and public sector AI projects, not just frontier AI priority projects.

Misconception 4: Any AI developer can claim this compute directly. Reality: Compute support under Article 9(1)-(2) flows to projects recognised under Article 8 — pioneering frontier AI projects carried out by an EDIC or other Union-funding-eligible entity covering at least three Member States, with Member States pooling computing time. Private providers typically participate as partners within such projects.

Misconception 5: "Available capacity" is an open-ended promise. Reality: The qualifiers "within the limits of available capacity" (9(1)) and "to the extent that sufficient AI computing capacity is available within the Union's share" (9(2)) are real constraints. If public capacity or the Union's EuroHPC share is committed, the obligations do not generate more, and Recital 35 protects existing EuroHPC users from being displaced.

Official sources

Related

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