Summary Under the proposed Cloud and AI Development Act (CADA) — COM(2026) 502 final, a proposal that is not yet in force — frontier AI priority projects would draw compute from two layers. Member States contribute AI computing resources, and under Article 9(2) the Union would "at least match" those contributions "to the extent that sufficient AI computing capacity is available within the Union's share of European high performance computing access time." That Union share runs through the European High Performance Computing Joint Undertaking (EuroHPC JU), set up by Council Regulation (EU) 2021/1173. Recital 35 says the EuroHPC JU access policy would be accommodated to reflect these allocations, but without prejudice to that Regulation's rules and to projects already holding EuroHPC AI compute. CADA does not rewrite the EuroHPC access policy or set quotas itself.

Detail

CADA is a proposal. Nothing below is in force, and the operational detail — how an allocation would actually be requested, scheduled and metered — is not written into the text. What CADA proposes is a coordination obligation layered on top of existing EuroHPC infrastructure, not a new supercomputing programme or a new access portal.

The gateway: recognition under Article 8

Compute matching attaches to a specific class of project. Under Article 8, the Commission would recognise, by decision, projects selected through open calls for expression of interest that support grand challenge 3 in Annex I. The criteria the proposal lists include that the project is a pioneering frontier AI project; that it is carried out by a European Digital Infrastructure Consortium (EDIC), established under Decision (EU) 2022/2481, or another entity eligible for Union funding, with at least three Member States participating; and — the criterion that feeds the compute pipeline — that participating Member States "pool computing time and other relevant resources."

That last point matters for how Article 9 works. The Member-State contribution is not a side effect of recognition; it is one of the conditions of recognition. A recognised frontier AI priority project is, by construction, one to which several Member States have already committed compute time.

The matching obligation: Article 9

Article 9 sets out how compute is secured once a project is recognised. The verbatim text of the first two paragraphs:

"1. 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. 2. 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."

Two things are worth reading carefully here, because both are routinely overstated.

First, the base layer. Article 9(1) is an obligation on both the Union and the Member States to allocate "sufficient AI computing resources from their compute capacities" — but expressly "within the limits of available capacity." This is not a guarantee of any particular volume. It is a best-efforts allocation duty bounded by what capacity exists.

Second, the matching layer. The Union "shall at least match" what Member States contribute, and this is the firmer of the two commitments — "shall at least match" is stronger language than the qualified allocation duty in paragraph 1. But it carries its own ceiling. The match is drawn specifically from "the Union's share of European high performance computing access time," and it applies only "to the extent that sufficient AI computing capacity is available" within that share. The Union's commitment is therefore conditional on, and capped by, the capacity actually available in its EuroHPC access-time share. It is not an open-ended promise to double whatever Member States put in; it is a floor ("at least match") that is itself gated by available Union capacity. If that share is largely committed, the realised match could be smaller than the nominal Member-State contribution.

For architects sizing a workload, the practical takeaway is that the matching figure is a planning assumption bounded by EuroHPC availability at the time, not a fixed multiplier you can underwrite a model budget against.

Where the Union's share comes from: EuroHPC JU

CADA does not build new machines. The "Union's share of European high performance computing access time" in Article 9(2) refers to the access time the Union holds in the supercomputers operated through the EuroHPC JU — the European High Performance Computing Joint Undertaking established by Council Regulation (EU) 2021/1173, a Council Regulation of 13 July 2021. Recital 26 notes, as context, that implementation of the Cloud and AI Leadership Initiatives could be entrusted to joint undertakings such as the EuroHPC JU established by that Regulation.

The EuroHPC JU already operates EU supercomputers and already runs an access-time allocation policy with its own rules and procedures. CADA's frontier AI matching has to fit inside that existing machinery rather than bypass it.

How CADA proposes to fit in: recital 35

Recital 35 is the hinge between CADA's new obligation and the established EuroHPC regime. It frames the Union's contribution as a match "on a proportional basis and within the limits of available EuroHPC capacity," and it states that this is "without prejudice to the rules and procedures laid down in Council Regulation (EU) 2021/1173."

On the access policy itself, recital 35 says:

"The EuroHPC JU access policy should be accommodated to reflect the allocation of such computing resources in an efficient, transparent and timely manner without prejudice to the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources."

Three precise points follow, and the wording carries each of them:

  • The access policy is to be adapted, not overridden. The verb is "accommodated." CADA does not itself rewrite the EuroHPC access policy, define new allocation tiers, or set percentages. It contemplates that the EuroHPC JU's own policy would be adjusted to reflect frontier AI allocations — leaving the mechanics to that Regulation's rules and procedures, which recital 35 expressly preserves.
  • Existing rules and procedures are preserved. The "without prejudice to the rules and procedures laid down in Council Regulation (EU) 2021/1173" clause means the matching mechanism is meant to operate through, not in derogation from, the EuroHPC legal framework.
  • Existing projects are protected. The accommodation is "without prejudice to the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources." Frontier AI priority projects do not displace, cancel or pre-empt the allocations of current EuroHPC users.

So the proposed integration is additive and forward-looking: the EuroHPC JU would adapt its policy to make room for frontier AI matching going forward, while existing allocations and the rights attached to them stay intact.

What the proposal does not say

Because this is a proposal and because recital 35 defers to the EuroHPC framework, several things are deliberately absent from CADA's text. There is no allocation percentage, no fixed quota for frontier AI projects, no application process or portal defined in CADA, and no service-level commitment on turnaround. The phrases "efficient, transparent and timely" are aspirations for how the EuroHPC JU access policy would be accommodated, not metrics. Anyone planning around CADA should treat the operational detail as something that would be worked out within the EuroHPC JU's access policy, not something the Act prescribes.

What this means for you

For CTOs and architects. If you are scaling frontier AI in the EU, the route the proposal sketches is consortium-based, not a direct compute application. Access to the matching layer flows from being recognised under Article 8 — which presupposes a pioneering frontier AI project, an EDIC or other Union-funding-eligible entity, at least three participating Member States, and pooled Member-State compute time. Treat the Article 9(2) match as a capacity-bounded "at least match" on top of those Member-State contributions, not as a guaranteed doubling. Watch for changes to the EuroHPC JU access policy, since recital 35 indicates that policy — not CADA — is where the allocation mechanics would land.

For SMEs and start-ups. The frontier AI priority project route is built for large multi-Member-State consortia, so direct access is realistically limited. CADA does not, in the verified text relied on here, create a separate matched-compute guarantee for smaller projects. If you need EuroHPC time, the existing EuroHPC JU access channels under Council Regulation (EU) 2021/1173 remain the route, and partnering with a consortium or national research body is the practical path into the frontier AI track.

For legal and compliance teams. The match in Article 9(2) is contingent on what Member States actually "contribute," so contribution definitions matter. Pin down what "computing resources" and "computing time" mean in any consortium agreement (for example, the metering basis), because the Union match is keyed to those contributions and capped by available capacity. Track how the EuroHPC JU accommodates its access policy, and note that recital 35 preserves the rights of existing EuroHPC projects — useful if your organisation already holds an allocation.

Common misconceptions

"Any AI project can get matched EuroHPC compute." No. The matching obligation in Article 9(2) attaches to frontier AI priority projects recognised under Article 8. A project that is not recognised under Article 8 does not trigger the Article 9(2) match.

"CADA doubles your compute." Overstated. Article 9(2) is an "at least match," but it is drawn from the Union's EuroHPC access-time share and applies only "to the extent that sufficient AI computing capacity is available" within that share. The realised match depends on EuroHPC availability; it is a capacity-bounded floor, not a fixed 2x.

"CADA bumps existing EuroHPC users." No. Recital 35 makes the accommodation of frontier AI allocations "without prejudice to the continuity of ongoing operations and the rights of projects already benefiting from allocated EuroHPC AI computing resources," and "without prejudice to the rules and procedures laid down in Council Regulation (EU) 2021/1173." Existing allocations are protected.

"The EU provides compute regardless of Member State input." No. The Union matches resources "contributed by Member States." Pooling Member-State compute time is also one of the Article 8 recognition criteria. Without a Member-State contribution there is no base for the Union to match.

"CADA sets the allocation rules and quotas." No. CADA proposes that the EuroHPC JU access policy "should be accommodated"; it does not itself prescribe quotas, percentages, or an application process. The mechanics sit within the EuroHPC framework.

Official sources

Related

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