AI Gigafactories: what does the EU call actually require on site connectivity?

On 30 July 2026 the European Commission opened the call to select up to seven “AI Gigafactories” across Europe. Press release IP/26/1708 is clear on the figures: up to €10 billion in EU and national funding, at least €20 billion in expected private investment, over €30 billion in total. The text that actually qualifies a candidate consortium — published by the EuroHPC Joint Undertaking under reference EUROHPC-2026-CEI-AIGF-01, deadline 12 November 2026 — puts a precise number on exactly one thing: how many AI processors the site must install, lot by lot and phase by phase. Everything that holds the site together around those processors — power, efficiency, and the network that gets data in and out — is described, in the public texts we were able to read, with adjectives, not thresholds. The question a specification must answer before the first processor arrives is not asked by the call. It is asked by whoever digs the duct.

The number exists for compute, lot by lot and phase by phase.

The Commission’s press release states this with a precision the rest of the text lacks. The call is split into two lots across two consecutive development phases. In Lot 1 — up to four projects, up to €100 million in EU funding in phase one and up to a further €400 million in phase two — “each Gigafactory selected in this lot must deploy at least as many of the most advanced AI processors as are currently installed in Europe’s most powerful AI factory”, and in phase two “capacity must increase to at least three times that level”. In Lot 2 — up to three projects, up to €200 million in phase one and up to a further €800 million in phase two — the phase-one ceiling is double that — “must deploy up to twice as many” — rising to “at least four times”. The EuroHPC JU call page summarises the overall threshold this way: every facility “must integrate at least three to four times the number of the most advanced AI processors currently available in the most powerful European AI factories”. That is a written, verifiable multiplier, tied to a public reference figure — the capacity of the largest AI factory operating in Europe today.

On the same page, in the very next sentence, the register changes: “In addition to computational scale, AIGFs must demonstrate adequate power capacity infrastructure and incorporate best practices in energy efficiency, water efficiency, and circularity principles in their design and operations.” “Adequate”, “best practices”: no figure, no threshold, no declared kW. Of connectivity, in this paragraph — the one that actually lists what a consortium must demonstrate to qualify — there is no trace.

Connectivity is named, not measured.

It is not that the call ignores the network: it names it, twice, in two different texts. The press release describes what an AI Gigafactory will combine: “advanced AI processors, software and cloud technology stacks, high-speed connectivity and energy-efficient data centres.” The EuroHPC JU page that defines what an AI Gigafactory is, citing Council Regulation (EU) 2026/150 of 16 January 2026, is more technical: it refers to supporting data centre infrastructures “including high-capacity storage and networking.” In neither place does the network receive a requirement: it is a listed ingredient, not a condition to be met with a figure. Compute has a multiplier, lot by lot and phase by phase; connectivity has a compound adjective, “high-speed”, that is neither measured nor tested.

”Distributed”: when the site is not a single place, the network is the only thing holding the facility together.

There is a detail in the press release that makes the gap more concrete, not less. An AI Gigafactory “may be established in a single Member State, either on one site or across several locations,” and may include cross-border developments “through distributed AI computing facilities.” A facility whose compute capacity sits physically in two or three places, possibly across two different Member States, is not held together by concrete: it is held together by the links that carry training data from one site to another. On that part — how many paths, with what physical diversity, with what latency and what declared optical budget — the call, in the public texts available to us, says nothing. That is exactly the point where a serious specification has to write down what the call leaves implicit.

The same gap shows up in the Italian dossier too.

This is not an isolated feature of this particular call. When Italy authorises a large-scale data centre, the PUCD single authorisation procedure is meticulous on the electrical side: a peremptory 90 days from the grid connection offer (STMG) for the environmental impact dossier, on pain of losing the queue position. MASE’s operational guidance lists the permits, the payments, the planning-compliance checks — and treats user connection networks as design deliverables that go into the file. But it is the electrical connection that has a clock which starts and a deadline that voids the application: for the fibre feed, the diverse paths, the cabling availability class, we found no Italian or European rule imposing an equivalent deadline. At every level, the system knows how to count megawatts. It is far less good at counting fibres.

What to write into a specification for a site of this scale.

The minimum number of physically distinct fibre entries into the site — not two cables in the same duct — with the point where the paths converge stated in metres, not left to a generic “guaranteed redundancy”.
The cabling availability class under EN 50600-2-4 stated in writing — class 3 or 4 for a facility of this scale, not “to industry standard” — consistent with the other two infrastructures under the same standard, because the site’s overall class is the lowest of the three.
Bidirectional OTDR certification of every entry span, with native traces handed over together with the report, not just an attenuation value within threshold.
As-built records and a register for every path, labelling included, delivered together with commissioning — not reconstructed from memory a year later.
If the consortium takes the “distributed” route across several sites, the optical budget and the redundancy of the inter-site links documented at design stage, not assumed to already exist somewhere.

The point.

The AI Gigafactories call puts a written multiplier on compute — one, two, three, four times the largest AI factory in Europe, lot by lot and phase by phase — because that is the figure the technological sovereignty the initiative claims to build actually rides on. On the network that carries data in, out and between the sites of a facility this size, we write the same rigour ourselves, in the specification: the cabling availability class becomes a clause, OTDR certification of every entry becomes an acceptance test, delivered together with the work — it is what we do from survey to certification on data centre projects, with tender specifications and compliance written before signature, not after. Every path, every OTDR trace, every as-built record — for a single site or for a consortium distributed across several Member States — goes into the same network map, on which an AI makes the diagnosis and a team closes the fault: for an operator, a data centre of this scale, an industrial site, a public authority, a healthcare facility or a defence site. Where AI is needed on network data, it runs within the client’s perimeter — on-premise or on dedicated cloud with the data centre located in Italy — together with CSIDIA, the group’s other company.

Are you assessing a specification for a large-scale data centre site, or do you need to document the connectivity of a project already under way? Talk to an engineer: the site survey is free of charge, and the cabling availability class gets written down before the digging starts, not after.