EuroHPC AI Gigafactories: sovereign access is planned, not yet a live service

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A manufacturer may want European control over AI workloads without buying its own very large compute estate. That need is not solved merely by locating servers in the European Union: operators, support access, logs, model services, commercial terms and human authority still matter.

A procurement model, not an operating cloud

EuroHPC AI Gigafactories (AIGFs) are a planned EU infrastructure and service-access initiative. The July 2026 announcement describes industry-led facilities combining large-scale AI compute, storage, high-bandwidth networks, secure cloud access and support for training, fine-tuning and inference for researchers, public authorities, industry, start-ups and small and medium-sized enterprises in EuroHPC Participating States. 1

By 27 October 2026, zero operational AIGFs with publicly evidenced customers or realised service outcomes were documented. The tender remained open under reference EUROHPC-2026-CEI-AIGF-01: Lot 2 was due 3 December 2026 and Lot 1 had been extended to 4 February 2027. Selected projects were expected in 2027, followed by up to 18 months of Phase 1 set-up. These are procurement milestones, not go-live dates or service-level commitments. 2 3

The initial procurement targets up to seven AIGFs. Planned Phase 1 public contributions are EUR100 million for a Lot 1 AIGF or EUR200 million for a Lot 2 AIGF, capped at 17% of project information-technology capital expenditure, whichever is lower. Expected private investment above EUR20 billion and these funding figures are plans, not awarded payments, built capacity or customer spend. 1 3

Three layers and two access routes

The intended workflow is: a user brings a workload or AI-service need; access is allocated from EuroHPC or a participating state, or a private organisation makes a commercial request; the AIGF supplies a service layer; and the user trains, fine-tunes, deploys or consumes inference. The documents do not yet provide a common end-user form, private-SME scoring rubric, queue service-level agreement or acceptance evidence. 3

Layer 1 is managed infrastructure and orchestration for expert users. Layer 2 is managed machine-learning operations and development platform services. Layer 3 is integrated AI, model-as-a-service or software-as-a-service, including hosted models, inference endpoints, guardrails and compliance tooling, aimed partly at less specialised SMEs. The planned public allocation mix is 50% Layer 2 (including Layer 1) and 50% Layer 3 for Lot 1; Lot 2 is 60% and 40%. 4

This is a public-anchor/off-take model rather than a published pay-per-use catalogue. EuroHPC and Member States would procure guaranteed access through framework and specific contracts. Bidders submit normalised hourly prices by hardware family, but no winning tariff, commercial price card, SME discount, minimum commitment or pay-as-you-go terms were public at the cutoff. 3 4

Sovereignty does not clean the data

The tender goes beyond EU hosting. The coordinator must be an EU-incorporated, effectively EU-controlled entity, and core operations—including orchestration, data handling, maintenance, disaster recovery, logs and telemetry—must be performed and processed in the EU under EU Member State jurisdiction. 4

That does not make every downstream service sovereign. International anchor customers may retain autonomy over value-added services, so a buyer must check jurisdiction, privileged support, model weights, backups, subprocessors and data egress for each layer and contract. Nor do the documents promise ERP, product-lifecycle-management, computer-aided-design, configure-price-quote, certification or logistics integration. Sovereign compute cannot prevent an obsolete drawing or stale price from reaching a recommendation without an authorised engineering or commercial review.

An illustrative SME test

Consider an illustrative industrial SME with a restricted library of drawings, inspection records and supplier documents. A bounded first test could use a future Layer 2 or Layer 3 service to classify and retrieve approved records. The SME would need canonical identifiers, revisions, permissions, material and tolerance fields, owner-approved source versions and current commercial data. An engineer remains authoritative for feasibility; a commercial owner approves price, terms and any binding quote.

Measure authoritative-record retrieval, wrong-revision rate, unsupported claims, review time and refusal when evidence is missing. Do not call a faster prototype a customer outcome. The adjacent programme reported 19 AI Factories and 13 Antennas offering free customised support to SMEs and start-ups, but that is not evidence of AIGF production access, industrial integration or revenue impact. 6

The practical conclusion is disciplined optimism: monitor awards and later operator terms, use an AI Factory for bounded discovery, and keep source/version control, deterministic checks and human approval between any future AIGF output and a quote or order.

Sources and next reading

AIGF tender notice and submission route — infrastructure-provider application, not an end-user sign-up.

EuroHPC AI Factories and Antennas — adjacent SME/start-up support programme, not an AIGF access portal.

Explore OfferSignal · Focused Business & AI Workshop · All research

References