Future-Proofing EU AI Gigafactories: Four Design Imperatives

EU AI Gigafactories: Between ambition and unresolved challenges

The EU has committed to building up to five AI Gigafactories (AIGFs), its largest planned compute infrastructure investment to date. The initiative reflects the EU’s recognition of frontier AI’s growing importance for the economy and national security. Each AIGF will host roughly 100,000 H100-equivalent chips, designed to meet frontier AI’s growing compute demands.

Yet, the EU AI Gigafactories risk falling short of the Commission’s ambitions, as several unresolved challenges threaten their success. For instance, by the time the first AIGFs come online in 2027–2028, the largest foreign data centres could already have roughly 50 times more compute capacity than each AIGF. By 2030, that gap could even reach 200 times.

In our new memo, ”Future-Proofing EU AI Gigafactories: Four Design Imperatives”, we identify four design choices that increase the likelihood that AIGFs will deliver on Europe’s frontier AI ambitions, and set concrete recommendations for each.

Read the full memo for a deeper dive into these design imperatives.

Four design imperatives for EU AI Gigafactories

Imperative 1: Hardware fit for purpose

AIGF hardware must remain useful as compute demand rises and chip generations turn over. Frontier AI compute trajectories are uncertain but trending sharply upward, and advanced chip supply is tightly constrained. If the EU aims for frontier AI competitiveness, AIGFs should be designed to meet three criteria.

1. Future-proofing: Procure state-of-the-art equipment (e.g., Blackwell-class chips) with a credible upgrade pathway for integrating next-generation hardware.
2. Expandability: Design sites with contingency plans to scale rapidly to 500,000–2 million H100-equivalents if compute demand continues to grow.
3. Supply assurance: Secure written chip supplier commitments specifying quantities and delivery timelines.

Imperative 2: A tiered security posture, with at least one high-security EU AI Gigafactory

AIGFs should adopt a tiered cybersecurity posture scaled to workload sensitivity. Low- to mid-sensitivity use cases can rely on RAND SL2-3-grade defences, appropriate for most commercial and research threats. High-sensitivity use cases, however, require SL4-5-grade defences designed to withstand operations by leading cyber-capable institutions. At least one AIGF should be solely dedicated to this higher standard, with hardware-anchored cryptographic verification that generates evidence of key properties like execution environment, data encryption, and model and software integrity.

A tiered security posture would provide the EU with several advantages. For example, beyond hardening AIGFs against severe threats for sensitive workloads, it would help the EU meet the security standards that frontier AI providers and governments increasingly attach to access as capabilities advance, as illustrated by the U.S. AI diffusion framework and Anthropic’s gated release of Claude Mythos Preview through Project Glasswing.

Imperative 3: Interoperability by design

Interoperability across three vectors is essential for a competitive EU AI ecosystem. Between AIGFs, it could enable decentralised training across sites as compute demand surges, for instance via a high-capacity fibre loop; with AI Factories, it would allow users to scale validated workloads to AIGF level without repurposing; and with private EU clouds, it supports multi-cloud strategies, stronger price negotiation, and improved resilience.

Two technical priorities underpin all three: chips natively supporting multiple precision formats (e.g., BF16/FP16/FP8 and emerging lower formats), and topology-agnostic networking and interconnect stacks built on open standards such as Ultra Ethernet and UALink rather than proprietary vendor architectures.

Imperative 4: A flexible operating model for EU AI Gigafactories

EU AI Gigafactories must implement a flexible operating model to hedge against different demand scenarios. While future compute demand is likely to grow, its trajectory remains uncertain, and anchor tenants may not materialise at the scale envisaged.

AIGFs should therefore be able to shift capacity across the full model lifecycle (training, fine-tuning, and inference) as demand evolves; operate as multi-client service platforms if needed; and monetise excess capacity through discounted, interruptible spot-instance models, public-sector procurement, or a designated geostrategic reserve. 

Throughout, European users must retain priority access through reserved capacity and curtailable foreign access, so that commercial viability does not come at the cost of European AI capability.

Four imperatives, one goal

The EU has committed to pursuing frontier AI leadership, and the AI Gigafactory initiative is to date its largest compute investment to deliver on that commitment. Yet EU AI Gigafactories must be purpose-built to match that ambition.

The four imperatives above will not on their own fully close the gap with leading foreign buildouts, but they are the design principles required to ensure that the AIGF investment translates into European compute that is competitive, sovereign, and safe, while meaningfully advancing Europe’s frontier AI vision.

Read the full memo for our complete analysis and recommendations.

For more insights, research, and expert analysis, follow The Future Society on Linkedin.