EU Sovereign-Compute Programs Will Announce 15+ GW by Q4 2027 But Energize Under 25% of It

The claim

By the end of Q4 2027, EU-linked sovereign-compute and AI-data-center programs announced during 2026 and 2027, taken collectively, will have committed to at least 15 GW of AI data-center capacity, while less than 25 percent of that announced capacity will be physically energized (grid-connected and operational) by December 31, 2027.

This is the announced-versus-energized gap stated as a falsifiable, dated prediction. It is specific in three ways: a floor on announced capacity (15 GW or more), a ceiling on energized capacity (under 25 percent of announced, i.e., under roughly 3.75 GW if the floor is exactly met), and a hard date (December 31, 2027).

Why I believe it

The announcement side is nearly certain to clear 15 GW. SoftBank alone committed up to 5 GW in France (3.1 GW firm by 2031). The EuroHPC AI-gigafactory program, expanded by Council regulation in January 2026, is provisioning facilities of roughly 100,000 processors each across multiple member states. Mistral plans 200 MW of self-build by end of 2027. Add the German JAIF federation, the French Alice Recoque exascale system, AMD's France deal, and the wave of national champions racing to plant flags, and 15 GW of cumulative announcements through 2027 is a low bar. Europe is in an announcement arms race, and announcements are cheap.

The energization side is where the constraint bites. Standard Western European grid-interconnection queues imply seven-to-ten-year timelines. France's CRE fast-track (deliberation 2025-120) compresses designated-site connections to three or four years, but that clock starts in 2025-2026 for the earliest projects, which puts energization in 2028-2030, not 2027. SoftBank's own firm target is 2031 for 3.1 GW. EuroHPC gigafactories are still in procurement. Mistral's 200 MW is the most likely near-term win but is a rounding error against 15 GW. The physics of commissioning power blocks, building substations, and stringing 400 kV lines simply does not deliver multi-gigawatt energization inside an 18-month window from a 2026 standing start.

So the structure of the prediction follows directly from the structure of the industry: announcements scale on a capital clock (fast), energization scales on a grid clock (slow), and the gap between them is large and persistent.

What would make this true

• Cumulative announced EU sovereign-compute capacity reaches or exceeds 15 GW (SoftBank France's 5 GW plus EuroHPC gigafactories plus member-state and private commitments).
• As of December 31, 2027, the sum of energized, grid-connected, operational capacity from these announced programs is under 25 percent of the announced total.
• Energization is measured as capacity actually drawing grid power for AI workloads, not capacity under construction, permitted, or contracted.

What would make this false

• A surprise wave of fast energization: if France's fast-track and pre-existing brownfield sites deliver more than 25 percent of announced capacity online by end of 2027, the prediction fails. This would require multiple gigawatt-class campuses energizing two to three years ahead of their stated 2030-2031 targets.
• Announcements fall short of 15 GW: if the EU announcement pace collapses (a macro shock, a capital pullback, or a political retreat from sovereignty), the floor is not met and the prediction is void on its first clause.
• Definitional drift: if "energized" is loosened to include partial commissioning or behind-the-meter pilots, the 25 percent ceiling could be breached on a technicality. I am holding to grid-connected, operational capacity.

Confidence reasoning

I set confidence at 64, tier 2. The announcement floor is close to certain; the real uncertainty is whether energization stays under 25 percent. The downside risk to the prediction is that I am underestimating how much existing French capacity (brownfield sites, the AI Factory France national assets, Mistral's near-term build) gets counted and pushes the energized share above 25 percent. The upside is that multi-gigawatt builds almost never beat their grid timelines, and 2027 is too soon for the 2030-2031 cohort. On balance, the gap holds, but the margin is narrow enough to keep this out of tier 1.

This prediction operationalizes the central thesis of the accompanying analysis: in AI infrastructure, gigawatts-announced and gigawatts-energized are different metrics, and the difference is the story of the era.