AI Data Centres: 5,500 Megawatts Announced — and the Question of How Much Stays in Germany

No property segment has seen demand dynamics over the past two years comparable to data centres. Bitkom's figures paint a picture whose scale sits outside conventional commercial real estate measures: roughly EUR 12 billion flowed into German data centres in 2025, 70 percent more capacity is planned by 2030, and AI data centres are expected to quadruple — from 530 megawatts to 2,020 megawatts of connected load (Bitkom). The 3,000 megawatt mark for total connected load was crossed for the first time in early 2026; 5,000 megawatts are expected by 2030.

In parallel, more than 70 announced large-scale projects with a combined volume of nearly 5,500 megawatts are on the table. That figure is the decisive point — not because it is high, but because it does not reconcile with the 5,000 megawatt capacity expectation for 2030. Announced projects substantially exceed expected delivery. The difference is the actual investment question.

The gap between announcement and grid connection

An announced data centre project is not a financed project, and a financed project is not a connected project. Three filters sit between letter of intent and productive operation, and each one alone can end a scheme.

The first is the grid connection. Connected loads in the triple-digit megawatt range are not a standard request; the network operator's timing commitment is frequently the binding path in the project schedule and runs into high single-digit years in critical regions. Securing land for a 100 megawatt project today means securing an option on a connection that may only become available at the end of the decade.

The second is the economics of the power price. In the second half of 2025, the commercial industrial power price in Germany stood at roughly 23 cents per kilowatt hour, against about 7 cents in Finland and just under 10 cents in Sweden (Bitkom). For an AI training cluster whose operating costs consist substantially of electricity, a differential of that magnitude is not a site detail but a business model difference. For latency-insensitive training loads — and training is typically latency-insensitive — there is no compelling reason to choose Germany.

The third is capital structure. KPMG addresses the tension between billion-euro investment and capital outflow explicitly: a substantial share of the value creation and ownership of German data centres sits outside Germany (KPMG). For property investors this means competing for land and grid connections against capital providers with fundamentally different return expectations and time horizons.

What latency means for site geography

The distinction between training and inference workloads is the most important geographic logic in this segment, and it is regularly skipped in property discussion.

Training workloads are compute-intensive, run for weeks and are largely insensitive to latency. They migrate to wherever power is cheap and cooling is efficient — Scandinavia, Spain, and increasingly locations outside Europe. Germany does not win that competition at 23 cents.

Inference workloads are the opposite. They serve end users in real time and are latency-critical, which enforces physical proximity to the user. Data residency requirements from GDPR, sector-specific regulation and increasingly the AI Act add to this. These workloads must sit in Germany, regardless of the power price.

This is precisely why the market is increasingly growing away from the established hotspots (Haufe). The Frankfurt concentration remains relevant for connectivity and peering, but inference capacity is dispersing — closer to user clusters, to available grid connections and to sites with waste heat offtakers.

For investors this is the more defensible thesis: Germany is structurally an inference market, not a training market. Underwriting built on hyperscaler training demand is a bet against the power price. Underwriting built on regulatorily bound inference and colocation demand is a bet on data residency — a considerably more robust foundation.

The valuation problem: an asset class without comparables

Transparency remains the unsolved problem of the asset class. Reliable site and transaction data barely exist (Haufe). For valuation practice this means comparable-based methods do not work, and income approaches face the question of how to value an asset whose technical fit-out ages faster than its structure.

In practice this leads to a separation that must be modelled cleanly in underwriting. The shell — land, structure and above all the grid connection — has a long useful life and a value that tends to rise as connection capacity becomes scarce. The technical fit-out — cooling, UPS, racks, power distribution — has useful lives of eight to fifteen years and may be superseded earlier when hardware generations turn over. AI clusters with power densities far above classic colocation levels require liquid cooling that cannot be retrofitted into older assets.

The residual value at the end of a 15-year lease is therefore predominantly the value of the grid connection and the land, not of the building. Underwriting models applying a property-typical residual assumption systematically overstate the exit.

Operations, waste heat and financing logic

The waste heat question increasingly determines consentability. Data centres generate substantial volumes of low-temperature heat, whose use is now regulated as binding for new facilities. The practical hurdle sits on the offtake side: injection only works where a heat network with a matching temperature level exists or an industrial offtaker sits immediately adjacent. Few available sites meet that condition. For site selection this means the availability of a heat offtaker carries weight approaching that of connected load — sites offering both are the actual scarcity in this market.

Tenant structure is equally relevant. A colocation asset with a diversified base of mid-market and public sector customers carries a fundamentally different risk profile from an asset with a single hyperscaler tenant. The single-tenant case delivers covenant strength and long lease term but concentrates the entire residual value risk on a renewal decision driven by the capacity planning of a global group — and that planning is not made at the German site. The multi-tenant model offers lower individual covenant strength but a re-letting capability that actually exists, and that is what makes the difference in an exit scenario.

Finally, financing. Data centres are increasingly funded by specialist lenders and infrastructure funds whose terms follow infrastructure rather than property logic: longer tenors, different covenants, different return expectations. For property investors this is a cost-of-capital disadvantage for as long as they structure the asset as real estate. Competing seriously in this segment means adapting the financing structure to infrastructure capital's expectations — and accepting that operational risk moves closer to the owner than classic property mandates contemplate.

What this means for principals: the 5,500 megawatts of announced projects are a demand signal, not a capacity forecast — the realistic magnitude is the expected 5,000 megawatts of total load by 2030. Germany does not win training workloads at 23 cents per kilowatt hour; the viable thesis is regulatorily bound inference and colocation demand away from the hotspots. In underwriting, the grid connection and land belong modelled separately from technical fit-out as the actual value driver — residual value after 15 years sits in connection capacity, not in the building.