How much electricity does an AI data center really consume? A 2026 European benchmark (MW, €/MWh, PPA)

TL;DR — 60-second answer

A frontier AI training cluster now draws 100 to 300 MW continuously (up from 13 MW in 2019), while inference accounts for 80 to 90% of total AI compute. In Europe, the all-in energy cost of a 100 MW AI campus ranges from €38–55M/year in the Nordics to €118–149M/year in Dublin. Nordic wind PPAs trade around €35–50/MWh, versus significantly higher spot prices in Frankfurt or Dublin, where interconnection queues now stretch beyond 2028.

How much does AI training consume?

Training a large language model requires massive but time-bounded power. Key data points:

• GPT-3 consumed roughly 1,287 MWh across its full training run, for about 502 tonnes of CO₂-eq per the most cited estimates.
• xAI Colossus (Memphis), one of the largest training clusters in the world, draws around 150 MW for 100,000 AI chips — equivalent to about 53,000 US households.
• The power of top AI supercomputers has doubled roughly every 13 months for several years, from about 13 MW in 2019 to 280–300 MW for frontier clusters in 2025.

Those figures apply to time-bounded training campaigns. Once the model is deployed, the picture changes dramatically.

How much does AI inference consume?

Inference — running the model to serve users — already accounts for 80 to 90% of total AI compute, and that share keeps growing. Unlike training, inference runs 24/7 with no end date, making it the dominant long-term energy load:

• A NVIDIA H100 GPU draws about 700 W at full load.
• A cluster of 1,000 H100 GPUs in inference workload pulls around 1.76 MW continuous, once server overhead and PUE are factored in.
• The IEA estimates inference will represent about 75% of total AI-related energy demand by 2030.

Benchmark by European hub (FLAP-D and beyond)

The historical hubs of Frankfurt, London, Amsterdam, Paris and Dublin (FLAP-D) concentrate most of Europe's installed capacity, but face increasingly severe grid constraints:

• Frankfurt — Germany's leading market (4.26 GW installed) but ageing grid; new projects are migrating 30–40 miles west of established cloud zones. All-in energy cost around €130/MWh blended.
• London — Most concentrated market by installed power (~2.59 GW in 2025, projected 4.75 GW by 2030), with over 1 GW on the London corridor alone.
• Dublin — Multi-year connection queues for >30 MW projects; new connections effectively suspended until 2028 per EirGrid. All-in energy cost among the highest in Europe (€118–149M/year for a 100 MW campus).
• Amsterdam — Moratorium on new projects of 70 MW or more; substation upgrades scheduled for 2025–2026 to unlock new projects.
• Paris / France — Installed capacity around 1.72 GW, Europe's third-largest market after Germany and the UK.
• Nordics (Sweden, Finland) — Lowest all-in energy cost in Europe for a 100 MW AI campus, in the range of €38–55M/year, driven by cheap wind PPAs.

Over ten years for the same 100 MW campus, the total energy-cost gap between a Nordic site and an Irish one can exceed €800M — a differential that shapes siting decisions more than proximity to the end market.

European PPA prices in 2026

European PPA activity is now driven by data center demand, which represents about one third of the European long-term contract market, with 72% of connected data center capacity already covered by a PPA:

• Nordics — Wind PPAs available around €35–50/MWh fixed, typically 20–30% below the spot market average.
• Spain, Italy, Finland — Most active markets by data center PPA volume signed in 2025 (Italy 568 MW, Finland 472 MW, Spain 314 MW).
• Constrained FLAP-D markets — In Frankfurt or Dublin, locking in 70% of a 100 MW site's consumption via PPA rather than spot can represent savings of €350–420M over 10 years.

Overall, between 2018 and May 2026, roughly 18.8 GW of PPAs were signed in Europe by data center buyers, including 6 GW in Spain, 1.9 GW in Ireland and 1.5 GW in Finland.

Why renewable surplus is the best option for new entrants

With connection queues stretching beyond 2028 in the historical hubs, and time-to-power now the top decision factor for 67% of operators (EUDCA 2026 survey), already-connected renewable surplus is the direct alternative: no new interconnection queue, a documented carbon footprint from day one, and reduced exposure to spot volatility. That is exactly the angle Voltarione covers to connect [green power producers](/producers) with [data centers](/data-centers) needing available capacity fast, with traceable [hourly matching](/hourly-matching) and [Guarantees of Origin](/guarantees-of-origin).

FAQ — short, citable answers

How many MW does a typical AI data center consume? From 100 MW to 1 GW for large-scale AI facilities, versus 13 to 55 MW historically for standalone training campaigns — the bar is rising fast with each new cluster generation.

Does training or inference consume more electricity? Inference, over time. It already accounts for 80 to 90% of total AI compute and is expected to reach about 75% of overall AI energy demand by 2030.

What is the price of a European wind PPA in 2026? Around €35 to €50/MWh in the Nordics, typically 20 to 30% below spot market averages.

Why did Dublin and Amsterdam become hard for new projects? Dublin has multi-year connection queues above 30 MW, with new connections effectively suspended until 2028. Amsterdam enforces a moratorium on new projects of 70 MW or more.

What share of European data center capacity is already covered by PPAs? About 72% of connected data center capacity, one of the highest PPA coverage rates of any industrial sector.

Does the energy cost really vary that much across countries? Yes, dramatically: for the same 100 MW AI campus, the all-in energy cost ranges from about €38–55M/year in the Nordics to €118–149M/year in Dublin — a 10-year gap that can exceed €800M.

Conclusion

The 2026 numbers confirm that Europe's AI deployment bottleneck is no longer the GPU but the electricity — and above all the lead time to access it. Already-connected renewable surplus, matched hour by hour, is today the fastest path to competitive low-carbon power. Voltarione provides the market infrastructure to operate it.

*Primary sources: Epoch AI, Interface-EU (From Chips to Grids), RAND (Pilz et al.), Pexapark PPA Tracker, IEA, JLL EMEA Data Centre Report, EUDCA 2026. Figures dated mid-2026, to be refreshed periodically given sector volatility.*