Aller au contenu
AIHyperscalers

AI data center energy demand: where we really stand in 2026

18 March 2026 7 min

GPT, Gemini, Claude and large-scale inference: the raw numbers on AI consumption and how to keep up.

Crossing the terawatt-hour line

The IEA estimates AI data centers will consume 540 TWh in 2026, i.e. 2.1% of global electricity production. By 2030 the figure could double.

Why inference outweighs training

A model like GPT-5 cost ~70 GWh to train. But a single user query consumes ~3 Wh, times 2 billion queries/day: inference dwarfs training by a factor of 50 in cumulative energy.

New 1 GW+ clusters

XAI Memphis (1.2 GW), Microsoft Stargate Wisconsin (1.8 GW planned), Meta Louisiana (2 GW): unit campus size now exceeds a nuclear reactor. Grid connection has become the #1 bottleneck.

The local co-sourcing advantage

Rather than waiting 5–7 years for a grid extension, AI operators now sign PPAs with surplus producers under 50 km away. That is exactly the friction Voltarione removes.

Explore the marketplace

25+ European sites mapped in real time with available surplus.

Open the map

Continue reading