Power & datacenter draw
The energy constraint at the base of the stack.
| name | metric | value | region | as of | src |
|---|---|---|---|---|---|
| GB200 NVL72 rack power | power draw | ~120 kW/rack | – | – | ↗ T1 |
| Data-center electricity (IEA) | share of global electricity | ~1.5% and rising with AI | Global | – | ↗ source |
| Frontier training compute | growth | ~4-5x/year (Epoch AI) | – | – | ↗ T1 |
| Hyperscaler AI capex | trend | record 2025-26, hundreds of $B/yr combined | US | – | ↗ T1 |
| Amazon–Talen Susquehanna nuclear PPA | contracted nuclear capacity (PPA) | 1,920 MW through 2042 (full volume no later than 2032; front-of-meter transition Spring 2026 with Susquehanna refuel outage) | Pennsylvania, US (PJM) | 2026-07-10 | ↗ T1 |
| Microsoft–Constellation Crane (TMI-1) PPA | contracted nuclear capacity (20-yr PPA) | 835 MW, restart 2028; ~7M MWh/yr | Pennsylvania, US (PJM) | 2026-07-10 | ↗ T1 |
| Meta nuclear offtake (Oklo, Vistra, TerraPower) | contracted/optioned nuclear capacity | up to 6.6 GW by 2035 | US (PJM) | 2026-07-10 | ↗ T2 |
How much electricity does AI use?
Data centres as a whole account for a small but fast-growing share of world electricity demand, and AI is the part growing fastest. The International Energy Agency tracks the trend and the projections in detail (IEA).
Why is power a problem for AI data centres?
A site can only draw what its grid connection allows, and new large connections take years to permit and build. Chips can be delivered far faster than power can be connected. For many announced projects the grid, not the silicon, is what sets the opening date.
How much power does an AI data centre use?
The largest new sites are planned in the hundreds of megawatts, and some campuses are sized above a gigawatt. Older facilities were typically tens of megawatts. A gigawatt is roughly the output of a large power station.
How much electricity does a single AI query use?
Far less than the data centre figures suggest: one short response is a small fraction of a watt-hour. The totals get large because of volume. Agents make it worse, because one task can mean many model calls rather than one.
What is PUE in a data centre?
Power usage effectiveness: total facility power divided by the power that actually reaches the computers. A PUE of 1.0 would mean no overhead at all; everything above it is cooling, power conversion and the rest. Large modern facilities run close to 1.1 to 1.2.