Datacenter Power Demand

The demand for datacenter power is the central thesis of the energy investment cycle. AI training and inference require dense, reliable, always-on electricity at scales that are straining the grid. Unlike prior waves of datacenter growth (cloud storage, streaming), AI workloads are GPU-dense and cannot tolerate interruption — a power event during a multi-week training run destroys enormous value. This makes reliability, not just price, the primary procurement criterion, and pushes hyperscalers toward nuclear and gas over intermittent renewables.

The numbers are staggering and frequently revised upward. Global datacenter electricity demand reached 132 GW in 2026 (up 27% from 104 GW in 2025), with US datacenters alone consuming 41 GW. Goldman Sachs projects US datacenter demand doubles to 66 GW by 2027. The IEA projects 1,000+ TWh annual consumption under high-growth scenarios — roughly Japan's entire electricity grid. AI-optimized servers (GPU clusters) now account for 31% of datacenter power consumption. Gartner forecasts 26.4% global demand surge in 2026.

The demand is geographically concentrated. US, Ireland, Singapore, and Northern Virginia (Loudoun County alone) represent disproportionate shares. Virginia's "Data Center Alley" drew 3,700 MW of new load in 2023. Phoenix, Chicago, and Atlanta are the next-tier markets.

Key Claims

• Global datacenter electricity demand: 132 GW in 2026, up from 104 GW in 2025. Evidence: strong (Gartner, IEA, Global Electricity)
• US datacenters: 41 GW in 2026 → 66 GW in 2027. Evidence: strong (Goldman Sachs)
• AI-optimized servers: 31% of datacenter power in 2026. Evidence: moderate (Gartner)
• IEA high-growth scenario: 1,000+ TWh by 2026 — comparable to Japan's entire grid. Evidence: moderate (IEA Energy and AI report)
• Hyperscalers (Meta, Amazon, Google, Microsoft) represent 76% of new corporate energy contracts. Evidence: strong (South Pole PPA Buyers Guide 2026)
• AI datacenter load is dispatchable-power-seeking — reliability requirements push toward nuclear and gas over intermittent renewables. Evidence: moderate (inference from deal structure)
• A single AI supercluster can drive multi-GW procurement: Meta contracted up to 6.6 GW of nuclear by 2035 for one site (the "Prometheus" supercluster in New Albany, Ohio). Latitude Media notes the scale approximates northern Virginia's entire "Data Center Alley." Evidence: weak (single news item, Latitude Media)
• Power is the gating constraint on AI buildout — hyperscalers procure nuclear specifically to secure firm, carbon-free baseload outside congested grid-interconnect queues; demand from this single buyer class is now underwriting the SMR/advanced-reactor commercialization curve. Evidence: moderate (multi-source synthesis, hyperscaler-nuclear-procurement)

Benchmarks & Data

• Global datacenter demand: 104 GW (2025) → 132 GW (2026) → 290 GW (2030 projection)
• US datacenter demand: 31 GW (2025) → 41 GW (2026) → 66 GW (2027)
• AI-server share of datacenter power: 31% (2026)
• Global datacenter energy consumption: ~500 TWh (2026 base) to 1,000+ TWh (IEA high scenario)
• Corporate renewable PPA volume: 46 GW new contracts globally in 2024 (record); hyperscalers = 76% of that

Open Questions

• How much of the demand is inference vs training? (Inference is growing faster but slightly more tolerant of intermittency)
• Will efficiency gains (Blackwell vs Hopper compute-per-watt improvement) offset volume growth?
• Does the US AI executive order (if any) mandate domestic nuclear procurement?
• At what load density does air cooling become insufficient and liquid cooling mandatory? (Affects facility design and power draw)

Related Concepts

• Power Purchase Agreements — how hyperscalers secure the power
• Baseload vs Intermittent — why AI load prefers nuclear/gas
• Grid Interconnect Queue — why new supply is constrained

Changelog

• 2026-06-15 — Initial compilation from web research (IEA, Goldman Sachs, Gartner, S&P Global)
• 2026-06-24 — Compiled new sources (meta-66gw-nuclear-deal, hyperscaler-nuclear-procurement)