RESEARCH

Global AI Data Center Power Capacity for 2025 Splits Between 8 GW and 44 GW Across Trackers

Four major trackers estimate global AI-dedicated data center power capacity for 2025 anywhere from about 8 GW to 44 GW -- a more than fivefold spread. Epoch AI (~30 GW) and RAND (~21 GW) build bottom-up from chip shipments and compute-to-power ratios; McKinsey (44 GW) models AI workload demand as a share of its own 82 GW total data center estimate; Goldman Sachs (~7.7 GW) derives its figure by applying a 14% 'AI share' to a smaller 55 GW total data center baseline. All four agree AI is the fastest-growing driver of global power demand -- they disagree by more than 5x on how much AI-specific capacity exists today, because each draws the boundary between 'AI' and 'general' data center capacity differently.

Findings

  • Epoch AI estimates global AI-dedicated data center power capacity reached approximately 30 GW by Q4 2025, built bottom-up from the rated power of AI chips sold, with a roughly 2.5x multiplier applied for servers, cooling, and infrastructure overhead.
  • McKinsey's 2025 analysis puts AI-specific data center demand at 44 GW in 2025 -- about 54% of its own 82 GW total data center demand estimate for the year -- and projects it will grow 3.5x to roughly 155 GW by 2030.
  • RAND's power-requirements report estimates AI data center capacity at roughly 21 GW in 2025, more than fourfold its 2023 baseline, using a compute-to-power conversion model distinct from both Epoch AI's chip-sales method and McKinsey's demand-side modeling.
  • Goldman Sachs frames the same year differently: it puts total global data center power usage at 55 GW with AI accounting for only 14% of that -- about 7.7 GW -- because its baseline splits capacity three ways (AI, cloud computing at 54%, traditional workloads at 32%) instead of isolating AI infrastructure end-to-end.

Delta Engine result

↔ Divergence Detected — Δ 0.4837 (threshold 0.05)

Evidence quality

avg 25 · min 25 · max 25 · spread 0