Global AI power consumption reaches 38GW

What AI 2027 Predicted

The scenario’s “KEY METRICS 2026” sidebar lists global AI power consumption reaching 38 gigawatts of peak power by end of 2026. This reflects the massive datacenter buildout depicted in the scenario, where infrastructure investment pushes AI compute capacity to unprecedented levels.

How We Track This

We monitor:

• IEA datacenter electricity demand reports and projections
• Goldman Sachs, SemiAnalysis, and other analyst estimates of datacenter power draw
• Individual datacenter project tracking (capacity announcements, construction timelines)
• Distinction between total datacenter power and AI-specific power consumption
• DOE and EIA reports on US electricity demand from datacenters

Current Evidence

Estimating global AI-specific power consumption is challenging because most reporting covers total datacenter power (which includes non-AI workloads like cloud computing, streaming, and traditional enterprise IT).

Total datacenter power (global):

• The IEA projects global datacenter electricity consumption to rise by approximately 240 TWh above 2024 levels through 2030, reaching roughly 945 TWh by decade’s end. In 2024, global datacenter consumption was approximately 400-450 TWh, equivalent to roughly 46-51 GW average load (IEA Electricity Mid-Year Update 2025).
• Goldman Sachs forecasts a 50% increase in global datacenter power demand by 2027 and 165% by 2030, relative to 2023 levels (Goldman Sachs Research).
• A CTVC database tracked 294 datacenter projects totaling 73.6 GW of planned demand as of mid-2025 — though only a fraction is operational.

AI-specific power:

• AI workloads are estimated at 30-40% of total datacenter power today, growing to 50%+ by 2027
• If global datacenter power reaches ~55-60 GW total by end of 2026 (Goldman Sachs trajectory), AI-specific consumption could be 20-25 GW — well short of 38 GW
• Reaching 38 GW of AI-specific power would require either much faster datacenter buildout than projected or AI workloads consuming a much higher share than expected

Stargate Project Pipeline: OpenAI’s Stargate project has expanded to approximately 10GW of planned capacity across multiple sites: Abilene TX (1.2GW, partially operational Sep 2025), five additional sites (~7GW announced Sep 2025), Michigan (1.4GW approved Dec 2025), plus a 4.5GW Oracle partnership (Jul 2025). This single project’s planned capacity exceeds the 38GW prediction’s share for any one company, though planned capacity ≠ operational capacity — most sites are under construction with multi-year build timelines.

US datacenter power:

• Pew Research reports US datacenters accounted for 4% of total US electricity in 2024 (Pew Research)
• The DOE projects US datacenter energy demand to double or triple by 2028 (DOE)

Sources:

• IEA Electricity Mid-Year Update 2025
• Goldman Sachs: AI to drive 165% datacenter power demand increase
• Pew Research: US datacenter energy
• DOE datacenter energy report

Counterevidence & Limitations

• 38 GW of AI-specific power is an aggressive target; most projections put total global datacenter power (all workloads) at 55-65 GW by end of 2026
• Datacenter construction timelines often slip — grid interconnection, permitting, and power procurement create bottlenecks
• Efficiency improvements (better chips, improved cooling, inference optimization) partially offset raw power demand growth
• The distinction between “AI power” and “total datacenter power” is blurry — the scenario may be using a broader definition
• If 38 GW means total datacenter power allocated to AI-capable facilities (not just AI workload power draw), the number becomes more plausible

What Would Change Our Assessment

• Upgrade to “on-track”: Credible estimates of 30+ GW of AI-specific datacenter power operational by mid-2026
• Downgrade to “behind”: If AI-specific power remains below 20 GW through end of 2026
• Reinterpretation: If 38 GW refers to total datacenter capacity (not AI-specific), the target becomes more achievable

Update History

Date	Update
2025-10	Anthropic’s 1M TPU deal (October 23) targets “more than 1 gigawatt” of capacity online in 2026. Q3 Big Tech capex at $114B/quarter with power availability cited as primary constraint — demand outpacing electrical infrastructure build-out.
2026-03	Industry estimates place AI-specific power consumption at 20-25 GW globally (end 2025). The 38 GW target for 2026 appears aggressive but datacenter buildout is accelerating rapidly.
2026-03-16	Programs.com (Mar 2026): AI workloads projected at 44 GW for 2026, non-AI at 38 GW — total 82 GW. If accurate, the 38 GW AI-specific target may actually be exceeded. Reuters reports largest US sites now consume >1 GW continuous. IEA projects >1,000 TWh datacenter consumption by 2026 under high-growth scenarios. Confidence adjusted 0.45 → 0.50 — newer estimates suggest the target is more achievable than previously assessed, though estimates vary widely.
2026-03-23	The Atlantic (Apr 2026 issue) reports OpenAI alone has announced plans for facilities requiring 30+ GW total, more than New England’s largest recorded demand (The Atlantic). White House event features xAI committing to 1.2 GW of dedicated power per supercomputer site (White House). UPI reports 10 largest US AI datacenter projects alone require 5.26 GW. Confidence adjusted 0.50 → 0.55 — planned capacity increasingly supports the 38 GW target, but gap between announced and operational capacity remains large.