AI Energy Consumption Statistics in 2027: Complete Data & Trends

Quick Answer

• AI-dedicated data center power (2027): 228 TWh (IEA)
• Share of global electricity: 0.8%
• US AI data center power: 118 TWh (Lawrence Berkeley Lab)
• GPT-5 training emissions (est): 18,400 tCO2e (Stanford HAI)
• Hyperscaler nuclear PPAs signed (2024–27): 24 GW

Key AI Energy Statistics

IEA's 2027 Electricity Report projects data center consumption at 945 TWh by 2030, with AI driving 54% of growth. Lawrence Berkeley National Lab estimates US AI data centers consume 118 TWh in 2027 — 2.8% of US electricity. Stanford HAI reports frontier training runs exceed 10 GWh each.

Market Data

Metric	2027 Value	Source
Global DC power	580 TWh	IEA
AI-specific DC power	228 TWh	IEA
US DC power	248 TWh	LBNL
US AI DC power	118 TWh	LBNL
Global share of electricity	0.8% AI	IEA
Frontier model training	10–30 GWh	Stanford HAI
Inference % of AI power	72%	SemiAnalysis
Hyperscaler nuclear PPAs	24 GW	BloombergNEF
Water use (DC cooling)	820B liters	UNEP
PUE average hyperscale	1.15	Uptime Institute

Top Hyperscaler Nuclear Deals

Company	Deal	Capacity
Microsoft	Three Mile Island restart	835 MW
Amazon	Talen Energy (Susquehanna)	1920 MW
Google	Kairos SMR	500 MW (multi-site)
Meta	1–4 GW SMR RFP	Multiple
Oracle	1 GW SMR plan	Texas

Training Emissions Estimates

Model	Training TWh	tCO2e
GPT-5	0.048	18,400
Claude 4.6	0.031	11,200
Gemini 3 Ultra	0.042	14,800
Llama 4	0.024	8,400

Sources

1. IEA — Electricity 2027 Report
2. Lawrence Berkeley National Lab — US DC Energy 2027
3. Stanford HAI — AI Index 2027 (Compute chapter)
4. SemiAnalysis — AI Power Consumption 2027
5. BloombergNEF — Nuclear + AI 2027
6. UNEP — Data Center Water 2027
7. Uptime Institute — Global DC Survey 2027
8. Microsoft / Amazon / Google / Meta — Sustainability Reports 2027
9. Goldman Sachs — AI Power Demand 2027
10. Epoch AI — Frontier Model Compute Tracker 2027

FAQs

How much power do AI data centers use in 2027? 228 TWh globally per IEA.

What % of US electricity? 2.8% from US AI data centers (LBNL).

How much does training a frontier model consume? 10–30 GWh per frontier run (Stanford HAI).

Are hyperscalers going nuclear? Yes — 24 GW of PPAs signed 2024–27.

Is inference or training bigger energy use? Inference at 72% of AI power (SemiAnalysis).

Conclusion

AI's energy footprint in 2027 is substantial and growing. More at misar.blog.