INTELLIGENCE
THE ENERGY COST OF INTELLIGENCE
FRONTIER TRAINING COMPUTE
GPT-6 Astra
OpenAI · 2026-09-03
1.0 × 1027FLOP
EPOCH CONFIDENCE · LIKELY
Training power draw ≈ 233 MW · Epoch AI estimate
Training compute values are Epoch AI estimates; each carries Epoch's confidence label.
≈5× per year since 2020 · Training compute for frontier language models has been growing at 5× per year since 2020
GPT-6 Astra
OpenAI · 2026-09-03 · Likely
1.0 × 1027 FLOP
Grok 4
xAI · 2025-07-09 · Speculative
5.0 × 1026 FLOP
GPT-4.5
OpenAI · 2025-02-27 · Likely
3.8 × 1026 FLOP
Grok 3
xAI · 2025-02-17 · Likely
3.5 × 1026 FLOP
Llama 4 Behemoth (preview)
Meta AI · 2025-04-05 · Likely
5.18 × 1025 FLOP
COMPUTE PER WATT
KAIROS
CALMIP / University of Toulouse - CNRS
73.282GFlops/W
×22.8 since 2013-06 · 13 years · CAGR 27.2%DERIVED
#1 system of each Green500 list (energy efficiency measured on HPL). GFlops per watt; 2013 lists reported MFlops/W and are divided by 1000.
DATA CENTRES · SHARE OF WORLD ELECTRICITY
- 2024 · 415 TWh · 1.5%MEASURED
around 1.5% of global electricity consumption
- 2025 · 485 TWhMEASURED
up 17% on 2024
- 2030 · 950 TWh · 3%PROJECTION
IEA projection: roughly doubling to 950 TWh, around 3% of global electricity demand
2025 share 1.56%ESTIMATE
IEA 2025 data-centre consumption (485 TWh) ÷ Ember 2025 world electricity generation (31,006.6 TWh). Consumption vs generation, so treat as approximate.
Data-centre electricity consumption as reported by the IEA. The 2025 share of world electricity is derived by Observatory from IEA consumption and Ember world generation, and is labelled as an estimate.
INTELLIGENCE IS BOUNDED BY ENERGY
Useful computation is a physical process, and it costs energy.
Frontier training compute has grown about 5× per year since 2020 (Epoch), while the best FLOPs per watt on Green500 rose ×22.8 over 13 years, about 27% a year. Efficiency is moving much slower than the training runs, so the energy those runs draw keeps rising. The bottleneck is power.
At the floor, erasing one bit costs at least kT·ln2 — Landauer's limit, about 2.9×10-21 J at room temperature. Today's digital chips operate many orders of magnitude above that floor. The gap is why some researchers build hardware that computes with physics itself, thermodynamic and probabilistic computing, rather than spending energy to fight noise. Experimental verification of Landauer's principle (Nature, 2012).