FEAR AI.
Home / Data centers

RESEARCH / ELECTRICITY & ENVIRONMENT

The cloud has
a footprint.

Electricity is physical. “AI demand” is a definition.
Start by separating the two.

FIRST, THE DENOMINATOR.

All computing
isn’t all AI.

Data centers also serve storage, business software and other computing. A company-wide electricity total is another boundary again. These categories can overlap; they cannot simply be added together.

The IEA’s April 2026 Base Case separates all data-center electricity from AI-focused facilities. Both figures below are 2030 projections, not current meter readings. IEA report, Figure 1.5, pp.25–26 ↗

950TWh

Worldwide electricity demand from all data centers in 2030.

IEA Base Case · Projection · April 2026

A subset, not an additional 465 TWh.
Training and inference are not separately allocated here.

THREE DISTINCTIONS

Read the small print.

Power ≠ energy.

A proposed gigawatt connection measures capacity. Annual terawatt-hours measure energy. Turning one into the other needs assumptions about use, ramp-up and timing.

A promise ≠ a plant.

A power agreement can cover existing generation, a proposed project or environmental attributes. Its nameplate capacity is not proof of an operating facility.

A share ≠ an impact.

A national electricity share cannot settle local questions about water, air quality, noise, bills or infrastructure. Place and time belong beside the annual total.

WHAT WE WANT TO FOLLOW

The company.
The place. The date.

Follow a number from a company disclosure to the relevant region or site. Keep actual consumption, requested grid service and proposed generation separate. Where an AI-only allocation is not disclosed, the honest answer is a visible gap.

Explore sources and notes →

Sources & scope

  1. International Energy Agency, Key Questions on Energy and AI, April 2026, Figure 1.5, printed pp.25–26. The 950 and 465 TWh values use the same Base Case and target year; AI-focused facilities include supporting facility demand. They are not AI-chip-only consumption.
  2. Google, Environmental Report 2026, p.93. A useful example of company reporting boundaries and historical restatements, not an AI-only electricity allocation.
  3. IEA, Energy and AI, 2025. Earlier forecast vintage; do not combine its scenarios with the 2026 update as one unchanged series.

Institutional modeling and corporate disclosures have different authority and incentives. This comparison covers the electricity projections listed here. Water use, emissions, and local impacts require separate measures.