The International Energy Agency (IEA) has released its latest forecast: by the end of 2026, global data center electricity consumption will exceed 1,000 TWh — equivalent to Japan's entire annual electricity usage.
For context, that figure stood at 460 TWh in 2022, meaning consumption has more than doubled in less than four years.
Where is the growth coming from?
The vast majority comes from AI.
Specifically, two types of workloads: AI model training (which fewer players are doing) and AI inference (which an increasing number of users are running).
A January 2026 report from Bloom Energy estimated that U.S. data center energy demand would surge from 80 GW in 2025 to 150 GW in 2028 — nearly doubling in three years, almost entirely driven by AI.
The U.S. Energy Information Administration takes a longer view, projecting that data centers could account for 12% of total U.S. electricity consumption by 2030, up from 4% today.
To put the scale in perspective: a single large AI training facility requires between 100 MW and 1,000 MW of dedicated power. The highest-end facility consumes as much electricity as 800,000 typical homes.
The grid is buckling under the strain
This is not just an abstract energy statistic — it is already affecting real people.
A March analysis by Consumer Reports found that in data-center-dense regions such as Virginia, Texas, and Georgia, residential electricity bills have risen by 8% to 15%.
This raises a politically charged question: who should pay for AI's electricity consumption?
Data center operators typically secure discounted bulk power rates, while residential rates continue to climb. The fairness of this pricing structure is now being debated in those states.
However, a new study released yesterday (April 9) found no direct evidence of a causal link between AI data centers and rising residential electricity bills — the debate remains unresolved.
The problem of power supply
With such a massive increase in demand, where will the electricity come from?
The major tech companies are pursuing different strategies:
- Amazon, Google, and Microsoft are signing large numbers of power purchase agreements (PPAs) to lock in wind and solar energy.
- Microsoft has reclaimed some power from the Three Mile Island nuclear plant.
- Google is exploring small modular nuclear reactors (SMRs).
- Some data centers are building nearby gas turbine units.
Nuclear power is regaining favor in the AI industry because of its stable baseload output (it is not dependent on weather). However, nuclear plant construction cycles are too long to fill the gap in the short term.
Will this become a hard constraint on AI development?
On the algorithmic side, model efficiency is indeed improving — Kimi, DeepSeek, and Meta Muse Spark all claim to achieve equivalent results with less compute. But total electricity demand is a function of volume: even if the cost per inference drops, the number of users and the frequency of calls are growing faster.
The real constraint may not be compute power, but electricity.
Some analysts have begun describing AI expansion's ceiling in terms of energy constraints rather than compute constraints. In certain regions, securing construction permits and finding a stable, high-capacity power source is harder than acquiring GPUs.
Sources: Energy demand from AI (IEA); CocoLoop, The AI Data Center Power Crisis: How Big Tech's 125 GW Problem Threatens the Grid (tech-insider.org); AI Data Center Energy Consumption Projections 2026 (ZestLab)