Data center energy is usually reported in terawatt-hours consumed, which is covered in the consumption report. This page tracks the other number: gigawatts of power capacity. Capacity, not consumption, is what the grid actually has to supply at peak, and it is the figure that determines whether a project gets built.
Capacity Trajectory
| Year | Global data center power demand | Change |
|---|---|---|
| 2025 | ~104 GW | Baseline |
| 2026 | ~132 GW | +27% |
| 2030 (forecast) | ~290 GW | ~2.8x on 2025 |
Gartner projects data center electricity demand growing roughly 26 percent in 2026 and power demand rising about 27 percent to 132 GW. For scale, 132 GW is comparable to the total generating capacity of a large industrialised country, dedicated to one category of building.
AI Share of the Total
| Metric | Value |
|---|---|
| AI-optimised server share of data center power, 2026 | ~31% (Gartner) |
| AI data center electricity growth rate vs overall, 2025 | 2.94x |
| US share of global AI data center capacity by power draw | ~45% |
The 2.94x ratio is the number to hold onto. AI-focused capacity is not merely growing, it is growing nearly three times faster than the data center sector it sits inside, which is why sector-level averages consistently understate the AI-specific picture.
Why Capacity Is the Binding Constraint
The limiting factor on AI expansion has shifted from chips to power, and from power generation to power delivery. A campus can have signed supply and still wait years for interconnection. Concentration makes this worse: Northern Virginia, Ireland, and Singapore carry enough local density that grid stress is acute regardless of national-level headroom. See the interconnection queue.
Consumption forecasts assume the capacity gets built. Increasingly it does not, either because the grid cannot deliver it or because the community declines to host it. See moratoriums and local opposition.
Brand Visibility Implications
Power availability is becoming a real constraint on inference supply, and constrained supply gets rationed by price. If inference costs stop falling, the cheap-model tier that currently absorbs most routed token volume becomes relatively more attractive, which shifts more brand-relevant answers toward open-weight models with different recall characteristics. Energy is upstream of the measurement problem in a way that is easy to miss.
Methodology
Figures compiled from Gartner, IDC, LBNL, grid-operator filings, utility rate cases, and press reporting through July 2026. Forecasts are cited to the forecaster because independent projections in this area diverge widely, and several of the underlying quantities are estimates rather than measurements. Where sources disagree, ranges are given rather than a single number. Updated quarterly.
How Presenc AI Helps
Presenc AI tracks brand visibility across the full model mix, so shifts driven by inference economics show up as measured changes rather than unexplained drift.