AI data center power demand is outrunning the grid
The short version
Power demand from US data centers is going from 31 gigawatts in 2025 to 41 GW in 2026 and 66 GW in 2027. Goldman Sachs expects only 60 percent of the capacity scheduled for next year to arrive on time. The IEA says around 20 percent of planned projects are at risk of delay because grids cannot connect them fast enough.
What changed
For three years the story was chips. In 2026 it is electricity. Goldman Sachs Research puts US data center power demand at 31 gigawatts in 2025, 41 GW in 2026 and 66 GW in 2027, based on capacity reaching roughly 95 GW by the end of 2027, more than double the level at the end of 2025.
The demand forecast is not the problem. The delivery is. Goldman expects about 60 percent of the capacity scheduled for the next year to materialize on time, falling to roughly 50 percent over the two-year horizon. Supply chain and labour shortages are the most common causes of delay.
The numbers
| Measure | Figure | Source |
|---|---|---|
| US data center power demand | 31 GW (2025), 41 GW (2026), 66 GW (2027) | Goldman Sachs |
| Share of peak summer demand | 4.1% (2025) to 5.3% (2026) to 8.5% (2027) | Goldman Sachs |
| Planned projects at risk of delay | About 20% | IEA |
| Growth in data center IT demand to 2030 | About 27% a year, reaching 121 GW | McKinsey |
| Projects postponed by community action, Q1 2026 | 4 GW | McKinsey |
Why the grid is the bottleneck
An AI-focused data center draws as much electricity as 100,000 households, and the largest ones more. Unlike an aluminium smelter, which can sit anywhere with cheap power, a data center has to sit near fibre and customers, which concentrates demand in a handful of counties.
Adding a new high-voltage transmission line in the US often takes more than a decade. A data center can be built in 18 to 24 months. That gap is the crunch: the buildings arrive, the wires do not.
McKinsey's read is that the near-term risk is underbuilding rather than overbuilding, and that the power infrastructure being built is unlikely to become stranded even if compute demand falls short of projections.
The local politics
Community pushback is now a measurable constraint. Four gigawatts of projects were postponed in the first quarter of 2026 because of local opposition, according to McKinsey. Water use, noise and rising household bills are the usual triggers.
For operators, that changes site selection. For everyone else, it changes electricity prices in the counties where the clusters sit.
What it means if you buy AI, not build it
Three practical consequences, all visible on invoices already.
Capacity is rationed. When a provider cannot add power, the cheapest tier sells out first and the queue moves to premium models. That is part of why some API bills rise while token prices fall.
Regional pricing appears. Inference in a constrained grid costs more to serve. Expect more providers to price by region or by time of day.
Long contracts come back. Reserved capacity with a committed spend is how the industry bought its way out of the 2024 GPU shortage, and it is happening again with power.
Questions people ask
Is the AI data center buildout a bubble?
McKinsey's analysis says short-term power shortfalls are more likely than stranded assets, because the infrastructure has uses beyond AI. That is not the same as saying valuations are safe.
How much power does an AI data center use?
A typical AI-focused facility draws as much electricity as about 100,000 households, and the largest exceed that.
Why can't they just build power plants?
Generation is quicker than transmission. New high-voltage lines in the US often take more than ten years to permit and build.
Read next: why AI token prices fell 94 percent while bills rose, and how to measure cost per task instead of cost per token.
Sources
Goldman Sachs Research via American Public Power Association (15 June 2026); IEA, Energy and AI, executive summary; McKinsey, Powering AI: how real is the risk of overbuilding (31 July 2026); Fortune on the $1.6 trillion buildout (19 October 2026).
Read from the primary source on 21 October 2026. Numbers change, so check before you commit.