Power infrastructure for the AI era
Compute is no longer the scarcest input for AI. Power is. This guide covers how much electricity AI needs, why the grid has become the bottleneck, and what developers, utilities and generators are building to close the gap.
How much power AI needs
Data centres have long been a steady, modest share of electricity demand. AI changes that. Training and serving large models runs on dense racks of GPUs and accelerators that draw far more power per rack than traditional servers, and the largest new campuses are planned in hundreds of megawatts to gigawatts.
Global averages hide the real pressure. Demand concentrates in a few regions where land, fibre and power line up, and there a single campus can match the load of a mid-sized city. For utilities that planned for flat demand, that is a sudden change in the planning baseline.
Why the grid is the bottleneck
A data centre can be built in one to two years. The power to run it often takes longer.
Interconnection queues
New large loads and new generation both have to be studied and approved before they connect. In many markets those queues now run for years, and data-centre developers increasingly pick sites by where power is available, not where land or customers are.
Transmission and equipment
New transmission lines face long permitting timelines. Transformers, switchgear and turbines have long manufacturing lead times, and supply has not kept pace with orders.
Round-the-clock reliability
AI training runs want steady power every hour of the year. That favours firm sources, or variable renewables paired with storage and backup, and it raises the question of who pays for the grid upgrades a single customer triggers.
Where the power will come from
| Source | Role for AI loads | Timeline |
|---|---|---|
| Grid upgrades and existing plants | Cheapest when capacity exists; limited by queues and transmission | Now to several years |
| Natural gas, often on-site | Firm power that is quick to deploy; emissions and turbine supply are the constraints | 1–3 years |
| Solar and wind with batteries | Low-cost energy; needs storage or backup for 24/7 loads | 1–3 years |
| Existing nuclear (restarts, uprates, long-term contracts) | Firm, carbon-free power bought under long contracts | Now to a few years |
| Small modular reactors (SMRs) | Firm, carbon-free and siteable near campuses; still pre-commercial at scale | Mostly 2030s |
| Geothermal and other firm clean power | Emerging options backed by hyperscaler offtake deals | Late 2020s onward |
The pattern across hyperscalers is the same: sign long-term power purchase agreements for whatever can deliver soonest, and place bets on firm clean sources for the next decade. Restarting a retired reactor and offtake agreements with SMR developers are both now part of that portfolio.
Behind-the-meter generation, where power is produced on or next to the campus instead of drawn from the grid, has moved from a niche idea to a mainstream option where grid access is slow.
Using less: cooling, chips and flexibility
- Liquid cooling. Direct-to-chip and immersion cooling handle rack densities that air cannot, and cut the energy spent on cooling.
- More efficient silicon. Each accelerator generation does more work per watt, though total demand still rises as usage grows.
- Flexible load. Some AI work, especially training and batch jobs, can slow down or move when the grid is stressed. Grid operators are starting to treat flexible data centres as a resource that can make faster connections possible.
- Heat reuse. Waste heat can feed district heating networks where the infrastructure exists.
Who is building it
AI power is turning into its own industry, with several groups converging on it:
- Data-centre developers and hyperscalers securing land and power together, often years ahead.
- Utilities and grid operators rewriting load forecasts, connection rules and tariffs for very large loads.
- Independent power producers building or repurposing plants alongside data-centre campuses.
- Nuclear and SMR developers signing offtake agreements with technology companies.
- Equipment makers of turbines, transformers, batteries, cooling and power electronics scaling up supply.
- "Power-first" AI campuses: companies that start with a power source and build compute around it.
FAQ
How much electricity do data centres use?
The IEA's 2025 Energy and AI report estimated about 415 TWh in 2024, around 1.5% of global electricity, and projected about 945 TWh by 2030, with AI the main driver.
Why is grid connection the bottleneck for AI data centres?
A large AI campus can need hundreds of megawatts or more, delivered reliably around the clock. New transmission, substations and generation take years to approve and build, so connection queues in many markets now take longer than building the data centre itself.
Will nuclear power AI data centres?
Partly, over time. Hyperscalers have signed long-term deals to restart or extend existing reactors and to buy power from small modular reactors, but most new SMR capacity is not expected before the 2030s. Until then, gas, renewables with storage and grid upgrades carry most of the load.
Latest news
Headlines on data-centre power, grids, nuclear and energy for AI from industry sources, updated through the day.
- Intro Group to invest $270m in Egypt’s Kemet Data Centerdatacenterdynamics.com · 2026-10-02
- AI data center planned for Karlsruhe Institute of Technology, Germanydatacenterdynamics.com · 2026-10-02
- NJ data center polluted for a year before regulators stepped inlatitudemedia.com · 2026-10-02
- Second part of construction permit application submitted for SMRs at Palisadesworld-nuclear-news.org · 2026-10-02
- NPCIL signs MoU for Gujarat nuclear power projectsworld-nuclear-news.org · 2026-10-02
- NYPA to own 51% of 240-MW solar project with EDFutilitydive.com · 2026-10-01
- Senate permitting bill would expand federal role in transmission sitingutilitydive.com · 2026-10-01
- North Carolina food bank gets microgrid backup for future stormscanarymedia.com · 2026-10-01
This domain is for sale
A short, strong name for an AI power company: a data-centre developer, an energy-for-compute platform, an independent power producer, an SMR or on-site generation venture, or a grid-flexibility product. Send your offer or questions below.