AI’s Electricity Demand: What the Latest IEA and DOE Numbers Really Show

Artificial intelligence does not consume electricity in isolation. It runs inside data centres that also power cloud storage, streaming, enterprise software, and traditional computing. That makes the energy debate vulnerable to two opposite mistakes: attributing every data-centre watt to AI, or dismissing AI because total global demand still looks small.

The International Energy Agency’s base case projects global data-centre electricity consumption of around 945 terawatt-hours in 2030—just under three percent of global electricity use. Within that total, accelerated servers used for AI are expected to be the fastest-growing component. The IEA’s 2026 update also reported that data-centre electricity use rose sharply in 2025 and highlighted shortages in grids and critical equipment.

A global share can hide a local shock

Electricity systems are regional. A new cluster of data centres may be manageable in a global chart and still overwhelm a particular substation, transmission corridor, or generation plan. Facilities also tend to concentrate where fibre, land, tax incentives, water, and power are available.

This concentration explains why utilities can face large connection queues even if data centres remain a minority of worldwide demand. The challenge is not only producing enough annual electricity. It is delivering power at the required location, reliability, and time.

Forecasts are scenarios, not meter readings from the future

Demand estimates depend on uncertain variables: how quickly AI adoption grows, how much computation each task requires, improvements in chips and cooling, the rate at which new capacity can connect, and whether software efficiency offsets part of the growth.

The US Department of Energy illustrates that uncertainty with a broad range. Its report estimated that US data centres used 176 TWh in 2023, equal to 4.4 percent of national electricity consumption, and could reach 325 to 580 TWh in 2028. The width of that range is not a defect. It is a warning to planners to prepare for more than one plausible future.

Efficiency can reduce cost and still raise total use

More efficient processors and models can lower the electricity required for an individual task. Yet cheaper, faster computation can increase the number of tasks performed. This rebound effect means efficiency is necessary but cannot be assumed to reduce total demand.

Operators need measures that can be audited: power usage effectiveness, water use, workload scheduling, hardware utilization, and the carbon intensity of electricity at the time it is consumed. Annual renewable-energy matching is useful, but it does not by itself prove that clean generation was available during every hour of operation.

The infrastructure response is broader than new generation

The IEA and DOE analyses point toward a portfolio rather than a single technology. New clean generation matters, but so do transmission, substations, transformers, storage, demand flexibility, and faster interconnection processes. Existing data centres can also shift some non-urgent workloads toward hours when electricity is less constrained or cleaner.

Planning should protect other customers. If infrastructure is built primarily for a concentrated new load, regulators must decide how costs are allocated and how reliability risks are managed. Transparency about forecasts and connection assumptions is essential.

What the numbers really mean

The strongest conclusion is not that AI will inevitably break the grid or that innovation will make the issue disappear. It is that a fast-growing, geographically concentrated demand source is colliding with infrastructure that takes years to build.

Policy should be flexible enough to handle uncertainty but firm enough to require credible forecasts, efficient design, and responsible connection. The question is no longer whether AI has an energy footprint. It is whether the electricity system and the technology industry can make that footprint measurable, financeable, and compatible with wider public needs.