Data centres: 415 TWh. Make AI earn its power.

Energy meets AI from two sides at once. Utilities use it to forecast wind and sun, balance grids and serve customers. And AI's own data centres have become one of the fastest-growing sources of electricity demand. The International Energy Agency counts 415 TWh used by data centres worldwide in 2024, about 1.5% of world electricity, and projects about 945 TWh by 2030. The sector that keeps the lights on now has to power its own most demanding customer.

415 TWh used by data centres worldwide in 2024, about 1.5% of world electricity (IEA, 2025)

Adoption is high, results are mixed

On paper, AI is everywhere in the utility. 81% of North American utilities use AI, according to Itron. In Germany, 79% of energy suppliers have introduced it (PwC, 2026), and 82% of municipal utilities now see AI as an instrument of their digital transformation, up from 61% a year earlier (BDEW and EY). The most common uses are energy-demand forecasting (62% of utilities), predictive maintenance and asset monitoring (53%) and customer service (48%), according to L.E.K. Consulting's 2025 Global Energy Study.

In practice, many projects stall. 61% of AI uses at German energy suppliers are still at pilot stage (PwC). And energy has a higher failure rate than other industries: 34% of energy AI initiatives were terminated, against 26% across industries, according to S&P Global Market Intelligence. That is not a reason to stop. It is a reason to decide earlier, with a business case and a stop/go gate before money goes into scaling.

34% of energy AI initiatives were terminated, against 26% across industries (S&P Global Market Intelligence, Mar 2026)

Where AI pays on the grid

The IEA's evidence shows where the gains sit when AI is deployed well. AI-based fault detection shortens outages by 30–50%. Better use of existing lines could free up to 175 GW of transmission capacity without building new ones. Satellites and sensors with AI detect incidents up to 500 times faster.

These numbers matter because the grid is the bottleneck. Germany spends about €3 billion a year on grid congestion management (Bundesnetzagentur, 2025). In the United States, only 13% of grid connection requests filed between 2000 and 2019 reached operation (Lawrence Berkeley National Laboratory). A forecasting model that cuts error by a few points, or a grid tool that frees capacity on an existing line, translates directly into money and into faster connection for new renewables.

The demand AI brings

The other side is growing fast. Data-centre electricity use rose 17% in 2025, and the IEA expects it to double by 2030; the US, China and Europe account for 85% of today's use. In Ireland data centres took 22% of metered electricity in 2024 and 23% in 2025 (CSO). German data centres used about 20 TWh in 2024, with about 31 TWh expected by 2030 (Bitkom and Borderstep).

This sets a test every AI project in the sector should pass: does it earn its power? Worldwide AI spending is forecast at US$2.67 trillion in 2026 by Gartner. A utility that runs AI should be able to show what each workload saves in forecast error, congestion cost or outage minutes, against what it costs in compute and electricity. The Chips & Compute brief maps the supply side of that equation.

More than half of the 23 AI offerings forecast, trade or operate grids

The Good AI Prism places 24 entries in Energy & Utilities, 23 AI offerings and one unclassified record (Tibber), each with a dated source from the provider's own pages. Seven manage flexibility, demand response and customer energy: Bidgely, GridBeyond, 1KOMMA5°'s Heartbeat AI, Kraken, Sympower, Tibber and Uplight. Six forecast renewables and trade energy, among them enercast, Fluence Mosaic, Meteomatics, Rebase, Tesla Autobidder and Volue. Six plan and operate grids: AiDASH, envelio, GE Vernova's GridOS, Siemens' Gridscale X, Hitachi Energy's Nostradamus AI and Utilidata. Three optimise buildings and industrial energy, including Trane's BrainBox AI and Schneider Electric's EcoStruxure, and two monitor renewable asset performance. The Energy & Utilities sector brief lists each group with its rule.

The rules: safety component or not

In the EU the line is drawn by function. Under the AI Act as amended by Regulation (EU) 2026/1744, AI used as a safety component in managing and operating the supply of water, gas, heating or electricity carries high-risk duties from 2 December 2027 (Annex III, point 2). Systems that only assist or optimise are not automatically safety components. Customers must be told when they interact with an AI system since 2 August 2026.

Energy is also a highly critical sector under NIS2 and the Critical Entities Resilience Directive, and data centres from 500 kW report their energy use under the Energy Efficiency Directive. Germany adds its NIS2 implementation act with Bundesnetzagentur security catalogues, the KRITIS umbrella act and the smart-meter rollout. The UK regulator Ofgem published version 2 of its guidance on ethical AI in energy in May 2026; Texas lets grid operators curtail loads of 75 MW or more in emergencies; Korea treats energy supply as a high-impact area under its AI Basic Act since January 2026.

Seven questions before AI touches the grid

Before a pilot starts, settle seven questions in writing. Does the system act as a safety component in supply, or only advise? Which NIS2 and critical-infrastructure duties apply to the system and to its provider? Which meter and customer data does it use, and how is privacy protected? How is the model validated, and what is the fallback when it fails at three in the morning? How does it integrate with operational technology, and can we exit without rebuilding the control room? Do customers know when prices, offers or controls are set by AI? And which measures decide scaling: forecast error, congestion cost, outage minutes or energy saved? A provider that answers these in writing is ready for a pilot. One that answers with a demo is not. The why AI does not scale article explains the operating model behind the answers.

The people who run the grid

76 million people work in energy worldwide, and more than half of 700 organisations surveyed by the IEA report critical hiring bottlenecks. In advanced economies, 2.4 workers are near retirement for every new entrant under 25. 80% of German municipal utilities name the skills shortage as a challenge (BDEW and EY). AI can carry part of that load, but only if control-room operators keep the right to override it and are trained to do so.

In practice

Best practice: an energy ledger for every model. Each AI workload gets one line in a ledger: what it saves in forecast error, congestion cost or outage minutes, and what it consumes in compute and the electricity behind it. The ledger turns the question "does it earn its power?" into a figure reviewed each quarter, and gives the control room and the board the same view.

Scenario: a solar forecast at a municipal utility. A municipal utility replaces its day-ahead solar forecast with an AI model. The pilot runs in parallel with the old method for a full season, on the same data, and is scored on forecast error and balancing cost. The model advises traders and does not act on supply, so it stays outside Annex III, and the operator keeps the right to override it at any hour.

What to do Monday

1. List every AI use and model, from forecasting to customer service, including tools staff bring themselves.
2. Mark the safety components. Separate tools that act on supply from tools that advise; only the first group carries high-risk duties under Annex III.
3. Set the stop/go gate first. Run an AI Opportunity Case with a business case before scaling, given the sector's high termination rate.
4. Name operators with override rights and document their AI literacy training.
5. Measure what decides: forecast error, congestion cost, outage minutes, energy saved and the electricity the AI itself uses.

This article describes market and regulation. It is not legal advice.