Capgemini: AI data centres are making grid forecasting harder

The rapid build-out of AI infrastructure is not simply adding load to electricity grids; it is making that load significantly harder to predict. That is the central finding of a new report from the Capgemini Research Institute, which surveyed 612 senior electricity executives across 21 countries in January 2026, alongside 175 data-centre operators.

The research identifies a growing mismatch between projected and actual demand. Some 67% of utility executives report encountering "phantom" load requests from data-centre developers, with around 19% of those requests ultimately never materialising. The result is a capital-allocation dilemma: utilities must decide how aggressively to invest in grid capacity without reliable signals about which demand will convert, risking either stranded assets or supply shortfalls. A separate 68% of respondents already anticipate shortages as data-centre demand outpaces their ability to expand supply.

The forecasting problem

Three-quarters (77%) of utility executives surveyed say they are struggling to forecast future electricity demand accurately, citing AI workloads as particularly difficult to model given their variable and bursty consumption profiles. Geographic concentration compounds the issue: more than half identify localised load concentration as a major obstacle to reliable service, with dense clusters of high-density facilities creating bottlenecks that affect both system stability and investment planning.

Looking ahead, the report projects that AI training and inference could account for 60% of total data-centre electricity consumption within three to five years, up from the current 25%. If accurate, that trajectory would represent a structural shift in how and where grid investment must be directed.

Claire Gauthier, Global Head of Energy and Utilities at Capgemini, said: "The challenge is no longer only how much power is needed, but whether it can be delivered reliably, where and when it is required."

AI as part of the solution

Despite being the primary source of demand uncertainty, AI is also viewed by the sector as a tool for managing it. Around 60% of utility executives expect advanced AI analytics to deliver efficiency improvements of more than 10% across failure reduction, operational productivity and outage response. Yet adoption remains thin: only 45% say they currently use AI for grid optimisation at all, and just 16% have deployed more advanced AI-driven approaches to real-time power-flow management.

On the supply side, data-centre operators are responding to grid constraints by accelerating investment in behind-the-meter power. Some 39% plan to add on-site or BTM generation within the next one to two years, and more than 70% expect these solutions to materially reduce grid dependence within five years. A large majority (86%) see grid-independent operation as a competitive advantage, a signal that the traditional utility-as-sole-supplier model is under structural pressure.

Market and regulatory context

The findings land at a moment of heightened policy scrutiny on both sides of the Atlantic. The UK National Grid has flagged multi-year connection queues for large commercial loads, while EU member states are grappling with how to reconcile data-centre growth ambitions against accelerating decarbonisation commitments under REPowerEU. The report's finding that 68% of executives globally view natural gas as a near-term transitional fuel sits uncomfortably with those targets, and is likely to attract regulatory attention in jurisdictions with binding clean-energy obligations.

The broader investment picture is drawing in a range of actors beyond traditional utilities: battery energy storage developers, small modular reactor vendors and behind-the-meter software providers are all competing for capital that was, until recently, directed almost exclusively at grid-scale renewables. Capgemini's report, drawing on a large, multinational sample, adds quantitative weight to a debate that has until now been dominated by anecdote and vendor projection.