Capgemini's latest survey of 600 utility executives finds that AI workloads are forecast to consume most data center electricity within five years, while the same operators plan to disconnect from the grid.
Utilities are being asked to pour billions into grid capacity for AI data centers whose operators are, at the same time, racing to stop depending on the grid at all. The forecasting systems meant to guide those investments are broken, and the bill for the gap is likely to land on everyone else.
That is the contradiction at the heart of a new report from Capgemini Research Institute, published June 25 and built on a survey of more than 600 senior electricity executives at organizations with annual revenue above $500 million. The accompanying press release frames the work as a study in "grid adaptation and investment." Capgemini, the global consulting and IT services firm, sells AI, data-center, and sustainability advisory work in the same market its report describes, which is worth keeping in mind when weighing any of its forecasts.
The headline number is striking. AI training and inference, which today account for about 25% of total data center electricity demand, are expected to climb to roughly 60% within three to five years, largely displacing other IT workloads. Behind that single figure is the entire reason the grid has become the bottleneck. AI workloads do not behave like the office computing that used to fill data centers. They pull power in dense bursts, run around the clock, and tolerate downtime poorly. The result, by Capgemini's tally, is that 77% of utilities now struggle to forecast future demand accurately and 68% expect outright supply shortages because data center demand is outpacing the grid's ability to expand.
Most of those demand forecasts are also being corrupted from the inside. Sixty-seven percent of electricity executives told Capgemini their data center customers routinely request power capacity that never materializes, which the firm labels "phantom load." Nineteen percent of those requests, on average, simply do not happen. That is not a small forecasting error. It is a systematic distortion of the signal utilities use to decide whether to build new transmission, substation capacity, or generation. Build for phantom demand and you overbuild, locking in costs that someone has to pay for decades. Underbuild and you brown out the customers who actually show up.
The second half of the contradiction is harder to ignore. The same data center operators whose phantom reservations are confusing utility planning are also, by their own executives' account, aggressively detaching from the grid. About 30% already operate behind-the-meter generation or on-site storage, and another 39% plan to add behind-the-meter capacity within one to two years. Eighty-six percent see independence from the grid as a competitive advantage, and more than 70% expect a significant reduction in their reliance on grid power within five years.
That is the recipe for stranded utility assets. Utilities are still expected to fund the transmission lines, substations, and reserve generation needed to serve peak AI demand, and to recover those costs through regulated rates, while their largest growth customers quietly build themselves off-grid escape hatches. If the AI boom delivers the demand that justifies the investment, the math works. If it does not, or if the behind-the-meter shift moves faster than expected, the cost gap does not vanish. It gets reallocated, and the most likely destination is the residential and commercial rate base.
There is a quieter signal inside the same numbers that complicates the picture. Less than half of utilities, 45%, say they currently use AI for grid optimization, and only 16% have deployed more advanced AI-driven approaches for power flow, resilience, and real-time performance. Roughly 60% of executives expect advanced AI analytics to deliver double-digit improvements in outage prevention, operational productivity, and failure reduction. The same technology that is breaking the grid's demand forecasting is also being pitched, by the same consultancy and many of the same vendors, as the fix for the grid's operational performance. That is not a contradiction on its face, but it is a reminder that the diagnosis and the prescription both come from firms with a commercial interest in the answer.
Geographic concentration makes the planning problem worse and more local. More than half of the executives surveyed identify clustering of data centers in specific regions as a major reliability obstacle, with localized bottlenecks affecting both system stability and investment decisions. A grid built to handle a national average is not a grid built to handle a Northern Virginia, a Phoenix, or a Dublin.
On the supply mix, the executives surveyed are notably skeptical of single-answer solutions. Seventy-eight percent of electricity executives and 73% of data center executives say renewables alone are not enough for continuous AI workload power, and both groups report active investment in battery energy storage. More than 68% acknowledge that small modular reactors and conventional nuclear remain a long-term option rather than a near-term fix. The emerging consensus is a diversified stack, with the grid increasingly acting as a flexible backbone rather than the primary source.
What to watch next: whether regulators in the most concentrated data center regions start treating behind-the-meter build-outs as a stranded-cost trigger, and whether utilities begin publishing demand-realization rates alongside reservation totals so the phantom load problem becomes auditable rather than rumored. Both are testable, and both decide who ends up paying for an AI power boom that may not need as much grid as the planning documents claim.