Summary

• Capgemini says AI datacentres are making electricity demand more volatile and harder to forecast.
• Utilities face “phantom” power requests that can distort investment planning, grid connection queues, and infrastructure buildout.
• The growth of AI infrastructure is turning power availability, permitting, flexibility, and energy mix into strategic technology constraints.

Capgemini says AI-driven datacentres are making electricity demand harder to forecast, creating a planning problem for utilities, governments, hyperscalers, and investors.

The consultancy’s latest research, AI meets the grid: shaping the datacentre power play, finds that nearly 80% of utilities expect more extreme and volatile demand patterns. Around one in five datacentre power requests may never materialise, while 67% of electricity executives refer to speculative or “phantom” load requests that distort forecasts.

The pressure comes from more than rising electricity consumption. Large AI facilities can require power on an industrial scale, but projects may be delayed, resized, moved, or cancelled as operators chase compute demand, grid access, equipment availability, planning approval, and commercial financing. Utilities then have to decide whether to invest in generation, transmission, substations, interconnections, and reserve capacity for demand that may not arrive as promised.

That uncertainty creates a capital allocation problem. Under-investment can delay datacentre connections, restrict AI infrastructure, and push projects into regions with faster grid access. Over-investment can strand assets, raise system costs, and create public resistance if grid upgrades are built around speculative private demand. In European markets, where planning, grid congestion, renewable integration, and local environmental concerns already slow large infrastructure projects, the margin for error is narrow.

The research also shows AI’s dual role in the electricity system. While AI workloads are driving new pressure on the grid, 60% of utilities expect AI to improve grid efficiency and unlock operational gains. Few have implemented advanced AI-driven approaches at maturity. The technology is therefore both load and tool: it increases electricity demand while offering ways to improve forecasting, maintenance, balancing, flexibility, and network optimisation.

For datacentre operators, power availability is becoming as important as land, fibre, tax incentives, and proximity to customers. Operators may need firmer commitments before asking for grid capacity, more on-site or near-site generation, better participation in flexibility markets, and clearer workload planning. Utilities need stronger ways to separate credible projects from speculative queue inflation before they commit capital.

The issue complicates climate claims around digital infrastructure. Datacentres often rely on renewable power purchase agreements to support sustainability targets, but always-on AI workloads place different demands on the electricity system. Renewables can support growth, yet they do not remove the need for firm capacity, grid reinforcement, storage, demand flexibility, and local planning consent. Nuclear, gas, storage, and local generation are all re-entering the discussion because operators and policymakers are trying to match intensive compute demand with reliable supply.

Europe’s AI and cloud ambitions depend heavily on this physical layer. The EU can fund model development, support AI factories, and encourage cloud sovereignty, but datacentre growth still has to fit into electricity systems that were not designed for sudden, concentrated AI loads. Ireland, the Netherlands, Germany, France, the Nordics, and the UK are all dealing with different versions of the same problem: compute demand is moving faster than grid planning cycles.

Capgemini’s report points towards a more disciplined phase of AI infrastructure development. Utilities need portfolio-level demand intelligence, faster and more selective interconnection processes, modernised tariffs, and closer collaboration with datacentre developers. Hyperscalers and AI companies need to treat power as a binding constraint rather than an operational detail to be solved after site selection.

The next bottleneck for AI may not be model quality or enterprise appetite. It may be whether electricity systems can distinguish credible demand from speculative signals quickly enough to build the right infrastructure in the right places.