Better chips do not make the power problem disappear

Efficiency and growth are different measurements

The International Energy Agency reported that global data-centre electricity consumption rose by 17% in 2025, compared with 3% growth in overall electricity demand. Its April 2026 update also describes continuing infrastructure constraints. These figures concern aggregate consumption; they are not a fixed energy cost for an individual AI request. IEA: 2026 energy and AI update ↗

This distinction resolves an apparent contradiction. An application can require less energy for a particular task while the total electricity used by the industry increases. More users, more requests and more demanding workloads can outweigh gains in efficiency.

A server order depends on infrastructure outside the server

For a buyer, the practical issue is not the global total alone. It is whether the intended site can support the proposed installation. Electrical distribution, cooling and available connection capacity can set a project timetable that has little relationship to the hardware lead time.

The IEA identifies bottlenecks involving equipment and infrastructure, including transformers and grid connections. That is a reason to involve facilities teams early, not a reason to assume every location faces the same constraint. Local conditions and the proposed configuration still determine the answer. IEA: Key Questions on Energy and AI — executive summary ↗

Nameplate figures do not establish operating costs

A power supply's rating is not a measurement of what a server continuously consumes. Nor does an accelerator's power limit describe the whole rack. Host processors, memory, networking, storage and cooling all contribute to the operating requirement.

Build a scenario around the workload's expected utilisation and the facility's actual terms. Separate peak electrical planning from average energy consumption. A system may need a supply sized for demanding conditions while spending much of its time at a different load.

That distinction also matters when comparing platforms. A faster machine that finishes a job earlier may use energy differently from a slower machine running longer. The appropriate comparison is completed useful work under stated conditions, with the surrounding infrastructure accounted for.

Plan capacity with the room included

The IEA's projections are forecasts, not a promise of any particular future demand or electricity price. For an individual purchase, a more actionable exercise is to map the equipment to the facility and the workload.

Confirm electrical feeds, cooling capacity, rack constraints and installation responsibilities before locking the configuration. Then measure consumption during acceptance and normal operation. Efficiency improvements are valuable, but they have to survive the full deployment. Power planning belongs inside the hardware decision because it determines whether the purchased computing capacity can be used at all.