Will AI Hit a Power Ceiling?
AI is not likely to hit a single, hard global “power ceiling” overnight. The more realistic constraint is local: transformers, substations, interconnection queues, cooling systems, and the pace at which reliable electricity can be delivered to specific sites. In other words, AI expansion may be delayed by infrastructure timing, not stopped by one universal limit. The key question is not only how much power is available in theory, but how fast compute can be deployed, how long grid connections take, and whether dependable power is available when the chips are ready.
One useful distinction is between MW and TWh. MW describes instant power capacity at a point in time; TWh describes energy used over a period, usually a year. Confusing those numbers can make the scale of AI demand look clearer than it is. It also matters that “data centres” includes more than AI facilities. General cloud storage, enterprise IT, and network equipment all sit inside the broader category, so forecasts for all data centres are not the same thing as observed AI-only consumption.
IEA attribution: the IEA’s 2025 Executive Summary says all data centres consumed about 415 TWh in 2024, and its 2025 base-case forecast for 2030 was about 945 TWh. That forecast is not a measured outcome; it is a scenario, not a fact on the ground. The difference matters because forecasts describe possible demand under assumptions, while actual supply still depends on where projects land and how fast infrastructure is built.
The bottleneck chain usually starts with transformers and grid queues. A company may be able to order servers quickly, but getting a new utility connection can take much longer. Then comes cooling: more dense racks raise heat-removal requirements, which can slow or reshape project design. Reliability is another issue. AI training clusters often want large, continuous blocks of power, while inference can sometimes be shifted into flexible batch workloads that run when electricity is cheaper or more available. That flexibility can reduce pressure, but only if the application can tolerate delays.
Onsite power is part of the tradeoff. Gas turbines, batteries, fuel cells, and other behind-the-meter options can improve speed or resilience, but they do not remove the need for fuel, maintenance, permits, or local interconnection. Annual renewable purchasing can also be misunderstood: buying renewable certificates or signing a yearly contract is not proof of continuous local supply at every hour. It helps with accounting and procurement, but it does not guarantee that the nearby grid is supplying clean power exactly when the facility needs it.
The most practical way to think about the issue is a race between three clocks: compute deployment time, grid connection time, and reliable power availability. If chips arrive before the grid is ready, projects slow down. If the grid is ready but cooling or backup systems are not, projects still stall. If reliable electricity exists but only intermittently, operators may need to redesign workloads or accept lower utilization. That is why the likely story is not a universal AI power wall, but uneven growth shaped by local infrastructure.
Source: https://www.iea.org/reports/energy-and-ai/executive-summary
Indicators to watch: confirmed grid connection, energised capacity, and commissioning milestones.
