October 16, 2026
An AI data center can have a site, financing and an order for advanced chips. None of those guarantees that it can switch on when its owner wants it to. The missing input may be a grid connection, a transformer or a credible agreement about who pays for the infrastructure.
That is the important shift in the AI buildout. Computing remains a constraint, but the ability to deliver reliable electricity to a particular location is becoming a constraint of its own. A megawatt available somewhere on the national grid is not necessarily a megawatt available at the proposed site.
The International Energy Agency's April 2026 update projects global data-center electricity consumption rising from 485 terawatt-hours in 2025 to 950 TWh in 2030, roughly 3% of global electricity demand by then. Those figures cover all data centers, not AI alone, and the 2030 figure is a forecast. The agency expects electricity consumption at AI-focused facilities to triple over that period. [1]
The headline is not that chips have stopped mattering. The same IEA report describes tightening constraints in advanced chip manufacturing and high-bandwidth memory. It is that ordering more chips does not solve a shortage of usable power. [1]
Electricity demand is measured in energy consumed over time. Grid connections and equipment have to accommodate power at the moment it is needed. Confusing the two can make a large annual supply number look like a solution to a very specific local capacity problem.
The IEA's 2025 assessment estimated that new transmission lines can take four to eight years to build in advanced economies. Its 2026 update describes continued pressure on connections, electricity equipment and planning systems. That mismatch matters: an ambitious data-center construction schedule can run ahead of the infrastructure required to serve it. [2] [1]
A developer therefore needs more than a low advertised electricity price. The practical questions are when service can begin, how much capacity is dependable, which upgrades are required and how the site will operate during constraints. A cheap location with an uncertain connection date can be less useful than a more expensive location that can actually support the workload.
Power availability also influences where the next generation of computing gets built. Areas with suitable infrastructure can attract projects that would otherwise cluster near existing technology markets. But a new connection is not a substitute for fiber, skilled staff, cooling and a workable local planning process. The deployment problem is a package, not a single commodity.
Ohio offers a concrete example of why demand forecasts need discipline. In July 2025, the state's utility commission directed AEP Ohio to create data-center-specific tariffs. The regulator said the settlement was intended to protect other customers from the costs of underused infrastructure built to serve data centers. [3]
In a February 2026 update, AEP Ohio reported that a group of prospective customers initially requesting more than 30,000 megawatts proceeded to paid engineering studies covering 13,022.7 MW. That process ultimately produced binding contracts for 5,642 MW under the new tariff. These are the utility's reported figures, not an independent estimate of market demand. [4]
The distinction is essential. AEP also had 12,219 MW of contracts signed before the new tariff, giving it 17,861 MW of total contracted data-center projects scheduled to come online progressively through 2035. Contracted capacity is not the same as capacity already operating. [4]
The lesson is not that the AI buildout is imaginary. It is that a request, a paid study, a binding contract and an energized facility are different stages. Financial commitments help reveal which projects are prepared to bear the costs their plans create. Treating every announcement as certain demand can encourage infrastructure spending that outlives the project behind it.
Slow connections make on-site generation attractive. A developer can try to produce electricity beside the servers rather than wait for the grid to deliver everything.
But that moves part of the constraint rather than removing it. The IEA's 2026 analysis finds that reliable on-site gas generation needs redundant capacity to serve critical, variable data-center loads. It also warns that gas-turbine supply constraints mean on-site generation is not necessarily a faster route to market at scale. [1]
Fuel supply, emissions requirements, maintenance and backup arrangements still have to work. Batteries can help manage power swings and short interruptions, but their duration and recharge requirements matter. A project cannot replace a dependable operating plan with a favorable headline about an energy technology.
Nor is every workload equally flexible. Work that can be scheduled or moved may tolerate different constraints from a service that must answer customers immediately. The useful question is not whether a data center is flexible in principle. It is which work can move, under what conditions, without breaking the service it exists to provide.
There is a real counterargument to the power narrative: AI can become more energy-efficient. The IEA reports rapid reductions in electricity used per individual task. Yet it also describes rising use and more demanding applications, including video generation, reasoning and agentic work. Total consumption depends on all three: efficiency, adoption and the mix of tasks. [1]
That makes neither endless demand growth nor effortless efficiency savings a safe assumption. Electricity infrastructure has long lives. Forecasts should distinguish committed projects from aspirations and should be updated as actual usage emerges.
The same caution applies to electricity bills. Data centers do not automatically make everyone's power more expensive. The IEA finds that the outcome depends on existing supply, required investment, the shape of demand and policy choices. Reliable new demand can improve use of underutilized infrastructure; poorly aligned expansion can leave other customers carrying costs. [1]
That is why tariffs and connection terms belong in the AI conversation. Who pays for upgrades? What happens if a project shrinks or leaves? Can the customer reduce demand when the system is stressed? These are not administrative details. They determine whether local communities see the buildout as useful investment or an obligation imposed on them.
The next stage of AI will still involve better models and faster processors. It will also depend on engineering, utility planning and credible commitments. The decisive milestone is not the announcement of a larger campus. It is the point at which the campus can reliably operate, with the costs and responsibilities understood.
Image: AI-generated editorial illustration, not documentary photography.
