For most of the internet era, computing companies could treat electricity as an input rather than a strategy. Rent servers, choose a cloud region, watch the bill. Power was essential but abstract: a utility problem several layers below the product.

Artificial intelligence is collapsing those layers.

The International Energy Agency reported that electricity demand from data centers rose 17 percent in 2025, while demand from AI-focused facilities grew faster still. It expects global data-center electricity use to double by 2030 and AI-focused use to triple.2 In the United States, the Department of Energy estimates data centers consumed 4.4 percent of electricity in 2023 and could consume between 6.7 and 12 percent in 2028.3

The range is wide because the system is moving faster than its measurement. It is also wide enough that precision can distract from the direction. After decades in which American electricity demand barely grew, large computing facilities are helping turn load growth into a central industrial question again.

An AI company can order chips in quarters. Much of the electrical system is planned in years.

That mismatch is creating a new geography of intelligence—not one organized only around talent and capital, but around where power can actually be delivered.

A megawatt is not a commodity if it arrives in 2031

Electricity statistics flatten time. A region may produce enough annual energy and still be unable to serve a large new facility at the place and hour it is required.

A data center needs more than generation. It needs a site, transmission capacity, substations, transformers, permits, cooling, backup systems, and an agreement governing connection to the grid. Every component has its own queue.

At the end of 2025, Lawrence Berkeley National Laboratory counted roughly 8,200 proposed U.S. generation and storage projects seeking grid interconnection: 1,312 gigawatts of generation and 749 gigawatts of storage. Among projects that reached commercial operation in 2025 where data was available, the median journey from interconnection request to operation exceeded five years.4

Those queues measure new supply, not data-center load, but they expose the same institutional constraint: the grid was not designed to evaluate a flood of large, overlapping requests at software speed. FERC’s Order No. 2023 attempts to move generation interconnection from “first come, first served” toward “first ready, first served.”6 That is an important reform. It does not make a substation appear.

For builders, the important metric is therefore not the advertised price of electricity. It is time to firm power: how long until a site can receive the quantity and reliability of power the workload requires, under terms that survive construction and operation?

This turns energy diligence into product diligence.

Compute has one advantage: some of it can move in time

Factories, hospitals, and homes consume power where they are. Compute is unusual. Some workloads can move between regions, shift by hours, slow down, pause, or run when electricity is abundant.

Not all of them can. Interactive inference has latency obligations. Training runs have expensive interruption costs. A half-finished checkpoint is not equivalent to a thermostat cycling for ten minutes. But the category has more temporal flexibility than its standard design admits.

That flexibility can become infrastructure.

A power-aware compute system might route interruptible batch work toward regions with surplus generation, delay non-urgent inference, coordinate checkpoints with grid events, or bid load reductions into electricity markets. Data centers can combine utility service with storage or behind-the-meter generation. Operators can design separate service levels for workloads whose real requirement is “completed by morning” rather than “running continuously.”

The dominant cloud abstraction treats a compute unit as interchangeable across time. The emerging energy reality rewards software that knows the difference between urgent, movable, interruptible, and thermally constrained work.

There is a catch. Flexible load is easy to describe and difficult to underwrite. Grid operators need confidence that a promised reduction will occur. Customers need guarantees that their work will finish. The coordination layer has to understand both power systems and computing systems well enough to make commitments to each.

That boundary is a company-building surface.

The firm boundary is moving down the stack

Major technology companies are already signing long-duration power contracts, backing generation projects, and hiring teams that resemble those of industrial developers. This is not simply vertical integration for its own sake. When a critical input becomes scarce and slow, owning more of its delivery can be rational.

The same shift will happen at smaller scales.

Site selection will increasingly start with electrical topology rather than real-estate price. Capacity rights, interconnection position, and transformer availability can matter as much as acreage. Developers with credible utility relationships may be more valuable than developers with a speculative parcel and a press release.

Software companies serving the sector will need to model infrastructure truth rather than merely aggregate public announcements. A proposed gigawatt is not an operating gigawatt. A queue position is not a completion date. A power-purchase agreement is not a physical connection.

The best products will answer operational questions:

  • Which sites can support a given load shape, and when?
  • Which network upgrades are required, who pays, and what could change the estimate?
  • Which workloads can be shifted without violating customer commitments?
  • How should storage, onsite generation, and grid service be combined?
  • What is the exposure to congestion, curtailment, fuel prices, and permitting?
  • Which project milestones are evidenced rather than announced?

This is less glamorous than a new model demo. It is also where large amounts of model capacity will either exist or fail to exist.

Efficiency will not settle the question

Chips and models are becoming more efficient. The IEA describes rapidly falling electricity use per AI task. It simultaneously projects rising total consumption because cheaper, more capable systems invite far more use, including long-running agents.1

Both claims can be true.

Efficiency determines how much useful computation a megawatt produces. Demand determines how many megawatts the market wants. Improvements in the first do not guarantee a decline in the second.

This means the energy argument should not collapse into moral panic about a single query or blind confidence that hardware will solve everything. The consequential question is what kind of electrical system accompanies computational growth.

If new demand is rigid, geographically concentrated, and planned through inflated requests, it can raise costs and strain institutions. If it is paired with credible new supply, transmission, flexible scheduling, and honest queue management, it can finance infrastructure that serves a wider economy.

The distinction is design.

Intelligence acquires an address

Software encouraged a generation of founders to believe scale was almost placeless. The marginal user could be anywhere; infrastructure expanded behind an API.

AI retains that global reach at the interface. Underneath, it is becoming intensely local. A model runs in a particular building, drawing from a particular grid, through equipment with manufacturing lead times, under agreements approved by particular institutions.

The next important AI infrastructure companies may therefore look unfamiliar. Some will resemble energy developers. Some will coordinate loads. Some will manufacture cooling or power equipment. Some will turn opaque interconnection processes into reliable development data. Some will build compute markets whose unit of account includes when and where electricity is available.

They share a premise: intelligence is not floating in the cloud. It is being converted from electricity, at a site, on a schedule.

The companies that understand that conversion will decide where the next layer of the digital economy can actually be built.