Why electricity, not chips, is the real limit

Early in the AI buildout the scarce thing was chips: everyone wanted GPUs and there were not enough. That bottleneck eased as production ramped. What replaced it surprised much of the industry. The binding constraint became electricity.

The framing that captures 2026 is blunt: for a growing number of projects, the limit on building a new AI data center is not capital, not land, and not even chips, but whether the local grid can deliver the power. You can buy the GPUs and pour the concrete, and still be unable to turn the machine on because the electricity is not available.

This is a genuine shift in what kind of problem AI is. A software company's growth was limited by talent and code. An AI company's growth is now limited by megawatts, a constraint shared with aluminum smelters and steel mills, not with other software.

The rest of this lesson is about why electricity, of all things, became the wall, and it comes down to a mismatch: AI demand arrived in a few years, and the electricity system moves in decades.

The heart of the problem is a difference in clock speed between two industries.

AI moves in months. A lab decides to build, orders chips, and wants a running cluster within a year or two. Model generations turn over on a similar timescale.

The electricity system moves in decades. A new power plant takes years to permit and build. A major transmission line can take a decade or more through planning, permitting, and land acquisition. The grid was designed for demand that grows slowly and predictably, a few percent a year, giving planners time.

AI demand does not grow a few percent a year. A single large AI campus can request hundreds of megawatts, the output of a substantial power plant, to arrive at once. Modern AI sites are commonly quoted in the range of 100 to 750 megawatts each, and the largest proposed campuses go into the gigawatts, the scale of a nuclear station.

So you have a demand that materializes in months meeting a supply system that responds in decades. That mismatch is the entire power problem, and everything else, the queues, the gas turbines, the nuclear deals, is industry improvising around it.

To draw power from the grid, a large new load must be studied and approved to connect, ensuring it will not destabilize the system. This process has become a years-long traffic jam known as the interconnection queue.

The backlog is enormous. Projects seeking to connect to grids, generation and large new loads like data centers, total on the order of 1,500 gigawatts in the United States alone and more than 2,500 gigawatts worldwide, according to grid and agency data. In the busiest hubs, new high-capacity connections in places like Northern Virginia, Dublin, Singapore, and Amsterdam now face waits of four to seven years.

Read that against the previous lesson's point about fast-depreciating chips. If your GPUs lose much of their value in a few years but you must wait years just to power them, the economics are brutal: you could be paying for silicon that is half-obsolete before it ever runs.

The queue is why location became a central strategic question. A site is now chosen less for land or tax incentives and more for one thing: can it get power soon? The scramble for places with available electricity, or the ability to make their own, defines where AI gets built.

If the grid cannot deliver electricity in time, the response is to bring your own generation and skip the wait. This is why AI, a digital industry, is suddenly deep in the power business.

The options divide by how fast they can be built:

• Natural gas turbines are the fastest path. They can be deployed on-site in roughly 12 to 18 months, far quicker than a grid upgrade, which is why many developers are advancing on-site gas generation, largely in the United States, as the IEA noted in 2026. Fast, but with emissions.
• Nuclear is the slower, cleaner bet with the most headlines. Deals have been struck to restart a retired reactor for a tech company on a multi-year timeline, and to build small modular reactors for first power around the end of the decade. These are long-horizon commitments to firm, carbon-free power.
• Renewables plus storage are cheap and clean but intermittent, so on their own they struggle to guarantee the 24/7 power a data center needs.

The pattern is telling. Companies that a few years ago worried only about software are now signing 20-year power contracts and reviving nuclear plants. The compute race turned technology firms into energy buyers and, increasingly, energy builders.

Behind the headline constraints sits an unglamorous one that quietly gates everything: the transformer, the device that steps voltage up for long-distance transmission and back down for use. You cannot connect any large load to the grid without them, and they have become scarce.

Lead times tell the story. Large power transformers averaged around 128 weeks, roughly two and a half years, to deliver as of 2025, and specialized units even longer. These are complex, custom-built machines made by a handful of manufacturers who cannot ramp production quickly.

This is a perfect illustration of a deep principle in any buildout: the constraint is always the slowest link, and it is often something boring. All the money, all the chips, all the ambition, and the project waits on a two-year-lead transformer. A system moves at the speed of its most stubborn component, and that component is rarely the glamorous one.

The transformer shortage also shows why the power problem cannot be solved by spending alone. You can pay any price for a chip and get it faster; you cannot easily pay to make a transformer factory, a grid connection, or a permitting process move at software speed. Some constraints are physical and institutional, and money moves them only slowly.

Power is the biggest physical constraint, but not the only one, and the others increasingly shape where AI can build.

Cooling and water. The heat from dense AI hardware has to go somewhere. Many facilities use large amounts of water for cooling, which becomes contentious in dry regions and puts data centers into direct competition with communities and agriculture for a scarce local resource. Liquid cooling helps with the chips but does not erase the overall heat-rejection problem.

Local grid strain and prices. A data center drawing the power of a small city can raise electricity costs or strain reliability for everyone else on the same local grid. That has turned some communities from eager hosts, chasing jobs and tax revenue, into wary or resistant ones, adding a political and social constraint on top of the physical ones.

The honest summary is that AI compute has become a heavy-industry siting problem. Where you can build is now governed by the availability of power, water, cooling, and community acceptance, exactly the questions that have always governed factories and refineries. The digital cloud turns out to be extremely physical, and it has to negotiate with the same real-world limits as any other large industrial facility.

Step back and the power problem reframes the whole AI story. The popular narrative is about algorithms and intelligence. The physical reality is that the pace of AI is increasingly set by the pace at which the world can generate and deliver electricity to run it.

That has three consequences worth holding.

First, energy policy is now AI policy. How fast a country can permit power plants, build transmission, and clear interconnection queues directly shapes how much AI it can host. Grid reform and AI competitiveness became the same conversation.

Second, the constraint is durable. Chips can be manufactured faster; power systems are physical and institutional and change slowly. So electricity is likely to remain the binding limit for years, not months, regardless of how much money is available.

Third, it makes the economics harder, which is the next lesson. If you must build or contract power years ahead, on decade-long commitments, to run chips that depreciate in a few years, the financial bet becomes far more demanding than it looks from the outside.

The simplest way to remember the whole lesson: AI runs on electricity, the world cannot make electricity as fast as AI wants to consume it, and that gap, not intelligence, is what currently governs the buildout.

AI demand arrives in months; the electricity system responds in years to decades. The gap forces the queue, the on-site gas and nuclear deals, and leaves projects waiting on slow physical components like transformers.