The Grid Wasn't Built for This

AI's appetite for electricity is rewriting American energy infrastructure — and the bill is coming due for everyone.

Abilene, Texas is a city of about 125,000 people, tucked into the flat scrubland of West Texas, better known for cattle ranching than cutting-edge technology. That changed when OpenAI broke ground there on the flagship campus of its Stargate Project — a data center complex occupying roughly 4 million square feet, billed as the largest single building in the world, powered by a natural gas plant capable of generating 360 megawatts, with plans to scale to 2 gigawatts of computing capacity. In a matter of years, a city-scale industrial installation appeared in the Texas plains, drawing power and water from a region that already has limited supplies of both.

Abilene is not an anomaly. It is a preview.

The build-out of artificial intelligence infrastructure is now one of the most consequential forces reshaping the American energy landscape, and most people have no idea it is happening. The data centers that train and run large language models, power cloud computing, and process the world's digital workloads are quietly becoming one of the largest sources of electricity demand in the United States — and the grid, built for a different era and a different scale of consumption, was not designed for what is being asked of it.

The Scale of What's Coming

Start with the numbers, because they are extraordinary.

U.S. data centers consumed 183 terawatt-hours of electricity in 2024, according to the International Energy Agency — more than 4% of the country's total electricity consumption, roughly equivalent to the annual demand of the entire nation of Pakistan. By 2030, that figure is projected to grow by 133%, reaching 426 terawatt-hours. BloombergNEF forecasts that U.S. data center power demand will more than double by 2035, rising from approximately 35 gigawatts in 2024 to 78 gigawatts. Average hourly electricity demand from this sector is expected to nearly triple over the same period.

To understand what those numbers mean in practical terms: a typical AI-focused hyperscale data center consumes as much electricity annually as 100,000 households. The larger ones currently under construction are expected to use 20 times as much. A single cluster of 100,000 H100 GPUs — a scale that multiple companies now operate — draws 70 megawatts continuously.

The four largest tech companies — Amazon, Microsoft, Google, and Meta — collectively spent over $200 billion on capital expenditures in 2024, a 62% year-over-year increase from 2023. Amazon's capital expenditure alone reached $85.8 billion, up 78%. Amazon's projected spending for 2025 surpasses $100 billion. These are not software companies buying servers. They are becoming infrastructure companies, and the infrastructure they're building runs on electricity at a scale the American grid was not designed to absorb.

Virginia and the Limits of a Local Grid

Nowhere makes this more visible than Loudoun County, Virginia — a stretch of suburban landscape outside Washington, D.C. that has become known as Data Center Alley, home to the largest concentration of data centers on the planet. The county has over 27 million square feet of existing data center space, and the growth is not slowing.

The strain is measurable. Between 2016 and 2024, data center electricity use in Virginia surged by 231 percent, according to Dominion Energy officials. The demand from Loudoun County's facilities was so intense that the transmission grid became constrained for high-energy customers, requiring new 500-kilovolt and 230-kilovolt transmission lines just to handle the load. Dominion has committed to investing $50.1 billion in capital projects between 2025 and 2029 — transmission lines, substations, and new generation — largely to keep up.

The costs of that build-out do not fall only on the companies consuming the power. In the PJM electricity market, which stretches from Illinois to North Carolina, data centers accounted for an estimated $9.3 billion price increase in the 2025-26 capacity market. The result is a concrete increase in residential electricity bills — an estimated $18 per month in western Maryland, $16 per month in Ohio. A Carnegie Mellon University study estimates that data centers and cryptocurrency mining combined could lead to an 8% increase in the average U.S. electricity bill by 2030, potentially exceeding 25% in the highest-demand markets of central and northern Virginia.

In other words, the residents of those markets are subsidizing the infrastructure of the AI industry, whether they know it or not.

The Nuclear Pivot

The power demands of AI are so extreme, so constant, and so concentrated that renewable energy alone cannot meet them — at least not yet. Wind and solar generation is intermittent. AI training clusters need continuous, reliable baseload power, 24 hours a day, every day. This reality has driven a shift that would have seemed improbable just a few years ago: the largest technology companies in the world are going nuclear.

The deal that marked the pivot was Microsoft's 2024 agreement with Constellation Energy to restart Pennsylvania's Three Mile Island Unit 1 — a reactor that had been shut down since 2019, now being refurbished and renamed the Crane Clean Energy Center. The 20-year power purchase agreement, worth $16 billion, will route the facility's 835 megawatts of output exclusively to Microsoft's data centers in the mid-Atlantic region. Nearly 50 years after a partial meltdown made Three Mile Island synonymous with nuclear catastrophe, the plant is being brought back to life to power artificial intelligence.

The deal opened the floodgates. Google signed the first U.S. corporate small modular reactor fleet deal with Kairos Power, targeting 500 megawatts of capacity by the early 2030s. Amazon invested $700 million in X-energy and signed a power purchase agreement with Talen Energy's Susquehanna nuclear plant. Meta committed to sourcing between 1 and 4 gigawatts of new nuclear capacity. Oracle announced plans for a gigawatt-scale data center powered by three small modular reactors. As of mid-2026, every major tech hyperscaler has signed at least one nuclear power deal. Thirteen deals, totaling over 9.7 gigawatts committed.

Nuclear power has gone from a regulatory burden to a competitive necessity. The logic is straightforward: whoever secures the power supply secures the AI infrastructure. Long-term nuclear contracts — typically 20 to 30 years — create structural advantages for incumbents that cannot be replicated by newcomers. Energy access is becoming a moat, alongside chip access.

The Water Nobody Is Talking About

Electricity is the headline resource. Water is the quieter one.

Data centers generate enormous amounts of heat, and the most common way to manage that heat is evaporative cooling — the same basic principle as sweating, just at industrial scale. Large data centers can consume up to 5 million gallons of water per day, equivalent to the daily water use of a town of 10,000 to 50,000 people. A study by the Houston Advanced Research Center found that data centers in Texas will use 49 billion gallons of water in 2025 — and as much as 399 billion gallons by 2030, equivalent to drawing Lake Mead down by more than 16 feet in a single year.

The geography of this problem is not accidental. Data centers are frequently built in dry regions because lower humidity reduces the risk of corrosion damage to servers. But lower humidity also means scarcer water. The facilities are being constructed in precisely the places least equipped to give up their water supply. In Newton County, Georgia, Meta broke ground on a $750 million data center, and the county is on track to face a water deficit by 2030. OpenAI's Stargate campus in Abilene sits in drought-prone West Texas. About 30% of data centers currently under construction are in regions where water scarcity is expected to intensify by 2050.

The community backlash is already arriving. In 2025, local opposition to AI data centers led to the delay or cancellation of projects totaling $156 billion globally. In Goodyear and Buckeye, Arizona, a $14 billion project was withdrawn after local authorities blocked rezoning. In Spain, citizen groups launched a campaign under the slogan "Tu Nube Seca Mi Río" — your cloud is drying my river — calling for a moratorium on new data centers. Google's plans for a $200 million facility in Chile were temporarily halted after environmental opposition and court action over water concerns.

A New Kind of Infrastructure

What is emerging from this moment is something genuinely new: technology companies functioning not as software businesses but as energy utilities. The distinction matters.

A software company writes code. It has no physical footprint that strains municipal water systems or reshapes a regional grid. The companies building AI infrastructure in 2025 and 2026 are doing something different. They are acquiring power plants, signing 20-year energy contracts, lobbying state governments for subsidies, and siting industrial campuses in communities that had no meaningful say in the decision. Texas committed over $1 billion in data center subsidies in 2025. Virginia offered $732 million in 2024. The states are competing for investment that reshapes their grids, their water tables, and their residents' electricity bills — often before those residents fully understand what is being built.

The irony embedded in this moment is worth sitting with. The technology being powered by all of this infrastructure is marketed as efficient, invisible, frictionless — a tool that lives in the cloud, that answers your questions instantly, that requires nothing from you but a prompt. But behind every prompt is a physical chain: a GPU cluster drawing 70 megawatts, a cooling system drinking millions of gallons of water, a transmission line cutting through someone's county, a reactor restarted after decades of dormancy.

The grid wasn't built for this. Neither were the aquifers, the transmission corridors, or the communities sitting in the shadow of facilities they didn't choose. The AI boom is, at its foundation, an infrastructure story — and infrastructure always has an address.

The question worth asking is who lives at that address, and whether they were consulted.