The Nuclear AI Boom: Why Big Tech is Reviving Dormant Reactors

Generating a single image with an AI model uses about as much energy as fully charging your smartphone. When you multiply that by billions of queries, you stop talking about software optimization and start talking about physical, grid-crushing gigawatts. The artificial intelligence revolution is running out of juice.

Before you read: We previously covered why inference compute is the new AI arms race with models like OpenAI o1. This article explores the physical, real-world consequence of that arms race: a terrifying spike in electricity demand.

Table of contents

• The Gigawatt Problem
• Waking the Dead
• A Critical Reality Check
• Fun Facts
• The New Hardware Reality
• Sources

The Gigawatt Problem

The math is sobering. A standard Google search requires about 0.3 watt-hours of electricity. A ChatGPT request consumes roughly 2.9 watt-hours. That is nearly a 10x increase in power draw for a single interaction.

According to the International Energy Agency (IEA), data centers accounted for around 2% of global electricity demand in 2022. By 2026, that number is projected to double. To put that in perspective, the AI industry is adding the equivalent power consumption of the entire country of Japan to the global grid in just four years. It is like replacing a global fleet of bicycles with M1 Abrams tanks overnight and wondering why the gas stations are suddenly empty.

Solar and wind cannot solve this alone. They are intermittent. AI data centers operate at 100% capacity, 24/7. They need baseload power—steady, unwavering, massive streams of electrons. And there is only one carbon-free source capable of delivering that at scale.

Waking the Dead

Frankly, I never thought I would see the day a software company resurrected the site of America’s worst nuclear meltdown just to generate better Python code.

In September 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy to revive Unit 1 of the Three Mile Island nuclear plant in Pennsylvania. They are renaming it the Crane Clean Energy Center. It will cost $1.6 billion to restart, and Microsoft is buying 100% of the 835 megawatts it produces.

Amazon went a different route. They flat-out bought a $650 million data center campus located directly adjacent to the Susquehanna nuclear plant in Pennsylvania, securing up to 960 megawatts of direct, behind-the-meter nuclear power.

Meanwhile, Google is playing the long game. They recently signed the world’s first corporate agreement to purchase power from a fleet of Small Modular Reactors (SMRs) built by Kairos Power, aiming to bring them online by 2030.

A Critical Reality Check

The tech industry loves a silver bullet, but nuclear power is not an agile software deployment.

Big Tech is writing massive checks for SMRs, a technology that is currently mostly vaporware. The only approved SMR design in the US, from a company called NuScale, saw its flagship project canceled in 2023 because the costs spiraled from $5.3 billion to $9.3 billion.

Restarting dormant plants like Three Mile Island or Palisades in Michigan is theoretically faster, but they face brutal regulatory hurdles from the Nuclear Regulatory Commission (NRC). Microsoft wants Three Mile Island online by 2028. Honestly, given the pace of federal safety reviews for a plant that has been sitting idle for half a decade, I will eat my keyboard if they hit that deadline.

Tech companies are promising investors infinite AI scaling today, backed by nuclear power that won’t realistically arrive until the 2030s. The math between now and then relies heavily on burning a lot more natural gas, completely torching the “net zero by 2030” climate pledges these same companies made five years ago. What is the difference between an AI startup and a nuclear reactor? One burns through a billion dollars and melts down in twelve months; the other actually generates electricity.

Fun Facts

• Three Mile Island Unit 1 (the reactor Microsoft is reviving) operated perfectly safely until its retirement in 2019. It was its neighbor, Unit 2, that suffered the infamous partial meltdown in 1979.
• Data centers in Ireland now officially consume more electricity than all urban homes in the country combined, accounting for over 21% of the national grid’s total output.

The New Hardware Reality

We spent the last decade abstracting the physical world away into “the cloud.” The AI boom is a brutal reminder that the cloud is just someone else’s computer, and that computer needs to plug into a wall. The next era of software dominance will not be won by the company with the best algorithms. It will be won by the company that owns the uranium.

Sources

• International Energy Agency (IEA): Electricity 2024 Report
• Electric Power Research Institute (EPRI): Powering Intelligence Report (May 2024)
• Constellation Energy Press Release: Crane Clean Energy Center (September 2024)
• Central Statistics Office of Ireland: Data Centre Metered Electricity Consumption 2023