OpenAI is buying a decommissioned nuclear plant in Ohio to power a 10-gigawatt AI campus. For comparison, that's roughly the electricity consumption of seven million American homes — dedicated to one company's AI workloads.

OpenAI has begun a sweeping build-out of a nuclear-powered AI campus in Ohio, anchored by the restart of Three Mile Island Unit 1 — renamed Crane Station. The deal with Constellation Energy commits roughly 10 gigawatts of dedicated capacity, enough to make this campus the single largest AI training site in North America.

The numbers, when you sit with them, are staggering. A 10-gigawatt continuous draw is roughly the output of ten large nuclear reactors. By comparison, that's the electricity use of about seven million American households — directed, full-time, into a single company's AI infrastructure.

Why It Matters

Modern AI data centers consume power unlike anything in commercial computing. Microsoft, Google, Amazon and Meta have all been quietly buying power agreements for years, but AI workloads have continued to outpace the grid. The Ohio deal is the moment when AI's appetite crossed the line from "concerning" to "structural."

Why nuclear? AI training runs are continuous, can't tolerate brownouts, and need stable baselines regardless of weather. Solar and wind are cheap per-watt-hours, but they are intermittent. Nuclear is weather-independent, carbon-free, and runs 24/7 at near-full capacity — exactly the profile an AI campus demands.

The facility itself is a story of industrial repurposing. Three Mile Island Unit 1 was shut down in 2019 for economic reasons, not safety. It sat dormant for five years. Restoring it for AI is faster and cheaper than building a new plant — and entirely separate from the damaged Unit 2 reactor, which remains in decommissioning.

The Core Idea

Total power consumption for an AI campus isn't a single number. It's the sum of three parts: compute (chips running), cooling (heat removal), and network/infrastructure (plumbing, networking, building overhead). For heavy training workloads, compute dominates. For inference-heavy campuses serving millions of users, cooling can become the largest line item.

OpenAI's Ohio build-out commits the energy envelope first, then fills the compute. By purchasing the entire output of a 10-gigawatt nuclear plant, they remove the largest single-variable cost from their capacity planning. The remaining problem — chips, buildings, networking — is solvable on engineering timelines. Power, before this deal, was not.

Key Findings

Power = P_compute + P_cooling + P_network; 10 GW ≈ 7M homes. The Ohio campus commits roughly the baseload of ten large nuclear reactors. Stable, continuous, dedicated. No grid stress. No solar-supplements-at-night gamble.

Vertical integration extended to power. OpenAI's strategy now mirrors Google's and Meta's: own the full stack. The bottleneck that the AI industry faces is no longer chips — it's watts.

Trends across the industry. Microsoft has reopened Three Mile Island Unit 1 for its own AI workloads. Amazon has bought a nuclear-powered data center campus in Pennsylvania. Meta has signed 20-year nuclear-power agreements. Energy is becoming a strategic asset on par with compute itself.

Crane Station timing. Constellation has stated the first power delivery from the restart is targeted for 2027. Until then, OpenAI is securing interim contracts and grid capacity to bridge the build-out.

What It Means for Practitioners

If you are an AI practitioner, the constraint that matters most two years from now might not be GPU availability — it might be power. Energy is becoming a strategic lever, and AI model design choices that improve performance-per-watt (sparsity, mixture-of-experts, quantization-aware training, model distillation) are going to be commercially valuable in a way they weren't a year ago.

If you are an end user of AI services, the price you pay for inference is increasingly shaped by power contracts. Efficiency is no longer just about cost — it's about who can deploy the largest models, and where.

Honest Caveats

Restarting a dormant nuclear plant is faster than building a new one, but it isn't instant. Crane Station is targeted for 2027, and substantial permitting, refueling, and turbine work sits between now and that first delivery.

There is risk in making the entire campus's reliability depend on a single energy source. A nuclear restart that runs into delays would cascade directly into the AI campus's timeline. The "weather-independent" advantage cuts both ways.

The 10-gigawatt figure is a commitment-to-capacity, not yet a delivered draw. The build-out will ramp. Power agreements of this size are usually staged over multiple years — so the headline number is a destination, not the day-one reality.

Close

The AI industry just bet half a trillion dollars that the binding constraint on intelligence is no longer silicon — it's watts. The Ohio announcement is the cleanest proof yet that this bet is real.