Why AI Data Centers Are Turning to Nuclear
AI power demand is outgrowing the grid, and hyperscalers are buying nuclear to keep up.
AI data centers are turning to nuclear power because AI needs huge, steady, round-the-clock electricity. The grid cannot add that power fast enough.
In the last two years, Microsoft, Amazon, Google, and Meta all signed nuclear deals. Some restart old reactors. Others fund new small modular reactors (SMRs).
This guide explains why the shift is happening, which deals are real, and what it means for your AI costs and risk. It is written for business leaders, not energy investors.
What is driving AI data centers to nuclear
AI data centers are turning to nuclear because AI electricity demand is rising faster than the grid can supply it. Nuclear offers large, constant, low-carbon power.
The scale is hard to overstate. The IEA estimates data centers used about 415 terawatt-hours (TWh) in 2024. That is roughly 1.5% of all global electricity.
The IEA expects that to roughly double to about 945 TWh by 2030. AI-focused data centers are the fastest-growing part.
The United States feels this first. A DOE-backed Lawrence Berkeley National Lab report found U.S. data centers used about 4.4% of national electricity in 2023.
That same report projects data centers could reach 6.7% to 12% of U.S. electricity by 2028. That is a very fast climb.
AI workloads run all day and night. They need firm power that never dips. Nuclear runs around the clock, which fits AI far better than weather-dependent sources.
- Data centers used about 415 TWh globally in 2024, near 1.5% of world electricity (IEA).
- Global data center demand may hit about 945 TWh by 2030 (IEA).
- U.S. data centers may reach 6.7%-12% of national electricity by 2028 (Berkeley Lab).
- AI servers are the single biggest driver of this growth.
Wondering how the nuclear-and-AI energy shift affects your own AI cost and risk? Book a consultation with Layer3 Labs to map your AI energy exposure and build an efficiency plan.
Book a ConsultationWhat is a small modular reactor (SMR)?
A small modular reactor (SMR) is a nuclear reactor that makes 300 megawatts of electricity or less. It is built in a factory and shipped to the site in modules.
Traditional reactors are giant custom projects. Each one is built on-site over many years. SMRs aim to be smaller, repeatable, and faster to assemble.
The NRC, the U.S. nuclear regulator, defines SMRs as light-water reactors under 300 megawatts. Some newer designs use different coolants and are called advanced reactors.
The promise is simple. Build many identical units in a factory. Lower the cost per unit. Place them near where the power is needed, like a data center campus.
Companies like NuScale, X-energy, Kairos Power, and Oklo are leading U.S. designs. We name them only as context, not as investment tips.
- SMR = a reactor producing 300 MW or less.
- Built in a factory, then shipped in modules to the site.
- Goal: repeatable units, lower cost, faster build than giant reactors.
- Can sit next to a data center for on-site or near-site power.
The hyperscaler nuclear deals so far
Every major cloud provider has now signed a nuclear deal to power AI. These deals fall into two buckets: restarting old reactors and funding new SMRs.
The restarts move fastest because the plants already exist. Microsoft signed a 20-year power deal in September 2024 with Constellation Energy.
That deal will restart the 835-megawatt Three Mile Island Unit 1 in Pennsylvania. It is being renamed the Crane Clean Energy Center. Constellation targets a 2028 restart at about $1.6 billion.
Amazon took two paths. It agreed to buy up to 1,920 megawatts from Talen Energy's existing Susquehanna nuclear plant.
Amazon also anchored a $500 million funding round for X-energy in October 2024. Together they plan the "Cascade" project in Washington state, starting near 320 megawatts and expandable to 960 megawatts.
Google chose new SMRs. It signed a Master Plant Development Agreement with Kairos Power in October 2024 for up to 500 megawatts by 2035, with a first reactor around 2030.
Meta issued a nuclear request for proposals in December 2024 seeking 1 to 4 gigawatts. In January 2026 it partnered with Oklo on a 1.2-gigawatt campus in Pike County, Ohio.
- Microsoft + Constellation: restart Three Mile Island Unit 1 (835 MW), target 2028.
- Amazon + Talen: up to 1,920 MW from the existing Susquehanna plant.
- Amazon + X-energy: new SMRs (Cascade), ~320 MW first phase, up to 960 MW.
- Google + Kairos Power: up to 500 MW of new SMRs by 2035.
- Meta + Oklo: 1.2 GW campus in Ohio, 16 Aurora units at 75 MW each.
Why not just use the grid, solar, or gas?
Nuclear wins for AI because rival options each fall short on speed, steadiness, or carbon. AI needs firm power now, and no single alternative delivers all three cleanly.
The grid is overloaded. In some regions, connecting new large power to the grid now takes more than eight years. Over 2,000 gigawatts of projects sit waiting in U.S. queues.
Solar and wind are cheap and clean, but they do not run all the time. AI data centers need steady output, so they would need huge batteries to fill the gaps.
Natural gas is fast to build and reliable. But it adds carbon, which clashes with the clean-energy goals of most large tech firms.
Nuclear is the one source that is firm, low-carbon, and land-efficient. Its downside is time and complexity, which the next section covers.
Here is a simple side-by-side of the main options.
| Power option | Reliability | Carbon | Timeline | Cost note |
|---|---|---|---|---|
| Grid hookup | High if available | Mixed | Often 5-8+ years to connect | Rising; capacity is scarce |
| Solar + storage | Variable, needs batteries | Very low | 2-4 years | Cheap energy, costly firming |
| Natural gas | High | High | 2-4 years | Low upfront, carbon exposure |
| Restart existing nuclear | Very high | Very low | ~3-4 years | ~$1.6B example (Crane) |
| New SMR | Very high (once built) | Very low | Late 2020s-2030s | High upfront, unproven at scale |
- Grid: interconnection can take 8+ years in busy regions.
- Solar/wind: clean but not steady without large batteries.
- Gas: fast and firm, but adds carbon.
- Nuclear: firm, low-carbon, land-efficient, but slower to build new.
Realistic timelines and limits
Most new SMRs will not deliver power until the late 2020s or the 2030s. That gap matters, because AI demand is here right now.
New reactors need design approval, licensing, and construction. The NRC review is thorough and slow by design. First-of-a-kind projects often slip.
Oklo targets its first small Aurora reactor at Idaho National Laboratory around late 2027. Kairos aims for a first plant near 2030. Meta's Oklo campus starts later still.
This is why restarts came first. Three Mile Island Unit 1 already exists, so Microsoft's power could arrive by 2028, years ahead of most new SMRs.
Costs are also uncertain. SMRs promise savings through mass production, but that only works after many units are built. Early units carry high, unproven price tags.
The honest summary: nuclear is a strong long-term answer for AI power, but it is not a quick fix for demand in 2026.
- Most new SMRs: online in the late 2020s to 2030s.
- Restarted plants (like Crane): power possibly by 2028.
- Licensing and first-of-a-kind builds often cause delays.
- SMR cost savings depend on building many units, not the first few.
What this means for your AI cost and risk
For your business, the nuclear pivot signals that AI power is now a real cost and supply risk, not a background detail. It can shape your AI vendor's capacity and pricing.
Here is the non-obvious part. Energy cost flows into the price you pay per token or per API call. When power gets scarce and expensive, that pressure reaches your bill.
There is also a capacity risk. If your AI vendor cannot secure power, it may delay new regions or ration compute. Your access to the newest models can slow down.
Power scarcity is already raising prices. One regional U.S. grid saw its capacity auction cost jump to $14.7 billion in 2025, up from $2.2 billion a year earlier.
You do not need to build reactors. You need to plan as if AI power will stay tight and slightly pricey for years. Build that into your budgets and vendor contracts.
Smart moves: track your AI usage, right-size models to the task, and ask vendors about their power and capacity plans. Efficiency is now a cost lever, not just a green goal.
- Energy cost feeds into per-token and per-request AI pricing.
- Power shortages can delay your vendor's capacity and model rollout.
- One U.S. grid's capacity cost rose to $14.7B in 2025 from $2.2B.
- Right-sizing models and tracking usage now saves real money.
The bottom line for AI adopters
The bottom line: AI's power hunger is driving a real nuclear pivot, and it will affect your AI cost and reliability over the next decade. Plan for tight, valuable power.
Restarts like Three Mile Island bring power sooner. New SMRs are promising but mostly land in the 2030s. Neither removes today's supply crunch.
For most companies, the action is not energy strategy. It is AI efficiency, smart vendor choices, and clear-eyed budgets that assume power stays a constraint.
Frequently Asked Questions
- AI data centers are turning to nuclear because AI needs large, steady, round-the-clock electricity that the grid cannot add fast enough. Nuclear supplies firm, low-carbon power that matches AI's constant demand.
- An SMR is a nuclear reactor that produces 300 megawatts of electricity or less and is built in a factory, then shipped to the site in modules. The goal is faster, repeatable, lower-cost builds compared to giant reactors.
- Microsoft, Amazon, Google, and Meta have all signed nuclear deals. Microsoft is restarting Three Mile Island, Amazon works with Talen and X-energy, Google partnered with Kairos Power, and Meta partnered with Oklo.
- Data centers used about 415 TWh globally in 2024, near 1.5% of world electricity, according to the IEA. That figure could roughly double to about 945 TWh by 2030 as AI demand grows.
- Restarted plants deliver soonest. Microsoft's Three Mile Island restart targets 2028. Most new SMRs from Oklo, Kairos, and X-energy are expected in the late 2020s to the 2030s, so they are a longer-term answer.
- Solar and wind are clean but not steady, so they need large batteries for round-the-clock AI loads. The grid is often years away from connecting new large demand. Nuclear is firm, low-carbon, and land-efficient.
- Yes. Energy cost flows into per-token and per-request AI pricing, so scarce, expensive power can raise your bill. Power shortages can also delay your vendor's capacity and slow access to new models.
- Modern U.S. reactors operate under strict NRC oversight, and SMRs are designed with added passive safety features. Every project still needs full regulatory review before it can run, which is one reason timelines are long.
- A restarted reactor is an existing large plant brought back online, like Three Mile Island Unit 1 at 835 MW. An SMR is a new, smaller, factory-built reactor under 300 MW that still needs to be licensed and constructed.
- Assume AI power stays tight and slightly pricey for years. Track your AI usage, right-size models to each task, and ask vendors about their power and capacity plans. Efficiency is now a direct cost lever.
Map your AI costs and energy exposure to your real business
Layer3 Labs helps you turn AI power and cost trends into a concrete plan for your workflows. Our free AI workflow audit shows where you can cut spend and reduce risk.
Book your free AI workflow audit