Nuclear SMRs for AI:
The $10 Billion Bet on Atomic-Powered Data Centers

1. The Nuclear Gold Rush

Something extraordinary happened in the data center industry in 2024-2025. Five of the world's largest technology companies — Microsoft, Amazon, Google, Meta, and Oracle — simultaneously turned to nuclear energy as the answer to AI's insatiable power demands. Not wind. Not solar. Not batteries. Nuclear.

The numbers tell the story. Microsoft signed a $1.6 billion deal to restart Three Mile Island Unit 1, rebranded as the Crane Clean Energy Center. Amazon locked in a 1,920 MW power purchase agreement with Talen Energy's Susquehanna nuclear plant running through 2042. Google partnered with Kairos Power for 500 MW of advanced reactors by 2035. Meta announced nuclear deals totaling 6.6 GW with three separate providers. And Oracle's CEO casually mentioned they'd already secured building permits for three small modular reactors.

As a data center engineer who has spent years optimizing facility power and cooling systems, I find this nuclear pivot both entirely logical and deeply uncertain. The logic is clear: AI training clusters consume megawatts around the clock, and nothing matches nuclear's combination of density, reliability, and carbon-free output. The uncertainty is equally clear: no commercial SMR has ever powered a data center, costs have a history of spiraling, and timelines measure in years, not months.

This article examines the technology, the deals, the economics, and the engineering reality behind the biggest energy bet in data center history.

2. The Power Crisis No One Predicted

The fundamental driver behind Big Tech's nuclear pivot is a power demand crisis that caught the entire energy industry off guard. When ChatGPT launched in November 2022 and GPU clusters started scaling exponentially, data center power projections became obsolete overnight.

The projections paint a sobering picture. The International Energy Agency projects global data center electricity consumption will double from 460 TWh in 2024 to roughly 945 TWh by 2030. Goldman Sachs projects a 165-175% surge in data center power demand by 2030. The Lawrence Berkeley National Laboratory found that US data centers alone consumed 176 TWh in 2023 — 4.4% of total US electricity — and could reach 325-580 TWh by 2028.

This is not a gradual scaling curve. This is the fastest-growing category of electricity demand on the planet — growing at 15% per year, four times faster than all other sectors combined.

Why renewables alone cannot solve this

Solar has a capacity factor of roughly 25%. Wind averages about 35%. Nuclear exceeds 90%. AI training workloads run 24 hours a day, 7 days a week, often for weeks or months continuously. You cannot pause a training run because the sun went down or the wind stopped. And the battery storage required to bridge a multi-day weather event for a 500 MW campus simply does not exist at economic scale.

Nuclear provides what AI actually needs: 24/7 carbon-free baseload power at the density required by modern GPU clusters. A single 300 MW SMR produces more reliable energy per year than a 1 GW solar farm.

3. What Are Small Modular Reactors?

A Small Modular Reactor is a nuclear fission reactor with electrical output below 300 MWe, designed for factory fabrication and modular deployment. Unlike traditional gigawatt-scale nuclear plants that require massive on-site construction over a decade, SMRs are built in factories and shipped to site — at least in theory.

Key advantages over traditional nuclear

Scalability: Deploy one module, add more as demand grows. A 4-pack of 80 MW Xe-100 reactors gives you 320 MW. Need more? Add another 4-pack. This matches the phased deployment model that hyperscalers already use for data center campuses.

Passive safety: Most SMR designs use passive safety systems — natural circulation, gravity-fed cooling, negative temperature coefficients — that shut the reactor down safely without human intervention or external power. This fundamentally changes the safety equation compared to older designs.

Flexible siting: Smaller footprint, reduced emergency planning zones, and some designs that work with air cooling (no cooling towers needed) mean SMRs can be located where data centers are, not where rivers are.

Co-location potential: Behind-the-meter or adjacent-to-grid deployment eliminates grid interconnection queue delays, which currently average 5+ years in the US with over 2,600 GW of projects waiting.

4. Big Tech's Nuclear Portfolio

The scale of Big Tech's nuclear commitments is unprecedented. Here is every major deal, with verified figures.

5. SMR Technology Showdown

Not all SMRs are created equal. Six designs are competing for the data center market, each with fundamentally different approaches to reactor physics, cooling, and deployment.

Which design makes the most sense for data centers?

From a data center engineering perspective, several factors matter beyond raw megawatts:

Oklo Aurora (75 MWe) is sized perfectly for a single data hall (60-72 MW typical). Its compact footprint and sodium cooling (no water towers) make it ideal for co-location. But it's unproven and faced an NRC rejection in 2022 before resubmitting.

GE-Hitachi BWRX-300 (300 MWe) offers the right scale for a medium campus and uses well-understood boiling water reactor technology. It is the furthest along in Western construction (OPG Darlington, Ontario, construction started May 2025). The drawback: first-unit cost is US $5.6 billion.

TerraPower Natrium (345-500 MW) has an innovative molten salt energy storage system that can boost output to 500 MW during peak demand — useful for burst compute workloads. Broke ground in Kemmerer, Wyoming in June 2024. Backed by $3.4 billion (including $2B from DOE and investment from NVIDIA).

X-energy Xe-100 (80 MWe) uses helium cooling and TRISO pebble fuel, enabling air-cooled configurations that eliminate water dependency entirely. For data centers in arid regions, this is a significant advantage.

6. The Grid Is Breaking

To understand why nuclear co-location matters, you need to understand how broken the US power grid is.

PJM Interconnection, the largest US grid operator serving 65 million people across 13 states, projects a 6 GW shortfall in reliability requirements by 2027. The grid interconnection queue contains over 2,600 GW of waiting projects with average wait times exceeding 5 years. This queue is growing faster than it's clearing.

The regional concentration is alarming. Virginia, the world's data center capital, already consumes roughly 26% of its state electricity supply for data centers alone. North Dakota follows at 15%, Nebraska at 12%, Iowa at 11%, and Oregon at 11%.

The economic impact is already visible. Data centers accounted for an estimated $9.3 billion price increase in PJM's 2025-26 capacity auction. Average residential electricity bills increased $18/month in western Maryland and $16/month in Ohio as a direct result.

Why co-located nuclear bypasses the queue

In December 2025, FERC issued a unanimous decision allowing AI facilities to plug directly into nuclear and gas-fired generators, bypassing traditional grid interconnection queues. This was explicitly because PJM's existing tariff was deemed "unjust and unreasonable" given the scale of demand.

Co-located nuclear eliminates three critical bottlenecks: the interconnection queue (5+ years), transmission construction (3-7 years), and distribution capacity constraints. For a hyperscaler who needs 500 MW in 2028, the grid simply cannot deliver it through conventional channels.

7. The Cost Problem

Nuclear power's history is littered with cost overruns, and SMRs are not exempt. The cautionary tale is NuScale's Carbon Free Power Project (CFPP) — the poster child for SMR deployment.

The CFPP was cancelled in November 2023 when UAMPS (the utility consortium) failed to secure the required 80% subscription from member utilities. The project's cost escalation from $3.6B to $9.3B — a 158% increase — destroyed economic viability.

The pattern echoes traditional nuclear. Vogtle Units 3 & 4 in Georgia, the only new US nuclear construction in decades, ran 7 years late and ballooned from $14 billion to $35 billion.

But hyperscalers are different customers

Here is what is genuinely different about Big Tech as nuclear customers, compared to traditional utilities:

Long-term PPAs: Microsoft signed a 20-year agreement. Amazon's runs through 2042. These multi-decade commitments provide the revenue certainty that banks and investors require to finance nuclear construction. Traditional utilities serve ratepayers who can push back on cost recovery.

Balance sheets: Microsoft, Amazon, Google, and Meta have combined cash reserves exceeding $300 billion. They can absorb construction cost overruns that would bankrupt a regional utility. Meta's 6.6 GW commitment is backed by a company with $65 billion in annual revenue.

Alternative cost: The cost of not having power is potentially measured in lost AI training revenue of tens of billions of dollars. A 6-month delay in GPU cluster deployment due to power constraints could cost more than the cost overrun on a reactor.

8. The Nuclear-Water Connection

Readers of my previous article on AI water consumption will immediately see the connection. Traditional nuclear power plants are among the most water-intensive energy sources, evaporating up to 3.0 liters of water per kWh produced. Replacing one water problem (data center cooling) with another (nuclear cooling) would be counterproductive.

But several SMR designs fundamentally change this equation:

X-energy Xe-100: Uses helium gas as coolant. Can be configured with dry (air) cooling, eliminating water dependency entirely. This is the design Amazon chose for the Cascade project.

TerraPower Natrium: Uses liquid sodium cooling with a molten salt thermal storage system. Can operate with air cooling. Sodium's superior heat transfer properties enable efficient thermal management without evaporative towers.

Kairos Power Hermes: Uses molten fluoride salt (FLiBe) at low pressure. Passive cooling options reduce or eliminate water consumption compared to conventional pressurized water reactors.

An air-cooled SMR co-located with a data center that also uses liquid cooling (rather than evaporative towers) could address both the power and water sustainability challenges simultaneously. This is the scenario that makes the strongest engineering case for nuclear-powered data centers.

9. The Global Nuclear Race

The SMR push extends far beyond Silicon Valley. Governments worldwide are treating SMR development as a strategic priority.

The geopolitical dimension is significant. Countries that develop and deploy SMR technology first gain a strategic advantage in exporting nuclear technology — a market projected to reach $150-300 billion by 2040. The US, UK, Canada, France, and South Korea are racing not just for domestic energy independence, but for global export markets.

10. Timeline & Engineering Assessment

When will SMRs actually power data centers? Here is my engineering assessment, separating what is likely from what is aspirational.

The honest engineering take

Microsoft's TMI restart in 2027 is highly likely because it is restarting an existing reactor with known technology, not building something new. Everything else is a bet.

The OPG BWRX-300 in Canada (2029-2030) is the most important SMR project to watch. It is under construction, fully funded, and uses mature BWR technology. If it delivers within 20% of budget and 12 months of schedule, it will validate the entire SMR concept. If it doesn't, it will send a chilling signal through the industry.

For hyperscalers planning data center campuses today that need power in 2028-2030, the nuclear option is not available. Natural gas, existing nuclear PPAs, and grid power remain the only realistic options for near-term demand. SMRs are a 2030s solution being planned today.

What I'm watching as a data center engineer

1. OPG Darlington BWRX-300 construction progress. This is the bellwether. Quarterly updates on cost and schedule will tell us whether factory-built SMR economics are real or theoretical.

2. TerraPower Natrium operating license timeline. The application expected in 2027-2028 will test whether NRC can process advanced reactor licenses at the speed industry needs.

3. FERC co-location rules in practice. The December 2025 ruling opened the door, but practical implementation — how data centers actually interconnect with nuclear generators — will determine whether co-location works at scale.

4. The water question. Whether SMR-powered data centers choose water-cooled or air-cooled nuclear designs will determine if this solves the AI water problem or merely relocates it.

5. Second-order costs. Security, waste management, decommissioning, insurance, and regulatory compliance costs that may not appear in headline LCOE figures but will appear in real facility operating budgets.

The nuclear bet is enormous, unprecedented, and not guaranteed to work on the timeline or at the cost that Big Tech needs. But the alternative — a future where AI progress is constrained by power availability — is a risk these companies have decided they cannot accept. Whether the engineering delivers will be one of the defining infrastructure stories of the next decade.