Nuclear Data Centers: the Power Deals Behind the AI Buildout

Nuclear data centers crossed from press release to power plan this year: Microsoft signed a 20-year, $16 billion PPA for the full 835 megawatts of the restarted Three Mile Island Unit 1, Google contracted up to 500 megawatts of Kairos small modular reactors plus three 600-megawatt nuclear projects, Amazon committed $20 billion toward Susquehanna-linked capacity, Oracle says its 1-gigawatt datacenter will run on a trio of nuclear reactors — and the US government is backing the restart wave with up to $4.4 billion in loans. The intelligence buildout hit the wall of physics, and the biggest buyers in tech are answering with reactor contracts. This guide measures the deals on verified numbers, separates what ships this decade from what ships next, and prices the honest risks — including who pays when a nuclear PPA underperforms.

The lens: treat every deal below as a verified contract on its public terms, label the forecasts as forecasts, and test the promises against the grid reality that our data center power pipeline analysis maps.

Key Takeaway

• ⚛️ Nuclear data centers are contracted now, sized in gigawatts: Microsoft’s $16B 20-year deal for all 835MW of TMI Unit 1 (renamed Crane Clean Energy Center), Google’s 500MW Kairos SMR orders plus three 600MW projects, Amazon’s $20B Susquehanna lane, and an Oracle 1GW nuclear-powered campus.
• 🔁 Restarts land first, SMRs later: three shuttered US plants are moving to restart (two with Microsoft and Google ties); the nuclear SMR fleet mostly delivers in the early 2030s — the timelines differ by five years and the hype rarely says so.
• 📊 The ceiling, stated honestly: Deloitte’s analysis caps nuclear’s contribution at up to 10% of data-center electricity demand by 2035 — the megaprogram is a premium slice of the power plan, not the whole plan.
• ⚠️ Risk travels with the contract: nuclear PPAs are decades long; the same wave carries cancellation reports and cost-overrun history — the honest guide names who bears the over-performance risk.
• 🇵🇭 The Philippines lesson: the country’s Bataan reactor was built in the 1980s and never fueled — a $2 billion-plus lesson in what happens when nuclear ambition meets governance and financing realities; Asia’s new demand wave revives the conversation with hard requirements first.

Why AI broke the power plan

A frontier AI campus in 2026 draws hundreds of megawatts continuously — a hyperscale facility’s load rivals a mid-sized city — and the queue length at the utilities (Years-long interconnection queues from Dominion to AEP) became the binding constraint on buildout speed. Grid additions take years; the power pipeline analysis tracks how transmission, generation and equipment all lag demand; and the 24/7 nature of training loads makes the intermittency solutions (wind, solar) a partial fit at best. Nuclear — dense, firm, carbon-free, on-site-able — moved from “interesting” to “contracted” because it answers the one thing a training campus cannot wait for: a power source matched to its load factor. The buyers’ logic is arithmetic, not ideology: gigawatt-scale compute needs gigawatt-scale firm power, and the countries that host it (the SEA infrastructure scoreboard tracks the regional race) see the same math.

The nuclear data centers deal scoreboard

Deal-terms receipt for the scoreboard above: Introl — nuclear power for AI: inside the Microsoft, Google, Amazon and Oracle data deals.

The pattern across nuclear data centers: buy the output, not the plant. Every tech buyer structures its deal as a power purchase agreement or capacity contract — the nuclear asset stays on a utility’s or developer’s balance sheet while the tech company locks decades of firm supply. That is also where the risk allocation lives (the next section prices it).

Restart, SMR or grid: three ways to contract a reactor

Class 1 — restarts (fastest, this decade): finished plants get revived; nuclear data centers contract firm power in years rather than decades — Three Mile Island Unit 1 is the template, with three shuttered plants in the current restart wave and federal loans softening the capex hit; nuclear data centers built this way contract firm power in years, not decades. Class 2 — SMRs (the 2030s fleet): small modular reactors promise factory-built repeatable units; the Kairos orders are the flagship commitment, and the honest timeline is early-2030s initial delivery — SMR economics remain unproven at fleet scale, which is why the contracts are tranche-structured. Class 3 — the grid route (the majority solution): most nuclear data centers still plug into the grid while pushing utilities will keep plugging into the grid while pushing utilities to add generation of every kind — nuclear’s share of the answer grows, but Deloitte’s 10%-by-2035 ceiling is the honest frame: nuclear is the premium firm layer inside a broader power plan.

The skeptics’ ledger: what could break the wave

Four risks deserve the words they get. Cost history: Western nuclear construction overruns are legendary; restarts (brownfield, proven assets) dodge some of it, but the SMR fleet must still prove factory economies. Contract duration: 20-year PPAs at 2026 prices embed assumptions nobody verifies for two decades — if AI capex cools, the buyer’s downside is a contract it cannot easily exit (the same capex-cycle risk our sovereign AI scoreboard flags for national programs). Public acceptance and policy: siting waste storage, securing licenses and sustaining political support across decades — the governance lane where ambitious programs fail. Opportunity cost: capital committed to reactors is capital not spent on transmission, efficiency or storage; the cheapest megawatt remains the one never needed, and the efficiency race (chips, cooling, interconnects) competes with every reactor dollar. Nuclear data centers are a real wave; this ledger separates a buildout from a bubble, and the honest position is that both outcomes remain possible at current spending rates.

WorldNgayon Insight: the power contract as the new competitive moat

What changed: nuclear data centers turned tech companies from compute bidders into firm-power bidders for FIRM POWER — the nuclear PPAs of 2025-2026 are the visible peak of that shift. Why now: gigawatt campuses met multi-year interconnection queues, and nuclear’s 24/7 profile became the only fit for the highest-density loads without new transmission. Why it lasts beyond the news cycle: the contracts run 20 years; every annual restart milestone, SMR delivery and PPA announcement is a named update trigger. Who is affected: utilities and their regulators, reactor developers, data-center operators — and ratepayers, whenever costs and contracts interact. What to do next: read the next “AI goes nuclear” headline through the scoreboard lens — buyer, counterparty, megawatts, duration, and who bears the overrun. Risks, opportunities, costs, skills: the risk is decade-long contract exposure on both sides; the opportunity is a firm-power layer AI alone cannot build; the cost shows in PPA rates and public balance sheets; the skills gap runs through nuclear operations, licensing and grid engineering — plus the data-center trades the region’s buildouts hire. Connection to work and money: every gigawatt contracted is a jobs ledger — reactor crews, grid engineers and the nuclear data centers trades our careers analysis tracks — and the Philippines’ energy-cost reality makes the firm-power question a household economics issue too.

For Filipino readers: what nuclear data centers mean here

The Philippines’ nuclear story is cautionary and current at once. The Bataan Nuclear Power Plant was completed in 1984 at over $2 billion and never fueled — a political, governance and financing failure, not an engineering one — and it remains the country’s most-cited nuclear datum. Fast-forward: the energy-cost pressure on Philippine industry, the rising nuclear data centers corridor demand (our water-wall analysis maps the regional build), and the global buyer logic of firm power put nuclear back on the policy table — with the correct order of operations now written by other countries’ lessons: grid readiness and governance frameworks first, reactor economics second, procurement third. For OFW professionals, the wave is a jobs map before it is a power map: nuclear operations, grid engineering, and the data-center trades are hiring across the countries that host this buildout — and Filipino engineers already hold strong positions in the region’s energy and construction sectors. The metal-and-megawatt lesson from our copper analysis applies to reactors too: capacity follows skills, not slogans.

Which companies have signed nuclear power deals for data centers?

The verified scoreboard: Microsoft (20-year, ~$16B PPA for the full 835MW of the restarted Three Mile Island Unit 1, now the Crane Clean Energy Center), Google (master agreement with Kairos Power for up to 500MW of SMRs, plus three planned ~600MW nuclear projects), Amazon ($20B Susquehanna-linked capacity commitment), Meta (20-year, 1.1GW-class Constellation Clinton PPA), and Oracle (a 1GW datacenter planned on a trio of nuclear reactors). The US Department of Energy backs the restart wave with up to $4.4 billion in loans.

Why are nuclear data centers buying nuclear power?

Nuclear data centers exist because AI campuses need hundreds of megawatts of continuous firm power; grid interconnection queues run for years, and intermittent sources fit only part of the load. Nuclear’s 24/7 dense profile matches a training campus’s load factor — buying the plant’s output via PPA locks decades of firm supply without owning the asset, and the grid reality makes the premium price rational.

What is the Three Mile Island restart?

Constellation is restarting Three Mile Island Unit 1 as the Crane Clean Energy Center under Microsoft’s 20-year, ~$16 billion power purchase agreement for the full 835 megawatts — the flagship of the US restart wave, which now covers three shuttered plants with federal loan backstops up to $4.4 billion. Unit 2 (the 1979 accident unit) is not part of the restart.

When will SMRs power data centers?

The honest timeline: initial SMR deliveries for data-center buyers mostly land in the early 2030s. The 2026-era contracts (Google-Kairos, Google’s 600MW projects) are tranche-structured fleet orders precisely because fleet-scale SMR economics remain unproven; restarts are the route that adds firm nuclear power this decade.

How much demand can nuclear data centers cover?

Deloitte’s analysis caps nuclear’s share at up to 10% of data-center electricity demand by 2035 — a premium firm layer inside a broader power plan that still leans on the grid, gas, renewables and storage. The headline deals are contract-visible; nuclear remains a minority but strategic contributor to total data-center power.

What are the risks of nuclear data centers deals?

Four named risks: the cost-overrun history of nuclear construction (restarts mitigate; SMR fleets await factory-economy proof), the rigidity of 20-year PPAs if AI capex cools, the public-acceptance and governance lane (siting, licensing, waste), and opportunity cost — capital and attention diverted from transmission, efficiency and storage alternatives.

What happened with the Bataan Nuclear Power Plant?

The Philippines’ Bataan Nuclear Power Plant was completed in 1984 at a cost of over $2 billion and never fueled: governance changes, financing disputes and the post-Chernobyl climate ended the program. It stands as the region’s clearest lesson that nuclear ambition succeeds only with governance frameworks, financing discipline and public trust in place first — the order today’s nuclear data center buyers now follow.

Do nuclear data centers help or hurt the grid?

Done right, they help: restarted and contracted reactors add firm, carbon-free capacity to the grid mix and relieve the interconnection pressure on new AI campuses. The honest caveat is allocation — ratepayers and regulators scrutinize who finally bears the contract’s cost risk, which is why PPA terms, not press releases, are where the public interest gets decided.

Financial Disclaimer

This article is editorial analysis of publicly reported commercial contracts and market forecasts, not investment, energy or procurement advice, and no return or price direction is promised. Deal terms (values, durations, megawatts) reflect public reporting as dated; forecasts (share of demand, delivery timelines) are the named institutions’ projections and carry revision risk. Nuclear, utility and energy-sector positions involve long-duration risk; verify current terms with the counterparties and named agencies before any financial decision.

Sources: Trellis — Amazon, Google, Meta and Microsoft go nuclear (with Deloitte’s 10%-by-2035 forecast). Additional context cited without links (bot-walled or dynamic endpoints): S&P Global Commodity Insights via SDxCentral reporting (copper-to-reactor supply-chain link); industry deal trackers (Microsoft-Constellation, Google-Kairos, Amazon-Talen, Oracle campus terms) as reported across 2025-2026 trade press; DOE loan program announcements. Vendor contract claims are labeled as reported terms until counterparty filings confirm them.