In the space of about two years, every major hyperscaler became a nuclear power buyer. Microsoft is restarting Three Mile Island. Google, Amazon, and Meta have committed gigawatts more across restarts and next-generation reactors. It is the most striking energy story in technology, and beneath the headlines it carries a quieter signal about the permanence of the AI buildout and the future shape of the infrastructure that will eventually retire. Here is the 2026 picture.
TL;DR
The AI industry has become nuclear power’s most enthusiastic new customer, and the scale of commitment is remarkable:
- Every major hyperscaler has signed at least one nuclear deal. As of 2026, tracking services counted roughly 13 announced projects committing over 9.8 GW of nuclear capacity to AI data center infrastructure.
- Microsoft is restarting Three Mile Island. Its 20-year, roughly $16 billion power purchase agreement will restart the former TMI Unit 1 (rebranded the Crane Clean Energy Center) for 835 MW, with commercial operation targeted for the second half of 2027.
- The strategy is two-pronged. Hyperscalers are securing immediate power from existing nuclear plants (restarts and PPAs) while underwriting next-generation small modular reactors (SMRs) for the early 2030s.
- The driver is physics. A single NVIDIA H100 draws roughly 700 watts, a rack of eight exceeds an average home, and a 100,000-GPU cluster draws around 100 MW. Nuclear offers carbon-free, 24/7 baseload power that wind and solar cannot match for constant AI loads.
- Power certainty is now the primary differentiator. Nuclear-backed sites reportedly command lease premiums of 15 to 25 percent over grid-constrained alternatives, and site selection increasingly considers nuclear proximity.
The infrastructure angle that matters beneath the energy headlines: 20-year power purchase agreements and multi-billion-dollar plant restarts are long-duration bets that signal AI demand is expected to persist for decades. That permanence reshapes where data centers get built and how they are planned, which in turn reshapes where and how the hardware inside them will eventually be retired. And a timeline mismatch sits at the center of it, because the reactors arrive years after the hardware they will power, which will refresh several times over before the first nuclear electron flows.
Note: This article is educational and covers a fast-moving energy and infrastructure topic. It is not investment or financial advice, and ROC Telecom is not a financial advisor. Details of announced nuclear projects change frequently; the specifics here reflect our understanding as of mid-2026.
Why AI Turned to Nuclear
The nuclear rush is a direct consequence of the physics of AI computing colliding with the limits of the grid.
The Power Problem
AI workloads are extraordinarily power-dense. A single NVIDIA H100 GPU draws roughly 700 watts, a rack of eight of them draws more than an average American home, and a hyperscale training cluster of 100,000 GPUs draws on the order of 100 MW, enough for roughly 80,000 homes. As these clusters multiply, data center electricity demand has surged. Data centers reached an estimated 4 percent of US power usage, a figure that could more than double by 2030, and in hubs like Northern Virginia they already consume more than a quarter of regional electricity.
That demand runs into a hard reality: the grid cannot deliver new power fast enough, and in constrained markets there is simply not enough of it. Hyperscalers needed a new source of large-scale, reliable power, and they needed it on a timeline the traditional grid could not meet.
Why Nuclear Specifically
Nuclear answers a requirement that renewables struggle to meet: constant, carbon-free baseload power. AI training and inference run 24 hours a day, and a data center needs power that is always on. Solar and wind, valuable as they are, run at capacity factors around 25 to 35 percent and depend on weather and time of day. Nuclear runs at capacity factors above 90 percent, delivering steady output around the clock. For an operator committed to carbon-free power but needing constant supply, nuclear is one of the few options that fits, and unlike a new gas plant it aligns with the sustainability commitments these companies have made.
The economics only work under specific conditions. Nuclear is expensive to build, with costs far above natural gas per kilowatt, so the deals pencil primarily where carbon-free requirements and the premium on power certainty justify the cost. For hyperscalers racing to secure guaranteed power for decades-long AI ambitions, that calculation increasingly favors nuclear.
The Deals Reshaping the Landscape
The nuclear commitments fall into two categories: immediate power from existing plants, and future power from next-generation reactors.
| Deal | Approach | Capacity | Timeline |
|---|---|---|---|
| Microsoft / Three Mile Island (Crane Clean Energy Center) | Restart of a shut-down reactor via 20-year PPA with Constellation | 835 MW | Commercial operation targeted second half of 2027 |
| Amazon / Talen (Susquehanna) | PPA drawing power from an existing plant for a co-located campus | Up to ~960 MW campus | Near-term, phased |
| Google / Kairos Power | First US corporate agreement to develop a fleet of SMRs | Up to 500 MW across six to seven units | First reactor targeted around 2030 |
| Amazon / X-energy | Investment in SMR developer for future units | Up to ~300+ MW (multiple Xe-100 units) | Early 2030s |
| Meta / multiple (TerraPower, Oklo, Vistra, Constellation) | Largest total commitment across restarts and advanced reactors | Up to several GW | Longest horizon, roughly 2032 to 2035 |
The pattern is clear. The first wave secures power fast by restarting or contracting existing reactors, which is why Microsoft’s Three Mile Island deal will deliver the first nuclear electrons to an AI data center around 2027. The second wave, the SMR commitments from Google, Amazon, Meta and others, underwrites next-generation reactor technology that will not deliver commercial power until the early 2030s. Hyperscalers are not just buying power, they are actively underwriting the future of nuclear technology to secure long-term supply.
The Symbolism of Three Mile Island
Microsoft’s restart of Three Mile Island Unit 1 is worth pausing on, because the symbolism is hard to miss. This is the site that became synonymous with nuclear failure after the 1979 partial meltdown of its other unit. Unit 1, a separate and undamaged reactor that operated safely until it was shut for economic reasons in 2019, is now being restarted specifically to power AI, rebranded the Crane Clean Energy Center. A facility that once symbolized nuclear retreat has become a symbol of AI’s power hunger and nuclear’s revival. Few developments capture the moment more vividly.
Why Nuclear Is Reshaping Where Data Centers Get Built
The most consequential effect of the nuclear rush, for infrastructure planning, is that it changes the geography of the data center industry.
For years, data center site selection prioritized connectivity, land, tax incentives, and grid access. Power availability was a factor, but rarely the deciding one. That has changed. Power certainty is now the primary differentiator in data center site selection, and nuclear-backed sites represent a scarce and valuable form of that certainty. Reporting on current transactions suggests nuclear-adjacent or power-certain sites command lease premiums of 15 to 25 percent over comparable grid-constrained alternatives.
This reshapes where capacity gets built. Sites near existing nuclear plants, or positioned for future SMR co-location, gain a strategic advantage. Amazon’s investment in a campus adjacent to the Susquehanna plant and the emerging model of co-locating SMRs directly with data centers point toward a future where reactors and data centers are planned together. As power becomes the binding constraint everywhere, the industry is reorganizing around where reliable power can be secured, and nuclear is becoming one of the anchors of that new geography.
For infrastructure planning, this matters because where data centers are built determines where they are eventually retired. A shift toward nuclear-anchored sites, potentially in new locations chosen for power rather than traditional connectivity hubs, gradually redistributes the future map of decommissioning and asset recovery demand.
The Timeline Mismatch That Matters for Hardware
Here is the dimension of the nuclear story that the energy coverage rarely connects to the hardware, and it is the one most relevant to infrastructure lifecycle: the reactors and the chips run on completely different clocks.
The nuclear power arriving to serve AI is mostly years away. Microsoft’s Three Mile Island restart, the earliest major deliverable, targets late 2027. The SMR commitments from Google, Amazon, and Meta largely target the early-to-mid 2030s. These are long-horizon infrastructure projects, appropriately so, since they are meant to power AI for decades.
The AI hardware, by contrast, runs on a compressed refresh cycle measured in a few years. NVIDIA ships a new architecture on roughly an annual cadence, and operators typically refresh high-density AI hardware every three to four years. That means the GPUs installed today will be retired, and likely refreshed two or three times over, before much of the contracted nuclear power even comes online.
The implication is striking. The nuclear buildout is being planned to power generations of AI hardware that do not exist yet, while the hardware running in these facilities today will cycle through disposition long before the reactors are switched on. The power infrastructure and the compute infrastructure are on fundamentally different timelines, which means the retirement and refresh of AI hardware is a continuous, near-term reality unfolding entirely in the years before the nuclear power is even available. For operators, the hardware lifecycle question is immediate and recurring, regardless of how long-dated the power strategy is.
What the Nuclear Rush Signals About the Buildout
Beyond the geography and the timelines, the nuclear commitments carry a signal worth reading carefully, especially alongside the ongoing debate over whether AI infrastructure is overbuilt.
Twenty-year power purchase agreements and multi-billion-dollar plant restarts are among the longest-duration bets a company can make. Microsoft’s 20-year PPA duration significantly exceeds typical solar or wind agreements, and the willingness to underwrite plant restarts and new reactor development signals a conviction that AI electricity demand will persist for decades. In the framing of the stranded-asset debate, the nuclear commitments are evidence for the bull case: companies do not sign 20-year power contracts and restart nuclear plants for a demand they expect to evaporate.
At the same time, honesty requires noting the counterpoint. Long-dated power contracts express confidence in aggregate, long-run AI electricity demand, which is a different question from whether any particular generation of GPU hardware holds its value, or whether the current pace of capital spending is sustainable in the near term. A company can be simultaneously convinced that AI will consume enormous power for decades (justifying nuclear) and exposed to the risk that specific hardware vintages strand quickly (the depreciation concern). The nuclear bet and the depreciation debate are about different time horizons and different assets. (For the disposition side of that debate, see our analysis of the AI bubble and the stranded-asset question.)
What the nuclear rush establishes with reasonable confidence is that the major players are planning for AI to be a permanent, power-intensive fixture of the economy. That permanence is precisely what makes the hardware lifecycle, the continuous refresh and retirement of AI compute, a durable and growing reality rather than a passing wave.
What This Means for Infrastructure Retirement
The nuclear story connects to infrastructure retirement in three concrete ways, even though reactors and asset disposition might seem worlds apart.
The Buildout Is Being Planned as Permanent
The nuclear commitments signal that AI infrastructure is being built for the long haul, which means the associated hardware refresh and retirement is a permanent, recurring function, not a one-time event. Operators planning decades-long power strategies are, implicitly, planning decades of hardware cycling through their facilities. Disposition is part of that permanent operational picture.
The Geography Is Shifting
As nuclear-anchored and power-certain sites gain advantage, the map of where data centers are built is gradually redrawing around power availability. Over time, this redistributes where decommissioning, asset recovery, and recycling demand will concentrate, potentially into new regions selected for power rather than the traditional connectivity hubs. Nationwide disposition capability, rather than dependence on proximity to legacy hubs, becomes more valuable as the geography spreads.
The Hardware Cycle Runs Ahead of the Power
Because the nuclear power is years away while the hardware refreshes continuously, the immediate reality for operators is a steady stream of AI hardware reaching end of life now, well before the reactors arrive. That ongoing retirement, of high-density, high-value GPU and networking equipment, requires certified data destruction, informed asset recovery to capture residual value, and responsible recycling for genuine end-of-life material. The long-dated power strategy does not change the near-term, recurring need to handle retiring hardware well.
The nuclear rush is, at its heart, a bet on permanence. And a permanent, power-intensive AI infrastructure is one in which handling the continuous retirement of hardware well, capturing its value and managing its data and materials responsibly, becomes a lasting part of running the operation.
Frequently Asked Questions
The following is general educational information about a fast-moving energy topic, not investment or financial advice.
Why are data centers turning to nuclear power?
AI data centers are turning to nuclear power because AI workloads are extraordinarily power-dense and require constant, reliable electricity that the grid increasingly cannot supply fast enough. A single NVIDIA H100 GPU draws roughly 700 watts, and a 100,000-GPU cluster draws around 100 MW. Nuclear offers carbon-free baseload power at capacity factors above 90 percent, running 24 hours a day, which matches the always-on nature of AI training and inference better than weather-dependent solar and wind. For hyperscalers with carbon-free commitments and decades-long AI ambitions, nuclear provides the guaranteed, round-the-clock power their facilities need.
What is the Microsoft Three Mile Island deal?
Microsoft signed a 20-year power purchase agreement, worth roughly $16 billion, to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania, rebranded the Crane Clean Energy Center. The deal provides 835 MW of carbon-free power for Microsoft’s AI and cloud operations, with commercial operation targeted for the second half of 2027. Notably, this is the site of the 1979 partial meltdown, though that involved a different unit (Unit 2); Unit 1 operated safely until it was shut for economic reasons in 2019. The restart is the clearest symbol of AI’s power hunger reviving the nuclear industry, and it will deliver some of the first nuclear electrons dedicated to an AI data center.
What are small modular reactors (SMRs)?
Small modular reactors are compact nuclear reactors that produce up to about 300 MW of power, compared with roughly 1,000 MW or more for a traditional reactor. They are designed to be scalable, require less land (around 50 acres), and can be built faster and sited more flexibly than conventional plants, including potential co-location directly with data centers. Their high capacity factor and ability to provide constant, carbon-free power make them attractive for AI workloads. However, commercial SMR deployment is largely expected in the early-to-mid 2030s, and challenges including limited specialized fuel supply and a thin nuclear-engineering talent pool continue to affect timelines.
Which companies have signed nuclear deals for AI?
As of 2026, every major US hyperscaler had signed at least one nuclear deal, with tracking services counting roughly 13 announced projects committing over 9.8 GW of capacity. Microsoft secured 835 MW through the Three Mile Island restart. Google committed to up to 500 MW of small modular reactors with Kairos Power. Amazon draws power from the Susquehanna plant via Talen and invested in SMR developer X-energy. Meta made the largest total commitment, up to several gigawatts across TerraPower, Oklo, Vistra, and Constellation, though on the longest timeline. The commitments split between restarting or contracting existing plants for near-term power and underwriting next-generation reactors for the 2030s.
When will nuclear power actually reach AI data centers?
The timelines vary significantly. The earliest major deliverable is Microsoft’s Three Mile Island restart, targeting commercial operation in the second half of 2027, because restarting an existing reactor is faster than building new. The small modular reactor commitments from Google, Amazon, and Meta largely target the early-to-mid 2030s, since SMRs still require regulatory approval, construction, and commissioning. This means most of the contracted nuclear power is years away. It is an important point for infrastructure planning: the AI hardware running in these facilities today, which refreshes every three to four years, will cycle through retirement and replacement well before much of the nuclear power comes online.
How does the nuclear buildout relate to the AI bubble debate?
The nuclear commitments are often read as evidence for the bull case in the AI-overbuild debate: companies do not sign 20-year power purchase agreements and restart nuclear plants for demand they expect to disappear, so the bets signal conviction that AI electricity demand will persist for decades. However, this expresses confidence in aggregate, long-run power demand, which is a separate question from whether any specific generation of GPU hardware holds its value, or whether the near-term pace of capital spending is sustainable. A company can rationally be confident that AI will consume enormous power for decades while still facing the risk that particular hardware vintages depreciate quickly. The nuclear bet and the hardware-depreciation debate concern different time horizons and different assets.
How does nuclear power affect where data centers are built?
Nuclear is reshaping data center geography because power certainty has become the primary differentiator in site selection. Nuclear-backed and power-certain sites reportedly command lease premiums of 15 to 25 percent over grid-constrained alternatives, and operators increasingly consider nuclear proximity when choosing locations. This favors sites near existing nuclear plants or positioned for future SMR co-location, and points toward a model where reactors and data centers are planned together. Over time, as capacity concentrates around where reliable power can be secured, the geography of the industry redistributes, which also redistributes where future decommissioning and asset recovery demand will concentrate.
What does the nuclear rush mean for AI hardware retirement?
It reinforces that AI hardware retirement is a permanent, recurring function rather than a passing wave. The nuclear commitments signal that AI infrastructure is being planned for decades, which means the continuous refresh and retirement of AI hardware is a durable, growing reality. Practically, because the nuclear power is mostly years away while the hardware refreshes every three to four years, operators face a steady, near-term stream of high-density GPU and networking equipment reaching end of life now, requiring certified data destruction, informed asset recovery to capture residual value, and responsible recycling. As the industry’s geography shifts toward power-anchored sites, nationwide disposition capability becomes increasingly valuable.
Is nuclear power for data centers actually carbon-free?
Nuclear power generation produces electricity without direct carbon emissions, which is why it appeals to hyperscalers with carbon-free and sustainability commitments, and why it is attractive for the constant baseload that AI requires but solar and wind cannot continuously provide. This is a primary reason operators favor nuclear over natural gas for round-the-clock AI power despite nuclear’s higher cost. That said, full lifecycle considerations (construction, fuel processing, and long-term waste management) are part of the broader nuclear discussion. For the specific need of constant, carbon-free electricity to power AI workloads without the intermittency of renewables, nuclear is one of the few options that fits, which is central to why the hyperscalers have turned to it.
The Bottom Line
In roughly two years, every major hyperscaler became a nuclear power buyer, committing to more than 9.8 GW across 13 projects, from Microsoft’s landmark restart of Three Mile Island to the small modular reactor fleets that Google, Amazon, and Meta are underwriting for the 2030s. The driver is the physics of AI computing: power-dense GPU clusters need constant, carbon-free electricity that the grid cannot supply fast enough, and nuclear is one of the few sources that fits. Power certainty has become the primary differentiator in the industry, reshaping where data centers get built and reorganizing the map around reliable supply.
For infrastructure, the deeper signal is permanence. Twenty-year power agreements and multi-billion-dollar plant restarts are bets that AI will be a lasting, power-intensive fixture of the economy. That permanence makes the continuous refresh and retirement of AI hardware a durable, growing reality rather than a passing wave. And a timeline mismatch sits at the center of it: the nuclear power is mostly years away, while the hardware inside these facilities refreshes every three to four years, cycling through retirement long before the reactors switch on. The power strategy is long-dated, but the hardware lifecycle is immediate and recurring, which makes handling the retirement of AI infrastructure well, capturing its value and managing its data and materials responsibly, a lasting part of the operation.
How ROC Telecom Helps
ROC Telecom is an R2v3, RIOS, NIST 800-88, and ITAR-compliant ITAD specialist built for the continuous hardware lifecycle beneath the AI buildout:
- Nationwide decommissioning and asset recovery, suited to an industry whose geography is redistributing around power availability rather than legacy hubs
- GPU and AI infrastructure asset recovery for the high-density hardware that refreshes every three to four years, well ahead of the long-dated power infrastructure, with speed-to-remarketing that protects value against generational decay
- Specialist asset recovery across routing, switching, optical transport, and compute with direct buyer relationships
- NIST 800-88 data destruction with per-asset serialized Certificates of Destruction and full chain-of-custody documentation
- R2v3 Appendix E materials recovery with in-house dismantling and direct-to-refiner processing for genuine end-of-life hardware
- Mass-balance recovery reporting for the ESG and sustainability disclosures that carbon-conscious operators increasingly require
15+ years of ITAD experience, $25M+ in client capital recovered, 45M+ pounds diverted from landfill.
ROC Telecom does not provide investment or financial advice. We help operators handle the continuous retirement of AI hardware that a permanent, power-intensive buildout generates.
Request a Free AI Infrastructure Disposition Assessment
Tell us about the AI infrastructure you are retiring or refreshing. A specialist will discuss recovery value, secure data destruction, and certified disposition, wherever your facilities are located. No commitment, no spam. Prefer to talk directly? Call 585-406-1249 or email info@roctelecom.com.
"*" indicates required fields
Related reading:
- The AI Bubble Debate and the Stranded-Asset Question
- Why AI Is Shortening Your Hardware Lifecycles (and What to Do About It)
- How GPU Decommissioning Differs from Standard Server Retirement
- Data Center Asset Recovery: The 2026 Strategic Guide
- The Northern Virginia Data Center Market: A 2026 Overview
- Hyperscale ITAD vs Enterprise ITAD: What’s Actually Different in 2026
- The Phoenix Data Center Market: A 2026 Overview
