Why Big Tech Is Betting Billions on Nuclear Power for AI Data Centers

Direct answer: Meta, Microsoft, Amazon, and Google have collectively committed more than $50 billion to nuclear power projects for AI data centers as of early 2026, betting on small modular reactors — factory-built units in the 5-to-300-megawatt range that can be constructed in 3 to 5 years, versus 7 to 10 years for a traditional large reactor — as a way to bypass grid interconnection queues that currently run 5 to 10 years in many US regions. China is moving in parallel at national scale: its Linglong One reactor is scheduled to begin commercial operation in the first half of 2026 as the world's first commercial onshore small modular reactor. It matters right now because AI data centers need a scale and reliability of power that most existing grids were never built to deliver quickly, and nuclear is the option enough major buyers now believe can actually close that gap on a usable timeline.

Fifty Billion Dollars, and a Bet on Reactors That Don't Exist Yet at Scale

Four of the largest technology companies in the world — Meta, Microsoft, Amazon, and Google — have collectively committed more than $50 billion to nuclear power projects tied to AI data center operations, as of early 2026. That is a genuinely large number for an energy technology that, in its small modular form, is still mostly unproven at commercial scale outside a handful of pilot and early-deployment projects. It's worth sitting with that tension for a moment: some of the most sophisticated capital allocators in the world are betting tens of billions of dollars on a class of reactor that, in most of these deals, hasn't yet been built and switched on at the scale being promised.

The logic behind the bet is straightforward even if the execution is genuinely uncertain. Small modular reactors, generally in the 5-to-300-megawatt range, are factory-built rather than constructed entirely on-site, which is the single biggest reason they're projected to take 3 to 5 years to build compared with 7 to 10 years for a traditional large nuclear plant. For a data center operator watching grid interconnection queues stretch 5 to 10 years in many constrained US regions, a reactor that could plausibly be operational in half that time, dedicated to a single facility rather than shared across an entire regional grid, is an attractive way to route around a queue rather than wait in it. Whether that plausible timeline survives contact with real permitting, fuel supply, and construction schedules is the open question running underneath almost everything else in this piece.

Why Nuclear, Specifically, for AI Data Centers

AI data centers have a fundamentally different power profile than the data centers that came before them, and that difference is most of the reason nuclear has entered the conversation at all. An AI-focused facility can require 80 megawatts or more of continuous power, more than double the roughly 32 megawatts a standard data center typically needs — and unlike a lot of general computing workload, AI training and inference tend to run GPUs at sustained, near-maximum utilization for long stretches rather than in bursty, variable patterns. That combination — much higher baseline demand, running continuously, with limited tolerance for interruption — is exactly the load profile nuclear power is best suited to serve, and exactly the profile wind and solar, with their weather-dependent intermittency, struggle to serve reliably on their own.

Nuclear's advantage here isn't primarily about carbon accounting, even though that's part of the pitch companies make publicly. It's about the physical shape of the demand curve. Solar generates during daylight hours and drops to zero at night; wind varies with weather in ways that don't reliably track a data center's actual consumption pattern. A gigawatt-scale AI training cluster running continuously needs firm, dispatchable, always-available power — the kind a nuclear reactor, once operating, can provide at a stable output level for years at a stretch without the variability that comes with weather-dependent generation. That's the specific gap Big Tech's nuclear bets are trying to close, distinct from (though often bundled together with) the broader corporate sustainability case for choosing a zero-carbon power source.

Baseload Power: The Term Behind Every One of These Deals

Baseload power is the minimum, continuous level of electricity demand a system needs met at essentially all times, and it's the single term that best explains why nuclear specifically, rather than any other single power source, keeps showing up across this entire trend. A grid, or a single large facility acting like its own micro-grid, needs its baseload covered by something that doesn't switch off overnight or drop out when the wind stops; everything above that baseline can be covered more flexibly, including by intermittent renewables, because short-term gaps above the baseline are easier to manage with storage, demand response, or backup generation than a gap in the baseline itself.

AI data centers have unusually high, unusually flat demand curves compared with most other large electricity consumers — a training cluster running continuously doesn't show the same daily or seasonal demand swings a typical commercial building, or even a conventional data center, might show. That makes almost their entire power draw look like baseload, rather than a smaller baseline topped up with more flexible peak demand, which is exactly why baseload-capable generation, historically nuclear, hydro, or fossil fuel plants, keeps coming up as the serious answer to AI power demand in a way it doesn't for other, more variable industrial electricity users. It also explains why the debate rarely frames nuclear against renewables as a straight substitute-for-substitute comparison — the honest comparison is nuclear or gas covering the baseload while renewables and storage handle everything layered on top, which is precisely the blended strategy Stargate UAE and several of the other projects covered in this piece are pursuing rather than picking one power source exclusively.

The Small Modular Reactor Pitch: Faster, Smaller, Built Differently

Small modular reactors are the specific technology carrying most of this bet, and the pitch behind them is genuinely different from how nuclear power has traditionally been built.

	Traditional large reactor	Small modular reactor (SMR)
Typical capacity	Often 1,000MW or more	Roughly 5MW to 300MW
Construction approach	Substantially built on-site	Factory-built in modules, assembled on-site
Typical construction timeline	7 to 10 years	3 to 5 years
Typical buyer	Utility serving a broad regional grid	Increasingly, a single large industrial user (e.g., an AI data center operator)
Deployment model	One large plant, one site	Multiple smaller units, potentially added incrementally as demand grows

That factory-built approach is the crux of the timeline advantage: instead of a bespoke construction project built entirely on-site, with all the site-specific engineering, permitting, and labor-scheduling risk that implies, key reactor components are manufactured in a controlled factory environment and shipped for on-site assembly, closer in spirit to how modern shipbuilding or modular construction works than to how nuclear plants have traditionally been built. The other genuinely new feature is the buyer profile: SMRs are increasingly being purchased not by a utility serving an entire regional grid, but by a single large industrial customer — an AI data center operator — dedicating a reactor's output to its own facility rather than feeding it into the shared grid. That direct-purchase model is part of why the current SMR wave is so closely tied to AI data center demand specifically, rather than being a broader utility-sector story.

The Interconnection Queue Problem Nuclear Is Trying to Solve

The single biggest practical driver behind this entire trend is grid interconnection delay. In many grid-constrained US regions, the queue to connect a large new power consumer to the existing grid now runs 5 to 10 years — a direct function of how much new demand, AI data centers prominent among it, is trying to connect to grids that were sized and planned around a very different demand forecast. Northern Virginia, Oregon, and Texas are specifically cited as regions under this kind of grid stress, all three being existing data center hubs where AI-driven demand growth is colliding hardest with existing interconnection capacity.

A dedicated SMR, built specifically to power a single facility, is one of the few available ways to sidestep that queue almost entirely — if a data center operator owns or contracts power directly from a reactor built on or near its own site, it isn't waiting in line for a shared grid connection at all, at least for that facility's own generation needs. That's the practical, unglamorous core of the pitch beneath all the nuclear-for-AI headlines: this is at least as much a queue-avoidance strategy as it is an energy-strategy or sustainability one, and the 5-to-10-year interconnection delay it's trying to route around is arguably the single most important number in this entire trend, more than any specific reactor design or company announcement.

Who's Placing the Bets

Microsoft, Amazon, Google, and Meta are all pursuing nuclear or SMR deals as of early 2026, and while the specifics of each company's approach vary, the underlying motivation across all four is the same: guaranteed, dedicated, always-on power for AI infrastructure that doesn't depend on winning a slow-moving grid interconnection queue. Amazon's approach includes a direct investment in X-Energy, a small modular reactor developer, as part of AWS's exploration of nuclear-powered cloud campuses — a structure that gives Amazon a more direct stake in how quickly and successfully that specific reactor technology gets built and deployed, rather than purely acting as an end customer signing a power purchase agreement after the fact.

The nuclear power purchase agreement has become the standard commercial structure underpinning most of these deals: rather than owning a reactor outright, a technology company commits to buying a reactor's output over a long contract term, giving the reactor developer the revenue certainty needed to secure financing and construction, while giving the technology company a guaranteed, contracted power supply without taking on the full construction and operating risk of owning a nuclear facility directly. That structure is part of why more than $50 billion in commitments could accumulate across just four companies relatively quickly — a power purchase agreement is a financial commitment that can be signed well before a single reactor component is manufactured, let alone installed.

Microsoft, Google, and Meta each frame their own commitments slightly differently in public communications, but the underlying commercial mechanics tend to converge on the same handful of structures: a long-term power purchase agreement with a reactor developer or utility, an equity or debt investment in the company building the reactor, or, in some cases, a direct role in restarting or extending the life of existing nuclear capacity rather than only backing brand-new small modular designs. Which structure a given company favors says something about its risk appetite — an equity stake carries more upside, and more exposure if the reactor developer stumbles, while a pure power purchase agreement is a cleaner, lower-risk commitment to buy output once it exists, with less influence over how quickly that output actually arrives.

Is This Substance, or Big Tech PR Getting Ahead of Reality?

Not everyone is convinced this wave of announcements will deliver what it promises on the timeline being promoted, and that skepticism deserves a fair hearing rather than being waved off. The Bulletin of the Atomic Scientists has published direct pushback on the framing around next-generation nuclear for data centers, cautioning against taking Big Tech's PR at face value. The core of that skepticism is straightforward: announcing a nuclear power purchase agreement or an equity investment in a reactor developer is a relatively low-cost, high-visibility move a technology company can make quickly, while actually designing, licensing, manufacturing, and commissioning a working reactor is a multi-year undertaking subject to regulatory review, supply-chain constraints, and construction risk that a press release can't shortcut.

That gap between announcement and delivery is exactly what critics point to when they argue nuclear-for-AI commitments often outpace what utilities and reactor developers can realistically deliver. A signed agreement is a real commercial commitment, but it is not the same thing as megawatts flowing to a data center, and the nuclear industry's own history includes plenty of projects that took considerably longer, and cost considerably more, than their original announcements suggested. None of this means the underlying bet is wrong — it means the honest way to read every announcement in this space in 2026 is as a stated intention backed by real capital, not as a guarantee that the promised timeline will hold.

The Global Picture: A Trend That Looks Very Different by Country

Nuclear power for AI data centers isn't a uniform global story — it's dominated by two very different models, US corporate power purchase agreements on one side and Chinese state-directed nuclear expansion on the other, with most other regions covered in this research sitting well behind both.

United States

The US is the epicenter of the corporate-nuclear-deal version of this story. Microsoft, Amazon (via its X-Energy investment), Google, and Meta are all pursuing nuclear or SMR deals, driven specifically by grid stress in data center hubs like Northern Virginia, Oregon, and Texas. On the deployment side, Crusoe and Aalo Atomics are targeting a nuclear-powered "AI Factory" proof of concept at Idaho National Laboratory in 2027, with full commercial deployment targeted for 2029 — a concrete, dated project worth watching as an early real-world test of whether the SMR-for-AI model can actually be delivered on the timelines its backers are promising.

United Kingdom

No distinct regional-specific reporting connecting UK small modular reactor policy directly to AI data centers turned up in this research pass. The UK does have separate, real momentum in SMR development more broadly, notably through Rolls-Royce SMR, but that program isn't documented here as being specifically tied to powering AI infrastructure the way the US corporate deals or China's Linglong One are.

UAE / Dubai

The UAE offers a concrete, if limited, data point: Stargate UAE is explicitly planned to draw on a nuclear, solar, and natural gas power mix rather than any single source. That's a smaller, more specific detail than a dedicated national SMR-for-AI program, but it confirms nuclear is a genuine part of the power strategy behind one of the largest single AI compute buildouts covered anywhere in this research.

Australia

Australia is reported to be building its AI-driven grid capacity "without a nuclear option," relying instead on renewables and gas to meet growing data center demand. That's a notable and explicit contrast with the US and China, and it puts Australia's AI power strategy on a genuinely different track from the nuclear-heavy approach gaining momentum elsewhere — a real policy choice rather than an absence of relevant reporting.

Germany

No distinct regional-specific reporting tying nuclear power to AI data centers was found for Germany in this research pass. Germany's broader, well-documented data center investment activity (covered in more depth elsewhere) does not appear, in this research, to lean on nuclear power specifically as part of its strategy.

Europe / France

No distinct regional-specific reporting tying nuclear power specifically to AI data centers was found for France either, beyond general EU nuclear-energy policy debates that fall outside what this research captured. Given France's own long-standing, comparatively nuclear-heavy national grid, it would be reasonable to assume nuclear plays some role in powering the country's substantial AI compute buildout, but that's an inference rather than a documented finding specific to this topic.

China

China is running the largest, most state-directed version of this trend anywhere in the world. The country holds 112 operational or approved nuclear units, roughly 125 million kilowatts of combined capacity, the largest such total globally, with two new units starting construction and seven more completing before the end of 2026. Its Linglong One small modular reactor is scheduled to begin commercial operation in the first half of 2026, a genuine milestone as the world's first commercial onshore SMR. The scale of demand behind this buildout is significant in its own right: Chinese data center electricity demand is expected to double to 400 terawatt-hours by 2030, and per IEA-cited projections, China's data center electricity mix — currently dominated by coal, at roughly 70%, with renewables around 20% and nuclear around 10% — is expected to shift so that nuclear and renewables together supply nearly 60% of China's data center electricity by 2035. That is a materially faster and larger-scale nuclear buildout than anything described in the US corporate-deal model, run through direct state planning rather than a series of individual corporate power purchase agreements.

Regulation, Supply Chains, and the Parts of This Story That Move Slowly

Even a well-financed SMR project has to clear a regulatory path that doesn't move at venture-capital speed. In the US specifically, a small modular reactor built for an AI data center still needs Nuclear Regulatory Commission licensing, environmental review under the National Environmental Policy Act, compliance with state-level siting rules that vary considerably by location, and liability coverage under the Price-Anderson Act, the federal framework that has governed nuclear liability insurance for decades. None of these steps are unique to AI-driven demand — they're the same regulatory path any new nuclear generation has to clear — but they are a genuine, non-negotiable part of the timeline, and they're a large part of why critics are skeptical that even a "fast" 3-to-5-year SMR construction estimate reliably holds once realistic permitting and review time is factored in alongside pure construction time.

Supply chains add a second layer of realistic friction. Advanced reactor designs, including many SMR designs, depend on specific fuel types and enrichment levels that aren't yet produced at large industrial scale, and fuel availability for these next-generation designs is a genuine constraint that could slow deployment independent of financing or construction progress. A reactor design can be fully engineered and permitted and still be waiting on a fuel supply chain that hasn't scaled to match the sudden surge in demand this wave of AI-driven orders is creating — which is exactly the kind of unglamorous, unresolved bottleneck that tends to get left out of an announcement but ends up determining the real delivery timeline.

How Big Could This Market Actually Get

Despite the genuine execution risk, the projected scale of the SMR market reflects just how seriously this bet is being taken across the industry. IDTechEx forecasts the global SMR market reaching $53.8 billion by 2036, and nearly $300 billion by 2046 — a trajectory that assumes SMRs move well beyond AI data centers into broader industrial power use over that period. More than 40 gigawatts of SMR capacity is already positioned globally for industrial users, data centers among them, which is a meaningfully large pipeline even accounting for the reality that not every positioned gigawatt will be built on schedule, or built at all.

Reading that forecast alongside the skepticism covered above is the most balanced way to think about where this trend actually sits in 2026: the capital commitment, the positioned capacity, and the long-term market projections are all real and large, and at the same time, the regulatory, supply-chain, and construction risks that could push actual delivery well past current announced timelines are also real. Both things are true simultaneously, and the projects worth watching most closely over the next few years — Crusoe and Aalo Atomics' Idaho National Laboratory proof of concept chief among them — are exactly the ones that will show, in practice, which side of that tension wins out.

It's worth being precise about what "positioned" capacity actually means in that 40-gigawatt figure, because it's easy to read a pipeline number like that as though it were already built. Positioned capacity typically means a project has reached some combination of a signed agreement, an announced site, or an early development stage — real commitments, but not megawatts anyone can draw on yet. The gap between a 40-gigawatt pipeline and 40 gigawatts of operating capacity is exactly the gap the regulatory and supply-chain constraints covered above have to be worked through before it closes, and it's likely to close unevenly: some projects, particularly those backed by governments with direct control over permitting and fuel supply, like China's, are positioned to convert pipeline into operating capacity considerably faster than projects depending on a slower, multi-agency US regulatory path.

The Sustainability Framing, and Why It's Only Part of the Story

Nuclear power's near-zero operational carbon profile fits neatly into the corporate sustainability commitments most of these same technology companies have separately made, and that fit is a real, additional reason nuclear is an easier internal sell than, say, building new natural gas capacity would be. But treating the sustainability framing as the primary driver risks missing the more urgent, more concrete problem doing most of the actual work behind these deals: the interconnection queue. A technology company chasing only a lower-carbon power source, with no urgency around grid-connection timelines, would have more options available to it and less reason to move as fast or commit as much capital as this wave of nuclear deals reflects. The pairing of "zero-carbon" and "bypasses a multi-year queue" is what makes nuclear specifically, rather than any other clean generation source, the option so many of these companies are converging on at the same time.

What This Means Going Forward

For a business that depends on AI infrastructure but isn't itself in the energy business, the nuclear-for-AI trend is worth understanding as a signal about the underlying power constraint shaping the entire AI infrastructure market, not as an area to invest in directly. The core takeaway is that power availability, not just chip availability, is becoming a genuine bottleneck on how fast AI compute capacity can actually grow, and the operators willing to commit tens of billions to unconventional power sources are the ones most likely to have reliable capacity available on a predictable timeline over the next several years.

That's a reasonable factor to weigh when choosing which cloud or AI infrastructure provider to build on for a workload that needs to scale over a multi-year horizon: a provider actively securing its own dedicated power supply, through a nuclear power purchase agreement or otherwise, is signaling a level of long-term capacity planning that a provider purely reselling whatever grid-connected capacity it can access in the short term may not be able to match. It's also one more reason to build AI-dependent products so the underlying compute provider remains a swappable component rather than a permanent foundation — exactly the architectural discipline behind the AI agents and automation work we do at Scult, where designing for provider flexibility is treated as a baseline requirement rather than an afterthought, given how much of the infrastructure layer underneath any AI product is still this actively in flux. Businesses evaluating longer-term AI infrastructure partnerships, particularly ones with sustainability commitments of their own to account for, may also find it worth reviewing how a given provider's energy strategy holds up against their own reporting and compliance obligations — our compliance page is a reasonable starting point for how those obligations tend to break down by framework.

The honest summary of where this trend stands in 2026: the money is real, the physical need for firm, always-on power at gigawatt scale is real, and the skepticism about whether announced timelines will actually hold is also entirely reasonable. Nuclear power for AI data centers isn't hype in the sense of being unfounded — the demand problem it's trying to solve is genuine and well documented. Whether it delivers on the specific timelines being promised in 2026 is a separate question, and one that will be answered by which projects actually reach commercial operation, not by how many more billion-dollar agreements get announced between now and then.

What Businesses Want to Know About Nuclear-Powered AI Data Centers

Why are Big Tech companies investing in nuclear power and SMRs specifically for AI data centers?

Big Tech companies are investing in nuclear power and small modular reactors because AI data centers need a scale and reliability of electricity that existing grids often can't deliver on the timeline these companies want to grow. AI-focused facilities can require 80 megawatts or more of continuous, sustained power, more than double a standard data center's roughly 32-megawatt need, and that demand has to be available essentially around the clock rather than in the variable bursts a grid might otherwise accommodate more easily. Nuclear power offers firm, weather-independent, always-on baseload generation that matches that demand profile far better than solar or wind can alone, and small modular reactors specifically offer a construction timeline, 3 to 5 years versus 7 to 10 for a traditional large reactor, that's fast enough to plausibly matter for a company trying to bring new AI capacity online within the next few years rather than the next decade. Collectively, Meta, Microsoft, Amazon, and Google have committed more than $50 billion to this strategy as of early 2026.

Is Big Tech's next-gen nuclear push for AI data centers substance or PR hype?

It's genuinely both, and the honest answer depends on which part of the announcement you're evaluating. The underlying problem — AI data centers needing more firm, reliable power than many grids can currently deliver on a fast timeline — is real and well documented, and the capital being committed, more than $50 billion collectively across four companies, is real money attached to real contracts. The Bulletin of the Atomic Scientists has specifically cautioned against taking the framing at face value, though, pointing out that signing a nuclear power purchase agreement or making an equity investment is a fast, low-cost, high-visibility move, while actually licensing, building, and commissioning a working reactor takes years and is subject to regulatory and supply-chain risk a press release can't resolve. The fairest read is that the intent and the financial commitment are genuine, while the specific timelines being promoted deserve real skepticism until projects actually reach commercial operation.

How much have Meta, Microsoft, Amazon, and Google collectively committed to nuclear power for AI as of early 2026?

Meta, Microsoft, Amazon, and Google have collectively committed more than $50 billion to nuclear power projects tied to AI data center operations, as of early 2026. That figure spans a mix of structures across the four companies, including direct equity investments in reactor developers, like Amazon's stake in X-Energy, and nuclear power purchase agreements, where a company commits to buying a reactor's output over a long contract term without owning the facility outright. The scale of that combined commitment is itself a significant signal about how seriously these companies view power availability as a constraint on their AI growth plans — $50 billion is a substantial sum to commit to a power source that, in its small modular form, is still largely unproven at commercial scale, and it reflects genuine urgency about securing dedicated capacity rather than a speculative side bet.

What is a small modular reactor (SMR), and how does its build time compare to a traditional large nuclear reactor?

A small modular reactor is a nuclear reactor generally in the 5-to-300-megawatt range, built substantially in a factory and shipped for on-site assembly, rather than constructed almost entirely on-site the way traditional large reactors are. That factory-built approach is the main reason SMRs are projected to take 3 to 5 years to build, compared with 7 to 10 years for a traditional large reactor, which is typically sized at 1,000 megawatts or more and requires far more bespoke, site-specific construction work. The other significant difference is who typically buys one: large reactors have traditionally been built by utilities serving broad regional grids, while SMRs are increasingly being purchased by a single large industrial customer, an AI data center operator among the most prominent examples, to power one specific facility directly rather than feed a shared grid.

Why do AI data centers need 80MW or more of power, more than double a standard data center's needs?

AI-focused data centers need substantially more power than standard data centers because of the density and utilization pattern of the hardware inside them. An AI-optimized facility can require 80 megawatts or more of continuous power, versus roughly 32 megawatts for a standard data center, largely because GPU-dense AI infrastructure runs at much higher power draw per rack than general-purpose computing equipment, and AI training and inference workloads tend to keep that hardware running at sustained, near-maximum utilization for long stretches rather than the more variable, bursty usage pattern typical of general enterprise computing. That combination of higher per-rack power draw and higher, more continuous utilization is what pushes total facility power demand so far past what a comparable non-AI data center would need, and it's the direct reason AI data centers have become such a specific, concentrated source of new grid demand in the regions where they're being built.

What is the Crusoe/Aalo Atomics nuclear-powered 'AI Factory' project, and when will it go live?

Crusoe and Aalo Atomics are targeting a nuclear-powered "AI Factory" proof of concept at Idaho National Laboratory in 2027, with full commercial deployment targeted for 2029. The project is a concrete, dated test case for the broader SMR-for-AI thesis running through this entire trend: rather than a company simply signing a nuclear power purchase agreement and waiting for a reactor developer to deliver years down the line, this is a specific, named proof-of-concept site with a real target date, sited at an existing US national laboratory with deep nuclear research infrastructure and expertise already in place. Because it has explicit, near-term dates attached — a 2027 proof of concept followed by 2029 commercial deployment — it's one of the more useful projects to watch as an early, real-world signal of whether the SMR-for-AI construction timeline promises made across the industry can actually be met, rather than slipping the way skeptics like the Bulletin of the Atomic Scientists have warned they might.

What is China's Linglong One reactor, and why is it called the world's first commercial onshore SMR?

Linglong One is a small modular reactor in China scheduled to begin commercial operation in the first half of 2026, and it holds the distinction of being the world's first commercial onshore SMR — meaning the first small modular reactor design to reach actual commercial operation on land, rather than remaining at the demonstration or pilot stage, or being a marine/offshore design. That "first" matters beyond symbolism: a commercially operating SMR gives the entire industry, including the US companies betting billions on similar technology, a genuine real-world data point on construction timelines, operating performance, and commercial viability, rather than relying entirely on projections and announcements. Linglong One's 2026 commercial operation date puts China meaningfully ahead of most Western SMR-for-AI projects, several of which, like the Crusoe/Aalo Atomics Idaho project, are still at the proof-of-concept stage targeting 2027 and beyond.

How does China's total nuclear capacity (112 units, 125 million kW) compare globally?

China holds 112 operational or approved nuclear units, with roughly 125 million kilowatts of combined capacity, securing the largest total nuclear capacity of any country in the world. That scale dwarfs the individual reactor and power purchase agreement deals driving the US corporate nuclear-for-AI story, and it reflects a fundamentally different model: rather than a series of individual companies each securing their own dedicated reactor capacity through separate deals, China's nuclear buildout is centrally planned and executed at a national level, with AI and data center electricity demand being one of several drivers behind that broader expansion rather than the sole reason for it. For anyone trying to gauge which country is actually best positioned to power a large-scale AI compute buildout with nuclear energy over the next decade, China's existing capacity and construction pipeline puts it in a substantially stronger position than any single Western company's set of nuclear power purchase agreements currently reflects.

How many new Chinese nuclear units are starting construction or reaching completion in 2026?

China has two new nuclear units starting construction in 2026, with seven more units reaching completion before the end of the year. That pace, nine units either breaking ground or coming online within a single calendar year, is a substantial rate of nuclear construction activity by any global standard, and it's happening as part of a broader, sustained expansion that has already given China 112 operational or approved units and roughly 125 million kilowatts of combined nuclear capacity, the largest national total in the world. Set against the roughly 400-terawatt-hour data center electricity demand China is expected to reach by 2030, this pace of nuclear construction reads as a deliberate, planned response to a specific, forecasted demand curve rather than a reactive scramble to catch up with unexpected AI-driven growth after the fact.

How is China's electricity mix for data centers expected to shift between now and 2035?

China's data center electricity mix is currently dominated by coal, at roughly 70%, with renewables contributing around 20% and nuclear around 10%, according to IEA-cited projections. That mix is expected to shift substantially over the following decade, with nuclear and renewables together projected to supply nearly 60% of China's data center electricity by 2035 — a significant reduction in coal's share and a meaningful expansion of both nuclear capacity (backed by the country's ongoing construction pipeline, including Linglong One and the broader 112-unit fleet) and renewable generation. That shift lines up with China's stated broader climate and energy-transition goals, but it's also a practical response to the sheer scale of new demand: with Chinese data center electricity demand expected to double to 400 terawatt-hours by 2030, continuing to meet that growth primarily with coal would run directly against the country's own decarbonization targets.

Why is Australia reportedly building its AI power grid 'without a nuclear option'?

Reporting on Australia's AI infrastructure buildout frames the country as pursuing grid capacity growth "without a nuclear option," relying instead on renewables and gas to meet rising data center demand. That's a genuine policy divergence from the US and China, both of which are leaning into nuclear power, including small modular reactors, as part of their AI infrastructure strategy. Australia's approach reflects its own longer-running energy policy context, where nuclear power generation has faced significant domestic political and regulatory barriers independent of the AI data center boom, rather than a judgment specifically about nuclear's suitability for powering AI workloads. The practical consequence is that Australia's AI-driven grid growth is being tested against renewables-plus-gas capacity in a way that other nuclear-embracing markets aren't, making it a useful comparative case for whether that combination can keep pace with the same kind of sustained, always-on AI power demand nuclear is specifically being pursued elsewhere to solve.

What nuclear/renewable mix is planned to power the Stargate UAE campus?

Stargate UAE, the UAE's flagship sovereign AI compute cluster, is planned to draw on a mixed power source combining nuclear, solar, and natural gas rather than depending on any single generation type. That blend reflects the same underlying tension driving nuclear investment everywhere else in this piece: gigawatt-scale AI compute needs firm, continuous baseload power that solar alone can't reliably provide, natural gas can supply quickly but doesn't sit comfortably alongside long-term decarbonization goals, and nuclear offers a stable, low-carbon middle path that takes longer to bring online than gas but doesn't carry the same emissions profile. Combining all three lets the project meet its power needs immediately through gas and solar while nuclear capacity is developed, rather than waiting years for a single power source to be fully ready before any compute can be powered at all.

Why is nuclear power considered well-suited to the 'always-on' load profile of AI data centers, compared with wind or solar?

Nuclear power generates a stable, continuous output level for extended periods, largely independent of weather or time of day, which matches how AI data centers actually consume electricity: GPU clusters running training or inference workloads tend to operate at sustained, near-maximum utilization for long stretches, not in the bursty, variable pattern that characterizes a lot of general computing demand. Solar generation drops to zero overnight and varies with cloud cover during the day; wind output shifts with weather conditions in ways that don't reliably track a data center's actual power draw. A data center that needs firm, dispatchable power around the clock has to either overbuild renewable capacity substantially and pair it with significant storage to smooth out the gaps, or rely on a generation source that's inherently steady in the first place — which is exactly the gap nuclear is positioned to fill, distinct from, though often discussed alongside, the broader corporate case for choosing a zero-carbon power source.

What regulatory hurdles (NRC licensing, NEPA, state siting rules, Price-Anderson Act liability) apply to SMRs built for AI data centers in the US?

A small modular reactor built to power a US AI data center still has to clear the same regulatory path as any new nuclear generation: licensing review from the Nuclear Regulatory Commission, environmental review under the National Environmental Policy Act, compliance with state-level siting rules that vary considerably depending on location, and liability coverage under the Price-Anderson Act, the decades-old federal framework governing nuclear liability insurance. None of these requirements are unique to AI-driven demand, but none of them move at the pace of a corporate press release either, and they represent a genuine, non-negotiable part of any SMR project's real timeline. This regulatory reality is a core part of why critics, including the Bulletin of the Atomic Scientists, are skeptical that even an optimistic 3-to-5-year SMR construction estimate reliably holds once realistic permitting and review time is added on top of pure manufacturing and assembly time.

What is Amazon's investment in X-Energy, and what is it meant to achieve?

Amazon has made a direct investment in X-Energy, a small modular reactor developer, as part of AWS's broader exploration of nuclear-powered cloud campuses. Rather than purely acting as an end customer signing a power purchase agreement after a reactor is already built, an equity stake gives Amazon a more direct interest in X-Energy's success, and potentially more influence over the pace and priorities of the reactor's development, than a purchase agreement alone would provide. The goal is straightforward: securing dedicated, reliable nuclear power for AWS's own AI and cloud infrastructure without depending entirely on winning a place in an already-congested grid interconnection queue. It's also a broader bet that being an early, invested partner in a specific SMR developer gives Amazon better visibility and more leverage over that company's delivery timeline than simply waiting as a customer further down the priority list would.

How long are current nuclear/SMR interconnection queues in grid-constrained US regions?

Interconnection queues, the wait time for a large new power consumer to actually connect to the existing electrical grid, currently stretch 5 to 10 years in many grid-constrained US regions. That delay is a direct consequence of how much new demand, AI data centers prominent among it, is trying to connect to grids that were planned and sized around a considerably lower demand forecast, and it affects specific, named data center hubs including Northern Virginia, Oregon, and Texas particularly acutely. A 5-to-10-year wait is, not coincidentally, roughly comparable to or longer than the projected construction timeline for an SMR, which is exactly why dedicated nuclear generation has become an attractive alternative for companies unwilling to simply wait in an interconnection queue that may not resolve on a timeline compatible with their AI growth plans.

How big could the global SMR market become by the mid-2030s and 2040s?

IDTechEx forecasts the global small modular reactor market reaching $53.8 billion by 2036, and growing to nearly $300 billion by 2046. That trajectory implies SMRs moving well beyond AI data centers into broader industrial and utility use over that period, with data centers serving as one of the earliest and most visible demand drivers rather than the market's only long-term customer base. Reaching that scale depends on the current wave of SMR projects, the Crusoe/Aalo Atomics Idaho National Laboratory project and China's Linglong One among the most concrete examples, actually delivering on their construction timelines and operating reliably once commissioned; a market forecast of this size assumes the current early projects succeed well enough to justify the much larger buildout that a $300 billion figure by 2046 would require.

Why do critics argue that nuclear-for-AI announcements often outpace what utilities can actually deliver on a realistic timeline?

Critics, including the Bulletin of the Atomic Scientists, argue that announcing a nuclear power purchase agreement or making an equity investment in a reactor developer is a fast, relatively low-cost, high-visibility action a technology company can take almost immediately, while actually designing, licensing, manufacturing, and commissioning a working reactor is a multi-year process subject to regulatory review, construction risk, and supply-chain constraints that don't move at the same speed as a corporate announcement. Nuclear power's own history includes numerous large projects that ran significantly longer and more expensive than their original timelines suggested, which gives this skepticism real grounding rather than making it purely reflexive caution about a new technology. The practical implication is that a signed agreement or an invested amount of capital should be read as a genuine intention backed by real money, not as a reliable estimate of when megawatts will actually start flowing to a data center.

What supply-chain constraints (e.g., fuel availability for advanced reactor designs) could slow SMR deployment for AI data centers?

Several advanced reactor designs, including many small modular reactor designs, depend on specific nuclear fuel types and enrichment levels that aren't yet produced at large industrial scale, and that fuel-supply constraint is a genuine bottleneck independent of how well-financed or well-permitted a given reactor project is. A reactor can be fully designed, licensed, and even physically constructed, and still face delays if the fuel supply chain hasn't scaled fast enough to match a sudden surge in demand from multiple SMR projects competing for the same limited fuel production capacity at the same time. This is exactly the kind of unglamorous, easy-to-overlook constraint that rarely makes it into a company's nuclear-for-AI press release, but that ends up mattering considerably for whether a project's promised timeline actually holds once construction is complete and the reactor is ready to be fueled and brought online.

How much SMR capacity is currently positioned globally for industrial users such as data centers?

More than 40 gigawatts of small modular reactor capacity is currently positioned globally for industrial users, data centers among the most prominent examples of that category. That figure represents announced and planned capacity across multiple projects and countries rather than capacity that's already built and operating, so it should be read as a measure of stated ambition and pipeline size rather than delivered, commissioned generation. Set against IDTechEx's longer-range forecast of the SMR market reaching $53.8 billion by 2036 and nearly $300 billion by 2046, that 40-gigawatt positioned figure represents an early, foundational slice of a market still expected to grow substantially larger over the following two decades, assuming enough of the current wave of projects clear the regulatory, supply-chain, and construction hurdles covered throughout this piece.