Every time you ask a chatbot a question or generate an image, electricity flows through a data center somewhere on the planet. The total volume of that electricity is now swelling at an unprecedented pace. The International Energy Agency (IEA) projects that global data center electricity consumption will nearly double, from about 415 terawatt-hours (TWh) in 2024 to roughly 945 TWh by 2030 [source: IEA Energy and AI, 2025]. The single biggest driver of that growth is AI.
The question of what will supply all that power has the world's richest companies pointing in the same direction: nuclear. Microsoft is reviving a shuttered reactor, while Amazon and Google are betting hundreds of millions of dollars on small modular reactors (SMRs). China has signaled that the world's first commercial land-based SMR will switch on in the first half of 2026. But an announced deal and a reactor that is actually running are two different stories. This article places expectation and skepticism side by side, separating what has been measured from what is still only a promise.
One habit makes the rest of this article easier to read: sort every number into one of three piles. Measured figures describe something that already happened and was recorded — the 415 TWh data centers consumed in 2024, or a reactor that passed a named test on a named date. Projected figures describe a modeled future, such as the 945 TWh for 2030, which the IEA presents as a scenario rather than a guarantee. Announced figures describe an intention: a signed contract, a capacity target, a governor's directive. All three appear below, labeled where they appear. Announcements are the easiest to publish and the least binding, and they make up most of what the phrase "nuclear comeback" currently describes.
How Big Is AI's Power Appetite, Really?
The Measured Baseline
Start with the numbers. According to the IEA, data centers consumed about 415 TWh in 2024, roughly 1.5% of the world's electricity, growing at an average of 12% a year over the past five years [source: IEA Energy and AI, 2025]. By 2030, that figure is projected to reach about 945 TWh, close to 3% of global electricity. Regionally, in 2024 the United States accounted for 45%, China for 25%, and Europe for 15%.
Two things about that baseline are worth separating. The share figure depends on rounding: the IEA's roughly 1.5% of world electricity [source: IEA Energy and AI, 2025] appears in some coverage of the same report as just over 1%, a difference of expression rather than of data. The growth rate is the more telling number — 12% a year, sustained across five years, is the pace that moves data centers from a rounding error in the global power mix to a line item grid planners have to build for. The load is also concentrated rather than evenly spread, which is why the strain, and the politics that come with it, land in a handful of places.
Where the Growth Is Coming From
What matters most is the direction of that growth. The IEA expects electricity demand from AI-optimized data centers to more than quadruple by 2030. AI's share of data center power, around 5–15% in recent years, could rise to 35–50% by 2030 [source: IEA Energy and AI, 2025]. In other words, the growth of data centers as a whole is increasingly being pulled along by AI.
These are two different claims, and it helps to keep them apart. The near-doubling to 945 TWh covers all data centers, including the storage, streaming, and enterprise workloads that were growing long before the current AI cycle. The more-than-quadrupling applies only to the AI-optimized subset. That subset is the smaller of the two today, at something like 5–15% of data center power in recent years, which is why its growth rate can be steeper without AI consuming the whole 945 TWh on its own [source: IEA Energy and AI, 2025]. The IEA's argument is about which component pulls the overall curve upward.
A Base Case Is a Scenario, Not a Schedule
It is worth being clear, though, that these numbers are not a settled future. The IEA itself calls this its "Base Case." If models grow more efficient, or if overbuilt data center contracts are cancelled, actual demand could fall short of the forecast. The 2024 figure of 415 TWh is measured; the 945 TWh is a projection resting on a stack of assumptions.
The two headline numbers carry different burdens of proof. The 2024 consumption figure is an accounting of electricity that was actually delivered. The 2030 figure is the output of a model published in April 2025, and it moves whenever its inputs move: how quickly models become more efficient, how many announced facilities are actually built, how much of the announced capacity was speculative to begin with. The IEA publishes a central case because a single number is easier to plan against than a range, not because the range has collapsed.
What Big Tech Actually Bought — Restarts and SMR Bets
Facing the fear of a power shortfall, Big Tech concluded that renewables alone would struggle to secure the round-the-clock, large-scale supply they need. So they turned to nuclear, which runs steadily all day while emitting almost no carbon. Yet when you examine what they actually bought, it falls into two distinctly different categories.
The instrument they reached for deserves naming before the deals themselves. A power purchase agreement, or PPA, is a long-term contract to buy a generator's output at agreed terms. It moves electricity and money; on its own it pours no concrete. A twenty-year PPA can keep an existing plant alive by guaranteeing it a buyer, or it can give a developer the revenue certainty needed to finance something new — two very different transactions, and the deals below include both.
Extending the Life of Existing Reactors
The most eye-catching deals are, in fact, not new reactors but efforts to revive or preserve large existing ones. In September 2024, Microsoft signed a 20-year power purchase agreement (PPA) with Constellation Energy to restart Unit 1 of Three Mile Island, which had been closed in 2019 for economic reasons. All 835 megawatts (MW) of its output will go to Microsoft data centers. The plant has been renamed the Crane Clean Energy Center, and after the U.S. Department of Energy (DOE) approved a $1 billion loan in November 2025, the target restart date was set for 2027 [source: CNBC, 2025].
Three details are easy to lose in the headlines. The reactor being restarted is Unit 1 — not Unit 2, the reactor that suffered the 1979 accident that made the site's name famous. Its output is contracted into the PJM grid region. And the DOE's $1 billion loan does not cover the whole bill: the restart is estimated to cost about $1.6 billion in total [source: CNBC, 2025]. None of this changes the character of the deal — a plant that already exists, brought back rather than built — but it does explain why the target date sits in 2027 rather than immediately after the signature.
Meta followed a similar path. In June 2025 it signed a 20-year deal with Constellation to receive 1,121 MW from the Clinton Clean Energy Center in Illinois starting in June 2027 [source: Constellation Energy, 2025]. That plant, too, is an existing facility once slated for closure, so Meta's contract effectively extended its life. The point worth underlining is that most of the physical reality behind the so-called "nuclear comeback" is not cutting-edge technology but the life extension of existing reactors.
The Clinton deal contains one small piece of genuinely new capacity: a 30 MW uprate, meaning additional output drawn from the existing plant rather than from a new reactor [source: Constellation Energy, 2025]. That is the whole category in miniature. The two agreements cover 835 MW and 1,121 MW of contracted output, and both are scheduled to begin flowing in 2027. Until then the electricity in question is a contractual entitlement rather than a delivered kilowatt-hour — and these are the near-term bets.
The SMR Bet on the Future
The second category is a bet on reactors that do not yet exist. In October 2024, Amazon anchored a $500 million funding round for the SMR developer X-energy and announced a goal of securing more than 5 gigawatts (GW) of new nuclear capacity by 2039 [source: TechCrunch, 2024]. That same month, Google agreed to support seven reactors from the SMR startup Kairos Power, aiming to bring the first online in 2030 and reach a total of 500 MW by 2035 [source: CBS News, 2024]. It drew attention as the first time a tech company had commissioned the construction of new reactors.
The two bets are not the same kind of commitment either. Amazon's is an investment position with a capacity goal attached: it anchored a funding round and set a target of more than 5 GW by 2039, a statement about where the company wants to be at the end of the 2030s [source: TechCrunch, 2024]. Google's is closer to an order book — support for a specific number of reactors, seven, with a first unit dated to 2030 and a cumulative 500 MW by 2035 [source: CBS News, 2024]. Public money sits alongside the private money, through DOE cost-shared support programs for SMR development. What none of these arrangements includes is an operating reactor.
The difference between the two approaches lies on the timeline. A restart uses an existing facility and can deliver electricity within a few years, but most SMR deals will not produce their first power until after 2030. The announcements are dazzling, yet a great deal of time still stands between them and electricity that actually runs a data center.
Four Stages Between a Signature and a Kilowatt-Hour
Almost every dispute in this field is really a dispute about which stage a project has reached. First comes the announcement — a contract, a funding round, a target year. Second comes regulatory approval, which in the United States means a design clearing NRC review. Third comes construction, the stage where nuclear projects have historically slipped. Fourth comes grid connection and commercial operation, the only stage at which anyone receives electricity. A deal at stage one and a plant at stage four are both reported under the word "nuclear," and the distance between them is measured in years and billions of dollars. Most of the figures below belong to stage one.
What Is an SMR, and How Far Along Is It?
What "Small" and "Modular" Are Supposed to Buy
A small modular reactor (SMR) is a compact reactor whose modules are manufactured in a factory and assembled on-site. Most have a generating capacity of 300 MW or less, smaller than a conventional plant. Advocates say standardized module production can cut construction costs and timelines, and that the reactors are easier to build right next to a data center. The trouble is that this promise still lives mostly on paper.
The two words in the name do different work. Small refers to output, the 300 MW-or-less range, well below a conventional large plant. Modular refers to construction: repeated factory-made units rather than a bespoke project. The economic case rests almost entirely on the second word. If standardization cuts costs and timelines the way advocates argue, the saving comes from repetition — the tenth unit should be cheaper and faster than the first. That is a claim about a production line, and settling it requires building one.
Linglong One, Milestone by Milestone
The one closest to commercialization is China's Linglong One (model name ACP100). Built by the China National Nuclear Corporation (CNNC) at the Changjiang site in Hainan province, this 125 MW pressurized water reactor completed cold functional testing in October 2025 and passed a non-nuclear turbine trial run in December [source: World Nuclear News, 2025]. CNNC is targeting commercial operation in the first half of 2026; if it succeeds, it would be the world's first commercial land-based SMR.
Two qualifications belong with those figures. The 125 MW is the gross rating; net output, what the plant sends out after its own systems take their share, is given as 100 MWe [source: World Nuclear News, 2025]. And both completed milestones are pre-operational by design. Cold functional testing, as the name indicates, takes place before the reactor operates at power, and the trial that followed on 23 December 2025 was explicitly a non-nuclear steam start-up: the turbine turned, the reactor did not. These are real, verifiable steps that come before the ones that decide the question.
Announced Date Versus Recorded Status
But here, too, one must separate the announcement from the reality. As of May 2026, the International Atomic Energy Agency's reactor database (PRIS) still classified Linglong One as "Under Construction," with no first grid connection or commercial operation date recorded [source: IAEA PRIS, 2026]. For reference, Russia has operated a floating (barge-based) SMR since 2020, which is distinct from the land-based first. In short, SMRs are at the "coming soon" stage, not the "already here" stage.
PRIS is best understood as a record rather than a verdict. Its status field changes when a reactor's condition changes, so "Under Construction" with no first-grid-connection date is not a prediction that the target will be missed — only a statement that the event has not been recorded. That is what makes it useful. A company's target and an international registry's status field are different kinds of statement, and the registry is the one that moves only after the fact. Akademik Lomonosov, the Russian floating plant, shows how narrow the record at stake is: it has been generating since 2020, so what Linglong One would win is the land-based first, not the first SMR to produce power at all.
The Skeptics — Delays, Costs, Water, and Overestimation
The hope that nuclear will solve AI's power crunch draws equally heavy counterarguments. They fall roughly into three strands: that construction is slow and expensive, that it does not solve the water problem, and that the demand forecasts may have been inflated in the first place.
The First-of-a-Kind Curse
The recurring problem in the nuclear industry is that the "first of a kind" busts its budget and schedule. The signature example is the Idaho project between NuScale Power and a coalition of Utah municipalities (UAMPS). Planned as six reactors totaling 462 MW, it was cancelled in November 2023. The target power price had jumped 53%, from $58 per megawatt-hour in 2021 to $89 in 2023, and the total cost estimate had ballooned from about $5.3 billion to $9.3 billion [source: Utility Dive, 2023]. Unable to line up enough buyers, the project collapsed.
The escalation started earlier than those two dates suggest. When the project was first proposed in 2015 the total cost was put at about $2.7 billion, and the $9.3 billion figure at the end was more than three times that [source: Utility Dive, 2023]. The money was not only private: the DOE had put more than $232 million into the project since 2020. The mechanism that ended it is specific and instructive — the plant depended on municipal utilities subscribing to blocks of its future output, and when too few signed up at the rising price there was no way to proceed.
The Western reality is colder still. As of early 2026, not a single SMR is operating commercially in the United States, and several designs are still working their way through the regulatory review of the Nuclear Regulatory Commission (NRC). The first grid-scale SMR set to connect in North America, the BWRX-300 in Ontario, Canada, is targeting commercial operation around 2029 [source: World Nuclear News, 2025]. Critics such as the Institute for Energy Economics and Financial Analysis (IEEFA) judge SMRs to be "still too expensive, too slow, and too risky."
The North American exception deserves stating precisely, because it is the one place where a stage has actually been cleared. Ontario Power Generation's BWRX-300 at the Darlington site began construction in 2025; the 2029 commercial operation date remains a target, but the construction is a fact on the ground rather than an announcement [source: World Nuclear News, 2025]. In the United States, designs from GE Hitachi and X-energy are among those still working through NRC review, a stage that has to be finished before construction can start, let alone before power flows.
The Water Blind Spot
Power is not the only issue. Data centers use enormous amounts of water to cool their servers. By one estimate, AI data centers consumed roughly 264 billion gallons (about 1 trillion liters) of water in 2025 [source: Lincoln Institute of Land Policy, 2025]. Such estimates vary widely by methodology and cannot be taken as definitive, but the direction, that water demand is rising fast, is shared across multiple studies.
One regional estimate gives the trend a shape. Data centers in Texas were estimated to use about 49 billion gallons of water in 2025, with projections running as high as 399 billion gallons by 2030. Both figures come from the same estimate-based body of analysis as the national number, and the 2030 figure is a projection layered on top of an estimate, so it deserves the widest error bars of anything in this article. What survives the uncertainty is the direction, and the fact that it grows fastest where the capacity is being built.
There is a fact often overlooked here: nuclear plants also use large volumes of water for cooling. Switching the power source from coal and gas to nuclear can cut carbon, but it does not automatically erase the combined water footprint of the data center and the plant. Nuclear, in other words, is not a cure-all that offsets every environmental burden of AI.
There is a second-order version of this point. If the demand forecasts are right, the water question scales with them: more compute means more cooling, and more generation to serve that compute means more cooling again, whatever the fuel. Nuclear's advantage is in carbon, and it is a real one. It is not an advantage in water, and treating a low-carbon supply decision as a complete environmental answer skips the part of the ledger that plays out locally.
Is the Demand Overstated?
The final counterargument aims at the premise itself. The 945 TWh projection above is not a measurement but a scenario built on assumptions. If AI models become efficient enough to deliver the same performance with less computation, or if speculatively announced data center plans are never actually built, power demand could come in below the forecast. This debate is bound up with the broader conversation about an "AI bubble." It is worth clearly recognizing the unverified portion, and asking whether tens of billions of dollars in nuclear bets rest on inflated demand estimates.
It is worth being precise about what the skeptics are and are not claiming. The measured trend is not in dispute; it is the recorded past. The dispute is over the slope from here, and it has testable content. If efficiency gains let the same work be done with less computation, if announced data center projects quietly leave the pipeline, or if speculatively announced capacity is never built, the 2030 figure comes down. If none of that happens, it holds. What makes this more than a forecasting argument is that the nuclear bets are sized against the forecast: commitments dated 2029 to 2039 are being justified by demand nobody has yet observed.
Where Policy Meets Place — New York
From a Directive to a Backbone
This tension shows up most concretely in regional policy, and the U.S. state of New York is a prime example. In June 2025, Governor Kathy Hochul directed the New York Power Authority (NYPA) to develop at least 1 GW of new nuclear capacity in upstate New York [source: Governor of New York, 2025]. Behind the move are data center demand and Micron's $100 billion semiconductor plant near Syracuse.
Note what a directive is and is not. It instructs a state authority to develop capacity; it does not name a site, select a reactor design, or authorize construction. New York's own sequence shows why that matters: the June 2025 instruction was followed in June 2026 by a request for qualifications, the step at which a state establishes which developers are even eligible to bid [source: NYPA, 2026]. The Micron plant and the data center pipeline are the demand-side reason for the urgency, and they are also why the state is planning against load that has been announced rather than metered.
What the 5 GW Is Made Of
The vision then grew. In January 2026 the state expanded it into a 5 GW "Nuclear Reliability Backbone," and in June 2026 it announced a request for qualifications (RFQ) for developers along with $40 million in workforce training support [source: NYPA, 2026]. Yet environmental groups such as the Sierra Club and some media outlets point to major bottlenecks in siting, cost, and schedule. Here again, a gap separates the announced target from actual construction and operation.
The 5 GW headline deserves a second look. The expansion is described as 4 GW of new capacity plus 1 GW that was existing or previously announced, which means the figure is not five times the original directive but the original target folded into a larger program [source: NYPA, 2026]. That is not a criticism of the plan; it is a reading instruction. Announced totals in this field routinely include capacity already counted elsewhere, and the only way to know is to look for the composition. The $40 million for workforce training points at a bottleneck money alone cannot clear quickly: even a fully financed reactor needs people qualified to build and run it.
What to Watch
What Is Settled
To summarize what is happening: the rise in electricity demand driven by AI is a trend confirmed by measurement, and it is also true that Big Tech is betting heavily on nuclear. But most of that betting is still at the stage of announcements and commitments, not electricity that runs a data center. That is why expectation and skepticism must be weighed together.
The asymmetry is the thing to hold on to. On the demand side the evidence is measured and the trend is recorded: consumption, growth rate, regional concentration. On the supply side almost everything is dated in the future — 2027 for the two restarts, 2029 for the first North American grid-scale SMR, 2030 for Google's first unit, 2035 and 2039 for the cumulative targets. A story whose problem is measured and whose solution is entirely scheduled is easy to tell badly in either direction: as a crisis nuclear has already solved, or as a bubble in which nothing is real.
Four Checkpoints
A few signals ahead will help gauge where this story goes. First, whether China's Linglong One actually connects to the grid and enters commercial operation within 2026, the touchstone for whether the world's first commercial land-based SMR becomes real as announced. Second, whether the Three Mile Island (Crane) restart holds to its 2027 schedule. Third, whether the SMRs backed by Amazon and Google actually deliver their promised power between 2029 and 2035. Fourth, whether the IEA's 945 TWh projection holds, or is revised downward by efficiency gains and cancelled contracts. Whether the nuclear comeback ends as a slogan or becomes a reality on the grid will ultimately depend on whether these announcements turn into operation.
None of these checkpoints requires waiting until 2039 to learn something. Each is a date on which a specific claim either becomes a recorded fact or slips, and each slip is informative in the way the NuScale cancellation was: it shows which stage the industry is actually stuck at. The most useful habit to carry out of this article is the smallest one — when the next nuclear-and-AI headline arrives, ask which of the four stages it describes, and whether the number in it was measured, projected, or merely announced.