Nuclear power for the machines.

The number I could not stop looking at was 550. Between 2010 and 2018, the amount of computing done inside the world's data centres grew by roughly 550 percent. Their electricity consumption over those same eight years barely moved. Masanet and colleagues published that finding in Science in 2020, and it is the single most useful thing I have read on this subject, because it kills the intuition almost everyone starts with. More computing does not automatically mean more power. For most of the last decade, efficiency ate the growth.

I bring this up because the story going around right now says the opposite, and it might be right anyway.

The thing efficiency did, and might not do again

The Masanet paper matters because of how boring its mechanism is. Servers got better. Old inefficient machines in company basements moved into hyperscale facilities that run closer to capacity. Virtualisation meant one physical box did the work of several. None of that is glamorous, and all of it bought the world about a decade of essentially free computing growth in energy terms.

The reason people think that era is over is a short paper by Alex de Vries in Joule, published in 2023. Working from projections of how many AI servers would actually be manufactured and shipped, he estimated that AI could add somewhere between 85 and 134 terawatt hours per year to global electricity demand by 2027. For scale, that is in the neighbourhood of what a country like the Netherlands, Argentina or Sweden consumes annually.

Notice what that estimate is built on. Not vibes about the future of intelligence. Hardware supply chains. You can only run the models you have chips for, and someone has to build the chips. It is a modest, checkable way to make a projection, and it is also why the number has a range in it rather than a single confident figure.

Then the reactors

The corporate side of this is easy to summarise and has been reported everywhere. Microsoft contracted to restart a unit at Three Mile Island, now renamed the Crane Clean Energy Center, buying the output for twenty years. Google signed with Kairos Power for small modular reactors. Amazon put money into X-energy. These are company announcements rather than research, so I treat them as what they are, evidence of what firms are doing, not evidence that it will work.

The logic is legible enough. A training cluster wants steady high quality power in one location every hour of the year, and that profile suits a reactor better than it suits solar without storage. Restarting an existing plant is the cautious version of that bet. Ordering reactors that have not been built at commercial scale in the West is the expensive version.

The part that is actually unresolved

Here is where I expected to find a settled answer and did not. The standard line is that nuclear construction costs always escalate, that this is a law of the technology. Lovering, Yip and Nordhaus challenged exactly that in Energy Policy in 2016, assembling overnight construction costs for a large share of the reactors ever built and arguing that the pattern varies enormously by country and era, with some programmes holding costs flat or reducing them.

The following year, in the same journal, Koomey, Hultman and Grubler published a formal reply arguing the analysis leaned on overnight costs that exclude financing, leaned on early demonstration reactors, and that the authors' own modern-era data pointed the other way. Both papers are in the peer-reviewed record. Neither one has been withdrawn. If you want to know whether new reactors will be affordable, the honest answer is that serious people looking at the same construction data still disagree about what it shows.

There is a broader challenge too. Sovacool and colleagues, writing in Nature Energy in 2020, examined national electricity systems and reported that countries leaning heavily on nuclear did not show the emissions reductions associated with countries leaning on renewables, and that the two strategies sat awkwardly together in the same grid. That paper drew substantial methodological criticism after publication. I include it because it exists and because pretending the critical literature is absent would be dishonest, not because I think it closes the question.

Nobody talks about the water

This is the part that made me sit up, because it is the seam where this whole subject touches the book I wrote. David Mytton published a review in npj Clean Water in 2021 pointing out that data centre electricity gets constant coverage while data centre water gets almost none. American data centres consume on the order of 1.7 billion litres a day, direct and indirect. Against total United States consumption of around 1218 billion litres a day, that is small. It is not the volume that worried me.

It was this. In some facilities a majority of the water drawn for cooling is potable, and fewer than a third of operators measure their water use at all. Thermal generation is thirsty, so a reactor built to cool a data centre adds water demand at both ends of the same chain. We are building an enormous new thing on top of a resource we have not bothered to instrument.

Why a novelist keeps a folder on this

My novel is set in 2150, after fresh water became the basis of political authority. The machine at its centre is a desalination plant, which is to say a machine that converts electricity into drinkable water, kept running by people who are never permitted to fail. An artificial intelligence called MADRE administers what remains of the system.

Nobody in 2150 argues about whether building her was worth the electricity. They live inside the answer. That is the only real trick fiction has here. It skips forward to the people who inherit a decision and asks how it looks from there.

So I am not going to tell you the nuclear revival is a catastrophe or a rescue. The research does not support either verdict. What the research does support is narrower and, I think, more uncomfortable. Allocation of firm low carbon power is now being decided through twenty year private contracts, in an industry whose own energy trajectory is disputed, drawing on a water system almost nobody is measuring. Watch who signs. That is where the futures get written.

Peer-reviewed sources

• Data centre energy · Science, Q1 ScopusMasanet, E., Shehabi, A., Lei, N., Smith, S. and Koomey, J. (2020). "Recalibrating global data center energy-use estimates." Science, 367(6481), 984-986. doi:10.1126/science.aba3758
• AI electricity demand · Joule, Q1 Scopusde Vries, A. (2023). "The growing energy footprint of artificial intelligence." Joule, 7(10), 2191-2194. doi:10.1016/j.joule.2023.09.004
• Water · npj Clean Water, Q1 ScopusMytton, D. (2021). "Data centre water consumption." npj Clean Water, 4, 11. doi:10.1038/s41545-021-00101-w
• Reactor costs · Energy Policy, Q1 ScopusLovering, J. R., Yip, A. and Nordhaus, T. (2016). "Historical construction costs of global nuclear power reactors." Energy Policy, 91, 371-382. doi:10.1016/j.enpol.2016.01.011
• The formal reply · Energy Policy, Q1 ScopusKoomey, J., Hultman, N. E. and Grubler, A. (2017). "A reply to 'Historical construction costs of global nuclear power reactors'." Energy Policy, 102, 640-643. doi:10.1016/j.enpol.2016.03.052
• Nuclear and renewables compared · Nature Energy, Q1 ScopusSovacool, B. K., Schmid, P., Stirling, A., Walter, G. and MacKerron, G. (2020). "Differences in carbon emissions reduction between countries pursuing renewable electricity versus nuclear power." Nature Energy, 5, 928-935. doi:10.1038/s41560-020-00696-3

Editorial note: the argument above rests on the six peer-reviewed sources listed, all in journals indexed in Scopus and Web of Science. Company announcements from Microsoft, Google and Amazon are used only as reporting of corporate behaviour, never as evidence of outcomes. Every DOI was verified against the Crossref registry and checked for retraction notices on September 1, 2026. The author is a novelist and doctoral researcher, not an energy analyst, and the closing section is signed opinion rather than a finding.