AI data centers make this situation far more extreme because they will soon require gigawatt-scale electrical supplies, and they want it now. Even where power may be available—in the Virginia data center corridor in the US, for example—a new data center may have to wait seven years for connection to the power grid.
The only way we know of to improve AI models today is to throw more computing power (and, therefore, more electrical power) at the problem. XAI, a startup company launched in March 2023, deployed a “100,000 GPU Cluster” data center outside of Memphis, Tennessee, earlier this year that will initially require 150 MW of power—the equivalent power consumption required by, on average, 120,000 US homes. The local power company had previously delivered just 8 MW to the site, and xAI is currently using portable generators powered by natural gas to help make up the shortfall, adding to the area’s smog issues and creating a significant carbon footprint.
The pursuit of power at all costs attitude is a direct result of the need to chase the “once in a lifetime” opportunities that AI is presenting to companies that can afford to be in the race. Erik Schmidt, former CEO of Google, said in a recent interview, “If you’re not using AI in every aspect of your business, you’re not going to make it.” For an AI provider, this means making your solution better—both in absolute terms and in relative terms vs. your competition.
CLEAN AI ENERGY OPTIONS
So, what happened to the lofty goal of carbon-neutral data center operations?
Hydro and geothermal are useful options if you happen to live in an area where they can be deployed. Neither solar nor wind power can offer reliable, gigawatt-scale production because they need battery storage to make that happen. Grid-scale battery costs have fallen by
99% since the early 1990s, but even the largest battery farm in the world (the Edwards & Sandborn plant in California) would only power a gigawatt-scale facility for around three hours.
Further, locating such a plant close to the data center would be impractical as it takes up 4,660 acres with the batteries and solar panels needed to charge them. We could locate the battery/solar farm elsewhere, of course, but that would need a gigawatt-scale grid connection. But how long might we have to wait for that?
THE NUCLEAR OPTION
It should come as no surprise that many of the hyperscalers are looking to nuclear power for their immediate and long-term needs. Nuclear is safe and clean, and can deliver gigawatt-scale energy from a compact facility that could be co-located with the data center. Nuclear waste is a complex but solvable problem, and the cost of nuclear energy is more than competitive with solar or wind energy if we include the batteries needed to make these renewables more reliable.
Amazon recently bought a 900 MW data center site next to a 1,800 MW nuclear power station. Microsoft has stated its intention to restart one of the reactors at the Three Mile Island plant, also in Pennsylvania. For those of you thinking that “safe” and “Three Mile Island” may be uncomfortable bedfellows, remember that nobody died in the 1979 accident, and atmospheric radiation release was about the equivalent of a chest X-ray.
When it comes to submarine cables and nuclear power stations you would think that the obvious location for the nuclear plant would be near to water—since conventional Pressurized Water Reactor designs need lots of water to operate (between 1,500 and 3,000 liters per megawatt hour generated). But, ironically, it’s less likely that we’ll find the mega data centers located near to the Cable Landing Station. The trend in recent years has been to optically express traffic through the CLS and on to the Data Center.
And the distances may not be short either. While the EllaLink cable system between Brazil and Portugal has a huge Data Center facility collocated with the CLS in Portugal, it also includes a 1,000 km optical backhaul between Lisbon and Madrid.
Fortunately, some modern reactor designs are far less dependent on supplies of water, which is a good thing given that we saw drought conditions in France this August resulting in half of their nuclear power plants being forced offline. Longer term, there are studies that conclude up to 30 existing nuclear sites around the world would be exposed to sea level rises between 50 and 88 centimeters—with over a third of these located in the USA.
But shutting down nuclear plants because we think they are dangerous may not be the logical option either. It’s estimated that there are between 800 (London School of Economics) and 1,100 (emLab at University of California, Santa Barbara and National Bureau of Economic Research) additional deaths in Germany each year because Germany closed down its nuclear reactors in response to public fears and was obliged to replace that power by extending the life of coal-fired power plants. On a smaller scale, the decision to replace the 2 GW of nuclear power generated by the Indian Point Energy Center in New York state with natural gas generators could result in more excess deaths each year from air pollution than the 30 first responders killed in the Chornobyl (Chernobyl being the Russian spelling) nuclear accident in 1986.
Of course, nuclear power comes with inherent risks, but perhaps it’s time to put our fears into perspective. We should certainly expect hyperscalers to be looking at reactivating more US nuclear plants in the near future and building massive AI data centers right next door to them.
SMALL MODULAR REACTORS
But what about the challenges of new nuclear generating capacity? To be useful for AI data centers, the industry must deliver small modular reactors (SMR) that can be co-located with data centers and avoid grid connections. SMRs are generally not at the deployment stage today (only two operational examples exist in China and Russia). Google is working with Kairos, Amazon is working with X-energy, Oracle is working with an unnamed SMR developer, and Microsoft is going very long-term by looking at Helion, a fusion reactor startup. None of these companies has a licensed reactor design, a process that could easily take 10 years to complete, and fusion may be several decades away from practical use.
But here’s the thing; hyperscalers and their investors have a lot of money to spend on the “AI opportunity.” Nuclear looks like the one viable option for clean, compact, reliable, gigawatt-scale energy that can be located next to an AI data center and avoid grid connection delays.
Why should you care how AI data centers are powered? Putting aside the potential benefits that AI can bring to the world, if hyperscalers can fund the development and certification of SMRs worldwide, they could unlock access to essentially limitless, reliable, clean energy to help combat global warming for all of us. And I think that’s worth caring about.
This feature appeared in ON&T Magazine’s 2025 Special Edition, The Future of Ocean Technology, Vol. 5, to read more access the magazine here.
