Energy hungry data centers in AI era puts nuclear back on the agenda

Energy hungry data centers in AI era puts nuclear back on the agenda

Nuclear energy is slowly but surely emerging from the doldrums. For years, the very idea of using more of this reliable, scalable, and relatively cheap energy source to reduce CO2 emissions was out of the question. Now that it is becoming increasingly clear that the energy needs of data centers running AI workloads cannot be met with renewable sources alone, the tide seems to be turning. Hopefully, everyone will benefit from this paradigm change.

Google will build data centers that consume more than 1 gigawatt of electricity annually. CEO Sundar Pichai said this during a speech he delivered at Carnegie Mellon University in Pittsburgh, once known for its steel industry. It’s clear that in the 21st-century economy, it is not so much the raw materials of yesteryear but the computing power in data centers that helps determine the economic success of the companies and organizations that depend on them. To put things in perspective, one gigawatt of energy is roughly equivalent to the annual energy consumption of more than 2 to 3 million households.

The Google CEO’s comment was accompanied by the suggestion that the energy necessary for such consumption may well come from small nuclear reactors, known as SMRs or small modular reactors. These nuclear power plants are smaller than traditional reactors and whose components can be manufactured elsewhere to be assembled on-site. Incidentally, Pichai left open the question of whether Google will actually use such power sources. But why else would he bring it up?

Google is not the only one

Moreover, Pichai’s comment cannot be separated from other developments in the world of data centers. For example, Oracle is already planning to build three such small reactors to power some of its 162 data centers currently in operation or being built worldwide. Oracle’s largest data center currently consumes 800 megawatts. It doesn’t look like that power demand is going to decrease in the near future. Exactly how this will work is not yet entirely clear, but the intention is clear.

Microsoft, too, sees a solution to its growing energy needs in nuclear power. For example, the company has very pragmatically commissioned an older nuclear power plant in the U.S. state of Pennsylvania. It produces 835 megawatts of energy, equal to the annual energy consumption of 800,000 (American) households. The plant’s operator closed it in 2029 because it was ‘too expensive’. Now that a party like Microsoft sees its appetite for energy growing, this operator is allocating 1.6 billion dollars to restart the plant. Microsoft promises to use this reactor’s energy for the next 20 years.

Reactor patched up

The reactor being restarted especially for Microsoft is the Three Mile Island reactor, by the way. A partial meltdown took place there in 1979. The reactor that Microsoft will use is still in good condition. It will return to operation after the necessary refurbishment work, probably around 2028.

In the same Pennsylvania is the Susquehanna nuclear power plant, where AWS purchased an existing data center right next door earlier this year. That plant accounts for 960 megawatts of annual output. The company is also posting job openings for nuclear engineers these days. These engineers will probably be hired to explore further options for using SMRs.

It’s not surprising that large tech companies have set their sights on nuclear power. The AI revolution is ever hungry for energy. For example, GPU superpower Nvidia has been calling for some time now for data centers to be turned into full-fledged ‘AI factories‘, which come with massive power consumption. Meanwhile, almost all tech companies are committing to carbon-neutral ambitions for 2030, but Goldman Sachs has calculated that by that year, data centers’ energy demand will have grown to 160 percent of what they consume today.

That’s largely due to AI workloads. By comparison, entering a prompt into ChatGPT consumes 10 times as much energy as a search query to ‘traditional’ Google Search. According to the business bank, the world is going to feel the effects of this energy demand in the coming years. To summarise, a lot more energy will be needed, and the infrastructure is yet to be built.

Tip: AWS shows AI use can be done without CO2 spike

Energy consumption rising since 2020

In the same analysis, the bank highlights that before the AI boom, data center workloads had already been increasing threefold from 2015 through 2019. However, their energy consumption remained almost unchanged, mainly due to efficiency gains. Now, however, the ‘law of diminishing return’ seems to apply, as consumption has been rising since 2020. Whereas in 2015 (and 2016, and 2017 etc.) data centers worldwide still consumed about 200 terawatt-hours per year, in 2023 that had grown to more than 400 terawatt-hours. It is worth noting that the U.S. accounts for about one-third of this energy consumption, then and now.

In Europe, energy demand has declined since the peak year of 2008. There are several reasons for this, including the financial crisis, the coronavirus pandemic, and ongoing deindustrialization. But according to Goldman Sachs, with the construction of new data centers in Europe, –not least to meet ‘data residency’ requirements–, energy demand could increase by 40 or 50 percent.

Significant investment required

The bank figures show that by 2030, the energy consumption of European data centers will equal the total annual power consumption of Portugal, Greece, and the Netherlands. Because Europe has an old energy network (another law in force here: that of first-mover disadvantage), substantial investments are necessary: 800 billion for transmission and distribution and 850 billion for power generation via renewable energy sources. In this analysis, Goldman Sachs seems to assume energy growth will mostly happen through solar and wind farms.

But let’s face it: no million wind turbines and solar panels will satisfy data centers’ enormous appetite for energy. Generating 1 gigawatt of energy requires about 100 wind turbines. The actual output depends, among other things, on location, placement, and the strength of the wind. Generating a gigawatt with solar power requires almost two million photovoltaic panels. A (small) nuclear power plant takes up less space and provides consistent power.

For years, nuclear power was the energy source that must-not-be-named. Disasters like Chernobyl and later the (tsunami-induced) meltdown in Fukushima, Japan, stifled enthusiasm for this reliable, cheap and enormously scalable energy source. This is understandable, but if we want to reduce CO2 emissions while taking advantage of the computing power that data centers unlock, there is really only one real solution: plenty of investment in nuclear power. Big tech companies know this and are taking the lead.

An expensive ‘wende’

This does require a shift in thinking. Germany led the way in a so-called ‘energy turn’ or ‘energiewende’ over the past decade. It decommissioned countless nuclear power plants that could have remained in operation for many years. Of course, this cost an enormous amount of lost output, which solar and wind energy were nowhere near able to absorb. Consequence: Germany, the country that wanted to be at the forefront of a green revolution, started burning extra coal.

France, which cherished its nuclear power plants, now pays much less for energy. Germany now even imports nuclear energy from France. This (somewhat older) article on Forbes explains it clearly: France generated less than a tenth of Germany’s CO2 emissions in recent years, and generating the energy also cost that country half as much.

France’s investments in renewable sources like solar and wind power (made under pressure from Germany and the European Commission) resulted in higher energy bills for French households. In other words, energy for France would have been even cheaper had it not listened to the criticism that it should cut back on nuclear power.

Circumstances dictate the course

Now, under pressure from big tech, circumstances seem to finally dictate the course. To meet ambitious climate goals, there is no way around nuclear power. Whether this is a ‘clean’ energy solution is up for debate. It is not renewable, but no CO2 is released during energy generation. It produces hazardous waste, but is highly reliable. A key issue, of course, is that the waste should be stored safely and responsibly.

It is conceivable that data centers located in Europe, or under construction, will eventually receive some of their energy from nuclear power plants. At least if policymakers allow this and do not insist against all better judgment on expensive and inefficient (because of their unreliable output) solar and wind energy and biomass, supplemented by polluting coal.

If we want to fully take advantage of the computing power that data centers harness for AI workloads and all that it makes possible, we must invest in nuclear energy. When we do, it is important that not only big tech companies but ordinary citizens and local businesses benefit from the clean and reliable power it generates as well. Now, that would be an energy revolution.

Read also: Google’s CO2 emissions double: AI demands high environmental price