AI datacenters want to go nuclear. Too bad they needed it yesterday

Silicon Valley's latest energy fixation won't stop the coming power panic

Analysis Atomic energy is becoming the preferred solution to address the projected bump in megawatts needed to charge AI in the future, but it simply won't come soon enough in many cases.

Much has been written about infrastructure and the electricity required to keep feeding the AI beast. Recent research from Goldman Sachs estimated that datacenter energy use will more than double by the end of the decade – just five years away.

Last year, a report from the International Energy Agency (IEA) said the world's bit barns account for just over 1 percent of global electricity consumption, and this rises to more like 2 to 4 percent in large economies including US, China, and Europe. In hotspots like Ireland, datacenters already account for more than 20 percent of electricity consumption, and that figure keeps on rising.

And it isn't just datacenters. Compounding the issue, the IEA foresees that electric vehicles may account for 6 to 8 percent of total electricity demand globally by 2035, up from just 0.5 percent today.

With access to enough energy now seen as a constraint on sustained growth, some of the larger cloud and datacenter companies have taken it upon themselves to seek a reliable source of power, and nuclear presents itself as an attractive option.

Atomic generators can provide continuous round-the-clock power, unlike some renewables, and have almost zero carbon dioxide emissions, which is a bonus if your goal is to become carbon-neutral or carbon-negative in the future, like all the big IT companies claim they want to.

Just this month, a group of large-scale energy users including Amazon, Meta, and Google backed efforts to ramp global nuclear capacity, aiming to triple it by 2050. This follows Microsoft's earlier involvement in a project to restart generation at the Three Mile Island atomic plant in Pennsylvania by 2028, and Amazon's purchase of a datacenter campus on the same site as the Susquehanna nuclear power plant, also in Pennsylvania.

Yet the issue with many of these nuclear projects is that they are too little or too late, with some of them a decade or more away, while huge increases in datacenter energy consumption are forecast for the next few years. The World Nuclear Association acknowledges that an atomic plant typically takes at least five years to construct, for example.

"Undoubtedly, nuclear energy will serve as part of the world's energy mix for years to come," says Canalys principal ESG analyst Elsa Nightingale told The Register. "However, investing heavily in nuclear energy doesn't address the core issue. For one, nuclear projects have long lead times while AI's energy demands are coming now."

More critically, Nightingale adds, many electricity grids around the world lack the capacity to deliver on the astronomical energy demands forecast for AI – even with grid expansions.

A report published last year warned that Americans could face a 70 percent hike in their electricity bills by 2030 unless action is taken to boost generation and transmission capacity. It estimated that billions need to be invested in energy grids to keep pace with the growth in demand from datacenters.

Management consultancy Bain & Company also forecast that power use will outstrip supply within the next few years in some places, and that utility companies have been slow to react because they have for some time faced flat or shrinking demand in the US market which led them to prioritize efficiency over scaling capacity and building new distribution lines.

This could mean we are heading for trouble, unless datacenters can generate some of their own energy to meet demand or greater efficiency can be prioritized in AI infrastructure. Nvidia's Blackwell GPUs, announced last year, consume over a kilowatt of power each, for example.

"Do we want to see a future where IT companies are hypothetically competing with hospitals for electricity?" Nightingale asked.

This is one of the scenarios forecast by energy infrastructure biz Schneider Electric earlier this year. It published research that modeled four futures, based on how demand by AI datacenters is managed, one of which points to energy conflicts with other critical sectors of the economy, leading to economic downturn and other negative outcomes.

"The more immediate challenge here is that most of the largest tech firms worldwide have net zero goals which become increasingly untenable as AI emissions rise at the rates forecast," Nightingale told us.

Remember those carbon-neutral or carbon-negative goals we mentioned earlier? Microsoft admitted last year that its CO₂ emissions have risen by nearly 30 percent since 2020, thanks to building more datacenter capacity for AI workloads, while Google admitted that its emissions are up 48 percent since 2019.

To address this, the tech industry needs to urgently address efficiency, as AMD president Victor Peng told the Hot Chips conference last year.

"Whatever power budget you think you're limited at, if you get higher performance, you could either train larger models and get to intelligence quicker, or you can serve it more cost effectively," he explained, calling on semiconductor firms to focus more attention on making the infrastructure more efficient.

The answer to this thorny problem is thus likely to come from a number of different directions, including bringing more renewables such as wind and solar online to feed into the grid, making energy grids more robust, and datacenters generating some of their own power, even if this means gas turbines in the short term.

"So, nuclear energy, while an interesting piece of the puzzle, avoids the real problem entirely: AI is not only thirsty for water requiring vast amounts for cooling and arguably more challenging to resolve, but it is also hungry for energy," said Nightingale.

"Instead of looking to nuclear as a silver bullet, tech companies must first find ways to dramatically reduce AI's electricity and water demands, before they can credibly claim their AI investments and net zero goals are tenable." ®