The executive orders also included plans to accelerate construction of AI data centres, used for processes including AI inference, training, simulation and synthetic data generation. Regulations will be cut and federally owned land will be made available.
Stoked by fears of Chinese AI supremacy, such ambitious aims will also require huge amounts of power, with the new facilities expected to create new loads of more than 100 megawatts (MW) each.
“AI workloads, particularly large language model training, demand extraordinary computational intensity,” says Yasir Arafat, CTO and co-founder of modular nuclear plant developer Aalo Atomics, to Professional Engineering.
“Training a single state-of-the-art model can consume hundreds of megawatt-hours, roughly equivalent to the annual electricity use of dozens of US homes. GPUs driving these tasks operate continuously at high power levels, often exceeding hundreds of kilowatts per data centre rack, which translates to tens of megawatts for ‘hyperscale’ facilities.” Other systems, including data processing and cooling, also add to the power requirement.
New sources of electricity are needed – and to prevent another catastrophic burden on the climate, they need to be low-carbon. Trump’s executive order mentioned natural gas and coal, but also geothermal power and nuclear.
Even before last week’s announcement, many data centre operators and future owners were planning to build their own power plants – and one technology in particular has so far dominated the conversation.
Small modular reactors (SMRs) have been selected as a promising power source by many tech leaders, including Google, Amazon and OpenAI. Designed to be manufactured in factories then assembled on-site, they could provide the acceleration needed to take the lead in the AI race.
“On-site SMRs provide reliable, baseload, carbon-free power directly at data centres, reducing transmission losses, avoiding grid strain, and enabling faster deployment in two to three years via modular construction,” explains Ed McGinnis, former assistant secretary for nuclear energy at the US Department of Energy, to Professional Engineering. Another advantage is their ability to scale with the power requirements of a data centre campus, he adds, enabling future growth without strain on the wider grid.
The main benefit of SMRs is their “unparalleled energy density,” says Arafat, “delivering approximately 10,000-times more energy per unit mass of fuel than fossil fuels, and significantly more than batteries or solar energy.”
Unlike “complex and monolithic” SMRs modelled closely on gigawatt-scale nuclear plants, Texas firm Aalo aims to provide what it calls ‘XMRs’ (extra modular reactors). Including entirely factory-built power plants, the company hopes its approach could compress deployment timelines from years to months.
A “cluster” of 50MW Aalo Pod reactors could power an entire campus with a fraction of the land area required by renewables, Arafat claimed. They would also provide baseload power, which is useful for AI systems that require near-constant uptime.
Another space-efficient option could be underground nuclear reactors, like those being developed at Deep Fission. The California company announced a partnership with AI data centre developer Endeavour Energy in January this year, to develop two gigawatts (GW) of nuclear capacity from SMRs placed in boreholes a mile underground.
The partners expect their first reactors to be operational in 2029, and are targeting a cost of $0.05-0.07 per kilowatt-hour (kWh) – roughly equivalent to natural gas, according to the Hamilton Project research group.
Small reactors, big demand
Despite the ambitious aims of data centre operators in the US and elsewhere, there is a significant drawback – SMRs are not ready for use, at least for the most part. There are so far only two operating worldwide, and while developers describe their systems in the present tense, most are at least several years away from being ready.
Challenges include high costs and regulatory delays, says McGinnis, who is also president and CEO of nuclear fuel recycling company Curio. Most designs are in the early stages, he adds, with “first-of-a-kind” risks and uncertainties involved. Planning and participation will be needed from the federal government to realise the White House’s ambitions.
“SMRs face many of the same challenges that conventional nuclear power faces. SMRs have mostly the same regulatory and permitting challenges, similar supply chain and fuel cycle challenges, challenges with community support and waste management problems,” says Joshua Loughman, a systems engineer and Arizona State University data scientist, to Professional Engineering.
“If new advanced nuclear technologies like SMRs are also going to take decades to develop, it may be too little, too late to meet the immediate demands for electricity being forecast.”
Published in January this year, before Trump’s latest announcements, a Goldman Sachs Research report projected a 160% increase in data centre power demand by 2030. Up to 85-90 GW of new nuclear capacity would be needed to meet that demand – but “well less than 10%” is likely to be available, it continued, and natural gas, renewables and batteries will all have a significant role to play.
“If we could count on many SMRs being ordered, prices could come down; if prices could come down, demand for more orders would go up,” Loughman says. “This reinforcing feedback loop, supported by a combination of aggressive policy changes (executive orders supporting the nuclear industry, the removal of renewable energy subsidies, and local support), could breathe new life into a mostly lethargic nuclear industry.
“However, the appetite for electricity is right now, and even the shortened delays in this technology’s development could have electricity customers looking for more immediate solutions in the form of renewables, energy storage, energy efficiency and natural gas generation.”
SMR developers are well aware of the short timeframes required, and aim to deploy quickly. Aalo’s pilot factory is already operating in Austin, for example, producing XMRs designed for rapid deployment. A larger, gigawatt-scale factory could deploy in two to three years.
As the effects of climate change accelerate, AI supremacy cannot be the only target of expansion plans. With widespread concerns about the technology’s environmental impact, growth of the sector should be as green as possible. A 100MW data centre running on natural gas can emit over 400,000 tonnes of CO2 each year, Arafat points out.
But as the competition heats up, SMR’s other advantages could push them to the front of the pack. “The case for large corporations… to explore operating their own SMRs comes from their desire for a stable, carbon-free electricity source to power data centres that wouldn’t be at risk of competition with other customers,” Loughman says.
Want the best engineering stories delivered straight to your inbox? The Professional Engineering newsletter gives you vital updates on the most cutting-edge engineering and exciting new job opportunities. To sign up, click here.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.