AI to go nuclear? Data center deals say it’s inevitable

AWS, Microsoft, and Google are going nuclear to build and operate mega data centers better equipped to meet the increasingly hefty demands of generative AI.

Earlier this year, AWS paid $650 million to purchase Talen Energy’s Cumulus Data Assets, a 960-megawatt nuclear-powered data center on site at Talen’s Susquehanna, Penn., nuclear plant, with additional data centers planned — pending approval by the Nuclear Regulatory Agency.  

Microsoft, Google, and NuScale Power, a maker of small module nuclear reactors known as SMRs, approached Constellation Energy this spring with a request for information about cooperating on a possible SMR and to secure contracts to access nuclear power from the Baltimore power company in the immediate future, says a spokesperson for Constellation, one of the nation’s largest nuclear power providers.

“The data economy and Constellation’s nuclear energy go together like peanut butter and jelly. And as such, we are in advanced conversations with multiple clients, large, well-known companies that you all know about powering their needs,” said Joe Dominquez, Constellation’s CEO during a company conference call in May.

Last year, Constellation signed a deal giving Microsoft the rights to receive up to 35% of its power from nuclear sources in addition to its existing solar and wind purchases from Constellation for Microsoft’s Boydton, Va., data center. Microsoft has also signed a nuclear carbon credits deal with Ontario Power Generation for its operations in Canada.

In addition to its purchase of the Cumulus data center, AWS will have access to nuclear energy as part of a 10-year Power Purchase Agreement (PPA) from the Susquehanna site.

Many of the deals under discussion are with existing nuclear power providers for hyperscalars to access energy or to employ SMRs with smaller carbon footprints that will be annexed to existing nuclear power plants. NuScale, Oklo, Rolls-Royce SMR, Westinghouse Electric, Moltex Energy, Terrestrial Energy, General Electric, Hitachi Nuclear Energy, and X-energy are among the roster of companies with SMRs under development to meet the growing needs of AI data centers.

Oklo, which is chaired by OpenAI founder Sam Altman, has built a fast fission nuclear reactor dubbed Aurora and intends to sell its energy and SMRs to the US Air Force and data centers by 2027.  

“This global arms race for power arose pretty quickly, and it’s like nothing we have ever seen before.”

— Peter Kelly-Detwiler, principal, Northbridge Energy Partners

Microsoft co-founder Bill Gates, a longtime advocate for nuclear innovation, co-founded TerraPower, which broke ground in Kemmerer, Wyo., this summer on a new nuclear power plant dubbed Natrium that uses salt for cooling and is intended to be operated as a commercial power plant.

To date, the Nuclear Regulatory Commission has not given the green light on any of these projects to go live.

Making the case for nuclear-fueled AI

One energy analyst does not expect nuclear SMRs to be operational until 2030, yet he and many others acknowledge the need for sustainable, carbon-free alternatives to electricity, wind, and solar is very pressing.

“Today’s electric grids are struggling to keep up with demand, even as datacenter companies are planning huge new additions to their fleets to power generative AI applications. As a result, companies like Google, Amazon, and Microsoft are increasingly taking matters into their own hands and getting creative. They are now looking at on-site nuclear-based SMRs, and even fusion reactors,” says Peter Kelly-Detwiler, principal of Northbridge Energy Partners. “This global arms race for power arose pretty quickly, and it’s like nothing we have ever seen before.”

“Power scarcity is a real concern,” notes Sean Graham, research director of cloud to edge data center trends at IDC. “Hyperscalars are absolutely making big bets on generative AI and generative AI is energy-intensive and requires an order of magnitude more power than general purpose computing.”

John Marcante, US CIO in Residence at Deloitte and former global CIO of Vanguard, isn’t surprised that hyperscalers are contemplating nuclear energy to ramp up for the AI era. But he also notes that there will be innovations within the data center itself — in the GPUs and agile architecture — to pinpoint and address consumption needs.

“AI and in particular the emergence of generative AI requires large, data-intensive compute power. This is why companies like Constellation Energy and others are looking at the possibility of smaller modular reactors along with additional renewable energy sources to power future demand,” Marcante tells CIO.com.

“The future will be dependent on increased innovation in hardware and chip efficiency and advances in data center technology like liquid cooling plus advances in edge computing, which reduces the dependency on large, centralized centers,” he maintains.

Mega data centers on tap

Still, nuclear-fueled data centers are not within reach of just anyone, CIOs included. Only major data center operators such as Amazon, Microsoft, Google, Meta, and Apple have deep enough pockets and industry influence to enable access to nuclear assets or build a nuclear-based SMR market, says Ashish Nadkarni, group vice president and general manager of worldwide infrastructure research at IDC.

“If they can demonstrate that the nuclear SMRs are quote-unquote ‘green’ and they are relatively low risk than some of the other carbon-based power generators, then I suppose it has a promising trajectory,” Nadkarni says. “But how are they going to take care of the black-eye situations that nuclear power has …. managing nuclear waste and spent fuel? On paper, at least, nuclear power is the best green power we can get and deploy but it can lead to catastrophe.”

All the hyperscalers have mega data centers under development in rural regions including Iowa, Wyoming, and throughout the Midwest, where the existing power grid and supplemental sources of energy such as wind and solar are more affordable. Unlike Data Center Alley in northern Virginia, the Rust Belt is becoming dotted with this new class of mega data centers designed for advanced AI and compute-intensive workloads.

“On paper, at least, nuclear power is the best green power we can get and deploy but it can lead to catastrophe.”

— Ashish Nadkarni, GVP and GM of worldwide infrastructure research, IDC

Aside from the large cloud providers, Meta, for instance, has invested $800 million on a data center in Cheyenne, Wyo., being built for the next era of generative AI. “The Cheyenne Data Center will be optimized for AI workloads,” according to a digital post by Fortis Construction, Meta’s builder in Wyoming and other states.

Aside from using unorthodox sources of energy, these mega data centers are being built to offer maximum agility and unprecedented capabilities for the new era of AI and computing services.  

IDC’s Nadkarni says these advanced data centers now under development, for example, will offer a “composable and disaggregated infrastructure,” in which all aspects of computing can be decoupled. In this more agile infrastructure, “you can decouple processors, memory, and connectivity, and this idea of composable is that you can compose units of compute and storage in a software-defined manner on the fly,” the IDC analyst says.

It’s unclear to what extent these advances in infrastructure will impact power consumption, but CIOs agree that data scientists and engineers have known that sustainability and efficiency are two major design consideration for generative AI applications and services.

“The ironic thing is that AI can be used to more efficiently manage smart grids and technology infrastructure, helping to solve some of the demand issue it creates,” Deloitte’s Marcante says. “Additional innovation with respect to distributed AI architectures is also on the horizon. Most certainly the future will encompass a culmination of creative solutions given the potential of AI.”

Bob McCowan, global CIO of Regeneron Pharmaceuticals in New York, says these mega data centers will be needed in the medium to long termso “getting a head start on these is a good idea. I also support the need to look at nuclear power plants. We have been using these safely for generations and with the right operating models they can be safe and secure.”

McCowan also claims there is an artificially high demand for AI computational power, as many people and organizations are experimenting with generative AI but demand for energy and computing will balance out with experience.

“The hype has created demand and often misuse, but it’s par for the course as the true use patterns and value is determined,” McCowan says. “I expect to see many organizations ‘right size’ in a similar way we saw cloud computing being right sized and demand will become more predictable. That said, demand will increase, and I think those investments in building infrastructure will pay dividends, even if the timing is still a little unclear.”

US Patent & Trademark Office CIO Jamie Holcombe admits he is glad the tech titans — and not enterprise CIOs — will be tasked with building and managing power for the next generation of data centers.

“That’s why CIOs depend upon the large cloud providers to be the experts on producing ample, redundant, and reliable power — so that we can focus on our core mission, which is not delivering electricity-as-a-power service,” he quips.

Krishna Prasad, CIO and chief strategy officer at UST, says SMRs are worth a shot to meet increasing energy demands and carbon-free requirements.

“Companies building data centers are going to have to find innovative ways to fulfill their energy needs, while meeting their sustainability goals,” he says. “While the long-term viability of distributed, smaller scale nuclear plants is still unclear, this is a solution that holds promise.”