Data center power constraints and burgeoning AI workloads have companies scrambling to find new sources of electricity.
Fully 40% of existing AI data centers will be “operationally constrained by power availability” by 2027, Gartner predicts, and the research firm estimates the demand from incremental AI servers will reach 500 terawatt-hours per year by then, which is 2.6 times the 2023 level.
“The explosive growth of new hyperscale data centers to implement [generative AI] is creating an insatiable demand for power that will exceed the ability of utility providers to expand their capacity fast enough,” said Bob Johnson, vice president analyst at Gartner. “In turn, this threatens to disrupt energy availability and lead to shortages, which will limit the growth of new data centers for genAI and other uses from 2026.”
While sources like wind and solar provide some relief, they are constrained, noted Jay Dietrich, research director of sustainability at Uptime Institute.
“The issue here is that as the quantity of wind and solar grows on the grid, it’s very effective when it has a high generation capacity. But you end up with about between 10% and 20% of the time, depending on what the capacity penetration is, where you don’t have carbon-free energy to keep the grid stable,” Dietrich said.
That’s why, in an effort to find new energy sources—and in the face of the push to make it clean energy—data center owners are turning to nuclear power.
“Where these nuclear systems fit is their reliable power,” Dietrich said. “They’re reliable carbon-free power, and the current reactors are proven.”
Hyperscalers make nuclear plans
AWS, Google, Meta, Microsoft, and Oracle are among the companies exploring nuclear energy.
“Nuclear power is a carbon-free, reliable energy source that can complement variable renewable energy sources like wind and solar with firm generation. Advanced nuclear reactors are considered safer and more efficient than traditional nuclear reactors. They can also be built more quickly and in a more modular fashion,” said Amanda Peterson Corio, global head of data center energy at Google.
“As the digital economy grows, the services requested by Google’s customers will require a greater amount of generation to power our data centers,” Peterson Corio said. “As a corporate buyer with an ambitious goal to achieve 24/7 carbon-free energy for our operations, Google aims to meet that growth with carbon-free electricity that can offer reliable ‘firm’ (aka around-the-clock generation) power. Nuclear power has the benefit of providing both the ability to scale without adding more carbon to the atmosphere and it can provide that power every hour of the day.”
Most data center operators are opting to partner with power generators in some fashion, not set up fission reactors on the back lot. Google, for example, has entered into an agreement with Kairos Power to buy energy generated by its small modular nuclear reactors (SMR), which Kairos plans to have online by 2030.
However some may be considering running their own installations—during Oracle’s earnings call in September, CTO Larry Ellison said the company has acquired building permits for three SMRs to power a data center with more than a gigawatt of AI compute capacity in an undisclosed location.
What’s an SMR?
A small modular reactor, as its name suggests, is a smaller version of the nuclear fission reactor in a commercial power plant, with a commensurately lower output—typically up to 300 megawatts of electricity output (MWe), compared to 1000Mwe or more. It is designed so that its systems and components can be built in a factory and transported to the site for installation. This makes an SMR more affordable to build than a larger reactor, according to the International Atomic Energy Agency (IAEA), since large reactors are often custom-designed for their location and are built from scratch onsite.
In addition, noted Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, “SMRs are not a smaller version of Three Mile Island or Chernobyl. In addition to being quite a bit smaller, they require less fuel (they can last up to 30 years on a fueling), have advanced and modern safety features, and are designed to use physical sciences to enhance that safety and resilience. They can also use a variety of coolants.”
“One of the more interesting aspects of SMRs is this ability to kind of right size for the operating environment,” he added. “While these reactors can support up to 300 MWe, they can also scale down to as low as 10 MWe. I hate to use the t-shirt size example—and maybe this isn’t even a fair [analogy]. This is more like going to a tailor to get exactly the right size.”
And when that size is no longer the right size, users can simply add another module to expand capacity.
Except that it’s not quite that simple. There are, at last count, more than 70 designs in various stages of development, very few of which are actually approved for use. Some are cooled with water, like existing nuclear plants, but others use inert gases, molten salts, or liquid metals as coolants. Google partner Kairos, for example, works with molten fluoride salt, which, said Peterson Corio, “in particular enhances safety because, unlike water-cooled reactors, it does not need to be highly pressurized, and it solidifies at ambient temperature.”
On the other hand, SMR manufacturer and nuclear fuel recycler Oklo, which is backed by OpenAI CEO Sam Altman, builds low-output (15MW or 50MW per unit) SMRs that are cooled with liquid metal. Its fast fission design uses recycled nuclear waste from traditional reactors. The company says that it has received letters of intent from two major data center providers, as yet unnamed, to purchase up to 750MW of its power.
AWS is putting its money into yet another SMR from X-energy, which also has developed TRISO-X, a custom version of a U.S. Department of Energy endorsed fuel, TRISO (TRi-structural ISOtropic particle). AWS has signed an agreement with Energy Northwest, which will construct, own, and operate an installation using the X-energy SMRs, to fund the feasibility stage of the project. It has also signed a memorandum of understanding with Dominion Energy to explore the use of SMRs in Virginia.
And yet another SMR vendor, NuScale—the only one so far with a design approved by the U.S. Nuclear Regulatory Commission (NRC), as it happens—uses pressurized water cooling. Its largest model can generate 77MWe per module, for a total of 924MWe, and can go without refueling for up to 21 months.
Data center colocation
Not all data centers are embracing SMRs, though. AWS, Meta, and Microsoft are all looking at colocation with standard nuclear power plants.
Colocation, in this context, means that the data center is not only close to the power plant, but also it receives its electricity directly, bypassing the electric grid. It’s a win-win situation for the co-located customer and the operator; the customer gets a long-term commitment for a supply of the power it needs, and the power operator has a known source of revenue.
Microsoft, for example, has made a 20-year power purchase agreement (PPA) with Constellation to purchase power from the currently decommissioned Unit 1 at the Three Mile Island nuclear facility, assuming Constellation can leap the many regulatory hurdles and revive it. Constellation plans to spend $1.6 billion to revamp and update the plant, to be renamed the Crane Clean Energy Center (CCEC).
“Restarting a nuclear reactor requires U.S. Nuclear Regulatory Commission approval following a comprehensive safety and environmental review, as well as permits from relevant state and local agencies,” Constellation said in a release. “Additionally, through a separate request, Constellation will pursue license renewal that will extend plant operations to at least 2054. The CCEC is expected to be online in 2028.”
“I can see this happening,” Kimball from Moor Insights & Strategy said. “It is a lot easier to modernize and restart an existing site than it is to build from scratch. In addition to Three Mile Island, there are another dozen or so deactivated sites that could be reactivated to meet the needs of data centers—and more!”
Unfortunately for Meta, one of its colocation projects had to be canceled in the wake of not only regulatory questions but also an intervention by Mother Nature, after a rare species of bee was discovered on the site where it planned to build its data center.
Amazon, on the other hand, merely bumped heads with regulators—or rather, the operators of the grid and the Susquehanna, Pennsylvania, power plant did when they asked regulators to let them up the agreed-upon amount of power directed to the Cumulus data center, which Amazon had purchased from Talen Energy for $650 million, from 300MW to 480MW. The U.S. Federal Energy Regulatory Commission (FERC) denied the request in a split 2-1 ruling. In this case, Amazon has said it plans to continue the project.
The regulatory issues
Regulators everywhere are keeping an eye on all of these activities, and they’re working to streamline the processes required to examine and approve the new technologies.
For example, the U.S. Nuclear Regulatory Commission (NRC) is working with its counterparts in other countries to develop rules around new nuclear power plants, and FERC is scrutinizing the electricity component.
“The NRC has, for the last few years, been reviewing both preliminary information and full applications for small modular reactors, including designs that cool the reactor fuel with inert gases, molten salts, or liquid metals (as opposed to today’s water-cooled designs). Our reviews have generic schedules of 2 to 3 years, depending on the license or permit being sought,” said Scott Burnell, public affairs officer at the NRC. “We’re also in the process of adding a review process that more easily applies to new designs, as well as adding a generic environmental impact statement that will make environmental reviews far more efficient.”
In addition, he said, the U.S. has signed a memorandum of cooperation with Canada and the United Kingdom to collaborate on technical reviews of advanced reactor and SMR technologies, and it also continues to work with Poland.
Burnell said that, while a single NuScale design is the only SMR currently acceptable for use in the U.S., the NRC is reviewing a larger version of the module. It has also issued construction permits for test units to two companies, including Kairos, and is expecting discussions with X-energy and Energy Northwest about the Amazon project.
“The utilities or reactor vendors would own and operate the facilities and be responsible for meeting NRC regulations without input from Amazon or Google,” he explained.
FERC and its stakeholders’ concerns have been two-fold: the impact of data centers sucking up excessive amounts of power, thus causing shortages for other customers and potentially raising rates, and the fact that co-located loads would not be paying for the necessary maintenance and expansion of the grid.
However, FERC’s chair, Willie L. Phillips, had another concern. “Today’s order [to deny the request to increase power to the Amazon data center] also creates a national security risk,” he wrote in his dissention to the ruling. “There is a clear, bipartisan consensus that maintaining U.S. leadership in Artificial Intelligence (AI) is necessary to maintaining our national security. Maintaining our nation’s leadership in this ‘era-defining’ technology will require a massive and unprecedented investment in the data centers necessary to develop and operate those AI models. And make no mistake: access to reliable electricity is the lifeblood of those data centers. I am deeply concerned that in failing to demonstrate regulatory leadership and flexibility we are putting at risk our country’s pole position on this critically important issue. That is simply unacceptable.”
What’s the future?
Analysts agree that nuclear is an essential part of a carbon-free, AI-burdened electric grid.
“The attraction of nuclear in a world where you’re trying to take the grid to carbon-free energy is that it is really the only proven reliable source of carbon-free energy, one that generates whenever I need it to generate, and I can guarantee that capacity is there, except for the refuel or the maintenance periods,” Uptime Institute’s Dietrich pointed out.
“[SMRs] will be the preferred approach, but recognizing that the economics still have to be worked out,” he said. Things like the cost of finding suitable sites and building the installations, the fueling cycle, and the integration of multiple modules and their controls to maintain constant power output, even during maintenance or refueling of one module, need to be defined.
“Data centers need SMRs, for the same reason the rest of the grid needs SMRs,” agreed John Annand, research practice lead at Info-Tech Research Group. “Modern society depends and evolves on the assumption of the availability of limitless (and cheap) energy. Not only do we need additional electricity generation capacity, but we also need it to be more efficient and more environmentally friendly. Nuclear, for the last point, is debatable by some, but there is no question that SMRs very neatly solve the first two problems. A smaller footprint (as low as 5 acres) means they can be built in a lot more places than traditional nuclear power plants, and because they can be located close to where the electricity is needed, not as much is lost in transmission. That could mean another eight to 12 percentage points of efficiency, which, given our increasing demand, has a huge impact. And data centers have a demand!”
He added, “We were already on an upward trajectory, and the power-hungry demands of GPUs and AI have only turned that conversation up to 11. SMRs can theoretically be placed right next to a data center, like portables [classrooms] next to a high school whose student body has swelled beyond expectations. It’s a very elegant solution as we’re bringing power to where it is needed, whereas historically, we located data centers where power was cheapest.”
“However,” he said, “increased capacity has a way of inducing more consumption. The bigger question is, are we getting value out of that consumption? AI has generated a lot of hype, which in turn demands a lot of electricity be generated. Dealing with the hype rather than the symptom might be a better long-term solution.”
Source:: Network World