Why AI goes nuclear

Should you’re of a sure age, the phrases “nuclear energy” in all probability conjure up dystopian photos of energy vegetation melting down, glowing radioactive waste, protesters, and different darkish scenes starting from the unlucky to apocalyptic.

The reality is, nuclear energy’s popularity has been principally unfairly blemished since Seventies and ’80s because of the Three Mile Island and Chernobyl meltdowns in Pennsylvania and Ukraine (at the moment, a part of the Soviet Union), respectively. Whereas horrible, these disasters belie nuclear power’s true security report, which is definitely significantly better for people and naturally, the atmosphere, than most different energy sources — even renewables, and even accounting for the truth that nuclear waste must go someplace.

Now within the yr 2024, among the largest expertise corporations on Earth are able to embrace nuclear energy once more — and the reason being due to synthetic intelligence (AI).

Which corporations are embracing nuclear to energy AI operations?

Wanting over the past 9-10 months, and particularly, the previous couple of weeks, Microsoft, Google, and Amazon have all introduced large-scale commitments to purchase, put money into, and/or assist construct new nuclear energy vegetation. It’s no coincidence these rivals are the three prime suppliers of cloud computing and cloud storage options on the earth, and have additionally been among the many largest to embrace and supply AI fashions and expertise to prospects, each different companies and end-users.

Particularly, the most important AI-nuclear tasks which have been introduced this yr embody:

1. Google has partnered with Kairos Energy to make the most of small modular reactors (SMRs) to energy its AI knowledge facilities. The deal is projected to ship 500 megawatts of carbon-free energy by 2035, as a part of Google’s broader purpose of working on 24/7 carbon-free power by 2030. These superior reactors provide a simplified and safer design, aligning with Google’s push for sustainability.
2. Microsoft has agreed to restart the dormant Three Mile Island reactor in Pennsylvania by 2028 via a partnership with Constellation Power. This plant will present 835 megawatts of energy, supporting Microsoft’s knowledge facilities as AI power consumption continues to rise. Moreover, Microsoft has signed a contract with Helion Power to discover fusion power, positioning it as a possible future power supply. Earlier this yr, The Data reported that Microsoft and OpenAI have been reportedly partnering on a $100 billion AI supercomputer codenamed “Stargate” that might require 5 gigawatts (5000 megawatts to energy), or simply beneath the quantity of energy consumed usually by New York Metropolis (all for one laptop!!)
3. Amazon introduced on October 16, 2024, that it signed three new agreements to assist nuclear power improvement via SMRs. In Washington, Amazon is working with Power Northwest to develop 4 SMRs, projected to generate 320 megawatts within the first part, with the potential to extend to 960 megawatts. The venture is anticipated to start powering the Pacific Northwest within the 2030s. Amazon is additional partnering with X-energy, which is able to provide the SMR expertise, enabling future tasks to develop greater than 5 gigawatts of nuclear energy. Moreover, Amazon is exploring SMR improvement with Dominion Power in Virginia, including no less than 300 megawatts to satisfy the area’s rising demand. Amazon’s present cope with Talen Power includes a $650 million funding in a Pennsylvania knowledge middle powered immediately by nuclear power, serving to protect an older reactor and creating jobs.

SMRs, as talked about in a number of of the offers above, are reactors with a most output of 300 MWe, producing 7.2 million kWh per day.

They’re smaller than conventional reactors, which exceed 1,000 MWe, and provide higher flexibility as a result of their modular design, permitting for manufacturing and meeting in factories somewhat than on the location of the particular energy station itself.

They’re cooled by gentle water, liquid steel, or molten salt and incorporate passive security techniques, using pure circulation for core cooling and lowering the necessity for operator intervention, which simplifies design and minimizes failure dangers.

What’s driving the transfer to nuclear?

Clearly, the most important cloud-turned AI mannequin suppliers see an infinite future for nuclear energy behind their operations.

However why and why now? To search out out, I reached out to Edward Kee, CEO and founding father of Nuclear Economics Consulting Group, a nuclear power consulting agency, who beforehand labored as a service provider energy plant developer and a nuclear energy plant engineer for U.S. Navy Nimitz-class plane carriers.

In line with Kee — who after all, is incentivized to see extra nuclear energy spin up — the reply is that knowledge facilities used to coach and serve up inferences of AI fashions to prospects require numerous power, and proper now, the one approach to ship it’s largely via a fossil fuel-powered electrical grid, which is able to impede the tech corporations from reaching their local weather and emissions targets.

“The value of clean, reliable electricity for these data centers is pretty high,” he instructed me in a videoconference interview earlier this week. “Most companies have committed to zero-carbon power by 2030 or 2035, but using renewable energy accounting methods is a bit fallacious because solar doesn’t work at night, and wind doesn’t work when there’s no wind.”

Certainly, AI is a very energy intensive {industry}. As Anna-Sofia Lesiv wrote for the enterprise capital agency Opposite final summer time:

“Training foundational AI models can be quite energy-intensive. GPT-3, OpenAI’s 175 billion parameter model, reportedly used 1,287 MWh to train, while DeepMind’s 280 billion parameter model used 1,066 MWh. This is about 100 times the energy used by the average US household in a year.”

And because the Worldwide Atomic Power Company (IAEA), a non-profit worldwide analysis and requirements physique devoted to nuclear power, wrote in a report launched simply this week:

“As electrical energy consumption by knowledge facilities, cryptocurrencies and synthetic intelligence corporations is anticipated to double from 2022 to 2026, these corporations are looking for the following era of fresh power applied sciences that may assist to satisfy their targets.“

Pushed partly by this growing demand from the tech sector, IAEA issued a high-end projection within the report that finds a 150% enhance in international nuclear era capability to 950 gigawatts by 2050.

Nonetheless, the IAEA cautions this high-end projection would require a $100 billion funding over the identical 25-year timeframe — “a fraction of what the world invests in energy infrastructure overall, but a big change from the level of investment in nuclear over the past 20 years.”

Tech corporations try to string a industrial and political needle to get the facility they want

Whereas one may suppose that tech corporations of all entities would don’t have any hassle acquiring energy from the present electrical grid (powered primarily by pure gasoline and coal within the U.S.), the truth in line with Kee is that municipal and personal energy utilities corporations are cautious of committing a good portion of their output to new knowledge facilities, which might pressure their potential to serve their present crop of residential and industrial prospects past tech.

The tech corporations are “talking about adding frankly enormous amounts of new demand in terms of gigawatts on the grid,” the nuclear knowledgeable instructed VentureBeat. “And increasingly, the states and the utilities where they’re going to put those data centers are saying, ‘Hold on a minute, guys. You can’t just show up here and connect and take hundreds of megawatts or gigawatts of power without us having a plan to supply the generation to meet that demand. It’s going to cause problems.’”

Subsequently, with the intention to even get approval for brand new knowledge middle tasks and huge AI coaching “superclusters” of graphics processing models (GPUs) from Nvidia and others — like the sort Elon Musk’s xAI simply turned on in Memphis, Tennessee — municipal and state lawmakers and regulatory companies could also be asking the tech corporations to provide you with a plan for a way they are going to be powered with out draining an excessive amount of from the present grid.

“Talking a lot about your nuclear plants could help you with that in terms of public perception,” Kee mentioned.

Why having nuclear energy positioned bodily and geographically beside knowledge facilities is so interesting

You may additionally suppose that tech corporations trying to nuclear to unravel their AI energetic issues can be completely satisfied getting energy from any nuclear plant, even ones far-off from the place their knowledge facilities can be located.

However regardless that we shoppers typically consider the “cloud” on which many AI servers run as some type of ethereal, nonphysical area of electrons floating above us or round us and that we dip into and out of with our gadgets as wanted, the very fact is it’s nonetheless enabled by bodily steel and silicon laptop chips and {hardware}, and as such, its efficiency is topic to the identical physics as the remainder of the world.

Subsequently, placing knowledge facilities as shut as doable to their energy sources — on this case, nuclear energy vegetation — is advantageous to the businesses.

“We think of this AC power network we have as being pretty much fungible so you can get power at one point and customers another point,” Kee defined. “But when you have huge hundred megawatt gigawatt scale loads, you’re going to have to upgrade and change your transmission system which means a building new transmission lines.”

As an alternative of doing that, the massive tech corporations can be higher off situating servers proper beside the facility era facility itself, avoiding the price of constructing extra infrastructure to hold the huge power masses they require.

What does massive tech’s sudden curiosity in nuclear imply for the long term?

Ever the techno optimist, I personally couldn’t assist however get a little bit huge eyed on the current bulletins of Amazon, Google, and Microsoft placing cash in the direction of new nuclear vegetation.

I actually have gone on a journey of being cautious about nuclear energy to being extra open to it with the intention to assist cut back emissions for the sake of our local weather and atmosphere — very like the environmentalist advocacy nonprofit group the Sierra Membership (based by former Bay Space outstanding resident John Muir), which just lately endorsed nuclear energy to the shock of many given its lengthy historical past of opposition.

A future the place highly effective AI fashions assist enhance the demand for, and possibly even optimize the protection and efficiency of recent nuclear energy vegetation sounds superior and compelling to me. If AI is what it takes the world to look once more at nuclear and embrace it as one of many main sources of fresh power, so be it. Might AI usher in a nuclear power renaissance?

Kee, for his half, is much less sure about that optimistic worldview, noting that whether or not constructing new small modular nuclear reactors (SMRs) or restarting outdated full scale energy vegetation like Three Mile Island, the U.S. federal authorities via the company the Nuclear Regulatory Fee will nonetheless must evaluation and approval all of the tasks, which is prone to take a number of years on the earliest.

“Some of these announcements may be a bit hyperbolic in there on their promises and expectations,” he instructed VentureBeat. “So you want to keep your seatbelt on for a while.”

Nonetheless, having been working within the nuclear sector for many years now, Kee is inspired by massive tech’s lofty guarantees and does consider it might spur new nuclear power funding extra typically.

“There’s been excitement around small and advanced reactors for a decade or more, and now it’s linking up with the big technology power demand world…That’s kind of cool,” he instructed VentureBeat. “I don’t know which other sectors might follow, but you’re right—it could happen. If some of these new reactor designs get built, which was always in doubt because the economics are questionable for the first one, it might become easier to build a whole fleet by other parties, including utilities or municipalities.”