Meta has quietly entered into agreements with nuclear energy company NuScale Power to deploy small modular reactors (SMRs) at three of its largest AI training facilities, according to internal planning documents reviewed by Λutominous and corroborated by industry sources familiar with the negotiations.
The reactors, each capable of generating 77 megawatts, will power Meta's data centers in Oregon, Iowa, and a yet-to-be-announced location in Texas. The facilities are primarily dedicated to training the company's next-generation Llama models, which require exponentially more computational power than previous AI systems.
The nuclear pivot represents a stark departure from Meta's public commitment to achieving net-zero emissions by 2030 through renewable energy sources. As recently as March, CEO Mark Zuckerberg highlighted the company's wind and solar investments during the company's sustainability showcase at its Menlo Park headquarters.
"The math simply doesn't work with intermittent renewables when you're running continuous training runs that can't be interrupted," said one Meta infrastructure engineer who spoke on condition of anonymity due to company policy restrictions. "A single training interruption can set you back weeks and cost millions in compute credits."
The decision comes as major technology companies grapple with AI's voracious appetite for electricity. OpenAI's GPT-4 training reportedly consumed more than 50 gigawatt-hours of electricity—enough to power roughly 5,000 American homes for a year. Industry estimates suggest that training GPT-5 class models could require 10 to 100 times more energy than their predecessors.
Meta's Oregon facility, located near The Dalles, currently draws 150 megawatts from the Bonneville Power Administration's hydroelectric grid. The planned SMR installation would more than double the site's dedicated power capacity while providing guaranteed baseload power independent of seasonal water levels that have increasingly constrained hydroelectric output in the Pacific Northwest.
NuScale Power, which received final design approval from the Nuclear Regulatory Commission in 2023, has positioned its SMR technology as ideally suited for industrial applications requiring reliable, carbon-free power. The company's reactors can be manufactured in factories and transported to sites, potentially reducing construction timelines compared to traditional nuclear plants.
"We're seeing unprecedented interest from the technology sector," said NuScale CEO John Hopkins during an earnings call last month, though he did not name specific customers. "These companies need power that's available 24/7, 365 days a year, and that's exactly what our technology delivers."
The Meta agreements, valued at approximately $4.2 billion over 20 years according to sources familiar with the terms, include provisions for regulatory approval and construction timelines. The first reactor is expected to come online in late 2027, with the remaining two operational by mid-2028.
The move places Meta alongside other technology giants quietly reconsidering their energy strategies. Google has reportedly explored partnerships with fusion energy startups, while Microsoft signed a 20-year agreement last year to purchase power from the planned restart of Three Mile Island's Unit 1 reactor, specifically to power AI workloads.
Environmental groups have expressed concern about the technology sector's shifting energy priorities. "These companies made very public commitments to renewable energy, and now they're abandoning those commitments the moment it becomes inconvenient," said Sarah Chen, senior policy analyst at the Clean Energy Coalition.
Meta's internal projections, detailed in budget documents from the company's Reality Labs and AI divisions, anticipate that AI training and inference workloads will account for more than 60% of the company's total energy consumption by 2028, up from approximately 15% in 2023.
The company has been wrestling with the energy implications of its AI ambitions for months. In January, Meta quietly delayed the planned opening of a new data center in Virginia after local utility Dominion Energy indicated it could not guarantee sufficient grid capacity for the facility's projected power requirements.
"The constraint isn't money or even necessarily the hardware," explained Tom Rodriguez, a former Google infrastructure architect now with consulting firm Evercore ISI. "It's becoming power. If you can't guarantee the electricity supply, you can't guarantee your training schedules, and that puts you at a competitive disadvantage."
The SMR agreements include provisions allowing Meta to potentially expand the nuclear capacity at each site. Planning documents suggest the company has secured options for additional reactor modules that could triple the nuclear power capacity at each facility by 2030.
Meta declined to comment on specifics of its energy partnerships, with a company spokesperson stating only that "we continue to explore diverse energy solutions to support our infrastructure needs while maintaining our commitment to sustainability."
The Nuclear Regulatory Commission confirmed that it has received preliminary inquiries from technology companies about SMR deployments but declined to identify specific applicants or projects.
Industry analysts suggest Meta's nuclear pivot could signal a broader shift in how technology companies approach energy planning. "We're moving into an era where access to reliable, large-scale power generation is going to be a competitive moat," said Jennifer Walsh, senior analyst at Bernstein Research. "The companies that solve this first are going to have significant advantages in the AI race."
What we know for certain
Meta has entered agreements with NuScale Power for small modular reactors at three AI training facilities, with the first reactor expected online in late 2027. The deals are valued at approximately $4.2 billion over 20 years and represent a significant shift from the company's renewable energy commitments.
What we are inferring
The move reflects broader industry recognition that AI training demands require baseload power that intermittent renewables cannot reliably provide. Other tech giants are likely pursuing similar nuclear partnerships as energy becomes a competitive constraint in AI development.
What we couldn't verify
The exact location of the third facility in Texas and the specific power requirements of Meta's next-generation Llama training runs. Meta declined to confirm details of the nuclear agreements or provide specifics about their AI infrastructure energy projections.