Meta turns to nuclear power for AI training — asking for developer proposals for small modular reactors or larger nuclear solutions

Nuclear Power Plant
(Image credit: Talen Energy)

AI training is consuming so much power — more than the current electricity grid can provide — that Meta is now actively looking for proposals from nuclear developers to build nuclear power plants for its AI dreams. According to the social media giant’s Sustainability blog, it’s releasing a request for proposals to find nuclear energy developers that will help it get 1 to 4 gigawatts of new generation capacity. The company also said that it’s “taking an open approach” to “partner with others across the industry” to increase the grid’s capacity using nuclear energy.

Meta says that any potential partner must have development opportunities in building small modular reactors (SMRs) or even larger nuclear reactors. They must also have strong community engagement as well as expertise in developing, applying for permits, and executing projects. Through this call for proposals, the company hopes to identify potential partners that will help it get nuclear power sooner and that can scale their output (i.e., deploy multiple units of SMRs) to make nuclear power cheaper.

While the company says that it’s looking for a partner in its nuclear energy project, it also said that the chosen firm must “ultimately permit, design, engineer, finance, construct, and operate these power plants” as it wants an organization that has is fully invested in nuclear technology. This means that Meta is likely not looking to out a significant amount of money to build a power plant — instead, it wants to sign a contract with its chosen provider saying that it will buy a specific amount of power generated at a fixed price.

In other words, Meta is jumping on the nuclear power bandwagon to find the power that it needs to train its AI. Mark Zuckerberg himself said that power will constrain AI growth, and other tech giants have already begun working on nuclear energy sources to supplement their electricity needs. This includes Amazon, Google, Microsoft, and Oracle. Even Westinghouse, which has traditionally been one of the builders of large nuclear plants around the U.S., is working on a microreactor as a means of deploying nuclear energy faster to where it's needed.

The aforementioned companies already inked deals for their nuclear ambitions in 2024, while Meta has set its proposal deadline for Feb. 7, 2025. It might seem that the social media giant is late to the party, but given that it takes years before a nuclear reactor can start operating, it’s probably not off by much. Nevertheless, the fact that Meta is looking to obtain nuclear power sooner means that it recognizes the importance of being among the first to deploy this source of energy for powering future generations of AI models.

Jowi Morales
Contributing Writer

Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.

  • Gururu
    Its almost too funny how we using nuclear for AI and soon AI will use nuclear for us.
    Reply
  • jp7189
    A gigawatt is a lot of juice. A single SMR produces about 1/3 of that. It sounds like the colossus AI supercomputer uses about 0.15 gigawatts. Meta is looking for up to 4 gigawatts? That's about 25,000 acres of solar panels.. give or take.
    Reply
  • palladin9479
    And thus Skynet is born.
    Reply
  • SomeoneElse23
    What could possibly go wrong?
    Reply
  • gg83
    jp7189 said:
    A gigawatt is a lot of juice. A single SMR produces about 1/3 of that. It sounds like the colossus AI supercomputer uses about 0.15 gigawatts. Meta is looking for up to 4 gigawatts? That's about 25,000 acres of solar panels.. give or take.
    Great breakdown. What kinda junk is Zuck dreaming up now. Oh, it's to run his stupid crud-verse.
    Reply
  • LolaGT
    AI or not, what we need are many more of these new gen reactors, dozens and dozens of them, unless people enjoy rolling blackouts.
    We really won't have a choice, unless we all want coal burners in our back yards. Solar and wind are not going to save us, fusion isn't happening any time soon and we are a long ways away from a Dyson sphere.
    Fission will save us.
    Reply
  • GoofyOne
    Y'all have it all wrong ... the AI will just use humans to power itself. Here comes the Matrix!!!
    Reply
  • stonecarver
    Okay I'm just trying to get some kind of idea of lets say Elon Musks 100,000 H-100 GPU's would cost on the grid.

    https://www.tomshardware.com/tech-industry/artificial-intelligence/elon-musk-plans-to-scale-the-xai-supercomputer-to-a-million-gpus-currently-at-over-100-000-h100-gpus-and-counting

    The Nvidia H-100 consumes 350 watts x 100,000.

    100,000 H-100's per hour is 35,000,000 watts per hour x 24 hours = 840,000,000 watt hours per day

    Take that 840,000,000 24 hour number and x it by 365 days = 20,160,000,000 watt hours per year.

    Just to pick an average to bounce off lets say the average household uses 30 kilowatt hours a day.

    30 kilowatt x 365 = 10,950 for a household per year.

    10,950 house hold kilowatts divided by the H-100's of 20,160,000,000 would feed
    1,841,095.8 homes for a year.

    So one million, eight hundred forty one thousand, and ninety five point eight homes per year.

    Forgive me if I'm getting numbers off over a conversion here or there but mind boggling to say the least.
    Reply