Meta is building AI data centers in tents and isn't slowing down — Zuckerberg reveals plans for 'several multi-GW clusters,' including one called Hyperion that's almost as big as Manhattan

News
By last updated

Move fast and build things.

Inside Meta's data center
(Image credit: Meta)

Meta is accelerating its investment in the infrastructure it believes will be needed to power the future of artificial intelligence training and inference, with a number of high-capacity, fast-deployment data centers. CEO Mark Zuckerberg outlined plans for gigantic, "multi-GW clusters" at truly gargantuan scale, and he's doing it extremely quickly by leveraging a number of cost and time-saving measures inspired by xAI. Meta is also building "tented" data centers, rather than traditionally roofed structures.

Even if no one has quite made the AI business model profitable yet, that isn't stopping the world's largest tech companies from investing a collective hundreds of billions of dollars in powering the AI future they are so sure is coming. Alongside Nvidia pushing $500 billion global datacenter investment, and Elon Musk's xAI shipping a power station across the world, Meta is building a number of gigantic datacenters of its own. One of those will reportedly be as large as a sizeable portion of Manhattan.

Although its Llama models might not be capturing the AI zeitgeist like Meta hoped, it may soon be leading the way in domestic AI compute power. Earlier this month, Semi-Analysis reported that Meta will be the first company to build out and deploy a gigawatt-scale datacenter for AI. That will give it over three billion TFLOPS of compute power, we're told. And according to Mark Zuckerberg, this is just the beginning.

Zuckerberg threads.

(Image credit: Meta)

The 1GW "Prometheus" datacenter will come online in 2026, but Meta is also building "Hyperion," he said, which will take a few years to build up to scale and will eventually require as much as 5GW of power. Meta is also working on several other data centers in the longer term, and each of these "titan clusters" could be as large as a sizeable portion of Manhattan.

Zuckerberg threads.

(Image credit: Threads)

The self-styled "Zuck" went on to say that "Meta Superintelligence Labs will have industry-leading levels of compute and by far the greatest compute per researcher," which may hint at the business model Meta is targeting. If it can't build the AI itself, perhaps it could join Amazon and other entities, like Groq, in providing the infrastructure to power it.

That competition will mean Meta needs to scale up fast if it wants to be competitive, and that seems to be exactly what it's doing. Semi-Analysis detailed how Meta has reinvented its datacenter strategy about building them with AI in mind. Meta will reportedly use a new "tent" style of structure that is faster to build and has less redundancy to further increase the pace of deployment. Where traditional data centers have backup diesel generators should they encounter a power outage, the new Meta sites will do away with that entirely.

They will also leverage pre-fabricated power and cooling modules and will draw power from on-site meta substations, with smart workload management tools helping to maximize the power available to each datacenter. That can also include on-site natural gas generation, too, with Meta already building two 200MW natural gas plants in Ohio.

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

TOPICS
Jon Martindale
Freelance Writer
3 Comments Comment from the forums
  • rluker5
    Building an AI datacenter nearly the size of Manhattan with tent style structures for at least parts of it, other companies using AI GPUs as borrowing collateral to buy more AI GPUs. While some AI datacenters sit idle in China and Amazon not being able to find a better way to make money with AI than to rent out capacity.

    A new gold rush. Other than finding several copies of all of our data, figuring out what we are likely to buy and how to charge us the most for it, replacing some human interaction and data processing jobs, government monitoring, and data security breaches, what is it good for other than FOMO hyped equipment sales?

    I only see a few labor efficiency benefits and mostly downsides otherwise. Not including short term financial gains of course. I wonder how long before there will be more and more laws and restrictions to keep the general populace from being exploited from those looking to the AI industry to profit from the losses of material wealth, IP, privacy, freedom, and security of the general population?
    Reply
  • Flemkopf
    So, they're planning to put billions of dollars of equipment powered by a gigawatt substation...under a tent? With one of those datacenters being put in a state that had over seventy tornadoes last year, plus flooding, and they're putting in a minimum of redundancy?

    I saw the remains of a twenty acre agricultural greenhouse that got shredded by a tornado a few years ago. Tomatoes and peppers are a heck of a lot cheaper to replace than GPU servers. Heck, for that matter what is Zuck planning to do when it rains? Ten square miles of tenting means that you're dealing with 500 acre-feet of water running right through your facility every time you get an inch of rain.

    Then you've got the heat. Every gigawatt averages out to about two square miles of solar exposure in Ohio. That's going to have an impact in a localized area and no, the locals are not going to be happy about you boiling their nearby lake.

    All this for a product that doesn't have a path to profitability, leading to a scrapped data center withing three or four years. But at least they won't have to take down a building when that happens.
    Reply
  • SomeoneElse23
    I have to wonder what the real reason is behind this "AI" hype.

    It's certainly not what they say it is.
    Reply