- Meta built a massive AI data center cluster in Ohio
- And it innovated several new infrastructure techniques for that data center
- Now, the company is building the world's biggest data center in Louisiana
OCP GLOBAL SUMMIT, SAN JOSE, CALIFORNIA — The Open Compute Project (OCP) conference has been around for about 15 years, but 2026 is definitely its biggest year, so far. About 15,000 people are attending the sold-out conference to get together and talk about compute hardware. And more specifically, to talk about compute hardware and infrastructure for AI.
A speaker from Meta Platforms kicked off the conference this afternoon with his presentation about Meta’s AI data centers. And he kind of blew everyones’ minds.
Santosh Janardhan, head of Global Infrastructure with Meta, first talked about solving multiple problems for the company’s huge data center complex in New Albany, Ohio, which is named Prometheus. And then he gave a preview of Meta’s new, gargantuan data center in Richland Parish, Louisiana.
Prometheus includes over 30 structures, millions of miles of fiber-optic cables and tens of thousands of GPUs. “Constructing it ended up being quite an epic journey,” said Janardhan.
One of the first problems Meta had to solve was the networking between all the different structures.
Historically, data centers have been built as a single building, under one roof. But in the case of Prometheus with its multiple structures, the distances between the buildings would have caused unacceptable time delays for connectivity. “When you do AI training, it moves forward at the speed of the slowest conversation that you can have between two GPUs,” said Janardhan. “The physics is simple. You essentially want to create a lossless network.”
To do this, Meta created what Janardhan called an “Ethernet super spine.” He said, “You take the spines of all the buildings in the region, string them together and aggregate them locally.”
The advantage is that there are no stops, so no single points of failure. “And the other thing to understand here is that this thing is fast — tens of petabits a second,” he said. “You can transmit the whole Library of Congress in less than 30 seconds.” (For a deeper technical dive into Meta’s data center inter-building networking, check out this blog.)
He also noted the massive amounts of fiber optic cable needed for Prometheus, showing the below photo of some of the cables.
“Each one of those pipes that you see are serving actually thousands of strands of fiber,” he said. “In fact, if you aggregate all the network fiber we have in Prometheus, it's about three to four million kilometers of fiber. You can go to the moon and back, four times. The physical manifestations of AI training truly boggles your mind.”
Another problem that Meta solved was speeding up the construction of buildings. Historically, data centers have been built with steel and concrete. But Meta re-thought that because it wanted to build its data centers faster.
So, it decided to build tents, instead.
“Each one of our tents is the length of minimum two football fields, he said. These are humongous. And this takes months to build, not years. Because time-to-market matters.”
But he said the tents are “fairly robust” — rated for a category 1 hurricane — “because you’re putting in billions of dollars’ worth of equipment in it.”
While it was reimagining everything about data centers, Meta figured — why not also reformulate concrete?
Janardhan said, “I don't think people understand how much concrete is used in the world.
And it’s not great for the environment. It takes a lot of energy and water to create concrete. And when cities are paved with too much concrete, it creates an unpleasant heat-island effect, which contributes to more global warming.
"So what we did is we fundamentally went after the composition of concrete,” said Janardhan. “We built an AI tool to rethink how concrete works, and the new concrete that we ended up getting uses 19% less carbon, cures 43% faster, and is completely cost-neutral.”
Meta is open sourcing its AI tool for the concrete formulation. “All of you can go and use it,” he said. “Use it in your data center. Use it in your backyard.”
Hyperion will be even bigger
Now, Meta is taking the lessons it’s learned from Prometheus and applying them to its next AI data center cluster — Hyperion, in Louisiana.
“That’s going to be absolutely mammoth,” said Janardhan. It's a five-gigawatt cluster at a single location. To the best of my knowledge, it’s the biggest data center on the planet. It covers about 4,000 acres.”
At this point in his speech, the audience was in awe with the magnitude of Meta’s innovation and the sheer scale of its projects.
But then, Janardhan said, “What matters is what runs on this, right? At the end of the day, every call, every message, every interaction, every post you do on one of the properties, that has real meaning to someone, right?”
And then the audience realized that this was all for properties such as Facebook and Instagram, and that blew the wind out of the sails for many of us.
RELATED STORIES:
Meta goes all-in on nuclear power with 6.6 GW plans
To Meta, data centers are the new consumer electronics
Opinion: Fiber’s next act: From broadband to the infrastructure of the physical AI economy