Latest / Minds, Bodies, and Terawatts / Anthropic signs $1.5B/month compute deal with xAI's Colossus data center

Transcript

Anthropic signs $1.5B/month compute deal with xAI's Colossus data center · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where your competitor is also your landlord. Anthropic has signed a deal to pay xAI one-point-two-five billion dollars a month for the entire output of the Colossus data center near Memphis, Tennessee.
• 0:16
  [Guest] Through May 2029, for every GPU in the building. The terms surfaced in SpaceX's S-1 filing with the SEC.
• 0:25
  [Host] Musk posted on X in February that Anthropic, quote, hates Western civilization. Now they're writing his company the biggest rent check in tech?
• 0:35
  [Guest] Rivalry is a luxury when you're GPU-starved. Colossus packs over two hundred thousand Nvidia GPUs into a converted Electrolux factory that Musk's team stood up in months. When the alternative is waiting years for your own facility, pride gets expensive.
• 0:53
  [Host] So xAI built Colossus to train Grok. Why lease it to a direct competitor?
• 0:59
  [Guest] Because Grok demand hasn't matched the capacity xAI built. Rather than let those GPUs sit idle ahead of the SpaceX IPO, Musk turned a cost center into a revenue engine. He built an AI lab and accidentally created a utility.
• 1:17
  [Host] Three hundred megawatts for a single customer. In human terms?
• 1:21
  [Guest] Enough electricity for roughly three hundred thousand American homes. One mid-sized natural gas plant, every electron routed to AI workloads. And xAI's full Memphis complex is scaling into the gigawatt range.
• 1:35
  [Host] But either side can terminate with ninety days' notice. That's a fragile scaffolding for a power plant's worth of compute.
• 1:44
  [Guest] It is. And it tells you how desperate the market is. Anthropic would rather build on a rival's infrastructure with a short exit clause than wait for capacity that doesn't exist yet.
• 1:58
  [Host] They already have a massive partnership with Amazon Web Services. Why isn't cloud enough?
• 2:04
  [Guest] Training a frontier model means tens of thousands of GPUs running nonstop for months. Shared cloud can't guarantee uninterrupted access at that scale. This doesn't replace AWS. It fills a gap AWS structurally can't.
• 2:20
  [Host] Meanwhile, Memphis residents have been protesting the gas turbines xAI installed to power Colossus, saying they worsened air pollution without federal permits.
• 2:29
  [Guest] That's the part that doesn't make the S-1. The facility ran dozens of unpermitted natural gas turbines. The community absorbed the emissions while the compute shipped to San Francisco. The IEA projects data center power demand hitting nine hundred forty-five terawatt-hours by 2030. Someone always pays for those electrons.
• 2:51
  [Host] And Musk ends up profiting from the safety lab he's criticized the loudest.
• 2:57
  [Guest] He built Colossus to win the AI race, and his first major external customer is his most vocal critic. But the facility may prove more valuable than whatever model runs inside it. Models get replaced every generation. A two-gigawatt campus in Memphis doesn't.
• 3:16
  [Host] Patrick Deglon makes this case in Unscarcity. The Foundation layer, energy, compute, the infrastructure of intelligence, can't stay this concentrated.
• 3:25
  [Guest] For more on compute infrastructure and who should own it, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.