Latest / Minds, Bodies, and Terawatts / SpaceX Partners with Anthropic for Compute Power

Transcript

SpaceX Partners with Anthropic for Compute Power · Minds, Bodies, and Terawatts

• 0:00
  [Host] Here's the revised script. Key changes: (1) split the heavy compute-deals data dump across two exchanges so the host interrupts mid-list, (2) split the dense opening guest turn, (3) added the IEA's "Energy and AI" report by title with specific findings to fill the Article Engagement gap. ---
• 0:21
  [Host] This is Minds, Bodies, and Terawatts, where your competitors become your landlords. The Wall Street Journal and Axios reported yesterday that Anthropic has signed a deal to use the entire compute capacity of SpaceX's Colossus 1 data center in Memphis, Tennessee.
• 0:40
  [Guest] That's 300 megawatts and 220,000 Nvidia GPUs, H100s, H200s, and GB200 accelerators, all online within the month.
• 0:48
  [Host] Elon Musk's company powering his direct AI competitor. How does that happen?
• 0:53
  [Guest] SpaceX absorbed xAI earlier this year in a merger Forbes valued at $1.25 trillion. xAI moved its training to the newer Colossus 2. The Information reports Colossus 1 was running at just 11% GPU utilization, versus about 40% at rival clusters. Leasing it converts dead capital into revenue.
• 1:15
  [Host] And Anthropic was desperate enough to take that call?
• 1:19
  [Guest] Dario Amodei told CNBC that his company planned for 10x growth in Q1. They got 80x, annualized, driven mostly by Claude Code. Rate limits tightened, users complained, and Fortune reported backlash over degraded Claude quality as early as April.
• 1:37
  [Host] But they already have massive compute deals lined up.
• 1:41
  [Guest] Five gigawatts with Amazon, five with Google and Broadcom, $30 billion of Azure with Microsoft, $50 billion with Fluidstack.
• 1:50
  [Host] So why lease from a rival?
• 1:53
  [Guest] None of it delivers before late 2026. Forbes calls it the largest forward compute commitment in AI history, but forward doesn't help when you're throttling paying customers today. Anthropic needed GPUs this month.
• 2:08
  [Host] What do 220,000 new GPUs actually change for users?
• 2:12
  [Guest] Same day as the announcement, Anthropic doubled Claude Code's usage limits and removed peak-hour restrictions for paying subscribers. Straight line from data center lease to product improvement.
• 2:24
  [Host] Does this reshape the competitive map?
• 2:27
  [Guest] It collapses the old boundaries. Anthropic competes with xAI on models, depends on SpaceX for compute, and partners with Amazon, Google, and Microsoft simultaneously. Compute scarcity overrides every strategic preference, including who you're willing to do business with.
• 2:44
  [Host] What does SpaceX get beyond rent?
• 2:47
  [Guest] A marquee AI customer right before its IPO. Reuters reported SpaceX is preparing to go public. Showing Wall Street your data center hosts the top AI lab is a better pitch than 11% utilization on idle hardware.
• 3:04
  [Host] 300 megawatts powers about 300,000 homes. Where does that energy come from?
• 3:10
  [Guest] Natural gas turbines that bypass the Memphis grid entirely. Civil rights groups have challenged the setup as potentially illegal. Two years ago this building made refrigerators. Now it draws more electricity than some small cities.
• 3:25
  [Host] The IEA's "Energy and AI" report projects data center electricity demand doubling to 945 terawatt-hours by 2030, about 3% of global electricity. Demand grew 17% in 2025 alone.
• 3:38
  [Guest] And this deal shows exactly why those forecasts keep getting revised upward. When Anthropic plugs into 300 megawatts overnight, that load wasn't in any grid model. Each new AI lab scrambling for capacity adds demand that planners didn't see coming.
• 3:57
  [Host] Both companies also floated data centers in orbit. Two years ago that's a punchline.
• 4:03
  [Guest] Amodei called the growth "too hard to handle." The announcement says Anthropic expressed interest in "multiple gigawatts of orbital AI compute." SpaceX owns the rockets, and in orbit there's no permitting board and no grid to overload.
• 4:20
  [Host] Compute infrastructure is starting to look like oil pipelines did in the twentieth century.
• 4:26
  [Guest] Nations with cheap, reliable power hold the advantage. Companies controlling both energy and delivery become gatekeepers to AI itself. This deal isn't about renting GPUs. It's about who owns the physical layer of intelligence.
• 4:42
  [Host] Patrick Deglon calls this the Electron Gap in the book Unscarcity. AI's real bottleneck was never algorithms, it was always energy.
• 4:51
  [Guest] For more on the Electron Gap and compute concentration, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.