Latest / Minds, Bodies, and Terawatts / SpaceX Signs $6.3B Compute Deal with Reflection AI, Cementing Memphis as AI Capital

Transcript

SpaceX Signs $6.3B Compute Deal with Reflection AI, Cementing Memphis as AI Capital · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the rocket company sells server time. SpaceX signed a six-point-three billion dollar compute deal with Reflection AI, an open-source startup backed by Nvidia.
• 0:11
  [Guest] Reflection gets GB three hundred chips inside Colossus 2, SpaceX's Memphis data center, at a hundred and fifty million dollars a month starting July first. Add Anthropic and Google, who already rent SpaceX compute, and the rocket company pulls in over two-point-three billion a month just in GPU access fees.
• 0:30
  [Host] How did a space company become the world's biggest AI compute landlord?
• 0:35
  [Guest] Speed. Musk built the original Colossus from an empty Electrolux factory to running in a hundred and twenty-two days when normal builds take years. The Memphis campus now runs over four hundred thousand GPUs.
• 0:48
  [Host] That is a lot of hardware. How much electricity?
• 0:51
  [Guest] Three hundred megawatts today, heading for three gigawatts. Picture three nuclear reactors feeding one campus in Tennessee.
• 0:58
  [Host] And the government noticed.
• 0:59
  [Guest] FERC issued show-cause orders on June eighteenth to every regional grid operator, demanding they justify how large loads connect. Commissioner David LaCerte told them, quote, I say this not as a threat but as a statement of duty. The AI buildout has turned electricity into a federal policy question.
• 1:18
  [Host] But here is the complication. Every one of these compute deals carries a ninety-day exit clause. Either side can walk away. That does not sound like the infrastructure monopoly the numbers suggest.
• 1:30
  [Guest] The clause is legal. The physics is not. Training a frontier model means tens of thousands of GPUs running for weeks without interruption. Walking away mid-run kills the run. You are free to leave, but no comparable facility is ready to take you on day one.
• 1:48
  [Host] Reflection is an open-source lab. They give the trained weights away for free. Why does open-source AI require the most expensive compute in the industry?
• 1:58
  [Guest] Training costs the same physics whether you lock the result or give it away. Reflection's CEO Misha Laskin has put it this way: you cannot do scientific discovery on a closed model. It would be like doing rocket science but you only get to look at the rocket instead of taking apart the engine. The openness is the point, but the training run that produces it is indistinguishable from a proprietary one.
• 2:25
  [Host] So open-source AI runs on closed infrastructure. The model is free but the factory is not.
• 2:31
  [Guest] And the concentration tightens with every contract. SpaceX, Amazon, Microsoft, Google. These are not tech companies anymore. They are infrastructure gatekeepers that happen to sell intelligence.
• 2:41
  [Host] Patrick Deglon calls this the Star Wars path in Unscarcity, where a few gatekeepers own the means of intelligence production and everyone else negotiates for access.
• 2:51
  [Guest] Reflection will give those weights away for free, and the check they write to do it is larger than most AI companies ever raise. That is the paradox Memphis is building, one GPU rack at a time. For more on the infrastructure layer beneath AI, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.