Quote · All-In with Chamath, Jason, Sacks & Friedberg
Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter
Where this was said
The math behind distributed compute and datacenters in space
At 1:11:50 · chapter starts 1:10:17
Tesla's Megapod trademark filing suggests modular AI data centers at Supercharger stations. Travis describes putting compute in Adams restaurant kitchens. Debate over distributed training vs. inference efficiency constraints [1] — Travis Kalanick "Distributed training across physically separate locations is orders of magnitude less efficient — even 2 kilometers of fiber kills performa…" 1:12:14 .
Standing up a 1 gigawatt terrestrial data center costs approximately $35 billion in NVIDIA semiconductors plus $25 billion in power and cooling equipment.
A terrestrial 1-gigawatt data center costs $60 billion ($35B chips + $25B power/cooling). A reusable Starship puts the same compute in space for $40 billion. As terrestrial costs inflate and launch costs deflate, space wins. Running costs in orbit: about $1 billion per year for power.
Distributed training across physically separate locations is orders of magnitude less efficient — even 2 kilometers of fiber kills performance. But everything that kills distributed training is actually fine for inference. Distributed inference clouds are coming, and they could recycle unused compute from restaurants to homes.
With a reusable Starship, it would cost approximately $5 billion in launch costs to put a gigawatt of compute into orbit, making the total around $40 billion versus $60 billion terrestrially.
Chamath estimated that roughly 40% of all new data centers have faced regulatory or community contestation since 2021, a number he expects to rise.
