Google thinks SpaceX’s Starship has to launch 1,600 times before space data centers get off the ground
TechCrunch Tim Fernholz
Google launched a test chip into space on a SpaceX rocket. It’s a step toward orbital data centers, if the power, cooling and launch math ever works.
Based on reporting by TechCrunch, Tim Fernholz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google has put its first advanced chip into orbit. A prototype satellite carrying a Tensor Processing Unit launched today from California on a SpaceX rocket, and the point of the flight is brutally simple: see whether Google’s chip can actually run in space without falling apart.
The satellite was built by Planet Labs and is based on one of the company’s standard platforms. Once it’s commissioned, it will run the TPU in 15-minute bursts so it doesn’t overwhelm the satellite’s power and thermal systems. Google says it has tested the hardware on the ground already, but this is the first real try under space conditions, where there’s no lab trick to hide behind.
This is only the opening move in Project Suncatcher, Google’s long-term plan for compute clusters in orbit. The company says future demos are already in the works for next year, and those satellites are supposed to be more purpose-built for advanced compute, with a laser link between them. Longer term, Google is imagining a network of 81 satellites flying close together and processing in parallel.
The paper that Google published on Thursday, and plans to publish in Joule, is just as interesting as the satellite itself. It argues that SpaceX’s launch costs could keep falling, and that prices near $200 per kilogram by 2035 are plausible if the company keeps up a learning curve of about 20% a year. But the math is stubborn: to move enough payload to orbit for this plan, Google estimates Starship would need to haul 370,000 tons, which works out to about 1,800 launches over ten years, or 180 a year, assuming 200 metric tons per flight.
That’s a lot to ask from a rocket that has never flown more than five times in a year. Google is betting that the radiation problem is manageable too. After redoing chip tests in a particle accelerator, the company found a bit more error in the logic circuitry than expected, but still says the TPUs should handle large inference workloads for a satellite’s five-year life. Training giant models, though, is another story. The errors start to matter there fast.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of moonshot: a useful one with nasty arithmetic attached. Space data centers sound elegant until the bill arrives in launches, power, and cooling, which is usually how hype meets the atmosphere. The real tell here is that Google is thinking about inference, not glorious sci-fi training runs; that’s what serious people do when the demo has to survive contact with physics.
Read more about this at: TechCrunch