Particle.news

Google Launches Prototype AI Satellite to Test Data‑Center Chips in Orbit

The mission will gather failure‑mode and thermal data to test whether orbital AI can scale as launch costs fall or cooling technology improves.

Overview

  • Google’s refrigerator‑sized prototype, called MVP, launched Oct. 1, 2026 and carries four Trillium Tensor Processing Units plus solar arrays that produce about 1 kilowatt of power.
  • The company confirmed contact with the spacecraft and says MVP will run short inference sessions of roughly 15 minutes then power down so onboard heat pipes and radiators can shed heat in vacuum.
  • Extensive ground tests at UC Davis showed the TPUs tolerate years of radiation but a retest with reduced shielding produced slightly higher logic‑error rates, so in‑orbit data will refine error‑rate and failure‑mode estimates.
  • Google plans two purpose‑built satellites in 2027 to test high‑bandwidth laser links between craft, and a peer‑reviewed Joule paper from the team estimates full data‑rack satellites would need 50–100 kW and depend on large cuts in launch cost.
  • The experiment addresses a human and systems problem: rising local opposition and power limits for Earth data centers push firms to seek abundant solar power in LEO while hard constraints—heat rejection, radiation‑induced bit flips, limited bandwidth and hardware obsolescence—will decide whether orbital compute can move beyond research prototypes.