Overview
- Google said it will send a prototype satellite called MVP into low Earth orbit on Oct. 1 that carries four Tensor Processing Units and will run simple queries using its Gemini model while operating for about one year.
- The mission is explicitly experimental and meant to gather failure‑mode and performance data on TPUs rather than to deliver a live orbital data‑center service.
- Engineers will study how launch forces, space radiation that can cause bit flips, and vacuum‑stage cooling limits affect the chips; on the satellite Google will rely on heat pipes and radiators because there is no air for conventional cooling.
- Google plans two follow‑on satellites in 2027 to test high‑bandwidth laser links for inter‑satellite networking and has conceptual designs for fleets of more than 80 satellites with a target of cost parity with ground data centers by the mid‑2030s, but outside experts say major engineering and cost barriers will keep commercial use years away.
- If orbital compute proves viable it could ease terrestrial power and siting limits for energy‑hungry AI workloads, but that outcome depends on cheaper launches, mass satellite production and reliable long‑term in‑orbit operations that have not yet been demonstrated.