Google Launches First Orbital Compute Satellite, Eyes Space Data Centers Contingent on Starship Scale-Up
Google’s first orbital TPU will run in 15-minute bursts while testing 1 kW continuous power and AI workloads in space. Its 81-satellite vision depends on launch costs falling to $200/kg by 2035, requiring Starship to fly about 1,800 missions.
Google launched its first orbital compute satellite on October 1, 2026, sending one of its Tensor Processing Units into space aboard a SpaceX rocket launched from California. The satellite, built by Planet Labs, marks the opening move in Project Suncatcher — Google’s plan to develop large-scale compute clusters in Earth orbit.
The satellite will operate its TPU in 15-minute bursts to manage power and thermal constraints. Its primary mission is to confirm that a Google TPU can function in space, including sustaining a kilowatt of continuous power and running AI models under real conditions. “We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing,” said Travis Beals, the Google executive managing Project Suncatcher.
Google and Planet Labs are also developing two more purpose-built compute satellites expected to launch next year, which will attempt to communicate via a laser link. The long-term vision calls for a network of 81 satellites flying in close formation and processing AI workloads in parallel.
Alongside the launch, Google published a peer-reviewed white paper in the journal Joule analyzing what it would take to make orbital data centers cost-effective. The paper identifies launch costs as the central bottleneck. Google projects SpaceX could reach roughly $200 per kilogram by 2035 — but doing so would require Starship to fly approximately 1,800 missions over the next decade, or about 180 per year. Starship has never flown more than five times in a single year.
On the hardware side, Google’s updated research suggests its chips can handle the radiation environment of space. Radiation testing produced a logic error rate of roughly one in a million for typical inference workloads — acceptable for a five-year satellite lifespan, though Beals noted it would be “problematic” for large-scale training runs involving thousands of chips running for months.
The gap between current Starship launch cadence and the scale Google’s analysis requires suggests orbital AI infrastructure remains a long-range prospect rather than a near-term deployment.