Google Says $200 Space Launch Costs Require 1,800 Starship Flights
Google’s first TPU reached orbit, but its 81-satellite data center vision may need 1,800 Starship launches to push costs near $200 per kilogram by 2035.
Summary
On October 1, 2026, Google sent its first advanced chip into space aboard a SpaceX rocket launched from California. Planet Labs built the standard-platform prototype, one of more than 100 payloads alongside missions from Satlyt and Cowboy Space Company. It will test whether Google’s Tensor Processing Unit, a rival to Nvidia GPUs, can operate with one kilowatt of continuous power, cooling and model workloads. After commissioning, the TPU will run in 15-minute bursts to limit power and thermal strain. A more purpose-built 2027 demonstration will connect two satellites by laser for larger, collaborative workloads. Project Suncatcher ultimately envisions 81 satellites flying in close formation and processing in parallel, with bandwidth and latency designed for multi-rack AI workloads expected within five years.
Google’s peer-reviewed orbital data center paper, released October 1 for publication in Joule, is not an economic feasibility study. Using Falcon 9 payload volumes and SpaceX’s roughly 20% annual cost learning curve since Falcon 1, it projects launch prices near $200 per kilogram by 2035. Achieving that trajectory would require Starship to place 370,000 tons into orbit through about 1,800 launches over 10 years, averaging 180 annually and 200 metric tons per mission. Starship has never flown more than five times yearly, although Elon Musk suggested hourly flights in 2029. Google, a major SpaceX investor, also repeated radiation tests after finding its original chip configuration provided excessive shielding. Errors increased slightly, but Google expects roughly one error per million typical inference operations and five years of orbital service. Training across thousands of chips for months remains problematic.
Positives
- Google’s first advanced chip reached orbit, moving Project Suncatcher from ground testing to an operational space experiment.
- Radiation retests indicate TPUs can support typical inference with roughly one error per million operations across a five-year satellite lifespan.
- A 2027 demonstration will connect two purpose-built compute satellites through laser communications and attempt larger collaborative workloads.
- The 81-satellite design targets close-formation parallel processing for future multi-rack AI workloads.
- SpaceX’s roughly 20% annual launch-cost learning curve supports Google’s projection of prices near $200 per kilogram by 2035.
Risks & concerns
- Reaching Google’s cost projection requires about 1,800 Starship launches in 10 years, averaging 180 annually at 200 metric tons per mission.
- Starship has never completed more than five flights in one year, far below the cadence Google’s trajectory requires.
- The prototype TPU must operate in 15-minute bursts because continuous use could overwhelm the satellite’s power and thermal systems.
- Radiation-induced logic errors remain problematic for training runs involving thousands of chips operating for months.
- Google’s Joule paper explicitly stops short of establishing the economic feasibility of orbital data centers.