Google’s Orbital AI Bet Depends on Chips Surviving Space — and Starship Scaling Fast

Google’s prototype orbital compute satellite launched on a SpaceX rocket from California, marking the first time the company has sent one of its advanced chips into space. The satellite, built by Planet Labs, is designed to test whether a Google Tensor Processing Unit can function in orbit.

The mission will evaluate core operational needs: supplying continuous power, cooling the chip, and running models to identify potential failures. For readers tracking AI infrastructure, the key point is that Google is moving its data center ambitions beyond theory and into flight testing.

The effort is part of Project Suncatcher, Google’s plan to develop large-scale compute clusters in orbit around Earth. Once commissioned, the satellite will run its TPU in 15-minute bursts to avoid overloading the satellite’s power and thermal systems.

Google’s longer-term concept envisions 81 satellites flying in close formation and processing workloads in parallel. A future demo expected next year would use two satellites built more specifically for advanced compute and attempt collaboration through a laser communications link.

Google’s research points to launch cost as a major constraint for orbital data centers. The company’s paper, set to be published in Joule, discusses a path in which SpaceX could reach launch prices near $200 per kilogram by 2035, while noting the analysis is not an economic feasibility study.

The article says Google’s researchers estimate Starship would need to fly 370,000 tons of payload into orbit to follow a similar cost-reduction trajectory. That would require about 1,800 launches over 10 years, or 180 per year, assuming 200 metric tons per mission — a sharp increase for a vehicle the article notes has never flown more than five times in a year.

Google’s updated research suggests its chips are likely to survive space radiation. After redoing particle-accelerator tests to better reflect actual shielding conditions, the company found more logic-circuit errors but remained confident the chips can handle large inference workloads during a satellite’s five-year lifespan.

The distinction matters: the reported error rate may be manageable for typical inference, but more demanding mega-scale training runs could be problematic. The near-term takeaway is clear: orbital AI inference looks more plausible than full-scale orbital training based on the information provided.