Google Sends TPUs to Orbit to Test Space Data Centers

Google's first orbital compute satellite launched on Thursday aboard a SpaceX Falcon 9, carrying four of the company's Tensor Processing Units into low Earth orbit. It is the opening test of Project Suncatcher, the "moonshot" Google revealed in November 2025 that aims to run machine learning workloads on solar-powered hardware in space.

What happened

The spacecraft, built with Planet Labs, rode the uncrewed Transporter-18 mission out of Vandenberg Space Force Base in California. Travis Beals, senior director of Project Suncatcher, said the team confirmed contact with the satellite and that it was "operating as expected."

The prototype is deliberately modest. The TPUs will run a version of Google's Gemma model for roughly 15 minutes at a time to stay inside thermal limits, an experiment designed to measure how the chips handle launch stress, radiation and extreme temperature swings rather than to produce useful compute at scale.

The bigger number arrived alongside the launch. Google released a peer-reviewed white paper, due in Joule, arguing that orbital data centers only become cost-competitive if launch prices fall to roughly $200 per kilogram by 2035. Getting there would require Starship to lift about 370,000 tons of payload, or some 1,800 launches over the next decade at 200 metric tons per flight. SpaceX has never flown Starship more than five times in a year.

Why it matters

Google's pitch is energy. In low Earth orbit, satellites see near-constant sunlight and can generate up to eight times more solar power than ground-based arrays, according to the company's post. Google's long-term target is a formation of 81 satellites processing in parallel, an architecture where inter-satellite bandwidth and latency between TPUs determine what workloads are even possible.

Radiation testing suggests current TPUs can handle inference across a five-year satellite lifespan, but error rates remain too high for large-scale training. Google's own paper frames orbital compute as preparation for workloads that do not exist yet, with Beals saying the company is looking at where demand will be in five years, not today.

Builder impact

For teams shipping AI systems, orbital compute is not a capacity plan yet, it is a signal about where the compute constraint is heading. Ground data centers remain limited by power, land and community opposition, which is exactly what pushed Google and SpaceX toward orbit. SpaceX has said it will deploy "supercompute in space" in 2027 and is building its own satellite swarms in Redmond. The practical takeaway is defensive: plan for heterogeneous, latency-bound infrastructure and keep agent workloads portable, because the next round of capacity may sit somewhere with very different bandwidth characteristics than a region in Virginia.