Google’s AI Satellite Is Up, but Orbital Data Centers Still Need a Rocket Miracle
Google’s AI Satellite Is Up, but Orbital Data Centers Still Need a Rocket Miracle
Google’s Project Suncatcher has moved from an ambitious white paper to a live experiment in orbit — but the satellite’s launch is only the opening test of a far more demanding proposition: putting data-center-scale AI compute above Earth.
Before Thursday’s mission, Alphabet cast the effort as a long-term attempt to determine whether space could host scalable machine-learning infrastructure. Its argument is straightforward: satellites in low Earth orbit can access near-constant sunlight, potentially generating far more solar power than terrestrial systems. But the company also acknowledged that its TPUs had yet to face the thermal, radiation and power constraints of orbit.
That test began when a Planet Labs-built prototype carrying four Google TPUs launched aboard SpaceX’s Falcon 9 Transporter-18 rideshare mission. Google DeepMind chief Demis Hassabis celebrated the flight as “the first step of Project Suncatcher,” describing it as a long-term research moonshot rather than a finished computing platform.
Sundar Pichai struck a similarly upbeat but measured note after the launch, saying the prototype showed “how far we’ve come…and how far there is to go.” The spacecraft will run its TPU in 15-minute bursts, a deliberately cautious approach intended to avoid overwhelming its power and thermal-management systems.
Google’s next target is more ambitious: purpose-built satellites that can handle heavier workloads and communicate by laser, eventually forming an 81-satellite cluster for parallel processing. Yet Project Suncatcher’s own analysis makes the bottleneck plain. To bring launch costs toward $200 per kilogram by 2035, Google estimates Starship would need to deliver 370,000 tons to orbit — roughly 1,800 launches in a decade, or 180 annually.
There are other obstacles: cooling, radiation resilience, debris and the economics of replacing hardware in space. Google says its chips appear capable of surviving typical inference workloads, with an error rate of about one in a million. But its project lead, Travis Beals, said that was already problematic for mega-scale training runs involving thousands of chips over months.
The launch proves a TPU can now be tested where it matters. It does not yet prove that an orbital data center can be built — much less that it can beat one on the ground.
Write a comment