Google’s Project Suncatcher: 5 things to know about its plan to put AI chips in space
Tech Funding News Abhinaya Prabhu
Google launched a test satellite with four AI chips on October 1. It’s a first crack at running real computing in orbit, where heat and launch costs could wreck the idea.
Based on reporting by Tech Funding News, Abhinaya Prabhu — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google has finally put some of its AI hardware into space. On October 1, a prototype satellite for Project Suncatcher rode SpaceX’s Transporter-18 mission from Vandenberg Space Force Base in California, carrying four Tensor Processing Units and about one server’s worth of computing power.
The company built the satellite with Planet, the satellite operator, and Travis Beals, a senior director at Google, said the team has already confirmed contact and that the craft is working as expected. For now, that is the story: a basic test of whether the chips survive launch, radiation, and the kind of temperature swings that make space unfriendly to electronics. Google has also published a peer-reviewed paper in Joule on the research.
The early signs on radiation are encouraging, at least on the ground. Google says its Trillium chips handled more proton-beam radiation at the University of California, Davis, than a five-year mission would need to survive. Brendan Burke of Futurum reads the paper as showing about a 20-times margin for the main processing die, while warning that high-bandwidth memory is the weak point. It behaved oddly at a lower dose, though still above the mission requirement. That still needs to be proven in orbit, and Google knows it.
Heat may be the uglier problem. A satellite has no air or water to pull heat away, so the chips have to dump it through radiators. Google is testing heat pipes and radiators, and the current setup pulls about one kilowatt from its solar panels, runs short queries in roughly 15-minute bursts, then shuts down to cool. The design target is a year in orbit, but James Manyika, Google’s senior vice president for research, told the New York Times he does not expect anything “usefully operational” in the next few years.
Meanwhile, the field is already getting expensive. Starcloud, formerly Lumen Orbit, says it has flown an NVIDIA H100 in space, trained an AI model, and run a Google language model in orbit. Cowboy Space, formerly Aetherflux, is chasing a different prize entirely: beaming solar power to Earth. The bigger question is not whether people can get chips into orbit. It’s whether they can make the economics even vaguely sane, and Google’s own paper says that means launch prices near $200 per kilogram. That is a long way from where things are now.
My take — AI-written commentary, not fact-checked reporting
Space AI already looks like a funding race dressed up as infrastructure. The hardware demo is cute; the real product is the launch bill, and right now that bill still has the last laugh. The usual Silicon Valley trick is to call it inevitable before it is affordable.
Read more about this at: Tech Funding News