TECHNOLOGY · AI · CYBERSECURITY · THE SYSTEMS SHAPING BUSINESS
Google's first Project Suncatcher satellite, a prototype carrying the company's tensor processing units, launched Thursday on a SpaceX rocket from California, TechCrunch reported. Alongside the launch, Google laid out in a peer-reviewed paper how cheap and frequent rocket flights would need to become before data centers in orbit make economic sense.
What Happened
The satellite was built by Planet Labs and flew on SpaceX's Transporter-18 rideshare mission, which The Primary Wire reported last week was scheduled for Oct. 1. It carries Google's Trillium TPUs, according to a Google blog post, and its job is to show that the chips can work in orbit. The spacecraft supplies about 1 kilowatt of continuous power and must keep the chips cool while they run inference workloads. To avoid overloading the power and thermal systems, the TPUs will operate in bursts of about 15 minutes, TechCrunch reported.
Radiation is one of the main unknowns. Google told TechCrunch it redid its ground radiation tests after finding that its test setup shielded the chips more than they would be in space. The revised results showed somewhat higher error rates that the company still considers acceptable. For typical inference, the error rate is very low, roughly one in a million, said Google's Travis Beals, who added that the chips remain a problem for large training jobs that spread work across thousands of processors for long periods.
RelatedBefore launch, Google said, the hardware passed vibration tests simulating rocket ascent and thermal vacuum tests of the heat pipes and radiators that carry heat away, and the Trillium chips tolerated a total radiation dose greater than they would see over a five-year mission.
Why It Matters
The paper, which TechCrunch said is being published in the journal Joule, turns Google's interest in orbital compute into explicit numbers. Google estimates that SpaceX's Starship would need to fly about 1,800 times over a decade, or roughly 180 launches a year, delivering about 370,000 metric tons to orbit. To hit its cost targets, launch prices would have to fall to about $200 per kilogram by 2035. The analysis assumes SpaceX keeps cutting prices by about 20% a year, the learning curve it has followed so far.
The concept's economics therefore rest on the progress and pricing of a rocket Google does not control. TechCrunch initially reported the figure as 1,600 launches and later corrected it to 1,800.
The Long-Term Design
Google's concept is clusters of 81 satellites flying in close formation and linked by short-range, high-bandwidth lasers so that they behave like a single data center. The draw is power: satellites in low Earth orbit can generate up to eight times more solar power than panels on the ground, according to Google. Space-based computing is being promoted as a way around the electricity, land and permitting limits that are slowing AI data center construction on Earth.
The single prototype tests none of the formation flying or laser links. It is meant to answer narrower questions about whether a commercial AI chip can run reliably on a small satellite's power and cooling budget. Running in 15-minute bursts also means the test will not show how the chips behave under the continuous, round-the-clock load a working data center would place on them.
What to Watch
Early data from the prototype on power draw, heat rejection and radiation-induced errors will show whether ground testing predicted orbital behavior. Google's next step is a 2027 mission with two purpose-built satellites that will test laser communication between spacecraft, and later satellites are expected to carry dozens of TPUs each. Starship's own launch cadence and pricing will be the clearest outside indicator of whether the paper's cost path is realistic.
