Before an AI data center can work in space, its chips have to survive the trip and keep operating in orbit. Google is about to test that first question.
In a September 24 Project Suncatcher update, Google said it expects to put its first Tensor Processing Units in orbit next week. The prototype satellite, developed with Planet, is scheduled for SpaceX’s Transporter-18 rideshare mission. Google plans to measure how Trillium TPU hardware handles launch vibration and acceleration, radiation exposure and the thermal extremes of orbit.
The launch has not happened as of this draft. Its date is a schedule, not a confirmed flight result. Calling the craft Google’s “first AI satellite” would also blur the story: Google has worked on other AI-related satellite initiatives. This mission is specifically an early test of TPU hardware for the company’s proposed space-based computing system.
The engineering questions behind the headline
Google says its team shook the hardware on three axes to simulate rocket launch conditions. At a UC Davis proton-beam facility, it ran AI workloads while testing radiation effects and reported that Trillium TPUs withstood a total ionizing dose greater than its estimated five-year mission exposure. Those are prelaunch test results from Google, not proof that a complete orbital computing service is viable.
Cooling is another open problem. There is no air to move heat away in a vacuum, so Google is testing heat pipes and radiators. The planned flight should provide evidence about how that cooling arrangement behaves in orbit. It will not resolve the economics of launching, powering, cooling and maintaining a large fleet of computing satellites.
Power and cooling are difficult even for ground-based facilities, which is why Google joined an AI energy-management alliance for data centers. Moving the chips to orbit changes the power source and cooling environment; it does not make those engineering constraints disappear.
What would count as the next milestone?
Google’s longer-term idea is to link satellites carrying many TPUs with high-bandwidth laser connections. Its new update places a two-satellite interconnection test in 2027. The original 2025 Project Suncatcher announcement also described a learning mission with Planet, but the September update makes clear that next week’s single prototype is the immediate test.
The useful way to evaluate the project is as a sequence: chip survival first, in-orbit cooling and operation next, then precise high-bandwidth links between craft. Only after those steps would a scalable orbital AI system become a practical proposition. A successful launch would be important evidence for step one, not the arrival of a working AI data center in space.
There is an easy category error here. WeatherNext 3 uses satellite observations to make forecasts on Earth; Project Suncatcher asks whether AI hardware itself can run in orbit. The satellite connection sounds similar, but the computing challenge is different.
Sources: Google’s September 2026 Project Suncatcher update; Google’s original Project Suncatcher announcement.
