Google says the first prototype satellite it launched for Project Suncatcher has reached orbit, established contact with ground stations, and is operating as expected. It's the first time the company has actually powered up its own TPU AI chips and run models in space.
The satellite rode to orbit on October 1 aboard a SpaceX Falcon 9 on the Transporter-18 rideshare mission, launching from Vandenberg Space Force Base in California. It was built by the satellite company Planet under the internal codename MVP (minimum viable product), and flies at an average altitude of roughly 650 km.
What's on board
The satellite carries four Trillium TPUs — Google's sixth-generation TPU (v6e) — modified for spaceflight but not purpose-built, radiation-hardened silicon. Solar panels supply about 1 kilowatt of power, and the chips are tasked with running short Gemini inference requests.
There's no air in space, so fans are useless; heat can only travel through heat pipes to infrared radiators and be shed by radiation. That limits each chip to roughly 15-minute runs before it has to shut down and wait to cool. Citing Google's own design documents, The Register reports that a single TPU needs about 1.3 square meters of radiator area — an engineering problem Google itself admits it hasn't solved yet.
This satellite is meant to answer three questions: can the chips survive the vibration and acceleration of launch, will orbital radiation degrade them, and can they run reliably through the extreme temperature swings of low Earth orbit? No lab on the ground can simulate all three at once.
What ground testing already showed
Before launch, Google bombarded a Trillium TPU with a proton beam at UC Davis's Crocker Nuclear Laboratory, at a dose equivalent to a shielded five-year space mission. The chip showed no hard failure from total ionizing dose; memory errors began appearing at roughly three times that five-year dose estimate.
Google also ran the hardware through vibration tables to simulate launch and thermal vacuum chambers to simulate orbit. The in-orbit data will now be checked against those ground results. Under Google's roadmap, a 2027 mission will test high-bandwidth laser links between satellites — the kind of link needed to tie dozens or hundreds of satellites into a single orbital compute cluster.
The launch-cost math still doesn't work
When Google announced the project last November, it set a condition: launch costs need to fall below roughly $200 per kilogram before a space data center's energy cost per unit of compute can match a ground-based one, something it expects could happen by the mid-2030s.
The Register puts today's launch price at around $7,000 per kilogram — about 35 times higher. Radiators alone illustrate the problem: at 1.3 square meters per TPU, a satellite carrying the dozens of chips Google eventually envisions would need tens of square meters of radiator surface, on top of solar panels and laser communication terminals, making it hard to keep a single satellite's mass down. Whether launch prices fall into that range depends largely on progress with fully reusable rockets like Starship — something outside Google's control.
Rivals are moving too. Starcloud put an Nvidia H100 into orbit last year, and SpaceX has said it wants to deploy its first orbital compute test systems by the end of 2027. Google's approach is more cautious: fly one small, four-chip satellite, gather a few weeks of data, and then decide what comes next.
Sources: The Register, CocoLoop, DIGITIMES, Forkast, Google Research's Project Suncatcher paper; the paper verifies the radiation test results and the $200/kg cost threshold, while The Register verifies the radiator area and current launch price.