Google Launches In-Orbit TPU Test Satellite with SpaceX
On October 1, Google partnered with Planet to launch a prototype satellite equipped with four TPUs into orbit. The rocket was SpaceX's Transporter-18 rideshare mission, taking off from Vandenberg Space Force Base in California. Google confirmed that it has established contact with the satellite, and operations are proceeding as expected.
This is the first step of Project Suncatcher. The project will be publicly unveiled in November 2025, aiming to verify whether low Earth orbit can support scalable machine learning infrastructure. Data will be collected in the coming weeks to see how the TPUs withstand launch stress, radiation, and extreme temperatures. Officials stated that there are no fully equivalent tests on the ground. The results will be used to modify subsequent designs. A peer-reviewed paper has been published in Joule.
The satellite is a modified version of an existing platform from Planet Labs, roughly the size of a refrigerator, housing Google's four TPUs. It is powered by solar energy, generating about 1 kilowatt. Once operational, the TPUs will run for short bursts of 15 minutes each to avoid overwhelming power and thermal management systems. The original plan was to launch two dedicated satellites in 2027, but this prototype was integrated into an existing satellite for early testing. The demonstration satellites planned for next year will be designed for heavier computational loads.
Transporter-18 carried a total of 130 payloads. Google describes low Earth orbit as having nearly continuous sunlight, with solar output potentially reaching up to eight times that of the ground. The long-term vision is for 81 satellites to form a close-knit constellation, processing in parallel and interconnected via high-speed lasers. Travis Beals, the project lead, stated that the focus for multi-platform payloads is on bandwidth and latency between TPUs, looking at missions five years out, rather than today's tasks.
This marks Google's first deployment of its self-developed accelerator chips into space. Ground tests were conducted at the University of California, Davis. Alphabet's stake in SpaceX is estimated to be over $82 billion. This rideshare mission only demonstrates that the chips can power on, not that an orbital data center has been established.
The buyers are model companies looking to avoid ground power constraints, while the sellers are Planet, providing the satellite platform, and SpaceX, providing the launch. This is an experimental-driven R&D expenditure, not a computational order. Beneficiaries include Planet, which received the modification task, and SpaceX, which sold the rideshare slots. The pressure is on data centers still priced based on ground facilities, contingent on proving that the 15-minute short runs can be extended.
The four TPUs, 1 kilowatt, and 15 minutes define the boundaries of this prototype. The 81 satellites in formation and laser interconnects are still in the papers and designs.
Source: Public Information
ABAB AI Insight
Google's TPU development began in 2016 to reduce reliance on Nvidia. Power and cooling constraints in data centers are expected to become bottlenecks after 2024. Project Suncatcher aims to move computational power to low Earth orbit by November 2025, citing longer sunlight hours and different thermal environments. Instead of waiting for dedicated satellites in 2027, four chips were integrated into an existing Planet platform. The path is to first prove that they can power on in orbit before discussing formations.
Funds are spent on modifying satellites and rideshare slots, not on building new orbital data centers. One kilowatt is roughly enough for a microwave, and the 15-minute short runs are designed to avoid overwhelming power and thermal control. The two demonstration satellites planned for next year will be designed for heavier loads. The 81 satellites in close formation will require laser interconnects and dedicated launches; the 130 payloads on Transporter-18 are merely a side note. Alphabet's stake in SpaceX is estimated at over $82 billion, indicating that the relationship between the launch provider and chip provider is not purely customer-based.
A comparison can be made to the 1960s communication satellites that first launched test satellites before discussing global networking, as well as Starlink's use of rideshares and bulk launches post-2020 to establish its constellation as infrastructure. The difference this time is that the payloads are training and inference chips, not transponders. The industry is still in the experimental phase. Ground data centers remain the primary source of capacity, with only a refrigerator-sized prototype on the orbital side. Planet sells the platform and on-orbit experience, while Google buys radiation and thermal data.
Structurally, this is a prelude to technological substitution, not the substitution itself. The constraints of ground data centers are power and cooling, while low Earth orbit exchanges sunlight for power and vacuum environments for thermal design. The mechanism is that chips must first survive radiation and thermal cycles; being able to run for 15 minutes qualifies them for discussions on sustained loads. Pricing power still resides with ground data centers. To take a portion of the orbital market, the 1 kilowatt and four TPUs must first be transformed into deliverable tokens.
ABAB News · Cognitive Laws
First prove they can power on, then discuss building factories.
Rideshare slots purchase data, not capacity.
Even with eight times more sunlight, they must first survive fifteen minutes.