SpaceX launched Google’s first Project Suncatcher test satellite October 1 at 2:32 p.m. EDT from Vandenberg Space Force Base, a Falcon 9 rideshare carrying 130 payloads including a refrigerator-sized prototype built with Planet that carries four of Google’s tensor processing units. It is the first in-orbit test for the research effort Google calls a long-term moonshot toward machine learning infrastructure in space.
The mission, Transporter-18, is a rideshare flight packed with dozens of small satellites from multiple customers. The Google payload, internally named MVP for minimum viable product, is built in partnership with imaging company Planet. Google says its compute roughly matches a single terrestrial data-center server and its solar array produces about one kilowatt.
What the satellite actually tests
The prototype is designed to gather in-orbit data on how TPUs handle the physical stress of spaceflight and the radiation and thermal extremes once they arrive. The team at Google describes three hazards it has been working through for this test. The first is launch itself: the ride to low Earth orbit lasts about 10 minutes, with sustained loads reaching 10 g on the vehicle. Individual components such as the TPU chips can experience 50 to 100 g. Vibration and shake testing on the ground was designed to model this phase, and the hardware survived.
The second is radiation. Solar events and cosmic rays can flip bits and degrade electronics in ways that are hard to predict from simulation alone. Google’s team ran Trillium TPUs inside a proton beam at the University of California, Davis Crocker Nuclear Laboratory while running live AI workloads, watching for errors such as a bitflip and their effect on the compute jobs. Travis Beals, Google’s senior director for paradigms of intelligence, reported initial results: the chips hold up, in Google’s words, remarkably well and can survive a total ionizing dose greater than what they would receive across a five-year mission. That figure is a ground-test result, not a space measurement, which is why the satellite is flying. Real orbits bring thermal cycling and particle environments no beam test can fully duplicate.
The third is heat. There is no airflow in a vacuum, so engineers developed a cooling method to move the heat TPUs generate when they run inference and training workloads. A thermal-vacuum chamber on the ground tested heat pipes and radiators. The orbital test confirms whether the design works when the satellite is actually in space and when the Sun moves in and out of view around the orbit, cycling the vehicle between hot and cold every 90 minutes or so.
Why Google thinks space makes sense eventually
The bet is simple physics. Solar power in space arrives without clouds or night, roughly eight times more energy collected over a year than the same panels on Earth, and waste heat can in principle be radiated away from a large constellation more easily than it can be cooled inside a land-constrained building with a chiller plant. Project Suncatcher was announced in November 2025 as a research effort to explore whether scalable AI compute infrastructure could live in orbit, not as a working product. This prototype, in other words, is the first time the company’s claims about TPU durability, thermal design, and orbital station keeping leave the lab.
The next milestone is already scheduled. Two more experimental satellites launch in 2027 to test how a potential constellation communicates, using laser links between spacecraft and back down to Earth. The prototype flying now is deliberately limited in what it can prove. It answers survivability questions, whether the hardware survives launch and whether the cooling architecture works. It does not settle the economic question, whether the launch mass and radiation shielding and maintenance overhead of a cluster in orbit beats a terrestrial data center on cost per useful compute.
| Test | Method | Result so far |
|---|---|---|
| Launch vibration | Ground shake table, launch loads modeled at 10 g vehicle, 50-100 g components | Hardware survived, orbital data pending |
| Radiation | Proton beam at UC Davis Crocker Nuclear Laboratory | Chips tolerated more than a five-year mission dose, initial result |
| Thermal | Thermal-vacuum chamber, heat pipes and radiators | Design held in vacuum, orbital test ongoing |
| Laser links | Two-satellite pair | Scheduled 2027 |
Scale, orbit choice and the real bottleneck
The constellation Google sketches in its research posts is not a handful of satellites. A useful cluster needs dozens of spacecraft co-flying in close formation, pointing laser inter-satellite links at one another while holding tight relative orbits, which is a harder control problem than what typical imaging or communications satellites do today. Low Earth orbit is the likely home because it keeps latency low and radiation manageable, but that also means every satellite passes through eclipse roughly half of each orbit, so the solar budget swings hard across every 90-minute lap. Batteries and duty cycling become part of the compute math, not afterthoughts.
Launch cadence and mass are the real bottleneck. A single terrestrial AI data center campus can hold tens of thousands of accelerators and draw hundreds of megawatts. Matching that compute with current ride-share prices would require thousands of launches carrying hardware designed for vacuum, radiation tolerance, and years of unattended operation with no repair crew. That cost gap is why more than one Google speaker calls this a moonshot rather than a roadmap. Getting the cost curve down is the kind of thing that takes a decade or a new launch architecture, and Transporter-18 is only the first hardware data point on that path.
Context: a race, not just research
Other companies are not standing still. Starcloud, a startup backed by Nvidia, plans a data center in orbit, as does Axiom Space on a smaller scale. Europe is looking at orbital compute through ESA-funded research programs. If Google’s prototype shows the chips and cooling both work, the field shifts from hardware feasibility to cost comparison, and launch economics become the constraint. Transporter-18 itself shows how that bottleneck has loosened: rideshare pricing packs dozens of satellites onto a single Falcon 9, which made development cycles like Suncatcher possible on a research budget rather than a flagship program.
The public data from this flight should start arriving over the coming weeks, once the satellite powers on and the TPUs run their first in-orbit workloads. The September 24 Google Research post that laid out Suncatcher describes the mission as designed to gather in-orbit data on physical stress, radiation and thermal extremes. The company has not committed to a date for any production constellation, and its own language stays at the level of research potential. The satellite is up, the data is coming, and the company that runs search from Earth is now testing whether the next chapter of computing lives off-planet.