Google will launch its first satellite carrying AI chips into orbit on October 1, sending four Tensor Processing Units to space aboard SpaceX’s Transporter-18 rideshare mission from Vandenberg Space Force Base.
The satellite, called MVP and built with the Earth-imaging company Planet Labs, is the first flight hardware for Project Suncatcher, Google’s long-term research effort to find out whether large-scale AI computing can run in space. The fridge-sized craft carries four Trillium-generation TPUs, roughly the compute of one data-center server, and draws about one kilowatt from its solar panels.
Google announced the launch on September 24. CEO Sundar Pichai framed the mission around a single question: whether the company’s TPUs can “survive and operate in space.”
Three ways the hardware can fail
The first hazard arrives in the first ten minutes. A rocket launch subjects a spacecraft to sustained acceleration of up to 10 g, and individual components can face 50 to 100 g. Google shook the satellite on all three axes to reproduce the vibration frequencies of a launch, and says the hardware held.
Radiation is the second hazard. Solar events and cosmic rays can flip bits in memory, corrupting calculations. Google ran AI workloads on TPUs inside a proton beam facility at UC Davis’s Crocker Nuclear Laboratory and tracked how errors affected the work. The chips survived a total ionizing dose greater than they would absorb over five years in orbit, and restarting the processors generally cleared the errors. No hard failures appeared up to the maximum tested level of 15 kilorad silicon.
Heat is the third problem, and the hardest. TPUs pack a lot of power into a small area, and in a vacuum there is no air to carry heat away. Google routes heat through thermal-interface materials, aluminum and copper layers, and out through a radiator panel. The company tested the setup in a thermal vacuum chamber. Travis Beals, Google’s senior director of product management for the project, told The New York Times that the chips can run for roughly 15 minutes before the system must shut down and cool.
What the satellite will actually do
MVP will run AI workloads in short bursts, including limited Gemini queries, then pause to cool. Google plans to operate it for about a year, though the satellite could stay in orbit for up to six years before atmospheric reentry.
The company is explicit that this is not an orbital data center. “This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions,” Beals wrote in a blog post. The mission exists to gather engineering data that no ground test can fully replicate: the combined effects of launch vibration, radiation, vacuum, sunlight, shadow and repeated thermal cycling.
| Test | Ground result |
|---|---|
| Vibration | Hardware held through three-axis launch-profile shaking |
| Radiation | Trillium TPUs survived a dose above a five-year mission level, no hard failures to 15 kilorad |
| Cooling | Heat pipes and radiators validated in thermal vacuum chamber; 15-minute runtime windows in orbit |
| Laser links | 1.6 Tbps bidirectional in bench tests; two-satellite orbital test planned for 2027 |
Why put AI in space at all
The draw is sunlight. In low Earth orbit, satellites can get near-constant sun, and Google estimates orbital solar arrays could generate up to eight times more energy annually than comparable panels on Earth. The company announced Suncatcher in November 2025 as a research moonshot, comparing it to its early work on autonomous driving and quantum computing.
The longer-term concepts are ambitious. Google has modeled formations of more than 80 satellites working together on AI requests, flying in tight clusters at about 650 kilometers altitude, and has considered a custom spacecraft potentially as large as a soccer field. Satellites in such a cluster would talk to each other by laser, and Google needs very high bandwidth over very short distances, a reversal of most space laser systems, which are built for low bandwidth over long range. Bench tests have already shown 1.6 terabits per second bidirectional transmission with a single optical transceiver pair.
Two follow-up satellites are planned for 2027 to test those laser links in orbit. Google says future satellites would each carry dozens of TPUs.
Years from anything useful
Google itself is managing expectations. “We don’t expect, to be perfectly frank, that we’ll have anything usefully operational in the next few years,” James Manyika, Google’s senior vice president for research, told The New York Times.
The gap between this test and the vision is wide. A working chip in orbit answers a narrow engineering question, not an economic one. It does not establish that space-based AI can compete with terrestrial data centers on cost, reliability or scale, and the engineering data MVP returns will matter more than the launch itself. Google frames the project as a decade-long research effort, not a near-term commercial service.
Still, the fact that the first hardware is flying within a year of the project’s announcement says something about the pressure behind it. Data-center power demand from AI is climbing fast, grid connections are scarce, and every large cloud operator is looking for leverage. Space is the most expensive leverage available, which is precisely why the first question Google is asking is the simplest one: do the chips survive.
The competitive context
Google is not the only company looking up. Amazon has studied orbital data centers, startups like Axiom and Starcloud have announced space computing plans, and China has launched experimental computing satellites. What distinguishes Suncatcher is that Google is testing its own production silicon, the same Trillium TPUs that serve its cloud customers, rather than specialized space-rated hardware.
That choice cuts both ways. If standard TPUs survive in orbit, the path to scale is far cheaper, because the chips are already manufactured by the millions. If they degrade in ways ground tests did not predict, Google will need a radiation-hardened variant, and the cost advantage of commodity silicon disappears.
The timing also reflects a real constraint on the ground. Google’s data-center expansion has run into grid interconnection queues, local opposition and rising electricity prices. In its most recent earnings cycle, capital expenditure guidance climbed again, driven by AI infrastructure. Orbital solar power is speculative, but so was landing rockets a decade ago, and the company is spending a comparatively small sum to find out whether the physics works before anyone argues about the economics.