Google's Project Suncatcher prototype satellite is in orbit with TPUs aboard
Google Research said Thursday its Project Suncatcher prototype satellite — built with Planet and launched on SpaceX's Transporter-18 rideshare — is in contact and operating as expected, the first flight test in a moonshot asking whether space can host scalable machine-learning infrastructure.
Earth-side power and cooling are the bottleneck everyone can see. Google just put TPUs on a real satellite to measure whether the physics of space will even let ML hardware survive — a hard engineering gate before “space data centers” stop being a slide.
Google Research said on Thursday, 1 October 2026, that its Project Suncatcher prototype satellite is in orbit. The post is titled “Our Project Suncatcher prototype satellite is in orbit.” The byline is Travis Beals, senior director of Paradigms of Intelligence. The page dates the post Oct 01, 2026. It does not print an hour. The satellite was built in partnership with Planet and launched that day aboard the Transporter-18 rideshare with SpaceX. A rideshare is one rocket carrying many satellites, rather than a launch bought for a single craft. Google’s team has confirmed contact with the satellite, and the post says it is operating as expected. Those lines are the Oct 1 post.
What the flight is for. The post calls this the first step in a long-term research moonshot exploring whether space could one day host scalable machine-learning infrastructure. A moonshot, as Google uses the word here, is a long research bet, not a product a customer can switch on. Over the coming weeks, the team plans to gather in-orbit data on how its TPUs handle the physical stress of spaceflight and the radiation and thermal extremes of space. A TPU, a tensor processing unit, is Google’s own chip for the heavy math inside machine-learning models. Machine learning is the method that trains a model on examples. Radiation, here, is the high-energy particles in space that can flip bits or wear a chip out. Thermal extremes are the heat and the cold. Those lines are the Oct 1 post.
What the post does not say. It does not say Google has opened a space data center. It does not say anyone can rent machine-learning compute from orbit. The craft is a prototype on a research flight. The post says some things can only be tested in space, that the experiments are beginning, and that the team will use what it learns to refine the designs and share more as the mission goes on. Those lines are the Oct 1 post.
What Google had already published about the same project, before the launch. A page dated Sep 24, 2026, is titled “Behind Project Suncatcher, our moonshot to put AI in space.” The byline is the same Travis Beals. The page does not print an hour. It says Project Suncatcher was announced last year, and that the question is whether space can host scalable machine-learning infrastructure. In low Earth orbit, a low path around the planet, satellites can sit in near-constant sunlight. The page says a panel there can generate up to eight times more solar power than on Earth. That eight-times line is the Sep 24 page’s figure for the sunlight, not a measurement radioed down from Thursday’s satellite. The page was written while the launch was still ahead. It calls Transporter-18 upcoming, and it says the team would put the first TPUs in orbit “next week.” The Oct 1 post is the later update: the launch happened, contact was made, and the satellite is operating as expected.
What the Sep 24 page says was tested on the ground. A rocket ride to low Earth orbit lasts about 10 minutes, the page says, with vibration and sustained acceleration up to 10 times the force of gravity. Individual parts, including the TPU chips, can see 50 to 100 times gravity. The team shook the satellite on all three axes to copy the shaking of a launch. Google says tests like that rarely go as planned, and that the hardware held up. Once a chip is in space, radiation is the next problem. Solar events and cosmic rays can scramble electronics. The team ran TPUs in a proton beam at the University of California, Davis, Crocker Nuclear Laboratory, while the chips did machine-learning work, and watched how errors such as a bit flip showed up. A bit flip is one stored bit changing when it should not. Google says the early results show its Trillium TPUs hold up well and can survive a radiation dose greater than a five-year mission in space. Trillium is the name that page uses for the TPU generation under test. Those results are ground tests. They are not readings from this week in orbit. The orbit readings are what the Oct 1 post says the coming weeks are for.
Cooling, and the flight that is still later. The Sep 24 page says TPUs make a lot of heat in a small space, and that a vacuum has no air to carry heat away. The heat has to leave through radiators. The team is working on heat pipes plus radiators, and it has tested that approach in a chamber that copies both the temperature and the vacuum of space. The page says the check of that cooling in orbit is still ahead. Future satellites, it says, would each carry dozens of TPU chips and fly in clusters, talking by laser so they can share a large job. Most laser links in space today, the page says, are built for lower bandwidth across long distances. Google’s would need very high bandwidth across a short gap, with the kind of aim it compares to hitting a coin-size target from miles away while both ends are moving. The page says that laser test is in 2027, when two satellites go up. It calls this first launch a way to see what works, find what fails, and carry those lessons to later missions. Those lines are the Sep 24 page. They are not a second satellite already in orbit, and they are not a cluster flying today.
The paper behind the idea. The Oct 1 post says a peer-reviewed paper on the research behind the mission is available in Joule. The link on the post goes to “Toward a future space-based, highly scalable AI infrastructure system design.” The author line names Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, and James Manyika, at Google. The paper’s summary describes fleets of satellites with solar arrays, free-space optical links, which are laser links through open space, and TPU chips. It illustrates formation flight with an example cluster of 81 satellites inside a radius of 1 kilometer. One kilometer is about 0.6 miles. That mile line is arithmetic on the paper’s kilometer. It says Trillium TPUs were radiation-tested and survived a total ionizing dose equal to a five-year mission life without permanent failures. It says a learning-curve look at launch prices suggests a ride to low Earth orbit may reach $200 a kilogram or less by the mid-2030s. A kilogram is about 2.2 pounds. Those figures are the paper’s design study. They describe a future architecture, not the single prototype that launched on Thursday. The PDF’s citation line reads Joule volume 11, article 102678, February 17, 2027. The Oct 1 post says the paper is available now.
The picture is Google’s Project Suncatcher graphic. The background runs from yellow to magenta. Centered type reads Project Suncatcher, with a small sun above the word Project. The frame does not print a calendar date. It is the official lockup from the Oct 1 post. It is not a photograph of the satellite, the rocket, or a chip.
In plain terms, Google said on Thursday that a Project Suncatcher prototype, built with Planet and launched on SpaceX’s Transporter-18, is in orbit, in contact, and operating as expected. The coming weeks are for measuring how the TPUs take the stress, the radiation, and the heat of space. A Sep 24 page had already described the ground work: a shake test the hardware passed, and Trillium chips in a proton beam that took a dose above a five-year mission. Cooling with radiators in a vacuum, and a 2027 flight of two satellites linked by laser, are still ahead on that page. The Joule paper is the research design, including an 81-satellite sketch and a launch-price estimate. Thursday’s news is one prototype in orbit. It is not a space data center open for business.