Google Sent AI Chips In Orbit With Project Suncatcher

Google has put AI chips in orbit for the first time. On the afternoon of October 1, 2026, a SpaceX Falcon 9 lifted off from Vandenberg Space Force Base in California carrying a refrigerator-sized prototype satellite built alongside the Earth-imaging company Planet, and as reported by Scientific American, the four Tensor Processing Units riding inside mark the first flight of Google's custom AI silicon in space.

The satellite, internally called MVP for minimum viable prototype, is not a data center. It is a survival test. Its four Trillium-generation TPUs — the same chips Google uses to train and run AI models on Earth — will run Gemini AI workloads in bursts of roughly fifteen minutes before shutting down to let the craft shed heat, since there is no air in orbit to carry warmth away. Contact with the satellite was confirmed the same evening, and it is healthy, according to a statement from Travis Beals, Google's senior director of Paradigms of Intelligence, who leads the project. "Some things can only be tested in space," he wrote on the company blog the day of the launch.

So why send computing off the planet at all? Because AI is ravenous. The data centers training and serving models like Gemini draw enormous amounts of electricity and water for cooling, straining power grids and sparking backlash in communities asked to host them. If you have ever watched a new AI data center get proposed near a neighborhood, this story is about the radical alternative: putting the computers where the sunlight never ends.

Why Google Wants Compute Off the Planet

Putting AI chips in orbit is really about one resource: clean power. In a dawn-dusk sun-synchronous orbit about 650 kilometers up, solar panels see near-constant sunlight and can generate up to eight times more electricity than identical panels on the ground, as the company detailed in a peer-reviewed paper published in the journal Joule alongside the launch. There is no grid connection to build, no cooling water to draw, and no local opposition to fight — advantages that are hard to buy on Earth.

Before this flight, Google put its chips through the wringer on the ground. Trillium TPUs faced a 67 megaelectronvolt proton beam at the University of California, Davis, simulating the radiation of years in orbit, and the company said they held up surprisingly well — withstanding more ionizing radiation than a five-year mission would deliver. Launch is its own ordeal: components experience sustained forces of 50 to 100 times Earth's gravity, according to Google. Over the coming weeks, the prototype will gather data on launch stress, radiation effects, and thermal behavior, and Beals described the mission to NPR as "a very minimal test" of whether the hardware can keep computing in the harsh environment of orbit.

The long-term vision, first laid out in November 2025, is far bolder: clusters of more than eighty satellites flying in tight formation, linked by laser beams moving data at over a terabit per second, effectively forming a solar-powered computer in the sky. Google plans to launch two follow-up satellites in 2027 to test those high-bandwidth laser links between craft, with later designs carrying dozens of chips each, as detailed in Scientific American's reporting on Google's orbital plans.

Skeptics Say the Hard Problems Remain

Cooling is the glaring weak spot. The MVP satellite's power budget is roughly one kilowatt — about what a hair dryer draws — and even that must run in short bursts before accumulated heat forces a pause. Reaching cost parity with Earth data centers by the mid-2030s would require launch costs to fall to roughly two hundred dollars per kilogram, and analysis cited by The Decoder estimates that about ten thousand satellites would be needed to match a single one-gigawatt ground facility. "It's usually the late entrants who finally figured out how to do it effectively," MIT's Kerri Cahoy told Scientific American, noting that SpaceX itself did not reach orbit until half a century after Sputnik — a reminder that the company that eventually scales orbital computing may not be Google.

Google is not alone in the race. SpaceX, the company that launched this very satellite, has sought permission to launch as many as one million satellites for its own orbital compute network, and in December 2025 the Nvidia-backed startup Starcloud sent its first AI-equipped satellite to space, according to recent reporting. Meanwhile, AI chips keep getting closer to everyday life too — see our coverage of NVIDIA's RTX Spark PCs bringing AI agents onto local machines, and browse more AI news on the topic page.

For now, four chips are circling the planet in short computing bursts — the smallest possible opening chapter of the Suncatcher experiment. If AI chips in orbit ever scale beyond this test, the cloud in your pocket could literally be in the clouds, running on sunlight instead of your local power grid and drinking water. That is a big if, but as Beals put it, some things can only be tested in space.