Google’s first data center test in space runs Gemini in 15-minute bursts

Google is about to put its AI chips in orbit, and the first thing they will have to learn is how to take a break. A refrigerator-sized satellite called MVP, built with the satellite company Planet, carries four of Google’s Tensor Processing Units, the same family of chips that answers Gemini queries in its data centers on the ground. Once in orbit it will run Gemini models for about a quarter of an hour at a time, then pause so its radiators can dump the heat into space.

That pause is the whole experiment in miniature. Project Suncatcher, the Google research moonshot behind the launch, asks whether the electricity AI devours could come straight from the sun instead of from the power grids that homes, hospitals and factories also depend on. In the right orbit a solar panel collects up to eight times more energy than the same panel on the ground, with no clouds and almost no night. If AI chips can survive and work up there, the most power-hungry part of the internet could one day move off the planet.

For now the ambition is modest. MVP’s solar panels deliver about one kilowatt, roughly what a microwave oven draws, and its four chips add up to about the computing power of a single server in a ground data center. Google frames the mission as a plain question: can its AI hardware operate in space at all? “Some things can only be tested in space,” the company said. “Putting our first TPUs in orbit next week will help us get data and learnings to inform future launches.”

The ground tests give some reason for optimism. Google shook the hardware on all three axes to mimic a rocket ride, in which individual components can feel 50 to 100 g. It fired a proton beam at its Trillium TPUs at UC Davis’s Crocker Nuclear Laboratory while they ran AI workloads, watching for flipped bits. According to Travis Beals, senior director of Google’s Paradigms of Intelligence team, the chips withstood a total radiation dose greater than they would absorb during a five-year mission. In an earlier research paper, Google reported that the chips’ high-bandwidth memory only began to show irregularities at nearly three times that expected dose, and no chip failed outright even at 20 times it.

Heat is the harder problem, and the 15-minute limit shows it plainly. On Earth, data centers blow air or pump water across chips to carry heat away. In a vacuum nothing carries it, so MVP routes heat through aluminum and copper heat pipes to a radiator that can only shed it as infrared light, a much slower process. A real orbital data center would also face repairs no technician can make, the delay of sending answers back to the ground, and launch prices far above what the economics require. Google has not said how much the project costs. Its own research concluded that computing in orbit would only rival the energy costs of a ground data center once launches fall below $200 per kilogram, which Google projects could happen around the mid-2030s.

Nothing changes for Gemini users yet, in any market. Every answer still comes from a data center on Earth, and MVP’s output is measurement data, not a new service. Google is also not alone in the race. SpaceX has outlined orbital AI data centers built with Nvidia technology that could launch by late 2027, Yahoo Finance reported, and its AI1 satellite design targets about 120 kilowatts of sustained computing, more than a hundred times MVP’s power budget. Jeff Bezos and OpenAI‘s Sam Altman have also backed the idea of computing in orbit.

MVP is scheduled to lift off on October 1 aboard a SpaceX Falcon 9 from Vandenberg Space Force Base in California, as part of the Transporter-18 rideshare mission. It will fly a dawn-dusk sun-synchronous orbit, tracing the line between day and night so its panels stay in near-constant sunlight, and it is expected to run tasks for about a year, although the hardware could stay up for as long as six years before it falls back into the atmosphere. In 2027 Google plans to fly two satellites to test the laser links between them that a real constellation would need. Its published design imagines clusters of 81 satellites flying within a kilometer of each other.

The number to watch after launch is not how fast MVP computes but how long it can keep computing. If 15 minutes stretches toward an hour, a data center in orbit starts to look like engineering rather than science fiction. If it does not, AI’s power bill stays firmly on Earth.

Tags: Google, Sam Altman, SpaceX, Gemini, data centers, Planet