Google Is Sending Its AI Chips to Space to See If They Can Handle It
Google is about to launch a satellite carrying its Tensor Processing Units, the special chips it uses for AI, to test how they work in space. The spacecraft is scheduled to fly on a SpaceX Falcon 9 rocket on October 1 as part of Project Suncatcher, Google's early experiment that could one day lead to AI data centers in orbit. The Verge described the flight after The New York Times first reported the plan.
This flight is only a test. Google wants to measure how the chips handle the shaking of launch, plus radiation and extreme heat and cold in space. The mission stays as a hardware check. Google has not said it will run any real work on this trip.
Heat is a central problem. Google will test a cooling system with heat pipes and radiators in orbit. Travis Beals, Project Suncatcher's senior director of product management, told the Times the chips can only run for about 15 minutes before they must be turned off to cool down. On Earth, fans move air to cool computers. In space there is no air, so heat must be moved to surfaces that release it, much like a car radiator, and computing must happen in short bursts.
The October flight starts a longer plan. Google plans to launch two satellites next year toward a group of chip-carrying satellites. An earlier schedule put the first test launches in 2027, as reported in January. NYT January The current plan moves the first chip flight up to October and keeps the multi-satellite tests for next year.
Google calls Project Suncatcher "a research moonshot to scale machine learning compute in space." The idea is a connected network of satellites powered by the sun. Google Blog Each payload pairs the AI chips with laser links so satellites can talk to each other, a setup detailed by Google Research. Google is working with Planet on a learning mission. Google announced the effort in November as a space data center project.
What matters here is not top speed in space. It is whether the chips can keep running steadily with limited power, heat limits and higher error rates than on the ground. The 15-minute limit means work would need to be split into short chunks with saved progress. The laser links would need very exact aiming if work is ever shared across satellites. Solar power removes one limit from Earth and adds another, power with no grid to smooth it out.
In my view, it helps to keep expectations level. This is a test of the environment and cooling, not a working data center. The useful result will be real measurements of how modern AI chips act outside a data center. If heat and radiation prove manageable, space systems could one day add to ground capacity using direct sunlight and space cooling. That is far off, which is why the data from October and next year matter more than the launch.
