Google plans first orbital AI chip test on 1 October SpaceX launch

Google said on 24 September in a Project Suncatcher update that it is preparing its first orbital test of Tensor Processing Units, with the prototype satellite scheduled to launch on 1 October, Ars Technica reported.

Four Google TPUs aboard Planet Labs’ MVP

The satellite, called MVP, is about the size of a refrigerator and contains four Google-designed TPU accelerators, according to Ars Technica. Its solar panels supply about one kilowatt of power. The spacecraft was built by Planet Labs PBC, while Google supplied the AI hardware, SiliconANGLE reported. SpaceX is due to carry it alongside other payloads on the Transporter-18 rideshare mission aboard a Falcon 9 rocket.

TPUs are specialised chips used both to train AI models and to run trained models when they produce answers. The units going into orbit are the same kind Google puts in ground-based servers. For this mission, four chips and a satellite’s solar panels replace the much larger collection of computing equipment and electrical supply available to a data centre on Earth.

Google first announced Project Suncatcher in November 2025 as an effort to design AI data centres in orbit, SiliconANGLE reported. The prototype has a narrower job: Google said the mission is intended to gather information about how its hardware behaves in space and where it might fail, rather than operate as an orbital data centre. Google plans to run Gemini models on the satellite’s TPUs.

SpaceX and Starcloud are also pursuing orbital data-centre plans that would draw on near-continuous sunlight, Reuters reported. Such proposals concern far more than whether a processor switches on in space: computing equipment must also survive the trip up, receive power, shed heat and communicate with other equipment. Google’s single-satellite flight is set up to examine some of those engineering demands at a small scale.

Crocker Nuclear Laboratory tests meet orbit

Google has already exposed its Trillium TPUs to a proton beam at the University of California, Davis’ Crocker Nuclear Laboratory while the chips ran AI workloads. Travis Beals, a senior director at Google, said researchers watched for errors and the workloads performed well despite a radiation dose greater than the chips would be expected to encounter during a five-year space mission, SiliconANGLE reported.

Radiation can damage electronics or change a bit of data while a chip is calculating, producing an error without visibly breaking the machine. The laboratory test put working chips in a radiation beam, so researchers could watch what happened to their calculations. The flight adds the conditions of an actual launch and orbit, including vibration, launch forces and changing temperatures. Google said orbital testing is needed to learn how its hardware performs in that environment.

Those conditions matter to companies considering hardware for space-based computing. A chip that completes workloads during a ground radiation test has cleared one particular test, while the planned flight will expose the satellite and its equipment to several stresses in succession. Google has described the mission as a way to find points of failure that can inform later designs.

Fifteen-minute runs before the TPUs cool

Heat is another constraint. On Earth, moving air can carry heat away from electronics; in the vacuum of space, Google’s design instead conducts it from the chips through heat pipes to radiators. A radiator releases heat outward, much as a household radiator transfers heat from hot water into a room, though the satellite must release its heat into space without surrounding air. Google has tested its cooling approach in a thermal vacuum chamber, Beals said.

Getting an answer to a written question would, if AI computing were put in orbit, depend on chips that can remain powered and cool while they work. This test would exercise that requirement only in short periods: the TPUs are expected to run workloads for about 15 minutes before shutting down to cool.

That operating pattern also limits what this satellite can test about a future cluster of AI hardware. Ars Technica reported that MVP will operate for just a few months, while Google expects it to take years for Suncatcher to develop from a research project into a product. Connecting multiple satellites would require another system beyond the hardware and cooling arrangement flying on MVP.

Google aims to launch two additional satellites in 2027 to test the high-bandwidth laser links it envisages using between future computing clusters, Reuters reported.