Google Suncatcher Will Test AI Tensor Chips in Orbit

Google has conducted ground-based tests of its Suncatcher project, but launching the satellite will reveal the full scale of the challenges involved in running artificial intelligence hardware in space.

Why Google needs to test Tensor chips in space

One key goal is to evaluate how Google’s Tensor Processing Units (TPUs) handle the conditions of orbit. Hardware used on many space missions is specially designed to withstand extreme temperatures and radiation, while the TPUs in Suncatcher are the same chips Google integrates into its ground-based servers.

Radiation can damage a chip or flip individual bits, disrupting calculations. Vibrations during launch and high gravity also create risks. However, missions such as NASA’s Ingenuity Mars helicopter have shown that commercially available hardware can survive challenging outdoor environments.

Cooling AI chips could be the biggest challenge

Cooling may be the most difficult issue, particularly as the AI industry explores the possibility of orbital data centers. Spacecraft already use radiator systems to remove heat, but those systems are generally designed to handle relatively small amounts of heat.

AI accelerators produce substantially more heat. Google’s approach uses a layer of malleable thermal interface material to connect the chip to aluminum and copper heat pipes. The heat pipes then carry heat to a radiator, which emits it into space.

Google’s TPU can run for only about 15 minutes

Google runs Gemini models on its TPUs during testing, but the chips cannot operate continuously. The cooling system can function for only about 15 minutes before the TPU must be shut down so the radiator can catch up.

Suncatcher could launch in 2027

Google hopes the initial test will show the team what works and what needs to change, helping improve the design for future missions. A launch is still planned for 2027, but Google expects Suncatcher will take years to develop from a project into a product.