Google prepares to test whether its AI chips can withstand orbital conditions

The chips running artificial intelligence in Google’s data centers will have to face a very different environment before the company can consider deploying them throughout space. On September 24, Google announced Project Suncatcher’s first orbital test: a prototype developed with Planet, scheduled to fly on SpaceX’s Transporter-18 rideshare mission the following week. The immediate task is to find out how the equipment withstands launch and keeps working in orbit.

The payload carries TPUs, electronic accelerators specialized in the mathematical operations used by neural networks. These processors are designed to perform the repetitive calculations underlying AI systems and currently operate in data centers on Earth. Moving them onto a satellite changes the conditions under which they must do that work: beyond electrical power, they need to tolerate energetic particles and have a way to remove heat without surrounding air.

The first obstacle is the journey itself. During ascent, the rocket subjects the assembly to acceleration and vibration. In ground tests, engineers shook the satellite along three axes, reproducing frequencies expected during launch. According to Google’s Travis Beals, the hardware held up. The flight will allow that controlled result to be compared with the mechanical stresses of the actual mission, rather than treating the simulation as the final answer.

After launch, radiation adds a less visible threat. Energetic particles can pass through components and alter electronic information. A bit flip, for example, changes a unit of digital information between zero and one; depending on where it occurs, it can affect a calculation in progress. Materials can also suffer cumulative damage. Withstanding the dose received over time and performing operations without errors are therefore different aspects of reliability.

To examine this behavior, the team exposed Trillium-generation TPUs to a proton beam at the University of California, Davis’s Crocker Nuclear Laboratory while running AI workloads. Engineers monitored errors during processing. Google says initial results show tolerance of a total ionizing dose greater than that expected over a five-year mission. The announcement provides neither an absolute dose value nor a comprehensive error rate, so this comparison does not yet describe the full performance expected in orbit.

The other test will be thermal. A chip concentrates heat production in a small area. On Earth, moving air can carry away some of that energy; in a vacuum, that option disappears. The system must move heat from the processors to radiators, surfaces that release it by emitting thermal radiation. Google tested a combination of heat pipes and radiators in a chamber that reproduces thermal and vacuum conditions. The prototype will show how this approach works during operation in space.

The prospect of harnessing solar energy in orbit motivates the project, introduced in 2025. Turning that energy into computing capacity, however, will require more than keeping a chip switched on. A phase planned for 2027 is intended to test two satellites and laser links capable of exchanging large volumes of data. That communication would be needed for separate machines to work together; it belongs to future development, not to the results the first flight can deliver.

For now, Suncatcher is taking an engineering question into space on three concrete fronts: mechanical resilience, behavior under radiation, and heat removal. The answers could guide the design of subsequent missions. The costs, maintenance, and energy balance of a larger infrastructure remain open questions. This prototype’s advance will be to turn laboratory tests into real operating data, clarifying which obstacles still separate chips in orbit from useful computing at scale.