Google has taken an early but significant step toward running artificial intelligence in orbit. The company has sent one of its Tensor Processing Units, or TPUs, into space aboard a SpaceX rocket launched from California.
The mission is part of Project Suncatcher, Google’s long-term effort to explore whether networks of satellites could eventually operate like data centers above Earth. However, turning that vision into practical infrastructure may depend on SpaceX achieving an extraordinary increase in Starship launches.
Testing an AI chip in real space conditions
The experimental satellite, built by Planet Labs, will determine whether Google’s custom AI processor can operate reliably in orbit. Engineers have already conducted laboratory tests, but space introduces challenges that cannot be fully reproduced on Earth.
The system must supply steady power, control the heat produced by the processor, and protect its electronic components from radiation. Once the satellite becomes operational, its TPU will run AI models in short sessions of approximately 15 minutes. These limited tests are designed to prevent the spacecraft’s power and cooling systems from becoming overloaded.
This first mission uses an existing Planet Labs satellite design. A more advanced demonstration is planned for next year, when Google and Planet Labs expect to launch two satellites built more specifically for demanding computing tasks. Those spacecraft will also test laser-based communication, allowing them to exchange information directly.
Google’s bigger orbital computing plan
Project Suncatcher is not intended to become a commercial data center immediately. Google describes it as a long-range research program designed around the computing requirements that may emerge over the next several years.
Its larger concept involves 81 satellites traveling in a carefully coordinated formation. Instead of working independently, the spacecraft would divide and process AI workloads across multiple chips.
That level of coordination creates a major technical challenge. Modern AI systems require processors to exchange large amounts of information with very little delay. If communication between satellites is too slow or unreliable, the network may struggle to handle the kinds of complex workloads routinely managed by terrestrial data centers.
The launch problem is enormous
Even if Google solves the computing, cooling, and communication problems, it must still get a huge amount of equipment into orbit at an affordable price.
Google’s research suggests that launch costs could fall to around $200 per kilogram by 2035 if SpaceX continues reducing expenses at a rate comparable to its historical progress. Reaching that level, however, may require Starship to deliver approximately 370,000 metric tons of cargo to orbit.
Assuming each mission carries 200 metric tons, Starship would need to complete about 1,800 launches over ten years. That averages 180 flights annually—a remarkable target for a rocket that has not yet demonstrated anything close to that operational pace.
SpaceX has presented extremely ambitious plans for Starship, including the possibility of very frequent launches. Nevertheless, there is a considerable difference between a future projection and a dependable transportation system capable of flying hundreds of missions every year.
Radiation brings mixed results
Google’s testing indicates that its TPUs may survive space radiation well enough for many common AI applications. Researchers repeated some experiments after discovering that the chips initially received more shielding than they would have in orbit.
The revised tests recorded additional errors, but the overall failure rate remained low. Google believes the processors could perform large inference workloads throughout a satellite’s expected five-year life.
Training massive AI models would be more difficult. A tiny error rate becomes much more serious when thousands of processors operate continuously for months. For now, orbital systems may therefore be better suited to running trained models than creating them.
A moonshot with years to go
Google’s satellite is an important experiment, but it is only the beginning. Space-based data centers will require reliable chips, efficient cooling, fast satellite communication, and dramatically cheaper launches.
The concept may be technically possible. Whether it becomes economically realistic will depend not only on Google’s processors, but also on whether rockets can begin flying at a scale the space industry has never achieved.
