Google sent its first advanced Tensor Processing Unit (TPU) processing chip into orbit aboard a SpaceX rocket launched from California, in a practical step to test whether artificial intelligence workloads can operate beyond Earth. The satellite, built by Planet Labs, carries a prototype of Project Suncatcher, which aims to develop large computing clusters in orbit.
Testing the Chip in Real Conditions
The test focuses on the TPU’s ability to operate while receiving approximately one kilowatt of continuous power, managing cooling, and running a set of models to monitor potential failures. Travis Beals, the Google executive responsible for the project, said that ground tests cannot fully replicate real conditions.
After the satellite is placed into service, the chip will operate in 15-minute intervals to reduce pressure on the power and thermal-management systems. The current model does not rely on a platform designed entirely for advanced computing; it is built on a standard platform from Planet Labs.
From a Prototype to an Orbital Network
Google and Planet Labs are working on another experiment expected next year that will use two satellites more specialized for computing, while attempting to connect them through laser communications and run larger workloads. The long-term vision for Project Suncatcher is a network of 81 satellites operating in a close formation and processing data in parallel.
Google believes that connection speed and latency between TPU units will be critical when running workloads distributed across several racks. It therefore designed its concept with the computing needs expected five years from now in mind, rather than current applications alone.
Why Does This Matter?
The step shows that space data centers are still at the feasibility-demonstration stage and are not a ready substitute for terrestrial data centers. The challenge is not limited to sending a chip into orbit; it also includes creating a low-cost launch infrastructure capable of transporting enormous quantities of equipment at a recurring pace.
In a peer-reviewed version of a research paper to be published in the journal Joule, Google estimated that reaching a launch cost of approximately $200 per kilogram by 2035 could require transporting 370,000 tons into orbit, based on the assumption that SpaceX’s cost-reduction curve continues at approximately 20% annually. Assuming a payload of 200 tons per mission, that would mean about 1,800 Starship launches over ten years, or 180 launches annually.
This rate represents a major challenge, as Starship had not flown more than five times in a single year as of the date of the article. SpaceX and Elon Musk’s projections for increasing the launch rate in the future, including the possibility of reaching one flight every hour in 2029, remain projections that have not been demonstrated in practice.
Limits of Computing in Orbit
Google’s updated tests indicate that the chips may withstand space radiation throughout a satellite’s estimated five-year lifetime, with a relatively low error rate in ordinary inference operations—approximately one error per million operations, according to a statement by Beals. But the same level of reliability may not be sufficient for training massive models using thousands of chips over several months, placing a clear limit on the applications possible at the current stage.