Google Tests TPU Chip in Earth Orbit, Paving the Way for AI Data Centers in Space

Google is preparing to launch a prototype satellite carrying a Tensor Processing Unit (TPU) to low Earth orbit (LEO) next week through SpaceX's Transporter-18 rideshare mission. The move marks the first trial of Project Suncatcher, an ambitious research initiative exploring the possibility of placing large-scale AI computing infrastructure in space. The satellite was developed together with Planet, a satellite imaging company, and will serve as an initial proof of whether AI hardware can survive the extreme environment of space.

The mission is not about operating an orbital data center commercially, but rather about collecting real-world data on TPU performance under launch stress, cosmic radiation exposure, and drastic temperature swings in LEO. Google disclosed that ground testing had been carried out, including multiaxial shock simulations equivalent to forces 50–100 times Earth's gravity. In addition, the Trillium TPU chip was tested with a proton beam at the Crocker Nuclear Laboratory, University of California, Davis, while running AI workloads. As a result, the chip withstood radiation accumulation exceeding the expected exposure over five years of an orbital mission.

The next challenge is heat management. In a vacuum, the heat generated by the TPU cannot be dissipated through convection, so it relies entirely on radiative surfaces. Google's current thermal design combines heat pipes with radiators, which have been validated in thermal vacuum chamber tests on the ground. This approach is crucial because the TPU operates at high power density that can cause overheating if not managed properly.

The next testing phase is scheduled for 2027, when Google plans to place two satellites to evaluate the high-bandwidth laser links needed by future TPU-carrying satellite clusters. Existing laser communication systems are generally designed for low bandwidth over long distances, whereas Google's design demands high bandwidth over short distances—a precision challenge the company likens to hitting a coin-sized target from miles away while both points are moving.

Low Earth orbit is already crowded with about 44,870 tracked objects, the majority of which are space debris. The full Suncatcher architecture envisions a cluster of 81 satellites flying in formation within a one-kilometer radius at an altitude of about 650 kilometers, one of the busiest orbital layers. Experts assess that the concept is still several years away from commercial viability given the high launch costs, engineering constraints, and satellite production hurdles.

"We are not only testing whether the chips can survive, but also how they behave under conditions that cannot be fully replicated on Earth," a Google spokesperson said, as quoted from the same source.

Google is not alone. Nvidia-backed startup Starcloud launched an H100 chip into orbit last November. SpaceX is also targeting the end of 2027 for its first Starmind AI satellite powered by Nvidia chips. Competition in the realm of orbital computing shows that the industry is starting to seriously consider space as a strategic location for energy-intensive AI workloads.

For Indonesia, this development offers both opportunities and challenges. As an archipelagic country with high connectivity needs, orbital computing infrastructure could accelerate AI services in remote areas, but it also demands regulatory readiness and human resources. The government needs to anticipate the implications for space governance, including the potential buildup of satellites in low orbit that could disrupt national satellite operations. In addition, dependence on foreign technology for orbital computing could become a digital sovereignty risk if not matched by local research.

Looking ahead, the big question is not only whether TPUs can survive in space, but whether the orbital data center business model will be economical enough to replace part of ground-based infrastructure. If Google and its competitors succeed in proving this concept, the global computing landscape could shift fundamentally—with major implications for countries that want to harness AI without being limited by the physical location of data centers.