Google is preparing to send its custom artificial intelligence chips into orbit for the first time, in an experiment that could pave the way for satellite-based data centres to power future AI systems.

The initiative, called Project Suncatcher, will see the technology giant partner with Earth-imaging firm Planet Labs on a series of tests.

Google is preparing to launch a prototype satellite carrying its Tensor Processing Units, or TPUs, in its first in-orbit test, while a separate mission involving two satellites is planned for 2027 to test the high-speed laser links needed to connect future computing clusters.

Google is not alone in eyeing the sky as a potential solution to AI's rising energy demands. Elon Musk's SpaceX and start-up Starcloud are pursuing similar concepts, betting that near-constant sunlight and an alternative to terrestrial power infrastructure could make orbit a new frontier for computing.

Starcloud has already flown a satellite carrying an Nvidia chip, while Reuters has reported that Google and SpaceX have held discussions about the orbital data-centre concept.

Space poses challenges no data centre on Earth has to contend with. Radiation can corrupt electronics or trigger so-called bit flips, in which a stored zero flips to a one.

Google has already tested its Trillium-generation TPUs in a proton beam facility at the University of California, Davis, simulating radiation exposure equivalent to more than five years in orbit, and says initial results showed the chips could withstand the dose while running AI workloads.

The testing was carried out at UC Davis' Crocker Nuclear Laboratory.

Cooling is the other major hurdle. With no air in space to carry heat away, conventional fans are useless. Google's design instead relies on heat pipes and radiators to draw heat from the chips and release it into the cold of space, a method the company still needs to prove works in orbit.

Future satellite designs could fly in close formation, connected by high-speed laser links designed to let them function as a single computing cluster. Google plans to test this technology with two satellites in 2027, while its longer-term research has explored configurations involving around 81 satellites operating within a one-kilometre radius.

Behind the ambition lies a straightforward problem: AI models are consuming ever more electricity, straining power grids and drawing political scrutiny on the ground.

Solar panels in orbit can generate several times more power than their Earth-based equivalents, an appeal that has drawn SpaceX, Starcloud and now Google into the same race.

Industry analysts remain cautious, however. High launch costs, the complexity of space-hardened hardware and bottlenecks in satellite manufacturing mean a fully operational orbital data-centre is likely years away, even if Google's prototype mission succeeds.

The company has been careful to describe Suncatcher as an early-stage research programme rather than a commercial product in waiting. Its first orbital test is intended to gather data on how the TPUs cope with the physical stresses, radiation and thermal conditions of space.

For now, the venture remains a modest but symbolically significant step: a prototype satellite, a set of chips, and a test of whether the next generation of AI infrastructure might not be built on Earth at all.