Google Puts Four AI Chips In Orbit On Suncatcher Test

Google has put its AI chips in orbit for the first time. A prototype satellite carrying four Tensor Processing Units, built with the Earth-imaging company Planet, launched on 1 October aboard SpaceX's Transporter-18 rideshare mission from Vandenberg Space Force Base in California, alongside around 129 other satellites.

"Our team has confirmed contact with the satellite and it is operating as expected," said Travis Beals, senior director of Paradigms of Intelligence at Google, in the company's announcement.

The flight is the first hardware step in Project Suncatcher, Google's research programme on whether machine learning infrastructure could one day run in space. Over the coming weeks the company will gather data on how the chips cope with the physical stress of launch, with radiation, and with the thermal extremes of orbit, and it says it will use whatever it learns to refine future designs.

What The Mission Is Actually Testing

Three problems decide whether this idea works at all, and the satellite is built to measure each of them.

Radiation comes first. Google tested its Trillium generation TPUs on the ground at the University of California, Davis, using a proton beam at the Crocker Nuclear Laboratory, and says the chips survived a total ionising dose greater than they would receive over a five-year mission. The company has not published the dose figures, and that testing has not been peer reviewed.

Launch stress is the second. The spacecraft itself endures sustained loads of up to 10g, while individual components can see 50 to 100g, which is a demanding environment for hardware designed for a data centre rack.

Heat is the third and hardest. There is no air in orbit to carry heat away, so the prototype uses heat pipes and radiators, tested beforehand in a thermal vacuum chamber. Rejecting the heat from a chip that draws significant power is the constraint most likely to limit how much computing can be packed into a satellite.

Google plans to monitor the chips over roughly a year.

Why Orbit Appeals At All

The attraction is power. A satellite in the right low Earth orbit sits in near-constant sunlight and can generate up to eight times more solar power than an equivalent panel on the ground, with no grid connection, no local objections and no water for cooling.

That matters as terrestrial AI data centres run into limits on electricity supply and planning permission, which is why operators are chasing power wherever they can find it.

What Google Has Not Said

The company has withheld several details that would help assess the test, including the orbital altitude, the exact workloads the chips will run, and the power budget available to them. Four TPUs is a research payload rather than anything resembling useful capacity.

The next step is more revealing. Google plans a two-satellite mission in 2027 to test laser links between spacecraft flying in close formation, which is the capability any orbital cluster would depend on. One description of the challenge puts it at hitting a coin-sized target from miles away while both ends are moving.

Later designs would carry dozens of TPUs each, connected by those laser links, working towards a vision of 81 satellites flying together.

The Economics Remain The Harder Problem

Alongside the launch, Google published peer-reviewed research in the journal Joule setting out what orbital computing would cost. The paper estimates that SpaceX's Starship would need roughly 1,800 launches over the next decade, about 180 a year, to bring launch costs near the $200 per kilogram that orbital data centres require.

Starship has flown 13 times in total. The research also found that while error rates are low enough for inference, long training runs lasting months would accumulate problems, which points to a narrower role than orbital computing's advocates usually suggest.

Google is unusually direct about this. The paper sets out what would have to be true rather than claiming it will be, and the company calls the programme a long-term research moonshot.

What Businesses Should Take From It

For companies planning where AI workloads will run, nothing changes. Orbital capacity is a research programme with meaningful capacity in the 2030s at the earliest, and only if launch economics transform.

The signal worth reading is why a company with Google's resources is testing this at all. Power, cooling and land are now the binding constraints on AI infrastructure on the ground, hard enough that orbit is worth a year of flight data. That pressure is already shaping where data centres get built, what they cost and which countries can host them.

Others are moving too. Starcloud, backed by Nvidia, has run a GPU in orbit and raised $250 million in August, and SpaceX has its own ambitions for AI capacity in space.

Four Chips Is A Start, Not A Data Centre

What is in orbit today is a test article: four chips, a year of telemetry, and a set of questions about radiation, vibration and heat that could not be answered on the ground. Google has been careful not to call it more than that.

The findings will matter either way. If the chips hold up, the 2027 laser test becomes the next hurdle and the economics stay the hardest one. If they do not, the industry learns something useful about the limits of putting computing where it cannot be repaired.