Google's Project Suncatcher runs an AI workload in orbit and beams results to Earth

A haiku from orbit

Google's experimental Project Suncatcher satellite has run an AI workload in orbit and transmitted the result back to Earth, a milestone in the company's effort to find out whether its AI chips can operate in space. Computerworld reported the development on Nov. 3, about a month after the spacecraft launched.

The prototype, called MVP, lifted off on Oct. 1 aboard a SpaceX Falcon 9 flying the Transporter-18 rideshare mission. According to Heise, Google brought the satellite's AI accelerators and other onboard hardware online and ran initial machine-learning tests after launch. The first data produced by the onboard hardware and software reached Earth after several weeks of commissioning.

Four TPUs and one kilowatt

The spacecraft carries four Tensor Processing Units, the chips Google built for running AI models, and solar panels supplying up to one kilowatt of electricity, according to Data Center Dynamics. It is a test platform for operating those processors in orbit, not a full-scale computing facility.

On Oct. 30, Google announced that Gemma 4, a model developed by Google DeepMind and loaded onto the satellite before launch, had produced its first successful AI response. During a pass over a ground station, engineers prompted the model to write a haiku about space. It replied: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

Google also tried space-themed limericks and prompts asking the model to report telemetry on the satellite's systems, Data Center Dynamics reported. These tasks show the onboard system can generate responses and send them home, but they say little about how the hardware will handle longer or more demanding computing jobs.

Power and heat constraints

Reports disagree on how long the processors can run at a stretch. TechRepublic said the satellite's available power limits operation to one TPU for about 20 minutes. Heise, citing The New York Times, reported the accelerators can run for 15 minutes before they must be shut down to cool.

Google's next tests will examine thermal stress and radiation exposure, both of which can affect processors in orbit. Cooling is a particular engineering problem because there is no air in space to carry heat away; before launch, Google said it was working on approaches including heat pipes and heat sinks. Heise also pointed to broader challenges for the concept, including the cost of transporting equipment and limits on data transmission back to Earth.

From one satellite to a cluster

Google plans to test two interconnected satellites in 2027, Data Center Dynamics reported. TechRepublic said the pair are expected to demonstrate optical links for direct satellite-to-satellite communication and to have greater power-generation capability. The longer-term vision, according to Data Center Dynamics, is a cluster of as many as 81 satellites sharing computing work.

The cost question

Economics remain the largest obstacle. Amanda Peterson Corio, Google's global head of energy and power, said space data centers are "incredibly, incredibly expensive, up to 100 times more expensive than building on Earth," and that the company should explore "all the terrestrial options before we go extraterrestrial," according to Data Center Dynamics.

Google has previously estimated that launch costs would need to fall to $200 per kilogram to low Earth orbit for launch expenses, spread over a spacecraft's lifetime, to be roughly comparable with data-center energy costs per kilowatt. The company speculated that level might be reached by 2035 if SpaceX's Starship begins flying and conducts 180 launches per year. That figure is a projection, and terrestrial power prices could shift by then as well.

Why it matters

AI workloads are straining the power, land and water budgets of terrestrial data centers, pushing operators to look for alternatives. Suncatcher is an early experiment in whether some of that compute can move off-planet, and this week's result is the first concrete evidence that Google's TPU hardware and a Gemma model can complete real work in orbit. The gap between a haiku generated by a single TPU running for minutes and a data center running nonstop is still enormous. The 2027 two-satellite test, the 81-satellite cluster concept and the $200-per-kilogram launch-cost threshold will determine whether orbital AI compute remains a research project or becomes actual infrastructure.