MOUNTAIN VIEW, September 26 — Google will put its first test satellite for orbital data centres into space next month, carrying the tensor processing units it designed in-house, the company has confirmed. The spacecraft is due to fly on SpaceX’s Falcon 9 Transporter-18 rideshare mission, scheduled for October 1. The New York Times reported the launch plan earlier.
The satellite is part of Project Suncatcher, a Google moonshot aimed at gathering the data needed to work out how the company could host machine-learning infrastructure in orbit. Google has been working with Planet on the test spacecraft.
Custom silicon, bound for vacuum
Under the hood, the payload is a set of custom application-specific integrated circuits designed by Google. The on-orbit trials are intended to establish whether those chips can handle the physical stress of launch, the radiation environment and the thermal conditions in space. That is a materially different set of problems from the ones a terrestrial data centre faces, and Google is not pretending otherwise.
“Can our TPUs survive and operate in space? Well, we’re going to find out,” said Sundar Pichai, chief executive of Google and its parent company, Alphabet Inc. He added a line of his own: “One small step for TPUs…” Maria Biggs, an engineering director at Google, framed the project’s central question.
“How can we efficiently scale AI infrastructure through orbital data centers by leveraging the abundant energy we get from the Sun?” she said.
The energy argument
Travis Beals, Google’s Project Suncatcher lead, said using solar panels in orbit is more energy-efficient than trying to do so on Earth. “If you put a solar panel in the right orbit, it generates five to eight times more power than that same panel down on Earth,” Beals said. Google has said it is looking at different ways for heat dissipation to keep the TPUs cool, among them a combination of heat pipes and radiators, with tests in vacuum.
The on-orbit trials are intended to see how the system may need to evolve. During ground trials, the company has been looking to assess whether TPUs could survive five years of orbital life.
The architecture Google is sketching out is not a single spacecraft but a constellation. “Future designs of our satellites will each carry dozens of TPU chips while orbiting the Earth in clusters,” Google said. Those clusters would be networked via laser links, to be tested next year to validate high-bandwidth data transmission over short distances.
Google next year plans to launch two further satellites in early 2027 to validate intersatellite laser links. SpaceX, which is providing the ride to orbit, has its own orbital data centre ambitions — a reminder that the launch provider and the payload customer are, in the longer term, potential competitors.
The detail that matters for now is modest and specific: one satellite, one rideshare slot, and a set of chips whose survival is an open question. What to watch next is the October 1 launch itself, then the early 2027 pair of satellites and the laser-link validation that would tell Google whether its clusters can actually talk to each other at the bandwidth the concept requires.
