Google launched its first experimental prototype satellite on October 2, 2026, carrying advanced Tensor Processing Units into low Earth orbit to test whether artificial intelligence infrastructure can survive the harsh physical and thermal realities of space.
Built in partnership with Planet Labs, the refrigerator-sized spacecraft lifted off on a SpaceX Falcon 9 rocket from Vandenberg Space Force Base in California, flying as part of the Transporter-18 rideshare mission alongside more than 100 other payloads. The satellite marks the initial in-flight test of Project Suncatcher, an ambitious long-term research moonshot evaluating whether space could eventually host scalable machine learning compute clusters. Travis Beals confirmed that the team established contact shortly after launch and that the hardware was operating as expected.
Testing Chips for Thermal and Radiation Limits
Before clearing the pad, Google subjected its Trillium chips to rigorous ground evaluations. Engineers rattled the hardware across all three axes to mimic launch frequencies, where individual components experience acceleration forces up to 50 to 100 times gravity during a 10-minute trip to low Earth orbit. Radiation resilience was tested at the Crocker Nuclear Laboratory at UC Davis, where the chips were placed in a proton beam.
According to Google, those tests showed the processors can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission
. However, repeat evaluations uncovered a slight increase in logic errors, pegged at roughly one bitflip per million for typical inference work. While that error rate is manageable for answering basic queries, it presents a substantial hurdle for months-long training runs spread across thousands of processors. Once fully commissioned in orbit, the payload will fire up its four Tensor Processing Units.
“Radiative cooling in space is more efficient than on Earth because of the low background temperature of deep space, but the convective / evaporative cooling mechanisms available on Earth in addition to radiative cooling here aren’t available in space at all,”
Dr. Ulrich Speidel, School of Computer Science, The University of Auckland, via Science Media Centre
Power Pressures and the Economics of Orbital Scale
The primary driver behind moving compute off-planet is energy.
Slava Turyshev, a researcher at NASA’s Jet Propulsion Laboratory, warned that deploying a single one-megawatt node would demand up to 6,000 square meters of solar panels and push total mass to roughly 75 tonnes.
Meeting Starship Flight Requirements for Orbital Networks
Realizing a commercial orbital network depends heavily on dramatic reductions in rocket payload costs. Alongside the hardware launch, Google released a peer-reviewed white paper accepted by the journal Joule, detailing the economic trajectory required to make space-based datacenters commercially viable.
The study projects that launch prices must drop to approximately $200 per kilogram by 2035. Achieving that price point relies on SpaceX continuing a historical cost-reduction learning curve of about 20 percent per doubling of cumulative mass since the launch of the Falcon 1 rocket. Based on Falcon 9 metrics, the authors calculate that SpaceX would need to fly its giant Starship rocket approximately 1,800 times over the coming decade, carrying 200 tonnes of payload on each flight and accumulating 370,000 tons of additional payload mass.
That projection assumes a flight rate of 180 missions annually—a demanding schedule for a launch vehicle that has historically flown no more than five times in a single year. Connecting hundreds of orbiting satellites into a functioning cluster demands ultra-high-bandwidth laser links capable of tracking targets from miles away while both points move. Google plans to test those inter-satellite laser links in orbit with a two-satellite test scheduled for 2027.
