Google has achieved its first AI response from an orbital data center prototype launched on a SpaceX Falcon 9 rocket. During a Svalbard ground station pass, its onboard model generated a verified space haiku, marking a milestone in a long-term research moonshot to power AI compute in low Earth orbit.
The first hardware for Project Suncatcher reached low Earth orbit on October 1, riding a SpaceX Falcon 9 rocket alongside more than 100 other spacecraft from Vandenberg or Cape Canaveral during a Transporter-18 rideshare mission. Named MVP by Google, the initial test craft carries four Google Tensor Processing Units and generates up to 1 kW of electricity from its solar panels. Because of its power demands, the system can operate only one chip at a time for roughly 20 minutes during a pass, creating a distinct hurdle for space-to-ground communications.
After a three-week commissioning process that tested out the satellite’s subsystems, engineers on the ground prompted the pre-loaded Google DeepMind Gemma 4 open model to write a poem about being in space. As the spacecraft passed Svalbard, Norway, the downlinked verses confirmed the hardware’s success.
Blue marble below
Gemma 4 open model, via Google Research and TechCrunch
Data streams in silent grace
Starlight feeds my core. ✨
Technical Hurdles and Space Radiation Testing for Orbital TPUs
Getting a response from orbit was only the initial hurdle. Google’s ongoing test campaign will evaluate how the custom Tensor Processing Units perform under thermal stress and the harsh radiation environment of low Earth orbit. While engineers have extensively tested the chips in particle accelerators on Earth, reality in space introduces unsimulated variables.
We’re sending TPUs to space (yes, really).After years of research, we’re launching a satellite to evaluate if and how Google Tensor Processing Units (TPUs) hold up in orbit.The test mission, as part of our latest moonshot — Project Suncatcher — is designed to gather data… pic.twitter.com/LaPuYonb3w
— https://x.com/google/status/2103229008343126519
Space radiation can trigger “bit flips,” altering stored values in computer memory and quietly corrupting data without tripping standard monitoring tools. Ground simulations suggested that radiation would impact inference workloads minimally while posing a greater risk to training models, a dynamic the team will now measure directly in flight.
Project Suncatcher Evolution and the Economics of Space Compute
Google Research formally announced Project Suncatcher on November 4, 2025, accompanied by a preprint outlining a scalable architecture for machine learning in orbit. The core proposition rests on bypassing terrestrial energy constraints by deploying solar arrays in space, where continuous sunlight avoids grid bottlenecks.
Travis Beals, the senior director leading Project Suncatcher, told Scientific American that the team began with deep skepticism. As a first step, we tried to find reasons that it was impossible,
Beals said, but we gradually became convinced that it might actually work.
Despite the technical progress, the economics remain a massive hurdle. Amanda Peterson Corio, Google’s global head of energy and power, admitted at an industry summit that space data centers are currently up to 100 times more expensive than building terrestrial infrastructure. Suncatcher calculations indicate that launch costs to low Earth orbit must drop significantly, targeting roughly $200 per kilogram to make amortization comparable to terrestrial energy costs of roughly $570 to $3,000 per kilowatt-year.
Timeline of Project Suncatcher Milestones
- November 4, 2025: Google Research announces Project Suncatcher and releases a preprint detailing orbital machine learning system designs.
- June 2026: SpaceX highlights orbital data centers in investor materials following its public listing.
- October 1, 2026: The prototype satellite carrying four Google Tensor Processing Units launches on a SpaceX Falcon 9.
- Late October 2026: The spacecraft completes its three-week commissioning phase and successfully executes its first AI prompt over Svalbard, Norway.
- 2027: Google plans to launch twin interconnected satellites to test higher power generation and optical inter-satellite links.
Frequently Asked Questions About Google Orbital AI
Is the Suncatcher prototype a fully operational space data center?
No. The spacecraft carries just four Trillium Tensor Processing Units and about 1 kilowatt of solar power, delivering roughly the compute capacity of a single cloud TPU slice rather than a commercial data center.
What are the next steps for the Suncatcher team after this initial test?
Engineers will continue evaluating how the chips handle thermal stress and space radiation before attempting to launch two interconnected satellites in 2027 to test high-speed optical links between spacecraft.
How does Google plan to achieve economical launch costs for future satellites?
Project Suncatcher researchers estimate that achieving target cost reductions will require SpaceX to fly its fully reusable Starship heavy-lift rocket roughly 1,800 times.