Google Project Suncatcher: The First Physical Step Toward Orbital AI Compute

Google's Project Suncatcher will launch its first physical test payload, a refrigerator-sized Planet Labs satellite carrying four Google Trillium Tensor Processing Units, on a SpaceX Falcon 9 rideshare mission from Vandenberg Space Force Base on October 1. The mission, announced in November 2025, will run the TPUs for roughly 15 minutes at a time before thermal shutdown forces a cooldown, gathering in-orbit data on launch vibration and acceleration loads up to 10 times the force of gravity, radiation, and thermal extremes. Google says ground tests at UC Davis's Crocker Nuclear Laboratory showed the Trillium chips survived a total ionizing dose greater than a five-year space mission would deliver, while SpaceX separately targets about 1 gigawatt per year of orbital compute by the end of 2027 via its Nvidia-powered AI1 satellites.

The Orbital Compute Frontier On October 1, a refrigerator-sized satellite built by Planet Labs will ride a SpaceX Falcon 9 into low Earth orbit from Vandenberg Space Force Base. Inside it: four Google Trillium Tensor Processing Units, a kilowatt of solar power, and a question that has been theoretical until now. Can AI hardware survive the physical and thermal realities of space? The satellite is called MVP, and it is the first physical artifact of Google’s Project Suncatcher https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/ , a research moonshot announced in November 2025 to explore whether space could one day host scalable machine learning infrastructure. The launch, part of SpaceX’s Transporter-18 rideshare mission, is not an operational data center. It is a survivability test. The TPUs will run compute for approximately 15 minutes at a time before thermal shutdown forces a cooldown. That constraint—15 minutes of work, then silence—is the most honest thing about the project. What the 15 Minutes Tell You Google’s own blog post, published September 24, frames the mission as “measured, deliberate steps” toward a long-term goal. The immediate objective is to gather in-orbit data on how Trillium TPUs handle the physical stress of rocket launch—vibration, sustained acceleration loads up to 10 times the force of gravity—radiation exposure from solar events and cosmic rays, and thermal extremes. In a vacuum, there is no airflow. Heat can only be dissipated via radiators, a fundamentally different cooling architecture than anything terrestrial data centers require. Prior ground testing offers some confidence. At UC Davis’s Crocker Nuclear Laboratory, Google’s team exposed Trillium TPUs to proton beam irradiation while running AI workloads. The chips survived a radiation total ionizing dose greater than what they would receive during a five-year space mission. Thermal vacuum chamber tests confirmed the heat-pipe-and-radiator approach works in simulated conditions. But some things, as Google’s team acknowledges, can only be tested in orbit. The Race Beyond the Atmosphere Google is not alone. SpaceX is separately targeting approximately 1 gigawatt per year of orbital compute by the end of 2027 via its own Nvidia-powered AI1 satellites. Starcloud and Axiom are pursuing similar solar-powered orbital AI data center concepts. The driving force is the same across all of them: terrestrial data centers are hitting physical and political limits. Power availability, cooling requirements, land use, and permitting delays are constraining the expansion of AI compute infrastructure on the ground. In low Earth orbit, satellites can access near-constant sunlight, generating up to eight times more solar power than on Earth’s surface. The physics are attractive. The engineering is brutal. Vertical Integration, All the Way to Orbit Google’s approach is notable for one reason: vertical integration. The payload is Google’s own silicon—Trillium TPUs, the same chips that power Google Cloud’s TPU v6e-4 slices. The satellite bus is built by Planet Labs, a partner, but the compute layer is entirely Google’s. This is the compute landlord thesis extending from terrestrial GPU factories to orbital platforms. We have traced this pattern through SpaceX’s $119 billion Terafab initiative https://forkast.news/spacex-and-tesla-are-building-their-own-119-billion-chip-factory-and-the-compute-landlord-thesis-just-got-its-most-expensive-chapter/ , through Nvidia’s $10 billion anchor in Anthropic’s IPO, and through the $517 billion in compute commitments Anthropic accumulated in 11 months. The entity that controls the chip, the cloud, and now the orbital infrastructure controls the stack from bottom to top. A 2027 mission will test high-bandwidth free-space optical laser inter-satellite links between two satellites—the precursor to distributed orbital compute. The long-term vision involves clusters of satellites carrying dozens of TPU chips each, connected by laser arrays. Google’s own modeling suggests cost parity with terrestrial data centers could arrive in the mid-2030s, contingent on launch costs falling below approximately $200 per kilogram. Gartner, in February 2026, called the concept “pie-in-the-sky.” What the Launchpad Tells You The skepticism is not unwarranted. The current mission can only run compute for 15 minutes. The thermal management challenge—how to dissipate the heat that TPU chips generate in a vacuum—remains the critical engineering hurdle. The economic viability depends on external aerospace market dynamics that Google does not control. And the timeline from “survivability test” to “operational orbital data center” is measured in decades, not years. But the signal matters more than the timeline. When a hyperscaler with its own silicon, its own cloud platform, and its own AI models begins physically testing that silicon in orbit, it is no longer a thought experiment. It is a capital allocation decision. The 15-minute compute window tells you how early this is. The four Trillium TPUs tell you how serious Google is about owning the entire stack. And the October 1 launch tells you the race for orbital AI compute is no longer theoretical—it is hardware on a launchpad.