TL;DR
Google launched a single-server TPU prototype into low Earth orbit on October 1, proving its AI chips can survive launch but revealing that economics, heat, and a head-start by well-funded rivals are the real obstacles.
What happened
- Project Suncatcher launched October 1 aboard SpaceX Transporter-18 from Vandenberg Space Force Base, carrying four Trillium-generation TPUs built with satellite operator Planet.
- The four chips represent roughly one data center server's worth of compute, orbiting at approximately 650 kilometers altitude.
- Project lead Travis Beals confirmed contact and nominal operation; the satellite runs Google's open model Gemma in 15-minute compute windows before shutting down to cool.
- Ground radiation tests at UC Davis showed the main processing die tolerating roughly 20 times the dose a five-year mission would deliver, but high-bandwidth memory flagged irregularities at lower doses.
- A peer-reviewed paper in Joule accompanies the mission; two more satellites with inter-satellite laser links are planned for 2027.
Why it matters
- Heat, not radiation, is the binding constraint: a vacuum cannot carry waste heat away, so radiators and heat pipes add mass and limit duty cycles, directly inflating per-launch cost.
- Google's own analysis says orbital compute reaches cost parity with terrestrial data centers only if launch falls to $200 per kilogram, a level it projects for the mid-2030s. Current Falcon 9 rideshare pricing runs near $3,600 per kilogram per Futurum Research.
- Startup Starcloud already flew an NVIDIA H100 in November 2025, has since trained a model in orbit, and has raised $450 million at a $2.3 billion valuation. Its backers include NVIDIA and Cisco Investments.
- SpaceX is simultaneously Google's only launch provider and a direct competitor: SpaceX has filed for a constellation of up to 1 million satellites and, after acquiring xAI, has its own orbital compute ambitions targeting 2028.
- Europe is a marginal player: the European Space Policy Institute counted under 70 million euros in private capital for space data center ventures since 2020, less than Starcloud's August funding round alone.
What to watch next
- In-orbit radiation and thermal data collected over the next several weeks will either validate or complicate Google's ground-test results, especially for high-bandwidth memory.
- The 2027 two-satellite laser-link test is the next concrete milestone: success would prove the inter-satellite networking needed for swarm-scale AI workloads.
- SpaceX Starship commercial pricing is the economic hinge: Starcloud's Philip Johnston argues cost competitiveness kicks in at $500 per kilogram, which he expects by 2028 or 2029. Watch for Starship launch cadence and pricing announcements as the real signal for when orbital compute becomes viable.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.