TL;DR
Google has confirmed its first TPU flew to orbit on a SpaceX rideshare mission, marking the opening move in a long-term bet that low Earth orbit could host scalable AI compute infrastructure.
What happened
- Google's Trillium TPU reached orbit aboard a Planet Labs prototype satellite on SpaceX's Transporter-18 rideshare mission from California.
- The chip will run in 15-minute bursts to protect power and thermal systems, drawing up to 1 kilowatt of continuous power when active.
- Hardware survived pre-launch stress tests including 50 to 100 g vibration forces and proton-beam radiation testing at UC Davis's Crocker Nuclear Laboratory.
- Radiation results show the Trillium TPU can survive a total ionizing dose exceeding a five-year space mission, making it viable for large inference workloads in orbit.
- Google published a peer-reviewed white paper (to appear in Joule) projecting SpaceX launch costs near $200 per kilogram by 2035, contingent on roughly 1,800 Starship launches.
Why it matters
- Orbital solar advantage is the core thesis: LEO satellites can harvest up to eight times more solar power than ground installations, potentially rewriting the energy economics of AI compute.
- A follow-up demo in 2027 will fly two purpose-built satellites linked by laser, the first step toward handling heavier AI workloads without ground infrastructure.
- The endgame is an 81-satellite close-formation data center, a scale that requires launch economics that do not yet exist, making the 2035 Starship cost projection a critical dependency.
- Google explicitly flags that orbital TPUs remain unsuited to massive training runs, positioning this as an inference and edge-compute play, not a training replacement.
- Planet Labs gains a credible role as a space AI hardware integrator, a meaningful expansion beyond its Earth-observation origins.
What to watch next
- In-orbit TPU performance data from this mission will determine whether the 2027 dual-satellite laser-linked demo proceeds on schedule.
- SpaceX Starship launch cadence and pricing trajectory: the $200 per kilogram target by 2035 requires roughly 1,800 Starship launches, a number that will be trackable year by year.
- Competitor moves: whether Amazon (with its Kuiper constellation) or other hyperscalers announce comparable orbital compute experiments in response to Google going public with Project Suncatcher.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.