Google Launches First TPU AI Chip into Orbit, Outlines Space Data Center Economics

Google launched its first prototype orbital compute satellite into space aboard a SpaceX rocket from California, marking the company's first deployment of its custom artificial intelligence silicon into orbit, according to reporting by TechCrunch AI (https://techcrunch.com/2026/10/01/google-thinks-spacexs-starship-has-to-launch-1600-times-before-space-data-centers-get-off-the-ground/).

Manufactured by Planet Labs, the demonstration satellite is designed to test whether Google's Tensor Processing Unit (TPU)—an accelerator built to compete with Nvidia GPUs—can operate reliably in space. The operational test involves delivering a continuous kilowatt of power, regulating thermal thresholds, and running AI models. Once operational, the satellite will fire its TPU in 15-minute bursts to prevent excessive strain on thermal and power systems.

The mission is the initial phase of Project Suncatcher, a long-term research program managed by Google executive Travis Beals. The SpaceX launch carried over one hundred payloads, including other space AI experiments from startups such as Satlyt and Cowboy Space Company. Google and Planet Labs are preparing a subsequent test flight scheduled for next year with two purpose-built satellites designed to communicate via an optical laser link, building toward a conceptual cluster of 81 satellites operating in close formation.

Alongside the launch, Google released a peer-reviewed study set for publication in the journal Joule evaluating orbital compute logistics. The paper outlines the launch scale required to reduce orbital transport costs to around $200 per kilogram by 2035, based on a historical 20 percent annual cost-reduction curve observed since SpaceX's Falcon 1.

According to the paper's projections, achieving that price threshold would require SpaceX's Starship to haul roughly 370,000 metric tons of payload to orbit over the next decade. Assuming a capacity of 200 metric tons per mission, that volume equates to approximately 1,800 Starship launches across ten years, or 180 flights annually. To date, Starship has flown no more than five times in any single calendar year.

The study also detailed radiation stress tests conducted by exposing TPUs to particle accelerator beams. While realistic shielding scenarios generated minor logic circuit errors, researchers estimated typical inference error rates at around one in a million, sufficient for standard inference tasks across a satellite's five-year lifespan, though impractical for massive multi-month model training across thousands of chips.