Google plans satellite launch to test AI data center in space

Google just revealed one of its most audacious infrastructure bets yet: putting AI chips in space and powering them with the sun. The initiative, called Project Suncatcher, envisions constellations of solar-powered satellites running Google’s custom Tensor Processing Units in low-Earth orbit, effectively turning space itself into a data center.

The company announced the project on November 4, 2025, framing it as a long-term answer to one of AI’s most stubborn bottlenecks. Training and running large machine learning models requires enormous amounts of energy, and terrestrial data centers are increasingly bumping up against power grid limitations.

What Project Suncatcher actually involves

The plan starts modestly, at least by space standards. Google has partnered with Planet Labs, the San Francisco-based satellite imaging company, to build and operate two prototype satellites. Each one will carry Google’s TPUs into a dawn-dusk sun-synchronous orbit, a trajectory that keeps the spacecraft in near-constant sunlight.

That orbital sweet spot is the whole thesis. Google claims satellites in this configuration could achieve up to eight times the solar panel productivity compared to ground-based solar installations.

The prototypes are scheduled for launch by early 2027. Their job is not to process production AI workloads but to serve as a “learning mission” that stress-tests TPU performance in the harsh conditions of space. The checklist includes thermal management in vacuum, radiation resilience, and high-bandwidth optical inter-satellite communication links, the kind of laser-based data transfer that would eventually let a constellation of these satellites work together as a unified compute cluster.

Planet Labs will handle satellite design and construction, integrating Google’s hardware into the company’s Owl satellite bus roadmap. Google parent Alphabet already holds an equity stake in Planet Labs.

The radiation problem, mostly solved

Google says it has already conducted radiation testing on its Trillium TPUs, the latest generation of its custom AI accelerators, and found them resilient enough to handle radiation doses equivalent to a five-year mission.

The most vulnerable component turned out to be High Bandwidth Memory, the stacked memory modules that feed data to the processing cores at speed. That’s not surprising. HBM is densely packed and operates at tight voltage margins, making it more susceptible to radiation-induced errors. But Google’s testing suggests the issue is manageable within the mission’s expected exposure window.

The economics of computing in space

Google projects that by the mid-2030s, the economics of launching and operating orbital compute infrastructure could reach rough parity with terrestrial energy costs. The magic number appears to be launch prices dropping below $200 per kilogram.

CEO Sundar Pichai has described Project Suncatcher as a foundational step toward eventual orbital data centers, not a product announcement. The full vision, fleets of satellites processing AI workloads with effectively unlimited solar power, is a mid-2030s aspiration at the earliest.