Google Put TPUs in Orbit: Project Suncatcher's First Satellite Is Flying

Google’s attempt to move AI compute off the planet left the ground on October 1: a Project Suncatcher prototype satellite carrying Trillium TPUs launched aboard SpaceX’s Transporter-18 rideshare mission, developed with satellite operator Planet, per Google’s research announcement and Mashable’s coverage. The goal of this first flight is narrow and honest: find out what breaks. Travis Beals, Google’s Senior Director of Paradigms of Intelligence, framed it as a moonshot: “This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions.” The destination is orbital data centers, at a moment when Google’s SpaceX relationship already runs to billion-dollar compute contracts and the constraints of terrestrial buildouts, power, cooling, and land, have hard physical ceilings.

The Three Physics Problems, and the Answers So Far

Getting a data center to orbit means surviving launch, radiation, and vacuum, and Google has tested each. Launch stress: roughly ten minutes to orbit at up to 10g acceleration, with individual components seeing 50 to 100g; three-axis vibration testing showed the TPU hardware held. Radiation: Trillium TPUs were bombarded at UC Davis’s Crocker Nuclear Laboratory proton beam facility while actively running AI workloads, and survived a total ionizing dose greater than a five-year mission would deliver. Cooling is the unsolved one: vacuum means no airflow, so heat leaves only by radiation, and Google is testing heat pipes plus radiators in thermal vacuum chambers. That last problem is why this prototype matters; a data center is a heat factory, and orbital thermodynamics is the difference between a demo and infrastructure.

The 2027 Test Is the Real Milestone

One satellite in orbit proves chips survive; a data center needs chips talking to each other. In 2027, Google plans to launch two satellites carrying dozens of TPUs each, positioned to test high-bandwidth laser inter-satellite links at short distances, with pointing precision Google compares to “hitting a coin-size target from miles away while both points are in motion.” That is the actual feasibility question for orbital compute, because distributed training requires the kind of bandwidth between accelerators that on Earth runs over optical cables measured in centimeters. If the laser links deliver, the payoff Google projects is stark: future constellations each carrying dozens of TPUs could harvest up to 8 times more solar power than equivalent ground installations, running on uninterrupted sunlight instead of a grid.

What This Buys Google Strategically

The economics of this decade’s AI buildout are land, power, and permits, and Google is the most constrained of the frontier labs on all three. Orbital compute is a long-dated option against all of them at once: solar power without a utility bill, cooling radiated to space, and scaling that adds satellites rather than substations. It also fits the existing Google-SpaceX relationship rather than requiring new infrastructure. The realistic assessment: nobody is training frontier models in orbit this decade, but the compounding value of having proven the supply chain, the radiation hardening, and the interconnects while competitors treat compute as a ground-only problem is exactly the kind of option that looks cheap now and decisive in ten years.

What to Watch

Three checkpoints. First, telemetry from the prototype: whether the TPUs power on, run workloads, and hold calibration in orbit, expected in the coming weeks. Second, the 2027 pair: laser interconnect bandwidth between moving satellites is the make-or-break number for the whole thesis. Third, the competitive response, because if orbital compute trends toward viable, the land-and-power bottleneck that defines frontier-lab strategy stops being a moat and starts being a self-imposed constraint, and the SpaceX compute ecosystem suddenly has a second product to sell.