Google is sending TPUs into orbit

While ground-based data centers run out of grid power and cooling water, Google is looking toward orbit.

Google has revealed technical details behind Project Suncatcher, an ambitious moonshot exploring whether space can host scalable machine learning infrastructure. Next week, a prototype satellite built with Planet will ride aboard SpaceX’s Transporter-18 mission to test how Tensor Processing Units (TPUs) survive and operate in low Earth orbit.

Surviving 100g shocks and cosmic rays

Getting silicon into orbit begins with surviving the violent ride up.

During a ten-minute launch into low Earth orbit, the spacecraft experiences continuous acceleration up to 10g. Individual components like TPU chips can face localized forces between 50 and 100g. Google put the satellite through extreme three-axis vibration testing to simulate rocket frequencies, and the hardware managed to stay intact.

Once in orbit, radiation poses an ongoing threat. Solar events and cosmic rays easily cause bitflips and corrupt AI workloads.

Before heading to the launchpad, Google brought its Trillium TPUs to the proton beam facility at UC Davis's Crocker Nuclear Laboratory. Engineers ran active AI workloads while blasting the chips with protons. Early results showed the Trillium silicon withstood total ionizing doses exceeding what it would encounter across a five-year mission in space.

Zero airflow cooling and laser targets

Surviving radiation is only half the battle. In the vacuum of space, you cannot simply blow air across a heatsink.

TPUs generate intense heat within tiny footprints. Without atmospheric convection, heat dissipation must rely entirely on specialized heat pipes and radiators venting warmth into space. Google tested this system in a thermal vacuum chamber, but next week's flight will provide the first real-world orbital telemetry.

If the hardware holds, the long-term vision calls for constellations where individual satellites carry dozens of TPUs. Tying those floating chips into a unified cluster requires ultra-precise laser links. Google notes the challenge requires maintaining high-bandwidth connections over very short distances while both satellites move, roughly equivalent to hitting a coin-sized target from miles away. A two-satellite demo dedicated to testing those optical links is slated for 2027.

Why it matters

Orbit offers an obvious advantage: near-constant sunlight generates up to eight times more solar power than identical arrays on Earth.

We think Project Suncatcher is a classic Google moonshot that makes strategic sense. Terrestrial AI data centers are already colliding with local power grids, water constraints, and regulatory hurdles. Exploring an environment with unlimited solar power removes one of machine learning's biggest long-term physical ceilings.

For developers and enterprise customers, there is nothing actionable today. Orbital compute remains far from offering usable API endpoints. But by tackling fundamental physics first—vibration survival, vacuum cooling, and radiation tolerance—Google is quietly mapping out the engineering playbook before the competition even starts looking upward.