I remember standing under a launch livestream, the feed jittering while the rocket climbed and every sensor screamed. You feel the tiny, human parts of a machine become fragile in that noise. I want to tell you why Google sending AI chips into orbit next week feels less like a stunt and more like a test of what reality will tolerate.
A prototype satellite will ride on SpaceX’s Transporter-18 next week.
Google’s Project Suncatcher is sending its first satellite into low Earth orbit to see how four Tensor Processing Units (TPUs) behave off-planet. The cubesat-class payload, developed with Planet Labs, will hitch a ride on SpaceX’s Transporter-18 rideshare and collect telemetry on physical stresses, radiation, and operational errors while running real AI workloads.
I read Google’s post and the company was blunt: this launch is about finding what breaks and fixing it. You’ll see the same bluntness in the telemetry they publish — numbers, not promises.
The ascent squeezes hardware with intense vibration and g-forces.
The rocket ride is roughly 10 minutes of extreme stress. During that climb the spacecraft experiences sustained forces up to 10 g, and some components can briefly see 50–100 g.
The physical shock is one challenge; the other is radiation. Once in orbit, solar events and cosmic rays can flip bits or damage circuitry. Google already ran proton-beam trials at UC Davis’s Crocker Nuclear Laboratory on its Trillium TPUs while they were running models. The result: the chips survived a radiation dose larger than what a five-year mission would normally deliver, and the team mapped which errors matter to real workloads.
Thermal management is equally tricky in vacuum. I won’t pretend the solution is simple. Google is testing heat pipes and radiators and has already used a thermal vacuum chamber to mimic orbital conditions. The ascent itself, meanwhile, feels like a pressure cooker for hardware — intense, fast, and unforgiving.
How will Google cool AI chips in space?
Google plans a mix of passive and active systems: heat pipes to move heat away from the chips and radiators to dump that heat into space. The team validated prototypes in a thermal vacuum chamber and will compare chamber data with live-orbit performance to refine designs.
Satellites could harvest near-constant sunlight and trade models with lasers.
Low Earth orbit offers long stretches of sunlight that terrestrial data centers cannot match; Google says satellites could generate up to eight times the solar power available on Earth at some latitudes. That surplus changes the economics of running large AI models outside the atmosphere.
The second technical pillar is communications: Google plans high-bandwidth lasers to shuttle data between satellites and to ground stations. Imagine a fleet not as isolated boxes but as a highway of light where models and weights move across nodes.
Other players are testing similar ideas. Starcloud launched an Nvidia H100 GPU into orbit in November 2025, and SpaceX has been selling the concept of space-based AI infrastructure as part of its broader strategy. Google’s plan is incremental: test chips, then test inter-satellite lasers in planned 2027 missions with two satellites.
Why put AI chips in orbit instead of on Earth?
Power density and cooling are the blunt answers: more sun means more energy per kilogram in orbit, and moving heat into space can be simpler than fighting ambient heat on a hot server floor. There’s also a political and environmental angle — shifting some compute off-planet could reduce local energy demand and the visible footprint of sprawling data centers.
This mission will map failures so future designs can change.
Google said the first flight is exploratory: gather telemetry, identify failure modes, and iterate. Data from launch vibration, radiation hits, thermal behavior, and run-time errors will guide hardware choices for the next flights.
I’ve followed projects where early failures rewrote the design playbook. You should expect misfires, partial failures, and surprising resilience — that’s the point. Google’s prior lab tests gave them confidence that Trillium TPUs can survive heavy radiation doses, but the gap between a chamber and true orbit is why this flight matters.
What happens to electronics in space radiation?
Radiation can change bits, corrupt memory, and physically degrade components over time. Engineers mitigate these risks with shielding, error-correcting codes (ECC), software-level retries, and by characterizing which failures are tolerable for a given workload.
Planet Labs, SpaceX, Nvidia, UC Davis, and others are all part of this fast-moving experiment in making compute mobile beyond Earth. I’m watching the telemetry you will see, and I’ll tell you the parts that matter. How much of our compute future are you willing to move off the planet?