Google's first Project Suncatcher satellite is preparing to take AI computing into orbit, carrying four TPU accelerators in an experiment that could help determine whether future AI infrastructure can operate in space.

Google is preparing to launch AI hardware into orbit as part of Project Suncatcher, its ambitious research effort exploring whether large-scale machine-learning infrastructure could eventually operate in space.

The company's first experimental satellite is scheduled to launch on October 1, marking the first orbital hardware test for the project.

The refrigerator-sized prototype, known internally as MVP, will carry four of Google's Tensor Processing Units, or TPUs, alongside the hardware required to power and cool them in orbit.

Google intends to run Gemini models on the processors during the mission.

But this isn't a miniature replacement for one of Google's terrestrial data centers. At least not yet.

The mission is designed to answer a more fundamental question: Can the same class of AI hardware used inside data centers survive and operate effectively in space?

Four TPUs are going to orbit

Google has already conducted extensive ground testing, but space introduces conditions that are difficult to reproduce perfectly on Earth.

The TPUs will have to survive the intense vibration and acceleration of launch before dealing with radiation and extreme thermal conditions in orbit.

Radiation is particularly problematic for conventional computing hardware because energetic particles can interfere with electronics and potentially cause errors in calculations.

Project Suncatcher will allow Google's engineers to observe how its AI accelerators behave under those conditions.

The company is also investigating how multiple computing satellites could eventually communicate using high-bandwidth optical links.

If that technology scales, future satellites could potentially operate together as enormous distributed AI computing systems.

Cooling might be the biggest problem

There is an obvious irony to putting power-hungry AI chips into the freezing environment of space:

They're still difficult to cool.

Space may be extremely cold, but the absence of an atmosphere means conventional cooling techniques don't work.

On Earth, data centers can move enormous amounts of air or water around computing equipment. In a vacuum, heat ultimately needs to be transferred away and radiated into space.

Google's experimental system uses thermal interface material connected to aluminum and copper heat pipes, which move heat toward a radiator.

Even then, the first system won't be capable of continuously running its four TPUs.

According to details released about the project, Gemini workloads will run for approximately 15 minutes at a time before the processors are shut down temporarily so the cooling system can catch up.

That limitation demonstrates how early Project Suncatcher remains.

The satellite also has roughly 1,000 watts of solar power available, putting its computing capabilities much closer to an experimental server than an orbital hyper scale data center.

For now, Google wants data.

Why put AI data centers in space?

The idea sounds excessive until you consider what is happening to AI infrastructure on Earth.

Training and operating increasingly large AI systems requires enormous amounts of electricity.

Data centers also require land, cooling infrastructure and connections to increasingly strained electrical grids.

Space offers one extremely attractive resource: sunlight.

Solar panels in the right orbit can receive significantly more consistent sunlight than terrestrial installations, potentially allowing future computing infrastructure to generate large amounts of power without competing directly with electrical grids on Earth.

A sufficiently advanced orbital computing network could theoretically use solar energy to power AI workloads before transmitting results back to Earth.

But enormous engineering problems remain.

Hardware needs to survive radiation. Cooling systems must dispose of heat without an atmosphere. Launching thousands of tonnes of computing equipment into orbit would be extraordinarily expensive.

And then there is networking.

Modern AI clusters rely on extremely fast connections between accelerators. Reproducing that kind of bandwidth across satellites requires sophisticated optical communication systems.

Project Suncatcher is Google's attempt to find out whether those problems are solvable.

This is only the beginning

Google isn't pretending that October's satellite represents an orbital data center.

It is an experiment.

The company says future missions will test increasingly capable configurations, with additional Project Suncatcher launches targeted for 2027.

The October mission will provide engineers with real-world information about radiation, thermal management, power delivery and TPU reliability in orbit.

Those results will determine what comes next.

For now, putting four AI chips into space might sound insignificant compared with Google's enormous terrestrial computing infrastructure.

But if Project Suncatcher works, those four processors could represent the first hardware test of something considerably larger:

AI infrastructure that no longer needs to stay on Earth.