Project Suncatcher, Google's effort to run AI compute in space, faces its first hardware test. The satellite, called MVP, goes into orbit on 1 October 2026 aboard a SpaceX Falcon 9, sharing the ride with other small payloads.

It looks nothing like the facility the name suggests: the satellite is about the size of a refrigerator and carries four of Google's own TPU chips. The power it draws from its solar panels is roughly 1 kilowatt, about what a hairdryer uses. Planet Labs built the satellite; Google supplied the chips.

Why space

The reason is energy. Electricity is the largest running cost of a data centre, and new sites queue for grid capacity. Google's alternative is simple: move the compute to where the sun never sets. The long-term picture the company sketched is orbital clusters reaching cost parity with ground facilities by the mid-2030s.

What the test measures

This flight is not an attempt to build a data centre; it is an endurance test. It measures:

  • Whether the chips survive launch vibration and forces up to ten times Earth's gravity.
  • Resistance to radiation from cosmic rays and solar events.
  • Whether heat pipes and radiators can keep the hardware cool in orbit.

Google had already exposed Trillium TPUs to a proton beam at the Crocker Nuclear Laboratory at UC Davis. According to Travis Beals of Google, the chips "performed remarkably well" despite doses beyond what five years in space would deliver.

The known limit: 15 minutes

The project's most concrete constraint is already public. The prototype can only run workloads in 15-minute bursts before it needs to cool down, and Google acknowledges this has to be solved. Cooling in space is harder than on the ground because heat can only be shed by radiation, not into air.

The company plans to launch two more satellites in 2027, this time to test high-speed laser links. In the longer-term design, satellites form clusters of 81 in kilometre-scale arrays.

Not alone

Others are chasing the same idea: SpaceX, Starcloud and Orbital have announced similar plans. But this launch is the first moment the claim leaves the slide deck and becomes a measurable test.

The scale is worth keeping in mind while reading: four TPUs amount to the power of a single server in a ground data centre. This flight does not measure whether AI can run in space, but whether the chips survive there. The harder questions, cooling and moving data, come after.