Four chips. One kilowatt of power. Six years in orbit before burning up on re-entry. Google’s first step toward putting AI infrastructure in space is, by any measure, extremely modest. But the questions it’s trying to answer are not.
According to Engadget, Google is sending a small satellite called MVP aboard a SpaceX Falcon 9 rocket on October 1 as part of its Project Suncatcher initiative. The goal of that project, long term, is to place large-scale AI compute infrastructure in low-Earth orbit, powered by solar energy. The short-term goal is far simpler: find out if the hardware survives up there at all.
The satellite carries four tensor processing units, Google’s own AI chips, which together approximate the output of a single data center server. Solar panels supply one kilowatt of electricity to run them. For context, that’s roughly what a hair dryer pulls. Google plans to have the chips answer basic AI queries for about a year while the satellite orbits Earth for six years total before atmospheric re-entry destroys it.
The engineering challenges here are real and worth taking seriously. Radiation is the obvious threat. In space, cosmic rays and high-energy particles can cause “bit flips,” where binary values flip from one to zero or back, corrupting data and destabilizing chip behavior. Google tested the TPUs at Crocker Nuclear Laboratory by blasting them with radiation to simulate space conditions. The results were apparently good enough to justify the launch. But the fix for bit flips in orbit is essentially the same as fixing a frozen Windows PC: restart the chips and hope for the best.
Cooling is the other hard problem. Google built a proprietary system that uses conductive materials to push heat out into space, but it only works for about 15 minutes before the chips need to shut down to cool off. That might be manageable with four chips generating minimal heat. It becomes a very different problem at scale. Standard convection-based cooling, the kind used in every data center on Earth, does not work in a vacuum. There’s no air to move.
The people actually building this are clear-eyed about the timeline. James Manyika, Google’s Senior VP for Research, Technology and Society, said directly: “We don’t expect, to be perfectly frank, that we’ll have anything usefully operational in the next few years.” Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon, told The New York Times that scaling from one satellite to a functional network would take years and enormous capital, and would introduce entirely new engineering problems along the way.
That’s worth keeping in mind when evaluating the broader hype around space-based AI compute. Elon Musk has suggested this kind of infrastructure would be straightforward to build. The actual physics and economics suggest otherwise. A single Falcon 9 launch runs around $74 million. Cosmic rays can physically destroy semiconductor junctions. And the cooling system that barely handles four chips would need a complete rethink to support anything resembling a real data center workload.
Still, Google is pushing forward. The plan is to launch two more similar satellites in 2025, with an eventual goal of 80 satellites flying in close formation to collectively handle AI queries. The company is also exploring a much larger satellite, roughly the size of a football field. No budget figures have been disclosed.
What makes this worth watching is less the near-term capability and more what Google is actually testing. Power availability, radiation tolerance, thermal management, and chip reliability in orbit are all open questions. If MVP answers even some of them, it gives Google data that no competitor currently has. Microsoft, Amazon, and the major cloud players are all racing to build AI infrastructure, but none of them are testing it in space. That’s either a smart differentiator or an expensive detour. This satellite is how Google finds out which one it is.
