[UPDATE] Google's AI Chips Are Now in Orbit

Google has put its first artificial intelligence chips into space. On October 1 a refrigerator-sized prototype satellite carrying four of the company's Trillium-generation tensor processing units reached low Earth orbit aboard a SpaceX Falcon 9, riding up on the shared Transporter-18 mission from Vandenberg Space Force Base. It is the first real-world test of Project Suncatcher, Google's plan to one day run machine-learning hardware on sunlight above the atmosphere rather than on the grid below it.

We reported last month that the launch was imminent. Now the hardware is up, and the experiment can begin.

The satellite, built with the Earth-imaging company Planet and known internally as the MVP, is a deliberately humble machine. Its solar panels supply roughly a kilowatt of power, and the four chips run short bursts of Gemini queries in windows of about 15 minutes before pausing to shed heat. All told it offers the compute of a single Google Cloud TPU v6e-4 slice, a rounding error next to a terrestrial data center. The point was never raw power.

What the test is really for

The point is survival. Google wants to know how its own silicon holds up against three conditions it cannot recreate together on the ground: the violence of launch, the radiation of orbit, and the swing between heat and cold as the satellite passes in and out of sunlight. Data sheets and vacuum chambers can approximate each in isolation. Only orbit delivers all three at once.

If the chips endure, the longer-term vision is a constellation of satellites flying in tight formation and trading data over free-space optical links, drawing on near-constant sunlight in the right orbit. That is years away and may never pencil out. For now the prototype is designed to work for about a year and to fall back into the atmosphere within six.

The orbital race widens

Google is not alone in looking up. The appeal of space-based compute is simple to state and very hard to deliver: abundant solar energy and a cold sink for waste heat, set against the cost of launch and the impossibility of sending a technician to swap a failed board. By flying real inference silicon rather than a mock-up, Google has moved the question from the whiteboard to the telemetry feed.

The data coming back over the next year will settle a lot of arguments. Either Trillium shrugs off orbit, and the economics of a solar-powered constellation start to look worth modeling in earnest, or it does not, and Suncatcher becomes a well-documented lesson in why the grid still wins. Google has at least stopped speculating and started measuring.