Google Launches TPUs Into Orbit: The Cloud Gets Very, Very Literal

Google Launches TPUs Into Orbit: The Cloud Gets Very, Very Literal

For years, "the cloud" was a marketing metaphor for someone else's computer in a warehouse in Virginia. Google has apparently decided that metaphor was underachieving. On October 1, it sent actual AI chips to space.

Four TPUs, One Fridge-Sized Satellite

As part of Project Suncatcher, Google put four of its Trillium-generation TPUs into orbit aboard a prototype satellite riding on SpaceX's Transporter-18 rideshare mission from Vandenberg Space Force Base. The refrigerator-sized spacecraft, built with Earth-imaging company Planet and modestly named MVP, runs on solar arrays rated at about 1 kilowatt, roughly what a gaming PC pulls under load.

Before launch, Google bombarded Trillium chips with a proton beam at UC Davis's Crocker Nuclear Laboratory and says they can survive a radiation dose greater than what a five-year space mission would deliver. The flight will test how the chips handle launch vibration, radiation during live AI workloads, and cooling in a vacuum, where the only way to dump heat is radiating it away.

A Hedge on Power, Not a Moonshot (Yet)

One kilowatt won't train a frontier model, and Google isn't pretending otherwise. It describes the mission as a way to gather in-orbit data and find failure points before anything bigger gets built. The real motivation is earthly: AI data centers are hungry for power, and the sun in orbit doesn't take nights off or file grid interconnection paperwork.

The longer roadmap gets ambitious fast, with a two-satellite laser-link test planned for 2027 and a concept for an 81-satellite cluster flying in tight formation. That vision depends on launch costs falling to around $200 per kilogram, so don't expect your next chatbot query to bounce off low Earth orbit just yet. Still, every wild infrastructure idea starts with a weird little prototype, and this one is weird in exactly the right way.

Step one of space data centers isn't scale. It's proving the chips don't fry on the way up.

Your AI workloads may not need orbital power just yet, but if you're figuring out where AI actually fits in your business here on the ground, I'm happy to talk it through.

Source: Hardware Busters