Google has completed the first live test of space-based machine-learning compute, booting one of its data-centre TPU chips aboard a prototype satellite and running a small language model in orbit. The effort, Project Suncatcher, launched its test satellite on October 1. This week the company confirmed the chip survived launch, powered up in orbit and returned a haiku generated by a Gemma 4 model, a deliberately modest output from an immodest idea.
Project Suncatcher: Why Put Data Centres in Space
The logic is energy. AI data centres on Earth are straining power grids and provoking local opposition, while in orbit, solar panels harvest sunlight almost continuously with no cooling water and no land. Google is not alone in the thought: the same week brought news of a startup claiming a billion dollars of orders for modular AI data-centre pods promising deployment in 48 hours, evidence of how desperately the industry is hunting for capacity anywhere it can find it, as we noted in our recent science coverage of the crowded skies.
Enormous Problems Remain
Radiation corrupts memory, heat cannot be radiated away easily, hardware cannot be repaired, and launch costs dwarf the price of the chips themselves. A single TPU running a haiku is a proof of concept, not an economics argument. The daily AI briefing treated Suncatcher as one item among many in a packed week, which is the right scale: Google itself frames orbital compute as a moonshot research direction, in the tradition of projects that take a decade to matter, if they ever do.
What the Test Actually Proves
What Suncatcher establishes is narrower and still valuable: that modern AI accelerators can function in the radiation and thermal environment of low Earth orbit, and that the software stack, from boot sequence to model inference, works end to end through a satellite link. Each future step, bigger chips, clustered satellites, optical interlinks, now has a tested baseline rather than a slide deck.
Space computing will not relieve this year’s capacity crunch, and it may never beat a well-built terrestrial data centre on cost. But the history of computing suggests that once a capability is demonstrated, economics have a habit of following. Google has put a chip in orbit and brought back a poem. The industry will take it from there.
