Google runs a data-centre-class TPU in orbit
Google announced on 30 October 2026 that its Project Suncatcher prototype satellite had generated its first model output in orbit. The satellite, built with Planet and launched on 1 October 2026, carries four Trillium-generation TPUs with the open-weight Gemma 4 model preloaded. During a ground-station pass over Svalbard, operators uplinked a prompt asking for a poem about being in space and the model's response was transmitted back. Travis Beals of Google Research described it as the mission's "Hello World" and as the first AI response from a data-centre-class TPU in orbit. Google said subsequent tests will measure the TPUs under thermal stress and radiation exposure.
Why It Mattered
The output itself is trivial; the location is not. Through 2026 the binding constraint on frontier AI shifted from model design to the physical siting of compute — electricity supply, grid interconnection, cooling water and the capital cost of new data centres. Project Suncatcher is the first working demonstration of the alternative that had been discussed in proposals and papers but never flown: putting accelerators in orbit, where solar power is continuous and heat is rejected by radiation rather than by water. This test established only that the hardware boots, survives launch, runs inference and returns results over a ground link. It does not establish that accelerators can endure radiation over months, that constellations can be interconnected with enough bandwidth to train or serve at scale, or that any of it is economic. Those are the questions the follow-on thermal and radiation tests are meant to address. The entry belongs in a history of AI infrastructure as the commissioning milestone that moved orbital compute from proposal to flown hardware, and as a dated reference point for claims made later in the decade about space-based data centres. It also marks a convergence of two industries: a model developer specifying launch hardware, and an Earth-imaging satellite operator building the bus for it. Whether this becomes a line of development or a one-off experiment will be settled by the radiation results, not by the first response returned.
Who Built It
Google (Project Suncatcher, with Planet)
Applications
- AI Infrastructure
- Space Systems
- Edge Inference