Google booted Tensor Processing Units on its Project Suncatcher satellite and generated a haiku from orbit—the first inference run on the experimental platform, according to a company blog post Friday.
The satellite, launched this month, was also tested with space-themed prompts and telemetry queries. These tests aim to answer a single question: can AI hardware survive and perform reliably in the thermal stress and radiation of low Earth orbit?
That answer matters to Google's capital allocation calculus. Ground-based data centers consume massive power for cooling and require expensive real estate near fiber routes. Orbiting compute could reduce latency for edge applications and potentially bypass terrestrial grid constraints—though launch costs, satellite redundancy, and operational complexity remain brutal obstacles. Google is betting on a 2027 twin-satellite launch, committing years and hundreds of millions to verify the concept works at scale.
The real test begins now: whether TPUs can execute sustained workloads in space without degradation. If Google succeeds, orbital infrastructure becomes a credible third pillar alongside hyperscale data centers and on-device AI. If margins don't pencil out, it's an expensive R&D write-off. That 2027 date is when investors will know which story is true.
