Google's AI Just Answered From Orbit

The first consequence is a planning horizon: if compute can run in orbit, the map of where your future capacity sits — and what constrains it — gets redrawn.

Google's Project Suncatcher, reported in this magazine's October 31 Morning edition, put four Trillium TPUs in orbit — running one at a time — and got an AI response back: a haiku about space, generated by Google's Gemma 4 model aboard the satellite and downlinked over Svalbard, Norway. As engineering, it is a proof. As strategy, it is a flare: the two costs that dominate every terrestrial data center — power and cooling — are exactly what orbit promises to change.

Nobody is migrating your workloads to space this budget cycle, and the honest limits are real: radiation, maintenance, bandwidth and the small matter of launching the hardware. But every hyperscaler roadmap is now a power roadmap. Google exploring orbit is the same fact as your vendor exploring new substations — the constraint moved from chips to energy.

For the manager, the lesson is nearer than it looks. Your capacity planning already prices electricity and cooling; the largest buyer on earth just signaled those two lines now drive architecture itself.