Earlier this month, Google sent its Project Suncatcher prototype into orbit aboard a SpaceX Falcon 9 launch shared with more than 100 other spacecraft. The satellite carries four Tensor Processing Units, though its roughly 1-kilowatt power supply means it can run only one at a time—and for about 20 minutes per cycle.¹

After roughly three weeks of commissioning, engineers sent a command from Earth, powered up a TPU and asked the onboard Gemma 4 model for a haiku about space. Back came: “Blue marble below / Data streams in silent grace / Starlight feeds my core.”Sundar Pichai (sundarpichai)² Google said it had also run space-themed prompts and telemetry tests, calling the successful exchange a first response from its AI chip-equipped satellite.³

That is a genuine “hello world” moment, as Google chief executive Sundar Pichai put it—not proof that orbital data centers are ready for business. The company’s next phase is to test whether the TPUs can withstand the thermal stress and radiation exposure of orbit. Radiation-induced “bit flips” can silently corrupt stored data, a risk Google had previously modeled on the ground but must now measure in the real environment.¹

Google’s longer bet is more ambitious: two interconnected satellites in 2027, with more power and optical links intended to support distributed workloads. That vision differs from companies focused on standalone-satellite inference, but it also depends on reliable hardware, workable heat and power management, and dramatically cheaper, rapidly reusable heavy-lift launches.¹

Musk, whose SpaceX would be central to any such launch economics, boosted posts declaring that Google had shown space AI data centers “actually work.”Elon Musk (elonmusk)⁴ The narrower conclusion is sturdier: one TPU in orbit answered one prompt. Now Google has to show it can keep answering.