Earlier this month, Google sent its Project Suncatcher prototype into orbit aboard a SpaceX Falcon 9 mission, alongside more than 100 other spacecraft. The satellite carries four Tensor Processing Units, though its roughly 1 kW power supply can run only one TPU for about 20 minutes at a time—a sharp reminder that orbital computing remains constrained by energy as much as ambition.¹

After roughly three weeks of commissioning its subsystems, engineers sent up a simple first task: ask the pre-loaded Gemma 4 model to write a haiku about space. The reply came back: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Google also used the spacecraft for telemetry requests and other space-themed prompts; a spokesperson said it was operating nominally.¹

Sundar Pichai framed the exchange as a breakthrough in miniature. “We’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth,” he wrote, calling it a “hello world” milestone.Sundar Pichai (sundarpichai)²

The more exuberant interpretation arrived through Elon Musk’s repost of a claim that Google had “proved AI data centers in space actually work.”Elon Musk (elonmusk)³ That stretches the result. Google has demonstrated that an inference chip can switch on, process a prompt and transmit an answer. It has not demonstrated persistent, networked, economical computing at data-center scale.

The hard phase now begins: testing the TPUs against orbital heat and radiation, including memory-corrupting “bit flips” that may be especially consequential for more demanding workloads. Google says reliable execution in space would lead to twin interconnected satellites in 2027.⁴ Those spacecraft are intended to bring more power and optical links—the ingredients needed to test whether this poetic first contact can become industrial infrastructure.