Google's first orbital data center satellite is online and running inference in space

Google's first data-center satellite has completed commissioning and is now doing real compute work in orbit, according to TechCrunch. The prototype spacecraft, part of the company's Suncatcher effort, launched earlier this month with more than 100 other spacecraft on a SpaceX Falcon 9. After roughly three weeks of subsystem checks, ground engineers sent it its first prompt this week.

That first job was more symbolic than taxing: they asked Gemma 4, the Google DeepMind model pre-loaded on the satellite, to write a haiku about being in space. The model obliged with a short verse about looking down at Earth and drawing power from starlight. TechCrunch reports that Google also tried space-themed limericks and, more practically, had the spacecraft report on orbital conditions around itself. A company spokesperson told the publication the satellite is operating nominally.

A single TPU on a tight power budget

The satellite carries one Google Tensor Processing Unit and can generate up to 1 kW of electricity, according to TechCrunch. Even so, the chip's power appetite means it can only run for about twenty minutes at a time, which complicates the already difficult business of moving data between orbit and the ground.

The larger ambition, TechCrunch notes, is a future network of closely flying satellites handling large inference workloads in space.

Radiation is the real experiment

The poetry was a demo; the purpose of the mission is to see how the hardware survives. The ongoing test campaign will examine how the TPU performs under thermal stress and in the space radiation environment.

Google has already put its chips through particle accelerator tests on Earth, but the team now needs to observe radiation's actual effects on the processor in flight. As TechCrunch explains, radiation can flip bits in memory, corrupting data without tripping the monitoring tools that would normally catch such faults.

Ground testing suggested radiation would have minimal impact on inference workloads and be more of a problem for training a model. The satellite now gives Google a chance to test those dynamics in reality rather than simulation.

Two more satellites in 2027

If the onboard TPU proves capable, the next step is launching two more satellites equipped with advanced chips in 2027, according to TechCrunch. Those spacecraft would generate more power and demonstrate inter-satellite optical links.

Per TechCrunch, that networking focus is what sets Google's Starcloud project apart from other space data-center visions: it plans high-speed links between satellites to support coherent distributed workloads. SpaceX's planned data centers are also expected to use laser links, but its satellites would orbit too far apart to share processing effectively, while other entrants in the field plan to run inference on individual satellites.

Why it matters

This is the first time Google has run its own data-center-style compute in orbit, turning orbital data centers from a concept into hardware generating results — however modest — above the atmosphere.

The economics still hinge on launch costs. In a peer-reviewed analysis published earlier this month, Suncatcher researchers estimated SpaceX would need to fly Starship 1,800 times to achieve the same learning-driven cost reductions it saw with earlier vehicles. Both Google's and SpaceX's orbital data-center plans depend on Starship becoming fully and rapidly reusable, and Alphabet, Google's parent company, holds a significant investment stake in SpaceX.

There is also a timing tension worth watching. TechCrunch characterizes Suncatcher as a classic Google moonshot, aimed at infrastructure that becomes economical in the 2030s. Yet Alphabet's stock fell this week after the company reported that inserting AI into its search business is hurting ad revenue growth — a reminder that decade-scale bets ultimately depend on the company's near-term financial footing.