Google's Suncatcher runs its first AI inference in orbit as Satlyt tests satellite task sharing

Google's Suncatcher responds from orbit

Google's experimental Project Suncatcher satellite has produced its first AI-generated output from space. According to The Verge, Google said in a blog post on Friday that it successfully powered up the satellite's Tensor Processing Units, allowing a Gemma 4 model running on board to answer a prompt by composing a haiku about life in orbit. The company also ran other space-themed limerick prompts and telemetry prompts during the test.

The satellite launched earlier this month. Project Suncatcher is an experimental effort that Google says could eventually lead to AI data centers in space. The next phase is a series of tests measuring how the TPUs hold up under thermal stress and radiation exposure in orbit. Google says proving the chips can run workloads reliably in that environment is a prerequisite for its next milestone: twin interconnected satellites planned for 2027.

Satlyt demonstrates onboard task planning

Separately, startup Satlyt reported progress on the software side of the same problem. According to TechCrunch, the company — founded in 2024 by Rama Afullo, who previously worked at Google and SpaceX — is building software meant to turn spacecraft into virtual data centers, running AI and other applications on board so data is processed close to where it is collected rather than downlinked for ground analysis.

After 13 months of work, Satlyt completed a successful test earlier this month alongside the University of Houston on a NASA-funded research project. The test showed the software can decide how to distribute tasks among computers in orbit based on compute and communication availability. One physical computer aboard a spacecraft evaluated a simulated network of four computing hosts, four users, and eight services, then returned its results to Earth. Satlyt's software is already flying on three satellites.

Afullo told TechCrunch that users of the envisioned system would only need to define permissions, priorities, and operating limits, with the software making decisions within those bounds. He noted that NASA missions face the same constraints the company is targeting: scarce computing resources, patchy communication, and pressure to deliver useful information to people quickly.

There are caveats. The work was funded through NASA research, but the agency has not selected the technology for any mission or deployment. The experiment also surfaced a bug in how the software referenced delay timing — it recorded when data arrived rather than how long the transfer took, which is a different metric. Afullo described the core finding as a trade-off, since the best-performing approach also consumed the most computing power. He nonetheless framed the test as one piece of the foundation needed for satellites to share computing work. Satlyt placed in the top five of TechCrunch's Disrupt Battlefield competition this year.

What still needs to be proven

Neither project is close to a working orbital data center. Satlyt has yet to demonstrate its software coordinating multiple physical computers exchanging data with each other in orbit, and says further testing is needed across different workloads, hardware, and connectivity conditions. Afullo listed what remains: usable communication links, permission from participating operators, secure transfers, resource scheduling, and recovery when a connection between computers drops. The company's next steps are broader validation and integration, followed by demonstrations of computing work shared across separate spacecraft. It is also researching applications beyond low Earth orbit, including 4K video streaming from the Moon through its NASA-funded work.

Google, for its part, has shown only that its accelerators can boot and run a small model in orbit, with durability tests still ahead.

Why it matters

Satellites collect far more data than they can send home during limited communication windows, and every hop to ground-based analysis adds delay. Onboard processing — whether Google's approach of flying AI accelerators or Satlyt's approach of orchestrating compute across existing satellites — could shorten the distance between an observation and a useful answer. Afullo's example to TechCrunch: a spacecraft that spots a wildfire could get relevant information to responders far faster. Google's involvement signals that large AI infrastructure players see orbit as a plausible extension of the data center, while Satlyt's test shows the scheduling and orchestration layer that fleets of satellites would need is being built now. Both efforts remain early, but the two milestones surfaced the same week, one on hardware and one on software, indicating genuine momentum behind moving computation off the planet.