Google's First Space-Based AI Data Center Sends Haiku Back to Earth

Last Friday (October 30), Google published a detailed update on what many had been tracking for weeks: the MVP satellite from the Project Suncatcher successfully transmitted its first response from Earth orbit, running an artificial intelligence workload and sending back a haiku. Described internally as the project's "Hello World" moment, this milestone marks the first time a data-center-class TPU (Tensor Processing Unit) chip has operated effectively beyond the atmosphere.

What happened

The satellite launched on October 1 aboard SpaceX's Transporter-18 rideshare mission into low Earth orbit. For two weeks, the team completed the commissioning checklist—powering up the host processor and TPU trays, running preliminary machine learning tests. Last week, during a pass over the Svalbard ground station in Norway, engineers sent the first prompt directly to the Gemma 4 model running onboard. The result was a haiku about being in space:

> Blue marble below > Data streams in silent grace > Starlight feeds my core.

The satellite also responded with limericks and telemetry prompts reporting the status of its onboard systems. A spokesperson confirmed the satellite was operating nominally.

What makes this transmission particularly noteworthy isn't the poetic content—it's the underlying infrastructure. Every time someone runs an AI inference, whether answering a question, translating text, or generating an image, the data needs to be processed by a dedicated chip. On terrestrial data centers, these chips sit in servers connected by fiber-optic networks. On Suncatcher, the same kind of chip runs inside a 35-kilogram satellite, powered by solar panels instead of a wall outlet.

The context behind the launch

Project Suncatcher was officially announced in November 2025 as a Google research "moonshot," inspired by earlier projects like autonomous vehicles and quantum computing. Blaise Agüera y Arcas, Google's vice president and AI researcher who leads a team studying intelligence, was the principal architect. The name comes from a British children's rhyme about "catching the sun"—and the goal is ambitious enough to justify the nickname: building a constellation of laser-linked satellites that would form orbital data centers capable of executing massive inference workloads.

The motivation is twofold. First, energy: in low Earth orbit, satellites access near-constant sunlight, generating up to eight times more solar power than anywhere on the Earth's surface. Second, space: there is growing opposition to terrestrial data centers over environmental and land-use concerns. Travis Beals, senior director and Suncatcher lead, told NPR that "the Sun puts out almost all of the power in our solar system. All of the other power sources that humanity has tapped into are just a tiny fraction of a percent."

The technical challenges that remain

The haiku transmission was just the first step. The engineering gaps between this prototype and a fully functional constellation are enormous.

Cooling in a vacuum: On a terrestrial data center, fans and air conditioning systems dissipate the heat generated by chips. In space, there is no air. The only way to dissipate heat in a vacuum is through radiators. Google is testing a combination of heat pipes and radiators to transfer chip heat into space. On the MVP prototype, the TPU chips have to power down every 15 to 20 minutes to avoid overheating. The two follow-up satellites, scheduled for 2027, should operate continuously.

Cosmic radiation: Outside Earth's magnetic field protection, cosmic rays and solar events can cause catastrophic errors in electronics. Google tested its Trillium TPUs in a proton beam at UC Davis's Crocker Nuclear Laboratory while simultaneously running AI workloads. The initial results showed that the TPUs withstand a total ionizing radiation dose greater than what they would receive during a five-year space mission—an encouraging result, but still unconfirmed in flight.

Interconnectivity: In the final design, each satellite would carry dozens of TPU chips, forming clusters that communicate via high-bandwidth laser links. Google has already published a paper in the journal Joule describing the constellation's design, including control, communication, and data architecture. But laser interconnection between satellites separated by hundreds of kilometers, through Earth's atmosphere, remains an engineering problem in development.

The competitive landscape

Google is not alone in exploring space-based data centers. SpaceX, in which Alphabet (Google's parent company) is a shareholder, has already announced plans to build "orbital compute satellites" equipped with thousands of GPUs, with a target deployment start of 2028. Cowboy Space, a startup founded by a former Rocket Lab executive, is sending its Reason-1 satellite to test laser data links in orbit and has its own rocket in development, whose upper stage doubles as a 1 MW data center. Star Catcher is testing power transmission between two spacecraft on the same Transporter-18 launch.

The difference between Suncatcher and the competitor projects is scale and ambition. While Cowboy and Star Catcher plan to run inference workloads on individual satellites, Google has described an ecosystem of interconnected constellations—a concept that resembles, in scale and complexity, the communications satellite constellations already operating today.

What's next

The next Suncatcher step is the launch of two additional satellites in 2027, which will test laser interconnectivity between them—the orbital equivalent of linking two servers on the same rack. If it works, the next step would be scaling to a constellation of dozens or hundreds of satellites, each carrying 50 to 100 kW of compute.

Google estimates that if launch costs fall to $200/kg—which could happen if Starship flies about 180 times per year, as the company projects—the cost to launch one kilowatt of solar power to orbit would be about $810 per year, comparable to what US data centers spend on electricity. It's an optimistic projection, but not impossible.

The question that remains isn't whether space-based data centers are technically possible—the Suncatcher mission just proved they are. It's whether they'll be economically viable, environmentally superior, and operationally simpler than terrestrial alternatives. The haiku that the satellite sent from orbit might be just the first verse of a much longer poem—or it might be the prologue to a completely new genre of computing infrastructure.

Sources: The Verge, Google Research Blog, TechCrunch

✓ Independent sources cross-checked and verified before publishing