Suncatcher tests Google's AI chips in orbit
Google has validated a first round trip between Earth and a TPU chip in orbit, with a haiku generated by Gemma. This demonstration serves as a milestone for Project Suncatcher, which aims for AI data centers powered by satellites.
In October 2026, Google successfully sent a prompt to a TPU chip placed in orbit and then retrieved a response generated by its Gemma model on Earth: a haiku. This demonstration, presented as a step in Project Suncatcher, shows that the company’s chips can boot up and compute in space. It is significant mainly because it tests a workaround for the very concrete limitations of terrestrial data centers: electricity, cooling, and land availability.
Un haïku pour valider la chaîne entière
Sundar Pichai announced on X that the teams had "successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth – a haiku!" The prompt sent requested a haiku from the perspective of a TPU in orbit. Gemma’s response was: "Blue marble below, Data streams in silent grace, Starlight feeds my core. ✨" The format is almost trivial. The message, however, is clear: the computation loop works end to end.
The "hello world" ritual is old in computing, but here it takes place several hundred kilometers above ground. This gives the operation a somewhat theatrical aspect, without masking the real challenge: verifying that hardware designed for server rooms can withstand launch, vacuum, thermal variations, and radiation.
Pourquoi Google regarde vers l’espace
Project Suncatcher starts from a now familiar observation in AI: demand no longer only bumps against chips, but also access to energy and the capacity to install ever more ground infrastructure. Google therefore wants to explore clusters of solar satellites connected by laser, capable of hosting AI processing in orbit. In this model, the sky is not just a backdrop but an industrial site. The idea is ambitious, and engineers clearly enjoy playgrounds that start with a space launch and end with a spreadsheet.
The prototype, launched on October 1, 2026 aboard SpaceX’s Transporter-18 from Vandenberg, California, was designed with Planet. Nicknamed MVP for "Minimum Viable Product," it carries four TPU Trillium, Google’s AI chips, and about one kilowatt of solar panels. Google first wanted to verify the hardware’s endurance under extreme conditions. According to the company, launch shocks can reach 50 to 100 g for some components.
Ce que les premiers essais ont déjà montré
Technically, the first results are encouraging. According to Google, solar panels in a so-called "dawn-dusk" heliosynchronous orbit at about 650 km altitude can produce up to eight times more energy than on the ground. The research paper published in 2025 envisions clusters of 81 satellites flying a few hundred meters apart, connected by laser. In testing, these optical links reached 1.6 terabits per second. Meanwhile, TPU Trillium chips bombarded with protons in an accelerator showed no outright failure up to 15 krad, well beyond the dose expected for a five-year mission.
These figures are not enough to conclude commercial viability, but they shift the subject. Google is no longer just testing an engineering curiosity; the company is evaluating an infrastructure scheme where computing distribution would depend on orbits, solar panels, and laser links. At this stage, the real question is not whether a TPU can recite a poem. It is whether a constellation can, tomorrow, support real computing loads without turning into an expensive orbital menagerie.
Le rendez-vous de 2027 dira si le modèle tient
The next decisive test is scheduled for 2027, when two satellites must exchange data by laser while flying in close formation. If this test succeeds, Project Suncatcher will move from proof of concept to a credible architecture for AI deployed in orbit. If it fails, the haiku will remain mainly an elegant symbol of a still very preliminary experiment. Google is not alone in this field: in December 2025, the startup Starcloud, supported by Nvidia, had already run Gemma on an onboard H100 GPU and trained nanoGPT on Shakespeare’s complete works. The first space AI expo has thus already opened, but the list of exhibitors remains short.
According to Sundar Pichai, interviewed by Fox News last December, data centers in space could seem "more normal" within about ten years. The phrase well expresses Google’s strategy: to shift an idea today considered marginal toward an ordinary infrastructure scenario. The economic equation remains to be proven alongside the poetry.
Key points
- Google sent a prompt into orbit and received a haiku.
- The MVP prototype launched on October 1, 2026.
- Four TPU Trillium chips are onboard.
- The next decisive test is planned for 2027.
- Starcloud has already run Gemma in orbit.
- Sundar Pichai aims for normalization within about ten years.
By the numbers
- October 1, 2026 — launch date of the prototype, SpaceX Transporter-18 mission.
- 4 TPUs — number of Trillium chips onboard the MVP satellite.
- 650 km — targeted altitude for the "dawn-dusk" heliosynchronous orbit.
- 1.6 terabits per second — data rate reached in testing for laser links, according to Google.
- 15 krad — radiation dose endured without outright failure by TPU Trillium during tests.