Google makes an AI satellite speak in low Earth orbit
Google's orbital prototype Suncatcher ran Gemma 4 in low Earth orbit and sent a haiku back to Earth. The test mainly serves to assess power, radiation, and communications for future satellite networks.
Google successfully ran Gemma 4, its AI model, aboard a prototype satellite placed in low Earth orbit earlier this month, then had it send a haiku back to Earth. The experiment, named Project Suncatcher, marks the first complete message received from space by this orbital testbed. Beyond the literary nod, the goal is more concrete: to validate whether AI computing can operate outside the atmosphere despite radiation, heat, and communication limits with the ground.
An orbital “hello world,” but not a gadget
The test took place during a station pass over Svalbard, Norway, after several weeks of commissioning. Travis Beals, senior director within Google's Paradigms of Intelligence team, explains that the satellite responded to a command sent from the ground: write a haiku from the perspective of a TPU orbiting the Earth. The reply was transmitted back as follows: “Blue marble below, / Data streams in silent grace, / Starlight feeds my core.”
This first exchange not only proves that a model can produce text from space. It also shows that the satellite's subsystems, TPU trays, and command chain work together in an environment where everything becomes more delicate. The system carries four Trillium TPUs tested against radiation and powered by direct solar energy. According to Clement Farabet, this is a true “offline” execution in low Earth orbit, without ground relay for computing.
Why Google is interested in AI in orbit
Google's logic boils down to a very down-to-earth equation: energy and cooling are expensive on Earth, and launching into orbit could become economically more attractive if prices drop significantly. In its research paper published this month, the company estimates that an orbital launch cost reduced to 200 dollars per kilogram could eventually make the amortized cost of the satellite comparable to the electricity bill of terrestrial datacenters in the United States. For now, we are far from that. Very far.
The strategic interest is therefore to anticipate. Google believes it is better to test now the technical building blocks needed for a future constellation of interconnected small satellites, rather than wait for the market or launch providers to do all the work. The company already plans, according to its own calculations, to move toward a pair of interconnected satellites next year, with optical links between devices and more ambitious demonstrations in 2027 [to be verified according to the exact timeline of announcements].
The real challenge is physics more than text
Poems sent from space are amusing. The constraints are less so. Google indicates that the next tests will focus on TPU behavior under thermal stress and in the orbital radiation environment. The company has already conducted laboratory tests with a 67 MeV proton beam to measure the robustness of its components. In its calculations, high-bandwidth memory remains the most sensitive element, with irregularities beyond 2 krad(Si), about three times the minimum required by Google for survival to ionizing doses.
The research paper also cites a rate of about one silent data corruption event per 17 rad, which would lead, at an estimated orbital dose of 150 rad(Si) per year, to a very low probability of error for inference, on the order of one case per 3 million inferences assuming a rate of one inference per second. For training, however, the impact of singular events and the effectiveness of correction mechanisms remain to be studied. Although the satellite writes poetry, it is not yet intended to improvise its own error corrections.
Satellites, TPUs, and much engineering still to prove
The current prototype remains modest. According to available information, the satellite weighs about 500 kg and can generate up to 1 kW of electricity. Power imposes its limits: some elements can only operate in short sequences, complicating ground-space communications and continuous operation. In other words, the idea of an “orbital datacenter” here is more of a flying laboratory than a giant computing center.
Google specifically wants to measure what such a system can withstand before considering a larger architecture. The next tests will focus on chip endurance to temperature, radiation, and inter-satellite links. If results are positive, the group could then move from a solitary prototype to a small cluster of satellites communicating with each other. Otherwise, the haiku will remain a costly nice memory, and orbital poetry will have made a brief appearance in cloud history.
Key points
- Google sent back a haiku from low Earth orbit on October 31, 2026.
- The Suncatcher satellite carries four Trillium TPUs tested for radiation.
- Gemma 4 ran offline, without ground relay for computing.
- Google aims to test interconnected satellites as soon as next year.
- The target cost mentioned for orbit is 200 dollars per kilogram.
- Upcoming tests will measure radiation, heat, and system reliability.
By the numbers
- 4 TPUs — onboard the Suncatcher satellite, according to Google.
- 1 kW — electrical power announced for the orbital prototype.
- 2 krad(Si) — threshold of irregularities observed in testing for high-bandwidth memory.
- 150 rad(Si)/year — orbital dose estimated by Google for reliability calculations.
- 1 in 3 million — approximate probability of failure mentioned for inference.