Google’s satellite AI chip answers its first prompt from orbit with a haiku

An AI chip aboard Google’s experimental satellite has answered a request from Earth and sent back a poem. Google disclosed the successful exchange on October 30, 2026, marking an early working demonstration for Project Suncatcher, its research effort to find out whether space can support large-scale AI computing.

The task was modest: write a haiku about being in space. Its significance was the complete exchange, not the poetry. Engineers sent a prompt to the spacecraft, an onboard Gemma 4 model generated an answer, and the satellite transmitted that answer as it passed Svalbard, Norway.

Google research director Travis Beals called it the project’s “Hello World” milestone. His claim was specific: the first AI response from a data-center-class Tensor Processing Unit, or TPU, in orbit. TPUs are Google’s AI processors; the announcement was not a claim to have run the first AI model of any kind in space.

Four chips, short operating windows

The prototype launched on October 1. Google then spent several weeks checking its systems, powering up the host computer and TPU trays, and running early machine-learning tests. TechCrunch reports that commissioning took about three weeks before the spacecraft was ready for the prompt.

The hardware has a tight power budget. According to TechCrunch, the satellite carries four TPUs and generates up to 1 kilowatt of electricity. It can run only one TPU at a time, for about 20 minutes. That constraint also complicates communication with the ground, which already depends on suitable contact windows.

Gemma 4 was already loaded onto the spacecraft. The exchange performed inference: using a trained model to produce an answer to a new input, rather than training a model. PPC Land notes that Google’s October 30 announcement supplied no performance measurements, so the result is a functioning demonstration, not a published speed or efficiency benchmark.

The returned poem read: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Google followed it with space-themed limericks and prompts asking for telemetry, information about the technical condition of the payload systems. Those requests extended the test beyond creative writing to reporting on the hardware itself.

Radiation can spoil an answer silently

Google’s next phase will test how the processors perform under thermal stress and radiation exposure in orbit. The company has already tested its chips in particle accelerators on Earth. The spacecraft now provides a way to compare that laboratory work with hardware operating in the actual space environment.

Radiation can cause bit flips, changes to values stored in computer memory. Those errors can corrupt data without triggering ordinary monitoring tools. Ground testing suggested radiation would pose less trouble for inference than for model training, but the orbital campaign will assess how the chips behave outside the laboratory.

The Register reports that Google’s laboratory work found high-bandwidth memory particularly sensitive to radiation. Researchers estimated a very low failure probability for inference under their assumed orbital exposure and workload. They also said the effect on training jobs, and the effectiveness of system-level protections, needed further study.

The larger experiment needs connected satellites

Google identifies twin interconnected satellites in 2027 as its next milestone. TechCrunch reports that those spacecraft are intended to generate more power and demonstrate optical links between satellites. The goal is to let nearby spacecraft share computing work, rather than leave each satellite operating as an isolated machine.

That design separates Suncatcher from approaches focused on inference aboard individual spacecraft. TechCrunch describes Google’s plan as a network of close-flying satellites with high-speed connections supporting coordinated workloads. The present prototype tests an essential component of that design; communication and shared processing across spacecraft are later steps.

Even reliable hardware would leave a separate economic hurdle. As The Register explains, Google calculates that launch prices around $200 per kilogram could make launch costs, spread over a satellite’s useful life, comparable to electricity spending for US ground-based data centers. That comparison concerns those particular costs, not proof that a complete orbital facility would be cheaper.

Google projects that lower launch-cost threshold could arrive around 2035, making Suncatcher a long-term infrastructure experiment. Its immediate work is more concrete: determine whether the processors can reliably keep executing workloads through the heat and radiation conditions that a successful poem alone cannot test.