Google’s Suncatcher AI satellite sends its first complete reply to Earth

Google says its experimental Project Suncatcher satellite has returned its first complete AI response to Earth. In an X post inspected on October 31, CEO Sundar Pichai said the team sent a command into orbit, started the TPU, ran a prompt and received a haiku. The announcement reports an end-to-end execution and return path from the prototype.

A TPU is a specialized chip that Google designs to accelerate machine learning. Its Cloud TPU documentation describes hardware built for large matrix operations and high-bandwidth memory. Those operations process many numbers together, a common requirement of machine-learning models. Suncatcher tests whether this class of AI computing hardware can operate in space. The terrestrial product documentation also describes grouping chips to scale workloads. Compiled machine-learning operations run on TPU hardware, while the rest of the program executes on a host computer. This provides context for the computing system being tested.

The response follows earlier spacecraft checks. Google announced on October 1 that the prototype, built with Planet, had launched on SpaceX’s Transporter-18 mission and established contact. Planet’s October 20 update said the spacecraft from that launch, including Suncatcher, were healthy and progressing through calibration. Receiving a computed reply adds a further reported milestone to that sequence. Google described the mission as collecting orbital evidence about launch stress, radiation and thermal extremes, then using the findings to refine later designs.

The longer-term concept combines solar power with computing shared among satellites. Google’s November 2025 research introduction says a panel in the right orbit could be up to eight times as productive as one on Earth, with nearly continuous generation. That design rationale depends on the orbit and connections between spacecraft, as well as the processors themselves.

Moving heat through a vacuum

Heat generated by the chips needs a path out. NASA’s small-spacecraft thermal-control reference explains that a vacuum eliminates convection: heat inside a satellite moves mainly by conduction, while exchange with its surroundings depends on thermal radiation. NASA describes radiators as dedicated surfaces for rejecting excess heat and identifies competition for limited exterior area, including solar cells, as a small-satellite constraint. Google reports thermal-vacuum testing of heat pipes and radiators.

Google’s ground tests ran AI on Trillium TPUs under a UC Davis proton beam, tracking bit flips. Its September account reports survival beyond a five-year mission’s expected cumulative radiation dose.

The design paper also addresses failures. Replacing a broken TPU by hand, routine in a terrestrial facility, is impracticable in space; the researchers propose spare capacity as the simplest response. They also favor reliable, preferably passive mechanisms to move concentrated heat loads from chips to dedicated radiator surfaces.

Google’s research introduction describes high-bandwidth optical links between closely spaced spacecraft. A ground demonstrator reached 800 gigabits per second each way with one transceiver pair, or 1.6 terabits combined. This terrestrial experiment addresses the data exchange needed to share AI workloads among satellites.

The paper examines a sun-synchronous low-Earth orbit and models an example formation of 81 satellites within a one-kilometer radius. Maintaining proximity while avoiding collisions is part of the control problem. That modeled cluster describes a possible future architecture; it is separate from the prototype that returned the newly announced response.

Google’s September plan schedules a two-satellite, high-bandwidth optical-link experiment for 2027.