Google has successfully tested an AI model aboard an orbiting satellite, demonstrating how onboard processors could support future space-based computing infrastructure without relying on ground-based data centres.
Google has achieved an important milestone in its exploration of space-based artificial intelligence by successfully running an AI model aboard a satellite in orbit. The experiment demonstrated that an AI system could receive instructions from Earth, process them using onboard hardware and transmit its response back without depending on ground-based computing resources.
The demonstration was conducted under Google's Project Suncatcher, which aims to explore the possibility of building data centres in space to support the growing computing demands of artificial intelligence.
For the experiment, Google installed its Gemma 4 AI model on a prototype satellite equipped with one of the company's Tensor Processing Units (TPUs). The satellite was developed in collaboration with Planet, a satellite technology company.
The spacecraft was launched aboard a SpaceX Falcon 9 rocket as part of the Transporter-18 rideshare mission, alongside more than 100 other spacecraft.
After reaching orbit, the satellite received a command from engineers on Earth to generate a haiku from the perspective of a TPU travelling around the planet.
Google CEO Sundar Pichai described the achievement as a "hello world" moment. The phrase refers to a basic first test used to confirm that a new computing system is functioning.
The satellite generated the following poem:
Blue marble below
Data streams in silent grace
Starlight feeds my core.
As the spacecraft passed over Svalbard in Norway, the team successfully received the AI-generated response on Earth.
Google also tested other prompts, including space-themed limericks and requests for information about the satellite's onboard systems. The company reported that the spacecraft's systems were operating as expected.
The prototype satellite carries four Tensor Processing Units, Google's specialised AI chips designed to handle machine learning workloads.
However, the spacecraft has limited power-generation capacity of approximately 1 kilowatt. As a result, it can operate only one TPU at a time for around 20 minutes.
These computing sessions must also be scheduled according to the satellite's communication windows with ground stations. This allows engineers to send instructions and receive the results during periods when the spacecraft can communicate with Earth.
The experiment is designed to assess whether specialised AI processors can reliably perform computing tasks in the challenging conditions of space.
One of Google's major challenges is determining whether its AI hardware can withstand prolonged exposure to radiation and extreme temperatures in orbit.
Cosmic rays and solar radiation can cause bit flips, a phenomenon in which stored data changes unexpectedly. Such errors could potentially affect calculations and disrupt the functioning of AI systems.
To examine this risk, Google conducted ground-based tests at the University of California, Davis. Its Trillium TPUs were exposed to a proton beam representing radiation levels greater than those expected during a five-year space mission.
The results indicated that AI inference tasks may be less vulnerable to radiation-related disruptions than model training.
Temperature management is another challenge for orbital computing systems. Unlike equipment on Earth, satellites cannot use conventional air-based cooling fans because there is no surrounding atmosphere.
Google's prototype instead relies on heat pipes and radiators to manage the temperature of its onboard hardware.
These tests will help the company understand whether AI chips can continue operating reliably during extended missions.
Google is planning to expand Project Suncatcher by launching two additional satellites in 2027. The upcoming spacecraft are expected to have greater power capacity and laser communication links.
These links could allow satellites to communicate directly with one another and share AI computing workloads, potentially creating the foundation for an interconnected orbital computing network.
Google also sees solar energy as a potential advantage of space-based data centres. According to the company, sunlight in orbit can provide up to eight times more solar power than on Earth under certain conditions.
However, establishing large-scale computing infrastructure in space will require significant investment and technological improvements.
Despite the successful demonstration, the commercial feasibility of orbital data centres remains uncertain.
The cost of launching heavy computing equipment, maintaining satellites and managing communication networks could make the technology expensive. Researchers estimate that SpaceX's Starship may need around 1,800 flights to make orbital data centres economically viable during the 2030s.
Google's current experiment is therefore an early-stage demonstration rather than proof that large-scale space-based data centres are ready for commercial use.
The performance of the prototype over the coming months will help the company assess whether its next planned satellite launches can move forward.
Conclusion
Google's successful AI experiment in orbit represents an early step towards exploring space-based computing. Although generating a poem is a relatively simple task, the demonstration shows that AI models can process information using onboard satellite hardware. Future missions will determine whether radiation-resistant chips, improved power systems and satellite communication networks can make orbital data centres technically and economically practical.