According to The Verge, the company’s Project Suncatcher prototype satellite booted its onboard Tensor Processing Units (TPUs), ran the Gemma 4 AI model and transmitted a response to engineers on the ground. After spending billions developing artificial intelligence, Google has now gone to the expense of launching it into space so it can write poetry.
The satellite launched on 1 October aboard SpaceX’s Transporter-18 mission and was developed in partnership with satellite manufacturer Planet. Google’s engineers spent approximately three weeks commissioning its systems before attempting their first complete AI inference test.
During a pass above Svalbard, Norway, engineers transmitted a prompt asking the onboard model to produce a haiku about being in space. The satellite generated a response and successfully transmitted it to Earth, demonstrating that Google’s data-centre-class silicon could execute an AI workload in orbit.
Google subsequently tested space-themed limericks and prompts requesting technical information about the satellite. Humanity has therefore reached the point where sophisticated orbital computing hardware is being used to produce questionable poetry, something previously accomplished much more cheaply by people with Substack accounts.
The prototype carries four Google TPUs and can generate approximately one kilowatt of electricity. However, its limited power supply means it can run only one TPU at a time for about 20 minutes, far short of the continuous operation expected from conventional data centres.
Power consumption is only one of several difficulties facing the project. Conventional data centres rely on elaborate cooling systems to remove heat from processors, but space provides no atmosphere for that heat to escape. Engineers must rely on heat pipes and radiators to dissipate the considerable thermal output of AI accelerators.
Radiation presents another headache. High-energy particles can corrupt calculations, damage semiconductor components, and cause memory errors. Google’s preliminary ground tests suggested its Trillium TPUs could withstand radiation exposure exceeding what would be expected during a five-year orbital mission, although extended testing in space remains necessary.
The appeal of orbital computing is its potential access to solar energy. Google estimates that satellites in suitable low-Earth orbits could receive enough sunlight to generate up to eight times as much solar power as comparable terrestrial installations. This could reduce dependence on electricity grids already struggling to accommodate the technology industry’s increasingly extravagant ambitions.
Launching processors, power equipment, communications systems and cooling hardware into orbit remains expensive. Maintaining the equipment is another matter entirely, since sending an engineer to replace a defective memory module would make even Apple’s repair charges appear reasonable.
Google hopes eventually to operate clusters of satellites connected through high-bandwidth laser communications. Such systems would allow multiple spacecraft to share computing workloads, although maintaining precise connections between satellites travelling at orbital velocities presents significant engineering difficulties.
The company plans to launch two interconnected satellites during 2027 to investigate whether these systems can operate reliably. Future prototypes are expected to carry more powerful computing hardware and test the high-speed optical links needed to distribute AI workloads across an orbital network.
Google has not demonstrated that orbital data centres can compete economically with facilities on Earth, and the first prototype has yet to prove sustained operation under space conditions. The next experiments will examine thermal behaviour and radiation tolerance, while commercially useful computing constellations remain several years away.
