Add preferred source
Google has succeeded in generating a generative AI response using its own AI chip, the TPU, aboard a prototype satellite under its Project Suncatcher. The satellite-borne Gemma 4 composed a haiku in orbit and transmitted it to the ground — the first case of data-center-class TPU AI computation in orbit. Starcloud, SpaceX, and China’s ADA Space have also launched space AI computing demonstrations, while South Korea does not plan to launch a verification satellite until after 2030. Challenges to commercialization remain, including heat dissipation, radiation, and launch costs.
Key Elements
Google has succeeded for the first time in receiving a generative AI response from Earth orbitnsor Processing Unit (TPU). As the power and land burdens of ground-based data centers grow, competition among Big Tech companies to move AI computing infrastructure into space — where solar power can be harnessed — has entered a full-fledged demonstration phase
Google said on the 30th of last month (local time) that it had received its first AI response from a prototype satellite under “Project Suncatcher,” its space data center research initiative. When the satellite’s onboard open model, Gemma 4, was asked to express in haiku what it feels like to be in space, the satellite — passing over a ground station in Svalbard, Norway — sent down a three-line poem: “A blue orb below / Data flowing through quiet grace / Starlight fills my core.” Google says this marks the first AI response generated in orbit by a data-center-class TPU.
The satellite was launched on October 1 aboard a SpaceX Falcon 9 rocket. It was built jointly with satellite company Planet and carries four TPUs. It generates up to 1 kW of power from solar panels, but due to power constraints it can reportedly operate only one TPU at a time for about 20 minutes. After roughly three weeks of commissioning, Google powered up the host and TPU tray and completed initial machine-learning tests. It subsequently tested generating a five-line space-themed poem and issuing commands to report payload status.
Google’s achievement shows that the global race over space AI data centers has already entered the real-world phase. Starcloud, a startup backed by Nvidia, launched its “Starcloud-1” satellite carrying Nvidia’s high-performance H100 GPU in November of last year, and a month later ran the Gemma model in orbit to obtain an AI response. SpaceX, led by Elon Musk, is also preparing a “StarMind” satellite equipped with Nvidia’s AI computing system. Musk has laid out plans to launch AI satellites in the fourth quarter of next year and to scale up significantly by 2028.
China is moving quickly as well. In May of last year, ADA Space launched the first 12 satellites of its “Three-Body Computing Constellation.” Its long-term vision is to link 2,800 satellites, process satellite imagery with AI in space, and downlink only the necessary information to the ground.
South Korea, by contrast, has yet to take its first step. The Korea AeroSpace Administration plans to launch a platform satellite for verifying a space AI data center only after 2030. At a parliamentary audit of the National Assembly’s Science, ICT, Broadcasting and Communications Committee on the 23rd of last month, it was noted that the government’s budget for space AI data centers in next year’s budget proposal amounts to only 2.1 billion won (approximately $1.6 million). Rep. Lee Joo-hee of the Democratic Party urged the Korea AeroSpace Administration to increase the related budget and accelerate technology development.
There remain no small number of technological hurdles to clear before space data centers can be commercialized. In the vacuum of space, heat from chips cannot be cooled with air or water, so it must be transferred to a radiator and released as infrared radiation. Google is testing a cooling method combining heat pipes and radiators. Space radiation can cause errors in chips and memory, and the cost of repairing and replacing equipment in the event of failure is enormous.
Forecasts on economic viability are mixed. Musk argued at the World Economic Forum in January that “within two to three years, space will be the cheapest place to run AI,” while OpenAI CEO Sam Altman countered at an event in India in February that “the idea of building data centers in space under current conditions is absurd.” The view is that, given enormous launch costs and the problem of repairing failed GPUs, meaningful-scale operation within the 2020s will be difficult.
Google, too, is cautious about near-term commercialization. Travis Beals, who leads Project Suncatcher, said “our chips have officially earned their space wings,” while nonetheless expecting that it will take considerable time for space data centers to become cheaper than ground-based ones. After completing verification of the TPU’s heat and radiation tolerance, Google plans to launch two additional satellites in 2027 and conduct an experiment linking them via laser optical communications. The goal is to confirm whether AI chips distributed across multiple satellites can exchange data and jointly handle large-scale computation.
Han Young-min, director of the Space Launch Vehicle Research Institute at the Korea Aerospace Research Institute, said that “space AI data centers and satellite internet will become very important space infrastructure from a national security standpoint,” adding that “we need to pick up the pace on core technology development and space demonstrations in preparation for intensifying competition to seize the market.”
If space AI computing infrastructure becomes a reality, it could partially ease the bottlenecks facing ground data centers over the power and land needed for AI training and inference. However, unless launch costs, hardware reliability, and maintenance systems are in place, securing economic viability will be difficult, so for the time being the competition is expected to remain one of technology demonstration.
