Google receives first response from TPU on prototype satellite
Google said on Oct. 30 (Oct. 31 Japan time) that a prototype satellite for "Project Suncatcher," its research into placing AI computing infrastructure in space, had generated an AI response using its onboard AI chips, "TPUs," and that the response was received on the ground. While the satellite was in contact with a ground station, Google asked its open model (a model with published weights) "Gemma 4," carried on the satellite, for a haiku about being in space, and received the response as the satellite passed over Norway's Svalbard archipelago. Google described it as the first AI response generated in orbit by a data center TPU.
The haiku that came back was "Blue marble below / Data streams in silent grace / Starlight feeds my core." It then also ran a five-line limerick about space and a status report on its onboard equipment. The satellite was launched on Oct. 1, and over several weeks had completed powering the board carrying the TPUs and initial machine learning tests.
A 'hello world' milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we've now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth – a haiku!
Google will next examine how well the TPUs withstand heat and radiation in orbit. In 2027, it plans to launch two satellites and test communication by linking them with lasers.
A year after unveiling the concept, one satellite launched first
On Nov. 4, 2025, Google unveiled a concept of putting TPUs on solar-powered satellites and linking many satellites with lasers to perform machine learning computation. It estimated that with the right orbit, solar panels could generate up to eight times as much power as on the ground, and that launch costs could fall below $200 per kilogram by the mid-2030s. It also presented an example configuration of 81 satellites flying together within a radius of 1 kilometer.
Initially, the plan was to launch two satellites with satellite company Planet by early 2027. In practice, one jointly developed prototype was put into orbit first on Oct. 1 on SpaceX's rideshare launch "Transporter-18." In ground tests, Trillium-generation TPUs withstood radiation doses exceeding five years' worth in space and also withstood vibration during launch, the company said.
- Google unveils the Project Suncatcher conceptOfficial
- Launch the following week announcedOfficial
- Prototype satellite launched and contact confirmedOfficial
- Response generated by TPU in orbit receivedOfficial
- Test linking two satellites with lasers plannedOfficial
Pichai's post did not mention the model name; the use of Gemma 4 was disclosed in Google's blog post published the same day.
Four TPUs, about 15 minutes at a time because of thermal constraints
The prototype is small in scale. According to The New York Times, the satellite is about the size of a refrigerator, carries four TPUs and has about 1 kilowatt of power. The four TPUs are equivalent to the computing power of a single server in a data center, the report said.
In space, there is no air or water for cooling, and heat can only be shed as infrared radiation from radiators. According to U.S. public radio NPR, the prototype's TPUs must be stopped and cooled after running for about 15 minutes at a time. BigGo Finance, relaying South Korean reports, said that because of power constraints the TPUs are run one at a time for about 20 minutes. The prototype circles in a standard orbit different from the future concept and also uses batteries.
Google has not published performance figures such as inference speed. Travis Beals, who leads the project, described the prototype before launch as a very small test. AI data centers on the ground use tens of thousands of chips, and matching that scale in space would require many satellites.
GPU-based Starcloud got AI into orbit first, SpaceX targets 2027
Google is not the first to run AI in orbit. NVIDIA-backed startup Starcloud in November 2025 launched the satellite "Starcloud-1" carrying an NVIDIA "H100" GPU, ran Gemma in orbit and also trained a small AI model. Google's "first" applies only to data center TPUs.
Starcloud raised $250 million in August. Its valuation was $2.3 billion, more than double the $1.1 billion in March, when it raised $170 million. NVIDIA and Cisco newly invested, and the company aims to place 88,000 satellites with a total of 20 gigawatts of computing capacity in orbit.
SpaceX, working with NVIDIA, plans to launch satellites that perform AI computation in the October-December quarter of 2027. Elon Musk has said the scale will expand significantly in 2028. According to reports, up to 1 million satellites are envisioned in the future.
The number of AI chips per satellite is one H100 for Starcloud-1 and four TPUs for Google's prototype. According to reports, the satellites SpaceX plans to launch in 2027 will each carry about 72 NVIDIA chips.
More than five years before it is cheaper than the ground
Google cites nearly uninterrupted sunlight as the reason for choosing space. On the ground, Google's parent Alphabet raised its 2026 capital spending forecast to $195 billion to $205 billion in its late-October earnings and indicated that 2027 would exceed $300 billion. Project Suncatcher is positioned as long-term research looking toward the 2030s, separate from this ground-based investment.
The cost outlook, however, is tough. At launch, Beals said he did not think space would be cheaper in the next five years. Launch costs are high, and repairing failed equipment is difficult. Another challenge is that when radiation rewrites data in memory, it is hard to detect with normal monitoring. OpenAI CEO Sam Altman said in February that building data centers in space under current conditions is "ridiculous." Musk, on the other hand, said in January that space will become the cheapest place to run AI within two to three years.
Shares of Planet (NYSE: PL), which jointly developed the prototype, rose as much as 7.8% on Oct. 2, the day after the launch. The company's annual revenue exceeds $300 million, but nearly all of it comes from selling satellite imagery.
