Google gets its first AI answer from a satellite in orbit: Project Suncatcher takes off
On October 30, 2026, Google announced on its blog that the Project Suncatcher prototype satellite had sent back the first response from an AI model in orbit. The post, signed by Travis Beals, Senior Director of Paradigms of Intelligence, shares the haiku generated by Gemma 4, Google DeepMind's open model preloaded on board: "Blue marble below / Data streams in silent grace / Starlight feeds my core. ✨". According to Google, the satellite launched on October 1, 2026 on SpaceX's Transporter-18 rideshare mission; TechCrunch specifies it was a Falcon 9 carrying more than 100 spacecraft. Google built the mission together with Planet, which, according to TechCrunch, built the satellite.
According to Google, the team completed the commissioning checklist, powering on the host trays and the TPUs and running the first machine learning tests. TechCrunch adds that subsystem commissioning took about three weeks. TechCrunch also reports that there are four TPUs on board, but only one runs at a time, each for about 20 minutes because of power draw; the satellite generates up to 1 kW. TechCrunch corrected its own article on exactly this point about the four TPUs.
The announced next steps include testing how the TPUs behave under thermal stress and radiation exposure in orbit; the following milestone is a pair of interconnected satellites in 2027, which, according to TechCrunch, will demonstrate optical inter-satellite links and higher power generation. Google has already tested its Trillium TPUs with a proton beam at UC Davis's Crocker Nuclear Laboratory, and says they withstand a total ionizing dose higher than a five-year mission would deliver. According to TechCrunch, ground tests point to minimal radiation impact on inference but a potentially bigger problem for training, because of bit flips: orbital data will have to confirm this.
Project Suncatcher aims to assess whether orbital data centers are feasible, taking advantage of the near-constant sunlight in low Earth orbit and potential laser networks between satellites. TechCrunch frames it as a first step toward an economic model for the 2030s, but viability will depend on future drops in launch costs, for example through Starship reusability, something that has not yet been quantified. — Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read Google's official post of 10/30/2026 (date, Gemma 4, haiku, Svalbard, next steps, Beals quote) and Google's official Suncatcher page (Planet, Transporter-18, radiation tests on Trillium, 2027 plans). I used TechCrunch's 10/30/2026 article as independent confirmation: Gemma 4, haiku, launch, four TPUs used one at a time (with its correction note), 1 kW, spokesperson. The key facts match across all three sources.
- Incertezze
- Google does not say which TPU generation is flying: Trillium is mentioned only for the ground tests. It does not say how many TPUs are on board either. The four TPUs, one at a time for about 20 minutes, and 1 kW come only from TechCrunch, which corrected its article on this point. There is no data yet on real performance under radiation and thermal stress, or on costs. The economics of orbital data centers depend on lower launch costs (Starship reusability) and remain a long-term bet.
- Perché pubblicarla
- It is the first AI model response generated in orbit by a data-center-class TPU: a concrete, documented step for space data centers and their power limits. None of our previous articles had covered it.