On October 30, 2026, TechCrunch reported that Google's Project Suncatcher prototype satellite is "online and communicating with the ground". The satellite, built with Planet, launched at the start of October on a SpaceX Falcon 9 alongside more than 100 other spacecraft. After about three weeks of commissioning, engineers sent the first prompt this week to the model pre-loaded on board, Google DeepMind's Gemma 4. They also asked the spacecraft to report on its orbital conditions, and a spokesperson said it was operating nominally.
The hardware budget
TechCrunch says the satellite "carries four Google Tensor Processing Units (TPUs) and generates up to 1 kW of electricity", but "can only turn one on for about 20 minutes at a time", a point the outlet corrected after first publication. NPR's report after the launch said the chips would run Gemma for 15 minutes at a time "due to heat management constraints". Either way, this is not a data center. It is a test of whether an accelerator and a model survive orbit, with a duty cycle short enough that downlink windows and thermal limits decide when it computes.
Why radiation is the experiment
Google's November 2025 Suncatcher paper describes ground tests of Trillium, its v6e Cloud TPU, in a 67 MeV proton beam. The High Bandwidth Memory was the most sensitive part and "only began showing irregularities after a cumulative dose of 2 krad(Si) — nearly three times the expected (shielded) five year mission dose of 750 rad(Si)." No hard failures were attributable to total dose up to 15 krad(Si) on a single chip. Those are total-dose results from an accelerator. In orbit the question is single-event effects: a particle flips a bit in memory and corrupts a value without an error the system notices. TechCrunch reports that ground testing suggested the impact "would be minimal for inference workloads and more of a problem for training a model." Inference reads fixed weights once per token. Training accumulates state over many steps, so one flipped value can spread.
What comes next
The 2025 paper planned "two prototype satellites by early 2027". TechCrunch now describes two more satellites in 2027 that will generate more power and test optical links between satellites, which Google needs for distributed workloads across a cluster. Google's economic case depends on launch prices that the paper projects could fall below $200 per kilogram by the mid-2030s, which is a projection, not a quote.
What remains uncertain. Google has not published on-orbit error rates, throughput or thermal data. The haiku Gemma 4 wrote is a demonstration, not a benchmark. The 15-minute and 20-minute figures come from different reports.
Analysis: For engineers, the lesson applies on the ground too. Silent data corruption is a real failure mode in large fleets, and inference tolerates it better than training only if outputs are checked. When Belgavi's Book Store runs a help-desk model, Geetha can add the same check the Suncatcher team needs: send a fixed prompt with a known answer on a schedule and compare the output. A model whose answers drift on a fixed input is a hardware question before it is a prompt question.
Sources and evidence
TechCrunch reported on October 30 that Google's Suncatcher prototype, built with Planet and launched in early October, is online and ran Gemma 4 after three weeks of commissioning; it carries four TPUs and up to 1 kW, but can power one chip for about 20 minutes at a time (NPR earlier said 15). Google's 2025 ground tests found Trillium HBM irregularities only after 2 krad(Si), and orbit now tests bit flips, which ground work suggested matter more for training than inference.