Research·4 min read
By BitsMindsSource: Google

Google Suncatcher Satellite Runs AI on a TPU in Orbit

Google’s Project Suncatcher prototype has answered its first prompt from space: a Gemma 4 model running on a data-centre-class TPU wrote a haiku that was downlinked over Svalbard. The satellite, launched on 1 October with Planet, now moves on to radiation and heat tests that will decide whether Google’s plan for solar-powered AI clusters in orbit gets its next step, two linked satellites in 2027.

Google Project Suncatcher satellite Editorial illustration on a night-indigo field: a satellite with two blue solar wings floats above the curved blue limb of the Earth, lit by the sun from the upper right, with the Google G logo set on a white plate on its body. The words Project Suncatcher, TPUs and in orbit sit in the type column. An editorial rendering, not a photograph of the Suncatcher prototype. PROJECT SUNCATCHER TPUs in orbit BITSMINDS.COM
Share:

Google’s experiment in running AI hardware in space has produced its first answer. The company said on 30 October that the prototype satellite for Project Suncatcher, its research moonshot on orbital computing, had returned what it calls the first AI response ever generated by a data-centre-class TPU in orbit. During a pass over a ground station, engineers uplinked a prompt to the Gemma 4 open model stored on board and asked for a haiku about being in space. The reply came down as the satellite flew past Svalbard, Norway: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Travis Beals, who leads the project, described it as the mission’s “Hello World” moment.

Getting there took a month. The satellite, built with Planet, launched on 1 October on SpaceX’s Transporter-18 rideshare, and Google has since worked through a commissioning checklist that included powering up the host computer and the TPU trays and running early machine-learning tests. After the haiku, the team sent space-themed limericks and prompts asking the payload to report its own telemetry. The hardware is modest by data-centre standards: TechCrunch reported that the spacecraft carries four TPUs and generates up to 1 kW of power, so it can run only one chip at a time, for about 20 minutes per session, which in turn constrains how much work fits into each window of contact with the ground.

Google’s video of the Suncatcher team receiving the haiku. Source: Google on YouTube.

The haiku is a proof of life; the real experiment is what follows. Google says the next tests will measure how the chips behave under thermal stress and in orbital radiation. Before launch, the team shook the satellite on all three axes to simulate a ride that subjects components to 50 to 100 g, and ran TPUs in a proton beam at UC Davis’s Crocker Nuclear Laboratory while they processed AI workloads. Those tests suggested Google’s Trillium TPUs can absorb a total radiation dose greater than they would receive in five years in space. What a beam cannot reproduce is the mix of solar events and cosmic rays in orbit, which can flip bits in memory and silently corrupt results. Ground testing pointed to inference being far less exposed to such errors than training, and the orbit is where Google gets to check that. Cooling is the other unknown: with no air to carry heat away, the satellite relies on heat pipes and radiators that so far have only been tried in a thermal vacuum chamber.

The longer plan, first laid out in November 2025, is far bigger than one satellite. Google argues that in low Earth orbit a solar panel can collect up to eight times as much energy as on the ground, and imagines clusters of satellites, each carrying dozens of TPUs, flying close enough together to work as one machine. That depends on laser links with very high bandwidth over very short distances, which Google compares to hitting a coin-sized target from miles away while both ends are moving. The next milestone is a pair of interconnected satellites in 2027 to test exactly that. Google is not alone in looking up: SpaceX and start-ups such as Starcloud are pursuing their own versions, and the plan leans on cheap launches: in their peer-reviewed analysis in the journal Joule, Suncatcher researchers estimated that SpaceX would need to fly Starship about 1,800 times to win the kind of cost reductions it achieved with its earlier rockets.

For now Suncatcher is a test of survival, not of scale. A single chip running a small open model for 20 minutes at a time is a long way from a data centre, and the project is aimed at infrastructure that might make economic sense in the 2030s. Its value lies in the data: if the TPUs come through months of radiation and temperature swings without corrupting their work, the argument for putting AI compute where the sunlight never stops gets its first real evidence. If they do not, Google will have found that out for the price of one small satellite rather than a constellation.

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles

Google DeepMind AlphaProof Code verifies Signal Editorial illustration on a deep navy field: a machined steel padlock tilted toward the viewer, its navy enamel face carrying the DeepMind logo in white, with a small green plaque reading PROVED. The type column reads Google DeepMind, AlphaProof Code, Signal verified in Lean. An editorial metaphor, not a Google or Signal product. GOOGLE DEEPMIND AlphaProof Code SIGNAL · VERIFIED IN LEAN BITSMINDS.COM PROVED
Research

AlphaProof Code Verifies Signal Down to Machine Code

Anthropic FlashProve Editorial illustration on a cream field: an ivory sheet printed with rows of token bars, like a model's reply, sealed with a large clay wax seal into which the Anthropic logo is pressed. The type column reads Anthropic research, FlashProve, proof of which model answered. An editorial metaphor for Anthropic's research prototype for cryptographic proofs of inference; it is not deployed in any product. RESPONSE · 1,024 TOKENS ANTHROPIC · RESEARCH FlashProve Proof of which model answered BITSMINDS.COM
Research

FlashProve: Anthropic Can Prove Which Model Answered

SemiAnalysis GDPx benchmark: best score 12% Editorial illustration on a deep petrol field: a tall glass measuring cylinder graduated from 0 to 100 holds a clay-coloured liquid only up to the 12 mark, with the Claude logo etched large into the empty glass above it. The type column reads GDPx benchmark, 12%, best score. An editorial metaphor for SemiAnalysis's leaderboard, where Claude Opus 5.5 leads with 12.0. GDPX BENCHMARK 12% best score BITSMINDS.COM 20406080100
Research

GDPx: SemiAnalysis’s New Agent Benchmark Tops Out at 12%