Google’s AI Satellite Wrote a Haiku From Orbit
Key Takeaways
- Google Project Suncatcher produced its first AI response from orbit on 30 October 2026, when a data-centre-class tensor processing unit on a prototype satellite ran Google’s Gemma 4 model and returned a haiku to a ground station.
- The satellite launched on 1 October 2026 aboard a SpaceX Transporter-18 rideshare mission, was built in partnership with Planet, and carries four TPUs powered by solar arrays rated at roughly 1 kilowatt.
- The haiku — “Blue marble below / Data streams in silent grace / Starlight feeds my core” — was received as the satellite passed over Svalbard, Norway, and Google calls it a “hello world” milestone, not a product launch.
- Power is the binding constraint: reports say only one of the four TPUs can run at a time, for roughly 15 to 20 minutes per session, while heat and radiation tests have not yet begun.
- Google’s next step is a pair of interconnected satellites in 2027. Orbital data centres remain an experiment, and nothing about this test changes what Gemini users get today.
Introduction
The first AI answer from space was a poem, and that is exactly the point. On 30 October 2026, Google announced that its Project Suncatcher prototype satellite had run an AI model in orbit and sent a complete response back to Earth. The prompt asked for a haiku about being in space. The model, Gemma 4, answered with three lines about a blue marble, silent data streams and starlight. It is a small piece of writing, and a large engineering claim: a chip designed for a liquid-cooled data centre rack survived launch, powered up in orbit, ran a model, and delivered its output to a ground station.
Google chief executive Sundar Pichai announced the milestone in a post on X, calling it a “hello world” milestone for Project Suncatcher. That phrasing is deliberately modest, and it is the right frame for reading the story. In software, a “hello world” program proves that the toolchain works end to end; it does not prove that the system is fast, cheap, reliable or useful. The same is true here. The haiku proves the chain — satellite, chip, model, downlink — works. It proves nothing yet about speed, cost, durability or scale, and Google’s own announcement gives no figures for any of those.
What Google actually ran in orbit
Project Suncatcher is Google’s long-term research effort to test whether space could one day host AI computing infrastructure. The prototype satellite, called MVP, launched on 1 October 2026 and was built in partnership with Planet, the satellite imaging company. After launch, the team spent weeks commissioning the spacecraft: powering up the host computer and the TPU trays, then running early machine-learning tests.
The hardware matters because it is not a small, ruggedised space chip. The satellite carries four tensor processing units — the chips Google designs to run AI models in its data centres — described by Google as data-centre-class TPUs. Travis Beals, Google’s senior director for Paradigms of Intelligence, wrote in Google’s announcement that with the host and TPU trays powered up and machine-learning tests completed, the onboard Gemma 4 open model achieved its “hello world” milestone by sending an inference response from orbit. Google describes the result as the first AI response from a data-centre-class TPU in orbit, a claim that is specific enough to be meaningful: it is not the first computation in space, and not the first AI model run off Earth, but it is, on Google’s account, the first time this class of data-centre AI chip has produced an inference response in orbit.
The test itself was simple by design. During a ground-station pass, engineers uplinked a prompt asking Gemma 4 to write a haiku about being in space. As the satellite passed over Svalbard, Norway, it downlinked the reply: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Google says it then tested the system further with space-themed limericks and prompts asking the model to report the technical status of the satellite’s own systems. A Google spokesperson told TechCrunch that the satellite is operating nominally, which in this context means the spacecraft is behaving as expected after its first AI run — not that it has passed any durability test.
The constraints the haiku does not show
The most important numbers in this story are the ones that describe limits, and they come from reporting around Google’s announcement rather than from Google’s blog post itself. The satellite generates up to about 1 kilowatt of solar power. TechCrunch reports that only one of the four TPUs can run at a time, and only for about 20 minutes at a time; other coverage puts the figure at roughly 15 minutes, citing heat-management limits. The honest way to state it is that single-TPU sessions last roughly 15 to 20 minutes, which is a severe constraint for any workload that needs sustained compute, and a reminder that a satellite is a power- and heat-limited box, not a data centre hall with a roof removed.
Heat is the deeper problem. On Earth, a data-centre TPU sits in a rack with liquid cooling and effectively unlimited power and airflow. In orbit there is no air to carry heat away, and a chip that runs hot must radiate that heat into space or shut down. Radiation is the second problem: high-energy particles can flip bits and corrupt computations, and a chip designed for a shielded building has not yet shown how it behaves after months of exposure. Google says the next phase will test how the TPUs perform under thermal stress and radiation exposure. Those tests have not started, which means the durability question — the question on which any orbital data centre idea actually stands or falls — is still entirely open.
That is why the commentary around the announcement has been careful. One analysis put it well: the haiku is a boot test, not a benchmark. It shows that a TPU survived launch, powered up, ran a small open model from prompt to answer, and got the reply home. It shows nothing about throughput, cost per token in orbit, error rates under radiation, or how long the chips last. Google’s announcement gives no figures for speed, cost or longevity, and any article that converts this test into a claim about orbital data centres replacing earthbound ones is inventing the next three chapters of a story that has only published its first page.
Why Google is even trying this
If the constraints are so severe, why bother? Google’s answer, given when it announced the project, is about the long term. The company has argued that space offers two things Earth cannot easily provide at data-centre scale: abundant solar energy, and a naturally cold environment for rejecting heat once the engineering is solved. It has also pointed to falling launch costs, and has suggested that if launch prices keep falling, the economics of putting compute in orbit could change over a decade or more — one report notes Google’s expectation that SpaceX could reach the relevant cost levels by around 2035. That is a forecast about a market in 2035, not a plan for 2027, and it should be treated as such.
There is also a strategic logic that has nothing to do with poetry. The companies that build AI infrastructure are all, in different ways, running into the same earthbound limits: power availability, grid connection queues, water for cooling, planning permission and community opposition to new data centres. Testing whether some of that infrastructure could one day move off Earth is a hedge against those constraints, and a way to learn, cheaply and early, what breaks when data-centre hardware leaves the data centre. A single prototype satellite is a cheap education compared with a gigawatt data centre campus, and the lessons — about thermal design, radiation tolerance and autonomous operation — are useful even if no orbital data centre is ever built.
The next milestone Google has named is concrete enough to check: two interconnected satellites in 2027. Interconnection matters because a single satellite can only talk to a ground station during a pass, while a constellation could hand work between satellites and keep processing going. Until that step is demonstrated, Suncatcher remains a single node proving it can answer, not a network proving it can serve.
What this does not change
Nothing changes yet for anyone who uses Gemini, and Google has not claimed otherwise. The model that ran in orbit was Gemma 4, Google’s open model, loaded onto the satellite before launch — not a live Gemini service, and not a system that users can reach. No pricing, latency or availability claim follows from this test, and no reader should expect their AI assistant to get faster, cheaper or greener because a satellite wrote a haiku.
It is also worth separating Suncatcher from the wider space industry. Planet built the spacecraft, SpaceX launched it on a rideshare, and the ground station at Svalbard received the reply. Those are established capabilities being used for a new experiment, not new capabilities invented for it. The novelty is narrow and specific: a data-centre AI chip, running a real model, producing a real answer, in orbit, and getting that answer home. Narrow novelties are how hard engineering problems get solved — one link in the chain at a time — but they should be reported as narrow, not inflated into revolutions.
The sensible way to follow this story is to watch for the test results Google has promised, not the poetry it has already published. Do the TPUs survive thermal cycling? What error rates does radiation produce, and can they be corrected cheaply? Can two satellites in 2027 actually pass work between them? Those answers, when they arrive, will tell us whether Suncatcher is a curiosity or the first step towards a genuinely new place to put computers. Until then, the haiku stands as what it is: a small poem, and a surprisingly difficult engineering trick, from a machine a few hundred kilometres up.
For context on Google’s wider AI push, see our guide to Google Search AI Mode and how it works, and for the chip race behind it, our analysis of US and China AI competition in 2026.
