Somewhere above the poles, a small satellite carrying Google's custom AI chips has done something that looks almost whimsical: it composed a haiku about its own situation and sent it home. Beneath the playfulness lies a quieter human question, one that people have long met with dismissal, namely whether the machinery we build to think can survive the hostile emptiness beyond the atmosphere. The early answer is a cautious yes, with heat and radiation still standing as the hard problems that every orbital ambition must eventually reckon with. What is taking shape is less a finished promise than
Google's Project Suncatcher Sends AI-Written Haiku From Orbit as Space Data Centres Take Shape
The most complex and expensive haiku ever written
So the headline is a poem. Why does a haiku matter?
Because it proves the chain works end to end. A TPU in orbit ran Gemma 4, produced output, and sent it to Earth. Google calls it the 'Hello World' milestone.
Worth saying what it doesn't prove. A haiku is a tiny workload. Nothing here shows a data centre's worth of computing running reliably.
What's the biggest obstacle they describe?
Heat. The TPUs can produce over 100,000 watts per square metre, and the radiators shed around 300. So the satellite pauses every 15 minutes to cool.
And that pause is a real limit on how much work it does. The article doesn't say how much throughput is lost, so I'd treat the economics as unknown.
What about radiation?
Bit flips and silent data corruption. Engineers reproduced the corruption in lab tests, but the satellite is still being monitored in orbit.
Right, so the long-term survival question is open. It's only been weeks, and the piece doesn't report results from that monitoring.
Is Google ahead of the field?
Starcloud flew an NVIDIA H100 first. Google's argument is that TPUs will scale better.
That's Google's belief, not a demonstrated result. The real test is the second satellite next year, linking them together.
Le Pouls
- A poem written by an AI model in orbit is a playful surface, but it masks a real test of whether delicate computing hardware can endure launch forces, radiation, and the silence of space.
- Heat is the sharpest tension: chips can produce more than 100,000 watts per square metre, while radiators shed only about 300, forcing the satellite to pause every 15 minutes.
- Engineers are reproducing radiation-driven data corruption in labs and tracking the satellite's health in its 800km pole-to-pole path, navigating a problem that has no air to carry heat away.
- Rivals including Starcloud, SpaceX, Blue Origin, and European projects are racing into the same orbital territory, raising the stakes for proving the concept.
- The trajectory now points to a second satellite next year, where engineers will test whether linked satellites can operate together as a single orbital compute constellation.
Somewhere above the poles, a small satellite carrying Google's custom AI chips has done something that looks almost whimsical: it composed a haiku about its own situation and sent it home. Beneath the playfulness lies a quieter human question, one that people have long met with dismissal, namely whether the machinery we build to think can survive the hostile emptiness beyond the atmosphere. The early answer is a cautious yes, with heat and radiation still standing as the hard problems that every orbital ambition must eventually reckon with. What is taking shape is less a finished promise than the first honest step toward thinking machines that draw their power from the sun itself.
On 1 October, a small Google satellite rode a SpaceX rideshare into orbit. Several weeks later, researchers asked the Gemma 4 AI model running aboard it to write a haiku about being in space, and the three lines came back to Earth: "Blue marble below / Data streams in silent grace / Starlight feeds my core." Google calls this the 'Hello World' milestone for Project Suncatcher, its effort to run AI data centres in orbit.
The satellite was built with Planet, whose engineers helped stress-test the compute modules before launch. Each satellite carries dozens of Google's Trillium TPUs, and the idea is to draw on the nearly constant solar energy available in space and eventually link many satellites into a larger computing network. Travis Beals, the Google senior director leading the project, admitted that people often dismiss the concept at first, and that his team's job was to find the reason it would fail. Their answer so far is that it can work.
The hardest problems are physical. Radiation can flip bits and corrupt AI calculations, and the team has reproduced that failure in the lab. Heat is worse. With no air to carry it away, the chips must shed heat through radiators that manage only about 300 W/m2 against more than 100,000 W/m2 from the processors, so the satellite pauses every 15 minutes to cool.
Google is not alone. Starcloud has already flown a conventional GPU, and many companies, including SpaceX and Blue Origin, are moving toward orbital computing, partly driven by worries about terrestrial energy rules. Planet's Eric Stevens called the project an experimental moonshot. The next test arrives next year, when a second satellite will show whether linked satellites can function as one.
On 1 October, a small Google satellite went up on a SpaceX rideshare flight, and several weeks later it did something deliberately modest: it wrote a poem. Researchers asked the Gemma 4 generative AI model, running on the satellite's onboard processors, to compose a haiku about being in space. It came back to Earth as three lines: "Blue marble below / Data streams in silent grace / Starlight feeds my core."
Google is calling the transmission the 'Hello World' milestone for Project Suncatcher, its prototype for running AI data centres in orbit. The satellite was launched with the space technology firm Planet on the Transporter-18 rideshare mission and placed in orbit around Earth's poles. Travis Beals, a senior director at Google who leads the project, announced that "our chips have officially earned their space wings." A photo released by Google shows the joint team celebrating what it called the most complex and expensive haiku ever written.
The idea behind Suncatcher is to tap the near-constant solar energy available in orbit. Each satellite carries dozens of Google's Trillium Tensor Processing Units, or TPUs, the company's purpose-built AI chips. Google envisions linking many such satellites into compute constellations that would eventually function as large-scale data centres. It plans to test that configuration next year, after launching a second satellite.
Beals said that most people who hear the concept react by calling it crazy and assuming there must be a reason it can't work. His team's first task was to find that reason. "It turns out that it is possible," he said.
Getting there took heavy engineering. Google worked with Planet, which operates more than 200 imaging satellites, to stress-test the compute modules before launch. Engineers watched how the hardware handled liftoff, which subjects it to up to 100 times the force of gravity, and are now tracking whether it survives an orbital environment full of space junk, solar flares and radiation. Radiation can cause 'bit flips' that disrupt calculations, and ionising radiation can trigger 'silent data corruption' events that affect AI workloads. Google and Planet engineers managed to reproduce that phenomenon in lab testing. The satellite orbits at 800km on a pole-to-pole path chosen to maximise sun exposure, and engineers are now monitoring how it copes with radiation and thermal stress.
Heat is the particular problem. Space is cold, but with no air, heat does not drift away from components as it does on Earth. Suncatcher uses heat pipes, closed pumped fluid loops and flat panel radiators to draw heat off the TPUs and emit it into space as infrared radiation. The numbers are lopsided: TPUs can generate over 100,000 watts of heat per square metre, while the radiators dissipate only around 300 W/m2. As a result, the satellite has to pause every 15 minutes to let the radiators shed the excess.
Google is not first to put AI hardware in orbit. Private firm Starcloud launched a conventional NVIDIA H100 GPU last year that was able to run Gemma. Google argues that its specialised TPU design will scale more easily and efficiently. Orbital solar panels can collect five to eight times as much power as ground-based ones, since there is no atmosphere to weaken the light, and cheaper, faster launches are speeding deployment. Planet's director of systems engineering, Eric Stevens, said the approach lets engineers take advances from the commercial industry and send them to space immediately, rather than waiting through long design cycles to make them 'space rated'.
Others are circling the same territory. SpaceX CEO Elon Musk has talked up orbital data centres, posting that they are "looking pretty appealing right now," which the source links to concern that Democrats could impose restrictive planning and energy obligations on tech firms. Projects expected to move in coming years include Blue Origin's Project Sunrise, SpaceX's Starmind, and efforts from Orbital, Lonestar, Axiom Space, Cowboy Space Corporation, Satlyt and the European Space Agency's ASCEND. Some people have also raised environmental concerns about the idea.
Stevens said his first reaction was that the project seemed big even for Planet, but called it an experimental moonshot and an important first step toward moving data centres into orbit. "Everything seems too hard until someone tries it," he said. The next test is the second satellite, due next year, and whether linked satellites can work as one.
Citations marquantes
Everyone's reaction when they hear this idea is 'that's crazy'... The first thing we tried to do is to find that reason, and it turns out that it is possible.— Dr Travis Beals, Google senior director and Project Suncatcher lead
Everything seems too hard until someone tries it.— Eric Stevens, Planet director of systems engineering