Somewhere above the Arctic, a computer chip that normally belongs in a data center answered a very small question. Engineers on Earth asked for a haiku about space. A satellite running Google’s Gemma 4 model generated three lines and sent them back as it passed over Svalbard, Norway. Nothing dramatic appeared on the screen. Just a little poem.
I know, AI writing a poem is hardly news anymore. You can get one in five seconds while sitting at home. But this poem came from a data-center-class AI chip that had survived a rocket launch, switched on in orbit, received a prompt from the ground and sent back a reply. Google announced the result on October 30, 2026, after launching its Project Suncatcher prototype on October 1. That is a much stranger story than another chatbot update.
And it raises a question I never expected to ask seriously in 2026. Are we going to start putting AI data centers above our heads because running them on Earth is getting too expensive? Google is researching exactly that. The recent test is real, although the jump from one short poem to a working orbital server farm is enormous.
Access without medium partner: Google Put an AI Chip in Space
The First Reply Came From a Satellite, Not a Cloud Region
Google’s satellite launched aboard SpaceX’s Transporter-18 rideshare mission. The spacecraft was built with Planet, a company that already operates Earth-observation satellites. After contact was confirmed, Google’s team spent weeks commissioning the onboard host computer, powering up its Tensor Processing Unit hardware and checking whether the software could run after the trip into space.
Then came the small test. According to Google’s October 30 update, operators sent a prompt during a ground-station pass and asked Gemma 4 to write a haiku. The reply opened with the words “Blue marble below” and finished with a line about starlight. Google then tried more short poems and prompts that asked the machine to report details about its own equipment.
That might sound almost silly. Why spend money sending expensive AI hardware into orbit and ask it for poetry? Because the content of the answer was the least important part of the test. Google needed to show that the chip could boot, run a model, accept instructions, produce output and communicate that output back to Earth. Every step had to work while the equipment was travelling around the planet.
The machine itself is limited. NPR’s October 1 report described a refrigerator-sized satellite carrying four Google TPUs, with model runs limited to around 15 minutes because of heat. TechCrunch later clarified that the satellite can operate only one of those four TPUs at a time. This is a prototype for studying the hardware. It is nowhere near a cloud service you can sign up for.
Also, Google was not the first company to put AI hardware in orbit. Starcloud sent an NVIDIA H100 GPU aboard Starcloud-1 in November 2025 and says it later ran and trained small AI models there. Google’s narrower claim is about receiving an AI response from a data-center-class TPU in space. That wording matters. Otherwise, a good engineering result turns into a misleading headline.
I actually like that Google started with such a basic test. A polished demo could have hidden the messy parts. A simple prompt followed by a real answer gives everyone a clear starting point, and it leaves plenty of hard questions unanswered.
Why Google Is Even Thinking About Moving AI Off Earth
To understand Project Suncatcher, look at what is happening to data-center electricity demand. The International Energy Agency’s 2026 outlook estimates that data centers used about 485 terawatt-hours of electricity worldwide in 2025. Its central projection puts that number near 950 terawatt-hours in 2030, or roughly 3% of global electricity demand. AI-focused data centers are projected to grow faster than the sector overall.
That is a lot of power. And the problem isn’t only how much electricity the world can produce. A new data center needs a grid connection, transformers, cooling equipment and enough land for the building. Getting all of that approved and delivered can take longer than ordering the computing chips. A company can have billions ready to spend and still struggle to find a location where the power is available.
So Google is exploring a fairly wild alternative. Put solar panels and AI chips on satellites, then connect those satellites into a large computing system. Space has no local electricity utility asking you to come back after the next grid upgrade. In the right orbit, a panel can receive sunlight for almost the whole day.
Google says a solar panel in a suitable orbit could produce up to eight times more power than a similar one on Earth. Clouds and the day-night cycle reduce what ground-based panels receive. This figure is a research estimate for a future system with the right orbit and design, not a measured eightfold power advantage from the October prototype. That prototype follows a different orbit and uses batteries.
At first, this sounds like an easy win: more sunlight, less pressure on local grids. Then you remember the chips still need cooling, networking and repairs. And before any of those things can happen, every kilogram of hardware must survive a rocket ride. The free sunlight comes with an unusually expensive delivery charge.
The Most Annoying Problem Is Heat
Here’s a detail people often get wrong about space. It is cold out there, but that doesn’t mean a processor is easy to cool. Your desktop fan pushes warm air away from the chip’s heat sink. In a vacuum there is no surrounding air to carry the heat away. The heat has to move through solid parts or cooling pipes and then leave the spacecraft as infrared radiation from radiator surfaces.
Now picture a high-power AI chip doing calculations for hours. Its electricity eventually becomes heat. On Earth, data centers use fans, liquid loops, chillers and carefully designed airflow to get rid of it. On a satellite, the radiator needs enough area to release that heat into space. Bigger radiators mean more mass and more hardware to launch. Solar panels take up room too. The whole arrangement starts getting awkward very quickly.
That short 15-minute operating window reported for Google’s current prototype tells us something important. It doesn’t prove the chip is slow. It suggests the early spacecraft has a heat-management budget that limits how long a TPU can run before the system needs to cool down. Google has said it is testing heat pipes and radiators and wants real measurements of the hardware’s behavior in orbit. The exact long-run performance figures have not been published yet.
I found this part harder to wrap my head around than the solar-energy pitch. We see pictures of satellites against a black background and assume they can simply dump heat into empty space. But radiation cooling is a very different process from blowing air across a computer. The U.S. Government Accountability Office warned in April 2026 that cooling large orbital data centers remains unproven and could become a major obstacle.
And suppose the radiator gets damaged, the heat pipe stops working or a pump fails. You can’t send the usual data-center technician down the aisle with a replacement part. Repairs in orbit are a separate space-engineering project. Itself a fairly good reason to keep your cloud backup on Earth for now.
A Haiku Is Not a Speed Benchmark
Google has confirmed that a model ran and produced a response. It hasn’t released the details that would let us compare this machine with a small terrestrial server. We don’t know the exact Gemma 4 configuration, how many tokens per second it generated, how long the round trip took, or how much energy the request used. We also don’t have a published record of how it behaves under thousands of requests or repeated thermal cycles.
For normal AI users, those numbers matter. Imagine waiting for a model to finish warming up each time you send a prompt. Imagine a service that can process requests for a few minutes, then has to cool down before taking another batch. It may still be useful for a satellite experiment, but you wouldn’t want to replace a working API with that setup.
There is a possible upside which doesn’t need giant orbital data centers at all. Satellites already gather images and sensor readings. If they can process more information on board, they could send down useful results instead of transferring every raw byte to a ground station. That could help with Earth observation and remote sensing, especially when a satellite has limited communication time. It is a narrower use case, and probably an earlier one, than serving millions of chatbot requests from orbit.
For now, Google’s test only shows a successful end-to-end interaction. The company hasn’t claimed a price, public access, production reliability or an inference-speed record. I wish more AI announcements were this clear about what has actually been tested.
The Bigger Plan Needs Satellites Talking to Other Satellites
A modern AI data center doesn’t run as one giant chip. It uses racks of processors connected with very fast networks. The chips exchange model data and intermediate results many times while handling a workload. If one part waits too long for another part, much of the expensive hardware sits idle. So even if an individual TPU survives in orbit, Google still has the job of making many satellites work together.
Its proposed answer is optical links. Think of tightly controlled laser connections between nearby spacecraft. In a peer-reviewed paper published in Joule on October 1, Google researchers studied a possible group of 81 satellites flying near one another, around a central altitude of 650 kilometres, within a cluster roughly one kilometre across. It is a mathematical design study. Google has not launched an 81-satellite AI cluster.
The paper estimates that some future links would need aggregate bandwidth on the order of 10 terabits per second. In a lab, the researchers showed 800 gigabits per second of data transfer in one direction over a short free-space optical path. That’s useful work, but an optical link across a laboratory bench is far easier to keep aligned than a high-speed connection between moving spacecraft.
And the spacecraft cannot simply fly wherever they like. They are travelling through Earth’s gravity field, and slight differences in position change their paths over time. They need precise formation control, collision avoidance and reliable communication while moving at orbital speed. Google is considering software models to help keep the formation stable. I can see the attraction of the idea, but the list of things that must go right is getting quite long.
The next proper test is already on Google’s roadmap. It plans to put two interconnected satellites in orbit in 2027 to study the networking side. That’s a more interesting milestone to watch than another poem. If the two spacecraft exchange real model data at useful speeds, we’ll have something concrete to discuss about distributed AI computing in space.
Radiation Doesn’t Care Whether Your Code Passed Its Tests
Heat isn’t the only problem. Space radiation can damage electronic parts and also cause small, temporary errors. For ordinary software, a flipped bit might cause a crash. During AI training, an undetected error might damage calculations without the system noticing immediately. You could run an expensive job and discover later that the result is unreliable.
Google’s research team tested Trillium TPUs under radiation on Earth. The paper estimates a five-year orbital dose of about 750 rad for its proposed environment. In lab tests, its high-bandwidth-memory subsystem began showing problems after about 2,000 rad, while other workloads continued operating at considerably higher cumulative doses. That doesn’t mean an entire satellite is guaranteed to survive five years. It means the chips met some important early test conditions.
The researchers also estimated a very low rate of radiation-related silent errors for certain inference workloads, around one per three million inferences under their stated assumptions. Treat that as a model-based estimate from laboratory data, not a measured error rate for the October spacecraft. The effect on sustained training workloads needs more study, and real space conditions can show faults that ground tests missed.
This is one reason I’m doubtful that huge training clusters will be the first useful product of orbital AI. Training runs can last a long time and require tightly linked processors. A short local inference job on one satellite is much easier to manage, even if the model gives the occasional strange reply. The difference becomes expensive when you’re talking about a model-training run that takes weeks.
Google’s Money Calculation Has a Very Big ‘If’ Attached
Now comes the price. Google’s paper explores a future where launch costs to low Earth orbit drop to around $200 per kilogram by the mid-2030s. Under one set of assumptions, it calculates that the launch-related cost of supplying a kilowatt of satellite power, spread across five years, could reach roughly $810 per kilowatt per year. The researchers compare that with a reported $570 to $3,000 per kilowatt per year of power costs for terrestrial AI data centers.
Those figures sound close enough to make the idea worth studying. But they are not a complete cost comparison between a space data center and one on Earth. They don’t prove that operating a satellite network would be cheaper after you pay for compute hardware, radiators, networking, replacements, ground stations, insurance and the work of managing a fleet in orbit. A lot of the bill is still missing.
The $200-per-kilogram figure is also a projection. The study’s learning-curve model suggests that getting there through repeated heavy launches could require about 1,800 Starship-class flights in total over the relevant growth period, roughly 180 a year on average. That is an estimate of what it might take for the launch industry to reach a particular price level. It does not mean Google has booked 1,800 rockets for Project Suncatcher.
Even Google’s own project lead has been cautious. In NPR’s October interview, Travis Beals said he did not expect space computing to become the cheaper choice within the next five years. A research project can still be sensible when the commercial case is ten years away. The trouble begins when people describe a cost model as though the savings are already sitting in the bank.
There is another bill worth thinking about. The GAO points to orbital collisions and the growing problem of debris. Satellites wear out, and failed ones must be removed safely. Rockets and re-entry have environmental effects of their own. Sending compute away from populated areas might reduce some pressure on land and local water supplies, but it doesn’t make those wider costs disappear. Google’s paper itself says that a full life-cycle environmental assessment is outside its scope.
I Would Watch the Next Two Satellites, Not the Next Poem
There is something genuinely fun about an AI model writing a haiku from above the Earth. I can imagine the engineers watching that first reply arrive and feeling relieved that all the parts worked. After weeks of checks, even a few lines of poetry must have felt good. The rest of us can enjoy the weirdness without pretending a full orbital data center has suddenly appeared.
But I wouldn’t move my AI workloads into space anytime soon. We still need evidence on sustained chip performance, power and cooling, fault rates, optical networking and what one useful request actually costs. Google has given a clear first result. Now it needs the boring numbers, the same numbers you’d ask for before buying a mini PC to run local models.
The 2027 two-satellite test should tell us more. I expect space-based computing to find practical jobs near satellites and space sensors before it becomes a real rival to ground data centers. That’s an opinion, of course, and Google is aiming much bigger. Its researchers are asking what a large connected system might look like once launch prices and hardware improve enough.
For now, I keep coming back to how small that first task was. A satellite with data-center hardware crossed the sky, received a sentence from Earth and answered with a poem. The next achievement won’t be about writing better poetry. It will be showing that the machine can keep working when the easy part is over.