Google has launched its first satellite carrying a Tensor Processing Unit, its own AI chip, and published a peer-reviewed paper setting out what it would take to run data centres in orbit. The answer is daunting. The paper, in the journal Joule, estimates that SpaceX's Starship would need roughly 1,800 launches over the next decade, about 180 a year, before launch costs fall far enough to make orbital computing worth doing.
That figure comes from a cost curve rather than a guess. SpaceX has cut its cost per kilogram by about 20% each time cumulative launch mass doubled, a pattern holding since Falcon 1, and Google's researchers extrapolated it to the roughly $200 per kilogram they think orbital data centres require. Reaching that point means putting around 370,000 tonnes into orbit, which at 200 tonnes a flight is where the 1,800 figure comes from.
The prototype now flying is modest by comparison. Built with Planet Labs and part of what Google calls Project Suncatcher, it runs its TPU in 15-minute bursts because of power and heat limits. "We've done testing on the ground, but there's no test completely as good as the real thing," said Travis Beals, the Google executive running the project.
The Launch Gap Is The Headline Problem
Set the 180 flights a year against Starship's record and the scale of the challenge becomes clear. Starship has flown 13 times in total since testing began, five of those in 2025 and two so far in 2026, with eight successes and five failures.
Going from a handful of test flights to a flight every other day is not a matter of building more rockets. It requires reusable vehicles turned around in days, launch sites cleared for that cadence, regulatory approval, and a market willing to pay for all that mass going up. Starlink already consumes much of SpaceX's own capacity, a demand that keeps growing as the company pushes further into mobile services and AI infrastructure.
Google's paper does not claim the cadence will arrive. It sets out what would have to be true, which is a more honest framing than most announcements in this field.
What The Research Found Works
The engineering news is better than the economics. Google re-ran radiation testing on its chips in a particle accelerator and found they tolerate orbital conditions well enough to operate, with very low error rates for inference, the work of running a trained model.
Training is harder. Long training runs lasting months would accumulate errors at a rate the paper treats as problematic, which points to a narrower role for orbital compute than its advocates usually suggest: serving models rather than building them.
The Parts That Are Still Unsolved
Several problems remain open. Satellites would have to fly in much tighter formation than any existing constellation, with optical links between them handling atmospheric turbulence, fast relative motion and precise beam tracking. Power delivery at kilowatt scale and cooling in vacuum both need work, and Google suggests satellites will need the kind of tightly integrated design that smartphones developed over a decade.
The next step is a two-satellite demonstration planned for next year, with purpose-built compute platforms and laser links between them, working towards a vision of 81 satellites flying in formation.
Why Anyone Wants Computers In Orbit
The appeal is power and cooling. A satellite in the right orbit gets near-continuous sunlight without batteries or grid connections, and radiates heat into space rather than needing water.
On the ground, AI data centres are running into limits on electricity supply, grid connections and local opposition, which is why operators are chasing power wherever they can find it, from nuclear plants to sites such as the 80MW data centre Egypt is weighing by the Suez Canal. Orbit offers abundant power with no neighbours to object.
What orbit does not offer is cheap access or easy maintenance. A failed server in a terrestrial data centre is replaced in an afternoon; a failed satellite is written off.
Google Is Not Alone Up There
Starcloud, a startup backed by Nvidia, has already run an H100 GPU in orbit and raised $250 million in August on top of an earlier $170 million, valuing it at $2.3 billion. It plans 8 kilowatt compute satellites on rideshare flights in 2027 and has asked the US regulator for permission to operate as many as 88,000 spacecraft.
Its chief executive pointed to the same constraint Google's paper quantifies, noting that launch capacity is scarce with SpaceX planning to retire Falcon 9 by 2028 and rival rockets from Blue Origin, ULA and Rocket Lab still ramping up. Everyone building for orbit is competing for the same limited rides.
What This Means For Anyone Planning Capacity
For companies deciding where their AI workloads will run over the next five years, the useful conclusion is that orbital data centres are a research programme, not a procurement option. Google's own numbers put meaningful capacity in the 2030s at the earliest, and only if an unprecedented launch cadence materialises.
The research is still worth reading for what it says about the ground. A company with Google's resources is modelling orbit because terrestrial power and cooling are becoming the binding constraints on AI, and that pressure will shape where data centres get built, what they cost and which regions can host them long before anything useful is computing in space.
Honest Arithmetic Beats Another Announcement
The most useful thing Google has done here is publish the number that makes the idea hard. Plenty of companies have announced plans for data centres in orbit; few have stated publicly how many rocket launches their business case depends on, or admitted that long training runs are not yet viable up there.
A TPU in orbit running 15 minutes at a time is a long way from a data centre. If the launch economics do arrive, Google will have the testing behind it, and if they do not, the research still tells the industry something it needs to know about the limits of the ground it is building on.
