Executive Summary
- Jack Collier, Chief Growth Officer at io.net makes a case against orbital data centres and how it’s going to make AI access a whole lot worse. SpaceX went public in the largest IPO in history, and within days, its value went past $2 trillion. This all rides on the idea that the future of AI compute is in orbit.
- Orbital compute means trillions in capital, increasingly raised through debt and underwritten by insurance models nobody has validated, and those costs eventually land on whoever rents the compute.
- It’s a bizarre idea where they are spending billions to put compute in orbit when so much of it is already sitting idle down here, waiting to be switched on. Maybe they should start there.
Wall Street is betting billions that AI’s future is in orbit. It’s the wrong bet, and it will shut out the very people who need compute most. Last month, SpaceX went public in the largest IPO in history, and within days the market had pushed its value past $2 trillion. A good chunk of that bet rides on a single, audacious idea: that the future of AI compute is in orbit. SpaceX has filed with the US Federal Communications Commission to launch up to a million solar-powered data-centre satellites; Blue Origin has filed for more than 51,000 of its own. Wall Street, it seems, has decided that the way to solve the AI compute crisis is to leave the planet.
The trouble is that it gets the problem backwards. For all the talk of a compute shortage, the real bottleneck was never physical room to build. It was access: who can actually get hold of compute, at a price they can afford, when they need it. And orbit is about the least accessible place you could put it.
The pressure on Earth-based data centres is real. Power and land are scarce, water and cooling are limiting factors in more and more places, and local opposition is mounting. Yes, we’ll need to build more. But orbit is the wrong answer. Putting compute in space is wildly expensive, and it does nothing for the people who can’t get access today. If anything, it shuts them out further.
The maths doesn’t close. The economics are the place to start. By SpaceX’s own numbers, one of these first-generation satellites puts out about 120 kilowatts – roughly what a single high-end server rack pulls. A big facility on the ground? Hundreds of megawatts. So a satellite manages maybe a hundredth of what an Earth-based data centre does, quite possibly far less, and it costs a fortune to keep it alive up there: you are launching it, shielding the chips from radiation, servicing it in orbit and dumping its heat into a vacuum.
And the doubters aren’t only those of us with something to sell on the ground. Amazon’s own cloud chief, Matt Garman, says there simply aren’t enough rockets to launch a million satellites yet, and OpenAI’s Sam Altman, no stranger to a big compute bill, called the whole idea “ridiculous” for now, reckoning it won’t count for much this decade. Even the people whose job it is to price that risk can’t do it yet. Space insurers have openly warned that there is no track record for this kind of infrastructure, and that the market is still relying on policies written for an earlier generation of satellites. One analysis put the cost of a single gigawatt of orbital compute at more than $40 billion. For all the ambition, this is a moonshot with a data-centre logo on the side.
Then there’s the question of who pays. Orbital compute means trillions in capital, increasingly raised through debt and underwritten by insurance models nobody has validated, and those costs eventually land on whoever rents the compute. The hyperscalers can absorb them. The startups, the researchers and the developers in emerging markets cannot, and they are the ones already being priced out. Demand for AI infrastructure has been growing exponentially while supply crawls along in a straight line, and when access gets tight it is always the smaller players who lose out first.
Building the most expensive infrastructure ever conceived and then passing the bill down the chain will only push more of them out of the market.
It also deepens the problem sitting underneath all of this: concentration. Three cloud providers already control around two-thirds of the global cloud infrastructure market, a small club at the top with everyone else stuck on a waiting list. Orbital data centres don’t prise that club open. They extend it into space. The same handful of companies that own the launch capability would own the orbital compute capacity too, which puts the whole market even further out of everyone else’s reach. There is nothing democratising about that.
What makes the orbital gold rush so frustrating is that we don’t actually need it to solve the access problem, because the compute is already here. Most of what’s already been bought sits idle. One 2026 study of 23,000 enterprise clusters found average GPU utilisation of just five per cent. Think of the clusters bought for a training run that finished months ago, the independent data centres running well below capacity, the university labs, even consumer hardware. The capacity is there. What’s missing is anyone joining it up.
The compute is already here
The next phase of AI infrastructure is about orchestration, not sheer scale. The advantage goes to whoever can take the latent, underused compute already out there and pool it, so it can be tapped anywhere, on demand. Decentralised, distributed networks are built to do exactly that. They gather up the fragments of capacity that already exist and turn them into something that grows with demand instead of straining against it.
There’s no 12-to-24-month buildout involved, and no launch window to wait for. A developer can provision a cluster in minutes and spread the workload across many locations and devices, so that if one goes down the job just moves somewhere else. It’s cheaper and faster than the alternative, and crucially, it’s available now.
None of this is an argument against building on Earth. We will keep building, we have to. But the industry is starting to change shape. For years, the working assumption has been that whoever builds or launches the most will win, and I’m not sure that holds any more.
What matters now is making better use of what we already have: pulling the scattered models, providers and machines that never get joined up into a single usable pool. A mosaic of infrastructure, in other words, spread across the world and coordinated properly.
The hyperscaler era is levelling off. And there’s something faintly absurd about spending billions to put compute in orbit when so much of it is already sitting idle down here, waiting to be switched on. Start there.
