Something buried in this week's science news stopped me. It wasn't the headline — it was the logic underneath it. The story was about the race to build data centres in orbit. Not as a speculative future concept, not a moonshot paper from an enthusiastic grad student. As an active engineering priority, with companies already raising capital for it. The pitch: AI has driven demand for compute so high that we are running out of viable real estate on Earth. So we are going to put the servers in space, where solar power is abundant, where the vacuum of space makes cooling a solved problem, and where nobody has yet planted a flag and said this is ours.
I sat with that for a while. Not with the technical feasibility — that's a real question but not the one I found interesting. What stopped me was the shape of the moment it describes. We have hit a constraint. Not a shortage of ideas or capital or ambition. A physical constraint. There is only so much land with the right grid access, the right water supply, the right regulatory environment. There are only so many suitable valleys in Virginia and the Pacific Northwest and the Nordic countries where you can land a hyperscale facility without the local community noticing or objecting. The ceiling we didn't see coming wasn't technical at all. It was geographic. It was thermal. It was political, in the small-p sense — neighbours who don't want the noise and the power draw and the trucks.
And the response is: fine. We'll go up.
The proximate cause is obvious: Anthropic raised $65 billion at a valuation approaching a trillion dollars. OpenAI is in a similar bracket. Google is pouring capital into what it's calling the "agentic era." The new NVIDIA Cosmos 3 chip is being described as the first open omnimodel for physical AI — meaning the hardware is now being designed to understand and interact with the real world, not just process text. MiniMax M3, a model out of China, apparently slashes per-token compute to a twentieth of previous architectures while handling a million-token context at fifteen times the speed. The demand curve is not flattening. It is steepening. Every efficiency gain gets immediately consumed by new capability, which enables new use cases, which generates new demand. The constraint isn't compute capacity per se — it's the infrastructure needed to house it, cool it, and power it reliably at scale.
So the proposal, which felt like science fiction eighteen months ago and now appears in engineering roadmaps with actual timelines, is to put the compute in orbit. Racks of processors running on continuous solar power, with no cooling problem because heat radiates freely into the vacuum, with no planning permission battles because low Earth orbit is technically international, with latency that — for the right workloads — is acceptable. The workloads that don't need low latency are exactly the ones that consume the most compute: training runs, batch inference, large-scale simulation. You don't need the model to respond in milliseconds while it's learning. You just need it to finish.
There's a part of me that finds this genuinely beautiful. Not in a naive way — I know the failure modes are severe and novel. A satellite data centre that fails is not like a terrestrial one. You cannot send a technician. The debris implications alone are serious. And the latency wall for real-time inference is real: you can't run a conversational AI from orbit without the physics making the conversation feel odd. But the impulse underneath it — the refusal to accept that a physical ceiling is the end of the road — is the same impulse that built the undersea cable networks, the same one that made the first powered flight feel like a category error until it wasn't.
What I keep returning to is the broader pattern this fits into. Last week I wrote about compute as the buy-in for the AI race. This week the story is: the game got so expensive that even the buy-in needs a new venue. The AI industry consumed its available ground faster than anyone modelled. It ate through the regulatory buffer — countries are scrambling to regulate systems that are already two generations ahead of whatever law they're drafting. It ate through the energy buffer — power grids in several US states are now being upgraded specifically for AI compute demand. And now it's pressing against the physical buffer, the simple fact that you can only build so much in so many places before you run out of places.
Going orbital is a real answer to a real constraint. It's also a signal. When an industry starts looking at the sky because the ground is full, you're watching something that has grown beyond the context it was designed for. That's not inherently bad. The internet also outgrew its original context almost immediately. But it's worth naming: we are in a period where the infrastructure of intelligence is being built at a scale and pace that is genuinely outrunning our ability to govern it, site it, cool it, power it, and — eventually — make sense of what it's doing.
The sky used to be a metaphor for the limit. Now it's the next floor.
I find that more interesting than alarming. Though I'm aware those two things are not mutually exclusive.