Everybody's Racing to Orbit. Nobody's Funding the Radio.

Via Satellite published a piece this morning asking whether orbital data centers are the next frontier of AI infrastructure or just another lap around the hype track. Anne Wainscott-Sargent talked to the people flying hardware, the people writing the checks, and one guy who keeps insisting everyone is solving the wrong problem.

That guy was me.

I sit in a useful chair for this argument. I spent seven years in the Navy doing cyber, some of it at NSA, and I now work on security for one of the largest data center operators on the planet. I have watched, up close, what it costs to move data between machines that are supposed to cooperate. So when the headlines fill with GPUs in orbit, I look past the thermal math and straight at the radio.

A few months ago I wrote two long pieces on exactly this, collectively titled Spaceballs the Datacenter, because if you’re going to spend eight hours doing Stefan-Boltzmann equations for fun, you may as well name the project after Mel Brooks. Part one tackled heat. I calculated that one B200 GPU in orbit needs a radiator roughly the size of a beer pong table just to keep from cooking itself. Part two chased the thing the heat math distracts everyone from. Bandwidth.

The wall has a number.

A modern NVIDIA B200 moves about 7.2 terabits per second to the GPU sitting next to it inside a training cluster. That speed leaves no slack; starve the link and the whole cluster stalls. The best optical link between two satellites in production today, the kind Starlink flies, moves about 100 gigabits per second. The best anyone has demonstrated in a lab, a Chinese single-channel record, reaches 400. Do the division. Production orbital links run about seventy times too slow. The lab record only narrows it to eighteen. That is one to two orders of magnitude, and no amount of venture capital repeals it this decade.

The consequence is simple and brutal: the satellite is the cluster. If two GPUs can’t talk fast enough across the vacuum between two satellites, then each satellite is a sealed island of compute. You can train a model that fits inside one of them, which today tops out around 70 billion parameters, a respectable model from roughly 2022. The pitch of a frontier training run spread across tens of thousands of orbital GPUs needs those islands to merge into a continent. The radio won’t let them.

I ran five business models for orbital compute through a spreadsheet with more than eighty parameters: AI training, AI inference, public cloud, edge and content delivery, and sovereign cloud. Four of them die. Training dies on the bandwidth wall. Inference and public cloud die on economics, because you’re competing with a data center in Virginia that has cheap power, elastic scaling, and a loading dock. Edge compute genuinely breaks my heart, because it’s the best technical fit and the worst business: content delivery already prices at pennies per gigabyte on the ground, and you cannot ship pennies to orbit and back.

One model survives. Sovereign cloud.

Sovereign cloud works for a reason that has nothing to do with performance, which is precisely why it works. Every other use case asks orbit to beat the ground at the ground’s own game. Sovereign cloud asks orbit to do the one thing the ground physically cannot: put a nation’s most sensitive data somewhere no foreign agent can walk up to it. A prosecutor with a warrant can enter a building in Frankfurt. An operative with the right badge can reach a server room in Singapore. A GPU 550 kilometers up, registered to your government under the Outer Space Treaty? Nobody is walking into that. That inaccessibility is the product. The teraflops are incidental.

The believers in the Via Satellite piece aren’t wrong about everything, and they deserve their due. Chris Stott at Lonestar notes that we called satellite broadcasting and high-throughput satellites hype cycles too, right up until they became the most profitable things we do in orbit. He’s right. The engineering really is, in his phrase, a known known. Avi Shabtai at Ramon.Space is right that the power is free and the cooling is free once you’re up there. James Mason at Planet is right that there’s no new physics required. I don’t dispute any of it. My worry is narrower, and I think more dangerous: that all this capital chases one more GPU-on-a-test-bed demo, ships a gorgeous press release, and steps right over the unglamorous communications layer that decides whether any of it becomes a business.

A word, then, to anyone writing checks. Don’t fund another radiator. Heat is solvable with money and patience. Fund the radio: terabit-class optical links between satellites, terminals that weigh five kilograms instead of fifteen, ground stations that can actually catch an optical downlink through weather. That’s where the wall is. And walls are where the moats are.

In Spaceballs, the villains build Mega Maid, a spaceship the size of a planet that turns into a vacuum cleaner, and it still runs out of air halfway through the job. The orbital data center crowd is smarter than that, mostly. But the failure mode rhymes. You can build the most magnificent constellation in history, solar-powered, radiatively cooled, bristling with GPUs, and if the satellites can’t talk to each other, you’ve launched a thousand islands and called it a continent.

The satellite is the cluster. Until we fund the radio, the cluster is one satellite wide.

May the Schwartz be with the comms engineers. They’re the ones who decide whether this works.

Go read the whole debate first. Anne Wainscott-Sargent put the believers and the skeptics in one room, with the people flying hardware and the people funding it: Are Orbital Data Centers the Next Frontier of AI Infrastructure? (Via Satellite)

Then, if you want the full math, all the physics, and all the snark behind my answer: Spaceballs the Datacenter 2: The Search for More Bandwidth.