Surveyed · 22 August 2026 · not a facility

Orbital compute architectures

Orbital compute is not one machine. It is a stack of proposed layers, a thermal problem, and a set of filings. This sitting cites that stack. It does not claim a satellite, a terawatt, or Terafab.

The stack

1 · On-orbit edge Infer on the spacecraft so raw sensor bits do not all come home. Starcloud-1 (November 2025) carried an Nvidia H100. Public engineering note: the radiator was too weak for full power.
2 · Constellation cloud Many small compute satellites, optical inter-satellite links, a Starlink-class backhaul. FCC paper in 2026: SpaceX up to one million; Starcloud 88,000; others in the tens of thousands. Filings are not a constellation.
3 · Large orbital plant Gigawatt-class, large solar, “data centre in space.” White papers and conference claims. Not a built plant.
4 · Lunar / far-side store Cold archive. Different physics, different job. Lonestar-class demos. Not training.
5 · Earth silicon for orbit Terafab / D3 as described in public commentary: logic, memory and packaging on Earth for FSD, Optimus, and orbital nodes. A United States manufacturing sentence. Not an Australian berth.
This sitting · catch, analogue, local brain If orbital compute becomes real, someone still recovers, refurbishes and fuels. Delay is physics: twenty minutes to Mars, and a dark continent that is still Country. CrystalCore is labelled, fail-closed, memory off the model. It is not an H100 in LEO.

Physics that designs the architecture

Space is cold. Vacuum is a poor coolant. There is no convection. Heat leaves only by radiation. Power rejected scales with radiator area and the fourth power of temperature.

A one-megawatt orbital box needs on the order of a hockey-rink of radiator at modest chip temperatures (World Economic Forum / Starcloud paper, 2026). The ISS rejects about 70 kW with tonnes of radiator. Scale that to a megawatt and the radiators can outweigh the GPUs. Run chips hotter: smaller area, worse leakage and errors. That is the trade. “Space is free cooling” is not an architecture.

Also binding: total-dose and single-event radiation; eclipse versus sun-synchronous power; optical downlink as a pipe, not a brain; servicing; debris; spectrum. Launch mass dominates cost until cadence collapses it. Europe’s ASCEND study: the climate ledger only begins to work if launch is about ten times cleaner.

Surveyed record (not ours)

ItemWhat it isWhat it is not
Starcloud-1First H100 class GPU flown (Nov 2025)A data centre
Starcloud FCCUp to 88,000 compute satellites proposedAuthority to occupy the south
SpaceX FCCUp to 1,000,000 orbital data-centre satellites proposed (Jan 2026)A southern node
Davos 2026 claimLowest-cost AI in space in two to three yearsA completed thermal design
NSPM-17 (20 Aug 2026)1,000 US launches and reentries a year by 2030Australian cadence
Terafab / D3 (public)Earth plant for orbital and vehicle siliconPilbara. Not this sitting

Why the sitting still holds

Earth data centres are hitting power, water and permits. Orbit sells continuous solar and no cooling towers. The integrated United States play is silicon, launch, model, and network in one stack. That may or may not close. Either way, tokens in LEO do not replace a brain that still works when the link is twenty minutes or gone.

Iowa is postage. A million satellites on a form are not sovereignty. If the stack flies, the industrial question for this continent is unchanged: who catches, who refurbs, who fuels, and whose Country the concrete sits on.

← Ambition