The Connectivity Question

Orbit	Altitude	Period	Round-trip latency
Low Earth orbit	~550 km	~95 min	~5 ms
Geostationary	35,786 km	24 hours	~240 ms

Radiator, 40 MW

60–80k m²

Mass to launch

1,500–3,000 t

Upfront — orbit

$2.3–4.5B

Upfront — ground

$1.2–1.6B

Upkeep/yr — orbit

$0.6–3.2B

Upkeep/yr — ground

$0.4–0.6B

Illustrative. The orbital regimes (ISS, LEO, MEO, GEO) are labelled with real altitudes but spaced for readability, not to true scale; the debris layer is a representative sample of the 25,000+ tracked objects from Module 5; the reality figures carry forward the 40 MW estimates from Modules 2–4.

Source: DC Atlas — orbit geometry and the Modules 2–4 mass/cost synthesis; Google Research (Project Suncatcher) for the constellation architecture

Illustrative. The orbital regimes (ISS, LEO, MEO, GEO) are labelled with real altitudes but spaced for readability, not to true scale; the debris layer is a representative sample of the 25,000+ tracked objects from Module 5; the reality figures carry forward the 40 MW estimates from Modules 2–4.

Source: DC Atlas — orbit geometry and the Modules 2–4 mass/cost synthesis; Google Research (Project Suncatcher) for the constellation architecture

Say the data centre is up there, powered, cooled and — improbably — maintained. It still has to talk to Earth, constantly and at enormous bandwidth, and this is where a nice question you might ask on a whiteboard turns into a real design constraint. Do the satellites hang fixed over a spot, rotating with the planet, or do they race around it independently? Does a single ground station stay in view, or does the connection hand off between satellites? And could you just borrow the network that is already up there? The answers are more interesting than the pitch lets on, and they narrow the design sharply.

The first fork decides everything: low orbit or geostationary. A geostationary satellite, at 35,786 kilometres, orbits in exactly 24 hours, so it appears to hang motionless over one point on the ground — permanently in view of the same station, which sounds ideal. But it is far away, and physics charges for distance: a round trip to geostationary and back takes about 240 milliseconds, an eternity for interactive compute and enough to ruin most real workloads. So geostationary is out. Every serious orbital-compute proposal instead chooses low Earth orbit, roughly 500 to 700 kilometres up, where the round trip is a handful of milliseconds — but where the satellite no longer holds still.

Why compute has to live in low orbit

Approximate round-trip signal latency by orbit

LEO (~550 km)5 ms

MEO (~8,000 km)50 ms

GEO (35,786 km)240 ms

Because a low-orbit satellite is overhead for only a few minutes before it disappears past the horizon, you cannot rely on one. You fly many, evenly spaced in the same orbital shell, so that as one races out of view the next races in and the connection hands off between them — the model Starlink uses for consumer internet. Then you knit the satellites together with inter-satellite laser links: each one talks optically to its neighbours ahead, behind and to the side, forming a mesh in which data can hop from satellite to satellite around the whole constellation without ever coming down to the ground. That laser mesh, a necklace of compute spinning around the Earth, is the "ring of constant connection" the concept needs — and it is a genuinely elegant piece of engineering.

Property	Low Earth orbit (~550 km)	Geostationary (35,786 km)
Orbital period	~95 minutes (~16 laps/day)	24 hours (fixed over one point)
Moves relative to ground?	Yes — races overhead in minutes	No — hangs in place
Round-trip latency	Tens of milliseconds	~240 ms
Constant link needs	A constellation + hand-offs + laser mesh	A single satellite in view
Who picks it	Starcloud, Suncatcher, essentially all	No serious compute proposal

There is one bottleneck the pitch glosses, and it is the one that bites. Meshing satellites to each other by laser is comparatively easy; getting the sheer volume of an AI workload's data down to Earth and back is not. Ground stations are a finite, expensive resource, each satellite is only over a given station for minutes, and moving training-scale datasets — petabytes — through that keyhole is a genuine constraint on where the compute can live and what it can usefully do. It also reintroduces, through the back door, exactly the terrestrial infrastructure the orbital pitch claims to escape: you still need a global network of ground stations, real estate, fibre and power, to talk to the thing in the sky.

Exploring a space-based, scalable AI infrastructure system design

Connectivity, in other words, is the part of the orbital story that most nearly works, and even it comes with an asterisk and a hidden ground network. Which brings us to the only question that finally matters: put the whole bill on the table — the radiators, the launches, the batteries, the dying hardware, the ground stations — and does any of it beat a building on the ground?