The old idea
In October 1945, in Wireless World, a young Arthur C. Clarke pointed out that three stations in geostationary orbit could cover the whole planet, and worked out the orbital mechanics to prove it. He did not patent it. The paper is the founding document of the idea that infrastructure not exploration, not science, but plumbing is a reason to go to space. Everything in this chapter is a descendant of that observation. The satellite is not the mission. It is the building.
The physics, three architectures, and the trade between them
There are only three ways to arrange compute in orbit, and each is a different answer to Chapter 8’s mechanism problem. The monolith. One large spacecraft, one enormous radiator, one very large array. Fewest interfaces, best mass efficiency, worst deployment risk, and a single failure takes everything. The modular cluster. Many small identical units flying in formation, linked optically, behaving as one machine. This is Google’s approach, clusters of many satellites holding formation at separations of a few hundred metres.8 It is also, in a different form, the tile architectures. The hosted payload. Put the compute on somebody else’s platform, a commercial station, or a standardised bus flying other customers’ payloads. Lowest capital, lowest control, fastest to orbit. The trade between the first two is quantifiable, and I have not seen it done in public, so here it is.
Figure 13.1 — Modularity is bought, not free
Radiator area is indifferent to how you split the system, a megawatt needs its square metres whether it is one machine or a hundred. But three things are not indifferent: Shielding scales with enclosure surface area, and surface grows as the two-thirds power of volume. Split one megawatt into ten modules and total shielded surface rises by about 2.2×; into a hundred, by about 4.6×. This is simply the square-cube law, and it is the strongest physical argument for packing compute densely. Optical terminals scale roughly linearly with module count, because every module must talk to its neighbours. Structure scales mildly. There is a fourth, which Chapter 5’s self-heating section quantifies: a cluster’s members occupy part of each other’s radiator sky, so every module needs a few percent more panel than the same power would need flying alone. Formation spacing is a thermal decision before it is a control decision. So the monolith is mass-optimal and the cluster is not. What the cluster buys in return does not appear on that chart: simpler deployment per unit, the ability to launch incrementally as capital allows, graceful degradation when one unit dies, and, critically for Chapter 8’s refresh problem, the possibility of replacing a fraction of the fleet each year with current silicon rather than flying one ageing monolith for a decade. Read that last point carefully, because it may be decisive. The modularity penalty is a one-time mass cost of roughly two. The refresh problem is a recurring revenue cost that compounds. If I had to bet on which architecture wins, I would bet on the one that can be upgraded, and pay the two.
What breaks
Formation-keeping propellant and control for clusters. Deployment for monoliths. For hosted payloads, dependence on a platform operator whose priorities are not yours. And for all three, the refresh problem from Chapter 8, which nobody has solved.
Who is attacking it
Starcloud is the flight-heritage leader: it flew a commercial accelerator in orbit in late 2025 and trained a model there, and its second vehicle carries a newer accelerator with roughly ten times the compute and a hundred times the power generation, using direct-to-chip liquid cooling into deployable radiators, with booked workloads from named cloud customers.12 That combination, flown hardware, real customers, and the first serious test of the thermal architecture at rack scale, makes it the single most informative company in the sector regardless of whether you own it.
SpaceX with xAI is the scale competitor: a filing for a very large compute constellation, captive launch, and its own AI workloads to fill it. Chapter 9 already made the point that the sector’s enabler is also its largest competitive risk. The strength test is whether an independent operator is distinguished enough that SpaceX cannot erase it by announcement. Google is the research leader, with published physics and prototype satellites planned. Axiom Space is the hosted-platform route, with data-centre nodes already on orbit and a business that is crosssubsidised by existing revenue, which in a sector full of pre-revenue narratives is a genuinely different risk profile. Sophia Space is the tile bet, and Lonestar occupies the lunar storage niche. Madari and similar sovereign-backed entrants matter more than their size suggests, for reasons Chapter 17 takes up. And there is a tail of very early entrants with enormous constellation filings and no hardware, which is where the sector’s promotional energy concentrates.
The investment stance
Ambition is not a differentiator in this sector. Flight heritage is. A hundred-thousand-satellite filing costs a lawyer’s fee. Getting one rack to work in vacuum for a year costs everything the company has. I hold this layer, deliberately, including companies that compete directly with SpaceX but only where something has flown, and never at a size where being wrong about which architecture wins is fatal. The reason to own it despite the competitive risk is that this is where the asymmetric outcomes are: if orbital compute works at all, the operators capture the revenue, and an acquisition by a larger player is a good outcome rather than a capped one.
What to watch
A second-generation compute satellite running paying workloads on orbit. This is the nearest thing the sector has to a verdict, and it is imminent. Anyone publishing an actual price for orbital GPU-hours. Until someone does, every economic model in this sector, including mine is theory. Whether cluster architectures demonstrate formation-keeping at the separations their designs assume.