The reference node Two vehicles carry the rest of the book: Reference Node A, 100 kW. One modest compute payload on a single spacecraft. This is roughly the scale the sector is actually flying towards in the late 2020s, and everything in it exists in flight-proven form today. Reference Node B, 1 MW. Ten times the compute, and the scale at which the sector’s stated ambitions begin. Not simply Node A ten times over, because some things scale better than linearly and one thing, heat transport does not scale at all with current hardware. Both are assumed to fly in a dawn, dusk sun-synchronous orbit at roughly 650 km, which is the orbit Google’s Project Suncatcher design selects and, for reasons that will become clear in the propulsion section, very close to the correct answer. One convention, stated once and used throughout: a modern AI rack is taken as 200 kW, 1.4 tonnes, and $3.5M. Reported figures range from about 130 kW to 230 kW per rack depending on generation and configuration, so this is a choice rather than a fact, and it is among the assumptions most worth changing when you run your own version of these numbers.
Figure 8.1 — Reference Node A, subsystem by subsystem
1. The power chain, sized backwards The first mistake an outsider makes is sizing the solar array to the solar array. You size it to the load, and then you walk every loss backwards. Figure 8.2, Sizing backwards from the chips
To deliver 100 kW to silicon you need roughly 102 kW off the power bus, about 111 kW from the array after distribution and conversion losses, and about 120 kW at beginning of life once you have reserved margin for five years of radiation degradation and imperfect sun pointing. At 408 W/m² of cell output, that is roughly 290 square metres of array intercepting about 400 kW of raw sunlight. Three engineering decisions hide inside that paragraph, and each is a company in Part IV. Voltage. Moving 120 kW around a spacecraft at the 28 or 100 volts traditional buses use means enormous currents and therefore enormous conductor mass. You want to go to several hundred volts. But low Earth orbit is a tenuous plasma, and high-voltage surfaces in plasma can arc, a failure mode that has damaged real spacecraft. Highvoltage power management for orbit is genuinely hard, which is why it is a differentiator rather than a commodity, and it is the specific thing that makes a high-power bus a product rather than a bigger version of a small one. Deployment. Two wings of roughly 145 m² each must survive launch folded and then unfold once, perfectly, with no opportunity for repair. Mechanisms are the classic spacecraft single point of failure. The physics never fails here. The hinge does. Cell chemistry and a live argument. Chapter 4 asserted that multi-junction cells at 30% are obviously correct for orbit, because area and mass are what you pay for. That reasoning is sound at Node A’s scale. It may be wrong at gigawatt scale, and the sector is now openly split: in February 2026 Rocket Lab introduced silicon solar arrays aimed specifically at space-based data centres, arguing that low cost per watt at industrial manufacturing scale is what gigawatt-class orbital power actually requires.10 Both positions can be right. Silicon is roughly two-thirds the efficiency, so it needs about 50% more area and carries more mass per watt, a real penalty when launch is expensive and the spacecraft is small. But it is manufacturable by the square kilometre, and when you need gigawatts rather than kilowatts, manufacturability beats elegance. Which cell chemistry wins is a proxy for whether this industry is building spacecraft or building infrastructure, and it is worth asking every company in the power layer which side of that they are on.
2. The number nobody in this sector wants to build up Which brings us to the cost of the array, and to the single weakest link in every published orbital-compute economic model I have examined, including the ones I find persuasive.
Figure 8.5 — The least-examined number in the entire thesis
Bespoke space solar array systems, cells, blanket, structure, deployment mechanisms, qualification have historically been quoted at figures on the order of $1,000 per watt, with thin-film approaches claiming a path below $200/W.11 Terrestrial utility-scale modules are around fifteen cents. At $1,000/W, the 120 kW array on Node A costs $120 million, which is forty times the cost of the computers it powers and roughly forty times the entire launch bill. The thesis dies instantly. At the $3/W that gigawatt-scale orbital compute plans implicitly assume, that same array costs $360,000, and nobody thinks about it again. Three orders of magnitude separate those two worlds, and no publicly disclosed contract closes the gap. That does not mean the gap cannot close, the historical figure reflects one-off programmes with heavy qualification burdens and no production volume, which is exactly the cost structure that mass manufacture destroys. Starlink already demonstrates that spacecraft components fall by orders of magnitude when built in the thousands. But it does mean this: When you assess a company in this sector, find out what it assumes power hardware costs per watt, and whether that number comes from a quote or from a slide. This is the assumption most likely to be doing silent work in a business plan, and the one least likely to be challenged in a room full of people discussing radiators.
It is also why the power layer is strategically more important than its share of the spacecraft mass suggests, and why the fund thesis’s concentration there is defensible on structure rather than on enthusiasm.
3. The compute payload, and the ways vacuum changes it Now the part that earns the money. A 100 kW payload at our convention is half a rack: roughly 700 kg and $1.75M of accelerators, memory, storage and switching. Four things change when you put it in vacuum. No air means no air cooling anywhere. This is more disruptive than it sounds. Terrestrial servers use liquid cooling for the hot components and air for everything else, voltage regulators, memory, drives, switch ASICs. In vacuum there is no “everything else” path. Every component that dissipates meaningful power needs a conduction path to a cold plate, which means a substantially redesigned board and chassis, not a standard rack bolted to a
satellite. The vehicles being flown today are the beginning of that redesign, and it is one of the more underappreciated engineering programmes in the sector. No convection means hot spots are permanent. On Earth, a badly cooled corner of a board is rescued by moving air. In vacuum, heat goes where the metal takes it, and nowhere else. Outgassing and materials. Plastics, adhesives and lubricants release volatiles in vacuum, which then condense on the coldest, most sensitive surfaces available, typically your optics and your radiators. Every material in the payload has to be qualified for this, which quietly rules out large parts of the commercial supply chain. Shielding is packaging. Chapter 6 established that you shield the box, not the ship. For Node A, roughly 10 mm aluminium-equivalent around a compact enclosure adds about 350 kg. That figure scales with the surface area of the enclosure, not its contents, which is a strong argument for packing compute densely, the same square-cube logic that makes big things cheaper per unit than small ones.
4. The thermal chain, in detail Chapter 5 sized the radiator. Here is everything between the chip and it. Figure 8.3, Every degree lost between chip and panel is area you must launch
The junction runs at perhaps 95 °C. The cold plate touching it sits at 80. The coolant leaving the payload is at 72. By the time heat reaches the radiator root you are at 65, and the mean radiating surface, because a fin is never isothermal, and the tip is always cooler than the root is nearer 58. That 37-degree cascade is not waste in a moral sense; it is the price of moving heat. But look back at Figure 5.1 and price it. Radiating at 58 °C instead of 95 °C costs you roughly 40% more panel area, which on Node A is about 40 square metres and 320 kilograms you launched purely to pay for temperature drops. This is why the thermal layer is a real engineering business and not a plumbing subcontract. A company that shaves fifteen degrees out of that chain has removed a quarter of the radiator from the vehicle. For Node A, the numbers are comfortable: 100 kW at around 58 °C surface temperature needs roughly 140 m² of two-sided radiator, about 1,120 kg of panel and another 300 kg of pumps, fluid and plumbing. All of that is within the envelope of flight-proven pumped single-phase hardware. For Node B, they are not comfortable. A megawatt needs roughly 1,400 m² and pushes you into pumped two-phase transport, which is where flight heritage runs out. The step from Node A to Node B is not a scaling exercise. It
is a technology programme, and any company that describes it as the former is telling you something about itself.
5. Talking: optical terminals and the ground Node A carries two or three optical terminals, roughly 90 kg all in, for inter-satellite links and, where the architecture allows, direct optical downlink. RF remains as a low-rate backup for command and telemetry, because you never want your only path to the spacecraft to depend on a mechanism holding a microradian aim. Chapter 7 established that a compute node’s own traffic is modest. What is not modest is the ground segment: a network of optical or Ka-band stations, sited for geographic diversity against cloud cover, is a capital programme in its own right, and it appears in almost no spacecraft cost estimate you will read. When someone quotes you dollars per GPU-hour in orbit, ask whether the ground stations are in it.
6. Pointing: three requirements that fight each other Attitude control on a compute satellite is more interesting than on almost any other spacecraft, because it has three simultaneous pointing demands that are geometrically incompatible: 1. The solar array wants to face the Sun. 2. The radiator wants to face deep space and specifically to avoid facing the Sun or the Earth, since both warm it and Chapter 5 showed how expensive warm is. 3. The optical terminals want to face other spacecraft and, periodically, the ground. You cannot satisfy all three with one rigid body, which is why real designs use gimballed arrays, gimballed terminals, and a body attitude chosen to keep the radiators edge-on to the Earth. That is three more mechanism sets, and mechanisms are where spacecraft fail. The dawn, dusk sun-synchronous orbit is popular precisely because it partially resolves this fight: the Sun stays roughly perpendicular to the orbit plane, so a fixed geometry can keep arrays sunward and radiators edge-on to Earth without constant slewing. The choice of orbit is not a detail. It is the decision that makes the rest of the spacecraft possible, and it is a good early question for any company: what orbit, and why that one? Hardware is modest, reaction wheels, star trackers, magnetorquers to dump accumulated momentum, around 120 kg for Node A.
7. Staying up: drag and station-keeping A compute node is mostly sail. Roughly 290 m² of array and 140 m² of radiator on a 4.7-tonne vehicle gives an areato-mass ratio far worse than a conventional satellite, and low Earth orbit is not empty. Work the drag calculation and the result is genuinely reassuring, with a sharp cliff at the bottom. Figure 8.4, Why everyone is choosing 600, 700 km
Altitude
Station-keeping Δv per year
Five-year propellant, Node A
400 km
~114 m/s
~180 kg
500 km
~26 m/s
~42 kg
600 km
~7 m/s
~11 kg
650 km
~4 m/s
~6 kg
800 km
~1 m/s
~2 kg
Assumes area-to-mass 0.02 m²/kg, drag coefficient 2.2, electric propulsion at 1,600 s specific impulse. Atmospheric density varies by roughly an order of magnitude over the eleven-year solar cycle; multiply accordingly for solar maximum. At 650 km, keeping a compute node in place for five years costs a few kilograms of propellant. Drag is a non-issue. At 400 km it costs nearly 200 kg and a much larger thruster, and at solar maximum considerably more than that. So the altitude choice is triple-constrained and lands in a narrow band. Below about 500 km, drag punishes you. Above about 700 km, orbital debris lifetime and radiation dose both rise, and you lose the natural decay that keeps the orbit self-cleaning. In between sits the band everyone is choosing, for reasons that are entirely physical rather than fashionable. Allow 150 kg of thruster, tank and propellant on Node A.
8. Fitting inside the rocket Node A stows into a conventional fairing without heroics. Node B does not stow into anything smaller than Starship, and even there the constraint is not mass but folding. A megawatt node carries roughly 2,900 m² of array and 1,400 m² of radiator. As flat sheets a few millimetres thick, that is only tens of cubic metres, volumetrically trivial against a fairing measured in hundreds of cubic metres. The problem is that all of it must be creased, restrained through launch loads, and then released in sequence without a single jam, and the mechanism count grows with the number of folds. This is the strongest engineering argument for modularity: rather than one 1 MW spacecraft with a thousand hinges, fly ten 100 kW nodes flying in formation and link them optically. Google’s design takes exactly this route, envisaging clusters of many satellites holding formation at separations of a few hundred metres.8 It is also the
argument that eventually pulls in-space assembly and servicing from science fiction into the supply chain, which is Chapter 15.
9. Dying properly, and the problem nobody has solved Two end-of-life issues, and the second is a business-model problem disguised as an engineering one. Disposal. From 650 km, natural orbital decay takes far longer than the five-year post-mission disposal window that US regulation now expects, so a compute node must actively deorbit. Lowering perigee enough to guarantee reentry costs on the order of 100 m/s, perfectly affordable, but it must be budgeted at the start and it must still work after five years of radiation. A dead 5-tonne spacecraft with 290 m² of sail area at 650 km is precisely the debris problem Chapter 16 is about. Refresh. Here is the awkward one. Spacecraft are designed for lifetimes of five to fifteen years. AI accelerators are economically obsolete in roughly three. A terrestrial data centre solves this by rolling in new racks; the building outlives twenty generations of hardware. An orbital data centre cannot do that today. It either accepts that it is flying depreciating silicon for its full mission life, which quietly destroys the return, since the revenue per rack collapses as newer hardware arrives or it develops the ability to replace payloads on orbit, which is servicing, which does not commercially exist at this scale. This, and not radiation or thermal, may turn out to be the sector’s real structural weakness. It is the question I would put to every operator: what is your hardware refresh strategy, and what does your model assume about revenue per rack in year four? I have not yet seen a satisfying answer, and Chapter 15 is where the companies who could provide one live.
The build-up Table 8.1, Reference Node A: 100 kW of compute Subsystem
Mass
Notes
Compute payload
700 kg
Half a rack at the 200 kW / 1.4 t convention
Enclosure and shielding
350 kg
~10 mm Al-equivalent around a compact box
Solar array
800 kg
120 kW BOL at ~150 W/kg
Power management and harness
250 kg
High-voltage PMAD, distribution
Battery
80 kg
Housekeeping only, full-sun orbit assumed
Radiators
1,120 kg
~140 m² two-sided at ~8 kg/m²
Heat transport
300 kg
Pumps, fluid, plumbing, cold plates
Structure and mechanisms
750 kg
Primary structure, deployment hardware
Attitude control
120 kg
Wheels, star trackers, magnetorquers
Communications
90 kg
Optical terminals plus RF backup
Propulsion and propellant
150 kg
Electric, five years at 650 km, plus deorbit
Total
≈ 4,710 kg
Now the economics, and an admission. Chapter 2 used 3,000 kg as a placeholder for this vehicle. The build-up says 4,710. That matters, because the placeholder made the launch bill look like half the value of the computers, and the real number does not:
7.5×
1.6×
0.4×
At Node B the ratio improves, roughly 37 tonnes carrying $17.5M of computers, so about $22M of launch against
the payload, or 1.3×, because structure and avionics scale sublinearly while compute scales linearly. Scale helps. It does not rescue. The honest summary is that at contracted Starship pricing, launch is roughly the same size as the hardware bill, and the sector’s economics only become comfortable at the target price nobody has yet paid. That is a materially more cautious conclusion than Chapter 2’s placeholder implied, and it is why the correction is printed here rather than fixed silently.
Where each part breaks and who is trying Subsystem
The thing that breaks
Where the companies are
Solar array
Deployment mechanisms; cost per watt;
Two-phase flow at megawatt scale
Ch 11, thermal
unsolved Radiators
Deployment; micrometeoroid puncture;
Ch 11, thermal
coating degradation Optical terminals
Microradian pointing; ground-segment
Ch 14, comms
capital Attitude control
Three incompatible pointing demands;
Two rows in that table have no adequate supplier: megawatt-scale heat transport, and on-orbit hardware refresh. Everything else has at least one credible company attacking it. That asymmetry, one badly-served constraint and one unsolved business model, in a sector where everything else is buyable is the map. Part IV walks it.