First: who buys this? Cost models are the easy part. The question that decides this sector is who signs the contract, and it is almost never asked in print. Start by being clear about who does not. A hyperscaler will not put production inference on a satellite in the foreseeable future. Enterprise procurement requires an availability commitment, a support path, a second source and a settled legal jurisdiction; Chapter 8 established that an orbital node cannot be physically touched after launch, and Chapter 17 established that the jurisdiction question is genuinely unresolved. The named cloud “booked workloads” reported by operators today should be read as what they almost certainly are, funded pilots, valuable as validation, not evidence of a procurement channel. So who is left? Three buyers, and they share a property that turns out to matter more than anything in the cost model: all three pay a premium, and none of them are buying commodity compute. 1. Data that is already in orbit. Earth-observation operators are downlink-limited, not compute-limited. Chinese researchers building orbital compute frame their entire rationale this way, and Chapter 7’s arithmetic supports it: raw sensor data is enormous, conclusions are tiny. For this buyer the alternative is not a cheaper terrestrial GPU, it is discarding the data. That is not a price-sensitive comparison, and it is available today at kilowatt scale rather than megawatt scale. 2. Defence and intelligence. Latency to the sensor, resilience against attacks on terrestrial infrastructure, and sovereign control of the processing chain. Government customers buy attributes rather than throughput, and have historically paid multiples of commercial rates for both. It is not an accident that the private companies in this sector with real revenue have defence-weighted revenue. 3. Sovereign data jurisdiction. A satellite falls under the jurisdiction of the state that licensed it, not of the territory it flies over. For a state wanting compute outside another power’s legal reach, that is a feature with no terrestrial substitute, and Europe’s public programme is explicitly framed around it. The consequence reorganises the rest of this chapter: Orbital compute does not need to beat the commodity price of a GPU-hour. It needs to beat it for buyers who are not buying a commodity. Any model that benchmarks orbital cost against spot cloud rental is measuring against the wrong customer and it is the harshest possible test, which is why this chapter uses it anyway.
Second: what is the real competitor? Part I framed the alternative as “wait four to ten years for a grid connection.” That is the wrong counterfactual, and it is the strongest attack on this book’s argument. The real alternative is behind-the-meter generation: put a gas turbine on site and skip the utility entirely. It is already the dominant workaround, and if it is fast, the availability argument collapses and nothing else here saves it. So how fast is it?
Figure 18.5 — The queue moved. It did not disappear
.
Three companies in the world build heavy-duty gas turbines at scale. All three are sold out on a horizon no other industrial market can absorb: order backlogs stretching into 2029 and 2030, lead times that have gone from two or three years in 2022 to roughly five today, with manufacturers now advising customers to plan on seven-to-eight-year timelines, and steep price inflation on delivery slots.22 That is the answer, and it is more interesting than a simple defence. The bottleneck did not vanish when the industry routed around the utility. It relocated from the interconnection queue to the turbine factory, which is the same structural story this book opened with: compute is limited by industrial capacity, not by physics and not by money. Two honest qualifications. Smaller reciprocating and aeroderivative units have much shorter lead times, on the order of eighteen months to two years, and are genuinely fast but they are less efficient, dirtier, permitconstrained on air quality in exactly the metros where demand concentrates, and they do not reach hyperscale. And an operator who ordered turbine slots two years ago has power today while an orbital operator does not. Behindthe-meter gas is a real and partial escape at tens of megawatts, and it meets the same wall at hundreds. It also sets a cost benchmark that is not soft: a behind-the-meter site costs roughly the same per megawatt-year as a grid-connected one, because cheaper fuel offsets the generation capital.
The inputs, all of them, on one page
Input
Value used
Source
Reference vehicle
1 MW compute, 37 tonnes
Ch 8
Compute hardware
6.8 t, $17.5M
Ch 8 convention
Spacecraft mass, everything else
30.2 t
5 years base (3, 7 tested)
Ch 6
Spacecraft service life
7 years base (5, 10 tested)
Ch 8
Cost of capital
12% base (8, 20% tested)
new, see below
Availability
92%
new, see below
Fleet attrition
2% of hull per year
Ch 16
Performance fade
1.5% per year
Ch 16
Operations
$2M per MW-year, including ground
estimate
segment Insurance
3% of hull per year
Ch 16
Revenue
$9M / $18M / $30M per MW-year
three worlds
Four lines are new relative to how this sector usually models itself, and each exists because leaving it out flatters the answer: Cost of capital. Earlier drafts used straight-line amortisation and called themselves “a cash cost model, not a valuation.” That was a dodge. A long-dated, capital-intensive, high-risk asset must carry a capital charge, and the honest instrument is a capital recovery factor, the annual payment that returns capital at a required rate over the asset’s life. At 12% over seven years that is 22% of capital per year, against the 14% straight-line implies. This single correction adds more cost than launch does. Availability. Terrestrial data centres sell 99.9%. An orbital node cannot be touched: one jammed deployment, one unrecovered latch-up, one radiator puncture and it is a total loss. 92% is a judgement, unproven in either direction, and it appears in the tornado so you can move it. Fleet attrition and performance fade. Chapter 16 established that millimetre impacts are certain and that coating erosion attacks emissivity, which sits inside a fourth-power law. A model assuming constant thermal performance for seven years contradicts its own Chapter 16. Attrition replaces 2% of hull value annually; fade derates output 1.5% a year. Availability and fade together give an effective output factor of 0.879, a node sells about seven eighths of its nameplate GPU-hours. Every unit cost below is per effective GPU-hour.
Read that slowly, because it is the honest snapshot and it is not yet a business. At today’s manufacturing costs and today’s cost of capital, orbital compute does not clear the commodity price. The base case loses money. The stretch case, which assumes manufacturing economics nobody has yet demonstrated for a 37-tonne deployable vehicle, lands at $3.72 against a $3.50 market, close, and not there. Now hold that beside two other snapshots. Utility-scale solar in 2010 cost roughly four times the wholesale power price it was competing with. Lithium cells in 2010 cost about seven times what the automotive industry said it needed. Falcon 9 in 2012 was more expensive per kilogram than the incumbent it was built to undercut. In each case the snapshot was correct and the conclusion drawn from it was wrong, because the snapshot measured a first-of-a-kind unit and the thing that mattered was the slope. The right question is therefore not does this clear today, it plainly does not but at what point on the manufacturing curve does it clear, and can the industry get there. That is a calculable question, and the next
section calculates it.
Figure 18.4 — Cost per GPU-hour across manufacturing cost and cost of capital
The matrix is the honest summary of this book’s economics. Green cells clear today’s commodity price; there are few of them, and they all sit in the bottom-left corner where a company manufactures at a thousand dollars a kilogram and borrows like a utility. The thesis needs cheap manufacturing and cheap capital simultaneously, and the sector currently has neither. Which brings demand back with force, and explains the sequence the industry will actually follow. The premium buyers come first because they clear earlier. Against the stretch case at $3.72, a buyer paying 30% over commodity for sovereignty, latency to the sensor or resilience makes the business work today. The commodity market clears later, when the curve has done its work. The verdict this arithmetic supports: orbital compute is a premium-buyer business now and a commoditycompute business later, and the interesting question is the size of “later.” A pitch promising cheaper AI in orbit this year is wrong. A pitch promising it by the early 2030s is arguing about a learning rate, which is a real argument with real evidence on both sides.
When does the math work? Three things move together over the next decade, and none of them requires a breakthrough. Each is an extrapolation of a process already running.
Figure 18.7 — The three curves underneath the crossover
One: manufacturing goes down a learning curve. Wright’s law, unit cost falls by a fixed percentage for every doubling of cumulative production is among the most durable empirical regularities in industrial economics, and it holds across aircraft, semiconductors, solar modules, batteries and launch vehicles. Observed learning rates cluster between 10% and 20% per doubling. The relevant question for this sector is not whether it applies but how many doublings it gets, and that depends on deployed capacity. Take a deployment path that reaches a gigawatt in the mid-2030s, aggressive, but slower than the sector’s own filings. From two megawatts of cumulative capacity to a thousand is about nine doublings. At a 15% learning rate, that takes spacecraft manufacturing from $2,000/kg to roughly $450/kg. Not by inventing anything: by building the ninth hundred of something rather than the first ten. Two: the cost of capital falls as the asset class matures. This is the driver nobody models and it is worth as much as the manufacturing curve. Today an orbital compute constellation is financed at venture rates because it is unproven, uninsured and unrepeated. That is exactly what solar farms and wind projects looked like in the 2000s, and both are now project-financed at single-digit rates against contracted offtake. The path is not mysterious: flight heritage, then contracted revenue, then insurance, then debt. Going from 20% to 9% cost of capital cuts the annual cost of the same hardware by roughly a third. Three: availability improves with flight heritage, as it has for every spacecraft class ever fielded, and every point of availability is revenue against a fixed cost base. Put the three together against a flat revenue assumption, no help from rising GPU prices, no help from improving performance per watt, both of which would accelerate it:
Figure 18.6 — When the math starts working
Year
Build cost
Cost of capital
Cost per GPU-hour
2027
$2,000/kg
20%
$8.53
2030
~$990/kg
15%
$4.52
2032
~$700/kg
~12%
$3.41
2035
~$460/kg
10%
$2.59
2040
~$300/kg
8%
$2.11
Orbital compute clears the premium market around 2030 and the commodity market around 2032, on a 15% learning rate. At 20% it happens a year earlier; at a pessimistic 10% it slips to 2034. The band in Figure 18.6 is that whole range, and the honest headline is: somewhere between 2031 and 2034, on assumptions that require nothing new to be invented. Three things worth saying about that result. It is conservative on revenue. Holding the price of a GPU-hour flat for fourteen years while performance per watt improves every generation is almost certainly wrong in the sector’s favour: orbital costs scale with watts, radiators, arrays, mass, while revenue scales with useful output. Every doubling of tokens per watt is a doubling of revenue against a fixed thermal and structural cost base. That single effect could pull the crossover in by years, and it is deliberately excluded here. It is self-referential in a way you should watch. The manufacturing curve needs deployment, and deployment needs the manufacturing curve. Industries do escape that loop, subsidised early demand, government anchor customers, a vertically integrated player who eats the early losses but they do not escape it automatically. This is exactly what the premium buyers are for. Defence, sovereign and sensor-adjacent customers pay enough to fund the first nine doublings. They are not a consolation prize for a business that cannot reach the commodity market; they are the mechanism by which it gets there. And it explains why the investable moment is now rather than in 2032. By the time the curve crosses, the companies that rode it down are priced accordingly. The window for an investor is the period when the physics is settled, the engineering is legible, and the arithmetic has not yet caught up, which is a fair description of the present.
What actually moves the answer
Figure 18.2 — What actually moves the answer
One variable at a time from the base case: 1. Spacecraft build cost, dominates everything. 2. Cost of capital, second, and absent from every published model in this sector I have read. A company that finances like infrastructure has a structural advantage over one financing like a startup, independent of its engineering. 3. Spacecraft service life, a ten-year vehicle nearly halves the capital charge. 4. Compute hardware life, Chapter 6’s depreciation argument and Chapter 15’s servicing argument arriving together. 5. Availability, larger than launch price, and completely unmeasured. 6. Launch price, sixth. The number this sector discusses most is the second least important input in its own economics. 7. Insurance, smallest today, most likely to move if Chapter 16’s capacity problem bites.
The queue argument, priced across three revenue worlds
Figure 18.3 — The argument, honestly, across three revenue worlds
Earlier drafts held revenue fixed at $18M per MW-year while stress-testing seven cost inputs. That was the second dodge, and it mattered: $18M derives from spot GPU rental rates set during a capital-expenditure boom, and those rates are likelier to fall than the cost inputs are. Three worlds, then. At $9M, rates halving as supply catches up, orbital compute is unviable at any manufacturing cost in this model, and so is a good deal of the terrestrial industry. At $18M, base-case orbital loses money and only the stretch case survives, with a long break-even. At $30M, premium buyers, sovereign contracts, defence rates, orbital clears terrestrial past about seven and a half years of waiting. A narrower window than Part I implied, at today’s costs. Run the same three worlds against the 2032 cost line from Figure 18.6 and the picture inverts: the $18M world clears comfortably, and the $9M world becomes the marginal case rather than the impossible one. The break-even queue length falls with the cost curve, which is the point of the cost curve. It also explains the shape of the actual sector today: the companies with revenue sell to governments, not to clouds. That is not a limitation of the thesis. It is the first rung.
How long does orbit itself take? The availability argument claims capacity is needed now. Orbit should be held to the same clock.
Set against Figure 18.5 that is sober but not bad. Orbit already beats a grid connection in a congested hub and beats a new heavy-duty turbine slot. It is slower than a small reciprocating unit, and it is not the instant capacity the marketing implies. The direction of travel matters more than the current position. Every one of those timelines is fixed by physical infrastructure except the orbital one, which is fixed by a production line that does not exist yet. Grid queues lengthen as demand grows; turbine backlogs lengthen as demand grows; spacecraft build times shorten as production scales. Time-to-power is the one competitive dimension where orbit’s trend line points the right way
while everyone else’s points the wrong way.
How big is a gigawatt, physically?