A guidebook that cannot argue against itself is a brochure. This chapter is the strongest case I can make against everything in this book, made as well as someone who disagrees with me would make it. One framing note before the list. Chapter 18 showed that this sector’s economics are a date, not a verdict: the arithmetic crosses somewhere between 2031 and 2034 on assumptions that require nothing new to be invented. Most of what follows therefore moves the date rather than cancelling the industry. Two items genuinely cancel it, and they are marked. The rest are schedule risk, which is the normal condition of an infrastructure business, and which is priced by when you buy rather than whether.
Figure 21.1 — How this thesis dies
The two that would genuinely cancel it, and the ones that move the date 1. Manufacturing cost never falls (cancels). Part V’s model is unambiguous: at $20,000 per kilogram of spacecraft, orbital compute costs $23 per GPU-hour against a $3.50 market. The entire thesis rests on a 37-tonne power-and-cooling spacecraft being manufacturable at Starlink-like economics and Starlink satellites are around 800 kg of relatively simple hardware, not thirty tonnes of deployable radiator and high-voltage power distribution. The analogy that carries the whole model may not hold, and this is the most probable route to the sector failing quietly rather than dramatically. 2. The queue unblocks everywhere at once (cancels). Chapter 18 shows the orbital case only clears terrestrial past roughly seven and a half years of waiting, and then only in the premium-revenue world. That is a narrow margin against a moving target. Behind-the-meter gas is already the standard workaround, and although Chapter 18 shows the turbine order book has simply become the new queue, that is a supply constraint and supply constraints resolve. Small modular reactors are being financed specifically for data centres; fuel cells are shipping; ERCOT has restructured its interconnection review; political pressure to accelerate permitting for AI infrastructure is intense everywhere. If time-to-power on Earth compresses to two or three years by any route, the availability argument, the strong form of the case, the one that survives when the cost case fails, evaporates entirely. Nothing else in this book saves it. 2b. Capital stays expensive (moves the date). Figure 18.4 shows the thesis needs cheap manufacturing and infrastructure-grade capital together, and Figure 18.7 assumes the second arrives as the asset class matures, flight heritage, then contracted revenue, then insurance, then debt, exactly the path renewables walked. If that path stalls, the crossover slips by years. This is a financing problem masquerading as an engineering one, and it will not be solved in a cleanroom. 3. GPU prices fall (moves the date). Every cost figure in Chapter 18 is measured against a commodity rental rate near $3.50 an hour, set during a capital-expenditure boom. Halve it and orbital compute is unviable at any manufacturing cost in the model, as is a large share of the terrestrial industry. The premium buyers of Chapter 18 are the hedge against this, and they are a smaller market than the sector’s addressable-market slides assume. 4. Demand growth slows (moves the date and efficiency may help rather than hurt). The whole edifice assumes compute demand outruns power supply indefinitely. Two things could break that: a slowdown in AI capital expenditure, or continued rapid improvement in performance per watt. Note that Nvidia claims roughly ten times the inference throughput per watt from one generation to the next. If efficiency gains outpace demand growth
for even a few years, the power crunch that justifies going to orbit could resolve on Earth. It is worth noting that this one cuts both ways, and Chapter 18 excluded the favourable half. Orbital costs scale with watts, radiators, arrays, mass, while revenue scales with useful output. Every doubling of tokens per watt doubles revenue against a fixed thermal and structural cost base, which improves orbital unit economics faster than terrestrial ones, where the power constraint was the thing being relieved. Efficiency is a risk to the rationale and a gift to the arithmetic, and which dominates is genuinely open.
The three that would be serious but survivable 5. Hardware refresh is never solved. The most likely of the failure modes, and the most under-analysed. If an operator must fly the same accelerators for a full spacecraft life, revenue per rack collapses as newer silicon arrives on Earth, and the returns in Part V’s model, which assume three-to-five-year hardware amortisation are simply wrong. This does not kill the sector; it caps it, and it shifts value to whoever solves servicing. 6. Thermal does not close at megawatt scale. Chapter 11 is honest that two-phase transport above a megawatt has no flight heritage. If it proves intractable, orbital compute stops at the hundred-kilowatt scale and becomes the edge-processing business of the 2040 downside case. 7. Hyperscalers vertically integrate. If SpaceX, Google and the large cloud providers build the compute layer themselves at a scale that makes independents irrelevant, the operator layer loses its value, though the power, thermal and relay layers keep selling into a larger market. This costs an investor the high-performance sleeve and leaves the picks-and-shovels intact, which is precisely why the portfolio in Part IV is weighted the way it is.
The ones I think are overrated as risks
Radiation was retired as a first-order risk by the testing described in Chapter 12; it is now a depreciation schedule. Launch cost is, per Figure 18.2, the fourth most important input and is already better than the model needs. Latency is a non-issue for the natural workloads. I could be wrong about all three, but each would need new evidence rather than new argument.
The falsifiers, what would change my mind, specifically
I would substantially reduce conviction if: a published spacecraft build cost for a large compute vehicle came in above $5,000/kg; time-to-power on Earth, by grid, turbine or reactor, fell below three years in two major US markets; commodity GPU rental rates halved without a corresponding premium market emerging; an operator disclosed a cost of capital above 18% for its constellation build; a second-generation compute satellite failed thermally on orbit; performance-per-watt improvements ran ahead of data centre demand growth for two consecutive years. I would substantially increase conviction if: an operator published a price for orbital GPU-hours at a premium to commodity and sold capacity forward to a named sovereign or defence buyer; an operator raised debt against an orbital compute asset, which would prove the capital-cost problem is solvable; a spacecraft build cost near $1,000/kg was demonstrated for a vehicle above ten tonnes; a two-phase loop was qualified above 500 kW; a payload swap was demonstrated on orbit in any application; an insurer wrote a policy on a compute node at a rate below five percent. Those are twelve specific, observable events. Write them down. Any of them arriving should move your view further than any amount of additional argument, including mine.