The picks and shovels trap

We started this work with a conviction, and it is worth stating at full strength before we take it apart. Every time humans have opened a genuinely new physical frontier, it has been among the largest wealth creation events of its era. And the wealth has concentrated not in the gold but in the picks and shovels, and specifically in transport. Nothing happens on a frontier until people and goods can reach it. Whoever moves them captures the value.

The American West is the cleanest case, and the market data is startling.

What the exchanges looked like

Before the frontier opened, American capital markets were a banking sector with a few other things attached. Finance stocks, meaning banks and insurance companies, represented over 95 percent of United States market capitalisation until 1825, and over 90 percent until the 1830s.⁴¹ The Second Bank of the United States alone accounted for between 50 and 80 percent of total market capitalisation at points in this period.⁴¹

Then transport arrived, first as canals and then, from the founding of the Baltimore and Ohio in 1828, as railroads.⁴² By the 1840s railroads had begun to dominate the exchange, and by the second half of the century the American stock market was, in capitalisation terms, a railroad market with some other companies listed alongside it.

Capital came from everywhere. American railroads marketed their stocks and bonds in London, Amsterdam, Paris and Berlin, and raised over a third of their capital in Europe, turning the United States into a net debtor in the process.⁴³

The fact that makes the case, and quietly undermines it

There is a cleaner way to see the dominance than any percentage. In 1845 the largest corporation in the United States was the Vermont Central Railroad. In 1865 it was the Illinois Central. A transportation company held the title of largest company in America continuously from 1845 until 1884.⁴⁴

Nearly forty years at the top of the market. Global Financial Data's own assessment is that no industry has ever dominated the stock market the way railroads did in the 1800s.⁴²

This is where our thesis stood when we began, and we suspect it is where most people stop. It is also, we now think, where the analysis goes wrong, and the clue is sitting in that last sentence. A transportation company was the largest company in America until 1884. What replaced it was Standard Oil, a company whose entire business depended on shipping things on somebody else's railroad.

Here is the distinction that undoes the thesis as we stated it. Share of market capitalisation measures how much capital an industry absorbed. It does not measure how much wealth that capital created. For railroads, the two numbers point in opposite directions.

The graveyard, in primary data

By 1876, half the nation's railroads, some twenty-one thousand miles' worth, had defaulted on their bonds and were in receivership.⁴⁵ That was the aftermath of 1873. The Panic of 1893 was worse. By June 1894, 192 railways were in receivership, of which 126 had entered in the preceding twelve months, together representing roughly $2.5 billion of capitalisation, around a quarter of the industry's total.⁴⁶ Between 1893 and 1897, companies owning about one third of all United States railroad mileage passed through bankruptcy.⁴⁷

The federal historical statistics record the mileage in the hands of receivers or trustees directly, and the series is worth looking at because it does not flatter anyone.

It was not a single bad decade

The obvious rebuttal is that 1893 was a macroeconomic accident rather than an indictment of the business. The best available study says otherwise. Examining railroads that went through receivership before the First World War, Lubben found they were more than two and a half times more likely to undergo another receivership or bankruptcy afterwards. The average reorganised railroad subsequently failed at more than twice the rate of railroads that had never gone through the process, and almost three times the rate of modern Chapter 11 debtors. J.P. Morgan's involvement made little difference to a railroad's ability to avoid distress.⁴⁸

Reorganisation did not fix these businesses. It postponed them.

The deflation defence, and why we do not accept it

There is a serious counterargument and it deserves setting out properly. Railroads were financed predominantly with fixed-coupon debt, much of it European money chasing yields well above domestic alternatives. Under the gold standard, prices fell steadily through the last third of the century. Fixed nominal interest against falling nominal revenues is a solvency machine that would sink any industry, and the record shows exactly that mechanism: overexpanded lines that did not generate the returns needed to service their bonds.⁴⁷

We think this explains the timing and not the outcome, for three reasons.

First, the leverage was endogenous. Capital-intensive assets producing an undifferentiated commodity output get financed with debt precisely because they lack pricing power. The contemporary record describes duplicative and uneconomical lines, routes paralleling established ones, and excess capacity feeding cutthroat competition that eroded freight rates.⁴⁶ That is not a monetary problem. That is what a commodity transport market does to itself.

Second, the failures continued after the deflation ended, as the receivership series above shows and as Lubben's post-war data confirms.

Third, and decisively: if deflation explains the railroads, why did airlines and telecoms fail identically under inflation?

The airline industry has never resolved this. IATA's own 2026 outlook projects a return on invested capital of 6.8 percent against a weighted average cost of capital of 8.2 percent, which is to say the sector does not cover its cost of capital even in a good year.⁴⁹ The fibre buildout of the late 1990s produced its own version. Telecom equities lost roughly $2 trillion of market value between 2000 and 2002, a figure corroborated independently by a Congressional hearing record, the ITU and the Economic Policy Institute.⁵⁰ Cisco, the definitive picks-and-shovels position of that cycle, peaked at about $80.06 in March 2000 and did not decisively surpass that level until 2025, roughly twenty-five years later.^51a

Three eras, three monetary regimes, one result. The constant is not the money supply. It is that all three sold an undifferentiated commodity, ton-miles, seat-miles and bits, out of an enormous fixed asset base.

Where the money went instead

Recall how Part I ended. Standard Oil displaced the railroads at the top of the American market in 1884, and Standard Oil was a customer of the railroads. Energy grew from 2 percent of United States market capitalisation in 1865 to 10 percent by 1901, and by 1901 Standard Oil was more than twice the size of the next largest company. Communications did not exist as a listed sector in 1861.⁴³

The railroads were built, they were paid for largely by people who lost money, and the wealth accrued to the businesses that the cheap freight made possible. That is the pattern we spend the rest of this essay trying to explain, and then testing against space.

So far we have a pattern and no mechanism. Three industries, across three monetary regimes, absorbed enormous quantities of capital, dominated their markets, and failed to return their cost of capital, while the businesses immediately downstream of them prospered. That is too consistent to be coincidence, and we need to name what is happening.

Our answer is a single idea, borrowed from classical rent theory and applied to frontiers.

The binding constraint is the input that is necessary for the output, cannot be reproduced with capital alone, and is scarce relative to demand. Economic rent flows to the most inelastically supplied factor. Everything else is competed down to its cost of capital.

This is not a novel claim in economics. What we think is underappreciated is how badly transport scores on it, and how reliably investors misidentify transport as the constraint when it is merely the most visible input.

Three diagnostic questions

One. Can it be reproduced with money alone? If yes, it will not hold rent for long. Track, hulls, fibre and airframes are pure capital goods. Anyone with a balance sheet and a few years can make more, and during a frontier boom a great many people have both. The contemporary record of the 1890s describes exactly this: duplicative lines, routes paralleling established ones, and excess capacity feeding rate wars.

Two. Does building more of it destroy its own scarcity? This is the trap, and it is where transport is genuinely unusual as an asset class. The entire purpose of building transport is to make distance cheap. Succeeding at that objective is the same act as eliminating your own pricing power. A railroad that has done its job has made the region it serves accessible, which is precisely the condition under which a competitor can build a parallel line into it.

Compare almost any other asset. A patent does not expire faster because you exploit it. An ore body does not become cheaper to mine because you have mined some of it. A brand strengthens with use. A licensed monopoly is enforced by a sovereign regardless of how much you use it. Transport is close to alone in that operational success is structurally self-defeating.

Three. What becomes more valuable when the constraint is relieved? This is the useful question, because it tells you where the rent has gone rather than where it was. Cheap freight makes land valuable. Cheap bandwidth makes attention valuable. Cheap launch makes orbital position valuable. In each case the transport builder hands a windfall to whoever owns the complementary scarce thing, and the complementary thing is usually something the builder does not own.

Two problems this leaves us with

We are being honest about the limits of our own framework, because it does not yet account for everything in the record.

It does not explain why anyone built the railroads. If the returns were this bad and the mechanism this predictable, the behaviour of the promoters looks irrational, and it was not. Part IV shows what they were actually being paid in.

Nor does it explain Suez and Panama, which are transport assets that have held rent for more than a century and should not exist under the framework as stated. Part V resolves that, and the resolution turns out to be the most useful thing in this essay for anyone allocating capital to space.

And it has nothing to say about businesses whose economics do not turn on physical scarcity at all. Software is the obvious case, and we set out that boundary explicitly in Part VIII rather than pretending the framework is universal.

We have argued that transport commoditises itself by design, and that the sector's history is a graveyard. That leaves an obvious puzzle. If railroads destroyed capital as reliably as the receivership data suggests, why did anyone build them, and why did a handful of railroad fortunes survive the wreckage entirely intact?

The answer is that those fortunes were not made in freight. They were made in land, and the land was awarded for laying track. Once you see the mechanism clearly in the nineteenth century, it becomes very hard to unsee in the twenty-first.

What Congress actually granted

The Pacific Railway Act of 1862 conveyed to the chartered companies every alternate odd-numbered section of public land, to the amount of five alternate sections per mile on each side of the route, within ten miles either side.¹ That is 6,400 acres for every mile of track laid. The 1864 amendment doubled the grant to twenty alternate sections per side.² Congress separately issued bonds per mile, escalating with terrain: $16,000 across level prairie, $32,000 across intermediate ground, $48,000 through mountains, raised in 1864 to $32,000 and $64,000.²

The structure matters more than the size. This was not a payment for transport services rendered. It was a conveyance of territory, released against completed track in forty-mile increments.³ Build the miles, receive the land. The railroad was the eligibility test.

The federal government had patented 131,230,358 acres to the railroads by 1933, supplemented by roughly 51 million acres from state governments.⁴ The Northern Pacific alone received 39.4 million acres across its system.⁴ The companies then monetised that land at a substantial premium to the statutory preemption price of $1.25 per acre at which unsold grant land reverted to open settlement.⁵

You did not get paid for laying track. You got paid in land, conditional on laying track.

The same instrument, in orbit

Spectrum and orbital rights are allocated on a first come, first served basis through the ITU. Priority attaches to whoever completes coordination procedures earliest, and later filers must design around earlier ones. Left alone, that rule invites exactly what the industry calls paper satellites: claims staked years before any hardware exists.

Resolution 35, adopted at WRC-19 and revised at WRC-23, closes that gap with a milestone schedule. A non-geostationary system must deploy 10% of its constellation within two years of the end of the regulatory period for bringing into use, 50% within five years, and complete deployment within seven.⁶

The enforcement provision is the part worth dwelling on. If a system misses a milestone, the Radiocommunication Bureau applies a deployment factor that shrinks the constellation recorded in the Master Register. Miss the first milestone and the recorded system is scaled to no more than ten times the number of satellites actually deployed. Miss the second and it is scaled to twice. At the final milestone, the constellation is recorded at exactly the number flown.⁷

A paper claim collapses to a multiple of physical hardware. That is the same instrument as a land grant released against completed track, expressed in the vocabulary of radio regulation rather than the General Land Office.

The claim now on the table

On 30 January 2026, SpaceX applied to the FCC for authority to launch and operate a non-geostationary system of up to one million satellites, the SpaceX Orbital Data Center System. The Space Bureau accepted the application for filing on 4 February.⁸ The satellites would occupy altitudes from 500 to 2,000 kilometres at 30-degree and sun-synchronous inclinations, within orbital shells spanning up to 50 kilometres each. SpaceX describes the system in its own application as a first step toward a Kardashev Type II civilisation, one able to harness the full power of its star.⁸

Read against the milestone regime, a filing of this size is not a deployment schedule. It is a claim, and under Resolution 35 it is worth only what SpaceX can physically launch against it.

The tell

Which brings us to the most revealing detail in the filing, and it comes from the Commission's own record rather than from trade coverage. SpaceX's application explicitly requested exemptions from the FCC's NGSO milestone requirements and deployment obligations under §25.164, and from its surety-bond requirements and obligations under §25.165, alongside waivers of the processing-round requirements at §25.156(d) and §25.157.¹⁰

That is the land speculator's move, made explicit. The most capable launch operator in history, filing for the largest constellation ever proposed, is simultaneously asking to be excused from the obligation to deploy it. We read this two ways, and both are useful. It is a candid signal of implementation uncertainty. It is also confirmation that the milestone regime is the binding instrument in the entire arrangement, because nobody petitions for relief from a rule that does not bite.

What this resolves

The framework also disposes of the awkward exception we left open in Part II. The Dutch East India Company paid dividends averaging roughly 18% of nominal capital for close to two centuries, which sits badly with a thesis that transport destroys capital. But the VOC was never really a shipping business. It held a chartered monopoly over a trade route, which is a licensed position enforced by a sovereign, and it operated ships in order to hold it.

The pattern is consistent across all three cases. The transport is the cost of admission. The asset is the licensed position that the transport qualifies you to hold, and the rent comes from a regulator, a legislature or a charter enforcing scarcity that the market would otherwise compete away.

Our argument so far has a hole in it, and we should walk straight into it. If transport commoditises itself by design, the Suez and Panama canals should not exist as profitable enterprises. They are transport assets, and they have borne economic rent for well over a century.

The resolution is a single distinction that we think carries the rest of this essay. A railroad is a route. A canal is a chokepoint. You can lay a parallel line to Chicago, and in the 1880s a great many people did, which is precisely how the rate wars started. You cannot build a parallel isthmus.

Suez beat its cost of capital for a century

This is testable, and it has been tested. A study in the Journal of Economic History computed internal rates of return on the Suez Canal from construction through to nationalisation, 1859 to 1956, using published accounts and imputing terminal values at nationalisation. For capital generally, for French shareholders and for the British government, returns ran at 8 to 9 percent against opportunity costs of 3 to 4 percent.¹¹

Set that against the route businesses. IATA's own 2026 outlook has the airline industry earning a return on invested capital of 6.8 percent against a weighted average cost of capital of 8.2 percent, in a good year.¹²

The concession ran 99 years from 1869 and allocated annual profits 75 percent to shareholders, 15 percent to the Egyptian government and 10 percent to the founders. Construction costs were recovered out of cumulative toll surpluses by the late 1870s, after which dividends settled at high levels relative to par.¹³

The detail that should discipline how we read the orbital land grab

The same study computed Egypt's own return. The Egyptian government earned 2 to 5 percent against an opportunity cost of 11 percent.¹¹ The sovereign that owned the ground the canal was cut through did materially worse than the concession holders who owned the right to charge for passage.

Owning the territory is not the same as owning the instrument that charges for crossing it.

We flag this because the space industry's language invites the opposite assumption. Orbital shells and lunar terrain are discussed as property. They are not. What is actually being allocated is a licensed right to operate, and the Suez record is a warning that these two things can come apart sharply, with the sovereign on the losing side.

Panama prices the rent in public, every week

Panama gives us something rare: a transparent, live market price for chokepoint access. Vessels without a reservation can bid at auction to jump the queue, and the auction fee sits on top of the standard transit toll of roughly $400,000.¹⁴

Before the recent disruptions, slots cleared at roughly $135,000 to $140,000. Through the 2023 drought, one vessel paid $2.4 million, and shipping companies paid around $235 million in auction fees across that year alone.¹⁴¹⁵ Through 2026, with the Strait of Hormuz closed and El Niño drawing down Gatún Lake, slots moved to roughly $385,000 in March and April, and average Neopanamax auctions reached $1.697 million.¹⁶ The cleanest single data point remains the 8 November 2023 auction in which one operator paid $3.975 million, just under $4 million, for a single priority transit slot.¹⁷

That is roughly a 28-fold move in the price of passage with no change whatsoever in the physical asset. The canal did not get better. The alternatives got worse.

Which gives us the actual formula

A chokepoint's rent is not unlimited. It is capped by the cost of the next best route. Suez can charge up to, and never more than, the cost of sailing around the Cape of Good Hope. Panama's auction price is simply the market computing that differential in real time, week by week.

And the cap binds, as 2024 demonstrated

We should not present chokepoints as unconditional monopolies, because Suez has just run the experiment. Rerouting around the Cape adds roughly 3,000 to 3,500 nautical miles and 10 to 14 days to an Asia to Northern Europe voyage, and up to $1 million in additional fuel per trip.¹⁸ For decades that differential was comfortably larger than the toll, and traffic stayed. When Houthi attacks from November 2023 added a risk premium that pushed the Suez route past the Cape route in expected cost, Suez trade volumes fell roughly 50 percent year on year within the first two months of 2024, while traffic around the Cape of Good Hope rose about 74 percent.¹⁹

Suez Canal Authority revenue fell from $9.4 billion in FY2022/23 to $7.2 billion in FY2023/24, and on a calendar-year basis 2024 revenue fell by more than 60 percent, a loss the Egyptian presidency put at close to $7 billion.²⁰ The chokepoint did not stop being a chokepoint. Its rent simply collapsed to zero the moment the alternative became cheaper in risk-adjusted terms.

So the correct formulation is narrower than the one we started with. A chokepoint is not a monopoly on passage. It is a monopoly on a differential, and it holds only for as long as the alternative route is worse.

The test this gives us

For any asset on a new frontier we can now ask a precise question rather than a rhetorical one: what is the alternative route, and what does it cost?

For an orbital shell at a given altitude and inclination, with spectrum priority attached, there is no alternative route in any meaningful sense. For lunar propellant, the alternative is hauling propellant up from Earth, and the cost of doing so is exactly and calculably what caps the price of the lunar product. That is a number we can put a range on, and we return to it in Part IX.

Until June 2026 this section would have been an argument. SpaceX listed on 12 June at $135 a share, pricing the company at roughly $1.77 trillion, and closed at a market capitalisation of $1.93 trillion on 17 August 2026 per stockanalysis.com, citing S&P Global Market Intelligence.²¹ The valuation is not the interesting part. The interesting part is that a company which had disclosed almost nothing for twenty-four years now files segment accounts, and those accounts settle the question this essay has been circling.

The launch business loses money

In the second quarter of 2026, SpaceX reported total revenue of $7.81 billion, up 92 percent year on year. The Space segment, which is the launch business, contributed $962 million of that. It spent $1,076 million on research and development in the same quarter, more than the segment earned, and posted an operating loss of $542 million.²² Launch cadence actually fell, to 38 missions against 46 in the prior-year quarter.²³

Set against that, the Connectivity segment produced $4,291 million of revenue, up 66 percent, with $1,656 million of operating income and $2,597 million of adjusted EBITDA. After its own segment capex of $1,367 million, Connectivity threw off roughly $1.2 billion of self-funded cash, and it is the only part of SpaceX that currently does.²² Starlink passed 12 million subscribers, double the prior year, though ARPU fell from $85 to $66 a month as the network expanded into lower-income geographies.²³

This is the thesis, stated in a filing

Recall the divergence we noted earlier, and let us be precise about it, because the precision is what makes it powerful. Falcon 9's published customer price has risen from $54 million historically to $74 million in 2026, over a period in which the underlying cost of access to orbit fell dramatically.²⁴ The published series runs $54m, then $62m in 2020, $67m in 2022, $69.75m in 2024 and $74m today, and SpaceX explicitly attributed the 2022 increase to inflation.²⁴

Three things are being conflated whenever this industry is discussed, and separating them is the whole argument. Technology cost is what production costs the operator. Market price is what a customer pays. Economic value is what the customer actually gets. SpaceX's technology cost has collapsed while its market price has risen, and those are not contradictory. They imply a widening spread between cost and price, sustained by scarcity, reliability and the absence of a competitor at scale.

We argued that SpaceX captured that spread as margin rather than passing it to customers, and that it could do so because most of its transport is never sold on an open market at all. It is consumed internally.

The segment accounts confirm the mechanism and go further than we expected. The cost reduction does not merely fail to show up in launch pricing. It fails to show up in the launch segment at all, which runs at a loss. It surfaces one layer up, as Connectivity margin, and increasingly two layers up. One sell-side analysis put it more bluntly than we would have dared: nobody owns the stock for the Space P&L, they own it for what Starship does to the marginal cost of the other two segments.²³

SpaceX did not solve the transport problem. It declined to be a transport company, and now runs launch as an internal cost centre.

The migration has already gone one layer further

In February 2026 SpaceX absorbed xAI in an all-stock transaction, converting itself into a satellite connectivity and artificial intelligence conglomerate that also launches rockets.²⁵ The AI segment reported $2,561 million of revenue in Q2, up 247 percent year on year and 213 percent sequentially, with segment adjusted EBITDA turning positive at $1,146 million even as the operating line stayed negative under $1,885 million of quarterly depreciation.²² Deployed compute capacity moved from 0.4 GW a year earlier to 1.4 GW, against a target of roughly 15 GW by the end of 2027.²³

Read against Part IV, the picture is coherent. Launch qualifies the company to hold orbital shells and spectrum. Those positions carry a connectivity business that funds the group. That business now funds a compute business, and the orbital data centre application would eventually place the compute inside the positions the launch capability defends. Three layers of value migration inside a single balance sheet, compressed into about fifteen years.

And now the part that should worry us

We would be writing propaganda rather than analysis if we stopped there, because the same filings show SpaceX running the exact pattern that Part II identified as fatal.

Capital expenditure in the second quarter alone was $18.37 billion, of which $15.83 billion went to AI compute. That is 2.4 times the revenue the company generated in the same three months.²³ Visible Alpha consensus projections, which are analyst estimates and not company guidance, have group capex rising from $48.7 billion in 2026 to $118.4 billion in FY2028, with total debt growing more than fivefold from $41.7 billion to over $218 billion across the same span.²⁶

Long-duration, capital-intensive infrastructure, built ahead of demonstrated demand, financed increasingly with debt. That is a precise description of the American railroad system between 1873 and 1893. The company currently holds roughly $100 billion of cash after raising $85.7 billion in the IPO and a further $25 billion in investment-grade bonds, so the near-term liquidity question is not close.²⁷ The medium-term question is whether the assets earn their cost of capital before the debt matures, and that is precisely the question the railroads answered badly.

The counter-argument, which deserves a hearing

SpaceX's CFO has an answer, and it is not a weak one. He argues that AI compute capital carries a payback of under one year, which makes it behave less like infrastructure and more like an item that would otherwise sit in cost of goods sold.²³ If that holds, the railroad analogy fails, because a one-year payback is working capital rather than a thirty-year bond against a fixed asset.

We think this is the single most important thing to test in the whole SpaceX story, and it is testable. A GPU cluster with a sub-one-year payback and a five-year useful life is a spectacular business. The same cluster with a three-year payback and a three-year competitive half-life is Global Crossing. The distinction turns on how long compute pricing holds, which is a question about the AI market rather than about space, and nothing in SpaceX's launch advantage answers it.

The market appears unconvinced so far. The stock priced at $135, opened around $150, reached $225.64 on 16 June and has since traded down to roughly $146, having touched an all-time low of $104.83 on 3 August.²⁸ A little over four percent of the company floated, so price discovery on this question is happening in a very thin market.²⁹

A framework that only confirms what we already believed is not worth much. So we want to take the most exciting story in the sector, orbital compute, and put our own questions to it. We should say at the outset that if this works it is one of the most interesting things happening anywhere in technology, and we are actively looking at the category. What follows is not a verdict. It is the list of things we would need to see resolved.

Why the case is genuinely compelling

The physics behind orbital compute is not in dispute and it is not marginal. Solar flux at the top of the atmosphere is about 1,361 W/m², and a dawn-dusk sun-synchronous orbit delivers a capacity factor above 95 percent against a US median for terrestrial solar closer to 24 percent.⁵¹ There is no water consumption, no grid interconnection queue, no permitting process and no community opposition. Set against terrestrial AI infrastructure that is currently constrained by all four, the appeal is obvious, and the leading estimates in the category put orbital energy at figures one to two orders of magnitude below terrestrial wholesale power.⁵¹

If launch costs fall far enough, this becomes a category with no terrestrial equivalent. That is why serious capital is moving into it and why SpaceX has staked its largest regulatory claim on it.

Our questions are four, in ascending order of how hard we think they are to answer.

One. Does the comparison include the computers?

The published models in this category tend to compare energy and cooling costs while holding hardware constant, and hardware is where almost all the money is. A 40 MW cluster is roughly 25,000 GPU servers, 300 all-flash arrays and 1,700 switches. The servers alone run something like $12 to $13 billion; include the shell, storage and networking and the figure approaches $24 billion.⁵² Ten-year energy and cooling savings in the models we have seen come to a few hundred million dollars.

Energy is not the cost driver of an AI data centre. Silicon is. That does not make the category uninteresting, but it does mean the pitch has to rest on something other than the power bill, and we would want to understand what that something is. The strongest version we have heard is not cost at all: it is access to capacity that simply cannot be sited terrestrially at any price. That is a better argument and a different one.

Two. What does the launch manifest actually look like?

Most models in the category assume launch counts and price points that are ahead of demonstrated capability. Sun-synchronous orbit, which the power case generally requires, carries a payload penalty against LEO of roughly 40 to 60 percent, and Starship's SSO capacity has not been publicly stated. The gap between current commercial pricing and the figures these models require is close to two orders of magnitude. On published prices, Falcon 9 at $74 million for 22,000 kg to LEO implies roughly $3,364 per kilogram, and Vulcan at a commonly cited $110 million for 27.2 tonnes implies about $4,044.⁵³ Models in this category typically require figures in the tens of dollars per kilogram, and the most aggressive of those numbers trace back to informal public statements rather than published price lists.⁵³ The European Space Policy Institute has described the $10 million per Starship flight assumption underpinning many of these models as unrealistic in the near term.⁵⁴

This is the most tractable of our four questions, because it resolves itself with cadence. Every Starship flight is evidence.

Three. Is it a facility or a consumable?

One widely circulated independent cost model for a 1 GW orbital installation prices roughly 4,300 satellites and about 30 million kilograms of mass to orbit, at somewhere between $42 billion and $51 billion depending on the run, against roughly $16 billion for a terrestrial facility of the same capacity. The structural difference is the replacement cycle: the orbital version is written off and replaced entirely after five years, where a terrestrial data centre runs for decades and is repaired piecemeal.⁵⁵

We note that GPU refresh cycles are themselves compressing toward five years, which cuts in favour of the orbital case rather than against it, and this is an area where we think the sceptics may be arguing from an outdated premise about asset life.

Four. Thermodynamics, which has no learning curve

This is the one we cannot get comfortable with, and it is different in kind from the other three because it is not an engineering or financing problem.

In vacuum there is no convection. Every watt a processor consumes must be radiated away as infrared from a surface, and radiated power scales with the fourth power of temperature under the Stefan-Boltzmann law. At temperatures electronics survive, radiative cooling tops out at a few hundred watts per square metre. ISS radiator panels manage roughly 130 to 200 W/m². Terrestrial convective cooling with air and water moves thousands to tens of thousands of watts per square metre.⁵⁶

Radiators and solar arrays together can consume 65 to 70 percent of total satellite mass, and thermal coating degradation implies launching at least 40 percent more radiator mass than the clean-sheet calculation suggests.⁵⁶ Space-grade photovoltaics also depend on germanium substrates whose supply is concentrated in China.⁵⁶

The escape route exists and it is worth naming, because it is where we would focus diligence. Radiated power scaling with the fourth power of temperature cuts both ways: a chip architecture that tolerates substantially higher junction temperatures shrinks the radiator dramatically. That is a semiconductor research problem rather than an aerospace one, which means the binding constraint on orbital compute may not sit in the space industry at all. Any team that has a credible answer there is solving the hard part.

The spread is the finding

We would not call the thesis dead, and we would be wary of anyone who does. Independent models converge on roughly three to four times terrestrial cost today, narrowing to a modest premium in the early 2030s.⁵⁷ That is a real trajectory rather than a fantasy, and categories that look four times too expensive have a habit of not staying that way. What we would resist is the framing that this is imminent, or that falling launch costs alone will close the gap.

Which brings us back to the pattern

Note what happens to SpaceX under either outcome. Even the most sceptical published model still requires 30 million kilograms of mass to orbit per gigawatt. If orbital compute works, SpaceX launches it. If it never earns its cost of capital, SpaceX still launches it, and somebody else absorbs the loss.

That is the picks-and-shovels position in its purest form, and it is precisely what Nortel and Lucent occupied in 1999 through vendor financing. It ended badly for them, and the reason is instructive: they had funded their own demand.

Which is the risk we would flag hardest. Following the xAI absorption, SpaceX is the launch provider to this market, the largest filer for orbital data centre capacity, and one of the largest buyers of terrestrial compute. It is on every side of the trade at once. The Q2 capex of $18.37 billion, of which $15.83 billion went to AI compute, is the company underwriting demand for its own launch capacity out of its own balance sheet. That can be vertical integration or it can be circular financing, and from the outside those look identical until they do not.

If the framework is right, the space economy should already contain a profit pool that looks like a recurring toll on a licensed position, and it should not be where the industry's own marketing points. Both turn out to be true.

First, what is not the cash cow

Earth observation is the sector's most persistent story and its smallest business. Market estimates run $5 to $7 billion depending on market definition, which is one to two percent of a global space economy of about $429 billion.⁵⁹ Planet Labs became the first of the new generation to reach positive adjusted EBITDA and free cash flow in late 2025, roughly fifteen years after founding, on quarterly revenue of $81.3 million.⁶⁰ It is now a genuine business with gross margins around 58 to 61 percent, more than 90 percent recurring revenue and a $900 million backlog driven substantially by defence demand.⁶⁰ It is not a cash cow. It is a niche that took fifteen years to become viable.

Ground equipment at $165.2 billion is the largest segment and the most misleading. It is overwhelmingly navigation chipsets embedded in phones and vehicles, which is consumer electronics revenue booked to the space column, and the GPS signal those chips receive is a free government utility on which nobody collects rent. Published estimates of the GNSS market span orders of magnitude, from about $3 billion to $335 billion, because the term can refer to receiver hardware, to downstream services, or to the entire enabled economy.⁶¹ That spread is not a disagreement about the world. It is a warning about category definition, and we treat it as evidence rather than as a number.

The government segment, meanwhile, is a cost-plus contracting pool. Reliable, but not a rent.

The cash cow was satellite television, and almost nobody says so

For thirty years the profit engine of the space economy was direct-to-home broadcast. Geostationary operators held scarce orbital slots, sold capacity to broadcasters on multi-year contracts, and ran adjusted EBITDA margins around 55 percent.⁶² That is precisely the structure Part IV described: a licensed position, allocated by regulators, generating a recurring toll from customers who had no alternative route.

It is now in structural runoff, and the operator disclosures are unambiguous.

Telesat attributed its decline principally to broadcast contracts not renewing.⁶³ The licensed position did not stop being licensed. The service riding on it became obsolete, which is a distinction we return to below.

The replacement is the same business on different hardware

Satellite service revenues are forecast to grow from about $101 billion in 2024 to $122 billion by 2034, with every dollar of that growth coming from data, which triples to $67 billion and overtakes video around 2033.⁶⁴ Recurring subscription revenue over a spectrum-and-orbit position you exclusively control, sold to consumers and enterprises instead of broadcasters.

Now look at the supply side, because this is where Part II comes back.

Total capacity supply rose three and a half times between 2023 and 2025, reaching 105 Tbps. Ninety-five percent came from non-geostationary systems, and 93 percent of that is Starlink. Non-geostationary systems are expected to account for over 99 percent of all new capacity added between 2025 and 2030. Capacity prices are falling as a direct result, and geostationary operators have largely stopped ordering large satellites.⁶⁴

This is a textbook overbuild. It is the fibre glut and the 1880s rate war, with the same mechanics and the same falling unit prices. By every argument in Part II it should be destroying value.

It is not destroying value only because one company owns 93 percent of the new capacity. Monopoly is what converts an overbuild from suicide into a moat.

We think this is the single most important thing an investor can understand about the sector. If two or three well-funded operators had built that capacity, satellite bandwidth would be a commodity by 2028 and none of them would earn their cost of capital. The economics of Starlink are not primarily a function of reusability or of vertical integration. They are a function of there being one builder.

Why the sector maps itself wrongly

The published ecosystem maps of this industry all organise along the same axis: access at one end, meaning launch and spacecraft manufacture, then the space segment, meaning what operates in orbit, then downstream, meaning terminals, analytics and applications.⁶⁵ It is a sensible taxonomy by technical layer, and it is the same axis this essay has been travelling along, which makes it a useful thing to hold our conclusions against.

Two observations follow, and we think both are more interesting than anything on the maps themselves.

There is no box for the asset. No ecosystem map has a category for spectrum filings, orbital shells or ITU priority, and the reason is simply that these are not companies. They are assets, and they sit underneath the map rather than anywhere on it. Part IV argued that this invisible layer is precisely where the rent accrues, and a taxonomy that cannot draw the thing we think matters most is quietly making the argument for us.

The density runs backwards. Access is consistently the most crowded column on these maps and downstream the thinnest. Our framework predicts returns in roughly the inverse order. Company count as an inverse indicator of expected return is a crude heuristic and we would not lean on it hard, but it is the same shape as the American railroad sector at more than 60 percent of market capitalisation on the eve of 1893.

Where our framework does not reach

We should be explicit about a limit rather than let a reader find it. Our argument is built on physical assets and inelastic supply, and there is a substantial part of this sector to which that machinery simply does not apply.

The software and data layers are the clearest case: onboard software, mission operations, space situational awareness, security, storage and the analytics platforms that sit on top of imagery. These are neither scarce assets whose supply a regulator enforces, nor commodity outputs produced from enormous fixed-cost bases. They may well be excellent businesses, and if they are, it will be for conventional software reasons, meaning near-zero marginal cost, switching costs and data network effects, rather than for any reason to do with orbits. We have nothing useful to add there.

Two smaller gaps deserve naming too. We dismissed ground segment and terminals as consumer electronics booked to the space column, which is a fair description of the GNSS chipset business and an unfair one applied to the largest revenue line on the map. And we treat defence and government demand as a cost-plus pool in a single sentence, when it is in fact the buyer underwriting most of Earth observation and a growing share of connectivity. Both deserve their own analysis and neither gets it here.

A framework about rent from inelastic supply, applied to a business with near-zero marginal cost, produces confident nonsense. We would rather mark the boundary than stretch across it.

The warning inside the good news

Before we let this become a bull case, note what the geostationary operators demonstrate. They held a near-perfect monopoly on a genuinely scarce licensed position, enforced by regulators, for three decades. They are being destroyed anyway, because streaming made their service obsolete regardless of how securely they owned the shell.

A licensed monopoly over a resource nobody needs is worth nothing. The position is necessary and not sufficient, which means the demand question has to be answered separately and on its own evidence. That is what Part IX does.

Every frontier in this essay created wealth for the same reason, and it was not that people went there. The American West sent grain, cattle, timber and gold east. The VOC sent spices to Amsterdam. Suez and Panama earn because cargo passes through them in both directions. A frontier is valuable when things come back.

This gives us a test that almost nothing in the space economy passes, and we think it is the most useful single filter an investor can apply to the sector.

The gravity well is a one-way filter

Getting mass back down from orbit is not free. Delivering small payloads to orbit and returning them currently runs somewhere between $25,000 and $100,000 per kilogram.³⁰ That sets a floor on the value density of anything worth manufacturing off Earth.

Only products where mass is nearly incidental to value clear the threshold. And one export clears it by an infinite margin, because it has no mass at all.

Bits are the only thing that comes down the gravity well for free. That is the entire reason Starlink is the only large, profitable, purely-space business in existence.

We would put this to any founder pitching a space business: what comes back, and what does it cost per kilogram to send? Applied honestly, the question dissolves most of the sector.

In-space manufacturing does not pass

The physics here is real and we do not dispute it: suppressing buoyancy-driven convection genuinely improves crystal growth and fibre drawing, and the resulting materials have properties that cannot currently be produced on the ground. Our concern is with the market rather than the science.

The category's best-funded venture has raised on the order of $329 million, and its revenue to date comes principally from hypersonic testing for defence customers rather than from manufacturing, with pharmaceutical royalties still at an early stage.³¹ Its most recent capsule return, in May 2026, was a government-funded reentry test rather than a production run.³² On the fibre side, more than seven miles of ZBLAN were drawn aboard the ISS in early 2024, and NASA was careful to note that the fourth objective, fibre ten times better than the ground equivalent, remained unproven pending analysis of the returned samples.³² Aerospace America's assessment is that the space manufacturing market does not yet exist.³⁰

Four decades of attempts, a genuine physics advantage, and a market that has not yet arrived. We read that as evidence about demand rather than about timing, but we hold the view loosely: a single qualified pharmaceutical product with a real royalty stream would change it quickly, and that is the milestone we watch for.

Space solar sidesteps the filter, and now has contracts

Beamed power is the one category that solves the return-flow problem by exporting photons rather than mass, and it has now moved from concept to commercial agreement. In April 2026 Meta signed a capacity reservation for up to one gigawatt of space-collected solar, under an architecture that places satellites in geosynchronous orbit beaming near-infrared light down to existing terrestrial solar farms so those assets can generate at night.³³ Orbital demonstration is planned for 2028 and commercial delivery for 2030.³³

We think this architecture is materially cleverer than classic space-based solar, because using existing solar farms as the receiver removes the need to build new rectennas and gives the whole thing a plausible path to revenue on infrastructure that already exists. The open question is simply time: first delivery is four years out, and the agreement is a capacity reservation rather than a purchase.

The strongest candidate, and the number that decides it

Lunar polar volatiles, monetised as propellant rather than as water, is the one beyond-Earth asset with the structure of a genuine chokepoint. The Moon's axial tilt of about 1.5 degrees produces permanently shadowed crater floors that have cold-trapped volatiles for billions of years, sitting immediately adjacent to peaks of near-continuous sunlight. NASA narrowed 13 candidate regions to 130 sites meeting Human Landing System requirements, and Artemis III crews can traverse only about two kilometres from the landing point.³⁴ The set of locations that simultaneously offers accessible shadowed terrain, continuous solar power, landable slope and direct-to-Earth communications is a handful of ridgelines, and China's Chang'e-7 is targeting the same ground.

Part V gave us the formula. A chokepoint's rent is capped by the cost of the next best route, and here the next best route is calculable: hauling propellant up from Earth. So the entire lunar thesis reduces to one comparison, and the published literature straddles it.

Jones et al. found the cheapest option was commercial launch at $40,000 per kilogram to cislunar space, against $78,000 for the cheapest lunar ISRU architecture, and concluded that launching from Earth was the more cost-effective route.³⁵ Bennett et al. reassessed that study with improved economies of scale on the lunar surface hardware and reached costs well below the commercial transport figure Jones had assumed.³⁶

The physics argument for the optimists is the strongest thing in the debate. Best-case payload mass fraction from Earth to geostationary transfer orbit is roughly 2 percent. From the lunar surface to the same destination it is about 48 percent, some 24 times higher.³⁷ Metzger's modelling finds lunar propellant holds an absolute advantage down to distant retrograde orbit and GEO from year one, reaching GTO by year five and approaching LEO by year thirty, with a gearing threshold above which lunar wins that one candidate technology exceeds by an order of magnitude.³⁷

We are not going to adjudicate this, and we would distrust anyone who claims to. What matters is that the question is now precisely posed rather than rhetorical. Lunar propellant is worth exactly what it saves against the Earth-launch alternative, and falling launch costs lower the cap every year. Starship is therefore simultaneously the thing that makes lunar operations possible and the thing that makes lunar propellant harder to justify.

The caveat that could end it

Permanently shadowed does not mean icy. Some shadowed regions show little or no hydrogen signature, while some sunlit regions unexpectedly do show hydrogen in the upper metre of regolith.³⁸ No commercial-grade in-situ assay has been conducted anywhere on the Moon. The entire thesis rests on a measurement nobody has taken.

Which gives us the right historical analogue, and it is not the railroads. It is the Northwest Passage: three centuries of capital and lives spent searching for a route that turned out to exist and not to be economically useful. That is the honest bear case, and an essay that names it is worth more than one that does not.

We began with a conviction that frontiers create wealth and that the wealth sits in transport. Half of that survived contact with the data. The half that did not is the more useful half.

Transport is where the capital goes first and where it dies, because building it destroys the scarcity that made it valuable. Over 40,000 miles of American railroad sat in the hands of court-appointed receivers in 1894, and over 42,000 miles again in 1934. Airlines still do not cover their cost of capital. SpaceX, which holds the most complete launch monopoly in the history of the industry, ran its launch segment at a $542 million operating loss in a single quarter. The exception is the chokepoint: Suez returned 8 to 9 percent against a 3 to 4 percent opportunity cost for a century, because you can build a parallel railway and you cannot build a parallel isthmus.

The cash cow is always a recurring toll on a licensed position, and transport is the qualification test rather than the prize. The railroads were paid in 131 million acres released against completed track. Geostationary operators were paid in orbital slots rented to broadcasters at 55 percent margins. Starlink is paid in subscriptions over spectrum it holds because it can physically launch against its filings. In each case the transport is the cost of admission, and the asset is the position a regulator or a legislature agrees to defend.

There is no permanent cash cow beyond Earth. There is one candidate with the right structure and no revenue, lunar polar propellant, whose value is capped precisely by what it saves against hauling propellant up from Earth, and whose entire premise rests on an ice assay nobody has performed. And there is one real business, which works because it found the only export that does not have to climb back down the gravity well.

What would change our minds

We would abandon this framework if any of three things happened. If SpaceX's launch segment turned durably profitable on external customers rather than internal transfer, transport would have escaped the trap on its own terms. If orbital compute reached cost parity before 2035, the thermodynamic constraint we treat as binding would have proved tractable and our reasoning about physics-versus-engineering constraints would be wrong. And if an in-situ lunar assay returned commercial-grade ice concentrations at an accessible site, the chokepoint would stop being hypothetical and the whole beyond-Earth section would need rewriting in a hurry.

We would also note the risk that our own subject runs. SpaceX is spending 2.4 times its quarterly revenue on capital projects, with consensus debt rising toward $218 billion, while occupying every side of the orbital compute trade at once. That is the 1873 structure with better technology, and the essay's own Part II applies to it.

The right analogy is not the railroad

For the beyond-Earth thesis specifically, we think the closest historical parallel is not the transcontinental railroad but the Northwest Passage. Three centuries of capital, ships and lives were spent searching for a route that turned out to exist and turned out not to be economically useful. The searchers were not fools and the passage was not a myth. It was simply worth less than the cost of using it.

SpaceX is priced as though the frontier beyond low Earth orbit will pay. It is funded entirely by the bits.

That is not a bear case. Starlink is among the best businesses created this century, and the orbital positions it defends are genuinely scarce and genuinely enforced. It is a statement about where the evidence currently stops. The company has proved it can win the transport layer and then decline to be a transport company, twice. Whether it can do so a third time, in a domain where the constraint is thermodynamics rather than economics, is the question we would want answered before underwriting anything past low Earth orbit.