There is a comfortable story making the rounds. It goes like this: Starlink has won. The constellation is up, the dishes are shipping, the subscriber count is climbing, and from here it throws off cash like a toll road — a piece of infrastructure that, once built, simply earns. The same story usually carries a corollary: fiber is finished. Why trench a neighborhood when you can beam bandwidth from orbit?
Both halves are wrong, and they are wrong for the same reason. They mistake a manufacturing operation for an annuity.
I spent the last piece on the valuation, A Trillion-Dollar Markup, on what $1.75 trillion implies and what it quietly assumes. This is the sequel, and it asks a narrower, more useful question. Set the price aside. Look at the asset. Does the thing underneath the number behave like an annuity at all?
It does not. And once you see why, the suburb question, the one everybody waves away, answers itself.
The annuity test
An annuity is boring on purpose. You spend the capital once. The asset sits there. It earns for a very long time, and the cost of keeping it earning is small relative to what it produces.
Fiber passes this test almost perfectly. The hard money is the civil works… the trenching, the permits, the labor of getting glass into the ground or onto the pole. That is brutal up front. But buried fiber has a service life measured in decades, commonly twenty-five to forty years, and the glass itself does not really wear out.1 Fiber is expensive to deploy, but relatively inexpensive to operate and maintain once built.2
You light it, and then you mostly collect.
Better still, the capacity is almost free to expand. The bottleneck in a fiber plant is rarely the fiber; it is the electronics on each end.3 Want more throughput on the same strand? Light another wavelength. The marginal cost of serving the tenth household on a street you already passed is close to nothing, and the marginal cost of giving every household ten times the speed is an equipment swap, not a new dig.
That is what an annuity looks like. Spend once. Earn for thirty years. Expand capacity for pennies.
Now hold Starlink up to the same test.
The five-year fact
A Starlink satellite is designed for roughly five years in orbit.4 At the operational altitudes the constellation uses, that is not a marketing figure; it is physics. The atmosphere is thin up there, but it is not absent, and a satellite that low is in a slow, permanent fight with drag. It burns propellant just to hold station. When the propellant runs low or the hardware ages out, the satellite is deliberately lowered and allowed to reenter and burn up. Then it is replaced.
There are something north of 10,400 working satellites in the constellation today.5 Five-year lives mean that, just to stand still, just to keep the network exactly the size it is now, SpaceX must build, launch, and deorbit on the order of 2,100 satellites every single year.6 Forever.
Read that again, because it is the entire argument. The revenue-generating base does not sit there. It falls out of the sky on a five-year clock, and the only way to keep earning is to keep manufacturing.7
This is not infrastructure in the sense that a toll road or a fiber plant is infrastructure. It is a high-throughput orbital telecom machine that has to keep producing assets, launching them, and disposing of them, in perpetuity, or the business stops. The capital expenditure is not a hump you climb once. It is a treadmill, and the treadmill never turns off.
The hard numbers
Here is where the bull and the bear actually have to meet, because the treadmill is only fatal if the per-lap cost is high. It isn’t — and that’s the part the doom takes get wrong as badly as the annuity takes do.
Start with the launch. SpaceX does not publish internal costs, but the convergent estimate from analysts and former employees puts the marginal cost of a reused Falcon 9, the kind that flies Starlink, at roughly $15 to $20 million per flight, against a $67 million list price to outside customers. The bottom-up math agrees: an expendable upper stage in the single-digit millions, a booster amortized to about a million dollars across thirty-plus flights, fuel that costs less than a nice car.
Call a Starlink launch $17 million internal. Each of those flights carries about twenty-five V2 Mini satellites. So the launch share per satellite is roughly $600,000 to $800,000.
Now the satellite itself. This is the softer number, but it is anchored. SpaceX’s own leadership has repeatedly said the per-satellite cost is well below $500,000, and some scale estimates run as low as $250,000. Take the honest band as $250,000 to $800,000 per bird, probably under half a million.
Add them. The all-in marginal cost of putting one more V2 Mini in orbit is somewhere around $1.0 to $1.5 million, manufacturing plus its slice of a launch. Reusability is the reason that number isn’t three times larger, and reusability is the reason this business exists at all.
Run the treadmill at that cost:
At today’s ~10,500 satellites, steady-state replacement is about $2 to $3 billion a year. That is not a crisis. It is a manageable, recurring bill for a company with real revenue.
But the constellation is not meant to stay at 10,500. The Gen2 ambitions run to 30,0008, and the broader long-range number often cited is 42,000, based on 12,000 previously authorized satellites plus paperwork for up to 30,000 additional spacecraft.9 Scale the same treadmill: a 30,000-satellite network on a five-year clock needs ~6,000 replacements annually10, $6 to $9 billion a year, before a single dollar of growth or upgrade.
At 42,000 satellites, the replacement requirement alone would be about 8,400 satellites per year before growth or upgrades.11
That is the shape of the thing. Cheap per unit, ruinous in aggregate, and permanent. The bigger the network gets, the larger the perpetual rebuild bill becomes, and the more the whole enterprise depends on driving the cost-per-launched-satellite down faster than the headcount of satellites goes up.
Annuity versus machine
The cleanest way to feel the difference is in one line of the income statement that nobody romanticizes: depreciation.
Fiber depreciates at maybe 3-4% a year. The plant lasts decades; the accounting reflects it. The asset is patient.
Starlink’s in-orbit base depreciates at something closer to 20% a year12, because the assets are gone in five years. Every year, a fifth of the revenue-generating fleet is written down to zero and physically dropped into the atmosphere. The company does not own a network so much as it rents one from the laws of orbital mechanics, and the rent is paid in new satellites.
So what is Starlink, as an economic object? It is not a utility. Utilities are patient capital with regulated returns and assets that outlive their financing. Starlink is closer to a chip fab or an airline, a business defined by continuous, expensive fleet renewal, where standing still costs billions, and the competitive position has to be re-earned every cycle.
- A fab is only as good as its next node.
- An airline is only as good as its next fleet refresh.
- Starlink is only as good as its next generation of satellites and its next reduction in cost-per-bit.
That is a fine business. It is not an annuity. And the gap between those two descriptions is most of the gap in the valuation.
What Starship actually changes
The bull’s rebuttal is real and worth stating fairly: Starship breaks the cost curve. This is the lever that determines whether the treadmill is a footnote or a noose.
The third-generation satellites change the unit economics dramatically.13 Each V3 satellite is built to deliver on the order of a terabit per second of downlink. A single Starship can carry up to sixty of them, which means one launch adds roughly sixty terabits per second to the network, more than twenty times what a Falcon 9 batch of V2 Minis contributes today.14 Starship V3 flew its debut in May, carrying dummy V3 payloads, and operational deployment is the next milestone.15
Notice precisely what this fixes and what it does not.
It fixes the cost-per-delivered-bit. If you can put twenty times the capacity in orbit per launch, the dollars-per-terabit can collapse even if the absolute capital outlay stays enormous.16 That is genuinely transformative for the economics of serving demand.
It does not fix the treadmill. The satellites still die in five years. A bigger, cheaper-per-bit constellation is still a constellation that must be rebuilt on a clock. Starship makes each lap cheaper and more productive; it does not remove the laps. The depreciation identity survives the upgrade intact. You can run the treadmill faster and carry more on it; you cannot get off it.
So the right way to hold the V3 story is this: it is the difference between a machine that is barely solvent and a machine that is comfortably profitable. It is not the difference between a machine and an annuity.
The business never stops rebuilding
The industry is already trying to solve the treadmill problem by reaching for a new, even more precarious lever: moving the compute itself into orbit.17
The pitch is pure seduction, free solar energy, twenty-four-hour uptime, and no terrestrial permitting battles. But if the Starlink constellation is a machine that requires a five-year rebuild, an orbital data center is a machine that requires an insurance policy that the market refuses to write.
When you scale the physics from a single satellite to a gigawatt-class orbital facility, you aren’t just adding power capacity; you are building a kilometers-wide target in a debris field that, at 550 km, has reached parity with the active population.
The free solar comes with a much larger orbital risk surface: collision exposure, debris proliferation, insurance modeling uncertainty, and potentially fragile economics at scale.18 The math doesn’t just suggest the risk is high; it suggests the infrastructure is fundamentally incompatible with the orbital environment.
The insurability crisis isn’t a footnote; it is the tell that the model, much like the annuity story, collapses under the weight of its own physical requirements.
Why the suburb was never the prize
Here is the payoff, and it falls straight out of everything above.
A fiber line serves a household with a dedicated, effectively unlimited pipe. A Starlink satellite serves a cell, a patch of ground, with a fixed, shared bucket of capacity, divided among everyone beneath it. Those are opposite physics. Fiber’s constraint is the cost of reaching you. Starlink’s constraint is the number of people already standing in your patch of sky.
In a dense suburb, fiber’s economics get better with density: more homes per mile of trench, lower cost per subscriber, trivial capacity expansion. Starlink’s economics get worse with density: every new subscriber in a cell degrades service for all the others, and the only fix is more satellites overhead — the most expensive capacity expansion in the entire telecom industry. Density is the one place Starlink can least afford to win, and the one place fiber is cheapest to defend.
This is not a forecast. It is already in the pricing, in plain sight.
In high-demand areas, Starlink has layered on congestion surcharges, reported demand surcharges have reached as high as $1,500 in high-demand areas.19 In low-congestion areas, the company does the reverse, discounting hardware to as little as $89 to fill the empty cells. That is not a pricing accident. It is overt geographic yield management. Starlink is pricing customers out of the cells it can’t afford to serve and into the ones it can.20 The congestion charge is the price mechanism doing the rationing that fixed orbital capacity cannot.
A company that intended to conquer the dense suburb would not be charging people more to live there. It would be subsidizing them in. The pricing tells you where the company actually thinks its margin lives, and it is not in the cul-de-sac with three other broadband options.
The tell to watch for the rest of the year
If you want to track whether this thesis holds, do not watch the headline subscription price. SpaceX just raised U.S. residential plans by $5 to $10 a month, effective mid-June, and a number like that is noise; it tells you nothing about the structural question.
Watch three things instead.
Watch the congestion surcharges in dense cells. If they climb, or if new activations in suburban cells keep going “sold out,” that is the company quietly conceding that density is a cost center to be managed, not a market to be taken.
Watch the competitive vector, because it cuts the other way. Amazon’s roughly $11.6 billion Globalstar deal21 and its capture of Apple’s iPhone satellite connectivity, a contract reportedly offered to SpaceX first, plus Amazon Leo, formerly Project Kuiper, beginning to scale its low-Earth-orbit broadband constellation22, all compress Starlink’s pricing power independent of capacity. So a price cut this year would not necessarily signal confidence. It could signal that competition is arriving exactly as the perpetual capex bill gets larger.23 Falling prices and rising rebuild costs are a vise, not a victory.
Watch the upgrade treadmill on the ground. The full V3 capacity requires upgraded user terminals; the current dishes can’t exploit it. A network that has to periodically re-equip its own customer base to deliver its own next generation is, once again, behaving like a machine that must keep rebuilding — this time on both ends of the link.
The narrative that survives
Strip away the annuity fantasy, and a sharper, truer picture is left standing, and it is still a good business.
Starlink does not displace fiber. It does the thing fiber cannot: it reaches the places where trenching never pays.24 Rural homesteads. Ships. Aircraft. Disaster zones. Remote enterprise and government sites where there is no alternative, and price sensitivity is low. In that territory, Starlink is close to a monopoly, and the willingness to pay is structural. That is the durable franchise.
But it is precisely not a franchise over the dense, fiber-served, competitively-supplied suburb, the place the comfortable story imagined as the cash cow. Starlink is structurally barred from where fiber already is, and dominant exactly where fiber can’t go. The two technologies are not substitutes racing for the same customer. They are complements that have already divided the map.
The asset underneath the trillion-dollar number is a manufacturing company that happens to sell bandwidth, running a five-year fleet-renewal treadmill that gets more expensive the bigger it grows, defended by a real moat in the places where density is too low for anyone else to bother.
That is worth a great deal. It is just not worth what an annuity is worth. And the difference between those two things is the difference between the price and the asset.