Stocks for Thoughts · The SpaceX Trade

Two flights. Tens of thousands of satellites.

What you're actually buying at a trillion and a half — and the five ways it ends.

May's S‑1 was the first audited look inside SpaceX in twenty‑four years, and it told you more than any keynote ever has. Three businesses under one roof. One of them works.

Starlink — $11.4B of revenue, $4.4B of operating profit, the kind of margins a software company would frame and hang on the wall. Ten million subscribers by March, twelve by June, in more than 160 countries. It's a software business that happens to own nearly ten thousand spacecraft.

Launch — $4.1B of revenue and a $657M loss. Read that loss carefully: it's Starship swallowing development money, not the Falcon rockets, which quietly print cash. Launch is the business everyone thinks they're buying, and it's the smallest one on the page — the only piece a normal person would even recognise. Of 165 Falcon 9 flights last year, 42 carried someone else's payload. The rest were Starlink launching Starlink. That's not a market. It's the company moving money from one pocket to the other.

AI — xAI and X, folded in back in February. $3.2B of revenue against a $6.4B loss. And the trend runs the wrong way: it lost about 60 cents on the dollar in 2024, then two full dollars on the dollar in 2025, and early 2026 looks worse again. Normally, the more of something you sell, the better the economics get. Here, selling more makes them worse. Sit with that one for a second.

One segment funds the other two. That's the money.

Take Goldman's bull case — $474B of revenue in 2030 — and pull it apart. Starlink, $144B. AI, $322B. Rockets, $8.3B.

That's not a rocket company with an AI side‑project. It's an AI company with a launch subsidiary — and two‑thirds of the whole forecast is compute that doesn't exist yet.

Morgan Stanley pencils in $330B: a third lower, and leaning a little less on AI (about 58% of the mix, against Goldman's 68%). Musk, for his part, says he'd be surprised if revenue came in under a trillion by the early 2030s. Three numbers, one bet, just sized differently.

The loudest version of that bet is orbital, and it turns on a single piece of physics. So: half a page, plain English.

A data centre's real product is heat. On Earth you get rid of it the obvious way — blow air over the chips, or run water past them, so something physically touches the hot thing and carries the heat off. It's why the big campuses sit next to rivers, and why cooling eats 30 to 40% of a data centre's power bill.

In orbit there's nothing to touch. No air, no water. Heat can leave one way only: by glowing — radiating infrared into the dark — and glowing is slow.

Here's the part almost everyone has backwards. Space isn't cold in any way you can use. Space is empty, and empty is exactly what a thermos is made of. The same vacuum that sounds like a free refrigerator is the thing that makes a first‑class insulator.

And the arithmetic is unkind. A panel at room temperature sheds roughly 600 watts per square metre — about a thousand times slower than water doing the same job down here. One Nvidia H100 needs something like a square metre and a half of radiator. One megawatt — a facility a thousand times smaller than the ones going up in Memphis — needs around 1,600 square metres of it. A hockey rink. And those panels weigh at least as much as the chips they're cooling, probably a few times more, which matters a great deal when every last kilogram has to be thrown off the planet first.

The obvious objection: the satellite already has enormous solar wings, so just use the back of them. Doesn't work. A solar panel wants to stare straight at the sun; a radiator wants to face the other way, at cold empty sky. One object, two directions. Pick one.

None of this makes computing in orbit impossible. Lighter panels are real engineering, not fantasy. It just means the loudest branch of the whole bet quietly collapses into a weight problem: kilograms of radiator per kilogram of chip, and the cost of lifting them.

It also means four of the five futures below don't need any of it.

One ticker, five very different companies hiding inside it. Cheapest to dearest — and only the last one needs the thermos to work.

1. The Unwind. Cadence never gets past the test‑flight phase. The FAA has cleared Starbase for 25 launches a year; the compute plan, worked bottom‑up, needs something closer to 400 by 2030. The gap never closes, the panels stay heavy, launch stays expensive, and the curves never meet. Orbital data centres become a footnote, then stop getting mentioned at all. Meanwhile revenue per user keeps sliding — $99 a month in 2023, $66 by early 2026, before May's price rises — and Amazon's Leo finally shows up at scale. What's left is a very good broadband utility carrying an expensive habit. ~$30–40B revenue · $300–500B.

2. The Utility. Starship works, and does for the industry what Falcon 9 did: makes launch cheap, reliable, and boring. Starlink turns its head start into something durable, the May price rises stick, and it settles at thirty or forty million subscribers. Orbital compute survives, but small and real — earth‑observation, defence, the odd latency niche where the weight tax is worth paying. It never troubles a hyperscaler. Grok stays a mid‑tier lab. The market reprices SpaceX as infrastructure. Good multiples. Just not these ones. ~$50–65B revenue · $600–800B.

3. The Landlord. The most under‑discussed branch, and the only one already throwing off cash. SpaceX stops trying to win the model race and sells picks and shovels instead: buy the power, build the sheds, fill them with GPUs, rent to whoever's winning this week. It's doing exactly this now — Colossus 1 in Memphis, 300 megawatts and 220,000 GPUs, leased to Anthropic for about $1.25 billion a month, call it $15 billion a year, while xAI moved its own training next door. A direct competitor is the segment's biggest tenant, which is embarrassing and profitable in roughly equal measure. How long it lasts is its own question: the filing runs the lease to 2029, but Musk has called it a 180‑day deal you can walk away from on 90 days' notice. That gap is part of the risk.

Scale that to a few gigawatts and you've got a genuine arms‑dealer business that needs no new physics at all. The catch is that it's a commodity. Oracle, CoreWeave, Microsoft and Crusoe are all selling the same thing, the capital bill is brutal, and GPUs age out in about three years. Landlords get landlord multiples. ~$80–100B revenue · $900B–1.2T.

4. The Conglomerate. Fifty to eighty Starship flights a year by 2030 — historically astonishing, and still an order of magnitude short of the full‑constellation plan. Starlink becomes the default layer for everything fibre can't reach. Golden Dome, and whatever follows it, make SpaceX a defence prime in all but name. Orbital compute is real, but at gigawatt scale rather than hundred‑gigawatt scale — inference and specialist work, not replacing Memphis. All three engines run at once. A magnificent company that has to grow into its own price for half a decade. ~$90–120B revenue · $1.3–1.6T.

5. The Orbital Hyperscaler. The curves cross. Starship gets cheap and frequent at the same moment, and the panels get dramatically lighter — the optimists argue many times lighter than anything that's flown, and that one assumption carries the entire case. Past the crossing point, compute in orbit is permanently cheaper than compute on a grid: free sunlight, free cooling, and nobody whose permission you need. SpaceX becomes the landlord of the only place compute scales without asking planning permission. Goldman's $474B turns out to be the cautious read. The binding limits stop being physics and turn into chip supply, debris rules, the astronomers' lobby, and the awkward fact that a satellite you can't visit is a satellite you can't upgrade. ~$200–350B revenue · $3.5–5T.

Weight the four branches that need no new physics, and whatever's left over has to be carried by the fifth.

Be generous to the four — high‑end values, a heavy thumb on the Landlord — and the orbital branch only has to carry a modest slice. Be hard on them — low‑end values, a plain commodity multiple on the leasing — and it has to carry a good deal more. Either way, the honest answer is a range, and the width of that range is the whole point: the number swings by a factor of two on assumptions nobody can settle today. Anyone quoting it to two decimal places — this note included — is showing more confidence than the inputs can support.

What survives all that imprecision is the shape. Even being generous about the earthbound businesses, a real chunk of a trillion‑and‑a‑half‑dollar company is being paid for something that has never once operated. Not for Starlink, which is excellent. Not for launch, which is a monopoly. For the thermos.