The SpaceX IPO and Data Centers in Space | Stratechery by Ben Thompson

ELI5/TLDR

SpaceX wants to go public at a $2 trillion valuation while making $18.67 billion in revenue and losing $4.9 billion a year. By any normal math, that’s insane. Ben Thompson agrees the numbers are insane — but argues you’re not buying the numbers, you’re buying a dream, and Musk has a track record of turning his dreams into reality. The specific dream this time: putting AI data centers in space, where there’s unlimited room and no neighbors to object.

The Full Story

Musk sells dreams, not spreadsheets

Thompson opens with a deliberately petty grievance — getting an Uber Black and being assigned a Tesla Model Y, a “pretty basic car” that nonetheless made it onto the premium tier. His point is that the Model Y got there on brand alone. Tesla means something.

The real payoff of Musk’s master plan is the fact that Tesla means something.

That brand halo is the throughline for everything Musk does. The recurring pattern: his companies don’t win by playing the game better, they win by changing the rules through sheer scale.

Musk companies at their best don’t win the game. They change the rules through scale, such that billionaires buy economy cars because they actually drive themselves with supervision, and airlines transform the consumer experience on their own dime.

The airline example: Starlink (SpaceX’s consumer arm, $8.7B revenue and $4.4B profit last year) is being installed on 500+ American Airlines planes. American billed it as an “elevated onboard experience,” which Thompson finds funny — fast in-flight Wi-Fi was a differentiator when United signed up in 2024, but now it’s just table stakes. That’s the Musk move. Make the new thing so widespread that everyone has to have it, and you become the default supplier underneath.

The S-1 is a joke (on purpose)

Then there’s the filing itself. SpaceX claims a total addressable market of $28.5 trillion — “the largest actionable total addressable market in human history.” Of that, $26.5 trillion is supposedly AI, more than 13 times the entire space-and-connectivity opportunity combined. Thompson can barely keep a straight face:

In all seriousness, the numbers are obviously absurd, but then again, everything about this IPO is absurd.

The real financials underneath: a $2 trillion valuation on $18.67B revenue, $4.9B in losses, and growth that actually slowed from 35% to 33%. The slowdown came despite folding in xAI (Musk’s AI company) — which is what tipped SpaceX from a small profit into a big loss, via $5.1B in AI R&D. And that R&D bought a model currently sitting in fifth place whose founding team just walked out the door.

Why he doesn’t dismiss it anyway

Here’s the pivot. Thompson explains why he never bothered covering Tesla’s earnings, and won’t cover SpaceX’s either. With Musk, the stock isn’t a claim on a company’s cash flows — it’s a meme, a way to buy into the dream. Counterintuitively, every time Tesla issued new stock (diluting existing holders), the price went up:

Issuing more stock was not diluting existing shareholders. It was extending the opportunity to propagate the TSLA meme to that many more people… instead of infrastructure leading to a movement, a movement, via the stock market, funded the building out of infrastructure.

The capital markets become the engine. A dream attracts believers, believers fund the build, the build occasionally turns the dream real. Tesla’s valuation made no sense right up until the Model 3 and Y actually worked. So the question isn’t whether the numbers add up today (they don’t), it’s whether the dream is possible and whether there’s reason to think it might happen.

The actual bet: data centers in space

This is where Thompson goes from skeptic to genuinely interested. Three questions.

Is it possible? Yes — but only if you stop imagining Earth data centers shipped into orbit. Forget the giant buildings, cooling plants, and power grids. Think of one satellite as one server rack. The math is closer than you’d expect: Starlink’s largest current satellite is roughly the size of an Nvidia NVL72 rack already. The hard part is power and heat dissipation — a Starlink satellite handles maybe 25 kW, an Nvidia rack wants 135 kW — but in space you skip most of the cooling and power-distribution overhead, so ~100 kW might do the job. Racks would be disposable like Starlink satellites, interconnected with lasers, each with its own solar panels and radiators.

Is there a use case (the carrot)? Yes, and it’s specific. Thompson splits AI into three workloads: training, answer inference (a human waiting for a reply), and agentic inference (computers working for other computers). Agentic work has no human in the loop, so latency stops mattering:

Lower speed isn’t nearly as important a consideration if there isn’t a human in the loop. If an agent is waiting around for a job that is being run overnight, the agent doesn’t know or care about the user experience impact.

If delay is fine, you don’t need cutting-edge speed — you need cheap, plentiful capacity. And agentic inference is the workload that “scales not with humans, but with compute,” so it’ll be the biggest market by far. That’s exactly the kind of work you can afford to run on a slow rack in orbit.

Is there a stick? Yes, and it’s the most surprising part. The binding constraint on building AI on Earth isn’t just power — it’s zoning. Communities are blocking data centers. Thompson draws a contrast with globalization: when factories closed and jobs moved to China, the affected towns had no say. With data centers, building requires permission, so people suddenly do have a say — and they’re using it to say no.

people who didn’t have a say in globalization are suddenly finding they do have a say about AI and it’s not a surprise they’re expressing their disapproval by blocking data centers.

If demand for compute grows enough that there’s literally nowhere left to build, space stops being an alternative and becomes the only option. At that point the $2 trillion starts to look defensible.

An IPO worth supporting

Thompson is clear the bet rests on a tower of ifs — Starship has to work, the chip supply has to materialize, agentic inference has to unbundle today’s architectures, the zoning opponents have to win. He still calls the IPO “nuts.” But he’s glad it exists: Musk has a record of pushing humanity forward, Thompson genuinely worries AI could repeat the tragedy of nuclear power (where failure to build foreclosed a whole future), and he likes that this is a throwback to what an IPO is supposed to be — ordinary people putting capital in to actually build something and sharing the upside if it works.

It’s a testament to SpaceX’s ambitions that retail investors get to play VC. And hey, you get Mars upside for free.

Key Takeaways

SpaceX is seeking a $2 trillion valuation on $18.67B revenue, with $4.9B in losses last year and growth slowing from 35% to 33%.
The loss was driven by absorbing xAI: $5.1B in AI R&D that produced a model currently ranked fifth, whose founding team just departed.
Starlink alone did $8.7B revenue and $4.4B profit last year — though how SpaceX allocates launch costs to it isn’t fully clear.
Starlink has 10,000+ active satellites in low Earth orbit; it’s being installed on 500+ American Airlines narrow-body jets starting Q1 2027.
The S-1 claims a $28.5 trillion TAM — “the largest actionable TAM in human history” — of which $26.5T is AI (it excludes China and Russia from estimates).
The space data center concept reframes one satellite as one server rack, not a building. Starlink’s largest satellite is already close in size to an Nvidia NVL72 rack.
Power/heat is the constraint: ~25 kW today vs ~135 kW for an Nvidia rack, but skipping Earth-style cooling and power distribution could close the gap at ~100 kW.
Racks would be disposable, laser-interconnected, with their own solar panels and radiator arrays (~200+ sq meters of radiators per rack).
The killer workload is agentic inference — computers working for computers, no human waiting — where latency doesn’t matter, so cheap/slow/abundant compute wins. It scales with compute, not population, making it the largest future market.
The non-obvious Earth constraint is zoning: communities are blocking data centers. Thompson argues builders should literally pay people for permission to build.
xAI’s first data center (Colossus One) is monetized at $15B/year for 300 MW — roughly 3,000 racks’ worth. Thompson estimates Anthropic earns ~3x that revenue on equivalent capacity.
Thompson’s analytical move: with Musk, the stock is a meme/dream, not a cash-flow claim — issuing dilutive stock historically raised Tesla’s price by spreading the meme.

Claude’s Take

This is Thompson at his best — taking something that looks like obvious froth (a $2T valuation on a money-losing company with a clown-shoes TAM slide) and finding the one load-bearing idea worth taking seriously underneath it. He doesn’t soft-pedal the absurdity; he says “nuts” twice and laughs at the S-1. That earns him the right to then say “but here’s why the dream might be real,” and the data-centers-in-space argument is genuinely the most coherent version of that case I’ve seen. The satellite-as-rack reframe is the key unlock, and the agentic-inference-doesn’t-care-about-latency point is sharp.

Two things to keep your guard up on. First, the “Musk has a track record so give him the benefit of the doubt” move is doing a lot of work and is unfalsifiable by design — it’s the same logic that justified every Tesla valuation that “made no sense until it did,” which conveniently ignores the times it just didn’t. Second, the zoning-as-stick argument is clever but speculative; it requires demand so vast that Earth genuinely runs out of room, which is a much bigger “if” than Thompson’s breezy tone admits. He’s honest about the tower of assumptions, to his credit, but the framing still nudges you toward “this could work” more than the evidence strictly supports.

Score: 8. Tight, original, intellectually honest about the financials, and it teaches you a real mental model (the three inference workloads, the meme-stock-as-funding-mechanism) you can reuse. Docked a couple points because the central thesis leans on Musk’s reputation and a chain of low-probability ifs that the confident prose papers over.