SpaceX's IPO and the AI-Infrastructure Valuation Thesis

On June 12, 2026, SpaceX went public under the ticker SPCX and did the one thing a mega-IPO is supposed to do and almost never does cleanly: it held. The stock priced at $135 a share — a roughly $75 billion raise, the largest IPO in history — and closed its first session up about 19% at $160.95, pushing the company’s first-day valuation above $2 trillion. No broken syndicate, no 40% moonshot-then-collapse, no halt-on-volatility circus. An orderly print at a colossal size.

Within hours, that orderliness was being repackaged as a thesis. The reasoning goes: if public markets will absorb $75 billion of capital-intensive, deeply unprofitable “infrastructure” at a two-trillion-dollar valuation without choking, then the bid for picks-and-shovels infrastructure — launch capacity, satellite networks, and by extension AI datacenters and compute — is structurally strong. That is the framing now being attached to Anthropic’s reported Q4 2026 listing, where bankers expect a raise north of $60 billion, and to OpenAI’s parallel planning.

We build and operate this kind of infrastructure for clients across four time zones, so the question is not abstract for us: does one clean debut actually validate the AI-infra valuation thesis, or is it being asked to carry more weight than a single data point can bear? Walk the numbers first.

What the print actually was#

Strip the headline and read the mechanics, because the mechanics are most of why the debut went smoothly.

• It priced into forced demand. SPCX wasn’t trading purely on fundamentals on day one. Index methodology had already pre-committed buyers. MSCI added SPCX to its World and ACWI indices effective June 13 — the first business day after the debut — with its standard and large-cap indices following on June 29. Nasdaq revised its rules in May to let a top-40-by-market-cap newcomer enter the Nasdaq-100 after just 15 trading days, which puts fast entry around early July. BNP Paribas estimated Nasdaq-100 inclusion alone could drive roughly $8 billion of passive buying in the first month, with total index-driven inflows potentially reaching $30 billion. When index funds are mechanically obligated to buy, “the market embraced it” is a weaker signal than it looks.
• The product wrapper was ready. A wave of SPCX-linked ETFs and leveraged products was queued before the bell, several structured for leverage. That manufactures incremental, often retail, demand on top of the float — and amplifies moves in both directions later.
• Retail showed up hard. Reported retail orders ran into the tens of billions, which helps a first-day tape and tells you very little about where the stock clears once the lockups and the leverage unwind.

None of that makes the debut fake. It makes it engineered — and an engineered float clearing cleanly is a statement about plumbing and index rules as much as about conviction in the underlying economics.

Why SpaceX could carry it: Starlink’s cash flows#

Here is the part the AI-infra crowd should sit with, because it cuts against their own analogy. SpaceX did not IPO on a promise. It IPO’d on a profitable, fast-growing cash engine bolted to a pile of losses.

The S-1 disclosed an accumulated deficit of about $41.3 billion since 2002, and a 2025 net loss near $4.94 billion — the kind of number that historically scares a listing. What made the timing work is Starlink. In 2025, Starlink generated roughly $11.4 billion in revenue, up about 50% year over year, at a ~63% EBITDA margin — around $7 billion of segment EBITDA and $4.4 billion of operating income, serving 10.3 million subscribers across 164 countries.

That is the load-bearing fact. Starlink is a real, scaled, high-margin recurring-revenue business that happens to subsidize the unprofitable, capital-hungry bets around it — including xAI, which consolidated into SpaceX and posted a ~$6.35 billion operating loss in 2025. Investors paid two trillion dollars for a story in which a proven cash machine funds the moonshots, not for the moonshots alone. The market was, in effect, underwriting demonstrated unit economics with optionality attached — not optionality with a hope of unit economics.

Where the analogy to AI infrastructure breaks#

This is where the extrapolation gets dangerous. “Capital-intensive infrastructure listed cleanly, therefore AI infrastructure will too” treats two very different cash-flow profiles as the same asset class. They are not.

Starlink’s revenue is proven and recurring; frontier-model AI revenue is neither, at the unit-economics level. A Starlink subscriber pays a predictable monthly fee for a service with a defensible cost-to-serve and a long replacement cycle. The closest AI-lab analog — API and subscription revenue — is growing fast but is sold into a market where the unit cost of inference is still falling by an order of magnitude every year or two, where competitors ship near-equivalent models within months, and where the largest customers are themselves trying to insource the capability. Falling cost-per-token is wonderful for adoption and brutal for anyone modeling stable, Starlink-like margins on today’s compute.

The capex doesn’t depreciate the same way. Launch vehicles and satellites are expensive, but a Falcon booster amortizes across dozens of flights and a Starlink satellite earns for years. An AI datacenter’s economics are dominated by GPUs that are arguably obsolete in three to four years and by power and cooling contracts that run for fifteen. You are pairing a fast-decaying compute asset with a slow, fixed infrastructure obligation. That mismatch is the central financial-engineering risk of the AI-infra buildout, and a launch company’s balance sheet tells you nothing about how it resolves.

The moat is different in kind. SpaceX’s advantage is physical and cumulative: reusable hardware, a launch cadence rivals can’t match, regulatory spectrum, and orbital real estate. A model lab’s advantage is talent, data, and a lead measured in months that has to be re-won every release. Both are real moats. They do not deserve the same multiple, and an investor pattern-matching “infrastructure” across them is buying a category label, not a comparable business.

So when the orderly SPCX tape gets cited as cover for an AI-lab raise, notice the substitution: the thing that made SpaceX financeable — Starlink’s boring, recurring, 63%-margin cash flow — is precisely the thing the AI labs are still trying to prove they have.

The “picks and shovels” framing is doing a lot of work#

There’s a second sleight of hand worth naming, because it shapes how operators get sold infrastructure deals. “Picks and shovels” is the comfortable trade: don’t bet on which prospector strikes gold, sell the tools to all of them. Applied to AI, it usually means: don’t bet on a model lab, own the datacenters, the power, the networking, the GPUs. The pitch is that the arms dealer wins regardless of who wins the war.

That logic survives contact with reality only if two things hold. First, the tool has to be scarce and slow to replicate — a railroad, a port, a launch pad. Second, demand has to be durable enough to amortize a long-lived, fixed asset. SpaceX’s launch business genuinely fits: reusable orbital-class lift is scarce, and the asset earns across years of flights. An AI datacenter half-fits at best. The power interconnect and the building are scarce and slow — that part of the analogy holds, and it’s the real reason hyperscalers are racing to lock grid capacity. But the part of the capital stack that actually costs the most, the accelerators, is neither scarce in the long run nor slow to replicate, and it depreciates faster than almost any infrastructure asset class in history.

The practical consequence for anyone building this out: the financeable, “pick and shovel” layer is the boring, slow, physical one — land, power purchase agreements, substations, cooling, fiber. The layer that looks most like AI and gets the richest multiple — the compute itself — behaves more like inventory than infrastructure. Conflating the two is how you end up underwriting a fifteen-year obligation against a three-year asset and calling it a railroad. SpaceX’s debut, read carefully, is a reminder of which layer the market actually pays a premium to own: the one with proven, recurring cash behind it.

What an Anthropic or OpenAI listing actually inherits#

The labs do inherit something real from this print, just not what the headlines claim.

They inherit a proven distribution mechanism for size. SPCX showed that index rules, ETF wrappers, and retail demand can clear a nine-figure float in the tens of billions without a disorderly tape. For an Anthropic raise targeting $60 billion-plus — the company most recently raised $65 billion at a ~$965 billion valuation — that plumbing matters. The pipes can move the capital.

What they do not inherit is the underwriting logic. SpaceX cleared because a profitable core de-risked the losses. An AI lab going out at a near-trillion-dollar valuation while still burning to train the next model is asking the market to underwrite the promise of future leverage, on compute assets whose cost curve is actively working against margin stability. That can absolutely work in a strong tape — money was pricing OpenAI near $850 billion in March — but it is a different trade, and the SpaceX comp flatters it more than it justifies it.

There’s also a tape-condition caveat worth stating plainly: an orderly debut in June tells you about appetite in June. Index-driven and leveraged demand cuts both ways. The same ETF and passive machinery that smoothed SPCX’s entry can accelerate the exit if sentiment turns before the labs print. One clean window is not a standing offer.

The takeaway#

SpaceX’s listing is a genuine data point, and a useful one: public markets, with the right index and product plumbing, can absorb capital-intensive infrastructure at historic scale without breaking. If you are building or financing AI datacenters, launch capacity, or compute networks, that is encouraging about access to capital — the rails exist and they held. But it is a data point about plumbing and appetite, not a proof about economics. SpaceX was financeable because Starlink throws off real, recurring, high-margin cash that pays for the moonshots; the AI labs leaning on this print for their own raises are still trying to demonstrate the equivalent, against a compute cost curve and a depreciation profile that make Starlink-style margins far from guaranteed. Read the debut as evidence the market wants to fund infrastructure, not as evidence it has correctly priced the AI kind. One orderly IPO validates the channel. It does not validate the thesis — and the gap between those two is exactly where the next twelve months of AI-infra valuations will be won or lost.