SpaceX Filed for a $1.75 Trillion IPO Claiming It's an AI Company. The S-1 Tells You More About AI Economics Than Any Fund Pitch Deck.

SpaceX's roadshow started June 3. The company targeting a $1.75 trillion valuation positioned itself in its S-1 amendment — filed June 1 — not primarily as a rocket company but as an AI infrastructure giant. One line item in that filing tells you more about the structural economics of AI than eighteen months of VC conference panels.

The line: a Cloud Services Agreement with Anthropic, signed May 2026, under which Anthropic will pay SpaceX $1.25 billion per month through May 2029 for access to Colossus 1 — the Memphis data center hosting more than 220,000 Nvidia processors drawing 300 megawatts of power. Fifteen billion dollars per year. Paid from an AI safety company to a rocket company that also owns, through its founder, the most valuable private AI lab in the world. Anthropic, valued at $965 billion in its most recent round and running $47 billion in annualized revenue, cannot maintain its model training capacity without renting infrastructure from a direct competitor. That's not a technology problem. It's a capital concentration problem — and the SpaceX S-1 is the first document to price it publicly.

Morningstar published a note June 3 calling the $1.75 trillion targeted valuation nearly twice fair value. The critique is specific: Starlink growth is decelerating; orbital AI compute satellites cited in the filing won't deploy until 2028 at the earliest; the xAI Colossus revenue is a contract, not a moat. The Anthropic agreement allows either party to exit with 90 days' notice. A $1.75 trillion valuation partially supported by a 90-day-cancellable contract from a competitor is a pricing argument worth holding through the roadshow.

But the Morningstar valuation critique misses the structurally more important observation. The Colossus deal reveals, in public filings with SEC-level precision, what AI training infrastructure actually costs and who controls it. Three entities control the training capacity that frontier AI companies actually depend on: Amazon Web Services, Google Cloud, and xAI/SpaceX. Microsoft Azure is closing the gap through its Stargate commitments. No one else is in the conversation at training scale. The S-1 didn't create this concentration — but it quantified it for the first time, at $1.25 billion per month for a single tenant at a single cluster.

The investment implication for the application layer is not subtle. When training compute is this concentrated, the barrier to entry for new foundation model labs is not talent or architecture — it is the capital required to build or rent infrastructure that only a handful of entities can supply. Building a frontier model lab without a hyperscaler relationship or a $15 billion annual compute budget is not a plan. It is a theory. The application layer, which depends on inference rather than training, purchases inference capacity from any provider and switches on price and quality. That structural difference — between foundation model economics and application model economics — is now a public record. Fifteen billion dollars per year versus an API call.

The SpaceX valuation question will be answered when the stock trades. The more durable question the S-1 poses is about the dependency structure of the entire AI stack — and whether trillion-dollar AI valuations are being built on infrastructure whose ownership is narrow, whose pricing is now visible, and whose exit provisions are measured in weeks rather than years.

What the filing answers for AI investors isn't whether SpaceX is worth $1.75 trillion. It's that the infrastructure layer of AI is priced, concentrated, and structurally expensive to escape — and that the companies whose economic model doesn't require escaping it are better positioned than the ones still trying to build their way out.