Everyone's Got SpaceX's IPO Wrong: What the Market Is Missing

SpaceX filed for what looks set to be the biggest IPO in history, and the headlines are already locked on one number: nearly $5 billion in losses last year. That number is real. It is also the wrong place to start.

Buried in the same S-1 is a single customer deal that flips the story. Anthropic, one of the most important AI labs on Earth, has agreed to pay SpaceX about $1.255 billion every month through May 2029. That is roughly $15 billion a year from one tenant renting compute. Either side can cancel with 90 days notice, so it is not a forever lock. With AI demand this hot, canceling looks unlikely.

Once you see that check, the loss stops looking like a company that got ahead of itself. It starts looking like a company spending hard to build infrastructure other people will rent for a decade, and already lining up the first big payer.

What the S-1 actually describes

On May 20, 2026, SpaceX filed its S-1. The raise talk sits in a range that could be about three times Saudi Aramco's old $25 billion record, at a valuation near $2.5 trillion. The document does not describe a pure rocket shop. It describes three businesses under one roof.

First, rockets: Falcon 9 and Starship flying for NASA, the Department of War, and private customers. Second, connectivity: Starlink. Third, AI: the xAI stack SpaceX absorbed earlier this year, including Grok, X, and the Colossus data centers.

That third bucket is what Anthropic is writing the giant check for. On February 2, 2026, SpaceX closed an all-stock deal for xAI. Overnight, the rocket company owned a frontier AI arm and more than half a million GPUs across Colossus 1 and Colossus 2. When you ask Grok or ChatGPT a question, the answer does not come from your phone. It comes from a building packed with expensive chips and a lot of electricity. The companies that own those buildings and that power control the real scarce layer. Everyone else rents.

SpaceX is betting it becomes the picks-and-shovels company of the AI rush. Not mainly the model brand. The compute, the power, and eventually the orbital machines other labs pay to use.

The Anthropic proof point

Anthropic builds Claude. It competes with Grok and ChatGPT at the frontier, and on a lot of benchmarks its models, especially with Mythos in the mix, sit at the top of the pack. In May 2026 the filing shows SpaceX signed to give Anthropic access to that computing power on the terms above: about $15 billion a year.

Elon put the thesis in public language after the deal: SpaceX is offering AI compute as a service at significant scale, talking to other companies, and over time, especially with orbital data centers, expects to serve AI at extremely high scale. Anthropic has the money and talent to build its own campuses. It still chose to pay SpaceX instead. That is the proof.

There is a second, quieter deal in the same paperwork. In April 2026 SpaceX signed a compute deal with AnySphere, the company behind Cursor, and took an option to buy Cursor at an implied $60 billion valuation. Reporting suggests that option gets exercised after the IPO. Rockets are still the brand. The filing is already telling a different story.

Three segments, one playbook

The scary loss makes more sense when you split the company the way the filing does.

The space segment brought in roughly $5 billion in 2025 and still posted an operating loss of about $657 million. Starship ate the margin. SpaceX poured about $3 billion of R&D into Starship in 2025 alone. The filing says paying payloads should start in the second half of 2026. Until then Falcon 9 makes money and Starship spends it.

Connectivity is the engine. Starlink brought in about $11 billion of SpaceX's $18 billion 2025 revenue, roughly 61% of the top line, up almost 50% year over year. Operating income topped $4 billion. Adjusted EBITDA cleared $7 billion on about $11.4 billion of revenue, a software-like margin on a hardware network. By the end of March 2026 Starlink had roughly 10 million subscribers across 164 countries and more than 10,000 satellites, plus direct-to-phone service already touching something like 7 million devices a month.

The AI segment is where the headlines get their fear. In 2025 it brought in about $3 billion and lost more than $6 billion from operations. The moment the xAI deal closed, those losses landed on SpaceX's books. That is also the segment Anthropic just agreed to flood with $15 billion a year. Recognize that revenue against AI, and the segment that looks broken can turn profitable fast enough to cover rocket losses and put the whole company in the black as early as 2026.

This is a pattern you should learn. Take a dominant cash cow. Use its cash to fund the next era. Amazon did it with AWS. Tesla does it with cars funding autonomy and Optimus. Google does it with search ads. SpaceX is running the loudest version: Starlink's cash feeding the AI moonshot, and now a frontier lab writing rent checks into that same moonshot.

Orbit is the far bet

Everything above is still mostly Earth: Memphis and Mississippi data centers, Starlink beaming internet down. The filing also says SpaceX expects to begin deploying orbital AI compute satellites as early as 2028.

Why put a computer in space? Power, first. In the right sun-synchronous orbit, solar panels can stay in sunlight almost constantly. On the ground, AI already hits grid walls, NIMBY fights, and permit queues. In orbit those problems shrink. Heat and room come second. Most of a terrestrial data center's cost and footprint goes to cooling and site constraints. In orbit you radiate heat into cold space and expand without buying another county.

It sounds wild until you remember Starship exists to make that mass cheap enough to try.

The risks the filing has to say out loud

Key person risk is the big one. The document says SpaceX is highly dependent on Elon Musk, and he does not work there full time. Cash burn is next. Cash fell from about $24 billion at the end of December 2025 to about $15 billion by the end of March, nearly $9 billion gone in a quarter. The Anthropic deal is the obvious brake, and more compute tenants would help more. Governance is third. Dual-class stock keeps Class B at 10 votes. Before the IPO Elon controlled about 85% of voting power. After the float he still runs a controlled company that can skip some Nasdaq independence rules. He has said that control protects the Mars mission. Investors who hate that structure will stay away. Investors who trust him will not.

Shares are set to list on Nasdaq on June 12. For people who have followed Tesla and Elon's companies for years, the filing forces a conviction question more than a spreadsheet question. The loss number will keep making noise. The customer rent is the signal. SpaceX is asking the market to price a rocket company, an internet company, and an AI landlord as one machine, already collecting checks from the lab many people think has the best model on the market.

Check the video here.