SpaceX has filed to go public, and the document is not quite a SpaceX S-1. It is a filing for SpaceX merged with xAI, a merger that took effect on February 2, 2026, and xAI had earlier absorbed X. The stock on offer therefore bundles reusable rockets, the Starlink satellite-internet business, the Grok model, the X platform, the COLOSSUS data centers, and a plan to put data centers in orbit. A single share of the combined company is therefore a claim on rockets, satellite internet, a chatbot, a social network, and the eventual settlement of Mars, which is an unusual amount of unrelated risk to hold in one position.
The most science-fiction item in the filing is the orbital one, and it is no longer vague. SpaceX says it expects to begin deploying orbital AI compute satellites as early as 2028. The stated goal is 100 gigawatts of compute in orbit, carried by what the filing describes as potentially millions of satellites, requiring thousands of launches a year and the transport of roughly one million metric tons to orbit annually. For about two years the idea of moving data centers into space has lived on conference slides and in reporting about who is talking to whom. An S-1 is a different kind of document. It is the place where a vision becomes a disclosure obligation.
The more grounded question the filing answers is who pays for all of this, and the answer is interesting. The AI segment, which SpaceX acquired in the xAI merger, loses a lot of money. In 2025 it generated about $3.2 billion in revenue against an operating loss of roughly $6.4 billion. Its capital expenditures were about $12.7 billion in 2025, and about $7.7 billion in the first quarter of 2026 alone, which is roughly three times the combined capital spending of the Space and Connectivity segments over the same quarter. The AI business is, financially, a furnace.
The furnace runs on Starlink. The Connectivity segment generated about $11.4 billion of revenue in 2025 and roughly $7.2 billion of segment adjusted EBITDA. It is a subscription internet business, the most boring thing in the building, and it is profitable and growing. The S-1 describes a company in which a steady utility-style cash machine underwrites a frontier-AI bet and an orbital infrastructure program that will not generate revenue for years. This is a reasonable way to run a holding company. It is also the thing public investors are being asked to buy: the right to fund the orbital compute plan with the proceeds of a satellite-internet subscription.
On the ground, the compute already exists in scale. The COLOSSUS and COLOSSUS II clusters in Memphis and Southaven together supply about one gigawatt of compute power. The filing also describes Terafab, a chip-manufacturing initiative with Tesla and Intel that aims to produce one terawatt of compute hardware a year, the company hedging against the same AI-chip supply constraints that everyone else is hedging against. The orbital plan sits on top of all of this as the part that has not been built yet.
None of this is unreasonable, exactly. Vertical integration from launch to chips to model is a coherent strategy, and a profitable subscription business is a sensible thing to fund a moonshot with. But an S-1 changes the status of a promise. Once “millions of satellites by 2028” is in a filing, it stops being an aspiration and becomes a number that has to come roughly true, audited against, on a schedule. SpaceX has spent two decades making improbable timelines real. It has now written one down where the lawyers can see it.
–
By the Control Plane Editorial Team