What SpaceX's S-1 Actually Says About Putting AI in Orbit

What SpaceX's S-1 Actually Says About Putting AI in Orbit | Omniscient Media

The registration statement Space Exploration Technologies Corp. filed with the Securities and Exchange Commission on June 3 defines a new line of business it calls "orbital AI compute" and describes it as satellite constellations that "act as orbital data centers, harnessing solar energy for power and leveraging the space environment for cooling," with a target of beginning deployment "as early as 2028."^[1] The proposal is a bet that the two constraints strangling terrestrial AI infrastructure - power and heat - are easier to solve two hundred miles up than on the ground. A frontier training cluster draws as much electricity as a small city, and most of that electricity ends up as waste heat that has to be pumped out of the building with water and chilled air. In orbit, there is no grid-interconnection queue and no air to trap the heat.

The company went public nine days later. It priced the largest initial public offering on record, and within two trading sessions its market value crossed two trillion dollars.^[4] The valuation is not the part of this story that will still matter in five years. The part that will matter is that a single founder has, through a sequence of mergers between companies he already controlled, assembled the launch layer, the compute layer, and the distribution layer of a proposed off-world AI system inside one publicly traded corporation, and the central engineering claim that justifies the whole structure cannot be tested until 2028 at the earliest.

The physics the filing is responding to

The premise of orbital compute is a bet on two resources that are scarce on the ground and abundant in orbit. The first is power: a satellite in the right orbit can sit in near-continuous sunlight, drawing on solar energy without the day-night cycle, weather, or grid-interconnection queue that constrains a terrestrial solar farm. The second is cooling. There is no air in orbit to carry heat away by convection, but there is an effectively unlimited cold sink, so a spacecraft sheds energy by radiating it into the dark. The filing names both directly: solar energy for power, the space environment for cooling.^[1]

The scale the company is describing is what makes the idea more than a thought experiment. Its stated goal, in the filing's own words, is "to launch 100 gigawatts of compute to space each year."^[1] The filing states that the generation resources needed to support 100 gigawatts of compute, if operated continuously, "could generate approximately one-fifth of the annual power production in the United States," which it puts at 4.4 thousand terawatt hours in 2025.^[1] A company is telling public investors it intends to move a fifth of a national power budget's worth of computing demand into orbit each year, because it has concluded the ground cannot supply it fast enough.

The stack assembled under one roof

The proposal only coheres because the same company controls every layer it depends on. Getting 100 gigawatts of hardware to orbit each year is a launch problem, and the company builds and flies its own rockets. Running a data center in space and selling its output is a compute and connectivity problem, and the company now owns both. The filing organizes the business into "three reportable segments in our vertically integrated innovation engine: Space, Connectivity, and AI."^[1]

The AI segment did not grow inside the rocket company. It was acquired. The filing states plainly that the "AI segment refers to our AI business, which we acquired in connection with our acquisition of xAI in February 2026, and includes our AI compute, Grok, and X."^[1] The mechanics matter. xAI was absorbed in a transaction the filing calls the "xAI Merger," effective February 2, 2026, and that deal also folded in X Holdings Corp., the entity behind the X platform, which xAI had itself acquired effective March 28, 2025.^[1] The filing accounts for both as transfers "between entities under common control," the treatment used when the same party owns both sides of a deal.^[1] One person owned the rocket company, the AI lab, and the social network, and consolidated them onto a single balance sheet under the rocket company's name. When the xAI deal was announced, it was reported as a step toward building data centers in space, the ambition the registration statement would later formalize.^[3]

The result is a closed loop. The Space segment supplies the launch capacity to deploy the satellites. The Connectivity segment, built on the Starlink network, supplies the global links that would move data to and from an orbiting cluster. The AI segment supplies the workload, the Grok models and the compute they run on, that the orbital data centers would exist to serve. The vertical integration is the entire argument, because no outside vendor sells "100 gigawatts of compute to space each year" as a service.

The link is the hard part nobody talks about

Power and cooling get the attention, but a data center is only useful if data can reach it and leave it. A cluster on the ground sits next to fiber. A cluster in orbit does not, which turns the connection into its own infrastructure problem: the link has to carry training and inference traffic between satellites and the surface at high bandwidth and low delay to justify the trip off-planet. This is the gap the Connectivity segment is positioned to fill. The Starlink network already moves data between space and the ground at global scale, and the same filing lists "our AI-related offerings" alongside Starlink connectivity services as businesses whose growth the company is counting on.^[1] The backhaul for an orbital data center is, in this design, a network the company already runs for ordinary internet customers.

What actually pays for it

The orbital business does not yet earn money, and neither does most of what the company does. On a consolidated basis in 2025, the company reported revenue of $18,674 million, a loss from operations of $(2,589) million, and Adjusted EBITDA of $6,584 million.^[1] The profit, and most of the revenue, comes from one place. The Connectivity segment, the filing notes, was "primarily driven by Starlink," and generated revenue of $11,387 million, income from operations of $4,423 million, and Segment Adjusted EBITDA of $7,168 million in 2025.^[1] That single segment is roughly 61 percent of total sales and the only one running a profit.

The two segments that carry the company's identity both lose money. The Space business generated revenue of $4,086 million at a loss from operations of $(657) million, and the AI business reported revenue of $3,201 million at the largest operating loss of the three, $(6,355) million.^[1] The financial picture inverts the marketing. The thing that pays the bills is ordinary satellite broadband: a subscription internet service sold to consumers, enterprises, and governments. The rockets and the AI are funded by the modem in a customer's backyard. A reader who knew only the segment economics, and not the name, might reasonably classify this as a telecommunications company that also does research. The regulatory classification points the same way. On EDGAR, the company is listed under Standard Industrial Classification code 7370, services-computer programming and data processing, not under any aerospace category.^[2]

Segment	Revenue ($M)	Operating Income ($M)	Adj. EBITDA ($M)
			SpaceX 2025 Segment Financials
Space	4,086	(657)	653
Connectivity (Starlink)	11,387	4,423	7,168
AI (xAI, Grok, X)	3,201	(6,355)	(1,237)

Source: SpaceX Form S-1/A (Amendment No. 2), filed June 3, 2026. All figures in USD millions. FY 2025.

The claim that cannot be checked

The orbital compute concept rests on engineering that has not been demonstrated at anything close to the proposed scale. Radiating 100 gigawatts of waste heat from spacecraft, launching hardware at the required cadence, and connecting it reliably are each unproven at the size the filing describes, and the document does not pretend otherwise. In its discussion of risk, the company writes that "the timeline for certain of our initiatives involving unproven or new innovations, including our goal of deploying 100 gigawatts of annual compute power to orbit," along with related ambitions and "the launch cadence required to achieve these goals," may be "difficult or impossible to determine."^[1]

That is the company's own language, in a legally binding document, about the premise an audience is being asked to value. The earliest deployment date it offers is "as early as 2028," and even that is presented as a target rather than a schedule.^[1] Until a satellite is dissipating real computing heat in real sunlight and returning useful work over a real link, the orbital claim remains unproven. Public markets do not usually get to price an idea this unsettled, because the companies pursuing them are usually still private. This one is no longer private.

What the market did with it

The reception belongs in the record as context. The offering priced at $135.00 per share and raised approximately $75 billion, which the financial press described as the biggest initial public offering in history, at an implied valuation near $1.77 trillion.^[4]^[5] On the first day of trading, Friday, June 12, the stock opened at $150, reached an intraday high of $176.52, and closed at $160.95, up about 19 percent from the offering price.^[4]^[5] On Monday, June 15, the first full session, shares climbed roughly 20 percent on volume of about 244 million shares, lifting the company's market capitalization above $2 trillion. The same move made the founder, on paper, the world's first trillionaire, and established the company as the sixth-largest public corporation in the United States by that measure.^[6]

A two-trillion-dollar valuation now sits on top of a business whose only profit center is satellite broadband and whose signature ambition the filing itself says may be impossible to schedule. The market is pricing the orbital stack as though it works. The filing is careful to say no one knows yet whether it will, or when. The broadband business will keep paying regardless; what remains genuinely open is whether the off-world compute layer, the reason the rest was assembled, ever leaves the page.

AI Infrastructure

What SpaceX's S-1 Actually Says About Putting AI in Orbit

Discussion

Industry

OpenAI's Revenue Miss Puts SoftBank's Borrowed Bet on a Tighter Clock