SpaceX: Orbital Data Centers Are The Big IPO Narrative – And an Unsolved Physics Problem

With the S-1 filing submitted to the U.S. Securities and Exchange Commission (SEC), it has become clear what the central growth story of the upcoming SpaceX IPO will be: not Starlink, not even Starship. It is AI compute in space. Elon Musk’s company is openly positioning its public listing as a massive bet on the proposition that the next generation of AI data centers will not stand in Memphis, Mississippi or Virginia, but in a sun-synchronous orbit around Earth.

And what is also in the prospectus: These Orbital Data Centers (ODCs) “may not achieve commercial viability.” These future plans should therefore be taken with a great deal of caution — but Elon Musk certainly believes in them. ODCs are a “no-brainer,” he said. The IPO narrative is so big that SpaceX is targeting a valuation of 1.75 trillion dollars against revenue of just 18.674 billion dollars (with an operating loss of 2.589 billion dollars). That is a price-to-sales ratio of around 94x on 2025 revenue (more on that here).

The Addressable Market: 26.5 Trillion Dollars for AI

In the prospectus, SpaceX puts its total quantifiable addressable market (TAM) at 28.5 trillion U.S. dollars — and that excludes China and Russia. The breakdown shows strikingly where the company sees the music of the future being played:

• AI: 26.5 trillion U.S. dollars — of which 2.4 trillion in AI infrastructure, 760 billion in consumer subscriptions, 600 billion in digital advertising, and 22.7 trillion in enterprise applications
• Connectivity (Starlink & Co.): 1.6 trillion — of which 870 billion in Starlink Broadband and 740 billion in Starlink Mobile
• Space: 370 billion — meaning classic launch and space-enabled services

Put differently: SpaceX values the AI market at 71 times its classic space business and 16 times its connectivity business. Starlink, framed for years as a cash cow, shrinks in the prospectus narrative into a supporting service — a global data backbone designed to connect the orbital data centers with the world.

Concrete ambitions: SpaceX is aiming to build satellite constellations of up to one million AI compute satellites in sun-synchronous orbits. The annual rollout target: 100 gigawatts of compute power per year into orbit, carried by “thousands of launches per year” and the transport of “approximately one million metric tons of payload annually.” First deployments are scheduled “as early as 2028.” Longer term, the company even speaks of terawatt-scale capacity and a lunar base as the next energy hub for AI.

Why Compute in Space Is Supposed to Be Cheaper Than on Earth

The economic logic in the prospectus is surprisingly clear and rests on a cool cost-per-token calculation. SpaceX decomposes the cost per token into three components — the AI model itself, the compute hardware, and energy — and claims a structural advantage in the latter two.

Energy is the killer argument. The prospectus states that the Sun contains roughly 99.8% of the entire energy of the solar system — and that harvesting energy in space is “considerably more efficient than on land.” Space-based solar arrays can generate more than five times the energy per unit area of terrestrial solar, because they are continuously illuminated, suffer no atmospheric losses, and can be optimally oriented. Consequently, SpaceX argues that the “marginal energy costs” for orbital AI compute satellites will be “minimal,” because they are powered by solar arrays in space.

The terrestrial problem is real. The prospectus cites data showing that U.S. electricity generation was effectively flat between 2008 and 2023 (CAGR of 0.1%) and has since grown by less than 3% per year — while China’s electricity generation has expanded at roughly twice that rate over the same period. Demand for AI data center power in the U.S. clearly exceeds supply. The “175 GW Crisis” report by Introl, which SpaceX cites as an industry source, captures this in its very title.

Time-to-power as a competitive edge. With orbital compute, SpaceX claims it can “consistently activate the highest performing hardware before our competitors,” because it does not have to wait for grid interconnects. The industry benchmark for bringing a 100 MW greenfield data center online is around two years. SpaceX brought the first COLOSSUS II cluster online in 91 days — terrestrially. In orbit, it wants to compress that window even further.

Starship as the economic lever. The entire orbital compute business hinges on a fully reusable Starship. The prospectus states it bluntly: “AI compute satellites at scale need full Starship reusability to be economically compelling.” Without full reusability, there are no economically viable AI data centers in orbit. SpaceX is aiming to reduce the cost per kilogram to orbit by 99% versus the historical average. For comparison: according to NASA, Falcon 9 already cut the price from 18,500 to 2,700 USD per kilogram, Falcon Heavy to 1,400 USD. With Starship, the goal is to reach $100 per kilogram.

The Hurdles: Cooling, Radiation, Repairability

The prospectus — as befits an SEC document — is more honest about the risks than the typical Musk tweets. The cooling question is answered through the concept of radiative cooling: passive heat radiation into space. In the glossary, SpaceX explicitly defines the method as “a cooling method that dissipates heat by radiating it into space, often passively, and is expected to be used in orbital AI compute infrastructure.”

The unresolved problems are listed by SpaceX itself in the Risk Factors section:

• No one has done it before. Verbatim: “We have not, and no one else has, previously operated or attempted to operate orbital AI compute, and the conditions of space on such AI infrastructure have not been tested.”
• No repair, no upgrade. Once in orbit, the hardware is not accessible. Component failures mean “permanent capacity loss, accelerated depreciation, decommissioning or need for replacement.”
• Harsh environment. Vibration and acoustics at launch, extreme thermal cycles, cosmic and solar radiation, micrometeoroids and orbital debris, geomagnetic storms — all factors that “testing cannot fully replicate” on the ground.
• Shorter lifespan than the hardware. Satellites have a shorter useful life than the IT systems they carry. That forces regular replacement launches.

Technically, the waste-heat problem is huge: “To dissipate just 1 megawatt (MW) of heat while keeping electronics at a stable 20°C, an ODC (Orbital Data Center) would require a radiator surface of approximately 1,200 square meters — roughly the size of four tennis courts”, says this report.

The Real Bottlenecks: Chips, Launches, Spectrum

This is where the prospectus gets remarkably concrete. To achieve the goal of 100 GW of annual compute deployment in orbit, SpaceX assumes satellites carrying more than 100 kW of compute power per metric ton. From this it follows: thousands of launches per year and the transport of roughly one million metric tons of payload to orbit annually. For comparison: in 2025, SpaceX achieved a total of around 2,213 metric tons of mass to orbit — meaning the target is around 450 times higher.

The GPU bottleneck. It is made explicit in the Risk Factors section: SpaceX has no “long-term or other material contractual arrangements” with its direct chip suppliers and procures GPUs on a purchase-order basis. The direct chip suppliers depend on a “concentrated group of advanced semiconductor fabrication facilities.” The ability to scale orbital AI “depends on our ability to access a sufficient number of AI chips, significantly more than are currently available to us.”

The answer is called Terafab. In March 2026, SpaceX entered into a framework agreement with Tesla for Terafab, a chip-manufacturing program with a long-term goal of producing one terawatt of compute hardware per year. In April 2026, Intel joined as a third partner to contribute expertise in design, manufacturing, and packaging. But SpaceX itself admits: Terafab “may not be successful,” and concrete projects, timelines, or capex have not yet been defined. A very vague project.

The launch hurdle. Falcon 9 manages 23 tons to LEO, Falcon Heavy 64 tons. Starship V3 is designed for 100 tons, and future generations are supposed to double that. But the FAA currently does not even permit return-to-launch-site reentries for Starship, requiring a waiver. Twelve test flights have been conducted; the first payload-delivering flight is only scheduled for the second half of 2026.

Regulatory hurdles. A constellation with “potentially up to one million satellites” requires an entirely new level of spectrum authorizations, debris-mitigation approvals, and international coordination — “no assurance that such approvals will be obtained on acceptable timelines, terms, or at all,” SpaceX writes itself.

What This Means for the Narrative

The capex breakdown in the prospectus shows where the money is actually going: in the first quarter of 2026, SpaceX invested 7.7 billion dollars in the AI segment, compared with 1.3 billion in connectivity (Starlink) and 1.1 billion in space (primarily Starship). The AI segment is currently burning cash at a massive rate — a 2.5 billion quarterly loss against 818 million in revenue — but this is precisely where the investment thesis is anchored.

The irony for the industry: while OpenAI, Meta, Microsoft, and Google are fighting over gas turbines, nuclear power plants, and hyperscale power contracts on Earth, SpaceX simply wants to retreat into orbit. If the calculation works out, the SpaceX IPO will be read in hindsight as the moment when AI infrastructure left the Earth’s surface. If not, it will have been the most expensive first-principles bet of the AI era to date. The S-1 leaves no doubt about which path will be taken — and makes no secret of the fact that the entire vision hangs on one single point: the fully reusable Starship.