SpaceX filed its long-awaited S-1 registration statement with the SEC on May 20, 2026, formally launching what it hopes will be the largest initial public offering in history - targeting a $1.75 trillion Nasdaq debut under ticker SPCX in late June. The 277-page document disclosed, for the first time in the company's 24-year history, the full financial picture: $18.67 billion in 2025 revenue, a $4.94 billion net loss, and an AI segment that burned $6.4 billion from operations on just $3.2 billion in revenue. The story carrying the offering across the line - and the gap between what that story promises and what the physics permits - is the subject of this article.
In the span of roughly ninety days between late October 2025 and late January 2026, a concept that had spent decades as a fringe engineering curiosity was transformed into the defining narrative of the most expensive stock offering ever attempted. SpaceX filed with the US Federal Communications Commission for authority to operate up to one million satellites as orbital data centers. Elon Musk announced an all-stock merger with his artificial intelligence company xAI at a combined valuation of $1.25 trillion. The SpaceX board approved a compensation package granting Musk 60.4 million additional Class B shares contingent on the company deploying orbital data centers generating 100 terawatts of computing capacity - the equivalent of approximately 100,000 large nuclear reactors - and reaching a $6.6 trillion market capitalisation. The company's implied private valuation roughly doubled in six months, from approximately $400 billion to the $1.75 trillion target set for the public offering. The S-1, filed May 20, confirmed the scale of what is being asked: investors are being invited to price a company generating $18.67 billion in annual revenue at 94 times trailing sales, with the justification residing almost entirely in the orbital data centre thesis.
The timing of each announcement tracks the IPO roadmap with a precision that is either coincidental or deliberate. The physics, economics, debris environment, and thermal engineering of space-based computing make a strong case for the latter.
SpaceX's FCC application, submitted January 30 on a Friday under file number SAT-LOA-20260108-00016, requested authority for satellites operating between 500 and 2,000 kilometres altitude. Its most-quoted line declared the system would be "the first step towards becoming a Kardashev II-level civilisation - one that can harness the Sun's full power." Musk stated publicly, including at Davos in January, that "the lowest cost to generate AI compute will be in space" and that this would be true "within two to three years, three at the latest."
What the filing contained no reference to was a deployment schedule, a satellite design, a thermal management architecture, a cost estimate, a launch cadence, a power budget, or a revenue model. It requested a waiver of the FCC's standard milestone rules that require 50 percent deployment within six years. The rocket on which every cost projection in the orbital data center thesis depends - Starship - remains a test vehicle in active development. As of January 2026 it had completed eleven integrated test flights and had not yet been certified or configured for commercial payload delivery to orbit. It has never carried a paying customer.
Amazon's Project Kuiper team filed a formal petition with the FCC on March 6 asking that the application be denied outright, describing it as "a lofty ambition rather than a real plan." The filing triggered what might be called a regulatory gold rush. Starcloud filed for 88,000 satellites the day after SpaceX. Blue Origin filed for 51,600 satellites under its "Project Sunrise" banner on March 19. Together with China's filings covering roughly 200,000 satellites, more than 1.34 million orbital data center spacecraft were formally proposed in a ninety-day window. The total number of operational satellites in all of human history, across all nations and purposes, is approximately 11,000.
Three physical constraints, each individually severe, combine to place the near-term commercial version of this concept well beyond any solution currently available to engineers. None has been resolved at the advertised scale. None is being discussed honestly in the promotional materials circulating ahead of the IPO.
The first is radiation. Cosmic rays, trapped protons in the South Atlantic Anomaly, and solar particle events deliver ionising doses of roughly 1 kilorad per year in shielded low Earth orbit, rising to 50 kilorad per year at geostationary altitude. The standard engineering solution - radiation-hardened processors such as the BAE Systems RAD750 - tolerates total doses of 200 to 1,000 kilorad but runs at 200 megahertz on a 150-nanometre fabrication process. A 2026 NVIDIA H100 runs on a 4-nanometre process and delivers orders of magnitude more throughput per watt. There is no radiation-hardened H100. There is no radiation-hardened Blackwell. There is no plausible engineering path to one within this decade. Google's Project Suncatcher tested a Trillium TPU against 67 megaelectronvolt protons in 2025 and found that high-bandwidth memory stacks degraded at just 2 kilorad of silicon dose - a level that would be reached in under two years at shielded LEO.
The second constraint is thermal management. Data centres on the ground move waste heat into the atmosphere through air or water, both available in essentially unlimited quantity at negligible marginal cost. In orbit there is no atmosphere and no conductive medium. The only available mechanism is radiation, governed by the Stefan-Boltzmann relation. At a chip-tolerable operating temperature of 20 degrees Celsius and best-case emissivity, net radiative heat rejection peaks at approximately 633 watts per square metre. The International Space Station rejects roughly 70 kilowatts across 422 square metres of radiator panels - an effective density of 166 watts per square metre. A single NVIDIA GB200 NVL72 rack consumes 120 kilowatts and would require approximately 300 square metres of deployed radiator to remain within operating temperature limits. A 10-megawatt orbital cluster needs roughly 25,000 square metres. A 1-gigawatt facility needs approximately 800,000 square metres of radiator surface.
The third constraint is latency and bandwidth. Geostationary orbit imposes a round-trip light-speed delay of 240 milliseconds - more than the entire response-time budget for real-time AI inference. Low Earth orbit at 550 kilometres reduces that to 1.8 milliseconds one-way, which is physically tolerable, but the internal bandwidth requirement of an AI training cluster remains an insuperable problem. A single NVL72 rack achieves all-to-all interconnect bandwidth of 130 terabytes per second across its internal NVLink fabric. The state of the art in space-grade optical inter-satellite links delivers 10 gigabits per second. The gap is five orders of magnitude.
The S-1 numbers, disclosed for the first time on May 20, make the data centre dependency of the valuation explicit. Starlink generated $11.4 billion in revenue in 2025 at a 63 percent segment adjusted EBITDA margin, producing $7.17 billion in operating cash. The AI segment - xAI, Grok, and X - recorded $3.2 billion in 2025 revenue but posted a $6.4 billion operating loss, burning through capital at a rate that accelerated sharply: $7.7 billion in AI capex in the first quarter of 2026 alone. The company's total Q1 2026 net loss was $4.28 billion on $4.69 billion of revenue. By any conventional measure, SpaceX is currently one profitable satellite internet business carrying two loss-making divisions.
The most telling line in the S-1 is the Anthropic deal. SpaceX disclosed that Anthropic - a direct competitor to xAI in frontier AI models - is paying SpaceX $1.25 billion per month for compute capacity at the COLOSSUS terrestrial data centre in Memphis, through May 2029. SpaceX is renting ground-based data centre capacity to its AI competitor, under a contract terminable by either party on 90 days' notice, at the same moment it is telling investors that orbital compute will make terrestrial data centres obsolete within three years. If the company believed its own orbital timeline, this contract would not exist.
Independent engineer Andrew McCalip, then at Varda Space Industries, published the most rigorous public cost model in early 2026. His baseline, assuming a launch cost of $1,000 per kilogram that SpaceX has not yet demonstrated at commercial scale, found orbital capex for a 1-gigawatt deployment of $31.2 billion against $14.8 billion for a comparable terrestrial facility. The levelised cost of energy in orbit came out at $891 per megawatt-hour against $398 on the ground. His summary: "It's not a 25 percent mismatch. It's 400 percent. Closing that is the whole job."
The full inventory of orbital data center hardware that has reached space is short. Starcloud-1, launched November 2025, was a 60-kilogram microsatellite carrying one NVIDIA H100 that ran the open-source Gemma model - meaningful for a press release, irrelevant at hyperscale. Axiom's AxDCU-1 was an edge-compute prototype on the ISS, behind the station's heavy radiation shielding. Lonestar's Freedom mission delivered an 8 TB SSD and a single FPGA to the Moon aboard Intuitive Machines' Athena lander, which toppled on its side in a crater within 24 hours. China's Three-Body Computing Constellation, with 12 satellites and 744 TOPS each, is the only multi-satellite computing constellation actually operating in orbit.
The 277-page S-1 confirms that Musk will serve simultaneously as CEO, CTO, and Chairman of the Board following the IPO. His Class B shares give him 85.1 percent of voting power while he holds approximately 42 percent of equity. SpaceX will claim controlled company status post-IPO, exempting it from requirements that a majority of the board be independent. The structure means that no shareholder vote, no board resolution, and no activist campaign can constrain any decision Musk makes about orbital data centres, Terafab, or anything else.
The S-1 sets out the orbital data centre plan in concrete terms for the first time. SpaceX says it plans to begin deploying AI compute satellites as early as 2028, targeting 100 gigawatts of annual orbital compute capacity as a long-term goal. It claims a total addressable market of $28.5 trillion across its businesses - a figure that exceeds the annual GDP of the United States. These claims sit alongside the risk disclosure language that characterises the orbital programme as involving "significant technical complexity and unproven technologies" that "may not achieve commercial viability." Both statements appear in the same document. One of them is marketing. The other is the legal record.
Aswath Damodaran of NYU Stern, whose independent DCF analysis is the most rigorous public valuation of SpaceX available, placed the company's fundamental value at approximately $1.22 trillion based on its demonstrated businesses - implying a premium of more than $500 billion that must be justified by the orbital data centre programme. Every credible independent assessment converges on the same conclusion: orbital data centres at hyperscale are not a 2028 or 2030 commercial product. They are a mid-2030s research ambition at best, conditional on breakthroughs in radiative thermal management, radiation-tolerant high-performance compute, and launch economics that no actor has yet demonstrated.
The bridge being sold in late June looks impressive from the approach. The far side has not been engineered. The only party with a verifiable financial interest in the public believing otherwise is the one collecting 94 times trailing revenue as the offering price - and retaining 85 percent of the voting power to ensure no one can challenge that assessment once the money is in.
Read the risk factors before you read the roadshow deck.
Related Links| Subscribe Free To Our Daily Newsletters |
| Subscribe Free To Our Daily Newsletters |
