2026年二季度SpaceX更新：星链打印，AI燃烧——S1剖析（英）2026

Digging through one of the mostanticipated prospectus filings of all time LATE-STAGE COMPANY RESEARCHSpaceX Update: Starlink Institutional Research Group Prints, AI Burns—A Dissectionof Its S-1 Franco GrandaSenior Research Analyst,Late-Stage Researchfranco.granda@pitchbook.com pbinstitutionalresearch@pitchbook.com Published on May 22, 2026 Digging through one of the most anticipated prospectusfilings of all time Contents PitchBook is a Morningstar company providing the most comprehensive, mostaccurate, and hard-to-find data for professionals doing business in the private markets. Key takeaways •SpaceX is pitching itself as an AI company.AI-related terms account for 47% ofsegment-specific language in the S-1, yet the AI segment generated just 17.1% of2025 revenue (6.7% excluding advertising) and posted a $14 billion free cash flowloss. Connectivity, which generated 61% of revenue and virtually all free cash flow,gets 23% of the narrative. Anyone buying SpaceX above $1.5 trillion is primarilybuying the AI thesis. References12 •The Anthropic deal validates the infrastructure, not the model.Anthropic is paying$1.25 billion per month ($15 billion annually) for access to Colossus and ColossusII, nearly matching combined space and connectivity revenue in 2025. A deleteddisclosure from the confidential filing reveals build costs of $2.7 million/MW versusan industry benchmark of $12.3 million/MW, implying capital expenditure paybackin under two months at contract rates. •Our core business estimates held up well.Combined space and connectivityrevenue came in 2% below our estimate, but EBITDA was 4% above, reflectingbetter-than-modeled profitability. The major surprise was Starlink Mobile ($632million actual versus $42 million estimated), driven by higher ARPU and upfrontMNO payments. The major miss was space, where Starship R&D ($3 billion in 2025)turned an otherwise profitable segment into a $657 million operating loss. •Starlink’s margins are software-grade.At 63% segment-adjusted EBITDA, Starlinkexceeds every telecom and satellite peer we track and approaches Microsoft (65%).Once the satellite constellation is deployed, the marginal cost of an additionalsubscriber is near zero, particularly when customers pay for their own hardware. •The Starship R&D-to-capital-expenditure flip is a potential catalyst.Over$15 billion in cumulative Starship costs are currently expensed to R&D. Oncecommercial payload delivery begins (we expect early 2027), the reclassificationto PP&E alone would add approximately $3 billion to reported operating income, aswing the market may underappreciate. •The balance sheet is not what secondary investors were pricing.Total debt standsat $29.1 billion, with net debt of $13.2 billion. A $20 billion bridge loan matures inSeptember 2027, creating a refinancing wall 15 months post-IPO. Related-partyleases with a board member’s fund total $20.2 billion in undiscounted payments atan implied 8.3% rate, above the bridge loan’s 4.58%. We expect proxy advisory firmsto flag the governance package, which includes dual-class control, mandatoryarbitration, and a corporate opportunities renunciation for Elon Musk. •Critical disclosures are missing.The S-1 omits subscriber churn, Falcon 9 uniteconomics, and AI segment granularity—no breakout of Grok subscription, Xsubscription, or API revenue; no utilization rate on 1.0 GW of deployed compute.The segment occupying 47% of the narrative remains effectively a black box. Introduction SpaceX filed its S-1 registration statement on May 20, 2026, ahead of what is expectedto be the largest IPO in history. Our initiation report, published in March, modeledthe core launch and Starlink businesses from 2010 to 2040 using triangulatedestimates; the S-1 replaces that triangulation with ground truth—at least partially, givensome glaring omissions. The key incremental data here is around the all-importantAI business. This note does not attempt to summarize the prospectus. Its purpose is to surface thedata points that matter most, identify what the filing deliberately omits, and assesswhere our estimates were right and wrong. We will publish a follow-up note with our revised segment-level financial model and adeeper analysis of the company’s valuation ahead of the IPO. SpaceX is now an AI company It is worth highlighting how the most important takeaway from the S-1 is aframing decision. SpaceX is presenting itself to public markets primarily as an AIinfrastructure company. We analyzed term frequency and strategic emphasis across the nearly 300-pagedocument. AI-related terms, including “AI,” “compute,” “xAI,” “Grok,” “datacenter,”“token,” and “inference,” appear over 2,650 times, making up about 47% of segment-specific language. Space and connectivity account for about 30% and 23%,respectively. “AI” alone (over 1,000 mentions) exceeds “Starlink” (381) and “Falcon” (184) combined. “Compute” (448) outpaces “Starlink” on its own. The ProspectusSummary conta