Business Overview

Space Exploration Technologies Corp.

Nasdaq: SPCX

6 September 2026

Evidence base: the 424(b)(4) IPO prospectus dated 12 June 2026; the Form 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026); and the Forms 8-K filed between 15 June and 14 August 2026.

Industry structure, competitor economics and verification of the company's own claims come from independent sources: FCC, FAA, BryceTech, GAO, ITU, and the filings of named competitors.

This is a business and industry overview. It is not a valuation and not a recommendation.

1. Executive Snapshot

2. What the Company Does

SpaceX solves three different customer problems with one set of physical assets. A government or satellite operator needs mass delivered to a specific orbit on a schedule, and buys a launch. A household, ship, aircraft or army unit outside the reach of fibre or cell coverage needs an internet connection, and buys a Starlink subscription. An AI developer needs GPU capacity and power, and rents compute. The first business builds the vehicles; the second is the largest customer of those vehicles; the third arrived by merger in February 2026 and now consumes most of the capital.

The reporting entity is not what its name suggests. Under the xAI merger, completed 2 February 2026, xAI and X were combined into SpaceX's accounts as a reorganization of entities under common control, because Elon Musk controlled all three throughout. GAAP therefore requires the historical financial statements to be restated as though the combination had always existed. Every pre-2026 figure in this document — including the fiscal 2023 and 2024 revenue and losses — already contains X and xAI, even though SpaceX did not legally own them at the time (424B4, Note 1). This is the single most important thing to understand before reading any growth rate.

The unit of economics: one Starlink subscriber

The company defines a Starlink Subscriber as a unique service line on a Starlink.com account without a negotiated sales agreement — that is, retail and small-business, excluding managed enterprise and government accounts (424B4, Key Business Metrics). There were 12.0 million at 30 June 2026, against 6.0 million a year earlier, growth of 101.2% (Q2 2026 10-Q, MD&A). ARPU was $66 per month in Q2 2026, against $85 a year earlier, a decline of 22.4%. The company states plainly that it expects ARPU to keep falling as international mix rises and lower-priced plans are added.

Trace one subscriber from sale to cash. The customer pays an upfront terminal price and then a monthly subscription. The terminal is a separate performance obligation, so its revenue is recognized at delivery, and the transaction price is split between hardware and service at each obligation's standalone selling price (424B4, Note 2). The service contract is month-to-month, so revenue recognized each month equals the amount billed that month; where a customer pays in advance, the cash sits in deferred revenue until earned. Consolidated deferred revenue was $14,286m at 30 June 2026, up from $12,116m at the end of 2025. The subscriber is therefore cash-positive early and low-margin per dollar of revenue — the opposite shape from a leased-equipment model, where the operator capitalizes the box, collects nothing upfront, and recovers it through a higher monthly fee over years. Section 5 quantifies what that split does to the reported margin and to cash.

What the company does not disclose is the terminal's unit cost or unit margin, and it makes no statement about whether hardware is sold below cost. Nor does it publish a churn rate. The only retention evidence given is qualitative: since 2023 no Starlink Enterprise customer contributing more than $750,000 of annual revenue has voluntarily discontinued service (424B4). For a business whose whole case rests on a recurring subscription, the absence of a churn metric is a material gap.

The unit of economics: one launch, and why the number is misleading

Space sells launches under two revenue models: Launch Services, where revenue and costs are deferred until the customer's spacecraft reaches its intended orbit, and Launch and Development, where longer government programmes are recognized over time on a cost-to-cost basis (424B4, Note 2). All U.S. government launch contracts are firm fixed-price with milestone payments, so cost overruns are absorbed by the company.

The critical distinction is customer versus internal launches. Of 165 Falcon launches in 2025, 43 were for customers and 122 were internal — flights carrying the company's own Starlink and, increasingly, its own AI-related payloads (424B4, Key Business Metrics). In Q2 2026 the split was 10 customer and 27 internal, out of 37. Dividing Space segment revenue by total Falcon launches therefore produces a number that falls as the business grows: roughly $37.1m per launch in 2023, $28.3m in 2024, $24.8m in 2025 (derived; the company does not publish this). Management addresses this directly, stating that while total Falcon launches rose from 134 to 165 in 2025, customer launches and average price per launch remained relatively flat.

The consequence is that launch is not primarily a revenue line. It is the delivery mechanism for the Connectivity constellation, and its cost is the true product. That is confirmed by the segment result: Space earned $962m of revenue in Q2 2026 and lost $542m at the operating line, with segment adjusted EBITDA of negative $205m. The launch business does not make money at the scale at which it dominates the world's launch market.

Segments, and where the money actually is

Sources: 424B4 segment note (FY2025); Q2 2026 10-Q Note 18 (Q2 2026). Segment adjusted EBITDA is a company-defined non-GAAP measure: segment operating income excluding depreciation and amortization, stock-based compensation, restructuring and impairment. Prior periods were recast when the three-segment structure was adopted in Q1 2026.

Connectivity is the company. It produced 61% of fiscal 2025 revenue and all of the profit, and within it, consumer subscribers were over 60% of segment revenue in 2025 (424B4). Its second engine is the enterprise and government tier — aviation, maritime, land mobility, Starshield for U.S. national security, and satellite-to-mobile through roughly thirty mobile-network partners — which grew from $983m in Q1 2025 to $1,806m in Q2 2026 and is the faster-growing half. These tiers matter economically because they are sold across the same satellites the consumer business already requires, so incremental revenue arrives against sunk capacity.

The AI segment reveals a business changing shape mid-flight. Advertising, the legacy X revenue line, is shrinking: $367m in Q2 2026 against $426m a year earlier, which management attributes to a transition to a new advertising platform. What replaced it is AI Solutions and Infrastructure, which went from $311m to $2,194m in the same quarter, of which $1,600m is described as new revenue from beginning to offer cloud services to customers (Q2 2026 10-Q, MD&A). Selling GPU capacity and selling advertising are not the same business, and the segment now does mostly the former.

3. Industry, Competitive Position and Moat

Two industries matter here, and they have opposite profit structures. In launch, the customer buys transport; in satellite communications, the customer buys a service. The evidence — including the company's own segment disclosure — says the profit pool sits with the service operator, not with the transport provider, and that within the operator it is consumed by the capital cycle.

Where the profit sits, and why launch is not it

Rocket Lab, the only meaningful listed pure-play launch and space-systems comparator, has never earned an operating profit: fiscal 2025 revenue of $601.8m, up 38%, gross margin of 34.4%, and an operating loss of $228.8m (Rocket Lab FY2025 filings). SpaceX's own Space segment lost money in Q2 2026 while flying more than half the world's launches. Satellite operators, by contrast, report high headline margins — Viasat 33.4% EBITDA margin in fiscal 2026, SES 45.2% in H1 2026, Eutelsat 51.2% in fiscal 2025-26 — but Eutelsat spent roughly €900m of capex against €1,236m of revenue in the same period. High margin, negative free cash flow. That is the industry's actual economic signature, and it is the shape SpaceX's Connectivity segment will take as its fleet ages.

Satellite manufacturing, the third stage of the chain, is structurally broken: only seven commercial geostationary satellite orders were placed in 2024, the fewest since 1994 (Analysys Mason), and Airbus, Thales and Leonardo signed a memorandum of understanding in October 2025 to merge their space units. The chain therefore rewards whoever owns the end customer and can absorb the transport and hardware stages internally. That is exactly what SpaceX is.

Market structure and named competitors

Global orbital launch reached 325 attempts in 2025, up 25%, deploying 4,544 spacecraft and roughly 2.7 million kilograms of upmass (BryceTech). SpaceX flew 165 of those launches, 51% of the world total, and deployed 85% of all satellites. But 122 of its 165 launches were its own Starlink missions — meaning roughly 38% of every orbital launch on Earth in 2025 was SpaceX launching SpaceX hardware. Global launch volume is not an independent demand signal; it is a derivative of four or five constellation capex decisions, SpaceX's own being the largest.

On the launch side the company names United Launch Alliance, Arianespace and Northrop Grumman in its own filing. Independent evidence on their positions is unflattering to all three and to the new entrants. ULA's Vulcan flew the USSF-87 mission in February 2026 with a solid-rocket-booster anomaly and has never approached its stated cadence ambitions; the U.S. Space Force's own contract awards price ULA at roughly $214m per National Security Space Launch Lane 2 mission against SpaceX's $142.8m, a premium of about 50%. Blue Origin flew New Glenn three times in nineteen months, lost a payload on NG-3, and on 28 May 2026 suffered a fuelled-vehicle explosion at LC-36, its only pad, traced to a BE-4 main oxygen valve; it has no fiscal 2026 national security missions assigned. Ariane 6 has recorded four successes in sixteen months and targets no more than eight launches in 2026. China's LandSpace landed a Zhuque-3 booster on its second flight and targets reflight within six months.

In satellite broadband the picture is different, because one competitor has both capital and a deadline. Amazon's constellation — named in SpaceX's own filing only as Amazon LEO — was required by the FCC to have 1,616 satellites in orbit by 30 July 2026 and had roughly 369 in June 2026. The FCC granted a limited waiver rather than the twenty-four-month extension Amazon sought, and the 30 July 2029 hundred-percent deadline stands, with post-deadline satellites demoted in spectrum processing priority. Amazon must average roughly 950 satellites a year for three years while New Glenn has no operational pad and Vulcan has never hit cadence. The likeliest resolution is that Amazon pays SpaceX to launch a competing constellation, which is a revealing statement about where the bottleneck is.

Legacy operators show what LEO does to incumbents. Viasat's U.S. fixed broadband base is down to roughly 130,000 subscribers at $113 ARPU; HughesNet fell to 681,000, down 20% year on year and down 58,000 in a single quarter, and EchoStar has warned of going-concern doubt. LEO did not compress the industry's price — it took the incumbents' customers. Both survivors pivoted to government and mobility, the same tiers SpaceX is now stacking.

Barriers that bind, and barriers that only sound impressive

Three constraints genuinely bind. Pads: FAA and Space Force environmental approvals cap SLC-40 at 120 launches a year, LC-39A at 20 and SLC-4E at 100, roughly 240 approved Falcon slots in total, and LC-36 demonstrated that a single-pad operator is a single point of failure. Engines and tanks: BE-4 valves, Neutron tank failures and eight Starship mishap investigations are all the same barrier expressed differently. Flight count: reliability is bought with flights, and flights cannot be bought with money. Capital, by contrast, does not bind entry — total global space start-up investment was $10.9bn in 2025 — and FAA licensing, frequently cited as a bottleneck, has not gated any U.S. programme in this dataset since Part 450 began permitting portfolio licences.

What would be hardest for a well-funded competitor to reproduce is not any single asset but the combination operating at rate: roughly 70 satellites a week from the Redmond facility, about 200,000 user terminals a week, over 23,000 inter-satellite laser links, more than 400 ground stations, market access in 167 countries, four operational launch pads with four more Starship pads targeted by end-2027, and the flight history behind a stated 99%-plus Falcon mission success rate. Reproducing any one of those is a funding decision. Reproducing all of them simultaneously is a decade.

Testing the company's claims against outside evidence

The industry rewards the vertically integrated operator that owns launch, spectrum, manufacturing and several demand tiers at once, because that is the only configuration in which the transport stage's lack of profit does not matter. SpaceX is the only company that is all four. The open question the evidence raises is not whether it is the right kind of company for this industry, but whether it is now attempting to be that company for a second industry — AI compute — on the same balance sheet, at the moment the first industry's replacement bill comes due.

4. Growth Engine

Consolidated revenue rose 91.9% in Q2 2026, to $7,814m from $4,071m. That number describes almost nothing about the business that existed a year earlier, and decomposing it is the most useful analytical act available on these filings.

Reported growth versus what the underlying businesses did

Source: Q2 2026 10-Q, Note 3 and Note 18. Percentages of total growth are derived. The AI segment is present in the prior-year comparative only because of common-control reorganization accounting; SpaceX did not own xAI or X in Q2 2025.

Half of the reported growth is a business line that did not exist as a revenue stream a year ago, sold largely to a single new counterparty. Management attributes $1,600m of the AI increase specifically to having begun to offer cloud services to customers, and Customer B — unnamed, AI segment only — is 19.5% of consolidated revenue in the quarter. Meanwhile the legacy advertising line inside the same segment is shrinking. Excluding AI entirely, the space and connectivity businesses grew 57.6%, which is the more meaningful rate and still a strong one.

The half-year is a useful corrective to the quarter. H1 2026 revenue grew 53.7%, but Space revenue fell 1.9% year on year, because customer launches dropped from 21 to 17. A business flying 77 Falcon missions in six months and recording lower launch revenue than the prior year is a business whose launch volume has decoupled from its launch revenue.

The drivers, ranked

• Starlink subscriber additions — structural. Subscribers doubled from 6.0 million to 12.0 million in twelve months. This is the durable engine, and it works because each new subscriber is served by capacity that already exists in orbit. The offset is that ARPU fell 22.4% over the same period, from $85 to $66, and the company expects further decline as international mix rises. Growth is real but each unit is worth less, so the segment must add subscribers faster than price falls — in Q2 2026 it did, by a wide margin.
• The enterprise, government and mobile tiers — structural. Enterprise and government revenue inside Connectivity rose from $867m to $1,806m year on year in Q2. These tiers monetize the same satellites at higher revenue per unit of capacity — aviation, maritime, land mobility, Starshield, and satellite-to-mobile through roughly thirty mobile-network partners covering about 1.9 billion people. This is the highest-quality growth in the company because it raises revenue per satellite without raising the satellite count.
• AI cloud compute — management-driven, and unproven. From $311m to $2,194m in a quarter, concentrated in one customer, on contracts the company discloses as terminable on 90 days' notice after an initial ramp period (Q2 2026 10-Q, Part II Item 1A). The revenue is real; its durability is a different question, and the filing itself is the source of the doubt.
• Space customer launches — cyclical, and currently negative. Customer launches fell from 21 to 17 in the first half. Space revenue growth in Q2 came from mix and a single additional customer launch, not from volume. Global launch demand outside constellation self-supply is thin, and BryceTech's own Q2 2026 data shows spacecraft deployed down 4% and upmass down 2% year on year.
• Acquisitions — temporary in effect, permanent in dilution. Cursor closed 14 August 2026 for 389.3 million Class A shares at a $60.0bn implied equity value, after the reporting period. Mesh Optical closed 6 July 2026 for about 3.8 million shares. The EchoStar spectrum transaction remains open. None is in any reported figure in this document.

There is no financial guidance anywhere in the prospectus or the 10-Q. The company also disclosed at IPO that it will not use wire services to distribute results, publishing instead through its investor site and its own X account (8-K, 15 June 2026, Item 7.01). An investor gets the filings and nothing else.

5. Margin, Cash and Capital Allocation

The margin question has a clean answer at segment level and a confusing one at group level, because three businesses with different economics are averaged together. Connectivity converted $4,291m of Q2 2026 revenue into $2,597m of segment adjusted EBITDA — a 61% margin — because once a satellite is in orbit and a terminal is in a customer's hands, the marginal cost of another month of service is close to the cost of ground infrastructure and payment processing. That is the software-like part of the business. But not every dollar of that revenue is the same dollar.

Two revenue dollars, and a mix that is still shifting

Connectivity sells two things and the filings separate them. Products revenue — Starlink Kits, all of which is attributable to Connectivity — was $1,093m in fiscal 2023, $1,470m in fiscal 2024 and $1,510m in fiscal 2025, a rise of 2.7% in a year when the subscriber base roughly doubled (424B4, Note 3). In Q2 2026 it was $461m against $403m, up 14% against 101% subscriber growth. Realized revenue per terminal is falling steeply, alongside the company's own statement that it has significantly lowered production costs and scaled output to roughly 200,000 kits a week. Whether hardware is sold below cost is not disclosed and cannot be established from these documents.

Sources: 424B4 Note 3 and segment note; Q2 2026 10-Q Note 3 and Note 18. The share is derived — the company does not present this ratio, and Q2 2026 is a single quarter against full years.

The series matters because the two dollars carry different margins. A service dollar is close to pure contribution once the satellite is in orbit and the terminal is in a customer's hands; a hardware dollar is at best thin. As hardware falls from 28% of segment revenue to 11%, reported Connectivity margin improves on mix alone, before any operating improvement at all. It also means most of that tailwind has already been taken — the share cannot fall much further, so mix will stop flattering the margin within a few years and the trend will have to be carried by the service line on its own.

The same split shapes cash. Because the terminal is paid for at sign-up and many service, enterprise and government contracts are prepaid, growth in the customer base converts directly into deferred revenue. In the first half of 2026 the movement in deferred revenue supplied $2,169m of $3,466m of operating cash flow — 63% of it, against $680m in the prior-year half. Management attributes the increase to upfront payments from Space and Connectivity customers, so launch deposits are included and this is not a Starlink-only effect. The consequence is directional and testable: a deceleration in customer additions would remove a working-capital contribution as well as revenue, and operating cash flow would decelerate faster than revenue does.

Depreciation is the business model

Broadband satellites are depreciated over five years and first-generation mobile satellites over three (424B4, Note 2). Flight vehicles are depreciated by flights rather than years: Falcon 9 boosters are engineered for up to 40 flights but carry an accounting life capped at 25, reflecting the transition to Starship and a five-flight cap under certain government contracts. The company discloses the sensitivity directly: a one-year change in average satellite useful life would have moved fiscal 2025 operating income by roughly $480m, and Q1 2026 alone by about $170m. By contrast, a five-flight change in average remaining booster flights is immaterial. The asset that matters is the satellite, not the rocket.

This is why the Connectivity margin should be read as a payment plan rather than as a profit. The obligation behind it is non-deferrable in a way that ordinary maintenance capital is not: an unreplaced satellite does not degrade, it deorbits, and its capacity disappears from the network on a schedule set at launch. Section 6 sizes what that obligation becomes at authorized fleet size.

The financial spine

Sources: 424B4 consolidated statements (FY2023-25); Q2 2026 10-Q (H1 2026). Comparability warnings: (i) fiscal 2023-25 include xAI and X under common-control reorganization accounting even though SpaceX did not own them; (ii) the fiscal 2023 loss is driven by a $3,775m impairment of the Twitter brand on the rebrand to X; (iii) the segment structure changed in Q1 2026 with prior periods recast; (iv) a five-for-one stock split in May 2026 restates all share and per-share data; (v) H1 2026 includes a $1,545m loss on debt extinguishment. Free cash flow is not a company-reported measure.

Read across that table and the story is unambiguous. Revenue has compounded, adjusted EBITDA has compounded, and free cash flow has gone from roughly break-even in 2023 to a $25.0bn outflow in six months. The gap is entirely capital expenditure, and capital expenditure is almost entirely one segment: of $28,476m spent in H1 2026, $23,551m was AI, against $2,699m for Connectivity and $2,226m for Space. In Q2 alone AI capex was $15,828m, 86% of the company total.

Connectivity, notably, self-funds. Its H1 2026 segment adjusted EBITDA of $4,684m exceeds its $2,699m of capex. The satellite business generates surplus cash at its current fleet size and age. That surplus, the $85,675m of net IPO proceeds and $51,812m of new borrowing are being routed into AI data centres.

Where the cash went, and what that reveals

• AI infrastructure — $23,551m in six months, the dominant use of capital by a wide margin, funding data centres and compute hardware. Nameplate compute draw rose from 0.4 GW to 1.4 GW year on year.
• Connectivity and Space capital — $4,925m combined, for satellites, launch sites and vehicles. The core space business now receives roughly one dollar of capital for every five that go to AI.
• Debt refinancing — $39,396m repaid against $51,812m raised, at a cost of a $1,545m extinguishment loss and $1,153m of premium, clearing the legacy X and xAI debt stack through a $20bn bridge and then a notes issue. Total debt and finance leases stood at $39,512m at 30 June 2026 against $93,522m of cash and $6,487m of securities.
• Equity issued for assets — 389.3m Class A shares for Cursor, 3.8m for Mesh Optical, and 261.8m committed to EchoStar. No dividends and no buybacks; $4,426m was spent repurchasing common and preferred stock in H1 2026, chiefly in connection with the pre-IPO restructuring rather than as a capital return.

The ranking says something plain about how management thinks: this is a company that has decided its scarce satellite and launch franchise is a funding source for an adjacent bet, and it has used a public listing to raise the capital to make that bet larger. Whether that is right is a judgement about AI, not about space.

Transactions that post-date the reported figures

Three material items sit outside every number above. Cursor closed on 14 August 2026 — after the 30 June balance-sheet date — for 389,289,254 Class A shares at a $60.0bn implied equity value, plus 1.75m shares for vested awards and roughly 29.1m assumed restricted units and 44.4m assumed options. No pro-forma financials were filed with the 8-K, and Cursor's revenue and losses are not disclosed anywhere in these documents.

The EchoStar spectrum transaction is mid-flight. Total consideration is approximately $19.6bn: about $11.1bn in equity, payable through roughly 261.8 million Class A shares at a fixed value of $42.40 per share, plus up to $8.5bn to retire designated EchoStar debt. The FCC approved it on 12 May 2026 and the licences transferred to a holding trust on 22 May 2026, but the final transfer to SpaceX is modelled at 30 November 2027, with a disclosed contingency of an additional $827m if it slips to 2028. $856m had been paid by 30 June 2026 and sits as a prepaid asset. Note the arithmetic on the equity leg: the share count is struck at $42.40, the company's October 2025 internal valuation, while the IPO priced at $135.00 — so a fixed 261.8 million shares delivers substantially more market value than the $11.1bn at which the consideration is stated (inferred from the fixed share count; the filing does not restate the consideration at the offer price).

Finally, $27,955m of non-cancellable contractual commitments sit outside the debt schedule, $22,244m of which falls due in 2027, relating principally to AI infrastructure, third-party cloud capacity and the spectrum transaction.

6. Cyclicality, Constraints and What to Monitor

Connectivity is the least cyclical part of the business: a month-to-month broadband subscription is closer to a utility bill than to a capital good, and demand for it does not turn with an industrial cycle. Space and AI are the opposite. Launch demand is a derivative of constellation capital budgets, and AI compute demand is a derivative of a handful of hyperscaler capex decisions. The company's cycle exposure is therefore concentrated in the two segments that do not currently make money.

Where the cycle sits now

Launch volume is at or near a cyclical top. After 25% growth in 2025 to 325 orbital launches, BryceTech's Q2 2026 data shows launches up only 3% year on year with spacecraft deployed down 4% and upmass down 2% — and Bryce's own commentary observes that quarterly totals now track a handful of commercial constellation calendars more than broad appetite for orbit. New reusable capacity from Neutron, Terran R, a rebuilt New Glenn and Zhuque-3 is arriving into that deceleration. SpaceX's own customer launches fell from 21 to 17 in the first half of 2026, its Q2 upmass fell from 652 to 485 metric tons, and Space segment adjusted EBITDA went from positive $131m in H1 2025 to negative $556m in H1 2026. On volume and on margin, Space is past its own recent peak.

Connectivity is in the opposite position — at a record on every metric, with subscribers doubling and segment adjusted EBITDA at $4,684m for the half against $3,200m a year earlier. But it is at a record on a young fleet. Legacy satellite communications, meanwhile, is at a thirty-year trough: seven commercial GEO orders in 2024, Airbus, Thales and Leonardo merging their space arms, HughesNet losing 20% of its base in a year, EchoStar flagging going-concern doubt. LEO is growing at lower margins than the GEO revenue it displaced — Eutelsat's EBITDA margin fell 3.2 points and SES's 8.1 points on LEO and mobility mix.

The constraint that compounds

The FCC has authorized SpaceX to operate 15,000 satellites, and roughly 10,200 are on orbit. At the authorized fleet size and a five-year broadband life, the steady-state replacement requirement is about 3,000 satellites a year — roughly 58 a week, or 82% of the stated 70-a-week manufacturing rate, and on the order of 107 Falcon 9 launches a year simply to hold the constellation flat (inferred from the disclosed fleet size, useful life and production rate; the company publishes no replenishment figure). Observed retirements today, at 260 in a half-year, run far below that only because the fleet is young. The gap between 520 deorbits a year and a steady-state 3,000 is the single most important number in this business that nobody publishes.

This reframes Starship. It is discussed as a growth vehicle, and it is one — Falcon 9 and Falcon Heavy are explicitly not capable of deploying the next-generation V3 and V2 Mobile satellites (424B4), so the constellation roadmap is structurally gated on it. But its first economic job is refinancing an obligation already on the balance sheet. Twelve flight tests have been executed, the twelfth in May 2026; the company expects payload delivery to begin in the second half of 2026; orbital propellant transfer has not been demonstrated or attempted, and GAO confirmed as of May 2026 that SpaceX has not yet demonstrated orbital propellant storage and transfer. Space segment research and development ran at $3,004m in fiscal 2025, expensed rather than capitalized because Starship is pre-commercial.

The downside case does not require a failure, only arithmetic. If Starship slips, replenishment of a growing fleet runs on Falcon 9 at a cadence approaching the approved pad limits; a credible reusable competitor makes the $82m national-security Lane 1 price contestable in a segment that has no margin buffer to defend it with; the U.S. addressable niche keeps shrinking as fibre and fixed wireless build out — rural wireline availability at 100/20 Mbps rose from 67% to 75.3% in two years, and only 10.6 million Americans, 3.1%, now lack it (FCC Section 706, August 2026); and LEO mix continues to dilute margins industry-wide. Deterioration would show up as Connectivity growth decelerating while replenishment capex compounds — visible in the deorbit-to-launch ratio long before it appears in the subscriber count.

Durable versus borrowed

Leading indicators, and where they are published

• The deorbit-to-launch ratio — SpaceX's semi-annual constellation report to the FCC, filed in the IBFS system. This is the direct read on the replenishment treadmill and precedes any margin effect by quarters.
• Global and provider launch counts, spacecraft deployed and upmass — BryceTech's quarterly Bryce Briefing; active satellite counts at planet4589.org.
• Starlink subscribers and ARPU together — the company's quarterly 10-Q key business metrics. Neither figure means anything alone.
• Amazon's satellite count against the 30 July 2029 FCC deadline — FCC IBFS filings. Determines whether Amazon becomes SpaceX's largest launch customer or its most damaged competitor.
• National Security Space Launch mission assignments and prices — Space Systems Command award announcements. The cleanest public series on launch pricing.
• Starship flight test outcomes and the orbital propellant transfer demonstration — FAA mishap determinations and GAO's annual NASA Assessments report.
• Blue Origin's LC-36 return to flight and Rocket Lab's Neutron first launch — the two events that would make reusable cadence contestable.
• U.S. broadband availability at 100/20 Mbps — the FCC's biennial Section 706 report, which measures the addressable market shrinking underneath the consumer business.

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality and the replenishment constraint are covered above. What follows is what those do not capture.

• Two businesses funded from one balance sheet, one of which does not yet cover its own capital. Connectivity generated $4,684m of segment adjusted EBITDA in H1 2026 against $2,699m of its own capex; AI consumed $23,551m. If AI compute demand disappoints, the loss is not only the AI segment's — it is the satellite capital that was not spent, at the moment the replacement cycle steepens. Management discloses the ability to reallocate capex away from data centres, which is the mitigant and also the admission.
• Revenue concentration in two undisclosed counterparties. Customer A was 18.3% of Q2 2026 revenue and spans all three segments; Customer B was 19.5% and relates entirely to AI. Neither is named. Together they are roughly 38% of a quarter's revenue. The AI cloud contracts are disclosed as terminable on 90 days' notice after an initial ramp; one termination would remove close to a fifth of consolidated revenue at current run-rate.
• Related-party financing at a scale that changes the leverage picture. Three equipment lease agreements with entities affiliated with Valor Equity Partners — whose founder and chief executive, Antonio Gracias, sits on the board — carry aggregate undiscounted payments of roughly $20.2bn, guaranteed by SpaceX and accounted for as failed sale-leasebacks. The recorded debt rose from $4,507m at December 2025 to $13,329m at June 2026, with $513m of related-party interest expense in the half. This is AI infrastructure financed by a director's firm, disclosed but not independently priced.
• Key-person dependency with no insurance and no ability to remove. The filing states plainly that no key-person life insurance is maintained on Mr. Musk, and that he does not devote his full time and attention to the business, holding roles at Tesla, Neuralink and The Boring Company. Removing him as chief executive or chairman requires a majority of Class B shares voting separately as a class — shares he overwhelmingly holds. Class B carries ten votes and elects 51% of the board. The charter also renounces corporate opportunities and expressly permits Mr. Musk and his affiliates to compete with the company.
• Compensation tied to outcomes with no financial content. Of 1,302,072,285 restricted Class B shares held by Mr. Musk, one billion vest only on fifteen market-capitalization tranches from $500bn to $7.5tn together with the establishment of a permanent human colony on Mars with at least one million inhabitants; the remaining 302.1 million vest on twelve tranches from $1.065tn to $6.565tn together with completion of non-Earth-based data centres delivering 100 terawatts of compute a year. Neither condition is a business result an investor can underwrite, and both point capital allocation at outcomes far outside the current earnings base.
• No insurance on the fleet. The company carries no satellite or launch vehicle insurance. A launch failure grounding the Falcon fleet stops both customer revenue and — more consequentially — constellation replenishment, at a moment when the replacement obligation is compounding.
• Single-site concentration and licence conditions. Starship capability is concentrated at Starbase; Blue Origin's LC-36 explosion in May 2026 is the live demonstration of what one pad means. Separately, many FCC licences carry deployment milestones, reporting and surety-bond conditions, and failure on orbital-debris requirements can bring monetary penalties or loss of licensing authority. The company does not disclose the specific numeric milestone or penalty schedule attaching to its own constellation authorization.
• Prioritizing its own payloads over paying customers. The filing discloses that the company may prioritize its own launches over U.S. government contracts or third-party customers, that this may limit Space segment revenue growth and affect regulatory relationships, and that it could invite litigation from customers or competitors. This is the conflict at the centre of being both the dominant launch provider and the largest constellation operator.
• Litigation and regulatory exposure arriving with the AI segment. A $354m accrual sits against a docket that includes a €120m European Commission fine under the Digital Services Act (under appeal), a $105m patent judgment plus $67m of interest (under appeal), multiple Grok deepfake-imagery suits, Dutch GDPR class actions, and a Clean Air Act suit over gas turbines powering a Mississippi data centre. None of this existed in the space business.

What the sources could not answer

These are gaps in the filings, not failures of research, and each would change the analysis materially.

• Starlink churn. No churn or retention metric is disclosed for consumer subscribers — only a qualitative statement about enterprise customers above $750,000 of annual revenue. For a subscription business this is the central missing number.
• Cost per satellite. Neither the filings nor any primary independent source establishes what a Starlink satellite costs to build. Competitor benchmarks range from roughly €6m per OneWeb satellite to about $45m for IRIS² and SDA Tranche 3. Without SpaceX's own figure, the replenishment bill cannot be sized in dollars.
• Terminal unit economics, and therefore subscriber payback. No terminal unit cost, no unit margin, and no statement of whether hardware is sold below cost. Combined with the missing churn rate, this means the payback period on a Starlink subscriber cannot be computed from these filings at all — the central unit-economic question of the business is unanswerable as disclosed.
• Current cost per kilogram to orbit. Only 2018-vintage NASA reference points and an aspirational Starship target appear anywhere.
• The identity of Customer A and Customer B, and the aggregate share of revenue from the U.S. government as a discrete disclosed metric — only the statement that approximately one-fifth of 2025 revenue came from U.S. federal agencies.
• Cursor's financial results. A $60.0bn acquisition closed with no pro-forma statements and no disclosed revenue or losses.
• Artemis Human Landing System contract value, milestones or payments received. The programme appears only as two sentences of business narrative.
• Segment-level backlog. Total backlog of $47,461m at 30 June 2026 is disclosed only on a consolidated basis, so the durable Connectivity portion cannot be separated from AI infrastructure commitments.

Before forming a thesis, three of these would need resolving above the others: consumer churn, the dollar cost of the replenishment cycle, and whether the AI cloud revenue is contracted for years or renewable at 90 days.

8. Investor Takeaways

• This is a satellite communications company that owns a launch business and has bought an AI business. Connectivity is 61% of fiscal 2025 revenue and all of the profit; launch is the delivery mechanism for it, and loses money at 51% global share.
• The engine is one Starlink subscriber, at $66 a month and falling, multiplied by 12.0 million and doubling. That produced a 61% segment EBITDA margin in Q2 2026 — before the five-year satellite replacement cycle reaches steady state.
• The main growth lever is revenue per satellite, not satellites. Enterprise, government, aviation, maritime and satellite-to-mobile stack additional revenue onto capacity the consumer business already funds, and that tier is growing faster than consumer.
• What could break the story is arithmetic, not failure. A fleet approaching 15,000 satellites on a five-year life needs roughly 3,000 replacements a year — about 82% of stated production capacity — and the vehicle meant to make that cheap has not yet delivered a payload or demonstrated orbital refuelling.
• Monitor the deorbit-to-launch ratio, subscribers and ARPU together, and whether AI capex is funded by Connectivity's cash or the balance sheet. Those three tell you whether the engine is compounding or paying for itself.