The rocket company is the smallest business in this rocket company.

Starlink is one of the best infrastructure businesses built this century. That is not the question. The question is whether this security, at this price, with these rights, pays us for what we would be taking on. It does not.

• It is not a space company. 61% of revenue is a satellite ISP, 17% is an AI lab, 22% is launch. You would be underwriting a broadband monopoly that is financing an AI capital programme.
• The profitable segment is on a treadmill. Starlink's 63% segment EBITDA margin falls to roughly 31% once you fund the ~2,000 satellites a year that must be replaced forever.
• The valuation already embeds the bull case. Our generous sum-of-the-parts tops out at $104. Spot is $113.50. There is no scenario in our model where we are paid to take the risk today.
• You would own economics without rights. 85.1% voting control with no sunset, mandatory arbitration of securities claims, and a derivative-suit threshold no client of ours could meet.

The name on the certificate is the least useful guide to what you own. On 2025 revenue, SpaceX is a broadband operator first, an AI laboratory second, and a launch provider third.

Read the last two columns together. The segment that produces all of the profit receives a fifth of the capital; the segment that produces none of it receives three fifths.

Starlink has quadrupled its base in two years. It has done so by moving down the income curve — GDP-adjusted pricing across Africa, South-East Asia and Latin America. Both of these facts are true at once, and only one of them is in the bull case.

A 33% ARPU decline against a 348% subscriber increase still compounds revenue. But it changes what the terminal business is worth: every incremental cohort is cheaper, and the satellite serving it costs the same. Direct-to-cell now has 10m users across 30+ carrier partners in 22 countries, underpinned by $19.6bn of EchoStar S-band spectrum — a genuine, separable moat no rival can replicate at any price.

A Starlink satellite has a design life of roughly five years, after which it is deliberately de-orbited and burned up. To hold a 10,400-satellite constellation flat, SpaceX must launch about 2,000 replacements a year — forever. That is not growth capex. It is the cost of standing still, and EBITDA does not see it.

We are not dismissing this business. It is remarkable. It is also 22% of revenue growing at 8%, and the market is not paying $1.5 trillion for it.

In February 2026 SpaceX absorbed xAI in the largest merger ever recorded — SpaceX at $1.0tn, xAI at $250bn. What that transaction did was point the cash flows of a satellite broadband monopoly at an AI compute build-out.

• You are funding, not harvesting. Between now and 2029 the equity's job is to absorb burn. There is no dividend, no buyback, and no earnings to anchor a valuation.
• Dilution is a live risk. The company raised $75bn of equity and $25bn of debt inside two weeks. If AI capex holds near a $40bn annual run-rate, a second equity raise before breakeven is a reasonable expectation, not a tail case.
• Musk's award adds ~1.3bn shares. A 1.0bn-share grant plus a 302m-share replacement award, roughly 10% dilution. The 1.0bn tranche only vests alongside a million-person Mars colony, so we treat it as benign; the 302m tranche, tied to market capitalisation and orbital data centres, is not.

This is the part of the business a credit analyst would like. It is also $28.4bn against a $1.48tn enterprise value: roughly 1.9% of what you are being asked to pay.

The honest read is neither "unassailable moat" nor "Amazon is coming". It is that Starlink has a five-year operating lead, that the lead is narrowing at the edges, and that the binding constraint on Starlink's economics is more likely to be terrestrial fibre than another constellation.

We rank this above Starship, above Amazon, above the multiple. Every other risk in this note can be monitored and re-underwritten. This one is structural, permanent, and was chosen deliberately.

The IPO premium above intrinsic value is essentially one isolated bet: that xAI plus space-based compute becomes a real business, not an open-ended capex sink. Aswath Damodaran, NYU Stern — valuing the equity at $1.3tn, 28% below the IPO valuation

• We raise the discount rate, not the narrative. Our 2030 model discounts at 12% rather than the 9–10% we would apply to a comparable infrastructure asset with ordinary governance. That spread is the governance charge, and it costs roughly a fifth of present value.
• We cap the position. A holding you cannot escalate, litigate or vote is a holding you must be able to lose. Hence the 3% ceiling in §12 — a constraint that binds before valuation does.
• Key-man risk is undiversifiable here. Musk is simultaneously the controlling shareholder, the chief executive, the chief engineer of Starship, the counterparty relationship with the US government, and a politically exposed person. There is no succession disclosure that mitigates this, because there is no structure that could.

The June price action was not price discovery. A $75bn offering into a 4% float, followed by a fast-tracked Nasdaq-100 inclusion forcing $4.3bn of index buying, produced a $225 print that no willing seller ever tested. The correction to $113 is that artefact unwinding — and it is only partly done.

Two independent routes are run and cross-checked. The first values the segments as they are today. Change the scenario and every figure below moves with it.

The second route projects each segment to 2030 and discounts back at 12%. It lands at $41 against the sum-of-the-parts' $49 — and the agreement between two different methods is why we hold the conclusion with confidence.

Discounted at 12% over four years — a deliberate premium to the 9–10% an infrastructure asset of this quality would ordinarily carry, reflecting the governance structure in §11.

To break even on a four-year view, SpaceX must reach roughly $2.4 trillion of enterprise value in 2030 — which on a 22× multiple means about $110bn of EBITDA, some two and a half times our base case and close to Alphabet's EBITDA today.

We are not claiming better information than Morgan Stanley. We are making a different judgement about one line: what a pre-scale AI business inside a satellite operator is worth today, and what discount rate a shareholder with no enforcement rights should demand.

• The AI segment. Goldman carries xAI revenue from $3.2bn to $322bn by 2030 — 100× in five years. We carry it to $60bn in our base case, still a nineteen-fold increase and already generous. Nearly the entire target gap sits in this one assumption.
• The Starlink multiple. The Street capitalises segment EBITDA. We capitalise segment EBITDA after the replacement capital required to keep the constellation in orbit — roughly halving the base. That is the treadmill in §05.
• The discount rate. We charge 12% where an infrastructure comparable would carry 9–10%, because the governance structure removes the remedies that ordinarily justify the lower rate.

Ordered by expected impact on value, not by likelihood or by how often the risk appears in press coverage. The two highest-ranked items are structural rather than operational — they cannot be resolved by good execution.

This is a price and evidence decision, not a quality judgement. We want to own this asset. We want to own it at a price that pays us for what sits around it, and after the shares that will be sold have been sold.

The 3.0% ceiling is set by governance, not by valuation. Even at $30 we would not exceed it, because the downside cannot be escalated, litigated or voted.