SpaceX

IThe business

SpaceX designs and flies rockets, runs the Starlink satellite network and, since 2 February 2026, owns xAI and X as well.³ It listed on Nasdaq on 12 June 2026 at $135.00 a share and raised $85.7B net.⁴ It has filed no annual report yet, so the audited record is three years, 2023 to 2025, restated to include xAI and X in every period.¹⁵³

The company reports three segments. Space launches Falcon rockets, flies Dragon and develops Starship. Connectivity sells Starlink to consumers, businesses, governments and, through partners, phones. Super Intelligence, called SI, runs the Colossus data centers, rents compute to other companies, and sells Grok, Cursor and X.³⁴

In the twelve months to September 2026 revenue was $32.0B, up from $18.7B in 2025. Q3 alone was $14.6B, up 161% on the year C (calculated).³⁴⁶ Connectivity was 50% of the twelve months, SI 37% and Space 13% C (calculated).

The three segments are three different economic engines, and the profit lines show it. Connectivity earned $6.5B of operating income on $16.0B of revenue and spent $5.5B of capital. Space lost $1.7B, mostly because Starship development cost $4.2B of research in the year. SI lost $5.5B at the operating line, after $6.9B of depreciation and $2.4B of stock pay, and spent $44.6B of capital C (calculated).³⁴⁶

What it sells

• Launch. SpaceX flew 165 Falcon rockets in 2025 and 110 in the first nine months of 2026, down 11% on the same months of 2025. Of the 2026 launches 74% carried SpaceX’s own satellites, and Space books no revenue on them.³⁴
• Starlink. 13.6 million consumer subscribers at 30 September 2026, up 89% on the year, paying an average of $69 a month, 17% less than a year earlier and 5% more than in Q2. Enterprise and government customers paid $2.1B in Q3, up 91% on the year C (calculated).⁴⁶
• Compute. Nameplate compute, the GPUs installed multiplied by their power draw, rose from 0.6 gigawatts to 1.7 in a year. Management plans to use about half of it internally and to rent the rest at close to spot rates. The cloud contracts run typically three years and either side can end them on 90 days’ notice; the two named contracts start with a locked period.^4204445
• Products. Grok, Cursor and X. Cursor was bought on 14 August for an implied $60B and brought $608M of revenue in 48 days. X advertising was $538M in Q3, up 19% on the year, but $1,248M in nine months against $1,323M a year earlier C (calculated).⁴⁶

Who pays

Two customers paid 51% of Q3 revenue. Customer A, 16.6%, buys from all three segments and is not named. Customer B, 34.5% or about $5.0B, buys SI compute, was below 10% in Q1, and is not named C (calculated).⁴⁵³ The prospectus names Anthropic as the counterparty to a $1.25B-a-month cloud agreement.³⁴ Anthropic’s S-1 puts its agreements, as amended, at about $3.1B a month once all capacity is in, expected in December.⁴⁴ Google agreed in June to pay $920M a month from October.⁴⁵ No filing says who Customer B is, so I do not equate it with either.

Contracted backlog was $63.0B at 30 September, up from $28.4B at the end of 2025, and 67% of it falls due within a year, about $42B C (calculated).⁴³ The CFO’s guide for the first quarter of 2027, $33B to $37B of revenue against $14.6B in Q3, comes from the call. It is in neither the 10-Q nor the Q3 release. The year-one section of the valuation sets the $42B against the revenue my base case needs.²⁰²¹

Q3 is not a run rate

Five items flatter the quarter or sit outside it. Customer B was 34.5% of revenue. Cursor was in for 48 of the quarter’s 92 days. Net income of $4.4B included a $561M gain on the Boring Company stake and a $390M tax benefit. The CFO said about three points of Connectivity gross margin came from one-time items, mainly tariff refunds. He also said fourth-quarter gross margin should look more like Q2 than Q3.⁴²⁰²¹ I use Q3 as evidence of direction. The level for year one comes from the December goal and the Q1 guide.

IIThe moat

I think SpaceX’s moat is wide in launch and in Starlink today. It is widening in Starlink and not in Falcon, whose launches fell 11% this year. Starship is the test of whether launch widens. In AI compute I do not think SpaceX has a moat yet. It has a speed lead.³⁴

The edge is scale and cadence, and the filings do not disclose a unit cost to prove more. SpaceX makes the rocket, the satellite, the dish and the ground station, and 74% of its Falcon launches in 2026 carried its own satellites, so launching a Starlink satellite is a transfer inside the company.³⁴ Spectrum is a barrier and a cost: SpaceX is paying $19.6B for EchoStar’s.⁴

The launch evidence is plain. SpaceX flew 165 of the world’s 325 orbital launches in 2025 and about 82% of the mass, and 11 of the 12 National Security Space Launch missions.³²² A booster has flown 34 times.³ The lead is also a captive customer: 83% of mass to orbit in the first nine months of 2026 was SpaceX’s own satellites, which pays for a cadence that rivals have to buy elsewhere.⁴

Evidence it is widening

Starlink scale. Starlink had about 9,600 satellites in March, about 75% of all active maneuverable satellites in orbit, and over 11,100 by 30 September. Connectivity revenue grew 57% in the first nine months of 2026 against the same months of 2025, and the segment earned a 40% operating margin and a 64% Adjusted EBITDA margin C (calculated).³⁴⁶

A new subscriber costs about $855 of capital spending and earns about $529 of EBITDA a year, a payback of 1.6 years. The capital figure is overstated, because the segment’s spending also covers enterprise, government and replacement satellites, and churn and acquisition cost are not disclosed C (calculated).⁴⁶

Starship. The V3 satellite is designed for 1 terabit a second of downlink, and one Starship launch can carry up to 60 of them.³ Flight 14 in the third quarter carried 26, added about ten times the downlink of a Falcon 9 Starlink launch, and was the first Starship flight to earn revenue. Management expects weekly flights in the first half of 2027.⁶²⁰ These are company statements. If they hold, the cost of each added satellite falls, and Falcon becomes the older product.

Evidence it is narrowing, or not yet there

Compute is a commodity input. SpaceX buys its GPUs from Nvidia on purchase orders, as its rivals do, and its cloud contracts end on 90 days’ notice.³⁴ It built the first Colossus II cluster in 91 days against a two-year industry benchmark, a company claim I could not test.³ Speed is an edge. Rivals with money can copy it: CoreWeave’s revenue grew 112% in its latest quarter and Amazon’s net capital spending was $169.0B in the twelve months to June.³¹³²

Prices are falling in Starlink. Subscriber ARPU was $91 in 2024 and $81 in 2025. On the narrower consumer definition it was $66 in Q1 and Q2 2026 and $69 in Q3, and the company expects ARPU to keep falling for a few years as plans get cheaper.³⁴⁶

Launch is not growing. Falcon launches fell 11% and mass to orbit 7% in the first nine months of 2026. Space revenue grew 2% in that period C (calculated).⁴ The company expects Falcon 9 launches to decrease as Starship takes over.³

Size is not a low cost of capital. SpaceX raised $85.7B of equity and $25B of notes in 2026, and the market priced the notes at 91.7% of par on 30 September, a yield I calculate near 7.2% against a 5.00% ten-year Treasury C (calculated).⁴⁹¹⁷ Rocket Lab’s whole market value is $34.4B, one fifty-third of SpaceX’s.^393927374

SpaceX has a moat in launch and Starlink, which together earn $4.8B of operating income, $6.5B from Starlink less $1.7B lost in launch, and none yet in the one that is 37% of revenue and 81% of capital spending C (calculated).⁴

IIIThe competitive advantage

SpaceX’s advantage is scale, which shows in launches flown, satellites in orbit and revenue in three businesses where the rivals are far smaller. It owns the rocket, the satellite and the dish, and it buys its GPUs, turbines and servers. The advantage is greatest where it owns the chain and smallest where it buys an input or the customer can leave on 90 days’ notice.³⁴

Launch

In 2025 SpaceX flew 165 launches against 68 for CASC, 18 for Rocket Lab and 17 for Roscosmos, according to BryceTech. Blue Origin and United Launch Alliance are not among the four launchers it lists with double-digit counts.²² SpaceX’s Space segment took in $4.2B in the twelve months, 5.4 times the $0.77B of Rocket Lab, which the market values at 41.6 times revenue on enterprise value³⁹³⁹ C (calculated).⁴²⁷ The Space segment loses money. Starship research was $3.0B in 2025 and $3.2B in nine months of 2026.³⁴ Starlink pays for the cadence.

Broadband

Connectivity’s $16.0B of revenue and 40% operating margin compare with Viasat’s $4.6B, where revenue fell 1% in the latest quarter, and Iridium’s $0.9B, up 8% C (calculated).⁴²⁸²⁹ Amazon LEO is the real rival. It had 396 satellites in orbit on 2 July, the latest count I found, and plans 3,200 by 2029. Starlink had over 11,100 on 30 September.²³⁴

AI compute

SI took in $7.6B in Q3, against $2.6B for CoreWeave in its quarter to June, and CoreWeave’s revenue was up 112% on the year. On compute and Grok alone SI billed $6.4B, 2.5 times CoreWeave C (calculated).⁶³¹ AWS took in $42.2B in its second quarter, $168.9B annualized. The first-quarter guide of $33B to $37B for all of SpaceX would equal 78% to 88% of that quarter C (calculated).²⁰³² The difference is the term: SpaceX’s cloud contracts are typically three years and end on 90 days’ notice.⁴

The market values the two companies’ gigawatts at different scales. CoreWeave’s enterprise value is $60B on 1.5 GW of active power, $40B a GW at the Q4 2026 close.³⁹ SpaceX’s is $1,778B on 1.7 GW of nameplate compute, $1,046B a GW, 26 times as much. The bases differ, and CoreWeave does not define active power, so I read the gap as 20 to 30 times. Even if Starlink is worth my base value of $187B, the rest of SpaceX is $936B a GW C (calculated).³⁷⁴³⁴

SpaceX flew 2.4 times as many launches as CASC, the next largest in BryceTech’s count, in 2025, and its SI segment billed 2.9 times CoreWeave’s revenue, 2.5 times on compute and Grok alone, comparing its quarter to September with CoreWeave’s quarter to June C (calculated).

IVManagement

Elon Musk founded SpaceX and has been chief executive since May 2002. He was 54 at the time of the prospectus and also runs Tesla, Neuralink and the Boring Company.³ He holds 6,068.5 million shares, 44.7% of the 13,581.1 million outstanding and 38.8% without his 1,302 million unvested performance shares. His Class B shares carry ten votes each, which gives him about 86% of the votes on my count of 15 October share counts and 30 June holdings C (calculated). The prospectus reported 82.4% at the IPO.¹⁰⁴³ He is the chairman. The certificate of formation requires a Class B majority, on top of board approval, to remove him as chairman, director or chief executive, and it has no sunset on the Class B shares.²⁵

The prospectus names three executive officers: Musk, Shotwell and Johnsen. It names no head of Connectivity, Space or SI, and the CEO is the chief operating decision maker for all three. SI is 37% of revenue and 81% of capital spending, and Connectivity earns all the operating income. That is a key-person risk with no named second layer.³⁴

Pay and incentives

Musk’s salary was $54,080 in 2025, unchanged since 2019, with no bonus and no stock grant that year.³ His incentive is two 2026 awards of restricted Class B shares, 1,302 million in all, 9.6% of the shares outstanding C (calculated). The first, 1,000 million shares, vests in 15 steps as market value rises from $0.5T to $7.5T, and also requires a permanent Mars colony of a million people. The second, 302 million, vests in 12 steps from $1.065T to $6.565T and requires data centers off Earth that supply 100 terawatts of compute a year.³

The shares vote now. At the cover-count market value of $1.8T, three steps of the first award and two of the second are passed on market value alone. No step vests until the Mars or orbital condition is also met, and neither has been C (calculated).³⁷³⁴ The company recognizes no expense for them until vesting is probable, and my cases carry none: the performance shares are in neither my share count nor my stock-pay cost, because the Mars and orbital-compute conditions are not met in any case. If they vest in full, my base value is $43.³

His wealth is the better incentive: each $1 on the share price changes his 6,068.5 million shares by $6.1B C (calculated). He also holds 350 million options struck at $8.40. His record includes a 2018 SEC settlement over statements about taking Tesla private, with a $20M penalty, and an April 2026 judgment against him personally over two May 2022 statements about the Twitter purchase. Post-trial motions were pending at the 12 June prospectus, and I did not find a ruling.³

His forecasts can be scored in part. The prospectus said Starship would start delivering payloads to orbit in the second half of 2026, and Flight 14 did so in the third quarter. The next tests are the $100B December ARR goal and the 13 GW outlook for the end of 2027, which is the outlook of a person whose pay depends on market value and is not guidance.³⁶²⁰

Gwynne Shotwell is president, chief operating officer and a director. She joined in 2002 and has been chief operating officer since 2008. Her 2025 pay was $85.8M, of which $83.0M was options, and on 22 September 2026 she sold 342,170 shares at $151.15 to $154.72, about $52M.³¹³ Bret Johnsen has been chief financial officer since 2011 and earlier worked at Broadcom. His four million performance options now pay on Adjusted EBITDA, 371,125 for each $10B across 2025 to 2029, and none were earned for 2025.³ No executive has an employment, severance or change-in-control agreement.³ Musk made no open-market sale after the listing in the filings I reviewed; his Form 4 of 17 June shows the conversion of Class B shares to Class A on 15 June.¹⁴

Read together, the pay rewards market value and not cash returned. Each $1 on the share price is worth $6.1B to Musk C (calculated), and the big awards also need a Mars colony or orbital compute. He owns 44.7% of the shares and, in the filings I reviewed, made no open-market sale after the listing.³¹⁴ On capital, 2023 to 2025 is mixed: Starlink worked, at 1.19 times its capital spending, and X and xAI lost money in every year shown.³⁴ That is three years, not a full cycle.

The board and the vote

The board has nine members. Class B elects five of them, including Musk, Shotwell and Antonio Gracias, who runs Valor Equity Partners. Six are called independent, but one is a Google executive while Alphabet holds 551 million shares, and another sits on Tesla’s board.³¹¹ Google is also the customer on the $920M-a-month cloud contract.⁴⁵ The company uses the controlled-company exemption, so its compensation committee is not fully independent. Directors receive no cash or stock pay.³⁸

Texas law limits what a minority can do. SpaceX is incorporated in Texas. A derivative suit needs 3% of the shares, about $53B at the IPO price. Disputes go to the Texas Business Court or to arbitration in Houston.³ Musk is chief executive of SpaceX and Tesla, and the board that could change that is the one he elects.³²⁵

Related parties

Valor, run by a SpaceX director, holds three failed sale-leasebacks of AI hardware that SpaceX records as debt: $12.9B at 30 September, with $864M of interest expense on them in nine months. SpaceX guarantees them. Tesla sold SpaceX $506M of Megapacks in 2025 and $480M in nine months of 2026, holds 19.0 million shares, and is the subject of a review by special committees of both boards. The terms are disclosed.³⁴

Capital allocation

The record shows in the table. Starlink is the build that worked: Connectivity grew from $3.9B of revenue in 2023 to $16.0B in the twelve months, and its operating income of $6.5B is 1.19 times its $5.5B of capital spending C (calculated).³⁴ X and xAI have lost money in every year shown. The AI segment lost $4.0B in 2023 after a $3.8B write-off, then $1.6B and $6.4B, and SpaceX took it in with 0.1433 shares for each xAI share and $2.9B in cash.³

Cursor was paid for at $42.5B of accounting consideration, and SpaceX will expense $23.4B of retention stock over 2.8 years, $66B in all, for a company with about $4.7B of annualized revenue. That is 9 times revenue on the consideration and 14 times with the retention stock C (calculated). I have no evidence yet on its margin. The filings do not name the Cursor leaders or their award terms beyond time-based vesting.⁴⁶²⁰

Capital spending rose from $4.4B in 2023 to $55.0B in the twelve months C (calculated).³⁴ Operating cash flow covered it only in 2023. Repurchases were $5.0B in the twelve months, bought from employees and holders at fair market value, none in Q3. There is no dividend.³⁴

Share supply

The IPO lock-up releases in steps. A release of 1.3 billion shares became eligible on 22 October, and the stock fell 9.1% in that session; the company does not link the two. The next 328 million shares became eligible on 26 October, and the stock closed that day at $146.90, 0.4% below the session before. Up to 798 million more become eligible on 8 December.³⁶ Musk’s 6.4 billion shares are free on 12 June 2027, and about 7.8 billion shares stay locked for more than a year.³¹

VFinancials

Quarterly revenue went from $4.7B in Q1 to $14.6B in Q3, 3.1 times, and in Q2 and Q3 SpaceX spent 3.2 times its operating cash flow on capital, $38.0B against $12.1B C (calculated).⁴⁶ Free cash flow was −$39.2B in the twelve months, against −$14.0B in 2025.

Revenue went from $10.4B in 2023 to $18.7B in 2025, 34.1% a year, and to $32.0B in the twelve months to September 2026.³⁴ Revenue less cost of revenue rose from 41.2% to 56.7% of revenue. Adjusted EBITDA, a company measure before stock pay, went from 36.8% to 46.6%. Operating margin was −13.9% in 2025 and −2.1% in the twelve months.³⁴⁶

The cash

Capital spending in the twelve months was $55.0B against $15.8B of operating cash flow. SI took $44.6B of it, Connectivity $5.5B and Space $4.8B C (calculated).³⁴ Management guides fourth-quarter capital spending to $35B to $40B, which would put 2026 at $83B to $88B in cash C (calculated), and a further $11.8B was in payables at 30 September.²⁰⁴

Earnings quality

Net income was a loss of −$2.8B in the twelve months. Q3 net income of $4.4B included a $561M gain on the Boring Company and a $390M tax benefit, and operating income of $2.8B included a quarter of cloud revenue at prices the CFO said will not repeat in full.⁴²⁰ Stock pay was $4.7B, 15% of revenue, before $23.4B of Cursor retention stock still to be expensed.⁴ I treat stock pay as a cost. The SEC’s XBRL data for revenue, operating cash flow and capital spending match the 10-Q tables.¹⁶

The balance sheet

Cash and securities were $90.2B at 30 September, against $39.0B of debt and finance leases, of which $12.9B is the Valor financings. Net cash was $51.2B before $11.8B of capital spending in payables.⁴ The cash includes $17.5B of customer advances, mostly for Space agreements and Connectivity enterprise and government contracts.⁴ The $25.0B of notes were valued at 91.7% of par on 30 September, a yield I calculate near 7.2% against coupons that average 5.855% C (calculated).⁴⁹ Goodwill and intangibles are $54.8B, 22% of assets, mostly Cursor and X.⁴ Unconditional purchase obligations are $37.8B, $25.6B of them in 2027.⁴

VIValuation

How it goes wrong

Working back from what would most likely wreck this, I see three things. First, the price of compute. AI compute is $31.0 of my $48, and if a rented gigawatt fades to $17.5B by year six my base value is $13 E (estimate).⁴³ Second, the money. The base case needs about $78B of new money,⁵² and lenders waived a technical default in March.⁴ Third, control. One shareholder has about 86% of the votes, with no sunset,³¹⁰²⁵ and both boards formed committees in October to review a combination in which he sits on both sides.⁴²⁰

What an owner can take out

Adjusted EBITDA was $14.9B in the twelve months, before stock pay.³⁴⁶ Free cash flow was −$39.2B³⁴ and stock pay $4.7B,⁴ so after both an owner is at −$43.9B C (calculated). Most of the $55.0B of capital spending is growth, and I cannot split it from standing still, so I give no company figure. Starlink alone made $6.5B of operating income on $5.5B of capital spending, about $1.0B before tax and stock pay C (calculated).³⁴ My base case pays an owner $29 a share to September 2036, about a fifth of the price E (estimate).

What I cannot know is what a rented gigawatt earns once today’s contracts end, so I show a range and not a point: $13 if it fades to $17.5B, $48 on my fade to $29B and $111 if $50B stays for ever E (estimate).

What the price implies

At $134.72 the market values SpaceX at $1,830B on the cover count of 13,581 million shares, which includes 1,302 million unvested performance shares held by Musk. Net of $51.2B of cash it is an enterprise value of $1,778B, or $1,603B without the performance shares.³⁷⁴ Nasdaq’s own quote page shows the same market value, $1,829.6B.³⁷

Value per share on this page uses 13,074 million shares: the 12,279 million outstanding without the unvested performance shares, plus the 533 million that the 10-Q adds for options, restricted stock units and the employee plan in Q3, plus the 262 million shares owed to EchoStar for spectrum.⁴ On that count the price values the equity at $1,761B, $68B below the cover-count figure. The cover count includes the performance shares, worth $175B at the price, and leaves out the 795 million shares from options, RSUs and EchoStar, worth $107B C (calculated).

The stock listed at $135.00, reached an intraday high of $225.64 in June, closed at a low of $108.27 in August and closed at $134.72 in Q4 2026, 0.2% below the IPO price and 40% below the high.^37371333 The stock fell 6.4% on 30 October in a market-wide sell-off over AI spending, with the Nasdaq-100 down 3.5% and the S&P 500 down 1.8%; SpaceX made no SEC filing during the session.^40414243 A second source, CNBC, shows the same close.³⁸

There are no earnings to start from, and free cash flow is negative, so the price has to be read against what has to be built.

The enterprise value of $1,778B is 12.7 times the Q1 2027 revenue guide at an annual rate, $140B (13.1 times on market value), so the price already assumes the guide is met and kept.

What a gigawatt earns

A gigawatt of nameplate compute cost SpaceX about $40.6B to build: SI capital spending was $44.6B in the twelve months while nameplate compute rose by 1.1 GW C (calculated).³⁴ Over the last six months the figure is higher, $46.1B, on $32.3B of SI capital spending in Q2 and Q3 for 0.7 GW C (calculated). Capital spending leads installation, so both readings are noisy. I use $40.6B and test $46.1B and $35B.

Only rented gigawatts bill a customer. Management’s goal is to use about half the compute internally and rent the other half, and it could move to 60-40 either way.²⁰ So there are two prices. One is revenue per gigawatt of the whole fleet, which is what the quarterly results show. The other is the price per rented gigawatt, which is what a contract sets. In Q3 the fleet earned $16.6B a gigawatt. Customer B paid $5.0B of the $6.4B of compute and Grok revenue, so everyone else, Grok included, paid at most $1.4B.⁴⁶

The evidence on the price of a rented gigawatt points several ways.

The prospectus gives the market’s sizing as $17.5B per rented gigawatt: a $2.4T market on 235 GW of data-center demand in 2030, from RAND, 70% of it AI, at the 2025 median neocloud rental rate of $3.33 a GPU-hour. That demand is facility power, so I divide by its efficiency factor of 1.2 to get chip power; leaving that step out gives $14.6B and understates the market.³ Anthropic’s May contract, as the prospectus gave it, pays $15.0B a year for about 325,000 GPUs, or $46,154 a GPU-year, $5.27 an hour. The prospectus gives the power per GPU in the Colossus clusters, 1.3 kW for H100 and 1.9 to 2.0 kW for Blackwell, so the contract covers 0.42 to 0.65 GW, which is $23.1B to $35.5B a GW-year C (calculated).³ Those terms have since been amended, and the S-1 gives no GPU count, so I cannot redo the sum.⁴⁴ Google’s June contract works out at about $50B on Blackwell, as section VII shows.⁴⁵

The CFO said pricing was above the top of a $30 to $50 per watt range and that new compute pays back in less than a year. The transcript does not state the unit of the pricing. If it means a watt-year, that is above $50B per gigawatt-year, close to what the Q1 guide implies at half rented and to Google’s contract on Blackwell. I cannot reproduce the payback from the filings: under a year at $40.6B a GW and a 60% EBITDA margin needs $68B of revenue per gigawatt-year across the fleet, four times the Q3 reading C (calculated). I use neither as a price, because the unit is unstated and the source is a third-party transcript.²⁰

The table prices one new gigawatt, bought today, with chips replaced every six years. At my 10.5% cost of capital it breaks even near $16.3B of revenue per gigawatt of the whole fleet. At a 50-50 split, with the internal half earning $8B, a rented gigawatt must earn $24.6B for the fleet to break even, and the low end of Anthropic’s May contract, $23.1B, is below that. The prospectus’s sizing of $17.5B gives a fleet that is half rented about $12.8B a gigawatt, which is below break-even. The Q3 reading earns 10.9% a year on cost and $27B earns 25.7% E (estimate).³⁴

For a mature fleet replaced evenly, which my terminal value assumes, break-even is near $19.7B. My base case’s long-run $18.5B sits just below that, so growth after year fifteen adds nothing, and the value is made in the early years.³⁴ That is why the question is how long a high price lasts.

Year one: the guide and the cover

Year one runs from October 2026 to September 2027. My base case has $141B of revenue in it. The table shows how the four quarters get there.⁴⁶²⁰

The first two quarters follow the company: about $20B from the December goal and $35B from the Q1 guide. The last two, $86B or 61% of the year, have no guidance and rest on installed compute going from 2.8 GW to 6.0 GW. The Q1 guide asks for $31B per average gigawatt, 89% above the Q3 reading. At the Q3 rate it would need 6.3 GW on average in the quarter, against 3.3 GW in my path and 2.5 to 3 GW guided for year-end. The guide comes from a third-party transcript and not from an SEC document.⁴⁶²⁰

Backlog at 30 September covers little of it. Backlog due within a year is about $42B across all segments, against $104B of SI compute and Grok revenue in my year one, so at most 41% is covered and at least $62B is not.⁴ The two named contracts would run at about $48B a year at full rate, 46% of the $104B C (calculated).⁴⁴⁴⁵ Either side can end them on 90 days’ notice once the locked periods pass, so that cover is thin too.

A year-one miss matters less than what follows it. If year one comes in 20% low and the price recovers, my base value falls to $47. If revenue per gigawatt never leaves the Q3 reading, it falls to $29 E (estimate).⁴⁶

Reverse DCF

Free cash flow is negative, so I cannot solve for a growth rate in it. I solve instead for the price of compute. Hold my base-case build of 22.7 GW at year ten, half of it rented, Starlink and Space as they are, and ask what each rented gigawatt must earn from year three to support $134.72. The answer is $57.5B a rented gigawatt-year, indefinitely. That is $32.7B per gigawatt of the whole fleet, and a return of 33.1% a year on each new GW’s $40.6B, against 10.5% for my cost of capital C (calculated).³⁷⁴

Scale is the other lever, and the grid shows both. At prices near the contract midpoint, more gigawatts add little: at $29B a rented gigawatt, 22.7 GW is worth $44 a share and 40 GW $52. At $35B, 22.7 GW is worth $63 and 40 GW $86. The price needs about 76 GW at $35B a rented gigawatt-year, 47 GW at $41B, 30 GW at $50B, or 23 GW at $57.5B E (estimate).³⁴

Musk’s own outlook sits above the grid. He told the Q3 call he would be surprised by fewer than 13 GW by the end of 2027, and that this implies annual recurring revenue of $500B to $1T by then.²⁰ On 13 GW that is $38B to $77B per gigawatt of the whole fleet, above any reading in the filings C (calculated). It is the outlook of the person whose pay depends on market value, and it is not guidance.

The discount rate

I use 10.5%. A beta cannot be measured from 93 trading days. A daily beta against the S&P 500 fund over the period is 3.3, and against the Nasdaq-100 1.8, on an R-squared of 0.22, so I use neither.¹ I assume a beta of 1.3, which on a 5.00% ten-year Treasury yield and Damodaran’s 4.20% equity risk premium gives 10.5% C (calculated).¹⁷¹⁸ The 1.3 is my assumption. Damodaran measured his 4.20% against a risk-free rate of 5.29% on 1 October; on that rate the cost of equity is 10.75% and my base value $45.¹⁸

The notes give a reference that does not depend on beta. They yield about 7.2%, and an equity holder should earn 4 to 5 points more than a bondholder in the same company, which is 11.2% to 12.2% E (estimate).⁴⁹ So the range is 10.5% to 12.0%, and I show both: the base value is $48 at 10.5% and $36 at 12%. I use one rate for all three engines. Starlink at 9.5% and SI at 11.5% give $44.

Three cases

I build each case from the engines up: Starlink and Space by revenue, margin and capital spending as a share of revenue; SI by gigawatts, the share rented, the price per rented gigawatt-year, what an internal gigawatt earns, and a margin. Free cash flow is adjusted EBITDA less stock pay, cash tax at 23% after $9.7B of federal loss carryforwards and $3.6B of credits, capital spending, and $9.7B of spectrum cash. That is the up-to-$8.5B EchoStar debt payoff plus $0.4B of 2026 and $0.8B of 2027 interest-bridge payments still to be made. It leaves out the 800 MHz purchase, whose price is not disclosed.³⁴ Each GW added costs $40.6B, 80% of it chips that I replace at what they depreciate from year 4, phased in by year 7. Year one runs from October 2026 to September 2027, and values are dated 30 September, the last balance sheet; dating from Q4 2026 adds $0.3 E (estimate).

The base case takes the company at its word for a year and then lets prices fade.

• Year one revenue is $141B, the quarters in the year-one table, with 6 GW installed at the end, one a quarter from fourth-quarter capital spending of $35B to $40B E (estimate).²⁰⁴
• Half the compute is rented, Musk’s stated goal. The price per rented GW is $50B in year one, which is what my $141B of year-one revenue implies at that split (the Q1 guide alone implies $59B), and fades in a straight line to $29B, about the middle of the range on Anthropic’s May contract as first filed, by year six. The May terms have since been amended, so I cannot redo that anchor, and Google’s June contract pays about $50B a rented gigawatt-year on Blackwell chips. I keep the fade because both contracts can end on 90 days’ notice once their locked periods pass E (estimate).⁴⁴⁴⁵ Each internal GW earns $4B of Grok and other revenue, rising to $8B; the Q3 ceiling, at most $1.4B a quarter over 0.78 GW, is $7B. Blended, that is $27B a gigawatt falling to $18.5B, 13% above break-even for a new GW and 6% below it for a mature fleet. Cursor and X grow from $8B to $24B in year ten E (estimate).
• Installed compute reaches 18 GW in year five and 22.7 GW in year ten, against 13 GW by the end of 2027 in Musk’s outlook E (estimate).²⁰ In September 2030 the base case has 16 GW, 12% of the 137 GW of nameplate AI demand for 2030 that the prospectus takes from RAND, an estimate for the whole world.³
• SI’s EBITDA margin falls from 64% to 60%, against 65.5% in Q3 and 45% in Q2 C (calculated).⁶
• Connectivity grows 53% in year one, 34% in year two and 7.5% in year nine, to $90B in year ten at a 56.5% EBITDA margin E (estimate).
• Space earns a 30% EBITDA margin by year eight as Starship launches for others, and Starship’s research cost keeps its capital spending at 100% of revenue in year one and 20% by year nine E (estimate).
• Free cash flow is −$100B in year one and −$68B in year two, turns positive in year three, and reaches $90B in year ten E (estimate).

More than half of the base value, 55%, arrives after year 15. Year-15 revenue of $635B is 4.3 times AWS’s $148B of the last four quarters, and in year ten the base case runs 22.7 GW, 17% of the 2030 demand figure. If share is to hold at that level the market has to grow past the prospectus’s 2030 sizing E (estimate).³³²

My base value for SI, $31 a share, has three parts. The 1.7 GW already built is worth $9, Cursor and X are worth $7, and the build to 22.7 GW is worth $14, about $9B for each GW added, a fifth of its cost. $7 of the total is the price above $29B in years one to five E (estimate).³⁴

My base value for Starlink is $187B, 11.7 times its trailing revenue and 18 times its Adjusted EBITDA. Charter, the largest broadband operator in my peer set, trades at 2.0 times revenue and 4.9 times EBITDA on revenue that fell 2%. Iridium trades at 7.5 times revenue on 8% growth and Viasat at 3.3 on a 1% decline. Starlink’s connectivity revenue grew 57% in nine months, and the base case pays for that growth C (calculated).^39394282935

The commodity-compute case is the bear. It is a stop-build case, because a rational operator whose customers reprice on 90 days’ notice stops building. Customer B leaves or reprices, the price per rented gigawatt falls to the prospectus’s sizing of $17.5B, and SI freezes at 3 GW, the top of the 2.5 to 3 GW the company guides for year-end 2026, which it has already paid for. The margin falls to 47%. Year-one revenue is $85B. Starlink still grows to $63B of revenue by year ten, and Starship slips. SI is worth $3 a share, and Starlink and cash carry the $12 E (estimate).³⁴ If SpaceX kept building to 6.5 GW into those prices, the value would be $6, which is why I treat stopping as the case.

The bull case has Musk’s plan arrive on time. Installed compute is 10.5 GW by September 2027 and 17 GW a year later, in line with his 13 GW at the end of 2027, and reaches 50 GW in year ten. Sixty percent is rented, at $48B per rented gigawatt in year one, a little under the base case’s $50B because 60% rented gives a higher blended price, $30B against $27B, falling to $35.5B, the top of the May contract’s range as first filed, and Starlink reaches $134B in year ten. Free cash flow in year ten is $327B E (estimate).²⁰ The bull case does not vest the performance shares: its 50 GW is on Earth, and the awards need a Mars colony and 100 TW of orbital compute.³

I weight the cases 25%, 50% and 25%, which gives $76 a share against a price of $134.72. The bear gets 25% because one customer paid 78% of compute and Grok revenue in Q3 and the named cloud contracts can end on 90 days’ notice once their locked periods pass. The base gets 50% because it takes the company’s guide for a year and a fade through the May contract’s range, as first filed, for the price after. The bull gets 25% because it adds 8.8 GW in year one, eight times the 1.1 GW added in the last twelve months E (estimate).⁴²⁰ The weighted cases return 8.0% a year at the price.

Where I differ from consensus

Consensus EPS is $2.46 for 2027, $7.91 for 2028 and $13.21 for 2029, from 11, 8 and 5 estimates. The 2029 range runs from $5.44 to $26.01.¹⁹ At 13,074 million shares, $13.21 is $173B of net income in 2029. My base case earns $84B in year 3, a calendar-adjusted $6.69 a share, 51% of consensus, and $5.22 for 2028, 66% E (estimate). Nasdaq does not publish consensus revenue, so I cannot back out the revenue or margin the sell side assumes.

I can say which of my assumptions explains the gap. Giving my base case Musk’s build, 24 GW in year 3 instead of 13.5, lifts calendar-2029 EPS from $6.69 to $11.01. Giving it the bull’s price and rented share lifts it to $8.66. Installed gigawatts explain 2.2 times as much of the gap as price. On Musk’s build my 2028 EPS is $8.57, against $7.91 for consensus, so the sell side looks to me like it is paying for the build.¹⁹

What the price pays for

Of the $134.72, my base case supports $14.3 for Starlink, $31.0 for SI, and $2.4 for Space and net cash, which is $48 with rounding. The other $87 a share, $1.1T, is for what my base case does not contain: more gigawatts, higher prices for longer, or the options. The prospectus plans orbital compute satellites from 2028, a chip factory with Tesla and Intel, lunar contracts and Mars, and none has a committed budget or a revenue line today.³⁶ To carry the other $87 they would have to be worth $1.1T together.

What closes the gap

No single change takes my base value to the price E (estimate).³⁷ The two that matter most are the build and what each rented gigawatt earns. Holding $50B per rented gigawatt for ever takes the base value to $111. Musk’s build through 2027, at my prices, takes it to $58. Both together give $196, above the price. The share rented is the third lever: with every gigawatt rented at my prices the base value is $127, and with 40% rented $32.

Smaller choices move the base value by less. Replacing each batch of chips six years after it is bought, instead of phasing in replacement at what the fleet depreciates, takes it to $54. Charging full replacement from year one takes it to $35. I use the phased schedule, which sits between them. Growth after year 15 of 4.0% gives $52. Deducting all $17.5B of customer advances from net cash gives $46 E (estimate).³⁴

The funding gap

The base case spends $168B more than it earns before it turns positive in year three, and the bull case $293B. Cash and securities are $90.2B, and a $5B revolver is undrawn. Debt maturities through 2030 are $12.8B, and $37.8B of purchase obligations come first. The revolver lenders waived a technical default on 2 March 2026 that arose when SpaceX took on xAI’s debt.⁴ My value assumes the money is raised on fair terms. The base case needs about $78B of new money beyond today’s cash, and the bull case $203B E (estimate). On 24 October SpaceX signed a commitment for a finance lease of about $40B on GPU equipment with an Apollo-managed entity, expected to close before year-end. If it closes it covers about half of the base case’s $78B, as lease debt with fixed payments.⁵²

The base case needs $78B of new money and the bull case $203B, against the $111B that the IPO and the first notes raised in 2026 C (calculated).

The margin of safety

• Price$134.72Close, Q4 2026 [37]
• Base value$48My estimate E (estimate)
• Margin of safety below$3135% under my base value C (calculated)

My rule is 25% below the base value for a business with net cash and no customer above a tenth of revenue, 30% when one cycle or two customers drive results, and 35% when fixed charges such as leases sit ahead of the shares. SpaceX meets the second and third tests. Two customers paid 51% of Q3 revenue, the cloud contracts end on 90 days’ notice, and $12.9B of related-party lease financing and $25.0B of notes rank ahead of the shares.⁴³ I use 35%. That gives $31, 23% of the price. At that price my base case returns 12.9% a year, my bull case 23.9%, the commodity-compute case 5.2%, and the weighted cases 16.2% E (estimate).

The margin-of-safety price covers a 35% error in my base value. It does not cover the commodity-compute case, which is worth $12, 39% of the margin-of-safety price. That case is the one in which compute reprices to the prospectus’s own long-run sizing, and it leaves Starlink growing. I anchor the margin on the base case and not on the weighted value of $76, because the weights are my judgment and the weighted value is a mean of three cases, not a case I can build. The price sits 335% above the margin-of-safety price.³⁷⁴

Where this could be in 5 and 10 years

My base value is $84 a share at September 2031 and $101 at September 2036, against $48 today E (estimate). It rises because the cash burn is behind it: the base case spends $168B more than it earns in the first two years, then reaches 18 GW and $64B of free cash flow at year 5, and 22.7 GW and $90B at year 10. The drivers are the price of a rented gigawatt, fading from $50B to $29B, the margin on SI revenue, 64% falling to 60%, and the 10.5% discount rate. At year 10, 71% of the value is the value after year 15.

The bear stops building at 3 GW with the price falling to $17.5B, so it stays near $12 to $13. The bull builds 50 GW at a price that only fades to $35.5B, which gets it to $417.

The most likely way the base is wrong is the long-run price of a rented gigawatt. Google pays about $50B a year on Blackwell chips until June 2029, and either side can end it after 31 December 2026.⁴⁵ A $50B price that stops with the contract changes neither the year-5 nor the year-10 value. If $50B were the price from year 6 on, the base would be about $93 higher at year 5 ($177) and $109 higher at year 10 ($210) E (estimate). Each $1B on the long-run price is worth about $4.4 at year 5 and $5.2 at year 10. Case values are unchanged.

These are estimates of value, not forecasts of the share price, and the price can sit above or below value for years. The 6.4% a year above holds the cash flows for ever; this return sells at my year-10 value, so it is lower.

The same dollar, and the odds

My base case returns −0.4% a year from $134.72 to September 2036 E (estimate), against 5.00% on a ten-year Treasury that takes no business risk.¹⁷ The closest comparable on this site is Rocket Lab, at −8.0% a year with its price about 465% above its base value, against about 180% for SpaceX. The business and the price are two questions: Starlink and launch are strong, but $87 of the $134.72 is for what my base case does not contain.

The price also needs things to go right together. My $136 needs $50B a rented gigawatt kept for ever and 30 GW built, about $1.1T. Say 30% for the first and 40% for the second. Multiplied, that is 12%, about one in eight E (estimate), and since a high price pays for the build they are linked, so call it one in six.

VIIThe case for the price

What most people miss

Google agreed on 5 June to pay $920M a month from October 2026 to June 2029 for about 110,000 GPUs.⁴⁵ On the power per GPU in the prospectus, that is about $50B a rented gigawatt-year on Blackwell chips and $77B on the older H100 C (calculated).³ The filing does not name the chip. Even the low end is well above the $23.1B to $35.5B on Anthropic’s first contract, and 87% of the $57.5B the price needs.

The second is in Anthropic’s S-1, filed after SpaceX’s 10-Q. Anthropic agreed to spend up to $120.5B with SpaceX into 2029, about $3.1B a month once all capacity is in, expected in December.⁴⁴ That is 2.5 times the $1.25B a month in the prospectus.³ The total is far above SpaceX’s whole backlog of $63.0B.⁴ I read that gap as cancel rights and October amendments E (estimate). The S-1 gives no gigawatts, so I cannot price it.

The bull case

Gavin Baker of Atreides laid out the case with the CFO in an interview SpaceX filed on 9 June. Starship cuts launch cost and takes Starlink from 10 million to hundreds of millions. Of the Anthropic deal he said “you have a top five AI infrastructure business by my math.”⁴⁶ This was IPO marketing. His firm reported $4.67B at 30 June, 33% of its 13F.⁴⁷

Brad Gerstner of Altimeter told CNBC on listing day that his firm took part in the IPO. He expects it to become the largest AI hyperscaler in the United States within a few years, on a third-party transcript.⁵³ Altimeter reported $303.5M of it at 30 June.⁵⁴

The ARK Invest and Mach33 model of 10 June 2025 puts 2030 enterprise value at $2,508B.⁴⁸ It has no AI line, and ARK says the file will not be updated. At my 10.5%, my sum and not ARK’s, $2,508B is about $1,650B today, against an enterprise value of $1,778B C (calculated).⁴ Cathie Wood told Fox Business in July it could become “the most important company in history.”⁵⁵ ARK’s flagship fund had 6.15% in it on 30 October.⁴⁹

Bill Ackman likes the business, not the price. He told Money News Network on 20 July that Starlink is a “near monopoly” and that price is the question. Pershing Square does not own it, he said, for lack of the “degree of predictability” he wants.⁵⁶⁵⁷

Musk told the Q3 call that short contracts are renewing “at higher dollars per watt.”²⁰ Finbold reported in October that Goldman’s Eric Sheridan raised his target to $230 and Morgan Stanley’s Adam Jonas kept his at $300.⁵⁰⁵¹ Both banks led the IPO.³

Against reality

If SpaceX keeps about $50B a rented gigawatt every year and reaches 30 GW by September 2036, half of it rented, my grid gives $136, so the price is possible E (estimate). On my 22.7 GW it takes $57.5B. Musk’s build at $50B gives $196 E (estimate).³⁷²⁰

The whole fleet of 1.7 GW earned $16.6B a gigawatt in Q3, against the 30 GW and $29B those prices need C (calculated).⁴⁶ The two named contracts come to $48B a year, 46% of the compute and Grok revenue in my year one, and both can end on 90 days’ notice once their locked periods pass C (calculated).⁴⁴⁴⁵ Building to 30 GW costs about $1.1T at $40.6B a gigawatt, before new chips C (calculated). Apollo’s $40B GPU lease, if it closes, covers about half of the $78B of new money my base case needs.⁵²

Where I come down

The Google contract moved me: I now treat about $50B as a price someone signed. It did not move my values, because both contracts can end on 90 days’ notice, so I keep the fade to $29B and my $12, $48 and $198. I would raise them on revenue above $31B a gigawatt in Q1 2027, more than 4.5 GW installed at March, and minimum terms of three years.

VIIIRisks, and what would change my view

1. The price of compute. Everything above turns on it. If the price per rented gigawatt fades to the prospectus’s $17.5B by year six instead of $29B, my base value is $13 E (estimate). Customer B was 34.5% of Q3 revenue and the contracts end on 90 days’ notice.⁴³
2. Funding. The plan needs $78B of new money in my base case, and the Apollo lease of about $40B, if it closes, would cover about half.⁵² The notes trade at 91.7% of par, Q4 capital spending is guided to $35B to $40B, and spectrum closing needs up to $8.5B of cash.⁴²⁰
3. Starship. V3 satellites, mobile service and orbital compute all depend on it. The first revenue flight was in Q3, Flight 15 is due in November, and weekly flights are a first-half 2027 aim.⁶²⁰
4. Control and key person. Musk holds about 86% of the votes, elects a majority of the board and can be removed only with a Class B majority. He is chief executive of SpaceX and Tesla. Texas sets a 3% threshold for derivative suits. The Class B shares have no sunset in the certificate of formation.³¹⁰²⁵
5. Tesla. Both boards formed special committees in October to review a combination. No terms exist. A deal would be a conflicted transaction in which Musk sits on both sides, and I assume none in any case.⁴²⁰
6. Supply of shares. 1.3 billion shares became eligible on 22 October, 328 million on 26 October with no break in the price³⁶, and up to 798 million on 8 December, with Musk’s 6.4 billion on 12 June 2027. 1,302 million performance shares and 262 million EchoStar shares add up to 12% of today’s count C (calculated).³⁴
7. Legal. Grok image suits, a European fine of EUR 120M on X, a Clean Air Act suit over the Colossus II turbines, and a construction dispute with claims of more than $500M. SpaceX has accrued $350M.⁴
8. Accounting. Goodwill and intangibles are 22% of assets, the auditor’s critical audit matter is cost to complete on fixed-price contracts, and a valuation allowance on U.S. deferred tax assets could release for a one-time gain.³⁴
9. What I could not verify. The call guidance comes from a third-party transcript and a second write-up, because the official transcript was not retrievable. Customers A and B are unnamed. Credit ratings and the 800 MHz price are not in the filings I read. I could not test the company’s build-speed and model-quality claims, or benchmark the related-party terms.²⁰²¹ Anthropic’s S-1 is a registration statement, and it can be amended.⁴⁴
10. Regulation and politics. Launch needs FAA licences, Starlink needs FCC spectrum and international landing rights, and SpaceX holds direct U.S. government contracts. A change in any of them is a revenue event.³

The table is the list I will check each quarter. I move my base value on four figures: revenue per gigawatt, installed compute, the share rented, and the size of Customer B.⁴⁶²⁰