SPCX

As of 2026-10-20 · partly unblinded, three revisions unblinded (the revisions were written after we saw the price) · blind commit 2fb2e247b · market pass, adversarial-review revision, Q3-call update and call-step review after it

We rate SpaceX Avoid at $147.79. Our central estimate of what a share is worth today is $81.61, 44.8% below the price. The bear case (a severe but plausible bad outcome) is $36.69 and the bull case (the matching good outcome) $243.67. Avoid means fair value is under 0.85 times the price, a line of $125.62; ours is well inside it, and only the bull case is above the price. The value rests mainly on one number: how fast the rent that new AI computing capacity earns falls each year. We assume 20% a year; the price needs 10.0%.

The central value was revised three times on 2026-10-20. The blind central, committed before we saw the price, was $21.38. An adversarial review (a reviewer sent to attack the work) moved it to $63.57. The Q3 call, held later that day after the release and after our $147.79 price, moved it to $121.95. A second review of that call step then found the exit capitalised a rent that was still falling and that capital per gigawatt had not been refreshed for the guided build; the central is now $81.61 (bear $36.69, bull $243.67). revision.txt bridges every dollar from $21.38. The largest change from the blind value is the rent on new capacity together with the guided build, +$66.6, and it rests on management statements made after the price.

Terms used below. Each report values a company three ways: the central case is our best estimate, the bear case a severe but plausible bad outcome, and the bull case the matching good outcome. SI is the Super Intelligence segment (the former xAI and X), which also rents computing capacity to other AI companies. A GW is a gigawatt of computing capacity; "rent" is the revenue a gigawatt earns in a year, in dollars per GW-year, and "new-capacity rent" is that figure for capacity built from Q4 2026 on. EBIT is operating profit and NOPAT is operating profit after tax. WACC is the weighted average cost of capital (the blended cost of debt and equity). The exit price is our estimate of the share price on 2031-12-31, the mean of two methods: the long fade, which lets growth and the return on new capital fall gradually toward normal levels over 15 years, and the outside view, which applies what comparable companies trade at. "The Street" means the published estimates and price targets of Wall Street analysts. The purchase price is the price at which a buyer would earn a 12% yearly return (the hurdle) on our central case. The report checker (scripts/research-check.py) raises numbered flags when a report shows a pattern that needs an answer; the flag number is given in brackets next to what it means.

"What the Q3 call changed" labels each call figure as guidance from after the price and says which were taken, discounted or left out; "What the call-step review changed" answers the second review. The work was written by an analyst who had seen the price, the reviewers' numbers and the call, so it is unblinded: each change rests on a dated source, listed in assumptions.md and data/q3-call-2026-10-20.txt.

Partly unblinded at the start: the IPO price is derivable from the 10-Q, and the masked S-1/A showed the dilution line and the AI CEO Award's grant value (assumptions.md lists each; none enters the value). Two pre-price documents the blind pass missed are now in: Google's $920 million a month agreement (FWP, 2026-06-05) and a $1.11 billion a month deal from December 1 (CFO, Goldman Sachs conference 2026-09-10). The market pass (the step after the blind commit that reads the price, the Street and the odds: market.py, market.txt, odds.json) changed no blind input.

Conflict of interest, with dollar effects. This report was written by Claude, a model made by Anthropic. SpaceX owns xAI (merged 2026-02-02), whose Grok models compete with Anthropic's Claude. Anthropic is also SpaceX's main SI customer: in May 2026 it agreed to pay $1.25 billion a month through May 2029 for compute at COLOSSUS and COLOSSUS II, an agreement either side may end on 90 days' notice (S-1/A, 2026-06-03), $15.0B a year of the roughly $39B a year of announced compute contracts (Google $11.0B, the December deal $13.3B, customer unnamed). That Anthropic is the unnamed customer at 19.5% of Q2 2026 revenue is our inference. A buyer's interest is a lower rent. The first review found the blind choices on SI that lacked measured support all leaned one way, low: cash costs held at 36% of revenue (+$11.19 once corrected), capital per gigawatt from two quarters that ran ahead of nameplate (+$28.03) and the Anthropic fee decaying as if it were spot (+$0.08). After the price, every revision moved the value up: $21 to $63 to $122. That run is a reason to doubt the revisions whichever way the writer's bias runs, and the two places a choice can hide are the exit-year rent and the pairing of the rent path with the build. The call-step review tested both, and both moved the value down: the exit-year pull to the rent that earns the WACC (finding R1 below) and the refreshed capital per gigawatt (finding R2) took the central from $121.95 to $81.61. The remaining choices with no dated support are the 20% yearly fall of the new-capacity rent (-$6.87 at -30%, +$66.58 at -10%), the $42B level and the $120B-a-year build, each answered below.

The first review raised the central from $21.38 to $63.57

Labels in the tables: F1 to F15 are the first review's findings, M1 is a method change, C1 to C5 are the steps of the Q3-call update, and R1 to R8 are the findings of the second review (of the call step).

The review's dollar effects below are on the model as it stood before the Q3 call (central $63.57, capital per GW $34.6B falling 5% a year); the call and the call-step review follow. The central moved from $21.38 to $63.57 (bear $12.88 to $12.83, bull $79.70 to $191.63, case-weighted $33.84 to $82.90). The bridge, from revision.txt (each step on top of the ones above):

Two changes carry almost all of it, and the range around them is wide. Capital per GW is the input the data fix least: $27.7B (servers, networking and data-centre infrastructure gross over nameplate at 2026-06-30, 10-Q Note 5, which leaves out construction in progress), $34.6B (cumulative SI spending of $58.8B over 1.7 GW, which includes it) or $46.1B (Q2 plus Q3 spending over 0.7 GW added, when construction in progress rose from $4.6B to $12.6B in H1). We use the middle measure; the central is $48.33 at $40B and $35.49 at $46.1B (the reviewer's $45 sits near these). The fall from 34.6 to 26.8 by 2031 at 5% a year is our assumption and is the same size as the rent fall. Cash costs: Q1, Q2 and Q3 2026 cash costs were $1.43B, $1.42B and $2.61B on revenue of $0.82B, $2.56B and $7.56B (S-1/A, Q2 and Q3 releases): the ratio fell from 55% to 34.5% because most of the base is fixed, not because Q3 was a lucky quarter. A flat 36% ratio would still give $46.97.

The Q3 call raised it to $121.95, on management statements made after the price

The call was held 2026-10-20 at 13:30 PT, after the 8-K and after the $147.79 price. Every figure in this section is management guidance or press report from after the price. The primary source for the call is the Investing.com transcript (in.investing.com, "SpaceX tops Q3 2026 estimates, shares jump after hours"), which describes itself as AI-assisted. We read it through a summarising fetch, not the raw page, so the figures below are as that summary gave them and the wording of any quote is unchecked against the recording; the transcript names the CFO "Brad Coffman" and sometimes "Bret", while the filings name Bret Johnsen, so its speaker labels are not reliable. Earlier in the day we had only Dow Jones headlines, the market data live blog and search summaries; they agree with the transcript on every number below. Saved in data/q3-call-transcript-2026-10-20.txt and data/q3-call-2026-10-20.txt. The contract terms and the two contracts the blind pass missed are from primary documents or a transcript of a conference before the price.

What we found that predates the price. Google agreed on 2026-06-05 to pay $920 million a month from October 2026 through June 2029 for about 110,000 Nvidia GPUs, with capacity ramping through September at a reduced fee; either party may end it on 90 days' notice after 2026-12-31 (FWP 2026-06-05, data/filings/fwp-2026-06-05-google.html). The CFO said on 2026-09-10 that a new hosting deal starts December 1 at about $1.11 billion a month, on "90-day with a 90-day out, so roughly six-month commits" (Investing.com transcript of the Goldman Sachs conference). Announced compute contracts are now about $39B a year (Anthropic $15.0B, Google $11.0B, December $13.3B), every one with a 90-day exit. The blind model had Anthropic only. Adding the two to Q4 2026 is worth +$0.04 (C1 below); their weight is in what they say about price per unit.

Per watt, and the level we use. The CFO's words: pricing is "currently tracking slightly above the high end of the $30 to $50 per watt range"; the range came from Mr. Musk on the Q2 call ("my guess would be somewhere between $30 and $50" for "monetization per watt"). We read it as annual revenue per watt of nameplate, so new capacity earns $50B or more per gigawatt-year against the legacy fleet's $18.1B. Four checks agree and we do not model the other reading: payback inside a year on a $37.3B gigawatt needs about $50B of revenue; the CFO's $100B of annualised revenue on the month of December over about 2.5 GW is about $40B per GW-year; Google's $920M a month on 110,000 GPUs (FWP 2026-06-05) is about $100,000 a GPU-year, which at 1.4 to 2 kW a GPU (our estimate) is $50 to $70 a watt-year; and the Q1 2027 guide of $33B-$37B needs about $48B on roughly 1.4 GW of new capacity once the legacy fleet and the rest of the company are counted. What is uncertain is the level and whether it lasts, not the unit. Our central is $42B in 2027. It is a judgement: the December run-rate is $40B, and the payback figure is $50B; we sit above the run-rate because Musk says renewals are higher, and below the payback figure because the contracts are 90-day and a seller states the best price. The bracket in the odds is $25B to $55B. For and against the level: Morgan Stanley (sell-side) has weighted pricing at $31 per watt in Q4 2026 falling to $17 in Q4 2027 (the case for a lower level and a faster fall); Ornn's spot index is rising (H100 +17.7% and B200 +14.5% in 30 days; one snapshot).

Testing the model against the guide. Before the call, our 2027 revenue was $87.1B in the central and $96.2B in the bull; a Q1 2027 guide of $33B-$37B annualises to $132B-$148B, 1.5 to 1.7 times that. The guide justifies a different shape, not a bigger version of the old one: the legacy 1.7 GW at about $17B, and new capacity at $42B. The revised central 2027 revenue is $164.7B (bull $203.5B, bear $144.9B), more than the Q1 guide annualised because the build roughly triples capacity during 2027; the Q1 test above is the one that can be checked in a quarter. Customers must fund $133B of 2027 SI revenue against $28B annualised in Q3; the backlog is $63B, about $42B of it within a year. That funding is untested: no signed contract or customer balance sheet we hold supports the remainder.

What the guide does not settle, and the contract structure's due. A 90-day contract with a 90-day exit lets a buyer reprice every six months. Supply is coming (Musk's 10 GW and 20 GW targets are for SpaceX alone). The bear therefore gives the new-capacity rent a 30% yearly fall from $40B (2031: $9.6B, under the $20.2B that earns the WACC at the bear's costs) and loses Anthropic at the end of 2028. The central falls 20% a year to $17.2B in 2031 (just above the $16.6B that earns the WACC); the bull falls 12% to $30.0B. The spread between them is the report: $74.74 at -30%, $81.61 at -20%, $148.19 at -10% (the exit-year pull to the normal rent fades away as the rent holds, below).

A second review of the call step cut the central to $81.61

A review of the Q3-call revision (central $121.95) found eight things. The bridge from $121.95, from revision.txt:

The largest change from the blind value (the checker's large-move flag, flag 1) is the pair C3 and C4, +$66.6. After the call-step review, the largest single move from $121.95 is R2b, -$21.84.

Rent and spend paired (market.txt; central value; 2031 marginal-gigawatt return):

A is coherent by the rule that the build stops where a new gigawatt only earns its cost of capital. B is coherent in the other direction (rent that holds invites more building, 24.7% return) and is the market's world: the price needs a fall of 10.0%, and B2 at that fall is $148.19, the price. C is the sell-side's world, with the build left running into a rent below cost, and C2 is its coherent version.

What SpaceX sells: Starlink makes the money and Super Intelligence is the bet

SpaceX reports three segments, and their segment operating income adds up to the company's; there is no corporate line (Q3 2026 release).

• Connectivity (Starlink) sells satellite broadband. 13.6M consumer subscribers paid $69 a month on average in Q3 2026; enterprise, government (Starshield) and mobile customers paid $2.13B in the quarter, up 91% in a year. Revenue was $5.25B in Q3 at a 45.1% operating margin, after depreciation of satellites over five years. This is the business that makes the money: $5.2B of the $0.7B nine-month operating income (the other two lost money).
• Space sells launches. Customers bought 12 launches in Q3; 23 more were internal, mostly SpaceX's own satellites, whose launch costs are capitalised in Connectivity rather than sold. Space revenue was $1.77B, and the segment lost $7M after $1.2B of research and development, mostly Starship. Customer launches have been flat at 33-45 a year since 2023 (S-1/A).
• Super Intelligence (SI) is the former xAI and X: Grok, the X advertising business, Cursor (bought 2026-08-14) and, since May 2026, compute rented to other AI companies. SI revenue jumped from $0.8B in Q1 to $7.56B in Q3, of which about $3.75B is Anthropic (inference from the contract) and $0.61B is six weeks of Cursor. SI earned $0.43B, or $1.66B before Cursor award expense, on $16.4B of capital spending in the quarter.

The company spent $48.1B on capital in the first nine months of 2026, $40.0B of it on SI, against $13.1B of operating cash flow. The $85.7B IPO (June 2026) and $25B of notes pay for it: cash and securities were $89.9B and debt $38.8B at 2026-09-30, including $13.3B of sale-leaseback debt owed to a fund run by a SpaceX director (10-Q note 17). Backlog was $47.5B at 2026-06-30 (10-Q note 3), 56% to be recognised within a year, and $63B at 2026-09-30 with $17.5B of it deferred revenue (the call, after the price); it counts enforceable contracts and does not say how much of the Anthropic fee lies beyond a 90-day notice, so it is not a model input.

Q3 2026 revenue nearly doubled from Q2, with no guidance to beat

Before the call SpaceX gave no numeric guidance, so there is no guidance record to measure; the four reported quarters are the record. Against Q2: revenue $14.58B against $7.81B, operating income $2.79B against -$0.14B, Starlink adds 1.6M against 1.7M, ARPU $69 against $66, nameplate compute 1.7 GW against 1.4 GW, launches 35 against 38. The vendor's figures (EPS $0.35, revenue $14.575B) match the release; The vendor's statement shows revenue as 0, an market data error.

Where we differ from the blind report: SI's margin depends on which rent it earns

The blind report argued that Q3's 21.9% SI margin (before Cursor award expense) was a young fleet's, and that a fully depreciated gigawatt would earn 8.4%. That used $46.1B of capital per GW and a 34.5% cash-cost ratio that is mostly a fixed base. Recomputed from the same Q3 rent of $18.1B per GW-year, with depreciation at 1/5.5 of capital (preview.txt): 19.9% at Q3's cash costs and $34.6B of capital, 29.5% at a 25% variable cost, 8% stock pay and $37.3B of capital. At the Q3 call's $42B per GW-year the margin is 50.8%, and 27.5% at the 2031 central rent of $17.2B. The model's 2031 SI margin is 25.5% (the fixed cost base and the depreciation on the whole fleet are in it). The rent on new capacity, not the cost, decides the value, and the only evidence for it after the Q1 guide is management's own word. Because the exit rent is the one that earns the WACC at the case's own costs, the cost assumptions now matter through the path, not the exit: variable cost at 35% moves the central by only -$1.39. That is a consequence of the exit rule, not a finding about costs.

In the revised central, SI is $66.9B of the $91.0B of 2031 operating income, Connectivity $22.6B and Space $1.5B; the blind central had Starlink as two-thirds. The value rests on whether compute rent holds. The company's own case, that it can rent compute at $50 a watt or more while the cost of a gigawatt falls, is the bull.

The forecast: 2031 operating income of $91.0B in the central case

model.py builds each segment from its own drivers; output.txt prints every year. All drivers are listed with sources in assumptions.md.

The bear is one world, with its timing separated: the contracts are honoured through Q1 2027 (the guide holds), Anthropic serves notice in 2028 and leaves with that year, the new-capacity rent falls 30% a year and the legacy fleet's 12%, SpaceX cuts SI spending to $10B a year by 2029, Starship slips and Starlink grows more slowly at lower prices. Its 2031 SI loss is depreciation on the 2026-2028 build and a rent of $9.6B; the exit then does two things for it: the depreciation correction replaces the harvested fleet's depreciation with the steady state of $10B of spending (+$27.2B of NOPAT in the fade), and the rent pull lifts the exit-year rent to the $20.2B that earns the WACC (+$24.7B). Both are the method's year-five rules applied the same way as in the central, and together they are why the bear is $36.69, not $12. The bull is management's case nearly as stated: $50B per gigawatt-year and 12% a year off it, building $140B a year to 20.4 GW, SI at 51% margin and 2031 revenue of $638B, with the exit rent pulled by the same rule as the others: 47% of the way from $30.0B to the $12.9B that earns the WACC at the bull's costs (-$83.0B of NOPAT). There is no exception for the bull. Its $243.67 is above the price; it needs every input at once, we do not expect it, and at 25% it contributes $61 of the case-weighted $110.90.

Exit. The method's default, unchanged: the mean of the long fade and the outside view at the peers' 22.5x EV/EBIT on reported year-five EBIT, with the new-capacity rent pulled to normal in both and the SI depreciation correction in the fade only (method commit 86cdb4522). SpaceX has no fiscal-year multiple of its own (losses before the IPO), so the ten peers' median stands alone. The central outside view ($160.44) is 60% above the fade ($100.05). The input behind the gap is the return on new capital: the peers' multiple prices peer economics (the compute peers' median after-tax return on capital is 35%, peers.txt), while the fade prices SpaceX reinvesting 29% of its profit at the 13.9% mature return to grow 4% a year, 13.8x year-six NOPAT, after starting at the volume-only 26.7%. The depreciation correction is +$11.4B of central NOPAT in the fade ($104.5B of SI depreciation to the $90.0B that year-five SI spending implies at 4% growth); the rent pull is -$4.5B in both. The multiple itself is not the lever (+$5.56 at 25.2x, -$5.89 at 19.6x). The ten-year extension (same drivers to 2036, SI spending held at the 2031 level, the new-capacity rent floored at the normal rent after 2031) gives the central $104.66, the bear $39.61 and the bull $139.59: EBIT rises to $155B in the central by 2036, so the five-year exit is not capitalising a peak.

Bridge. Free cash flow is kept to the exit and negative free cash flow is borrowed at 6.03% (central net debt peaks at $17.6B in 2027, 0.17x EBITDA, against a 3.75x covenant). Cash taxes follow 100% bonus depreciation on 90% of capital spending (IRS Notice 2026-11) with the $10.8B opening loss carryforward used at 80% of taxable income; the deferral is repaid evenly over the 15-year fade, $22.5B of present value deducted at the exit in the central. Less one month of year-five cash costs as operating cash ($9.6B) and the litigation accrual; plus digital assets. Operating leases are not deducted (the lease cost is in EBIT). Today: no Cursor claim, because the unvested Cursor shares are in the count (below). Shares 13.072B: today's class A and B less Mr. Musk's 1.302B restricted shares, whose vesting needs a million-person Mars colony or 100 TW of space data centres (S-1/A), plus the diluted increment and the EchoStar shares. Counting the restricted shares would lower the central value to $74.22.

The share count, checked at the market pass (market.txt; 10-Q Q2 2026 cover and notes 12-14, Q3 release). Every class: class A 8.240B and class B 5.339B outstanding at 2026-09-30; class C was reclassified into A and all preferred converted at the IPO, so none is left (the 10-Q cover at 2026-07-28 shows 7.696B A and 5.485B B). Mr. Musk's 1.302B restricted class B are issued and inside the class B count: the SpaceX CEO Award (1.000B, 15 tranches from a $500B to a $7.5T market cap, each also needing a million-person Mars colony) and the AI CEO Award (0.302B, 12 tranches from $1.065T to $6.565T, each also needing 100 TW a year of compute off Earth). We leave both out because each tranche needs its non-market condition. The 10-Q excludes 1.331B performance awards from diluted shares, so about 0.029B belong to others, with no terms given; not counted. Options (0.472B) and RSUs (0.122B) at 2026-06-30 are in the count through Q3's 0.533B diluted increment; counting them gross instead is a small change (-$0.38). EchoStar's 0.262B shares are counted now, with its 2027 cash. Q3 basic weighted shares (11.942B) are about 0.27B fewer than the Q3 additions would give if all counted in basic, so about 0.27B of the new shares look like unvested restricted stock, most likely the Cursor holders' time-vesting awards (our inference). The blind model charged them twice, in the divisor and in a $22.2B dollar claim; the revision drops the claim (+$1.70). If they are outside the count, adding them gives $79.95 (-$1.67). The Q3 10-Q's equity note, not filed at 2026-10-20, settles it. Central equity value is $1,066.8B ($81.61 x 13.072B); at the price, $1.93T.

Cost of equity. 9.41% from the method's 5.29% risk-free rate and 4.20% premium and a 0.98 beta, the Blume-adjusted median of the ten peers. The five compute peers alone give 1.14 (cost of equity 10.08%) and $76.64 (-$4.98): printed because the compute business is the value, and we kept the ten-peer set fixed before the blind commit. A beta of 0.60 gives $97.54.

The drivers that move the value most

The fall in the rent has a lopsided effect (-$6.87 against +$66.58) because the exit rent is pulled to normal when the rent falls and left alone when it holds. At -10% the pull is a third of the way and the exit capitalises a rent of $27.5B. Spending more on SI now raises the value at the margin (20% more is +$9.20): the next gigawatt at 2031 economics earns 10.0% after tax against a 9.18% WACC (a $17.2B rent against $16.6B to earn the WACC). That conclusion is the rent assumption restated.

The price needs the rent on new capacity to fall 10.0% a year, not 20%

At $147.79 the market pays $1.93T for our 13.072B diluted shares (the vendor's market cap, $1.95T, uses 13.18B). Holding our central operating path and cost of equity, the price needs an exit of $235.86 on 2031-12-31 against our $130.25: $2.90T of enterprise value after the central exit bridge, 32x our 2031 EBIT of $91.0B, or 22.5x an EBIT of $129B, 1.4 times ours and 40% of our 2031 revenue (market.txt).

The figure that closes the gap is the yearly fall of the new-capacity rent. On our build (16.4 GW in 2031), the price needs the rent to change -10.0% a year after 2027 instead of -20%, which ends 2031 at $27.5B per GW-year instead of $17.2B; SI revenue is then $409B, total revenue $467B and EBIT $189B at 40%. Building more helps when rent holds and hurts when it falls (market.txt; the pull to normal fades to zero by -5%):

At a 10% yearly fall the price needs the 2027 rent at $41.9B, our central, so the two readings of the gap are the same one: if the rent falls 10% a year from $42B, the price is fair.

The Street through our model (Street analyst estimates, read 2026-10-20; the file does not say whether they include the call). The Street's revenue is below ours: $109.0B in 2027 and $201.1B in 2028 against our $164.7B and $259.5B, and its EBITDA is $35.1B and $67.0B against our $102.1B and $165.5B. Its EBIT field does not reconcile with its own EPS (2028 EPS $4.76 is $62.2B of net income against a $16.7B EBIT); our 2028 after-tax operating income is $85.6B. The Street's average target ($213.04, $194.71 discounted one year at 9.41%) needs, on our model, the new-capacity rent to fall 8.4% a year after 2027, everything else central. If the estimates predate the call, the Street is catching up to the guidance and its target is the stale half of a moving number; either way our central is below both the price and the Street.

The odds (odds.json): the Monte Carlo puts P(fair value at least the price) at 14%, with a 10th to 90th of $58 to $173, a median of $79 and a mean of $99, the median 3.5% under the central. The mean is above the central and the median below it because the rent-change triangle (min -40%, mode -20%, max +5%) is wide on the side where the value is steep: it explains 76% of the variance alone, the legacy rent change 36%, SI spending 23%, capital per GW's yearly change 17%, variable cost 15%. The simulation uses the same exit rule as the central, so a persistent-rent draw is not pulled back to normal (the pull fades to zero by -5%). The +5% top is Musk's "higher dollars per watt, not lower" taken as a rising rent; the first version of the range stopped at -5%, which left management's own statement outside it, and the result then was 2.9%. These odds follow from our ranges and are not an independent check. The blind version of this paragraph said "under 1%"; the review showed that no case had rising rent and that the result held by construction.

Is a risk charged twice? (the checker's flag for a value far below the price, flag 3). SI depreciation is charged once, and the bear's harvest depreciation is corrected in the fade only. The new-capacity rent pull to normal sets the exit rent, not a cost, so it does not repeat the rent path's fall. Stock pay is a cost in each year and the share count is held flat, so it is not also diluted. Debt for the build is charged once, as interest in the net cash path. The Cursor awards were charged twice in the blind model (shares and a dollar claim) and are now charged once. Anthropic leaving is in the bear only, with its timing separated from the rent fall. The deferred tax is charged once, as a present value at the exit. Short contracts are charged once, in the rent's yearly fall; they are not also in the cost of equity. Known limit: with the exit rent set by the WACC at the case's own costs, cost assumptions affect the value through the path and not the exit.

The case for the price, argued as hard as we can

The price is right in a world where SpaceX's new compute keeps earning close to what the CFO says it earns today. Built in full, on our own model:

1. The rent holds. The price needs the new-capacity rent to fall 10.0% a year, against our 20%. The case for it: frontier models earn more per token as they get better, the buyer pays for finished capacity, and the CFO says pricing is above the top of last quarter's range with payback in under a year. Musk says renewals come "at higher dollars per watt, not lower". Google ($920M a month) and the unnamed December customer ($1.11B a month) signed after Anthropic's $5.27 a GPU-hour, at higher implied rents, and the Ornn B200 index read $8.83 on 2026-10-20 (one snapshot; H100 +17.7% and B200 +14.5% in 30 days). If compute stays scarce through 2031, rent holds with the value of what it trains, and the build answers it: B in the pairing table is $164.45.
2. The software sits on top. Grok, X and Cursor (about $4.7B a year, bought 2026-08-14) earn software margins on capacity SpaceX owns. Our model lumps them with rented compute at one fixed-plus-variable cost structure; the Street's EPS needs SI cash costs near 19% of revenue. Musk puts internal use at about half of compute. If most SI revenue becomes Grok and Cursor subscriptions rather than rent, a lower cost ratio is reachable, but the exit rule makes the cost ratio matter little to the central (35% variable cost is -$1.39).
3. Scale and cost. SpaceX builds data centres fast (COLOSSUS II's first cluster in 91 days, S-1/A) and owns its power and launch; Musk says it may deploy power and cooling faster than all others combined. At the Q4 pace of build (1.15x) the central is $88.5; at his 10 GW for 2027 it would be higher still if the rent held.
4. Options the model leaves out. Starship has one revenue flight; the AI CEO Award's 100 TW condition says management means data centres in orbit; Starlink Mobile on EchoStar spectrum closes in 2027; its revenue sits only inside the enterprise and government path. Musk's merger comments about a Tesla combination (no timeline given), and chips from a Terafab by 2028, are not modelled.
5. The record. Starlink went from nothing to 13.6M subscribers and a 45% margin; the people now building SI built that.

Why we still say Avoid. The central ($81.61) is 44.8% below the price and $44 under the 0.85x line ($125.62): this is not a margin call. The case-weighted value ($110.90) is also below the price; only the bull ($243.67) is above it. The value is a bet on one number, the yearly fall of the rent on new capacity, between -10.0% (the price) and -20% (ours), and the evidence for either is one CFO statement, one quarter's guide and one CEO remark against contracts that the buyer can end in 90 days; the sell-side's own path (C) is lower than ours. The purchase price ($72.28) is 51% under the price. We hold Avoid with 12% confidence. What settles it: Q1 2027 revenue against $33B-$37B (our central implies $28.5B), SI revenue per average gigawatt-year in Q4 2026 and Q1 2027, the price of the first contracts that renew after their 90 days, SI capital per GW added over the same quarters, and two quarters of SI cash costs under 25% of revenue.

The bear could be too kind or too harsh: it keeps Starlink growing and leaves the 800 MHz spectrum purchase (price not disclosed) out, assumes the guide holds through Q1 2027, and is lifted by the same exit rule as the central.

What would change the view

Falsifiers, each moving the view toward a case:

• Q1 2027 total revenue below the $33B low end of the guide moves us toward the bear; at or above $37B toward the bull ($243.67).
• The first renewals of the 90-day contracts priced below $30B per GW-year, or any notice ending the Anthropic, Google or December agreements, is the bear's world.
• Renewals at or above $50B per GW-year with 2027 capacity delivered: the rent persists and the exit rent should not be pulled to normal (central $148.19 at a 10% fall).
• SI cash costs above 40% of SI revenue for two quarters (Q3 34.5%).
• SI capital per GW added above $45B in Q4 2026 or Q1 2027 (central $37.3B; the Q4 flow is $43B) is -$2 to -$4 a share at a $43-50B base.
• Q4 2026 capex outside $35B-$40B, or nameplate compute under 2.3 GW at the year end.
• Starlink net adds under 1.2M a quarter for two quarters, or ARPU under $62.
• Connectivity capital spending still above 30% of Connectivity revenue in 2027 while V3 satellites fly on Starship says Starship's cost cut is late.

Next results preview: Q4 2026

Due about 2027-02-25. The company has not announced the date; 2027-02-25 is the Earnings calendar's estimate and is not confirmed. The company gave a Q4 capex range ($35B-$40B) and a Q1 2027 revenue range ($33B-$37B) on the call (after the price); it gives no other numeric guidance. Values from model.q4_2026(); ranges and reasons from preview.py (preview.txt).

The most important is SI revenue, because it carries the key driver. Q4's guide for Q1 2027 is the test: our central implies $28.5B. A $1B surprise in Q4 SI revenue is $1.90B of revenue per gigawatt-year, 10.5% of the legacy level; held as a new level it moves fair value by $1.31 a share, while as a one-quarter event it is worth $0.06 (output.txt). The new-capacity rent is the number to watch, and the first quarter in which the 90-day contracts have renewed once is the first read of it.

Sources

• SpaceX S-1/A, filed 2026-06-03 (EDGAR, CIK 1181412)
• SpaceX 10-Q for the quarter ended 2026-06-30, filed 2026-08-04
• Q2 2026 release, 8-K exhibit 99.1, 2026-08-04; Q3 2026 release, 8-K exhibit 99.1, 2026-10-20
• 8-K on the Cursor merger closing, 2026-08-14
• Free writing prospectuses: Google cloud service agreement, 2026-06-05 (data/filings/fwp-2026-06-05-google.html); CFO interview, 2026-06-09 (data/filings/fwp-2026-06-09-cfo.html)
• Market data statement endpoints for SPCX and ten peers, and the peers' market data profile betas, fetched 2026-10-20
• Exit anchor history, data/exit-anchor.json, scripts/research-anchor-data.py, fetched 2026-10-20
• Method inputs (risk-free rate, premium): method.py, Damodaran 2026-10-01
• Market pass, read 2026-10-20 after the blind commit: market data quote (20:00:01Z), Market data profile, Street analyst estimates and price-target consensus; market data get_equity_quotes, read-only (20:44Z); all in data/market/
• Revision, 2026-10-20: Ornn GPU rental price index, https://data.ornn.com/gpu-cloud-price-comparison (data/ornn-2026-10-20.txt); IRS Notice 2026-11, interim guidance on the additional first year depreciation deduction (data/irs-notice-2026-11.pdf)
• Q3 call, 2026-10-20, after the price: Investing.com earnings call transcript (AI-assisted; read through a summary; CFO name mislabelled), Dow Jones Newswires headlines via news feed, the market data live blog; Goldman Sachs Communacopia, 2026-09-10 (Investing.com transcript); Q2 call 2026-08-04 (Motley Fool transcript); all in data/q3-call-transcript-2026-10-20.txt and data/q3-call-2026-10-20.txt. Third-party figures used as counter-evidence, not inputs: Morgan Stanley's per-watt path (sell-side, as reported), SemiAnalysis's cost per gigawatt (as reported). Not read: the recording and any 8-K slides.

Report-checker flags and how each was answered

• Flag 1 (large net move, +$60, from the blind value): the largest change is the pair C3 and C4, new capacity at $42B per GW-year and the build at the Q4 run-rate, +$66.6; it rests on the CFO's per-watt figure and Q4 capex guide, management statements made after the price, tested against payback arithmetic, the Q1 2027 guide and Google's fee per GPU. After the call-step review the largest single move is the capital-per-GW fall removed (-$21.84).
• Flag 3 (value -44.8% from the price): "Is a risk charged twice?" above.
• Flag 4 (outside view 60% above the fade): the return on new capital, under "Exit".
• Flag 14 (simulated median under central): the median of $79 is 3.5% under the $81.61 central because the rent-change triangle is wide on the side where the value is steep; see "The odds".
• Flag 16 (P(price justified) low): 14%, a product of our ranges; the world where the price is right is the rent falling 10.0% a year, under "The case for the price".
• Flag 17 (fair value against the discounted target): the Street's target needs an 8.4% yearly fall of the new-capacity rent; see "The Street through our model".
• Flag 18 (narrow range against targets): the ranges were widened in the review revision and again for the call (new-capacity rent -40% to +5% a year, $25B to $55B in 2027) and again here (capital per GW $30B to $50B, yearly change -5% to +5%).
• Blind-commit warning (the price 147.79 appears in the blind data): the match is a CSS coordinate, left:147.79pt, in the S-1/A's HTML, not a price.

SPCX — Space Exploration Technologies Corp. — research view: Avoid
As of 2026-10-20 · partly unblinded, three revisions unblinded

Fair value today      $81.61 per diluted share (cost of equity 9.41%, beta 0.98); equity value
                      $1,066.8B on 13.072B shares (blind $21.38; $63.57 after the first review;
                      $121.95 after the Q3 call, before the call-step review)
Price                 $147.79, market data quote, 2026-10-20T20:00:01Z (regular-session close; market data
                      last regular trade $147.95, after-hours $150.95 at 20:44Z)  →  fair value
                      44.8% below the price
Five-year exit price  $130.25 on 2031-12-31 (mean of the long fade $100.05 and the outside view $160.44;
                      a 10- or 20-year fade gives $129 / $131; value-driver formula 13.7x as a check)
Five-year return      -2.4% a year from today's price on our central case (bear -16.3%, bull +20.5%);
                      under the 12% hurdle in the bear and central cases
Purchase price        $72.28 to earn the 12% hurdle; the 0.85x view line is $125.62
Bear / bull today     $36.69 / $243.67, each a consistent business outcome; case-weighted $110.90 (25/50/25, our weights)
Confidence            12% that value at the target date lands within ±20% of the central (subjective, not calibrated;
                      the yearly fall of the new-capacity rent alone moves the central from $75 at -30% to $148 at -10%)
Moat                  Starlink: the largest low-orbit network (about 75% of active maneuverable satellites,
                      S-1/A) launched on the cheapest rockets; SI: none shown yet, compute rented on 90-day terms
Market expects        an exit of $235.86 (32x our 2031 EBIT of $91.0B, or 22.5x an EBIT of $129B, 1.4 times ours);
                      on our build, the rent on new capacity falling 10.0% a year after 2027 instead of 20%
                      ($27.5B per GW-year in 2031 against our $17.2B)
Price needs           the new-capacity rent falling 10.0% a year (central -20%); no other single driver reaches
                      the price within its range
Odds in our scenarios 14% of simulated scenarios have fair value at or above the price (P(price justified);
                      10th-90th $58 to $173, median $79); a result of our ranges, not an independent measure
Key driver            rent on new capacity, yearly change after 2027: -30% is -$6.87, -10% is +$66.58 a share
Street                average target $213.04 (median $211, range $75-$450; market data, 2026-10-20);
                      revenue $109.0B in 2027 and $201.1B in 2028 against our $164.7B and $259.5B (the
                      estimates may predate the call)