← The canon · AItopiaOrAImageddon?

Flight-proven — the first reflight of an orbital-class booster, and the cost curve that did not follow

moment · SpaceX; Elon Musk set the goal, the schedule and the hundred-fold cost claim; Lars Blackmore ran entry, descent and landing for Falcon 9 from 2011 to 2018 and co-invented the G-FOLD powered-descent algorithm with Behçet Açıkmeşe, Jordi Casoliva and a third JPL co-author, whose surname is redacted here and named in Sources; Gwynne Shotwell set the commercial terms and the 30% figure; Martin Halliwell, chief technology officer of SES, was the first customer willing to put a satellite on a rocket that had already flown · 2017

Something that happened and changed what people expected next.

The finding, stated up front so the rest of the file can be checked against it:

The engineering claim was a hit, delivered about four years late. The economic claim — a hundred-fold reduction in the cost of access to space — is a miss by roughly a factor of fifty at the price a customer pays, and it failed for a nameable reason the claim itself contains: the part you throw away sets the floor, and the Falcon 9 upper stage is still thrown away on every flight. The skeptics were wrong about feasibility, which they mostly did not claim, and right about demand elasticity, which they did — and they were made irrelevant not by being refuted but by a launch provider that became its own largest customer. Fifteen years after the claim, every published SpaceX price is higher in nominal dollars than before reuse existed and roughly flat in real terms. The cost curve fell. The price curve did not. The gap is margin, and it financed Starlink and Starship.

All five of those are true at once, and a file carrying fewer than five is worth less than this one.

1. What it is

The event

On 30 March 2017 at 22:27 UTC, a Falcon 9 lifted off from Launch Complex 39A at Kennedy Space Center carrying SES-10, a 5,300 kg communications satellite bound for geostationary transfer orbit. The first stage, booster B1021, had flown before: it launched the CRS-8 cargo mission to the International Space Station on 8 April 2016 and landed on the droneship Of Course I Still Love You, the first time an orbital-class booster had been recovered at sea. On SES-10 it flew a second time, and was recovered a second time.

That is the entry's date. It is the first occasion on which an orbital-class rocket stage did the thing the entire argument was about — flew, came home, and flew again with a paying customer's payload on top.

The more famous date is the wrong one, and the distinction is the point. On 21 December 2015, on Falcon 9 flight 20 and the debut of the Full Thrust variant, booster B1019 lifted eleven Orbcomm OG2 satellites from Space Launch Complex 40 at 8:29 p.m. EST and landed upright at Landing Zone 1 at Cape Canaveral about nine and a half minutes later. That is the image everyone has: the long exposure, the tail of fire coming back down. It proved recovery. It did not prove reuse, and reuse was the disputed proposition. B1019 never flew again; it was retired to display outside SpaceX headquarters. Fifteen months separate the picture from the fact, and a canon that files the picture as the milestone is filing the marketing.

How it was got to

The program ran roughly as follows, and the shape of it matters more than any one date.

Where it stands as of this file

As of 16 August 2026, three days before this entry was written: 679 Falcon 9 launches, 676 successful (99.6%), 643 landing attempts and 632 successful landings. The fleet-leading booster, B1067, flew its 35th mission on 8 June 2026 and its 36th on 9 July 2026 — more flights than any orbital booster in history, three short of the Space Shuttle orbiter Discovery's 39. B1067 debuted in June 2021 on CRS-22 and later carried two NASA crew rotations.

Annual Falcon 9 cadence: 25 (2020), 31 (2021), 61 (2022), 96 (2023), 134 (2024), 165 (2025). In 2025 SpaceX passed its 500th landing and its 500th flight on a used booster, put more than 3,000 Starlink satellites in orbit, and accounted for roughly 85% of all US orbital launches.

Put that against the world Richard Webb was describing when he wrote the best statement of the skeptical case: in calendar 2015 there were 86 orbital launches worldwide, flown by 21 launch systems from 7 countries. One company now flies roughly twice the entire planet's 2015 rate, on hardware it uses more than once. Whatever else this file concludes, that number is real and it is not small.

Boosters are designed for 40 flights and depreciated in SpaceX's own filings over 25 — the fleet leader is, on the company's books, long since worth nothing.

The turnaround is the number that did not arrive. The record for the same booster flying twice is nine days, set by B1088 in March 2025, beating fourteen days (B1080, November 2024). The goal Musk stated was twenty-four hours. The fleet turns around fast — two Falcon 9s launched 38 minutes apart on 15 August 2026 — but that is pad and fleet cadence, not vehicle turnaround, and the two are routinely confused.

2. The kind, and the header

Filed moment, and the proposal filed it correctly — canon/proposals.md describes "orbital-class booster recovery and reflight… and the launch-cost curve that followed," which is an event with a consequence. On the house definition set by eliza-1966 and used since by dartmouth-1956, llama-weights-2023, clippy-1996 and expert-systems-collapse-1987 — a moment is a date on which something visibly happened in public, and build.py states it as something that happened and changed what people expected next. 30 March 2017 has a date, a place, a livestream, a paying customer and a very large change in what people expected next. sputnik-1957 observed that the hard part of an obvious moment filing is finding something useful to do with the rest of the file; that applies here.

prediction was the serious alternative, and rejecting it creates a problem this file has to name.

The argument for prediction is strong. Section 4 grades seven dated claims with the full four-part discipline — claim, date made, date due, what happened — and LENSES.md is explicit that a principal's forecasts "are canon entries of kind prediction, dated and graded against what happened." Half this file is a prediction entry.

The argument against is the one tesla-volume-production made on 19 August 2026 and I am following it deliberately: "the artefact at the centre of it is not a claim — it is a factory that reached a rate… The canon needs the thing that was built to be the entry, and the forecasts about it to be graded inside — not the other way round." Substitute "a booster that flew thirty-six times" and the argument transfers without loss. If this file were prediction, the graded object would become Musk's cost forecasts, and 632 landings would shrink to a fact mentioned while scoring somebody.

But here is the problem, and it is now a canon-level problem rather than a file-level one. proposals.md commissioned tesla-volume-production and this entry for one stated reason: "the canon's grading of predictions is only worth something if it carries the hits." Both files have now been written, both are hits, and both were filed moment. Meanwhile musk-robotaxi-2019, a miss, is filed prediction. The per-file reasoning is sound in each case and the aggregate is a bias: anyone who computes a base rate by reading the entries whose kind: is prediction will now systematically over-count failures, because this canon's convention files a forecast that came true as the thing that came true, and a forecast that did not as a forecast.

That is exactly the failure the prediction spec warns about in the opposite direction — "a base rate assembled from self-assessments would be worthless" — arriving through the filing system instead of through flattery. The correction is not to misfile this entry. It is that the canon's forecasting base rate must be computed over graded dated claims wherever they live, not over the prediction shelf. Section 4 of this file and section 3 of tesla-volume-production are written to be extractable for that purpose, and this paragraph is here so the next reader does not have to notice the bias on their own. A future canon job could reasonably be asked to write an entry that holds the base rate itself; it does not exist yet and I am not permitted to invent the id.

interpretation was not considered. Nobody is arguing a position here; a booster flew twice or it did not.

descends_from: [], by argument

Nothing in canon/ is an ancestor of this, and I checked before writing the brackets empty.

lighthill-1973 is the pairing and is not an ancestor. proposals.md names it explicitly: cite this entry "beside lighthill-1973, already written, where the expert consensus was right." That is a citation instruction, not a descent claim. Nothing about the 1973 survey of AI research funding fed into a rocket landing algorithm. Section 5 does the pairing work in prose, and the pairing is more interesting than the proposal makes it sound, because the two consensuses failed in different ways.

sputnik-1957 is the canon's only other entry about getting to orbit and is still not an ancestor. That file is about the "Sputnik moment" as a rhetorical premise — a demonstration read as a gap read as a mandate for a buildout. There is a real edge in the neighbourhood, and it belongs to that file rather than this one: 21 December 2015 was widely narrated as exactly such a moment for European and American launch incumbents, and the response was exactly the buildout the phrase implies. But a shared trope is not descent.

tesla-volume-production is the sibling and the descent, if any, runs backwards. SpaceX's landing program predates the Fremont ramp by years, and tesla-volume-production asks, in its own §2, that this file be read beside its section 4 and that section 4 "probably be shortened" once this exists. I have read it. I do not think it should be shortened; I think the two files disagree usefully about how much of a forecaster's record schedule slip explains, and section 4.8 below says where.

musk-robotaxi-2019 should probably descend from this file, and I cannot write that edge. tesla-volume-production made the identical observation about itself: the credibility in the room on 22 April 2019 was manufactured elsewhere first. By April 2019 the same man had landed 34 boosters and reflown a dozen. A canon job writes exactly one file, so: a future revision of musk-robotaxi-2019 should consider adding both tesla-volume-production and spacex-booster-reuse to its descends_from.

Ids I wanted and did not invent. moore-1965 is the canon's natural home for cost-curve and learning-rate arguments and is still unwritten; section 5.1 has to make the argument itself rather than cite it. google-index-1998, human-genome-project-2003 and twitter-acquisition-2022 are proposed and unwritten. There is no Space Shuttle entry, which is the single largest gap this file ran into — the Shuttle is the prior that the skeptics were reasoning from, and I have had to summarise it inside section 4 rather than point at it.

3. Why a reading would cite it

This is an AI tracker. A rocket entry earns its place only if a reading actually reaches for it, so here are the occasions, each tied to a lens in LENSES.md.

3.1 — Compute and infrastructure: capex justified by demand that does not exist yet. This is the strongest and most immediate cite. Readings through August 2026 have recorded twenty-year datacenter leases, hundred-billion-dollar backstops and commitments with nothing on the other side of the ledger, and the open question in every one of them is who buys the output. SpaceX is the worked case where the answer was: the builder. Gatti and D'Ottavio (Intereconomics, 2025) measured it — 69% of SpaceX's missions were self-provisioned; strip out Starlink and the company flew 30 launches. The reusable fleet was not filled by a market that grew to meet it. It was filled by a satellite constellation the launch provider built in order to have something to launch. That is not a debunking, and a reading must not use it as one: it worked, spectacularly, and the constellation is a real business with real subscribers. It is a precondition list. It worked for an entity that owned the whole stack, had a decade of patience, and could absorb the losses. A reading weighing a buildout can ask, concretely, which of those three the buildout's owner has.

3.2 — Concentration: what a cost breakthrough does to price when nobody else can do it. The single most citable fact in this file. SpaceX's estimated internal cost per Falcon 9 launch is $15–28 million; Musk's own best-case marginal figure was $15 million (Aviation Week, May 2020), of which about $10 million is the new upper stage and about $1 million is refurbishing the booster. The list price is $74 million, and NASA has paid an average of about $106 million per Falcon 9 over the last decade. The surplus is enormous, it is legitimately earned, and almost none of it reached the customer — because after 2020 there was nobody else to buy from. This is the concentration lens's cleanest historical instance: a genuine order-of- magnitude cost breakthrough, landing inside a firm with no competitor, produces a margin and not a price cut. Cite it whenever a reading meets the claim that falling model-serving costs will reach users.

3.3 — The cost/price distinction, which the readings already make on their own. The 19 August 2026 midday reading observed of GLM-5.3 that "flat token prices are not flat workload prices," and that only an outside party produces that number. This entry is the fifteen-year version of the same discipline, in the other direction: falling costs are not falling prices, and only an outside party — a government auditor, a competitor, a customer with leverage — makes the two meet. Where a counterparty did have leverage, the pass-through was real and measurable: in 2020 the US Space and Missile Systems Center modified two GPS III launch contracts to permit reuse, and the military estimated the change saved $64 million across two flights, about $32 million each. Where the counterparty had none — NASA's dedicated bespoke-orbit missions — it did not.

3.4 — Robotics and embodiment: a measured demo-to-deployment interval. LENSES.md insists that "demonstrated under controlled conditions" and "deployed and working" are different findings and that a reading must say which. This entry supplies an unusually clean number for a hard physical control problem — supersonic retropropulsion and pinpoint powered landing of a 70-metre object onto a moving barge. First success 21 December 2015; first commercial reflight 30 March 2017; routine by the Block 5 debut on 11 May 2018. Call it two and a half to three years from first success to routine — and note that the two failures immediately preceding first success were in January and April 2015, so the "nearly worked" phase was under a year. That is a real base rate to hold against a humanoid-robot demonstration video, and it cuts both ways: it is fast enough to be a warning against dismissing demos, and it is slow enough to be a warning against treating one as deployment.

3.5 — When an expert consensus is being weighed against a builder. The proposal's stated occasion, and section 5.2 argues it is more interesting than "the experts were wrong."

4. What it got right, and what it got wrong

Not required for moment. It is why this file was commissioned, so it is the longest section here. Every claim carries four things: the claim in the claimant's own words where I could get them, the date it was made, the date it was due, and what happened. Both sides are graded. A file that grades only Musk would be the mirror of the failure musk-robotaxi-2019 warns about in its section 9.

The builder's claims

4.1 — The core economic claim, and the largest miss. Claim: "If one can figure out how to effectively reuse rockets just like airplanes, the cost of access to space will be reduced by as much as a factor of a hundred. A fully reusable vehicle has never been done before. That really is the fundamental breakthrough needed to revolutionize access to space." Made: SpaceX.com, 31 March 2013 (quoted in Webb 2016, Figure 1); the substance dates to the 29 September 2011 National Press Club speech. Due: no date given — treat it as due once reuse is routine, which it has been since roughly 2018. Mechanism he gave: fuel and oxygen cost about $200,000 against a $50–60 million vehicle, so a rocket flown a thousand times carries about $50,000 of capital cost per flight.

What happened: The published price of a Falcon 9 went from $54–59.5 million in 2012 to $74 million in 2026. US consumer prices rose about 45% over that span (CPI-U 229.594 → 333.918), so $54 million in 2012 is about $78.5 million in 2026 dollars — the real price fell by roughly 6%. Not 99%. Six.

On price per kilogram the picture is better and still nowhere near the claim. Falcon 9 v1.0 in 2012: about $54 million for 8,500–9,000 kg to LEO, ≈ $6,000/kg, ≈ $8,700/kg in 2026 dollars. Block 5 in 2026: $74 million for ≥22,800 kg expendable (≈ $3,250/kg) or ≥17,400 kg in reusable configuration (≈ $4,250/kg). That is a real improvement of roughly 2× to 2.7×, and most of it comes from the rocket getting bigger rather than from flying twice.

Even on SpaceX's own most favourable number — Musk's $15 million best-case marginal cost against the ~$54–62 million list price of the pre-reuse era — the factor is about 4.

Grade: Miss, by roughly a factor of fifty at the customer's price and a factor of twenty-five against the company's own marginal-cost claim. This is the largest miss in the file.

4.2 — Why 4.1 failed, which is the part worth citing. The claim contains its own refutation and the arithmetic is short. Two assumptions carried it: a thousand flights per vehicle, and full reusability. The fleet leader has 36. And on 17 November 2018 SpaceX abandoned Falcon 9 upper-stage reuse entirely. So every Falcon 9 flight still throws away a stage that costs roughly $10–12 million to build — which by Musk's own 2020 breakdown is two-thirds of the entire marginal cost of a launch, against $1 million to refurbish the booster.

The part you throw away sets the floor. Booster reuse can drive the recoverable fraction of the cost toward zero and the flight still costs an upper stage plus range and recovery operations. This is the transferable lesson and it is not about rockets: an efficiency that addresses one term of a cost function is bounded by the terms it does not address, and the bound arrives much sooner than the headline factor suggests. A reading meeting "we cut the cost of X by 90%" should ask what the remaining 10% is made of and whether the same technique touches it.

4.3 — Full reusability. Claim: a fully reusable launch system — both stages and the capsule — under power. Made: 29 September 2011. Due: Musk gave no date in that speech; the 2011 animation showed the Falcon 9 second stage doing it. What happened: abandoned for Falcon 9 on 17 November 2018 and moved to Starship. As of 24 July 2026, Starship has flown 13 integrated test flights; a Super Heavy booster was caught by the launch tower in October 2024; on flight 13 the ship made a controlled Pacific splashdown and the booster a powered soft ocean landing. No Starship upper stage has been recovered and reflown. Grade: Outstanding at fifteen years and counting, on its second vehicle. The canon should hold this open rather than score it; the honest statement is that the specific claim announced in 2011 has not been demonstrated by any vehicle by August 2026.

4.4 — Twenty-four hour turnaround. Claim: "we intend to demonstrate two orbital launches of the same flight vehicle within 24 hours no later than next year." Made: Block 5 prelaunch press call, 11 May 2018. (An earlier version targeted 2018, stated July 2017.) Due: end of 2019. What happened: the record is nine days (B1088, March 2025); before that fourteen (November 2024). Grade: Miss by roughly two orders of magnitude on the vehicle, and it has not been retried. Note the confusion this generates: SpaceX launched two different Falcon 9s 38 minutes apart on 15 August 2026, which is a genuine and impressive result about pads and fleets and is not the claim.

4.5 — Booster life. Claim: Block 5 designed for ten or more flights with no refurbishment between flights, and 100 flights with some refurbishment. Made: 11 May 2018. Due: open-ended. What happened: the design target is now 40 flights, the fleet leader is at 36, and SpaceX depreciates boosters over 25 flights in its own accounts. Grade: Hit on the first half, and the best-calibrated claim in this file. The ten-flight figure was beaten comfortably; the hundred-flight figure is still plausible and unproven. A forecaster who is out by a factor of fifty on cost and roughly right on hardware life is telling you where the model is good.

4.6 — The commercial discount. Claim: Shotwell, March 2016: about 30% cost savings from reusing the first stage. Against the then-current $61.2 million list price, full pass-through would have meant $42.8 million. Due: on the first reflights, i.e. 2017 onward. What happened: the first reflight customer got a real discount — SES's chief technology officer Martin Halliwell said on 13 September 2016 that the deal was "closer to Gwynne than to Karim", i.e. nearer Shotwell's 30% than SES chief executive Karim Sabbagh's asked-for 50%, and that SES "did receive a good discount on the baseline price." Neither side disclosed the number. The list price then went up, and by 2026 stands at $74 million. Grade: Half-hit that decayed. The discount was real to the customer who took the risk in 2016–17; it did not survive into the published price once reuse stopped being a risk and started being the only option.

4.7 — The rideshare price, for completeness. SpaceX's Transporter smallsat rideshare entered at $5,000/kg in 2019 and stands at $7,000/kg in 2026 ($350,000 for 50 kg). $5,000 in 2019 is about $6,530 in 2026 dollars, so the cheapest ticket to orbit is about 7% more expensive in real terms than when the program opened, on a manifest that is reliably oversubscribed. There is no product line in which SpaceX's published price fell.

4.8 — The pattern claim, and where I disagree with the sibling file. tesla-volume-production states the pattern as "schedule badly wrong, direction right, thing delivered" — a forecaster habitually late and habitually right about direction. On this evidence that is too generous, and in a specific way. The schedule error here (4 years on landing, 7+ and counting on full reuse, 2 orders of magnitude on turnaround) is the familiar part. But 4.1 is not a schedule error. It is a magnitude error that has been settled for eight years and is not going to improve, because the thing that would fix it — reusing the upper stage — was cancelled. And 4.6 is a different failure again: a claim that came true briefly and was then withdrawn by the market rather than by the claimant.

So the honest pattern across both files is three failure modes, not one: schedule slip, which resolves; magnitude inflation, which does not; and pass-through claims, which depend on a competitor existing. A reading grading an AI principal's forecast should sort it into one of those three before deciding how much to discount it, because they have very different half-lives. That distinction is this file's main contribution to the base rate, and it is why I do not think section 4 of the Tesla file should be shortened.

The skeptics' claims

4.9 — The strongest skeptical claim, and it is a miss. Claim: "demonstrated that there is no scenario in which the economics of reusable launch vehicles makes them preferable in cost to expendable launch vehicles." Claimant: Dr Stephen Book, chief technical officer at MCR and one of the field's most decorated cost estimators, summarising his paper "Costs of Reusable Launch Vehicles: Should We Pay Up Front to Build in High Reliability or Pay Later to Buy More Vehicles?" (STAIF-99, Albuquerque, 1999). Made: as quoted, PARAMETRIC WORLD, Winter 2012. Due: open. What happened: SpaceX flew 165 times in 2025 on recovered hardware and holds roughly 85% of US orbital launch. Grade: Miss, and the cleanest one in the file — "no scenario" is a universal, and one counterexample settles it. **Sourcing caveat: I have this quotation only through Webb (2016, Figure 1). I could not reach the Winter 2012 Parametric World issue, and I have not verified that Book's underlying analysis is fairly summarised by that sentence.** Treat the attribution as second-hand.

4.10 — The conditional skeptical claim, and it is a hit on its own terms. Claim: "The payoff of a reusable rocket is only possible if the launcher is flown many times, and market outlooks in the commercial sector and with European institutional missions do not add up to tip the scales in favor of reuse." Claimant: Stéphane Israël, chairman and chief executive of Arianespace. Made: 11 January 2016, three weeks after the first landing. Due: continuing. What happened: both halves came true. Reuse did require a very high flight rate. And European commercial and institutional demand did not supply one — Europe flew 3 launches in the year Gatti and D'Ottavio measured. ArianeGroup's chief executive Martin Sion later said a reusable Ariane 6 was "not economically interesting" for Europe when the programme was created, and Arianespace's leadership said in 2023 that Europe would not have a reusable rocket until the 2030s. Grade: Hit on the economics, applied to the right market, and it did not save them. Israël correctly predicted that reuse would not pay for Arianespace. He did not predict that a competitor would manufacture its own flight rate and take the market anyway. The claim was right and the company was destroyed by it being right, which is a distinct and instructive outcome that "the experts were wrong" erases completely.

4.11 — The demand-elasticity claim: right about the market, overtaken by events. Claim: **"LEO vehicles have a demand elasticity above 1.0 and GTO vehicles have a demand elasticity of 0.5. In other words, our data show that within the price range of today's vehicles, lowering price will not result in increased business for heavier lift, GTO vehicles. For lighter lift LEO vehicles, price decreased may lead to more customers, but not in large numbers." Claimants: Henry Hertzfeld, Ray Williamson and Nicolas Peter, Space Policy Institute, George Washington University, Launch Vehicles: An Economic Perspective, September 2005. Due: the following two decades. What happened: essentially correct about existing demand. The commercial geostationary comsat market did not expand as launch got cheaper; it contracted. What exploded was LEO constellations — the elastic segment they named — and the largest of them was built by the launch provider itself. Grade:* Hit on the segment call, and made moot by vertical integration. They said cheap launch would not summon buyers. They were right. SpaceX did not wait for buyers.

4.12 — The best-argued skeptical document, and what it actually concluded. Richard Webb (KAR Enterprises), "Is It Worth It? The Economics of Reusable Space Transportation," ICEAA International Training Symposium, Bristol, October 2016, written under the direction of NASA Marshall Space Flight Center's Engineering Cost Office. Ten months after the first landing, with SpaceX at "eleven attempts, six successes," Webb's answer to his own title question was: "IT DEPENDS!" — on stakeholder motivation, demand elasticity, degree of reuse, fleet sizing and vehicle size. The paper explicitly grants that "the technical feasibility of at least some degree of reusability has been demonstrated."

Grade: Not a miss. It named the correct variables. Its most useful single finding, which nobody quotes, is about attrition: a 5% recovery attrition rate reduces a booster's effective life to about 20 flights per year regardless of whether the design life is 100, 200 or 500 flights. SpaceX's landing success rate is now 632 of 643 (98.3%) — an attrition rate of about 1.7% — and that number, not the landing itself, is what makes the fleet economics work. Webb's framework predicts SpaceX's outcome correctly if you feed it SpaceX's attrition and flight rate. What the skeptics got wrong was not the model. It was treating flight rate as exogenous.

5. Commonly misused as

Not required for moment. It is where most of this entry's practical value sits, because the version of this story in circulation is wrong in four specific ways and each one produces a bad AI argument.

5.1 — "Reuse collapsed the cost of space by 20×," shown as a chart. The familiar plot runs from Shuttle-era figures (~$54,000/kg) down to Falcon 9 (~$3,000/kg) and is used to attribute the whole drop to reuse. Most of that drop happened before a single booster landed. Falcon 9 v1.0, fully expendable, flying in 2010–2013, was already around $6,000/kg at list — roughly a 9× improvement over the Shuttle with zero reuse, produced by vertical integration, cheap engines built in quantity, and competing against cost-plus incumbents. The improvement attributable to the period after reflight became routine is the 2× to 2.7× in section 4.1, and much of that is a larger rocket. Anyone using the chart to size a reuse effect is crediting reuse with a curve that manufacturing produced. (moore-1965 is the canon's proper home for this class of error and is unwritten; this paragraph is standing in for it.)

5.2 — "The experts said it was impossible." They mostly did not, and this is the misuse that matters most for AI, because the same move is made constantly against safety researchers and against anyone who models a cost curve. Webb's paper — the most careful statement of the skeptical case, written for NASA's own cost office — explicitly granted technical feasibility and concluded "it depends." Israël's claim was conditional and both of its conditions held. Hertzfeld and colleagues were right about elasticity. The one flat "no scenario" claim, Book's, is a genuine miss, and I have it only second-hand.

What the consensus actually got wrong was one modelling assumption: that the flight rate was a property of the market rather than something a supplier could manufacture. That is a real and serious error, and it is a much more useful thing to carry than "experts are wrong about what's impossible" — because the AI version of the error is not "they said it couldn't be done," it is "they forecast demand from the existing market." This is the sharp end of the lighthill-1973 pairing the proposal asked for: Lighthill was right about the field he was surveying and wrong about the century; the launch skeptics were right about the market they were surveying and wrong about who would be in it. In both cases the analysis was sound and the boundary of the system was drawn too small. That is one failure mode, seen twice, and it is worth more to a reading than two anecdotes about consensus.

5.3 — "Costs fell, therefore prices fell, therefore users benefited." The most consequential misuse and the one a reading will meet most often, dressed as "inference costs are dropping 10× a year." Fifteen years, 632 landings, and every published SpaceX price is higher in nominal dollars than before reuse existed — Falcon 9 $54M (2012) → $74M (2026); rideshare $5,000/kg (2019) → $7,000/kg (2026); both roughly flat in real terms. NASA's average of about $106 million per Falcon 9 sits against an internal cost of $15–28 million. The cost curve was real and it went almost entirely to the producer, who spent it on Starlink and Starship rather than on price cuts — which is a defensible use of it, and is not the same as the customer benefiting. Pass-through requires a counterparty with leverage, and where one existed (the Space Force on GPS III) the saving showed up as ~$32 million a launch. A reading should treat "cost" claims and "price" claims as different claims and say which one a source is making.

5.4 — "SpaceX proved the demand was there." It did the opposite. The demand was not there; 69% self-provisioned, 30 launches without Starlink. Using this story to argue that supply creates its own demand skips the actual mechanism, which was vertical integration financed by a decade of patient capital inside a private company answerable to essentially one shareholder. That is a live and precise cite against any AI buildout whose demand case is "the applications will come" — it did come, once, for a firm with that exact shape.

6. Where this file could be wrong

Sourcing. Two of my four skeptic quotations (Book; Israël) come through Webb's Figure 1 rather than from the originals, which I could not reach. The Hertzfeld/Williamson/Peter elasticity quotation likewise comes through Webb. The Webb paper itself I read in full from NASA's technical report server.

SpaceX is private and its cost figures are estimates. The $15–28 million internal-cost range and the ~$12 million upper-stage figure are third-party estimates and claimant statements, not audited numbers. The $300,000 refurbishment figure that circulates in 2026 coverage conflicts with Musk's own 2020 figure of about $1 million, and I have used the higher, sourced one. The list prices, the launch and landing counts, the NASA average, the GPS III saving and the cadence figures are public and I regard them as solid.

The $/kg comparisons in 4.1 and 5.1 are list price at maximum payload, which is a configuration almost nobody buys. They are the fairest like-for-like I could construct across fourteen years and they overstate the improvement for a typical customer, who is not filling the fairing.

The inflation adjustments are mine, using CPI-U annual figures of 229.594 (2012), 255.657 (2019) and 333.918 (2026). Anyone preferring a different deflator will get a different number; the direction is not sensitive to the choice.

Where I am least confident: section 4.8's three-failure-mode taxonomy is my inference from two files, not a documented finding. It is labelled as judgment. Section 5.2's claim that the consensus erred by treating flight rate as exogenous is also mine — Webb does not say it, Gatti and D'Ottavio do not quite say it, and it is the load-bearing idea in this entry. If it is wrong, section 5.2 goes with it.

Rule 7 disclosure. The corruption risk in this file runs the same way musk-robotaxi-2019 describes: the subject is not one this project is tempted to flatter, and the cheap grade is the hostile one. There is a second and opposite pressure here, named in tesla-volume-production: an entry commissioned to supply the canon with a hit is under structural pressure to find one. I have tried to answer both by making every grade arithmetic where arithmetic exists — launches, landings, list prices, CPI, flight counts — and by labelling the rest. The finding is that this is a hit, and that it is a smaller and stranger hit than the story in circulation. This file deposits nothing in the evidence ledger, writes no digest, touches no other entry, places no needle, no score and no landmark.

7. Sources

Primary and near-primary

Independent analysis

Record and figures

Not reached. PARAMETRIC WORLD, Winter 2012 (Book); Stephen Book's STAIF-99 paper in the original; the SpaceNews cost-savings article directly (HTTP 403); SpaceX's internal cost accounts, which are not public.