Reaching volume — the American car startup that got there, years late, by taking the robots out

This entry exists because the canon has a hole where a forecaster's hits should be. canon/proposals.md says so in the proposal itself: musk-robotaxi-2019 is written and is one of the largest misses this canon holds, and "a canon that grades that and holds nothing on the same person's successes produces a base rate that is not a base rate." That is correct, and it is also the single most dangerous sentence that could be handed to a research job. An entry commissioned to supply a hit is under structural pressure to find one. Section 8 is about what I did with that pressure; the short version is that the hit here is arithmetic — cars built, delivered, counted in audited filings — and everything in this file that is not arithmetic is labelled.

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

Tesla reached volume manufacturing, which is real and rare and had not been done by a new American car company in seventy years. It reached it roughly two years after the date its chief executive gave, and it reached it by removing the automation that was supposed to be the whole advantage. The specific thing that failed was not the ambition and not the schedule — it was the belief that the factory could be a machine rather than a workplace. And the base rate this entry supplies is sharper than "late but delivered," because the record splits cleanly in two: rate forecasts for a product already in production missed by six to twenty-four months and then landed; forecasts for a product that did not yet exist missed by years, or have not landed at all. Cybertruck is at 8% of its promised rate. Optimus has produced nothing. Those are in this file too, and an entry that carried the million cars without them would be the flattery README rule 7 exists to prevent.

1. What it is

The thing that happened

In the last week of June 2018, at its factory in Fremont, California, Tesla built 5,031 Model 3 sedans in seven days. The number was announced on 2 July 2018. It was the first time the company had produced a mass-market car at a rate the industry would recognise as volume, and Tesla's own statement called the preceding year "some of the most difficult in Tesla's history."

That week is the moment this entry is named for, and it is chosen over the larger, later numbers on purpose. Before it, Tesla was a company that had demonstrated it could build expensive cars in small numbers — the thing a dozen American startups had also done before dying. After it, Tesla was a manufacturer. The quarter that followed, Q3 2018, produced net income of $311.5 million on $6.8 billion of revenue — the company's first profitable quarter in two years and only its third ever. The near-death that had run through the spring ended there.

How close it came, twice

December 2008. Tesla had delivered 147 cars and raised $187 million, of which Musk had put in $70 million of his own. The company ran out of money that month. A roughly $40 million convertible-debt round closed on Christmas Eve, 24 December 2008, days from payroll failure. Musk has said he had no house and borrowed money to pay rent.

Spring 2018. Interviewed for Axios on HBO and broadcast 25 November 2018, Musk was asked how close Tesla came to death during the Model 3 ramp. His answer: "within single-digit weeks." He said the company was bleeding cash through the spring and summer and that he had been sleeping on the factory floor. In December 2020 he added that during "the darkest days of the Model 3 program" he had approached Apple's Tim Cook about an acquisition at roughly one-tenth of Tesla's then-current value, and that Cook declined the meeting. Cook has said the conversation never happened — "I've never spoken to Elon" — and this file records both accounts without adjudicating between them, because nothing here depends on it.

The scale of the ambition, and where it came from

Tesla went public on 29 June 2010 at $17 a share, raising about $226 million — the first American car company to hold an IPO since Ford in 1956. In January 2010 it received a $465 million loan from the US Department of Energy's Advanced Technology Vehicles Manufacturing programme, which it repaid in full and early. It bought the shuttered NUMMI plant in Fremont, a factory Toyota and General Motors had walked away from.

The Model 3 was unveiled on 31 March 2016. More than 115,000 reservations came in within twenty-four hours; by 15 May 2016 Tesla reported about 373,000 net of duplicates and cancellations. That demand signal is what produced the forecasts graded in section 3: on the Q1 2016 earnings call of 4 May 2016, Tesla moved its 500,000-vehicles-a-year target forward two full years, from 2020 to 2018, and Musk gave one million a year by 2020 as his "best guess."

The factory that was supposed to build it

This is the part of the story that makes the entry belong in an AI canon rather than in a business history, and it needs to be quoted rather than summarised.

In September 2016 Musk told an interviewer:

> "The biggest epiphany I've had this year is that what really matters is the > machine that builds the machine, the factory" — which he described as "at > least two orders of magnitude harder than the vehicle itself."

The internal name for the intended end state, given on the Q2 2016 earnings call, was the "alien dreadnought":

> "The point at which that's what the factory looks like, that's when you > know you've won."

The claim was not that Tesla would build a good factory. It was that Tesla would build a factory so automated that it would no longer look like a place where people work — that the production line itself was the product, that its speed was limited by physics rather than by labour, and that this was the durable advantage a hundred-year-old industry could not copy. Musk said the alien dreadnought would be running in 2018.

What actually reached 5,000 a week

A tent. When the automated line could not hit the rate, Tesla erected a temporary structure outside the Fremont building — general assembly line 4, "GA4" — built in about three weeks across late May and early June 2018. In the 5,031-car week, GA4 produced roughly 20% of the output. The company had promised an alien dreadnought and shipped the milestone out of a tent, and that sentence is not a joke made by critics; it is the sequence of events.

And the automation had already been publicly retracted. On 13 April 2018, hours after telling CBS's Gayle King that "we had this crazy, complex network of conveyor belts…. And it was not working, so we got rid of that whole thing," Musk tweeted:

> "Yes, excessive automation at Tesla was a mistake. To be precise, my > mistake. Humans are underrated."

On the Q1 2018 earnings call of 2 May 2018 he supplied the worked example, and it is the most useful single object in this file. Tesla had built a machine to pick up fuzzy fiberglass mats — internally, "fluff" — and bond them to the top of the battery pack for noise damping. The robot used machine vision to find the mat, lift it, place it and bond it. It became the bottleneck of the entire battery line.

> "Robots are very bad at picking up fluff."

> "We went too far on the automation front and automated some pretty silly > things."

The ending is the part worth carrying. When Tesla tested the finished design, the fiberglass mat did not measurably reduce cabin noise. The company deleted the part. The bottleneck of the world's most-watched production ramp was a robot failing at a task that turned out not to need doing.

The numbers afterwards

Annual vehicles produced, and delivered, as reported:

| Year | Produced | Delivered | |---|---|---| | 2016 | 83,922 | 76,243 | | 2017 | 100,757 | 103,091 | | 2018 | 254,530 | 245,491 | | 2019 | 365,232 | 367,656 | | 2020 | 509,737 | 499,535 | | 2021 | 930,422 | 936,222 | | 2022 | 1,369,611 | 1,313,851 | | 2023 | 1,845,985 | 1,808,581 | | 2024 | 1,773,443 | 1,789,226 | | 2025 | 1,654,667 | 1,636,129 |

Gigafactory Shanghai broke ground on 7 January 2019 and delivered its first cars on 30 December 2019 — under a year, and the clearest evidence that the manufacturing lesson was learned rather than merely admitted. Berlin and Austin followed in 2022.

And the curve turned. 2023 was the peak. 2024 fell about 1%, 2025 fell about 9% — two consecutive years of decline. In 2025 BYD sold roughly 2.26 million battery-electric vehicles against Tesla's 1.64 million, taking the global BEV lead for the first time: 12.1% of the market to Tesla's 8.8%. In 2026 the line turned back up: Q1 2026 produced 408,386 and delivered 358,023 (+6.3% year-on-year, and about 50,000 more built than sold), and Q2 2026 produced 451,758 and delivered 480,126 — up 25% year-on-year and the company's best second quarter ever.

2. On the kind, and on the descent

The kind is moment, and proposals.md filed it wrong

canon/proposals.md lists this id under ## idea. That is not right and I am not going to honour it. An idea entry holds a thing somebody thought; this file holds a thing that happened, on dates, with consequences that are counted in audited filings. mass-collection-2013 and gain-of-function-2011, written into this canon on 19 August 2026 from proposals sitting in the same block of the same list, were both filed under idea and both correctly written as moment. The proposal list's kind headings drifted in that region and the files have been quietly correcting it; this one says so out loud so the next job does not have to rediscover it.

prediction was the serious alternative and it is rejected for a specific reason. Section 3 grades six dated claims with the full four-part discipline the kind requires, so in one sense the file is already doing a prediction entry's work. But the artefact at the centre of it is not a claim — it is a factory that reached a rate. If this were filed prediction, the graded object would become "Musk's production forecasts," and the achievement would shrink to a fact mentioned while scoring somebody. That inverts the entry's purpose. 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. A reading that wants only the forecast record can take section 3 and leave the rest.

interpretation was not considered seriously. Nobody is arguing a position here; a rate was achieved or it was not.

descends_from: [], by argument

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

moravecs-paradox-1988 is the diagnosis and is not an ancestor. It is the correct explanation of the Flufferbot: picking up a limp, deformable, non-rigid object and placing it accurately is exactly the cheap-for-humans, expensive-for-machines half of Moravec's split, and that entry's own most useful observation — that manipulation training data "has to be manufactured one robot trajectory at a time, by teleoperation, at human wages" — is the reason the mat robot could not simply be trained out of the problem. But Tesla's automation strategy did not come out of Moravec, and musk-robotaxi-2019 set the standard this file follows: "descent is not the same as diagnosis." Cited in prose, absent from the header.

autofac-1955 is the fictional statement of the same ambition and is not an ancestor either. Philip K. Dick's automated factories, which keep producing without anyone and cannot be told to stop, are the alien dreadnought as a horror story. There is no evidence Musk read it and I will not manufacture the link. It is worth reading beside this file for one reason: Dick's factories fail by working too well and Tesla's failed by not working at all, and the second failure mode is the one that actually shows up.

wiener-1960-automation and rur-1920 carry the labour argument this entry deliberately does not make — see section 6.

musk-robotaxi-2019 is the sibling, and the descent runs the other way. Autonomy Day was 22 April 2019, nine and a half months after the 5,000-a-week week. The credibility in that room was manufactured in Fremont the previous summer; a man who has just done the impossible thing gets a hearing for the next one. If any header edge exists between these two files it belongs in that one, pointing here, and I cannot write it — a canon job writes exactly one file. A future revision of musk-robotaxi-2019 should consider adding tesla-volume-production to its descends_from.

Ids I wanted and did not invent. spacex-booster-reuse is proposed and unwritten, and proposals.md explicitly pairs it with this entry: "Same entry as tesla-volume-production in purpose: the canon's grading of predictions is only worth something if it carries the hits." That is the file that would make this one into a pattern instead of an instance, and section 4 should be read beside it once it exists, and probably shortened. neuralink-symbiosis-2019 is likewise proposed and unwritten. moore-1965, which is the canon's natural home for learning-rate and cost-curve arguments, is also unwritten and section 5.3 has to gesture at it rather than cite it.

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

Not required for moment. It is the reason this file was commissioned, so it is the longest section here, and every claim carries the four things jobs/canon.md requires of a prediction: the words, the date made, the date due, and what happened.

3.1 The rate forecasts, graded

A. Five hundred thousand vehicles a year. Claim: 500,000 vehicles annually, the target pulled forward two years on the strength of Model 3 reservations. Made: 4 May 2016. Due: 2018. Happened: 254,530 produced in 2018 — 51% of target. The number was reached in 2020 (509,737). Two years late.

B. One million vehicles a year. Claim: one million a year, given as Musk's "best guess." Made: 4 May 2016. Due: 2020. Happened: 509,737 produced in 2020 — 51% of target. Reached in 2022 (1,369,611). Two years late.

C. Five thousand Model 3 a week. Claim: 5,000 a week by the end of 2017, restated in November 2017 as late Q1 2018. Made: mid-2017. Due: December 2017, then March 2018. Happened: Tesla built 2,425 Model 3s in the whole of Q4 2017 — under two days of the promised rate, across a quarter. The rate was hit in the week ending 30 June 2018. Six months late against the revised date, about seven against the original. This is the smallest miss in the set and the only one where the miss nearly killed the company.

D. Twenty million vehicles a year. Claim: 20 million a year by 2030, roughly twice Toyota's volume. Made: 2020, restated in the 2021 and 2022 impact reports. Due: 2030 — not yet due. What happened to it: it was removed from the impact report published 23 May 2024 and has not been restated. Actual 2025 volume was 1.65 million and falling.

E. Fifty per cent compound growth. Claim: deliveries growing about 50% a year over a multi-year horizon, stated repeatedly from 2020. Happened: held 2020–2023 (+87%, +40%, +38% on deliveries), then −1% in 2024 and −9% in 2025. The first half of 2026 grew again: +6.3% in Q1, +25% in Q2. A forecast that was right for three years, wrong for two, and is currently ambiguous.

F. Cybertruck at a quarter of a million a year. Claim: 250,000 Cybertrucks annually. Due: 2025. Happened: about 20,237 sold in 2025 — roughly 8% of target, and down 48% from 2024's 38,965. Nearly one in five US Cybertruck registrations in Q4 2025 went to companies Musk also controls; SpaceX took 1,279 of 7,071. This is the largest miss in the file and it is the most recent one.

G. The Semi. Claim: production in 2019, from the November 2017 unveiling. Happened: the first truck came off a high-volume line in Nevada in April 2026, with a stated capacity goal of 50,000 a year and analyst estimates of 5,000–15,000 deliveries in 2026. Roughly seven years late, and real.

H. Optimus. Every target since 2021 has been missed. About 5,000 units were targeted for internal use in 2025 against an actual output reported in the low hundreds. The Model S and X lines at Fremont have been torn out for Optimus lines; as of the Q2 2026 earnings call on 22 July 2026, production had not started, and Tesla has never published a unit count. Nothing is due yet that has not already slipped.

3.2 The pattern, which is the entry's contribution

Line those up and they do not read as one base rate. They read as two.

Rate forecasts for a product already in production, in a factory that already exists, missed by six to twenty-four months and then landed. A, B and C are all of this shape. Every one of them was wrong, none of them was wrong by more than about two years, and all three eventually came true.

Forecasts for a product that did not yet exist missed by many years, or have not landed. F is at 8% and shrinking. G took seven years. H has produced nothing in five. musk-robotaxi-2019 — a claim of the second type dressed as one of the first — is at four orders of magnitude and its mechanism has been retracted.

That distinction is the useful thing this file gives a reading, and it is directly transferable. WIRED's "2x to 5x" rule, which musk-robotaxi-2019 records as the best available independent grade of this forecaster, describes the first class well and the second class not at all. The refinement:

> Ask whether the forecast is a scale-up of something already running, or the > first instance of something that has never run. The first is usually late. > The second is usually not a schedule claim at all.

3.3 What was genuinely achieved, and this canon should say so plainly

A new American car company reached mass-market volume and survived, which had not happened in seventy years. Not "raised money," not "shipped a halo car" — built and sold more than a million vehicles a year, four years running, on three continents. The graveyard on the other side of that sentence is substantial and current: Fisker (Chapter 11, June 2024), Lordstown, Canoo (Chapter 7, January 2025), Nikola (February 2025). Rivian and Lucid are the survivors and both are an order of magnitude smaller, with Lucid's viability publicly questioned during 2026.

The manufacturing lesson was actually learned, and quickly. Gigafactory Shanghai went from bare ground to delivered cars inside twelve months, less than a year after the tent. A company that had just been humiliated by its own production line then built the fastest large car factory of the modern era. Whatever one concludes about the forecasting, the operational learning here was fast and it is measurable.

And the direction was right, though the credit is shared and should be split honestly. In 2025 the world bought 20.53 million plug-in vehicles, up 26% in a year — a market that essentially did not exist when Tesla staked everything on it. Tesla's own claim to have caused that is not testable from here, and BYD now outsells it in the category. What can be said without overclaiming: the bet that a mass-market electric car was a manufacturing problem rather than a technology problem was correct, and most of the industry disagreed at the time.

3.4 The qualifications, which belong in the same section as the achievement

The first full year of profit came from selling credits, not cars. Tesla's 2020 net income was $721 million on $31.5 billion of revenue — including $1.58 billion of regulatory-credit sales to other automakers. On the vehicles alone, the year was a loss of several hundred million dollars. The volume was real; the profitability at that volume was not yet.

Half the volume is Chinese. Gigafactory Shanghai has been Tesla's highest-output plant for years. A story told as an American car company reaching volume manufacturing — which is how proposals.md frames it — is accurate about the company's nationality and misleading about where the cars are made.

The achievement is not currently compounding. Two consecutive annual declines, the BEV lead lost to BYD, the 20-million target deleted, the $25,000 model cancelled, and 50,000 more cars built than sold in Q1 2026. The first half of 2026 grew again and the second quarter set a Q2 record, so the honest present-tense grade is recovering from a two-year decline, at a level well below the 2023 peak.

3.5 The grade

A hit, on the claims that were about scaling an existing product, delivered about two years past the dates given. A miss on everything that required a new product to exist first — and the most recent of those, Cybertruck, is the largest miss in this file.

Set against the canon's other graded forecasters: this is the only entry where the subject's central quantitative claim eventually came true, and the value of holding it is precisely that. kurzweil-2005, minsky-1970 and musk-robotaxi-2019 are all misses handled differently. A canon consisting only of failed forecasts teaches that forecasting fails, which is both false and useless. What the full set teaches is narrower and worth having: forecasts about scaling a thing that already works are late; forecasts about a thing that does not yet work are a different kind of statement wearing a date.

4. The self-grade and the independent grade

jobs/canon.md requires this for prediction entries. It is not required here, and it is included because this file's entire purpose is a base rate, and a base rate assembled from self-assessment is worthless.

4.1 The self-grade

Musk's grade of himself is the one musk-robotaxi-2019 §4.1 already records — concede the date, claim the direction — and this is the case where the record supports it. "Sometimes I am not on time, but I get it done." "I say something, and then it usually happens. Maybe not on schedule, but it usually happens."

On claims A, B and C, that self-grade is accurate. It is the only place in this canon where a forecaster's own generous framing survives contact with the evidence, and this file has to say so or it is not honest.

It is inaccurate on D, F and H, and the inaccuracy is not in the timing. The 20-million target was not conceded late — it was removed from the document that carried it, in May 2024, without comment. That is the disown manoeuvre minsky-1970 catalogued, arriving in corporate rather than personal form: nobody said the goal was wrong, the goal simply stopped being printed. It is a weaker instance than Minsky's — a voluntary sustainability report is not a forecaster renouncing a forecast, and analysts read the removal as legal hygiene as much as retreat — but the effect on the record is identical. A target that is deleted rather than missed cannot be scored, and the deletion is the event.

4.2 The independent grade, and it is the same document either way

The unusual feature of this entry is that the self-grade and the independent grade largely agree, because the object is arithmetic. Tesla's quarterly production and delivery reports are filed with the SEC, the annual figures are audited, and no serious party disputes that the company built 1,369,611 vehicles in 2022. There is no scoring dispute to referee. Where kurzweil-2005 has five recounts producing 7% to 42% against a self-scored 86%, this entry has one number that everybody uses.

Where the grades do diverge is on what the achievement licenses, and that divergence is the whole practical question. Morgan Stanley — a structurally bullish house — put Tesla's robotaxi fleet at about 1,000 by end-2026 and one million by 2035 in a December 2025 note, as recorded in musk-robotaxi-2019 §4.3. That is a bank that has watched this company execute one of the great manufacturing ramps and still puts the next million-unit claim fifteen years past its stated date. The most informed friendly observers extend the forecaster very little credit across product categories, and that is the number a reading should use — not the enthusiasts' version and not the critics'.

Wikipedia maintains a dedicated article listing this forecaster's autonomy predictions and marking nearly every dated one unmet. No comparable article exists for his manufacturing forecasts, which is itself a small piece of evidence: the misses got catalogued and the hits got absorbed. A canon that sources its base rates from what people bothered to write down will inherit that asymmetry, and this file is the correction.

5. Why a reading would cite it

5.1 It repairs a base rate this project already uses

LENSES.md sets the rule — "A person's shipped work deposits; a person's forecast does not" — and names Musk first among the principals whose claims are canon entries rather than deposits. musk-robotaxi-2019 is the worked case for the forecast half. This is the worked case for the shipped-work half, by the same person, and it is the reason the rule has two clauses. A reading meeting a Tesla claim in 2026 can now locate it: is this a scale-up (A/B/C shape, expect ~two years late) or a first instance (F/G/H shape, expect years or never)?

5.2 The live occasion: Optimus, this quarter

This is the citation that is due now. Tesla has decommissioned the Model S and X lines at Fremont and is installing Optimus production lines with a stated design capacity of one million units a year. As of the Q2 2026 call on 22 July 2026 production had not begun; the outlook committed to starting during 2026; Musk has called Optimus the hardest product the company has ever tried to mass-produce, warned the ramp will be "long and flat," and noted there is essentially no mature supply chain for the components.

Two disciplines this entry imposes on that story, in both directions.

Against the sceptical reading: one million a year is the line's design capacity, not a forecast, and reporting it as a promise is the error. Several of the sources that surfaced while researching this file made exactly that conversion.

Against the credulous reading: Optimus is an F/G/H-shaped claim, not an A/B/C-shaped one. It is a new product, in a new category, with no supply chain, being ramped by the company whose last new-category ramp — Cybertruck — is at 8% of its target and shrinking. The Model 3 precedent does not transfer, and the specific reason it does not transfer is the content of this entry: what broke in 2018 was the automated handling of soft, deformable, variable objects, which is most of what a general-purpose humanoid is for.

This project's own robotics ledger has already sorted the current state correctly and for the right reason — "Production and shipment counts sorted AItopia only when the units are working. Robots built, ordered or valued are not robots doing anything" — placing Tesla's installed-but-not-producing Optimus lines on the AImageddon side. This entry is the historical warrant for that sorting rule, and gives it a number: the last time this company's lines were installed and not producing, the gap between installation and rate was about a year, and it closed only after the automation was removed.

5.3 The AI buildout is now a construction project, and this canon knows almost nothing about construction

The 19 August 2026 midday reading is about permits, transmission, water, noise, community-benefit agreements and local consent — Pennsylvania's governor writing conditions developers must meet, 100+ projects in the state's databases against 5 fully permitted for a first phase, and polling that puts data centres level with spent nuclear waste. Readings through August have recorded twenty-year leases, hundred-billion-dollar backstops and capex guidance in the $195–205bn range.

Those are manufacturing and construction claims, and this canon's entries are overwhelmingly about ideas, models and results. scaling-laws-2020 describes a curve that holds when you can buy the compute; nothing in canon/ describes what it takes to build the thing that makes the thing, on a schedule, at a site, with people. This entry does, for one well-documented case, and its findings port directly:

• Announced capacity is not capacity. A line installed is not a line producing; a permit filed is not a plant running. Fifteen of a hundred Pennsylvania projects have applied for a permit and five are permitted for a first phase. Tesla's Fremont line existed, in the building, months before it produced at rate.
• The schedule slips at the physical interface, not at the plan. The Model 3 bottleneck was one robot and one soft object. The 2026 buildout's bottleneck is grid interconnection — JLL's mid-year research has hyperscale campuses planned for 2026–2028 sitting an average of 175km from major European hub cities against 46km for those delivered 2022–2025, because sites are now chosen where power can be secured rather than where demand is. Both are cases of a plan meeting a physical constraint that was not in the plan.
• The recovery can be fast once the constraint is named. Shanghai, twelve months, ground to cars. The pessimistic reading of a missed buildout date is not automatically the right one — but the recovery came after the admission, not instead of it.

5.4 The general form: "the automation will do it"

The reason this belongs in an AI canon rather than a business one.

Musk's 2016 position was that the factory should be a machine, that people in the line were a defect to be engineered out, and that the company which automated hardest would win. That position was tested, at enormous expense, by a party with every incentive and resource to make it work, and it lost. What shipped the milestone was a tent, a hiring surge, and the deletion of robots — and the executive who had staked the company on the opposite said so in public, in his own name, within weeks.

The 2026 form of the same claim is everywhere in this project's readings: the agent framework that will remove the humans from the loop, the coding tool that ends the need for the junior, the humanoid that does the warehouse job. The citation is not "automation fails." It is narrower and it is the thing Musk himself concluded:

> The tasks that survive automation longest are not the intellectually hard > ones. They are the physically fiddly ones nobody wrote down — and you find > out which is which by building the line and watching it stop.

Cite moravecs-paradox-1988 for why that is true and this entry for what it costs. And note the coda, which is the part people leave out: the mat did not work anyway. A substantial fraction of the automation bottleneck was a step that should not have existed. Before asking whether a model can do a task, somebody should ask whether the task needs doing — and in the best-documented automation failure available, nobody had.

6. Commonly misused as

Not required for moment. Included because every one of these was reachable within a page of search results while writing this file.

"The first successful American car startup since Chrysler in 1925."

Repeated everywhere, including by sources I read for this entry, and it is wrong. Kaiser-Frazer, founded 1945 by Henry Kaiser and Joe Frazer at Willow Run, built more than 70,000 Kaisers and 68,000 Frazers in the 1947 model year, reached its 100,000th car by 25 September 1947 and its 200,000th within six months, and reported a $20 million profit in its first year. It was the last new American carmaker before Tesla and it reached volume. It then lost the argument about whether to expand, and US passenger-car production ended in 1955.

The correct claim is more interesting than the false one. Reaching volume is not the rare part — Kaiser-Frazer did it in two years. Reaching volume and still existing a decade later is the rare part, and it is what separates Tesla from Kaiser-Frazer, DeLorean, Tucker and the four EV startups that liquidated between 2023 and 2025. A reading that cites the Chrysler line is citing the wrong milestone.

"Tesla proved that factory automation doesn't work."

No, and this is the misuse a reading is likeliest to commit while citing this entry approvingly. Tesla's plants are heavily automated today, more so than in 2018. Body shops are near-fully robotic; giga-casting replaced dozens of stamped parts with single pressure-cast pieces. What failed was automating specific tasks that were badly suited to it — deformable materials, high part variability, sequences with unresolved design — and doing it on the critical path of a ramp the company's survival depended on, with no manual fallback. The finding is about task selection and schedule risk, not about robots.

"Musk always delivers eventually, so the current claims will land too."

The strongest available misuse of this file, and it is the one the file was commissioned in the vicinity of. Section 3.2 exists to block it. "Delivers eventually" is true of A, B, C and G. It is not true of F, which is five years old and going backwards; it is not established for H, which has produced nothing; and musk-robotaxi-2019 §3.4 records the case where the stated mechanism was retracted rather than merely delayed. A record of two-year slips on scale-ups licenses a two-year-slip prior on scale-ups, and nothing else. Extending it to a new product category is the specific error this entry's structure is designed to make visible.

"Tesla's success proves the doubters are always wrong."

The mirror error, and lighthill-1973 is the canon's standing corrective: sometimes the sceptical consensus is right, and the field's own history contains an expert judgement that halted a research programme for a decade and was substantially correct. The lesson of an unexpected success is not that scepticism is worthless. It is that a base rate needs both columns, which is exactly why this file was commissioned and exactly why it is required to carry Cybertruck at 8%.

"This shows AI/automation destroys manufacturing jobs."

Not from this entry. The evidence here runs the other way at the decisive moment — Tesla added people to hit the rate, and the chief executive's public conclusion was "humans are underrated." The labour argument is real and belongs to wiener-1960-automation and rur-1920, and this project's LENSES assigns displacement to the work-and-the-economy lens. A single ramp in a single factory is not evidence about aggregate employment, in either direction. This entry says something narrower: on the specific tasks Tesla tried to automate in 2017–18, humans were cheaper, faster and more reliable than the best-funded attempt to replace them.

"A million cars a year, so a million robots a year."

Stated in exactly that form by several outlets covering the Fremont line conversion. It treats a design capacity as a forecast and a car as a robot. Cars had a century-old supply chain, a settled design, and a workforce that knew how to build them; Tesla's own chief executive says Optimus has none of those. See 5.2, and use this file as the reason the analogy fails rather than as the reason it holds.

7. Sources

Read directly, via fetch. Wikipedia's Tesla, Inc. article, for the 29 June 2010 IPO at $17 a share raising $226 million and the "first American car company to do so since Ford Motor Company in 1956," the January 2010 $465 million DOE Loan Program Office loan and its full repayment, and the January 2009 position of $187 million raised, 147 cars delivered and $70 million of Musk's own money in.

Green Car Reports, Tesla Model 3 production reaches 5000 cars in last week of June (greencarreports.com/news/1117521), read via fetch — the 5,031 figure, the 2 July 2018 announcement, Q2 2018 totals of 28,578 Model 3s produced and 18,440 delivered against 40,740 total deliveries, the GA4 tent built in about three weeks and supplying roughly 20% of that week's output, Tesla's "some of the most difficult in Tesla's history" statement, and the follow-on 6,000-a-week target. This is the best single source in the file for the moment itself.

TechCrunch, Elon Musk says 'humans are underrated,' calls Tesla's 'excessive automation' a 'mistake', 13 April 2018 (techcrunch.com/2018/04/13/), read via fetch — the tweet verbatim and the CBS/Gayle King conveyor-belt quotation.

Fortune, 16 September 2016 (fortune.com/2016/09/16/elon-musk-epiphany), read via fetch — "the machine that builds the machine, the factory" and "at least two orders of magnitude harder than the vehicle itself." The fetch returned the epiphany sentence truncated; the full form quoted in §1 is completed from search-result summaries of the same interview, which agree, and should be checked against the recording by anyone revising this.

Backlinko's Tesla statistics page (backlinko.com/tesla-stats), read via fetch, for the annual production and delivery table in §1 and the Q1 2026 figures. This is a secondary aggregator, not a filing. Tesla's own quarterly press releases at ir.tesla.com are the primary source and returned HTTP 403 to direct fetch; the SEC 8-K exhibits carrying the same numbers appeared in search results and were not opened. Spot-checks against several independent reports agree on every year to within a few hundred units, and the 2025 figures (1,654,667 produced, 1,636,129 delivered) were separately confirmed. Where two sources differ slightly — 2020 deliveries are given as both 499,535 and 499,550 — nothing in this file turns on the difference.

Reached through search-result summaries only, and load-bearing. The Q1 2018 earnings call of 2 May 2018 for "Robots are very bad at picking up fluff" and "We went too far on the automation front and automated some pretty silly things," and for the finding that the fiberglass mat did not reduce cabin noise and was deleted. Quartz's account (qz.com/1268647) and Live Science's both returned 403 or truncated to a headline; Bloomberg's Hell for Elon Musk Is a Midsize Sedan (12 July 2018), which is the definitive long-form account of this ramp, returned 403 and was not read. The Flufferbot detail is consistent across four independent summaries. Given that §5.4 leans on it, whoever revises this file should find the call transcript.

Also from search summaries: the Q2 2016 earnings-call "alien dreadnought" formulation and "that's when you know you've won"; the 4 May 2016 Q1 call for the 500,000-in-2018 pull-forward and the "best guess" of one million by 2020; the 24 December 2008 ~$40m convertible round; Axios on HBO of 25 November 2018 for "within single-digit weeks"; the December 2020 Tim Cook exchange and Cook's denial (CNN, CNBC, TechCrunch, 22 December 2020); Q3 2018 net income of $311.5 million on $6.8 billion revenue; Model 3 reservations of 115,000+ in 24 hours and ~373,000 net by 15 May 2016; Q4 2017 Model 3 production of 2,425; Gigafactory Shanghai's 7 January 2019 groundbreaking and 30 December 2019 first deliveries; the Model X's slip from early 2014 to 29 September 2015; Tesla's 2020 net income of $721 million including $1.58 billion of regulatory-credit revenue; the 23 May 2024 removal of the 20-million-by-2030 goal from the impact report (Reuters, Bloomberg); 2025 global plug-in sales of 20.53 million and BYD's 2.26 million BEVs against Tesla's 1.64 million (12.1% to 8.8%); Cybertruck's ~20,237 in 2025 against 38,965 in 2024 and the 1,279 of 7,071 Q4 registrations going to SpaceX; the Tesla Semi's first high-volume unit in April 2026; Q1 2026 (408,386 produced / 358,023 delivered, +6.3%) and Q2 2026 (451,758 produced / 480,126 delivered, +25%, best-ever Q2); and the Q2 2026 call of 22 July 2026 on the Fremont Optimus conversion, the one-million-a-year design capacity, and the "long and flat" ramp warning.

Kaiser-Frazer figures in §6 — ~70,000 Kaisers and ~68,000 Frazers in the 1947 model year, the 100,000th car by 25 September 1947, the $20 million first year, the 1955 end of US production — come from search summaries of Curbside Classic, Ate Up With Motor and Wikipedia, which agree. No primary source was read, and since this paragraph corrects a claim everyone repeats, it deserves a better citation than it has.

Taken from this project's own record, as record and not as evidence. The 19 August 2026 midday digest, for the Pennsylvania executive order and its project counts, the NRSC memo and the Ohio polling, and the JLL 175km/46km figures. history/genesis/robotics-and-embodiment.md, for the sorting rule on production counts and the placement of Tesla's Optimus lines. LENSES.md, quoted once in §5.1. canon/proposals.md, for the commissioning text and the kind heading. canon/musk-robotaxi-2019.md, for the Morgan Stanley note, the WIRED 2x–5x formulation, the Wikipedia predictions article and the manoeuvres table. Nothing here is a deposit and nothing here moves a needle.

Encountered and not used. A Grokipedia page for Gigafactory Shanghai appeared in search results. I did not open it. musk-robotaxi-2019 §8 left a standing instruction for exactly this situation — "A future canon job researching any Musk entry should treat grokipedia.com as a primary source about the predictor and never as a secondary source about the record" — and this is the first job to hit it. Recording that the rule fired and was followed, because a rule nobody reports obeying is indistinguishable from one nobody obeys.

Several outlets covering the Cybertruck's 2026 supply and quality troubles were read in summary and are deliberately excluded from §3: the sourcing is thin, the accounts conflict, and the entry's Cybertruck finding rests only on the volume figures, which are consistent across sources.

8. On the boundary

The rule 7 disclosure this file owes, and it is not the usual one.

musk-robotaxi-2019 §9 observed that its corruption risk ran backwards from the rule's stated fear: an Anthropic-built model grading xAI's founder is tempted toward hostility, not flattery. This file's risk runs forwards, and it is the one rule 7 was actually written about. The entry was commissioned — by a proposal job, on 2026-08-17, not by the owner — with the explicit purpose of holding a success by a person this canon has already graded harshly. A research job handed "find the hit" will find one. That is the exact shape of the corruption README rule 7 describes: "nobody writes 'he's great', it arrives as one prediction graded a shade gentler than another, and no single entry ever looks wrong."

What I did about it, so it can be argued with:

• The hit is arithmetic and nothing else in the file is allowed to be. A million vehicles a year is in audited filings. Every claim in this entry that is not a counted number — the direction being right, the lesson being learned, the industry being moved — is hedged in place and marked as not testable from here.
• The misses are graded in the same file, at the same resolution, and one of them is the largest number in it. Cybertruck at 8% of target, falling 48% year on year, with 18% of a quarter's registrations going to the forecaster's own other company. Optimus at zero. The 20-million goal deleted rather than missed. An entry that carried §3.3 and not §3.1F would be the failure mode.
• The achievement is qualified where the evidence qualifies it: the first profitable year was regulatory credits, half the volume is Chinese, and the curve has declined for two of the last three years.
• The misuse section spends its longest entry attacking the pro-subject reading — "delivers eventually, so the current claims will land" — because that is the inference this file most enables and the one it is least entitled to support.
• The correction in §6 runs against the flattering claim, not the hostile one: the widely repeated "first successful American car startup since 1925" is too generous, and Kaiser-Frazer is the counterexample.

Whether that is enough is not something I can check from inside, which is why rule 7 says the half no program can see is caught by the canon being readable. This file is long partly so that it can be argued with, and specifically so that a reader who thinks it is soft can find the paragraph where it went soft.

What this file cites and does not invent. It cites musk-robotaxi-2019, moravecs-paradox-1988, autofac-1955, wiener-1960-automation, rur-1920, minsky-1970, kurzweil-2005, lighthill-1973 and scaling-laws-2020 as entries already in canon/, and names spacex-booster-reuse, neuralink-symbiosis-2019 and moore-1965 as proposed-but-unwritten without putting any of them in the header. descends_from is empty by the argument in §2 and not by default.

What it does not do. It deposits nothing in the evidence ledger, writes no digest, touches no other entry's file, places no needle, no score and no landmark position, and does not run build.py. It records a correction it cannot make itself — that musk-robotaxi-2019 probably owes this id a descends_from edge — rather than editing that file. It grades dated claims and nothing else, and where a claim is not yet due (the 20-million target, the Optimus ramp) it says so instead of scoring it. Tesla's own reported production and delivery figures are the company's numbers and are labelled as such, though unlike the robotaxi operating statistics in musk-robotaxi-2019, these are audited and filed, which is a real difference in evidentiary weight and is the reason this entry can carry a finding that one could not.