MBAs Considered Harmful? — Sources

External companion to the long-read. Citation chain for every load-bearing claim, organised by section. Derived figures carry their method; unverified claims are marked. Where the essay describes a company, a product or a person without naming them, the name is given here.

§1 The problem

Ford has moved roughly 350 engineers — rehired, newly hired or promoted — into vehicle quality roles over three years, replacing AI-based quality systems that failed to catch the defects they were installed to catch. Trade and mainstream coverage, 2025–26 (Bloomberg, CNBC and others). The executive’s account — that Ford assumed introducing AI and feeding it the design requirements would be enough — is paraphrased rather than quoted, because the wording varies across outlets. Unverified: the speaker’s name, the date of the statement and the originating outlet are not pinned down. The essay attributes the remark to the head of Ford’s vehicle hardware engineering; the source material for this file recorded it as the chief quality officer. Those are different roles, and the essay’s version is the one to treat as provisional until the original is located.

Ford’s 2023 quality reorganisation ran under Kumar Galhotra; the company subsequently took a top mainstream ranking in the J.D. Power Initial Quality Study, its first in sixteen years, and the chief executive has cited warranty savings in the hundreds of millions. Copy desk: this cuts against the essay’s use of Ford as the most-recalled automaker of the 2020s. Both are true — the recall record and the improvement — and the improvement is partly what the rehiring bought.

Gartner predicts that by 2027 half the companies that attributed customer-service headcount reductions to AI will rehire for similar roles, under different job titles, and finds that only 20 per cent of those who cut did so primarily because of AI. Gartner press release, 3 February 2026, based on an October 2025 survey of 321 customer service and support leaders. The Kathy Ross quotation is paraphrased from the release. Unverified: the survey’s geography is not confirmed from the release.

Robert Half tracker: 29 per cent of US hiring managers had eliminated a role for AI and then rehired for it as of April 2026, on a survey of more than 2,000 hiring managers; 32 per cent as of July 2026. Robert Half data supplied to CNBC. Presented as two dated levels, not as a trend. Copy desk: this figure appears in the briefing, not in the essay. Unverified: the exact survey titles for both waves.

§2 How we got here

The Whiz Kids: ten US Army Air Forces veterans of the Office of Statistical Control, hired by Ford as a group with a start date of 1 February 1946, led by Charles “Tex” Thornton with Robert McNamara as his deputy. McNamara became president of Ford in November 1960 and US Secretary of Defense in January 1961. Standard histories, including Byrne, The Whiz Kids, and contemporaneous accounts.

The McNamara fallacy was codified and named by Daniel Yankelovich in the early 1970s, as a four-step progression: measure what is easily measured; disregard what cannot be; presume that what cannot be measured is not important; say that it does not exist. The essay paraphrases the progression rather than quoting it. Attribution to Yankelovich is standard.

The quantification turn in the business schools followed the Gordon–Howell report (Ford Foundation) and the Pierson report (Carnegie), both 1959. The case method is Harvard-originated pedagogy. Stated at the level of general history.

IDC forecast $38bn a year in IA-64 systems revenue by 2001. The forecast was issued in June 1997; IDC’s June 1998 revision put it at $30bn. Itanium (Merced) shipped in June 2001, years late, and ran x86 software at roughly a tenth of native speed. Actual Itanium server sales were $479m in 2003 and $1.4bn in 2004 on IDC’s own numbers, with systems sales peaking at $4.5bn in 2008. Contemporaneous coverage; ZDNet, “Itanium: A cautionary tale,” 2005; The Register’s documentation of the forecast series running 95 to 96 per cent off.

Above the board sit investors, and above investors the market, with each layer able to read fewer of a claim’s properties than the one below it. The formulation about disclosure being “filed somewhere tedious enough that most people won’t look” is Patrick Boyle’s. The underlying finance: Sloan (1996), The Accounting Review, on the accrual component of earnings being less persistent than the cash component and the resulting predictability of returns, an anomaly that attenuated once it became a recognised trade; and Shleifer and Vishny (1997), Journal of Finance, on the limits of arbitrage.

Dijkstra, “Go To Statement Considered Harmful,” Communications of the ACM, March 1968. A letter. The essay characterises the argument rather than quoting it.

On the discounting of expert objections, the consultancy’s comply-over-explain lesson, and the claim that every part of the machine is locally reasonable. Unverified: these are the essay’s analytical frame. No external source is claimed for them.

§3 What it costs

Intel announced the sale of its XScale division to Marvell on 27 June 2006 — roughly 1,400 employees, $600m plus assumption of liabilities, with customers including RIM, Palm, HTC and Motorola. Intel retained its ARM architectural licence. The stated rationale was focus on core x86 and server. Intel and Marvell releases; contemporaneous coverage. The essay refers to this as “its ARM division” without naming XScale or Marvell. Derived: $600m against revenues of $35.4bn that year is 1.7 per cent.

2006 context: AMD took 5.3 points of x86 share in 2005; Intel’s Q1 2006 profit fell 38 per cent; a restructuring followed. Contemporaneous reporting and filings.

Intel’s 2005 results were the best in its history to that point — revenue $38.8bn, net income $8.7bn, announced by Otellini in January 2006. Intel Q4 2005 release.

Paul Otellini on declining the iPhone, to Alexis Madrigal, The Atlantic, 16 May 2013. Both quotations are verbatim from the interview: the cost and volume account ("…the forecasted cost was wrong and the volume was 100x what anyone thought") and the data and gut account (“while we like to speak with data around here… My gut told me to say yes”).

The revisionist case — that the economics genuinely did not work at the price Apple was offering — is put by The Chip Letter, a semiconductor history newsletter. Characterised and conceded in the essay.

The manufacturing flywheel argument — that volume funds process learning, learning delivers the best factories, and the best factories win everything else. Developed at length in Up for Scraps. Copy desk: earlier drafts carried an explicit cross-reference to that essay at this point. The current prose does not.

Vindication: revenue $35.4bn (2006) → $43.6bn (2010) → $54.0bn (2011). Otellini, January 2011: “2010 was the best year in Intel’s history. We believe that 2011 will be even better.” 2011 set records on revenue, operating income, net income and earnings per share, at a 62.5 per cent gross margin, with $14.1bn of buybacks across 642m shares against $1.5bn in 2010. Intel annual results releases and 10-Ks, FY2006–FY2011.

Amazon acquired Annapurna Labs in January 2015 for approximately $350m. Reuters and contemporaneous reporting. The essay calls it “a small chip-design firm”. Derived: Intel’s 2011 buybacks were about forty times that figure.

AWS silicon cadence: Nitro shipped November 2017; Graviton was announced November 2018; Graviton2 followed in December 2019 with a published claim of 40 per cent better price-performance against comparable instances; Trainium was announced on 1 December 2020 at re:Invent — the same event at which Intel supplied Habana Gaudi instances — and Trn1 shipped in 2022. AWS announcements. The essay names none of these products, referring to “its own server processors” and the 40 per cent claim. Unverified: the exact comparator wording behind the 40 per cent claim has not been checked against the AWS release.

Apple announced the Mac transition on 22 June 2020; the first M1 Macs shipped in November 2020; the transition completed with the Mac Pro in June 2023. Apple newsroom.

Microsoft announced Cobalt 100 at Ignite in November 2023, with 128 Neoverse N2 cores. Google announced Axion on 9 April 2024 at Cloud Next, reaching general availability in October 2024, with a claim of 50 per cent better performance than comparable x86 VMs; YouTube ads, Spanner and BigQuery were already running on Arm internally before the announcement. Google Cloud blog. The essay refers to this last point as Google’s internal systems “running on its new architecture before the chip even had a public name,” without naming Axion or the services.

Intel data-centre segment revenue: DCG $15,977m (2015), $17.2bn (2016), $19.1bn (2017), $23.0bn (2018, a record on the FY2018 10-K narrative), $23.5bn (2019). 2020 was the peak, DCG up 11 per cent. 2021 fell 1 per cent, which Intel’s FY2021 10-K attributed to product mix and, in the filing’s words, “a competitive environment.” 2022 fell roughly a quarter to $16.9bn on the DCAI basis, and 2023 fell again to $15.5bn. Intel 10-K segment tables. Copy desk: Intel restructured its segment reporting in 2021 (DCG to DCAI) and again in 2024 (Altera carved out), each time restating prior years, so any series spanning the decade splices definitions. On the post-Altera basis, DCAI runs $12.8bn (2024) to $16.9bn (2025) — numerically coincident with 2022 on the old basis. The two must never be spliced. The shape survives the restatements: growth through 2020, stall, cliff. Unverified: exact 2016–2019 figures were pulled at displayed precision rather than from each year’s segment table.

The cliff has more than one cause: post-pandemic digestion of cloud capacity, AMD taking server share on merit, and data-centre budgets rotating toward GPU systems in which Intel had no position. Stated in the essay.

Brian Krzanich diagnosed the 10nm process failure on Intel’s Q1 2018 earnings call, 26 April 2018, verbatim: “we understand the yield issues. They’re really tied to this being the last technology tied to not having EUV, and the amount of multi-patterning, and the effects of that on defects.” An analyst on the same call proposed skipping to 7nm on Intel’s EUV experience; EUV was not production-ready for a 2019 ramp. Intel had made four first-in-industry bets — SAQP interconnect, a 36nm metal pitch against competitors’ 40nm, cobalt interconnects, and no EUV — and Intel 4 subsequently dropped cobalt and adopted EUV. Call transcripts (Motley Fool, Nasdaq, Seeking Alpha); contemporaneous AnandTech. The essay refers to Krzanich as “its own chief executive”.

The telephone industry’s engineers and the Internet: 1990s data-networking history, ATM against IP. Unverified: characterised at a general level, drawing on the author’s professional domain rather than a cited source.

Pat Gelsinger was appointed in January 2021, effective 15 February, and removed in December 2024. The 18A volume ramp came in the second half of 2025 and Panther Lake in January 2026; the record quarter landed in 2026 under Lip-Bu Tan, who was appointed in March 2025 — a semiconductor veteran, Cadence’s chief executive from 2009 to 2021, with a venture background — and who places the growth inflection in 2027, calling 2026 “an execution year” in remarks at Stanford. Intel’s Q2 2026 revenue was up 25 per cent year on year, on 18A capacity committed under Gelsinger’s 2021 five-nodes-in-four-years programme. Intel results releases; Tan’s remarks as reported. The essay names neither Tan nor the programme, referring to “his successor” and “his factories”.

§4 Digging out

Gelsinger on what went wrong, All-In podcast, July 2026: “you need technologists running technology… went off the rail… the bean counters, the finance people.” The essay flags him as an interested party offering the crude version. Unverified: the exact episode date.

The renaming of Intel’s process nodes: 26 July 2021, at Intel Accelerated, under Gelsinger — 10nm Enhanced SuperFin became Intel 7, 7nm became Intel 4, with a roadmap of Intel 3, 20A and 18A. The stated purpose was alignment of perception. Intel announcement. This is what the essay means by engineers debasing the one metric in their industry that tracked physical reality.

Node names once denoted gate length and decoupled from it in the FinFET era. General history, stated at that level.

Intel’s 10nm was roughly at parity with TSMC’s 7nm: contacted gate pitch of 54nm on both, and transistor density of 100.8 MTr/mm² for Intel 10nm against roughly 91 to 96.5 MTr/mm² for TSMC N7 non-EUV. AnandTech; WikiChip (Intel 10nm poly pitch 54nm, minimum metal pitch 36nm, density 100.76 calculated against 100.8 Intel-reported); Angstronomics (90.64 for N7 2-fin). The parity claim is verified as written.

Ford’s rehiring as the simplified version of wiring the readers back in. As §1.

The dark fibre account: a gigabit between buildings at €2,000 a month against €500 for two overbooked megabits on 5:1 contended SDSL, in the early 2000s. Unverified: the author’s professional experience. No external source is claimed, and none of the surrounding argument about escalation paths, symmetry of justification, or scoring across the decision lag rests on a citation.

Chart 1 — segment revenue against the defection timeline

Exact DCG figures from SEC primaries: $15,977m (2015, FY2015 10-K); $19,064m (2017), $22,991m (2018), $23,481m (2019) (FY2019 10-K segment table); $26,103m (2020, FY2020 10-K segment table). 2016 at $17.2bn and 2021 at $25.8bn are confirmed at displayed precision (FY2016 10-Q series; Q4 2021 results and FY2021 10-K narrative), with exact figures available from the FY2017 and FY2021 10-K tables if needed. On the DCAI basis: $23.4bn (2020) and $22.7bn (2021) per Intel’s resegmentation exhibit of January 2022; $16,856m (2022) and $15,980m (2023) from 10-K-derived segment data. The basis break is rendered as two overlapping series, with the footnote carrying the splicing statement.

A note on what is not here

A longer earlier version of this essay carried several further lines of argument, sourced but now cited nowhere in the text. They are recorded here so the research is not lost: Gartner’s Magic Quadrant as an instrument with two axes and no units, compiled by a firm whose revenue comes from the vendors it places; Intel’s x86-S specification, published May 2023 and withdrawn in December 2024 citing “inputs from the ecosystem,” with the x86 Ecosystem Advisory Group formed with AMD and the major Windows OEMs in October 2024; a vendor-commissioned survey (Orgvue) finding 55 per cent of leaders who cut for AI considered the decisions wrong; Citigroup credit research of 10 June 2026 on the spread widening of the QTS bond financing a Microsoft-linked data centre; SpaceX as a control case on funding and failure tolerance, with the Falcon 1 sequence and the August 2006 NASA COTS award; and Intel’s buyback history, including roughly $152bn repurchased since 1990 and the halt after Q1 2021.

The essay also no longer carries its cross-references to The Token Sink and Up for Scraps, both of which it previously cited directly.