"The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox"

Filed under interpretation in canon/proposals.md, and that is where it belongs. Worth one paragraph anyway, because the near miss is instructive. The paper is seven pages long and contains no new data, no method anyone can run, and no result. What it contains is a reading of an episode — the forty years between Edison's first central power station and the moment electrification showed up in American manufacturing productivity — offered as the right frame for a different episode happening while David wrote. The house test for idea, argued out in human-computers and applied in nordhaus-price-of-light-1994, is that the thing is implementable and gets built on rather than argued about. This is the opposite on both counts: nothing here is implementable, and it has been argued about continuously for thirty-six years, including by David himself. It is not a moment either — the paper's publication changed nothing; the thing it describes is the moment, and that moment is 1882–1929. interpretation is exact.

On the year field. 1990 is the American Economic Review 80(2), AEA Papers and Proceedings, May 1990, pp. 355–361. The paper carries a footnote saying it "draws upon material developed in a longer work — my 1989 paper," which is Computer and Dynamo: The Modern Productivity Paradox in a Not-Too-Distant Mirror, Center for Economic Policy Research Publication No. 172, Stanford, July 1989; the 1989 paper was later published in the OECD's Technology and Productivity (1991), which is why much of the literature cites this argument as David 1991. All three dates refer to one body of work. This matters more than a citation nicety, and it is the first caveat this entry has to carry: every quantitative claim in the 1990 paper is a pointer into the 1989 working paper — "see my 1989 paper, Table 3"; "Table 4, Fig. 14"; "Table 5, and pp. 26–27"; "Appendix A." The famous version of the argument is a summary whose evidence lives somewhere else. I did not obtain the 1989 paper (§5 records the attempt), so where a number below is attributed to David it is attributed to a seven-page précis of a table I have not seen.

On descends_from. Empty, and not for want of ancestors. The real ones are Robert Solow's 1987 book-review sentence, Warren Devine Jr.'s 1983 Journal of Economic History reconstruction of the shift from shafts to wires, Richard DuBoff's Electric Power in American Manufacturing 1889–1958 (1979), Thomas Hughes's Networks of Power (1983), Bresnahan and Trajtenberg's then-unpublished general-purpose-technology paper (1989), and Freeman and Perez's regime-transition thesis. None of them is a canon entry, and minting ids for them would be worse than leaving the field empty. Several entries that do exist are siblings a reading will often want alongside this one: printing-press-1440, which is the same puzzle — a transformative technology that economists could not find in the aggregate data — run on a five-hundred-year-old case; nordhaus-price-of-light-1994, which attacks the measurement from the price side where this attacks it from the diffusion side; and expert-systems-collapse-1987, which is what the other branch of the same decade looks like when the disappointment turns out to be real.

What it is

In July 1987 Robert Solow reviewed a book in the New York Times Book Review and wrote one sentence that outlived everything around it: "You can see the computer age everywhere but in the productivity statistics." By the late 1980s that sentence had become the name of a problem. American business was investing heavily in data processing equipment; measured total factor productivity growth had been slow since the mid-1970s and showed no sign of turning. Economists trained to read the TFP residual as "the advance of knowledge" were looking at a residual that had gone flat during the arrival of the microprocessor.

David's answer was to say that this had happened before, in a case where we already know how the story ends.

His argument runs in four moves. First, the parallel is structural, not decorative. Computer and dynamo are both what he calls general purpose engines — "key functional components embodied in hardware that can be applied as elements or modular units of the engineering designs developed for a wide variety of specific operations or processes." Both sit at the node of a physically distributed transmission network, both are surrounded by strong complementarities that make standardisation a strategic question, and both diffuse slowly through a long trajectory of incremental improvement rather than arriving.

Second, the dating. In 1899, electric lighting was in a mere 3 percent of American residences and 8 percent of urban dwelling units, and the horsepower capacity of all electric motors installed in manufacturing establishments was under 5 percent of factory mechanical drive. It took roughly two more decades to reach 50 percent diffusion. Electrification "did not acquire real momentum in the United States until after 1914–17," when regulated regional utility rates fell substantially against the general price level and central-station capacity came to predominate over generating capacity in isolated industrial plants. Factory electrification "did not reach full fruition in its technical development nor have an impact on productivity growth in manufacturing before the early 1920s" — and, in the paper's most quoted clause, "This was four decades after the first central power station opened for business."

Then he closes the parallel with the sentence the whole entry is built around: in 1900, an observer of the "Electrical Age" stood as far from Edison and Swan's carbon filament lamp and the first Edison central stations as an observer in 1990 stood from the memory chip and the silicon microprocessor. "In 1900, contemporary observers well might have remarked that the electric dynamos were to be seen 'everywhere but in the productivity statistics!'"

Third, the mechanism, which is the part worth having. The delay was not inertia and not stupidity. It had a specific engineering shape. The nineteenth-century factory was built around a single prime mover — waterwheel or steam engine — turning long overhead line shafts that drove machines through belts. Because power losses rise with shaft length, the building itself was a consequence of the power system: multi-storey, braced to carry heavy overhead shafting and belt housings, laid out to keep shafts short rather than to keep material moving.

The first thirty years of electrification did not touch any of that. Under the "group drive" system, which "remained in vogue" from the mid-1890s to the eve of the 1920s, electric motors were added to turn separate sections of the same shafting. David is precise about what this does to the accounts: retrofitting a steam- or water-powered plant "typically entailed adding primary electric motors to the original stock of equipment," while the belts and shafts stayed in place as available capacity. "The effect would be to raise the capital-output ratio in manufacturing, which militated against rapid gains in total factor productivity — especially if the energy input savings and the quality improvements from better machine control were left out of the productivity calculation." A firm can adopt a superior technology, be right to adopt it, and make the productivity statistics worse in the process.

The gains came from "unit drive" — a motor on each machine — and they were mostly not about power at all. David lists four: lighter factory construction once the bracing for overhead shafting could be dispensed with; the shift to single-storey buildings, since the reason for stacking floors had been shaft length; layouts optimised for materials handling and reconfigurable as products changed; and the modularity that meant maintenance or retrofitting no longer required shutting down the whole mill. Elsewhere he adds the qualitative ones that the accounts never caught: better machine control from variable-speed DC motors, and lighter cleaner workshops with skylights where the transmission apparatus used to hang and without "the myriad strands of rotating belting that previously swirled dust and grease through the factory atmosphere, and, where unenclosed within safety screening, threatened to maim or kill workers who became caught up in them."

His single quantitative claim in the whole paper — and it is a claim about a cross-section, drawn from a table in the 1989 paper — is that industries whose installed secondary electric motor capacity grew most between 1919 and 1929 also saw the largest rise in TFP growth, and that "approximately half of the 5 percentage point acceleration recorded in the aggregate TFP growth rate of the U.S. manufacturing sector during 1919–29 (compared with 1909–19) is accounted for statistically simply by the growth in manufacturing secondary electric motor capacity during that decade."

Fourth — and this is the half of the paper almost nobody quotes — he then attacks his own analogy. Section IV opens "Computers are not dynamos" and does not soften. Human-computer interfaces are far more subtle than wiring a motor. Information is not electric current: it lacks superadditivity, has negligible marginal costs of transfer, and can produce "overload," a congestion effect arising from inhibitions on free disposal that no ordinary commodity has — transmitters are encouraged to broadcast indiscriminately, screening is costly, and "the growing duplicative allocation of human resources to coping with information overload may displace activities producing commodities that are better recorded by the national income accounts." And the sharpest one, which cuts directly against optimism: conventional plant and equipment wears out, and wearing out is what forces the occasion to rebuild. "Information structures per se do not automatically undergo significant physical depreciation… one cannot depend on the mere passage of time to create occasions to radically redesign a firm's information structures and operating modes. Consequently, there is likely to be a strong inertial component in the evolution of information-intensive production organizations." The dynamo's lag had an end date because steam plants rusted. Nothing guarantees the computer's does.

The paper's actual thesis statement is not "be patient." It is the sentence opening section IV: closer study of the dynamo revolution "should help us avoid both the pitfall of undue sanguinity and the pitfall of unrealistic impatience into which current discussions of the productivity paradox seem to plunge all too frequently." Both pitfalls. Section 4 of this entry exists because that clause is the one that gets dropped.

Why a reading would cite it

When a buildout is defended by a lag. This is the standing occasion and it is already live in the project's record. The 2026-08-20 midnight reading carried Dell'Oro roughly doubling its 2030 datacentre capex outlook to above $3tn in seven months, and OpenAI's CFO describing a 2027 IPO to a company whose operating loss had widened to $12.3bn against $6.7bn of quarterly revenue. Set that beside NBER Working Paper 34836 (February 2026), which surveyed nearly 6,000 senior executives in the US, UK, Germany and Australia and found 69 percent of firms actively using AI while "nine-in-ten reporting no impact on employment or productivity" over the past three years. A capital programme of that size against a measured effect of that size demands an explanation, and "the dynamo took forty years" is the explanation reached for. This entry is what makes that argument checkable instead of rhetorical: David does not say wait, he says here is the specific reason the number is quiet, and here is what you would have to observe for it to stop being quiet. Group drive is an observable condition. A reading can ask whether the firms buying the capacity are reorganising around it or bolting it to the existing shafting — and, per §4, can ask whoever invokes the lag to name the anchor and the due date, which is the request the argument usually cannot survive.

When a disappointing measurement is offered as proof that nothing is happening. The mirror occasion, and equally common. Ninety percent of executives reporting no effect is a real finding and this project should record it as one; it is not, on its own, evidence that there is no effect. David's contribution is the demonstration that a technology can be genuinely transformative, genuinely adopted, and statistically invisible or worse than invisible at the same time, because the retrofit raises the capital-output ratio while the reorganisation that pays for it has not happened yet. Brynjolfsson, Rock and Syverson formalised this in 2021 as the productivity J-curve: while intangible complements are being built, measured productivity is understated; later, when the hidden stock starts yielding, it is overstated. Their 2017 paper — which is the modern restatement of this entry and cites David explicitly — lists four candidate explanations for the AI paradox (false hopes, mismeasurement, redistribution, implementation lags) and argues lags dominate. A reading that wants to say "the absence of a productivity signal is not yet informative" has a citation for it here, and does not have to invent one.

When a good productivity number is offered as proof that it is. This is where the entry earns its keep in 2026, because the data have turned and the temptation has flipped. US nonfarm business labor productivity rose 2.2 percent in the year to Q2 2026 and has grown at a 2.1 percent annualised rate since Q4 2019 — the pace of the 1995–2004 IT acceleration, not of the 2005–2016 slump. The Dallas Fed (July 2026) reports 2.4 percent annualised since the start of 2024 against 1.6 percent in the five pre-pandemic years, and notes that information, finance and insurance, and professional and technical services — the three most AI-exposed sectors — account for 16 percent of hours worked and 40 percent of productivity gains since the start of 2024. That is a striking correlation and it is not attribution, and the reason to cite David here is his own 1999 correction to himself: the 1920s surge, he and Gavin Wright concluded, was not a technological event but "the confluence of two largely independent streams of development" — the culmination of the dynamo revolution and the restructuring of American manufacturing labour markets after mass European immigration was cut off in 1914. The canonical case for attributing a productivity surge to a general purpose technology turns out, on its author's own later reading, to be a case of attributing it to a technology and a labour shock at the same time. Any reading tempted to hand 2024–2026 to AI should have to pass that test first.

And when the analogy itself is the thing being asserted. "AI is the new electricity" is one of the most-repeated sentences in this industry. It is a citation to this paper by people who have generally not read it, and what they are citing is one clause of it. A reading that wants to engage rather than dismiss needs the other clauses — the mechanism, the both-pitfalls warning, and above all "Computers are not dynamos." The no-forced-depreciation argument in section IV is the strongest thing in the paper and it points away from the optimistic reading: factories rebuilt because their power plants wore out, and an organisation's information structures never wear out, so there is no equivalent forcing function and the inertia can run indefinitely. METHODOLOGY.md §4b already lists "the dynamo and the productivity paradox" among the canon candidates that argue for widening this project's question from AI to technology-and-civilisation; that widening is exactly the move David makes, and it comes with this caveat attached.

Where this entry does not reach. It is not evidence that AI will pay off. It is not evidence that AI will not. It is a demonstration that a particular class of measurement is uninformative during a particular kind of transition, and that demonstration is symmetric: it forbids reading a quiet quarter as absence and a loud one as arrival with equal force. Anyone citing it in only one direction is using half of it. It is also not evidence in this project's sense at all — nothing here deposits into the ledger and nothing here moves the needle.

What it got right, and what it got wrong

Not a prediction, so there is no self-declared due date to grade against — and §4 argues that the absence of one is the paper's central defect, not an excuse for skipping this section. The claims are dated and checkable. Graded one at a time.

Claim: factory electrification had no impact on manufacturing productivity growth before the early 1920s, four decades after the first central power station. Made May 1990. The aggregate discontinuity held and was nailed down harder; the "long lag" reading of it is now under direct empirical challenge.

David and Wright reconstructed the break in 1999 and it is real and large. After a "productivity pause" of about three decades in which gross manufacturing output grew at under 1 percent per annum relative to capital and labour inputs, TFP in US manufacturing expanded at more than 5 percent per annum between 1919 and 1929. On Kendrick's estimates the decadal TFP gain was about 22 percent for the whole private domestic economy and 76 percent for manufacturing; labour-productivity trend growth jumped from 1.5 points per annum in 1899–1914 to 5.1 in 1919–1929. Thirteen of fourteen major manufacturing categories accelerated between 1909–19 and 1919–29 — broad, "yeast-like" growth in Harberger's terms, not a few "mushroom" industries carrying an average. The event David was pointing at happened, and it was bigger than the 1990 paper claims.

The challenge is to the lag, and it is recent and serious. Fiszbein, Lafortune, Lewis and Tessada built a sector-county panel of US manufacturing from 1890 to 1940 and identified electricity's effect off the interaction of pre-electrification energy intensity across industries with proximity to early hydroelectric plants. Their finding, in their own abstract: labour productivity gains from electricity "were relatively rapid and long-lasting, in contrast with established narratives suggesting moderate impacts with long time lags," and firms "changed their production processes relatively quickly, becoming more capital-intensive, employing a different mix of labor skills, and increasing occupational diversity." Within ten years of electricity's arrival, not forty. They name David directly as the narrative they are contrasting with, and conclude that "barriers to change may have been lower than previously thought." If they are right, the canonical case for long implementation lags loses its canonical case, and every modern argument resting on it — including Brynjolfsson, Rock and Syverson's — loses its historical anchor while keeping its logic. Their status is a working paper, NBER 28076, circulated November 2020 and revised February 2024, and I could not confirm journal publication; it should be carried as a strong unsettled challenge and not as a verdict. A separate and older challenge, Moser and Nicholas (2004), questioned whether electricity qualifies as a general purpose technology at all on patent-citation evidence, and drew a published defence in reply; I did not read either and note them only as evidence that the case is contested rather than closed.

Claim: the analogy transfers — the computer in 1990 stands where the dynamo stood in 1900. Made May 1990, no due date given. Directionally right, arrived earlier than the stated clock, then failed to stay right, which is the interesting part.

The US productivity acceleration came in 1995–2004: aggregate labour productivity averaged 2.8 percent a year, against 1.3 percent from 2005 to 2016. Brynjolfsson, Rock and Syverson overlay the two eras and find them "remarkably similar" — roughly a quarter-century of slow growth in each, then a decade-long acceleration, 1915–1924 in the portable power era and 1995–2004 in the IT era. On David's own anchoring the payoff should have landed around 2010, and it landed around 1995; he had said the computer's diffusion was running faster than the dynamo's, so early arrival is consistent with what he wrote. Call that a hit.

Then it un-happened. Growth fell to 1.3 percent for 2005–2016 in the US and 28 of 29 other OECD countries decelerated with it, from an unweighted 2.3 percent to 1.1 percent. The entire argument had to be run a second time, for a different technology, by different authors, in 2017. That is the finding that should govern how this entry is used. An argument that accommodates the payoff arriving in 1995, and also accommodates it vanishing in 2005, and also accommodates a fresh one in 2024, is doing less work than its confidence suggests. Brynjolfsson, Rock and Syverson are honest about this: they note that the historical series also has a 1924–1932 slowdown and then a further acceleration averaging 2.7 percent a year in 1933–1940, and draw from it only that "productivity growth driven by a core GPT can arrive in multiple waves" — which is a description, not a prediction, and is compatible with essentially any future series.

Claim: the delay was caused by the survival of an obsolete factory architecture, and the gains came from reorganisation rather than from the motor. Made 1990. Held, and it is the durable contribution. Devine's 1983 reconstruction, which David leans on and which I did not read directly, is the source of the group drive/unit drive account, and the underlying diffusion numbers are not in dispute: electricity supplied under 5 percent of primary horsepower in US manufacturing in 1899, about 50 percent by 1919, about 75 percent by 1929. What David added was the reading of why the middle of that curve produced nothing — that the intermediate technology was a genuine improvement adopted for genuine reasons that nonetheless could not deliver the capital savings, and that the productivity accountant looking at 1910 was measuring a real thing correctly. Fiszbein and co-authors dispute the timing and explicitly do not dispute the reorganisation; they find "substantive changes in the intensity of physical capital use, the skill composition of the workforce, and the complexity of the production process," just sooner. Thirty-six years on, the mechanism is better supported than the schedule.

Claim: conventional productivity measures miss quality change and new categories during a regime transition. Made 1990. Held, and grew into a literature. His examples were the ones the accounts of the 1890–1914 era could not catch — brightness, fire safety and maintenance for electric light against gas; faster trip speeds and shorter waits on electric streetcars; the residential amenity of being able to live somewhere salubrious and commute. The modern version is the intangible capital argument, which David himself moved on to (with Abramovitz, 1996) and which Brynjolfsson, Rock and Syverson formalised: expenditure on complements is expensed rather than capitalised, so the output of building them does not appear in GDP while the inputs do, and measured productivity is depressed exactly when real investment is highest. This is the part of the 1990 paper with the clearest descendants.

Claim: "Computers are not dynamos," and information's special properties may make the lag longer rather than shorter. Made 1990. Held, underquoted, and the most useful thing in the paper for a 2026 reading. The no-physical-depreciation argument in particular has aged into something sharper than he can have intended: the case for AI transformation now rests almost entirely on organisations voluntarily rebuilding processes that are working, with no forcing function equivalent to a steam plant wearing out. The information-overload argument — that resources spent coping with abundant low-cost information displace activity the accounts do measure — was speculative in 1990 and is not obviously speculative now.

Errors of detail, small but load-bearing. The paper dates its computer-side anchors as "the 1043 byte memory chip (1969) and the silicon microprocessor (1970) by Intel." The Intel 1103 was a 1,024-bit DRAM introduced in October 1970 — the printed figure is a garbled unit as well as a wrong year — and the Intel 4004 microprocessor was announced on 15 November 1971. Both anchors run about a year early. On the other side he gives "the Edison central generating station in New York and London (1881)," where Holborn Viaduct began operating in January 1882 and Pearl Street on 4 September 1882. Each error is a year. They matter only because the anchors are the clock: "four decades" is a subtraction between two dates chosen by the author, and §4 is about what happens when the choosing is left unexamined. And his evidence for the computer's diffusion — "only 10 percent of the world's 50 million business enterprises today are using computers, and only 2 percent of the world's business information has been digitized" — is cited to a New York Times column of 31 December 1989. In a paper arguing that measurement is the hard part, the measurement of the thing being explained is a newspaper column.

And the grade he did not give himself, which is a gap in this entry. The prediction guidance in jobs/canon.md asks for the predictor's own grade beside an independent one. David wrote at least two retrospectives — "Digital Technology and the Productivity Paradox: After Ten Years, What Has Been Learned and What Do We Need to Know?" (White House Conference on Understanding the Digital Economy, 25–26 May 1999) and "Productivity Growth Prospects and the New Economy in Historical Perspective" (EIB Papers, 2001) — and I obtained neither, so no self-grade is recorded here and none should be inferred. What I can record, because I read it, is his posture rather than his verdict: writing with Gavin Wright in September 1999, four years into what would later be dated as the 1995–2004 acceleration, he was still describing the ICT payoff as ahead, closing on "contemporary portents of a future phase of more rapid total factor productivity growth" and listing candidates for it — inter-organisational computing for purchase ordering, invoicing, shipment tracking and payments; the growth of tele-working; and "digital information technologies embedded in hand held devices — or other robust and specialized tools that are carried on belts, sown into garments, or worn as head-gear — that are linked through sophisticated networks." The wearable-and-networked guess reads remarkably well from 2026. The timing does not: the economist who wrote the canonical account of why payoffs are recognised late did not recognise this one in real time either. That is not a gotcha, it is the base rate. The lag argument tells you the signal will be hard to see; it does not hand you a detector.

Commonly misused as

Not required for interpretation. The entry earns the section anyway, because canon/proposals.md asks for exactly this — the id was proposed as "the strongest argument that a real general purpose technology looks like a disappointment for decades, and the strongest warning against using that to excuse anything indefinitely" — and because the misuse is now the dominant use.

Misused as: a licence for indefinite patience. The paper contains no due date, no failure condition, and no statement of what would show the analogy had broken. That makes it, as deployed, unfalsifiable: any disappointing quarter is early, any good quarter is the arrival, and no observation can count against it. The trick is usually done at the anchor. David dates the lag from the first central station — 1882 to the early 1920s, "four decades." Brynjolfsson, Rock and Syverson, summarising the same author, write that at least half of US manufacturing establishments were unelectrified until 1919, "about 30 years after the shift to polyphase alternating current began." David and Wright date the real momentum to after 1914–17, which is about thirty-five years, or about three years before the surge depending on which end you measure from. One episode, three clocks, all honestly derived, spanning a factor of more than ten in the implied remaining wait when transposed onto AI. Whoever picks the anchor picks the answer. The honest use requires stating the anchor, stating the due date, and accepting the grade later — which is to say, converting the argument into the kind of thing this canon files under prediction. Almost nobody who invokes the dynamo does this.

Misused as: proof that the technology caused the surge. David's own 1999 verdict is that "a purely technological explanation of the productivity surge is inadequate," because it would neglect the concurrent restructuring of US labour markets after the closing of mass European immigration in 1914, and the interrelations between managerial innovation, the new factory technology, and the macroeconomic conditions of the 1920s. He and Wright treat the surge as the confluence of two largely independent streams. The popular version keeps the dynamo and drops the immigration restriction, the management revolution and the 1920s capital boom — which converts a careful historical argument into a claim that buying the machine is sufficient if you wait.

Misused as: proof that the lag must be long. The most recent empirical work on the very episode says the opposite, and says so pointedly (Fiszbein, Lafortune, Lewis and Tessada, above). It remains a working paper and the question is open, which is precisely the point: the historical fact being used to settle a modern argument is itself unsettled. Citing "forty years" as an established quantity in 2026 overstates what the literature currently supports.

Misused as: a prediction about AI. It is not a prediction, it is not about AI, and it declines in advance to be either. Its subject is why a residual computed from national accounts went quiet during one regime transition, and its explicit purpose is to warn against both over-optimism and impatience. The version that travels — "the dynamo took forty years, therefore be patient about AI" — deletes the first pitfall, the mechanism, and the whole of section IV.

And the reverse misuse, which this project is as exposed to as anyone. The 1920s surge did arrive, and it is tempting to read that as proof the payoff always arrives eventually. David and Wright's cross-national section exists to complicate exactly that: they set the US experience against the contemporary UK to bring out "differences between leader and follower economies in the dynamics of GPT diffusion and its relationship to the strength of surges in productivity growth," and to underscore "the important role of the institutional and policy context." I read their abstract and conclusions and not that section, so I record the claim as theirs rather than as verified here — but the shape of it is enough to block the inference. Whether a general purpose technology produces a surge somewhere is not the same question as whether it produces one here, and the second question has institutions in it. expert-systems-collapse-1987 is in this canon as the standing reminder that some disappointments are just disappointments.

Sources

Read directly, as rendered page images.

• **Paul A. David, "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox," American Economic Review 80(2), AEA Papers and Proceedings, May 1990, pp. 355–361 — retrieved from the copy hosted at almendron.com (a ResearchGate scan uploaded by David himself in May 2014) and read as page images, the whole paper including the references and the author footnote**. Source for everything attributed to the paper above: the general-purpose-engine definition; the 3 percent / 8 percent / under-5-percent figures for 1899 and the two-decade path to 50 percent diffusion; the 1879/1881 and 1969/1970 anchors as printed, including "1043 byte"; the 10-percent-of-50-million and 2-percent digitisation figures and their attribution to Peter Lewis in the New York Times, 31 December 1989; "four decades after the first central power station opened for business"; the group-drive and unit-drive accounts and the four advantages of unit drive; the capital-output-ratio passage; the half-of-the- 5-percentage-point TFP claim for 1919–29 and its citation to "my 1989 paper, Table 5, and pp. 26–27"; the 1914–17 utility-rate passage; the belting and skylights passage; the whole of section IV including "Computers are not dynamos," the overload argument and the no-physical-depreciation argument; and the "undue sanguinity / unrealistic impatience" sentence. Also the source for David's rendering of Solow as "We see the computers everywhere but in the productivity statistics" — a paraphrase, which he marks as "attributed to."
• **Paul A. David and Gavin Wright, General Purpose Technologies and Surges in Productivity: Historical Reflections on the Future of the ICT Revolution, University of Oxford Discussion Papers in Economic and Social History No. 31, September 1999 (dated August 1999 in the series list on its own back matter; presented to the Feinstein symposium, Oxford, 2–4 July 1999) — the copy at nuffield.ox.ac.uk, read as page images, pages 1–6 and 18–25 only: title, abstract, introduction, section 1, the personal-computer section, section 5 and the references. Source for: the abstract's framing and its UK comparison; the under-1-percent productivity pause and the above-5-percent 1919–29 TFP figure; Kendrick's decadal 22 / 76 / 41 / 14 percent figures; the 1.5 → 5.1 labour-productivity trend jump in footnote 1; the thirteen-of-fourteen categories and the Harberger yeast/mushroom contrast; the "confluence of two largely independent streams" including immigration restriction after 1914; the explicit statement that a purely technological explanation is inadequate; and the closing list of ICT portents including inter-organisational computing, tele-working and the belt-carried, garment-sewn and head-worn networked devices. I did not read sections 2 through 4**, which contain the detailed US–UK comparison, so nothing is attributed to that material beyond what the abstract and conclusions state.
• Erik Brynjolfsson, Daniel Rock and Chad Syverson, "Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics," NBER Working Paper 24001, November 2017 — the NBER PDF, read as page images, pages 1–6 and 20–31. Source for: the abstract's four explanations and the claim that lags dominate; the 1.3 percent 2005–2016 versus 2.8 percent 1995–2004 US figures and the 28-of-29 OECD comparison at 2.3 and 1.1 percent; the computer-capital-stock plateau at about 5 percent of nonresidential equipment capital in the late 1980s; their summary of David (1991) as "about 30 years after the shift to polyphase alternating current began"; Jovanovic and Rousseau's "a GPT does not deliver productivity gains immediately upon arrival"; Figure 8 and the 1915–1924 / 1995–2004 overlay, the 1924–1932 parallel slowdown and the 2.7 percent 1933–1940 recovery; and the J-curve derivation. I did not read pages 7–19 or 32 onward.
• Martín Fiszbein, Jeanne Lafortune, Ethan G. Lewis and José Tessada, "Powering Up Productivity: The Effects of Electrification on U.S. Manufacturing," February 2024 — the author copy at ethang.host.dartmouth.edu, read as page images, pages 1–3: title, abstract and introduction only. Source for the verbatim "in contrast with established narratives suggesting moderate impacts with long time lags," the naming of David (1990, 1991) as the narrative in question, the hydropower-proximity identification strategy, the within-ten-years reorganisation finding, the 10 percent 75th-versus-25th percentile productivity gap by 1920, and "barriers to change may have been lower than previously thought." I did not read the results. Circulated as NBER Working Paper 28076 (November 2020); I could not establish journal publication.

Established by search against primaries, without reading the full document. For each of these I can vouch for the fact and the citation and not for having read the source end to end.

• **Robert M. Solow, "We'd Better Watch Out," New York Times Book Review, 12 July 1987, p. 36** — the definitive citation and the exact wording, "You can see the computer age everywhere but in the productivity statistics," per the Standup Economist's citation hunt, which also records that the review is widely and wrongly attributed to the New York Review of Books, and that Solow later told the New York Times in 2000 that "you can now see computers in the productivity statistics."
• Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Philip Bunn, Steven J. Davis et al., "Firm Data on AI," NBER Working Paper 34836, February 2026 — abstract read at nber.org/papers/w34836. Nearly 6,000 senior executives across US, UK, German and Australian firms; 69 percent of firms actively using AI; executives' own usage averaging 1.5 hours a week; "nine-in-ten reporting no impact on employment or productivity" over three years; and the same executives predicting a 1.4 percent productivity gain, 0.8 percent output gain and 0.7 percent employment cut over the next three.
• Salomé Baslandze, Zachary Edwards, John Graham et al., "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives," NBER Working Paper 34984, March 2026 — abstract read at nber.org/papers/w34984. A separate, smaller survey of roughly 750 executives reaching a partly different answer — positive labour productivity gains concentrated in high-skill services and finance, and an explicitly named "productivity paradox, in which perceived productivity gains are larger than measured productivity gains." Recorded here because the two 2026 surveys disagree and a reading citing one should know the other exists.
• **US Bureau of Labor Statistics, Productivity and Costs, Second Quarter 2026, Preliminary, released 6 August 2026** — nonfarm business labour productivity up 2.2 percent from Q2 2025 to Q2 2026 on 2.5 percent output growth and 0.2 percent hours growth; 1.4 percent at an annualised rate for the quarter; and 2.1 percent annualised over the current business cycle since Q4 2019. The BLS site refused my requests, so these figures come from search results reporting the release rather than from the release itself, and should be re-checked before being quoted in a reading.
• Scott Davis, "International comparisons show AI effect on productivity," Federal Reserve Bank of Dallas, July 2026 — 2.4 percent annualised US labour productivity growth since the start of 2024 against 1.6 percent in the five pre-pandemic years; information, finance and insurance, and professional and technical services at 16 percent of hours and 40 percent of productivity gains since the start of 2024; and the author's own statement that this is correlation rather than causation. The Dallas Fed site returned 403 to me; taken from search results reporting it.
• Daron Acemoglu, "The Simple Macroeconomics of AI," NBER Working Paper 32487, April 2024 — the ceiling estimate of no more than a 0.66 percent TFP increase over ten years, revised below 0.55 percent once hard-to-learn tasks are accounted for, built from roughly 20 percent of US labour tasks AI-exposed, 23 percent of those profitably automatable, and average task-level labour cost savings around 27 percent. Included as the standing quantitative counterweight to the lag argument: it says the payoff is small rather than late.
• **Warren D. Devine Jr., "From Shafts to Wires: Historical Perspective on Electrification," Journal of Economic History 43(2), June 1983, pp. 347–372** — David's principal source for the drive-system account. The diffusion figures (under 5 percent of primary horsepower in manufacturing in 1899, about 50 percent by 1919, about 75 percent by 1929) are established; David's own citations to Devine's Tables and to pp. 351, 354, 362ff and 363ff are recorded from the 1990 paper. I did not obtain Devine.
• Intel 1103 and Intel 4004 dates — the 1103, a 1,024-bit DRAM, introduced October 1970; the 4004 announced 15 November 1971. Used only to date David's anchors, not to make any claim about the chips.
• Edison central station dates — Holborn Viaduct, London, operating from January 1882; Pearl Street, New York, from 4 September 1882. Same purpose.
• **Petra Moser and Tom Nicholas, "Was Electricity a General Purpose Technology? Evidence from Historical Patent Citations," American Economic Review 94(2), May 2004, pp. 388–394**, and the reply literature including "In Defense of Electricity as a General Purpose Technology" (CIRCLE/Lund working paper 2007/06). Cited only to establish that the GPT status of electricity has been contested in print. I read neither.

Attributed, not endorsed.

• The 2025–2026 genre of dynamo-invoking AI commentary — including the AEI essay "The Dynamo, the Computer, and ChatGPT: Explaining Today's Productivity Paradox" and a large volume of consultancy and newsletter writing making the same move. Referenced in section 4 as the object of the criticism, not as a source of fact. Titles seen in search results; I did not read them, and section 4's characterisation of the genre rests on the anchor-shifting that is visible in the primary literature itself rather than on any one of these pieces.
• MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025" — the source of the widely repeated "95 percent of enterprise GenAI pilots fail." Recorded here as a caution rather than as evidence: it is preliminary work, not peer-reviewed, built on 52 structured interviews and 153 survey responses with a narrow six-month ROI definition of success, and it has drawn substantial methodological criticism. It is the number most often paired with a dynamo citation, which is why a reading should know its provenance before using either.

What I could not verify, and left out.

• The 1989 CEPR paper itself. Computer and Dynamo: The Modern Productivity Paradox in a Not-Too-Distant Mirror, CEPR Publication No. 172, Stanford, July 1989 — the full-length work containing Tables 3, 4 and 5, Figure 14 and Appendix A, which is where every number in the 1990 paper actually lives. The AgEcon Search copy returned 403 and I found no other. Every quantitative claim attributed to David above is therefore attributed at one remove, from a seven-page summary of tables I have not seen. A reading that leans hard on the half-of-the-5-percentage-points figure should obtain the 1989 paper first.
• David's own retrospectives. "Digital Technology and the Productivity Paradox: After Ten Years, What Has Been Learned and What Do We Need to Know?" (White House Conference on Understanding the Digital Economy, 25–26 May 1999) and "Productivity Growth Prospects and the New Economy in Historical Perspective" (EIB Papers, 2001). Both exist, neither was obtained, and consequently no self-grade is recorded in section 3. This is the largest gap in the entry. The rule jobs/canon.md states for prediction — that an entry quietly repeating one grade is worth less than one showing both — is not binding on an interpretation, but it is the right standard here and this entry does not meet it. There is one grade above and it is mine.
• Whether Fiszbein et al. has been published in a peer-reviewed journal. I found NBER WP 28076 (November 2020) and a February 2024 author revision and no journal version. This matters for how much weight section 3 should carry.
• Any direct comparison of the 2024–2026 productivity acceleration against the 1995–2004 one on like terms. The Fed pieces cited above measure different windows on different bases; I did not find a single source putting the two accelerations side by side with a consistent method, which is the thing a reading would most want and is worth knowing does not appear to exist yet in a form I could reach.
• Paul David's dates. He died in January 2023, aged 87. The British Academy memoir gives 24 March 1935 – 22 January 2023; other summaries give 24 May 1935 and 23 January 2023. I did not resolve the discrepancy, and since neither date bears on anything in this entry, neither is asserted in the prose.

On the boundary. This entry names human-computers, nordhaus-price-of-light-1994, printing-press-1440 and expert-systems-collapse-1987, all of which exist in canon/, and creates no ids. It places no needle, no score and no landmark, deposits nothing in history/, and refers to the 2026-08-20 reading and to METHODOLOGY.md §4b by date and section only. Nothing here is evidence.