AI as a General-Purpose Technology

AI’s capability is racing ahead of organizations’ ability to absorb it. Frontier-model training compute has grown about fivefold a year since 2020,⁹ while the share of U.S. work hours actually assisted by generative AI sits between 1% and 5%.⁸ Two readings of that gap are on offer. The general-purpose-technology literature reads it as a stage in a known process, because every economy-transforming technology has opened the same gap for the same structural reason and closed it through decades of complementary invention.¹⁴ The skeptics read it as an early verdict, and argue the gains will be small¹⁸ or that no transformative application has yet appeared.¹⁹ Deciding between the two takes a kind of evidence the adoption statistics do not carry, and the corpus behind this paper holds only its beginnings.

A general-purpose technology is defined by the complementary invention it forces

The formal concept comes from Timothy Bresnahan and Manuel Trajtenberg, whose 1995 paper General Purpose Technologies: “Engines of Growth?” gave the idea its working definition.¹ They argued that “whole eras of technical progress and growth appear to be driven by a few ‘General Purpose Technologies’ … such as the steam engine, the electric motor, and semiconductors.”¹ Three properties mark the category. A GPT is pervasive, with potential uses across a wide range of sectors. It is technologically dynamic, with an inherent potential to keep improving over a long period. And it generates what the authors call innovational complementarities: “the productivity of R&D in a downstream sector increases as a consequence of innovation in the GPT,” so advances in the core technology raise the payoff to invention everywhere it is applied.¹

The third property carries the rest of the argument, because it says where the value comes from. Bresnahan and Trajtenberg were explicit that a GPT arrives unfinished. “Most GPT’s play the role of ‘enabling technologies’, opening up new opportunities rather than offering complete, final solutions.”¹ For illustration they reached for electricity, where “the productivity gains associated with the introduction of electric motors in manufacturing were not limited to a reduction in energy costs. The new energy source fostered the more efficient design of factories.”¹ The motor delivered energy savings; the redesign delivered the value.

They also warned that the same property creates a coordination problem. The complementary innovations are “widely dispersed throughout the economy,” so a decentralized market can produce “too little, too late” innovation, and “institutions display much more inertia than leading technologies.”¹ That institutional inertia is where the historical lag lives.

Steam and electricity paid off through their co-inventions, and economists file AI in the same category

AI invites comparison to steam and electricity because economists place all of them in one formal category. Brynjolfsson, Rock and Syverson state it plainly. “The steam engine, electricity, the internal combustion engine, and computers are each examples of important general purpose technologies. Each of them increased productivity not only directly but also by spurring important complementary innovations.”⁴ Steam pumped water from mines, and it also “spurred the invention of more effective factory machinery and new forms of transportation … standard time, which was needed to manage railroad schedules.”⁴ Most of the value sat in those co-inventions.

The economic historian Nicholas Crafts built his 2021 review on the same comparison, examining steam, electricity and ICT as GPTs that had big effects but only with a lag, substantial in the first two cases, and arguing that AI, like the First Industrial Revolution, may matter most as “an invention of a new method of invention” that raises the productivity of R&D itself.¹⁰ That phrasing traces back to Zvi Griliches on hybrid corn, the lineage Bresnahan and Trajtenberg invoke.¹

The technology industry adopted the same frame. In a March 2017 talk at the Stanford Graduate School of Business, Andrew Ng argued, in his words: “Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don’t think AI will transform in the next several years.”⁷ The compressed slogan that travelled, “AI is the new electricity,” is the Stanford writeup’s framing of that talk; the quotation above is the verbatim line.⁷ Ng himself paired the optimism with constraints, naming data and talent scarcity as the two principal obstacles to adoption.⁷

Electrification lifted manufacturing productivity four decades after the dynamo, once factories were rebuilt around it

The most-cited historical anchor for the lag is Paul David’s 1990 essay The Dynamo and the Computer.² David documented that electricity stayed a curiosity in the productivity data long after the technology was proven. “In 1899 in the United States, electric lighting was being used in a mere 3 percent of all residences (and in only 8 percent of urban dwelling units); the horsepower capacity of all … electric motors installed in manufacturing establishments … represented less than 5 percent of factory mechanical drive.”² It would, he wrote, “take another two decades, roughly speaking, for these aggregate measures … to attain the 50 percent diffusion level.”² His memorable line — that in 1900 the dynamos were “everywhere but in the productivity statistics!” — became the template for every subsequent paradox.²

The span itself is the finding. Factory electrification “did not … have an impact on productivity growth in manufacturing before the early 1920s … four decades after the first central power station opened for business.”² That sentence is the scholarly basis for the figure often rounded to “about forty years”; David’s own word is “four decades.”

The delay had a mechanism, and David named it. Early adopters retrofitted electric motors onto the existing “group drive,” the central shafts and belts inherited from steam, and captured only energy savings.² The large gains waited until factories were rebuilt around “unit drive,” with a motor on each machine, which freed the floor plan from the geometry of the old transmission line.² David and Wright’s 1999 companion paper tracks the diffusion. The secondary-motor share of installed manufacturing horsepower rose from just over 50% in 1919 to nearly 80% in 1929, and in that same window manufacturing total factor productivity surged “more than five percent per annum between 1919 and 1929,” with labor-productivity trend growth jumping from 1.5 percentage points a year over 1899–1914 to 5.1 over 1919–1929.³ The surge kept the rewiring’s timetable, four decades behind the invention’s.

The paradox recurs because co-invention is invisible to the statistics while it is being paid for

The lag has a name in the modern era — the “Solow paradox,” after Robert Solow’s much-quoted late-1980s observation that the computer age was visible everywhere except in the productivity statistics. (That exact Solow line is not held as a primary source in this corpus and so is not reproduced here as a verbatim quotation; the verifiable in-corpus antecedent is Paul David’s parallel remark that in 1900 the electric dynamos were to be seen “everywhere but in the productivity statistics!“²) Brynjolfsson, Rock and Syverson revived the paradox for AI in 2017, observing that “systems using artificial intelligence match or surpass human level performance in more and more domains … yet measured productivity growth has declined by half over the past decade.”⁴

They weighed four candidate explanations for the clash: false hopes, mismeasurement, redistribution, and implementation lags.⁴ The first of those deserves more attention than it usually gets, because it is the null hypothesis every lag argument has to defeat. If the technology simply cannot deliver, patience will never turn the statistics around. Brynjolfsson, Rock and Syverson argued against it on general-purpose-technology grounds, concluding that “lags have likely been the biggest contributor to the paradox” because, “like other general purpose technologies, their full effects won’t be realized until waves of complementary innovations are developed and implemented.”⁴ They were nearly as firm about mismeasurement, finding that “a set of recent studies provide good reason to think that mismeasurement is not the primary explanation for the productivity slowdown.”⁴ False hopes did not die in that paper; the hypothesis returns below in its strongest current form.

Their 2018 follow-up formalized why the statistics mislead in the meantime. The productivity J-curve shows that when firms adopt a GPT they divert measured resources into unmeasured intangible capital, “new processes, products, business models and human capital.”¹¹ Early on that depresses measured productivity, because effort is going into things the national accounts cannot see; later, when “the benefits of intangible investments are harvested,” measured productivity is flattered.¹¹ The error in measured total factor productivity therefore traces a J. Empirically, the authors’ intangibles-adjusted TFP measure ran 11.3% above the official measure at the end of 2004 and 15.9% above it at the end of 2017, driven mainly by software intangibles. Those are software-era estimates, from the last GPT cycle.¹¹ For AI, the same paper assessed how AI-related intangible capital may currently be affecting measured productivity and found the effects “small but growing.”¹¹ They place electrification directly inside the frame: “it took a generation for the nature of factory layouts to be re-invented in order to fully harness the new technology’s benefits.”¹¹

Read together, the two papers fit. The productivity slowdown is real, and lags are likely its biggest contributor; the J-curve describes the measurement shadow a lag casts while intangible investment accumulates. For AI, that shadow is so far small.

The co-invention bill lands on the organization, and three-quarters of companies have yet to show value

Co-invention is the name for what the lag buys. The J-curve paper is explicit that realizing GPT potential “requires large intangible investments and a fundamental rethinking of the organization of production itself. Firms must create new business processes, develop managerial experience, train workers.”¹¹ Every item on that list is acquired inside the adopting firm, slowly and at its own expense.

The sharpest evidence on what that work returns comes from a field experiment. Brynjolfsson, Li and Raymond followed 5,179 customer-support agents after a generative-AI assistant was deployed and measured a 14% average productivity gain — 34% for novice and lower-skilled workers, “minimal” for the most experienced — because the system “disseminates the best practices of more able workers.”¹⁴ The gain is real and measured. Its size and its placement were set by deployment decisions, by who got the assistant and how the work was organized around it, and those decisions belong to the adopting firm.¹⁴

The management surveys corroborate this at company scale, and they carry lighter evidentiary weight; the most-quoted of them is a consultancy press release. BCG’s 2024 survey of 1,000 executives across 59 countries found that 74% of companies “have yet to show tangible value” from AI, and that the leaders who do generate value spend their effort on a 70-20-10 split: roughly 70% on people and processes, 20% on technology and data, and 10% on algorithms.¹² The earlier MIT Sloan Management Review–BCG study of 3,000-plus managers points the same way. Only about 1 in 10 organizations got significant financial benefit from AI; “getting basics right (data, technology, talent, strategy)” lifted the odds only to 20%, and the differentiator was the organization’s capacity to learn with AI.¹³

Capability is compounding at about 5× a year while 1–5% of work hours use AI

On the capability side, one number carries the point. Epoch AI reports frontier-model training compute “growing at 5× per year since 2020,” a doubling roughly every 5.2 months and a cumulative increase of about 10,000× among the top five models.⁹ Investment has followed the capability curve, with global corporate AI investment more than doubling in 2025.⁵

On the realized side, the most careful measure is Bick, Blandin and Deming’s nationally representative survey. As of late 2024, “between 1 and 5 percent of all work hours are currently assisted by generative AI”; users report time savings of 5.4%, which works out to 1.4% of total work hours once non-users are counted.⁸ The firm-side data confirm the same order of modesty. The Census Bureau’s Business Trends and Outlook Survey put actual AI use at 17–20% of U.S. businesses over December 2025–May 2026, with a steep size gradient running from 37% of firms with 250-plus employees down to under 20% of the smallest,⁶ and the Federal Reserve’s synthesis of adoption surveys lands at about 18% of firms on a firm-weighted basis.¹⁵ These are different kinds of number — a growth rate of an input on one side, shares of a stock on the other — and the figure below carries the same warning about its own bars. What they jointly establish is direction. The capability and investment curves are compounding; the realized-use figures are low and moving slowly.

One apparent contradiction in the adoption data resolves cleanly, and resolving it shrinks the gap without closing it. Stanford HAI’s 2026 AI Index reports 88% of surveyed organizations using AI in at least one business function, up from 78% a year earlier,⁵ while the Census counts 17–20% of firms. The Federal Reserve’s note explains the spread as a weighting effect. Most U.S. firms are small, so a firm-weighted count runs low, while surveys that skew toward large organizations run high; on the Fed’s own numbers, about 18% of firms had adopted AI by the end of 2025 while 78% of the labor force works at a firm that has.¹⁵ The honest version of the gap survives the reconciliation. Adoption reaches most of the economy’s employment, and depth of use stays thin nearly everywhere.

The depth measures agree with each other. The AI Index finds 70% of organizations using generative AI in at least one function, yet “AI agent deployment was in the single digits across nearly all business functions,” topping out at 24% scaled use in software engineering.⁵ Anthropic’s Economic Index, which maps roughly a million Claude conversations onto U.S. Department of Labor task data, shows the same shape at task level. About 36% of occupations demonstrate AI use across at least a quarter of their tasks, while only about 4% use it for three-quarters or more.¹⁶ Use is wide and shallow. On the historical pattern above, that is what a GPT looks like mid-co-invention, and it is also what an over-hyped technology can look like from the inside; the skeptics’ reading below takes that second possibility seriously.

One more number belongs in the ledger, because the firm-side statistics cannot see it. The AI Index estimates U.S. consumer surplus from generative AI at $172 billion annually by early 2026, up from $112 billion a year earlier, and notes that the figure dwarfs estimated U.S. generative-AI revenues; most of the tools “remain free or close to it.”⁵ Consumer surplus is value received above price paid, so by construction neither revenue lines nor productivity statistics capture it. Real value is accruing where the standard instruments do not point. That cuts in the measurement thread’s favor, and it settles nothing about the firm-side lag.

The skeptics argue the gains will be small, and the record cannot yet refute them

The size of the eventual payoff is genuinely disputed, and the corpus holds both bookends. On the optimistic side, McKinsey estimates generative AI could add $2.6–4.4 trillion annually across 63 use cases, rising to $6.1–7.9 trillion including broader knowledge work, with about 75% of the value concentrated in customer operations, marketing and sales, software engineering, and R&D.¹⁷ McKinsey’s own productivity figure is explicitly adoption-gated, at 0.1–0.6 percentage points of annual labor-productivity growth through 2040, “depending on the rate of technology adoption and redeployment of worker time.”¹⁷ The trillions are a potential; the per-year productivity number is what survives the implementation lag.

Daron Acemoglu’s objection is a different kind of claim from a dispute over timing, and it deserves stating on its own terms. His task-based model works from task exposure and average cost savings per task, aggregated through Hulten’s theorem into a macroeconomic total. On those inputs, AI delivers “no more than a 0.66% increase” in total factor productivity over ten years, and less than 0.53% once the estimate adjusts for early AI’s concentration in easy-to-learn tasks.¹⁸ The lag does no work in that arithmetic. The total is small because the task coverage and per-task savings he estimates are small.¹⁸

Jim Covello of Goldman Sachs goes further, into the false-hopes branch itself. In the bank’s 2024 Top of Mind issue (captured for this corpus through secondary coverage, as reference 19 discloses), Covello set more than $1 trillion in projected AI capex against his reading of the applications: “eighteen months after the introduction of generative AI … not one truly transformative — let alone cost-effective — application has been found.”¹⁹ That is a denial that electricity is replaying at all. If he is right, the historical analogy fails at its first premise, and the gap this paper has been describing is a write-down in progress.

The lag thesis has two answers on the record, and an honest gap. Brynjolfsson, Rock and Syverson weighed false hopes directly against the alternatives and judged lags the larger contributor, on the general-purpose-technology grounds set out above.⁴ And the task-level evidence shows the technology already delivering measured value where the organizational work has been done; the 14% lift among support agents came out of a deployed system, and the consumer-surplus estimate is realized value at population scale.¹⁴⁵ Against Acemoglu specifically, the lag thesis can note that his numbers rest on estimates of today’s task exposure and per-task savings, while the defining GPT property, innovational complementarity, works by raising the payoff to downstream invention and so expanding what the technology can profitably touch.¹ That reply is only as strong as the classification behind it, and the corpus contains no measurement that refutes his arithmetic. Acemoglu’s model and Covello’s observation stand as the live alternative, and the aggregate statistics that would retire them do not yet exist.

Separating a mid-J-curve from false hopes means watching co-invention, and that evidence is only starting to arrive

Every failed technology also ran a gap between capability claims and realized value, so the gap alone proves nothing. The two stories separate on one variable. Under the lag story, the quiet years are accumulation, with firms building the unmeasured complements the J-curve names: new processes, business models, human capital.¹¹ Under false hopes, the same quiet years are just spending. The place to look is therefore the organizational side.

Adoption speed, in particular, cannot decide it. Generative AI’s work adoption has run “as fast as the personal computer,“⁸ and population adoption reached about 53% within three years, “faster than the personal computer or the internet.”⁵ But fast access is a fact about distribution; the tool arrives through a browser, where the dynamo arrived through a capital installation. Co-invention keeps its own clock. The corpus supports the claim that adoption is unusually fast; it does not yet contain evidence settling whether the organizational lag is correspondingly short.¹¹

And the electrification record itself was not uniform across economies. David and Wright declined to offer “a purely technological explanation of the productivity surge of the 1920s” and built their account on “both the generic and the differentiating aspects of U.S. industrial electrification in comparison with that of the contemporary UK.”³ The same dynamo diffused into two economies along two different paths. The four decades of the American case measure one episode, and they set no law; the corpus holds no measured second lag to compare.

What the corpus already holds are early readings of the co-invention variable. AI-related intangible effects on measured productivity are “small but growing.”¹¹ Leaders in the BCG survey report the 70-20-10 effort split that co-invention predicts, and the depth measures, with agent deployment in the single digits, remain where a young J-curve would put them.¹²⁵ Those readings are consistent with the early years of a lag. They are also consistent with a modest technology being modestly used. The variable that will separate the stories is whether the intangible investment keeps growing and the depth measures follow it.

What the verified record supports

Read through the GPT lens, a wide gap between capability and realized value is the profile of a technology mid-diffusion; electricity showed the same gap for decades before the 1920s surge.²⁴ The measurement corollary should be stated exactly as strongly as its source states it. AI-related intangible effects on measured productivity are so far “small but growing,“¹¹ and the software era shows how large such effects eventually became, with adjusted TFP running 15.9% above the official measure by the end of 2017.¹¹ Today’s aggregate statistics are therefore a weak verdict in either direction. If the lag story is right, they understate what is being built. If the skeptics are right, they are already telling the truth.

On where returns concentrate, the corpus speaks with one voice. The field experiment and both management surveys find value accruing to the firms that do the organizational work, with leaders putting roughly 70% of their effort into people and processes.¹²¹³¹⁴ That is Bresnahan and Trajtenberg’s co-invention and David’s unit-drive rewiring, visible in current data.¹²

The magnitude of the eventual payoff stays honestly uncertain, and the corpus deliberately holds bookends an order of magnitude apart, from Acemoglu’s sub-0.66% to McKinsey’s trillions.¹⁷¹⁸ The narrower claims are the ones the verified record will carry. AI fits the formal definition of a general-purpose technology.¹⁴ Its capability is compounding far faster than firms are absorbing it.⁹⁶ And on the record of the last two GPT cycles, the step that gates the payoff is the slow, expensive, mostly intangible work of rebuilding the organization around the tool.¹¹ Whether that work is now under way at scale, or being skipped, is the question the next few years of depth data will answer.

References