AI as a General-Purpose Technology

The argument in one line

AI’s capability is racing ahead of organizations’ ability to absorb it. That sentence is not a prediction; it is what the verified data already shows. Frontier-model training compute has grown about fivefold a year since 2020⁹ while the share of U.S. work hours actually assisted by generative AI sits between 1% and 5%.⁸ The interesting question is not whether the gap exists but whether it is a sign of failure or a stage of a known process. The general-purpose-technology (GPT) literature argues it is the latter: every economy-transforming technology has shown the same lag, for the same structural reason. This paper builds that case from the primary sources and then tests it against the most recent adoption data.

What makes a technology “general-purpose”

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,” and that such technologies share three properties:¹

• Pervasiveness — the potential for use across a wide range of sectors rather than one niche.
• Technological dynamism — an inherent potential to keep improving over a long period.
• 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 is the load-bearing one. Bresnahan and Trajtenberg were explicit that a GPT is rarely a finished solution: “Most GPT’s play the role of ‘enabling technologies’, opening up new opportunities rather than offering complete, final solutions.”¹ Their illustration was electricity itself — “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 value was unlocked by redesign, not by the motor.

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

The steam / electricity / internet analogy

The reason AI invites comparison to steam and electricity is not rhetoric; it is that economists place all of them in the same 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 did not only pump water from mines; it “spurred the invention of more effective factory machinery and new forms of transportation … standard time, which was needed to manage railroad schedules.”⁴ The co-inventions, not the engine, were where most of the value sat.

The economic historian Nicholas Crafts makes the same comparison the spine of his 2021 review, 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

This is also the frame the technology industry adopted. 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 rather than a verbatim sentence Ng spoke; the quotation above is the verbatim line.⁷ Notably, even Ng paired the optimism with constraints, naming data and talent scarcity as the two principal obstacles to adoption.⁷

The dynamo and the lag: roughly four decades

The single most-cited historical anchor for the lag is Paul David’s 1990 essay The Dynamo and the Computer.² David documented that electricity was 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 key figure is the gap itself. 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 This is the scholarly basis for the figure often rounded to “about forty years”; David’s own word is “four decades.”

Why so long? Because the payoff required re-architecting the factory. 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 came only when 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 it was in that same window that 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 (1899–1914) to 5.1 (1919–1929).³ The surge followed the rewiring, not the invention.

The Solow paradox and the productivity J-curve

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 their 2017 paper, 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 explanations — false hopes, mismeasurement, redistribution, and implementation lags — and concluded “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

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 (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 was 11.3% higher than the official measure at the end of 2004 and 15.9% higher at the end of 2017 — a large, persistent mismeasurement driven mainly by software intangibles.¹¹ They place the electrification case directly inside this 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

Co-invention: rewiring the organization, not installing the tool

If the lag is structural, its cause is co-invention: the complementary, mostly intangible investments — redesigned processes, new roles, retrained people, fresh business models — that a GPT requires before it pays. 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 The contemporary management evidence converges on the same point from a different direction.

Boston Consulting Group’s 2024 survey of 1,000 executives across 59 countries found that only 26% of companies “have developed the capabilities to move beyond AI proofs of concept and generate tangible value”; the other 74% “have yet to show tangible value.” Just 4% qualified as mature value leaders.¹² Crucially, BCG found leaders spend their effort on a 70-20-10 split: roughly 70% on people and processes, 20% on technology and data, and only 10% on algorithms — while laggards over-index on the technical layer.¹² The earlier MIT Sloan Management Review–BCG study of 3,000-plus managers reached a compatible conclusion: only about 1 in 10 organizations got significant financial benefit from AI, and “getting basics right (data, technology, talent, strategy)” lifted the odds to only 20%; the differentiator was the capacity to learn with AI, not the model itself.¹³

Field evidence shows the same texture at the level of individual work. Brynjolfsson, Li and Raymond’s study of 5,179 customer-support agents found a generative-AI assistant raised productivity 14% on average — but 34% for novice and lower-skilled workers and “minimal” for the most experienced, because the system “disseminates the best practices of more able workers.”¹⁴ The gain is real, but it is mediated by how the tool is deployed and to whom — which is an organizational design choice, not a property of the model.

What the lag means in numbers: capability vs. realized value

Place the current data side by side and the gap is unmistakable. On the capability side, Epoch AI reports frontier-model training compute “growing at 5× per year since 2020” — a doubling roughly every 5.2 months, a cumulative increase of about 10,000× among the top-5 models.⁹ Investment has followed: global corporate AI investment more than doubled in 2025, with private investment up 127.5% and generative-AI funding up more than 200%, per Stanford HAI’s 2026 AI Index.⁵

On the realized side, the picture is far more modest. Bick, Blandin and Deming’s nationally representative survey found that as of late 2024, “between 1 and 5 percent of all work hours are currently assisted by generative AI,” with self-reported time savings of 5.4% among users — equivalent to just 1.4% of total work hours once non-users are included.⁸ The U.S. Census Bureau’s Business Trends and Outlook Survey put actual firm-level AI use at 17–20% of businesses over December 2025–May 2026, with a steep size gradient: 37% of firms with 250-plus employees but under 20% of the smallest firms.⁶ The Federal Reserve, synthesizing multiple surveys, reported about 18% of firms (firm-weighted) had adopted AI by year-end 2025 — even as 78% of the labor force, employment-weighted, works at a firm that has.¹⁵

Adoption breadth and adoption depth diverge sharply. The 2026 AI Index reports that 88% of surveyed organizations use AI in at least one business function (up from 78% a year earlier) and 70% use 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, mapping roughly a million Claude conversations to U.S. Department of Labor task data, similarly found that only about 4% of occupations used AI for three-quarters or more of their tasks, with 57% of usage augmenting work and 43% automating it.¹⁶ Wide, shallow adoption is exactly what a GPT in its co-invention phase looks like.

How big is the prize — and how contested?

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 But McKinsey’s own productivity figure is explicitly adoption-gated: generative AI adds just 0.1–0.6 percentage points to 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.

On the skeptical side, Daron Acemoglu’s task-based model puts the total-factor-productivity gain from AI at “no more than a 0.66% increase … over 10 years,” falling below 0.53% once one adjusts for the fact that early AI targets easy-to-learn tasks.¹⁸ Goldman Sachs’s 2024 Top of Mind issue framed the same tension as a question — more than $1 trillion in projected AI capex against thin near-term returns — with Goldman’s Jim Covello arguing that “eighteen months after the introduction of generative AI … not one truly transformative — let alone cost-effective — application has been found.”¹⁹ These are real disagreements about magnitude and timing. They are not, however, disagreements about the mechanism: skeptic and optimist alike locate the constraint in adoption and task-level deployment, not in whether AI is a GPT.

Implications

Three things follow from reading AI through the GPT lens, each grounded in the verified record rather than forecast.

First, treat the gap as a stage, not a verdict. A wide gap between capability and realized value is the signature of a GPT mid-diffusion, not evidence the technology has failed — the same gap was visible in electricity for decades before the 1920s surge.²⁴ Judging AI by today’s aggregate productivity statistics repeats Solow’s error in real time, and the J-curve says those statistics are biased downward precisely now, while intangible investment is being made and not yet harvested.¹¹

Second, the binding constraint is organizational, and it is where the returns concentrate. The convergent finding across BCG, MIT-BCG, and the field-experiment evidence is that value accrues to the firms that rewire people and processes — leaders put ~70% of effort there — not to those that buy the best model.¹²¹³¹⁴ This is Bresnahan and Trajtenberg’s co-invention and David’s unit-drive reorganization, restated for 2026.¹²

Third, the magnitude of the eventual payoff is honestly uncertain. Credible estimates span an order of magnitude — from Acemoglu’s sub-0.66% TFP gain to McKinsey’s trillions — and the corpus deliberately holds both.¹⁷¹⁸ What the sources agree on is narrower and more useful: AI fits the formal definition of a general-purpose technology,¹⁴ its capability is advancing far faster than its absorption,⁹⁶ and on the historical record the rate-limiting step is the slow, expensive, mostly intangible work of rebuilding the organization around it.¹¹

References