Economists call a technology like the electric motor or the computer a general-purpose technology: one that spreads across many industries, keeps improving, and enables further inventions. History suggests such technologies reach the productivity statistics slowly. Paul David found that in 1899 electric motors supplied under 5 percent of US factory mechanical drive, and it took roughly two decades to reach half, about four decades after the first central power station opened.
Key takeaways
- Timothy Bresnahan and Manuel Trajtenberg describe general-purpose technologies as pervasive, improvable over time, and able to create complementary innovations.
- US factories needed about two decades after 1899 to move from under 5 percent to roughly half of mechanical drive powered by electric motors.
- The delay came from sunk investment in steam and water power and from the need to redesign factories, not from a lack of working motors.
- A 1987 remark by Robert Solow about computers and productivity statistics echoed the electricity pattern, according to David's 1990 paper.
- The electricity precedent is an analogy and does not predict what AI will do or when.
What "general-purpose" means to economists
In an NBER working paper issued in August 1992, Bresnahan and Trajtenberg proposed that whole eras of technical progress and growth are driven by a few key technologies. They gave the steam engine and the electric motor as past examples and semiconductors and computers as then-current ones. Such technologies share three features: they are pervasive, they can be improved over time, and they create what the authors call innovational complementarities, through which improvements in the technology spread across the economy and raise productivity broadly. The label is a description of a pattern, not a verdict on any one technology. Whether a technology qualifies is judged after the fact, by how widely it spread and what it made possible.
Applying the label to artificial intelligence is therefore a hypothesis. It can be tested against the evidence of earlier cases, and electricity is the case most often used.
How slowly electricity reached the factory floor
Paul David, writing in the American Economic Review in May 1990, collected the timeline of US electrification. The figures below come from his paper.
| Year | Event or measure |
|---|---|
| 1879 | Edison and Swan introduce the carbon filament incandescent lamp |
| 1881 | Edison central generating stations open in New York and London |
| 1899 | Electric lighting in 3 percent of US residences (8 percent of urban dwellings); electric motors under 5 percent of factory mechanical drive |
| 1914 to 1917 | Regulated regional electricity rates fall substantially against the general price level, and momentum builds |
| Early 1920s | Slightly more than half of factory mechanical drive capacity is electrified |
The invention itself was not the bottleneck. A working lamp and a central station existed by 1881. What lagged was the spread through the capital stock of the economy.
Why the gains waited for new factory designs
David identifies two reasons. The first was money. Replacing still serviceable plants built around water and steam power was unprofitable, so the fastest adopters were industries that were expanding and building new plants. The second was design. Early electrification often used the group drive system, in which a motor turned a section of shafting that ran several machines, so old belts and shafts stayed in place and added to capital without much gain in output per unit of input.
The unit drive, with an individual motor for each machine, allowed more. Factories no longer needed heavy overhead bracing for shafting, so they could be built lighter and on a single story. Layouts could follow the flow of materials rather than the line of the shaft, and a stoppage in one department no longer shut down the whole power system. Those gains required building differently, which took years.
The computer comparison and its limits
David opened his paper with the remark attributed to Robert Solow that computers are seen everywhere but in the productivity statistics, and argued that the situation was less unprecedented than it looked. In 1900, he wrote, an observer might equally have said that dynamos were everywhere but in the productivity statistics. He also warned that the analogy between information technology and electrical technology would have many limitations if taken literally. He noted, for instance, that computers had spread faster in their early decades than dynamos had in theirs.
Both points carry over to AI. The precedent shows that a long gap between invention and measured gains is possible. It does not show that a gap will occur for AI, how long it would last, or what organizational change would end it.
Indicators that separate a pattern from a slogan
- Diffusion: what share of organizations, and which kinds, use the technology in daily work, from an official survey with a stated question.
- Complementary change: whether workplaces, job roles, or processes have been redesigned, not only supplied with a new tool.
- Measured output: whether productivity statistics for the sector in question moved, and over what time span.
- Duplication: whether old and new methods still run side by side, which David noted in early computer use.
- Source type: whether a claim comes from a controlled study, an official statistic, or a vendor forecast.
FAQ
Is AI officially classified as a general-purpose technology?
No agency assigns the label. It is an academic category, and economists apply it with the benefit of hindsight. For AI the question is still being argued.
Does the electricity case mean AI productivity gains will take 40 years?
No. The 40-year gap describes electric power in US factories under its own costs, prices, and building stock. It is an example of how long a transition can take, not a forecast.
What did Solow actually say?
David attributes to Solow a remark that computers can be seen everywhere but in the productivity statistics. The remark dates from 1987 and is usually cited as an observation about measured productivity, not as a claim that computers were useless.
Why does the factory layout matter to the argument?
Because it shows that gains came from rebuilding work around the new technology, not from swapping one power source for another. The same question can be asked of any tool introduced into an unchanged process.
Related reading
- AI in data science work
- AI across several industries
- Ten AI technologies in use
- Deloitte AI Institute and industry applications
More in AI Basics.
Last verified: October 2026
Sources
- The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox, David P A, American Economic Review 80(2), May 1990
- General Purpose Technologies "Engines of Growth"?, Bresnahan T and Trajtenberg M, NBER Working Paper 4148, August 1992; Journal of Econometrics 65(1), 1995