There’s a 1990 paper I keep coming back to whenever someone tells me AI isn’t showing up in the productivity data. It’s not an AI paper. It’s about the electric dynamo.

Robert Solow wrote it about computers in 1987:

“You can see the computer age everywhere but in the productivity statistics.”

Swap “the computer age” for “AI” and the sentence still works, nearly forty years later. That’s the thing worth sitting with.

Forty years everywhere but in the statistics #

Paul A. David’s paper, The Dynamo and the Computer, makes a simple but unsettling argument. The electric dynamo was commercially viable by the 1880s. Edison’s Pearl Street station, the first central power plant in the United States, opened in New York in 1882. By 1900 the technical case for electricity was settled. And yet electric motors provided less than 5% of mechanical drive power in US factories as late as 1899. Productivity gains from electrification didn’t arrive until the 1920s, four decades after the first power stations.

“In 1900, contemporary observers well might have remarked that the electric dynamos were to be seen ’everywhere but in the productivity statistics.'”

The delay wasn’t because the dynamo was overhyped. It was because unlocking the technology’s real value required redesigning the systems around it. Nobody wanted to do that while existing infrastructure still worked.

The retrofit trap #

Early factories didn’t redesign around electricity. They retrofitted it. They kept the belt-and-shaft transmission systems in place, dropped electric motors in where steam engines used to be, and called it electrification. This was the “group drive” phase: rational in the short term, structurally limiting in the long run. The result was a hybrid that couldn’t fully exploit the new technology and looked worse on paper than before, because you were now capitalising two overlapping systems.

The productivity explosion only came when factories were redesigned from scratch around the “unit drive”: individual motors per machine, which made single-storey layouts possible, collapsed the overhead shaft infrastructure, and allowed production lines to be reconfigured without shutting everything down. The gains were architectural, not incremental.

AI is in its group drive phase #

This is where we are with AI. Most teams are in the group drive phase. AI is being bolted onto existing workflows: ticket systems, review processes, standups, IDEs designed for a different model of work. The process architecture underneath hasn’t changed, so the needle doesn’t move.

The real shift isn’t using AI more. It’s rebuilding the process under the assumption that AI is load-bearing infrastructure, not an assistant. That looks like async-first knowledge work, specification-driven development, pipelines designed to be read and acted on by agents. Fewer handoffs. Smaller coordination surfaces.

Brynjolfsson, Rock, and Syverson applied David’s framework directly to AI in 2017, examining the same gap between technological capability and measured output.

“Lags have likely been the biggest contributor to the paradox. The most impressive capabilities of AI […] have not yet diffused widely. Like other general purpose technologies, their full effects won’t be realized until waves of complementary innovations are developed and implemented.”

Their follow-on work formalized what David only implied. They called it the Productivity J-Curve: when a General Purpose Technology arrives, firms invest in reorganization, retraining, and new processes that are real and costly but don’t appear in national accounts. Measured productivity doesn’t stall. It falls before it rises. Recent field data fits the shape: a study of 5,172 customer support agents found a 15% average productivity gain from AI assistance, with gains around 30% for the least experienced workers and minimal impact for the most experienced. Task-level gains exist. The problem is diffusion, not absence.

Why the redesign won’t be forced #

David’s paper ends with a warning I find more useful than any of the optimism:

“Closer study of some economic history of technology… should help us avoid both the pitfall of undue sanguinity and the pitfall of unrealistic impatience.”

And there’s one asymmetry the dynamo analogy doesn’t fully cover. Physical factories depreciate. They age, become obsolete, get torn down, and that forces redesign. Information systems don’t work the same way. As David puts it, “one cannot depend on the mere passage of time to create occasions to radically redesign a firm’s information structures and operating modes.” The inertia is stronger. The forcing function is weaker.

Which means the redesign has to be deliberate. It won’t be structurally forced.

The two harder versions #

There’s a harder version of this. History suggests the productivity gains from technology transitions don’t just accrue to incumbents who adapt fastest. They often accrue to new entrants who build without any legacy to defend. When Paul David wrote this paper, a handful of TV networks commanded the nation’s attention, newspapers operated in every major city, and advertising meant Madison Avenue. None of those were displaced by competitors who adopted computers better. They were displaced by companies that didn’t exist in 1990. For AI, that asymmetry may be sharper still. The question isn’t only whether you redesign in time. It’s whether the upside goes to you at all, or to someone starting from a blank slate.

There is a harder version still. Daron Acemoglu’s 2024 macroeconomic analysis estimates total factor productivity gains from AI at no more than 0.66% over the next decade — well under 0.1% per year. His argument is not that AI doesn’t work. It’s that current gains come from easy-to-learn, well-defined tasks. Tasks where context matters, where there’s no objective outcome to optimize, are structurally more resistant. The dynamo analogy is directionally correct, but it assumes the unit drive factory delivers the same productivity multiple as electrification did. Acemoglu’s numbers suggest the ceiling may be lower, and the lag longer. That changes the calculus: the redesign is not just urgent. Its payoff is uncertain.

Either you build the unit drive factory now, or you spend the next decade wondering why the statistics never moved, and why a company you’d never heard of is doing your job.

Sources

• Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, American Economic Review, May 1990.
• Timothy F. Bresnahan and Manuel Trajtenberg, “General Purpose Technologies: ‘Engines of Growth?’”, Journal of Econometrics, 1995.
• Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics”, NBER Working Paper 24001, 2017.
• Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies”, American Economic Journal: Macroeconomics, 2021.
• Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work”, Quarterly Journal of Economics, 2025.
• Daron Acemoglu, “The Simple Macroeconomics of AI”, Economic Policy, 2025.