General-Purpose Technologies and the Productivity J-Curve

In 1899, electric motors supplied about 5 percent of the mechanical drive in American manufacturing. By 1919 the share was 55 percent, and by 1929 about 80 percent. The large productivity gains did not track the installation curve. They arrived after 1919, once factories stopped bolting motors onto line-shaft layouts designed for a steam engine and rebuilt around unit drive, with a motor on each machine and floors arranged by workflow rather than by proximity to a shaft (David, 1990, The Dynamo and the Computer, AER 80(2):355-361). David's paper took up the paradox Robert Solow had named in 1987, that the computer age was visible everywhere but in the productivity statistics. The same question is now asked about AI.

What makes a technology general-purpose

Bresnahan and Trajtenberg defined general-purpose technologies by three properties: pervasiveness across many sectors, inherent scope for continued technical improvement, and innovational complementarities, meaning that improvements in the GPT raise the returns to innovation in the sectors that use it, and vice versa (Bresnahan & Trajtenberg, 1995, General Purpose Technologies: "Engines of Growth"?, Journal of Econometrics 65(1):83-108).

The third property has the economic bite. Complementarities create a coordination problem: the GPT sector under-invests because it cannot capture users' gains, and users wait for the GPT to mature. Diffusion is slow and depends on co-invention by users, exactly the factory redesign David documented.

The J-curve mechanism

Brynjolfsson, Rock and Syverson formalised why that co-invention makes productivity look worse before it looks better (Brynjolfsson, Rock & Syverson, 2021, The Productivity J-Curve, AEJ: Macroeconomics 13(1):333-372). Firms adopting a GPT spend heavily on intangibles: new processes, training, data, reorganisation. National accounts largely do not record these as output. Later, the intangible stock produces measured output, but it is not counted as an input.

Let \(g_A\) be measured total factor productivity growth (the standard Solow residual) and \(g_A^*\) true growth when intangibles are counted. Write \(Y\) for measured output, \(I_U\) for unmeasured intangible investment with shadow price \(\phi\), \(U\) for the intangible capital stock with rental rate \(r_U\), \(rK\) and \(wL\) for the payments to measured capital and labour, and \(g_x\) for the growth rate of \(x\). Their decomposition is

The first term is negative when intangible investment grows faster than measured productivity: output is being produced and not counted. The second is positive when the intangible stock grows: its services are credited to productivity rather than to an input. Early in adoption the first term dominates; later the second does. Plot measured minus true productivity over time and you get the J.

A toy calculation shows the scale. Let \(Y = 100\), \(\phi I_U = 5\) growing at 20 percent a year, a measured residual of 1 percent, and intangible capital income \(r_U U / Y = 3\) percent growing at 10 percent. The first term is \(\tfrac{5}{105}(1 - 20) \approx -0.90\) points, the second \(\tfrac{100}{105}(0.03)(10) \approx +0.29\), so measured TFP growth understates the truth by about 0.6 points a year. Once investment plateaus (\(g_{I_U} = 0\)) with \(r_U U/Y = 8\) percent growing at 5 percent, the terms become \(+0.05\) and \(+0.38\): an overstatement of about 0.4 points.

Applied to US data, the authors estimate that adjusting for intangibles correlated with computer hardware and software puts the TFP level 15.9 percent above official measures by the end of 2017. They also compute that explaining the 0.55-point gap between US GDP growth in 2000-2007 and 2010-2017 would require about $107 billion a year of unmeasured investment.

Where economists disagree

The J-curve says a dip is consistent with a transformative technology. It does not say a dip implies one. Syverson himself showed that mismeasurement cannot explain the post-2004 US productivity slowdown in general: the slowdown appeared across dozens of countries regardless of their ICT intensity, and estimates of unmeasured consumer surplus fall far short of the missing output (Syverson, 2017, Challenges to Mismeasurement Explanations for the US Productivity Slowdown, JEP 31(2):165-186). The J-curve is a narrower claim about intangible capital, not about free digital goods, but the two are often conflated.

On AI specifically, Acemoglu's task-based estimate combines exposure shares with task-level cost savings and finds total TFP gains of no more than about 0.66 percent over ten years (Acemoglu, 2024, The Simple Macroeconomics of AI, NBER w32487). The disagreement is structural. A task-based model counts the tasks that exist; a GPT model expects the task list itself to be redesigned, as unit drive redesigned the factory. The productivity studies measure the first and cannot see the second.

When it breaks

The theory is hard to falsify in real time. A low productivity reading is consistent both with a J-curve in progress and with a technology that is not general-purpose. Only the later upswing discriminates, and by then the question is historical. Measuring diffusion honestly means refusing to cite the dip as evidence.

The intangible estimate rests on market values. The method infers unmeasured intangibles from how much firms' market value exceeds their measured assets. That premium also contains market power, expectations and mispricing, so the adjustment inherits whatever the stock market gets wrong.

Historical lags are not a forecast. Electrification's lag ran for decades after the first central stations of the early 1880s. The analogy says complementary investment will be required; it does not say how long AI's version will take, or that the upswing is guaranteed.

Aggregates hide who pays for the dip. The reorganisation phase shifts which tasks and skills are valued, and the adjustment costs fall on specific workers and firms well before any aggregate gain appears.