The AI J-Curve: The Solow Paradox, Politely Restated
We are being told that AI is transforming the economy. We are also being told that aggregate productivity data does not yet show it. Both statements are true, and the reconciliation is not a mystery – it is a well-documented economic pattern with a name.
Erik Brynjolfsson, Daniel Rock and Chad Syverson called it the Productivity J-Curve. General-purpose technologies (GPTs) require large complementary investments in intangible capital – process redesign, human skills, new organizational structures – that are systematically under-measured in national accounts. During the build phase, measured productivity is depressed by the resource cost of those investments. Only later, when the intangible capital is put to work, does productivity accelerate. Their adjusted TFP measure is 15.9% higher than official figures by the end of 2017, once intangibles are capitalized.
Robert Solow’s original 1987 quip – you can see the computer age everywhere but in the productivity statistics – was made 25 years into the IT era. The productivity acceleration eventually arrived in 1995. Brynjolfsson, Rock and Syverson quantify the historical pattern: roughly a quarter century of slow productivity growth followed by a decade of acceleration, both for electrification and for IT. That is the base rate for GPTs. AI is currently five years into that curve at best.
The 2026 central bank commentary is remarkably consistent with the J-curve reading. St. Louis Fed President Alberto Musalem, on 28 May 2026, updated Solow almost verbatim: we see AI everywhere but in the aggregate productivity statistics, with St. Louis Fed staff estimating only a 25% probability that Q1 2026 was in a high-growth regime. San Francisco Fed President Mary Daly’s February 2026 letter argues that current AI deployments are analogous to replacing a steam-powered motor with an electric one but leaving the factory floor unchanged – good progress, not yet transformative. ECB Chief Economist Philip Lane observes that the aggregate signal remains elusive even as microeconomic evidence is encouraging. Bank of England Governor Andrew Bailey put it most bluntly in Sheffield: productivity may slow down in the early phases, because it is the experimenting phase. We may well be there on AI.
The data itself supports the same reading. US total factor productivity grew just 0.8% in 2025 after 1.5% in 2024, and the euro area continues to describe its productivity growth as “still relatively sluggish” (ECB Annual Report 2025). US labour productivity has held up better at 2.1% in 2025, but the BLS notes that the current cycle is running at the 1947-2026 long-run average, not above it. Meanwhile, capex on AI is unmistakably real: our own AIMG estimate places 2026 AI spending near USD 2.5 trillion. The investment is visible; the payoff is deferred by design.
Acemoglu’s counterview – a mere 0.66% TFP gain over 10 years – is a useful bound. Goldman’s 1.5pp 10-year global productivity boost is another. The IMF’s WP/25/76 sits between them, with a high-scenario 5.6% US GDP lift over a decade. The width of these ranges is itself the point: the aggregate return is uncertain, but the temporal shape – depressed then rising – is not.
AIMG View: The Solow paradox is not a rebuttal to AI. It is the historically normal shape of a GPT diffusion curve. Boards mistaking the trough for the destination will underinvest, then acquire the winners at a valuation set by others.
AIMG Bottom Line: Manage the J-curve as a portfolio: fund the intangible investments now, measure via microeconomic operating metrics rather than aggregate TFP, and expect the visible payoff on a five-to-ten-year horizon, not a two-year one.