The Productivity J-Curve: The AI Payoff Looks Like Nothing, Then Everything
In the four quarters ending with the first quarter of 2026, utilization-adjusted total factor productivity in the US grew by 0.07 percent. Effectively zero. And yet, in the same period, the share of corporate earnings-call discussion devoted to AI rose to roughly 15 percent of all productivity talk, and 95 percent of it was in the future tense.
The gap between those two numbers is the modern sequel to one of the oldest puzzles in economics.
The Paradox That Keeps Coming Back
In 1987, Robert Solow quipped that you can see the computer age everywhere except in the productivity statistics. The joke aged into a law. Economists later gave it a pattern: when a genuinely general-purpose technology arrives, measured productivity first dips, then rises. Brynjolfsson, Rock, and Syverson formalized this as the Productivity J-Curve.
The shape is the point. A new GPT (general purpose technology) does not pay off immediately. Firms must first commit real capital and labor to things the national accounts cannot see: reorganized business processes, retrained workers, newly invented products, rewritten workflows. None of this shows up as output. It shows up as cost. So measured productivity looks worse than reality during the early phase, and only later, when the hidden assets start producing measurable output, does the statistic catch up and even overshoot.
History is littered with the dip. Electrification took a generation to pay off because factories had to be redesigned around the new power source, not merely rewired. The British industrial revolution produced a half-century of wage stagnation before the gains arrived. The pattern is so reliable that the J-curve has been proposed as a test: if a new technology does not generate large hidden intangibles, it probably was not general purpose to begin with.
The Invisible Balance Sheet
Why does the accounting miss so much? The mechanism is subtle. In growth accounting, productivity is a residual: output minus measured inputs. When a firm's market value rises by more than its observable investment, the gap reflects capital that was built but never recorded. These are the "intangible correlates" of visible spending, and they are large.
The paper's estimates are striking. Software carries intangible multipliers near ten: every visible unit of software investment drags roughly ten times as much invisible organizational capital along with it. Computer hardware is significant but smaller. Research and development, surprisingly, is nearly invisible in the mismeasurement, because it is a mature asset whose investment growth tracks its capital stock closely. The J-curve is not evenly distributed across all intangibles. It is concentrated where the complementary investment is largest, which today means software, and increasingly, AI.
The newest work pushes the same argument further. Brookings' "Counting AI" makes the case that generative AI is intangible capital that the national accounts still record as ordinary operating expense, and lays out a measurement agenda: an intensity index that tracks AI adoption from provider telemetry, and a reform of the national accounts to separate AI capital from AI operating cost. Older estimates put about 800 billion dollars a year of US business intangibles outside the accounts, leaving trillions of dollars of invisible capital stock. AI is the largest thing to ever sit on that invisible balance sheet.
The Twist Nobody Predicted
The July 2026 St. Louis Fed study adds a genuinely new wrinkle. The authors analyzed roughly 490,000 earnings call transcripts from 5,198 firms, using an LLM to classify how executives talk about productivity and AI. They found the near-universal future tense, and the overwhelmingly positive sentiment: 95 percent of AI-related productivity talk describes gains, versus 75 percent for non-AI talk. Optimism is not the interesting part. The interesting part is the authors' explanation for why even realized gains may stay invisible.
When AI makes some output radically cheaper to produce, that output simultaneously becomes less valuable. Generate a marketing campaign, an animation, a serviceable news article with a keystroke, and the price of each collapses precisely because everyone can now produce it. The task got easier; the output got cheaper. A real gain on one side of the ledger is erased by falling prices on the other, and the productivity math cancels itself out. Some things are going to become more abundant, which means they are also going to become less valuable.
This is a different failure mode from the one the J-curve describes. The J-curve hides investment that will pay off later. Abundance hides value that has already been transferred to consumers, and consumer surplus does not live in the productivity statistics at all. Both effects point the same way: the aggregate numbers understate what AI is doing, and they will keep understating it for a while.
The Dip Is the Signal
The temptation is to read the flat productivity line as evidence that AI is overhyped. The J-curve literature reads it as the opposite: the dip is the signature of a genuinely transformative technology mid-adoption. The firms are not reporting fantasy. They are reporting the early phase of a curve they have seen before, in the form of the computers they could see everywhere except in the statistics.
When the hidden intangibles finally mature, the same numbers that look flat today will overshoot. The mismeasurement runs both ways, and the over-correction is already priced into how we should read the current silence.
One second take-away
If you are looking for AI in the productivity statistics and not finding it, the problem is probably not AI and not the statistics. It is that you are measuring a J-curve at its dip, and mistaking the shape of adoption for the absence of value.
Data sources: The Productivity J-Curve (MIT IDE/NBER) | AI and Productivity: What Firms Are Saying on Earnings Calls (St. Louis Fed) | Counting AI (Brookings) | AI, productivity, and work (CEPR VoxEU)