AI Spending Is Surging. Productivity Growth Is 0.1-0.6%.

In 1987, Robert Solow wrote that you can see the computer age everywhere except in the productivity statistics. Brynjolfsson (1993) turned that observation into a research program, tracing the paradox through the 1970s and 1980s when American companies poured money into computing infrastructure and macroeconomic productivity barely moved. That was computers. We are now running the same experiment again with AI, and the early numbers look uncomfortably familiar.

Gartner forecasts worldwide AI spending will reach $2.5 trillion in 2026, up from $1.5 trillion in 2025. McKinsey's analysis of generative AI's economic potential estimates that GenAI will add 0.1 to 0.6 percent annually to labor productivity growth through 2040. Those two numbers sit next to each other and ask a question that no one has fully answered yet: where is all that spending going, and why is the productivity signal so small?

I want to be careful about how I frame this, because Brynjolfsson and Hitt (1996) actually resolved the original paradox, at least at the firm level. They showed that IT capital has a higher marginal product than non-IT capital, particularly when paired with complementary organizational changes like decentralization and worker empowerment. The economy-level paradox was partly an artifact of aggregation and partly a measurement problem. Value was being created; it just was not showing up in aggregate statistics because competitive markets were passing the gains to consumers as lower prices rather than to shareholders as profits. As Brynjolfsson and Hitt demonstrated, IT value and IT profitability are not the same thing.

But the resolution of the 1990s paradox also contained a warning that I think is going unreceived right now. The firm-level gains Brynjolfsson and Hitt found were concentrated in organizations that made complementary changes alongside their IT investments. The technology alone did not produce the gains. Reorganizing around the technology did. Organizations that bought computers without changing how decisions were made, how information flowed, or how workers were empowered saw weaker returns. The lesson was not that IT investment works. It was that IT investment works when paired with organizational restructuring.

That condition still holds, and I do not think most organizations deploying AI today are meeting it. McKinsey's 2025 State of AI report finds that 79 percent of organizations are using generative AI, up from 33 percent in 2023. But the same report says only 7 percent have fully scaled AI. The gap between those two numbers is the gap between using a tool and restructuring an organization around it. Buying a license and changing how decisions get made are different things.

The 0.1 to 0.6 percent productivity figure from McKinsey covers the period through 2040. That is a 15-year window. Brynjolfsson's lag explanation from 1993 said that IT benefits require organizational learning and restructuring, which take years to materialize. If the same lag logic applies to AI, and I think it probably does, then some of what we are spending on AI now will not show up in productivity statistics until the mid-2030s. Organizations are still learning how to restructure around AI, still building the data infrastructure, still developing the management practices that allow AI outputs to feed into actual decisions.

There is also a measurement problem that feels specific to AI in a way that goes beyond what the 1990s paradox involved. AI is concentrated in knowledge work: writing, coding, analysis, research, customer interaction. These are exactly the domains where productivity is hardest to measure. Output from a software engineer is not a widget count. Output from an analyst is not a line item in a ledger. When McKinsey maps automatable activities and estimates their economic value, the methodology necessarily involves some assumptions about how to price work that has never been priced at the task level. Whether a 30-minute analysis task that AI now does in two minutes represents a productivity gain depends entirely on whether the saved 28 minutes go into more valuable work or into attending another meeting.

I wrote about related questions in a post on the gap between AI potential and AI scale, where the focus was on organizational capability. Here I want to sit with the measurement problem a little longer. The productivity paradox in the 1990s was partly resolved by better measurement, including firm-level data that aggregate statistics missed. For AI, the equivalent would be measuring not just whether workers use AI tools but whether the time freed by AI gets redirected into demonstrably higher-value work. That is a harder study to run, and most organizations do not have the data infrastructure to run it even if they wanted to.

My read is that the 0.1 to 0.6 percent figure is a floor estimate under current conditions, not a ceiling on what AI will eventually produce. The lag is real. The measurement challenges are real. The complementary organizational changes are real and are only beginning. What Brynjolfsson taught me to watch for is not whether the spending is large. It is whether the organizational restructuring is following the spending. If it is not, the pattern from the 1980s will repeat, just with much larger dollar amounts on the left side of the equation.