Economics — Noozify Original — September 17, 2026

You Can See the AI Age Everywhere but in the Productivity Statistics

By Adam Canissario

Watch the story in half a minute

In July 1987, the economist Robert Solow — who would win the Nobel Prize three months later — reviewed a pair of books about the computer revolution and closed with a single line that has outlived everything else in the essay. "You can see the computer age everywhere but in the productivity statistics." American companies had spent two decades pouring money into computers, and the government's measure of output per worker-hour — the number that ultimately decides how fast living standards rise — had barely moved. Economists gave the puzzle a name, the Solow paradox, and spent the next decade arguing about it.

It is worth revisiting that argument now, because we are living through a nearly perfect rerun. Gartner projects $2.59 trillion in worldwide AI spending for 2026 — up 47 percent in a single year. Nvidia, the company selling the hardware underneath the boom, reported a $96 billion quarter, in August, more than double a year earlier. Meanwhile, one standard measure of US productivity, utilization-adjusted total factor productivity, grew just 0.07 percent over the four quarters ending in the first quarter of 2026, a near-standstill. And 95 percent of enterprise generative-AI pilots have produced no measurable effect on the bottom line. You can see the AI age everywhere but in the productivity statistics. The sentence works today without changing a word.

The instinctive explanations are the same ones offered in 1987, and they divide into two camps. Either the technology is overhyped, or the measurements are wrong. Both camps have a point, and both miss the more interesting answer, which comes from history.

The best guide is a paper the economic historian Paul David published in 1990, at the height of the original paradox, called "The Dynamo and the Computer." David pointed out that electricity had gone through exactly the same embarrassment. The electric dynamo was commercialized in the early 1880s; measurable productivity gains from electrification did not show up in American factories for roughly forty years. The reason was not that electricity didn't work. It was that factories had been built around steam, and steam shaped everything. A steam engine delivered power through a central shaft running the length of the building, so machines crowded around the shaft, floors were stacked vertically to stay close to it, and workflow was dictated by the plumbing. Early adopters simply unbolted the steam engine and wired an electric motor into the same layout — and got roughly nothing for it.

The gains arrived only when a generation of engineers designed factories around what electricity actually made possible: a small motor on every machine, arranged in whatever order the work required, in cheap single-story buildings organized around the flow of materials instead of the location of a shaft. The technology took a few years to install. The reorganization took four decades, because it required replacing buildings, retraining workers, rethinking management, and — not incidentally — waiting for the people who had grown up with steam to retire.

Computers followed the same script on a faster clock. The paradox Solow named in 1987 quietly dissolved in the mid-1990s, when US productivity growth roughly doubled and stayed elevated for a decade. What changed was not the computers; it was the reorganization catching up — supply chains, inventory systems, retail logistics and entire firms rebuilt around information flowing instantly instead of on paper. The economist Erik Brynjolfsson, who has spent his career on this question, calls the pattern the "productivity J-curve." A general-purpose technology first depresses measured productivity, because companies are sinking enormous effort into intangible things the statistics barely count — new processes, new skills, new organizational plumbing — and only later does the curve turn upward, often steeply.

If the historical pattern holds, today's dismal numbers are not evidence that AI is a bust. They are what the early years of a general-purpose technology have always looked like, with heavy visible spending, invisible reorganization, and output statistics that stubbornly refuse to cooperate. The 95-percent pilot failure rate reads differently through this lens — less like proof the technology is empty, more like a census of companies wiring the new motor into the old steam layout. A chatbot bolted onto an unchanged workflow is the line shaft with better marketing.

That said, the rerun comes with two honest complications, and a serious piece has to sit with both. The first is that the lag is not guaranteed to resolve on schedule, or at all — history offers one full precedent in electricity and a partial one in computing, which is a pattern, not a law. The second is a newer worry, raised by the Computerworld columnist Mike Elgan in the piece that prompted this one, is that AI might be different because its output is communication itself. When a tool makes it nearly free for every worker to generate more email, more reports, more code, and more meeting summaries, one person's productivity gain arrives in everyone else's inbox as work. There are early signs of exactly this. LinkedIn data shows applications per US job posting have roughly doubled since spring 2022 as candidates automate applying, and the share of federal court filings containing AI-generated text rose from 1 percent in 2023 to 18 percent in early 2026. Electrified factories produced more goods; AI-equipped offices may simply produce more messages — and the productivity statistics, whatever their flaws, are correctly reporting that messages are not output.

Which future we get probably depends on the same thing it depended on in 1910: whether organizations are willing to redesign the factory rather than rewire the old one. The companies that eventually broke the computer paradox in the 1990s did not use computers to type faster memos; they rebuilt what a retailer or a supply chain was. The equivalent work for AI — deciding which jobs, approvals, documents and meetings simply should not exist anymore — has barely begun, and it is organizational surgery, not software installation. It is also, not coincidentally, the part no vendor can sell.

Solow lived to see his quip refuted — he died in 2023, well after the computers showed up in the statistics at last. The safest prediction is that his sentence will be retired again someday, and the most useful question in the meantime is not "is AI overhyped?" but the one Paul David taught: how long does it take to stop building steam-engine factories, and has anyone actually started?