$250 Billion in AI Spending, So Why Isn't It in the Data?

Opening

Hello, dear reader. This is Oswarld’s Knowledge Talking.

These days, AI is everywhere you look. Turn on the news — AI. Sit through a company meeting — AI. Open an investment report — AI. So a recent headline from The Economist caught my eye: “The AI productivity boom is not here (yet).” Between 2024 and 2025, companies around the world poured more than $250 billion into AI — roughly ₩350 trillion in Korean currency terms. And yet Torsten Slok, chief economist at Apollo, put it this way: “You can see AI everywhere except in the macroeconomic data.”

That line matches word for word something someone said 40 years ago, and that’s what I want to dig into today. If we’re pouring this much money into AI, why isn’t it showing up in the economic indicators?

The Ghost of Solow: The Same Question, 40 Years Ago

In 1987, Nobel laureate economist Robert Solow wrote a now-famous line in a New York Times book review: “You can see the computer age everywhere but in the productivity statistics.” From the 1960s through the 1980s, transistors, microprocessors, and memory chips advanced at an explosive pace, and PCs started showing up in every office. But something strange happened. US productivity growth fell from an average of 2.9% a year between 1948 and 1973 to just 1.1% after 1973. Extraordinary technology had arrived — and productivity growth had been cut roughly in half.

This phenomenon became known as the “Solow Paradox” — technology is clearly advancing, yet it doesn’t show up in the economic data. And in 2025, economists are watching AI trace exactly the same pattern.

There’s an important concept here: general-purpose technology (GPT)¹. It refers to technologies like the steam engine, electricity, and the computer — technologies that ripple through the entire economy. What they have in common is a broad range of application, continuous improvement over time, and a wave of further innovation that follows in their wake. AI fits squarely into this category.

The Numbers Illusion: 90% of GDP Growth Comes from Building AI, Not Using It

Look at 2025’s US economic data and there’s a curious puzzle. GDP growth came in at a respectable 2.2%, while employment grew by just 15,000 jobs a month on average — about 0.1% annually. At first glance you might think, “We’re producing more with fewer people — that must be AI.” But the reality is something else entirely.

According to Harvard professor Jason Furman’s analysis, about 90% of GDP growth in the first half of 2025 came from data center and infrastructure investment. The economy didn’t grow because AI made things more productive — it grew because companies were building buildings and buying servers to make AI in the first place. That’s a completely different story.

Research from the San Francisco Fed backs this up: once you strip out the investment effect, the underlying productivity gain is close to zero. Martha Gimbel, a researcher at Yale’s Budget Lab, urges caution too — productivity data is inherently noisy, and one or two quarters of data shouldn’t be over-interpreted. In fact, since 1950, the gap between output and employment growth has exceeded 2 percentage points in a third of all years.

Here’s the core point: the simple equation “the economy is growing, therefore it’s thanks to AI” may not hold. Dig into the data, and AI’s pure contribution is still barely visible.

”I Tried It” and “It Changed Everything” Are Two Different Claims

At this point you might reasonably ask: “But people are clearly using AI a lot — ChatGPT, Claude, all of it. Surely that has some effect?” You’re right — usage is rising. Real-time survey data from a research team led by Alex Bick at the St. Louis Fed shows the share of US workers who have tried generative AI climbing from about 39% in August 2024 to 45% by the end of 2024.

But there’s another number that matters just as much: the share of total work hours actually spent using AI is only 1–5%. Out of 100 hours of work, at most 5 hours involve AI. And only about 10% of workers use it every single day.

BCG’s 2025 global survey (10,635 respondents across 11 countries) reveals something even more striking. Among leaders and managers, 78% use AI several times a week, while regular use among frontline employees has plateaued at 51% and isn’t climbing. BCG calls this the “Silicon Ceiling.” Leadership keeps pushing AI, but daily use on the ground has stalled.

At the level of individual tasks, the effect is undeniable. An MIT study (2023) found that writing tasks were completed about 40% faster with ChatGPT, and a Harvard Business School study of BCG consultants found productivity gains of 12–25%. But translate that to the macroeconomy as a whole? According to The Economist’s calculations, it amounts to only 0.25–0.5 percentage points a year. That’s because only a fraction of workers use AI at all, and even those who do spend only a sliver of their working hours on it. And even that estimate rests on the optimistic assumption that every minute saved gets redeployed productively.

Add to this the phenomenon of “AI slop”² — low-quality output generated by AI that humans then have to review and fix, which eats into the time saved. The hours AI frees up end up being spent correcting what AI got wrong.

The Electric Motor Lesson: Buying a Tool vs. Rebuilding the Factory

This might sound discouraging, but history tells us this isn’t a new story. When the electric dynamo was invented in the 1870s, factories simply swapped electric motors in for steam engines — while leaving the layout unchanged: one giant central power shaft, driving every machine through a web of belts.

The result? Almost no efficiency gain. Electricity’s real advantage was that power could be distributed — and nobody was taking advantage of that. The real revolution came 30–40 years later, in the 1920s, when factories gave each machine its own individual motor and redesigned their entire floor plans around the actual flow of work. Only then did productivity take off. Economic historian Paul David’s famous 1990 paper, “The Dynamo and the Computer,” is the classic analysis of this exact process.

The PC followed the same pattern. Adoption began in the 1970s and 80s, but the productivity payoff didn’t show up until the late 1990s — and even then, it wasn’t Silicon Valley itself but retailers like Walmart, using computers to overhaul logistics and inventory management, that produced the real productivity boost.

Stanford’s Erik Brynjolfsson has a framework for this pattern: the “Productivity J-Curve.” When a new technology is introduced, productivity often dips at first, because reorganizing the business is costly and people need time to adapt. Only once that adaptation takes hold does the curve turn sharply upward — the “J” shape.

A recent paper by Kristina McElheran of the University of Toronto and Brynjolfsson’s research team, analyzing AI adoption in US manufacturing, confirmed exactly this pattern: productivity dipped in the early stages of AI adoption. But over the longer run, adopters pulled ahead of companies that hadn’t adopted AI at all.

Why Organizations Move Slower Than Individuals

A large-scale survey released by the NBER in February 2026 makes this gap clearer than anything else. It surveyed roughly 6,000 executives — CEOs, CFOs, and others — across the US, UK, Germany, and Australia. 70% of companies were using AI, yet fewer than 10% reported any measurable change in productivity or employment over the past three years. The executives themselves spend just 1.5 hours a week using AI.

Raising an individual’s productivity and raising an organization’s productivity are fundamentally different problems. An individual can finish a writing task 40% faster with ChatGPT. But the decision-making structure, approval processes, cross-department collaboration, and performance-measurement systems of the organization that individual belongs to stay exactly the same. Adopting an AI tool and redesigning how work gets done around AI are two entirely different challenges.

It was the same story with electricity. Buying an electric motor didn’t transform the factory — the factory’s entire layout had to be redesigned, and workflows rebuilt from scratch. AI is no different. Setting up a ChatGPT account or installing Copilot company-wide doesn’t raise organizational productivity on its own. The real work is figuring out how to redesign business processes around AI in the first place.

Oz’s Lens

Honestly, I don’t see this as a “failure.” I think it’s exactly the process you’d expect.

Having built go-to-market strategies for years, I’ve seen this pattern play out again and again. When a new tool arrives, companies almost always follow the same sequence: buy the tool, bolt it onto existing workflows, wonder “why isn’t this working?”, and only much later start redesigning the process itself. It happened with SaaS adoption. It happened with the shift to cloud. It happened with CRM rollouts.

The problem is that most companies today are still stuck at step one. They’ve “adopted” AI tools, but very few have moved on to actually “redesigning” how work gets done around AI. That’s exactly why BCG’s survey shows such a stark performance gap between companies that redesign entire workflows and companies that merely deploy tools.

So I think the more essential question in this whole debate isn’t “does AI work or not” — it’s “is the organization ready to change around AI.” The bottleneck isn’t the technology’s capability; it’s the organization’s speed of adaptation. How fast you climb out of the bottom of the J-curve depends not on how powerful the AI is, but on management’s resolve and how deeply they’re willing to redesign the organization.

Closing

Here’s the core of today’s issue:

First, AI’s technical capability and its economic impact run on different timelines. The technology is advancing at a stunning pace, but for that to translate into economy-wide productivity, organizations and entire industries need to be restructured.

Second, history shows a clear common pattern for general-purpose technologies. Electricity took 40 years. The PC took 20. Going by Brynjolfsson’s J-curve, we’re probably somewhere near the bottom of that curve right now.

Third, the gap between “adopting” and “redesigning” is the whole ballgame. Just as the real breakthrough wasn’t buying an electric motor but rebuilding the factory, the real turning point with AI isn’t buying the tool — it’s redesigning how work gets done.

After reading this, here’s one thing worth checking for yourself: is your organization “adopting” AI right now, or is it “redesigning” itself around AI? The answer will tell you exactly where you sit on the J-curve.

References & Further Reading

• The Economist, “The AI productivity boom is not here (yet)”, Feb 22, 2026. The article that sparked today’s issue — it lays out the gap between AI investment levels and macroeconomic data.
• Yotzov, I., Barrero, J.M., Bloom, N. et al., “Firm Data on AI”, NBER Working Paper No. 34836, Feb 2026. A survey of 6,000 executives across 4 countries — the source of the key data showing almost no measurable productivity or employment impact despite high AI adoption rates.
• Bick, A., Blandin, A. & Deming, D., “The Rapid Adoption of Generative AI”, St. Louis Fed Working Paper, 2024 (revised 2025). A real-time survey showing the gap between AI adoption rates and the share of actual work hours spent using it.
• Brynjolfsson, E., Rock, D. & Syverson, C., “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies”, American Economic Journal: Macroeconomics, 13(1), 2021. The original source of the J-curve framework — explains why productivity dips early in the adoption of general-purpose technologies.
• McElheran, K., Yang, M., Brynjolfsson, E. & Kroff, Z., “The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s)”, US Census Bureau Working Paper, 2025. An empirical study documenting the early productivity dip from AI adoption on the ground in US manufacturing.
• David, P.A., “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, The American Economic Review, 80(2), 1990. The classic paper comparing the productivity paradoxes of electricity and the computer — a single essay that explains why general-purpose technologies work slowly.
• Yale Budget Lab, “An AI Productivity Boom? Don’t Count Your (Productivity Data) Chickens”, 2025. A clear-eyed breakdown of the noise and measurement problems in productivity data.
• BCG, “AI at Work 2025: Momentum Builds, But Gaps Remain”, June 2025. The report that coined the “Silicon Ceiling” concept, illustrating the gap in AI use between frontline employees and leadership.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. General-purpose technology (GPT): a technology like the steam engine, electricity, or the computer that ripples through the entire economy. Such technologies share a few traits — a broad range of application, continuous improvement, and a wave of further innovation that follows in their wake. Note: this GPT has nothing to do with the GPT in ChatGPT. ↩

2. AI slop: low-quality output produced by AI. Because humans need extra time to review and fix it, the net time savings from using AI end up shrinking. ↩