AI Has Paid Off Outside Tech. Apollo Is Reading the Wrong Scoreboard.

On October 24, Torsten Slok, the chief economist at Apollo, published a short note with a big headline: no signs of AI payoff outside the tech sector. His first sentence said AI “has delivered zero financial payoff to corporate America outside of tech.” Bloomberg picked it up, then TechCrunch, then about half of LinkedIn. By Sunday afternoon my feed was a wall of people quoting the word zero like it had been carved on a tablet.

Slok is a serious economist with a serious team, and his charts are real. I read the note three times and looked hard at every chart. I still think he’s wrong.

His numbers are fine. What they can see is the problem. He looked at the place where AI’s payoff should show up last and declared it missing everywhere. That’s a little like a factory owner in 1900 who bolted an electric motor where the steam engine used to sit, kept the same belts and shafts, saw no change in output, and announced that electricity was a bust. Hold onto that factory. We’ll come back to it.

What did Apollo’s Torsten Slok claim about AI’s payoff?

The note is five charts and four paragraphs. The first chart shows revenue per worker, adjusted for inflation. For the Magnificent Seven, the seven tech giants that dominate the index, it climbs from the mid-500s in thousands of dollars per employee in 2023 to nearly 800 by the third quarter of 2026. For the other 493 companies in the S&P 500, it sits near 500 the whole time.

The second chart is profit margins. The Mag 7 now keep roughly 30 cents of profit from every dollar of sales. The S&P 493 keep about 10 cents, close to where they were before the pandemic.

The third chart tracks productivity growth back to 1985 and finds no visible AI bump. The last two come from Ramp, the corporate card company, and show AI spending piled up in a small number of firms, most of them venture-backed.

From that, Slok concludes that mainstream corporate America is “pulling back” and starting to treat token spending as a cost center that eats margins. He closes with a line about judgment I’ll get to later, and a bottom line: “the AI story will be decided in the S&P 493, not the Mag 7.”

I agree with that last sentence. I disagree with nearly everything he infers from it.

Where Slok gets AI right

Let me give him his due first, because some of this I’ve been saying myself.

The money pouring into AI infrastructure is out of proportion to the revenue coming back, and I think the people pumping it are setting up a crash. I wrote a whole piece on the AI hype cycle and who’s causing it, and nothing in Slok’s note changes my mind on that. The sellers of AI are making the money right now. The buyers mostly aren’t, at least not where an accountant can point to it.

He’s also right that a lot of companies bought AI the way they buy office chairs. They handed everyone a license, changed nothing about how the work flows, and waited for magic. A license handed to every employee on top of the same old process is a subscription, and a CFO will treat it like one. If that’s what most of the S&P 493 did, of course the token bill looks like a cost center. It is one.

So the warning is fair. Anyone pricing AI stocks as if the whole economy will see a margin boom next year should read his charts and sweat a little. My argument is with the word zero, and with the idea that these five charts could detect a payoff if one existed.

Why are profit margins the wrong scoreboard for AI?

Warren Buffett learned this lesson in a textile mill, and he told the story in his 1985 letter to Berkshire Hathaway shareholders. His managers kept bringing him proposals for new machinery. Each one looked like a winner on paper. The new looms would cut costs, and the return on investment beat what he was earning in candy and newspapers.

He passed on most of them, and he was right to. Every competitor in the industry could buy the same machines. Once enough of them did, the lower cost became the new floor for prices across the whole industry. The savings went straight through the mills and out the other side to the people buying cloth. Each company’s decision made sense on its own. Added together, they canceled each other out, and margins never moved.

AI is the most widely available business tool in history. A dentist’s office in Ohio can rent the same model that a Fortune 50 company uses. If it makes everyone faster at answering customers, writing proposals and processing claims, competition will push most of that gain into lower prices and better service. Customers win. The margin chart stays flat.

A flat margin line in that world doesn’t mean nothing happened. It means the gain went to the people who don’t show up in an S&P earnings report. That’s all of us, every time we buy something.

Margins also get squeezed by everything else going on at once. Wages, tariffs, interest costs, insurance and the cost of implementing the AI itself all hit that same line. A company can save ten million dollars with AI, spend six on wage increases, and report a margin that looks exactly like last year’s. The savings were real. They just paid for something else.

The Mag 7 margins tell you something too, and it isn’t flattering to the argument. Four of those seven companies sell the chips, the cloud and the models everyone else is buying. Their margins are rising partly because the S&P 493 are paying them. That’s the AI payoff showing up as revenue for the sellers. Call it “tech’s payoff” and the buyers’ outlay “no payoff,” and you’ve counted the same dollars twice, once as a win and once as a loss.

Is AI spending concentrated in only a few companies?

Slok’s fourth chart comes from Ramp’s spending data, and it does look dramatic at first glance. The top 10 percent of firms account for 95.6 percent of all spending on OpenAI and Anthropic. For model serving and inference, it’s 99.5 percent.

Now read the rest of the same table. The top 10 percent of firms also account for 95.7 percent of cloud computing spending, 91.9 percent of paid advertising, 91.8 percent of software subscriptions that have nothing to do with AI, and 84.2 percent of CRM spending. Nobody looks at those numbers and concludes that cloud computing has delivered zero payoff to corporate America. Big firms spend most of the money on almost everything. That’s what big means.

The fifth chart says 80.4 percent of the heaviest AI spenders are venture-backed, against 13.1 percent of the median spender. Slok calls this circularity, money from AI investors flowing into AI startups and back to AI vendors. Some of it is. But startups always adopt new tools first, because they have no old process to protect. In the 1990s the heaviest spenders on internet infrastructure were also venture-backed startups, and a lot of them died. The internet did fine.

Those numbers describe the order in which adoption happens. They don’t say whether the tool works.

Why does new technology take years to show up in productivity statistics?

Back to the factory.

The first central power station in the United States opened in 1882. By 1899, according to the economic historian Paul David, electric motors still supplied less than 5 percent of the mechanical drive in American factories. It took roughly another two decades, into the early 1920s, before slightly more than half was electrified. The big productivity gains showed up then, about forty years after that first power station.

The delay wasn’t stupidity. A working steam plant was expensive to tear out, so owners waited until the old one wore down. Plenty of the early adopters did exactly what I described at the top: they took out the steam engine, put a big electric motor in its place, and kept the same shafts and belts running the whole building. Nothing improved, because nothing about the work had changed.

The payoff came when factories were redesigned around electricity. Every machine got its own small motor. Buildings went from tall and narrow, stacked around one power source, to long and single-story, laid out by the flow of the work. That redesign took engineers, architects and managers who could think without the shaft, and it took a generation to train them. David published that history in 1990 in a paper called “The Dynamo and the Computer,” and the title tells you why.

In 1987 the economist Robert Solow wrote a line that still gets quoted: “You can see the computer age everywhere but in the productivity statistics.” Computers had been in offices for twenty years by then. Productivity growth didn’t pick up until the mid-1990s, when companies finally rebuilt their processes around networks, databases and email.

ChatGPT went public in November 2022. We’re less than four years in, and Slok is reading the productivity series for a verdict. He knows better, too. In a note on June 30 of this year, he wrote that the key issue is “the length of the ROI runway outside the tech sector,” and worried about the productivity hockey-stick taking five years instead of five months. That’s a sensible question about timing. Four months later the answer became zero, and I don’t see what changed in between except the headline.

What does the San Francisco Fed say about AI and productivity?

Slok’s productivity chart uses total factor productivity, a utilization-adjusted measure the San Francisco Fed publishes. It tries to capture how efficiently the whole economy combines workers and equipment. It’s the stricter measure, and on it, nothing much is happening yet.

But the San Francisco Fed’s own economists have been looking at this, and their reading is more interesting than flat. In a May 2026 Economic Letter, Hamza Abdelrahman and Andrew Foerster compared total factor productivity with the simpler measure, output per hour worked. Labor productivity has been pulling ahead. Their model put the chance that the economy has entered a high-productivity era at about 57 percent using output per hour, and about 21 percent using total factor productivity. Their explanation: workers have better tools, and the broader efficiency gains haven’t caught up.

That’s the dynamo pattern again. Better motors first, redesigned factories later.

In September, Huiyu Li at the same bank reported that business output per hour grew at a 2.5 percent annual rate from early 2023 through the middle of 2026, against 1.5 percent from 2005 to 2022. She was careful to say AI’s share of that is unclear, and some of it comes from all the money being poured into equipment, including the data centers themselves. Fair enough. Nobody serious is claiming a boom is proven. “Not proven yet” and “zero” are very different claims, though, and only one of them is in Slok’s headline.

What does the evidence say about AI at the level of a single company?

Macro statistics are a blunt instrument for a new tool. The sharper evidence comes from inside individual companies, and it isn’t zero.

The best-known study is “Generative AI at Work” by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, first released through the National Bureau of Economic Research and later published in the Quarterly Journal of Economics. They followed roughly 5,000 customer support agents at a Fortune 500 software company as an AI assistant rolled out.

Agents with the tool resolved 13.8 percent more issues per hour. The newest and least skilled agents improved by about 35 percent. Customers dealing with AI-assisted agents expressed more positive sentiment and were less likely to demand a supervisor, and the agents with the tool quit at a lower rate.

Think about what happens to that gain in a competitive industry. A company that resolves 14 percent more calls per hour can handle growth without hiring, or answer customers faster, or cut prices. Maybe all three. None of those choices has to show up as a fatter profit margin. Most of them would show up as nothing at all on Slok’s charts, because the company didn’t hire the people it would otherwise have needed, and you can’t chart a person who was never hired.

That’s also why I don’t trust revenue per worker as a scoreboard here. A business that grows without adding staff should see that number climb. A business that holds revenue steady while its competitors cut prices should see it sit still, even if its people are getting far more done.

Where is AI paying off outside the tech sector?

Mostly in places Apollo doesn’t measure.

The S&P 493 are the largest, most established companies in America. They have the most legacy process to protect, the most lawyers reviewing every change, the most layers between a good idea and a changed workflow. They’re the factory owners with the biggest steam engines. Of course they’re last.

The payoff is landing first in small businesses, solo operators and startups, the businesses with no old factory to tear down. Those firms aren’t in the S&P 500, and most of them aren’t in Ramp’s data either. When a two-person accounting practice handles twice the clients, no economist’s chart notices. It shows up in that practice’s bank account and nowhere else.

I’m one of those businesses. AI works for me every day as a digital assistant. Picture a virtual assistant, only digital, while the writing stays mine. It turns my interview recordings into transcripts and summaries, and it gives me first-pass outlines I tear apart and rebuild. A book I write with that kind of AI help takes about a month for every 30,000 words. That gain is real money to me, and it will never appear in an Apollo chart, because The Writing King isn’t on anybody’s index.

I see the other side of it too. More and more of my clients show up with a manuscript an AI wrote for them, and it reads badly. They tried to bolt the tool onto a job it couldn’t do alone, the way that factory owner bolted a motor onto a line shaft. That’s the strongest form of Slok’s argument, and I’ll grant it: AI used lazily produces junk, and junk has no payoff. AI used by someone who knows the work produces a lot more of the work, faster.

And look at what one person can build now. Brandon Thomas, a developer, built free clones of Adobe and Microsoft Office software with Claude’s help. Whatever you think of that, a single person doing what used to take a building full of programmers is a productivity gain. It just doesn’t land on an S&P 493 income statement. If anything, it threatens a couple of them.

Can AI replace human judgment in business?

Slok’s best paragraph is his last one. He writes that AI “may be a great calculator, but in business, intelligence is mostly judgment.” He says AI doesn’t understand accountability, trust, culture, unwritten rules, ethical trade-offs or fairness, and those decide a company’s reputation and its margins.

I agree with every word of that. Twenty years of running a retailer’s computer operations taught me that the hardest problems were never technical. They were about which risk to take, who to trust at two in the morning, and which promise to the business I could keep. No model makes those calls for you, and anyone who hands one that job is going to get hurt.

But that paragraph argues against a use of AI nobody sensible is proposing. AI doesn’t need to understand accountability. The person using it does. A great calculator in the hands of someone with good judgment makes that judgment cheaper to exercise and faster to act on. That’s what I see every day in my own work, and it’s the whole idea behind using AI as an assistant. The human directs. The tool does the grunt work. The judgment stays where it belongs.

If the S&P 493 are buying AI hoping it will supply the judgment, they’ll be disappointed, and Slok’s charts will stay flat. That’s an argument about how to use the tool. It isn’t evidence the tool pays nothing.

Is the AI bubble going to pop?

It might, and it wouldn’t change my argument.

Elon Musk thinks a good chunk of it already is. On October 26 he replied on X to the venture capitalist Chamath Palihapitiya, who had called the neoclouds disorganized and poorly capitalized:

Pretty much.

As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers.

SpaceX closed $40B of debt for GPUs, but we’re investment grade and have Starlink & launch revenue.

That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many.

Neoclouds are the companies, CoreWeave and Nebius among them, that rent out GPU computing power to AI firms. Business Insider reported that CoreWeave’s stock fell about 37 percent over the past month and Nebius about 30 percent. Nvidia’s Jensen Huang keeps insisting demand is “off the charts.” When a reply on X asked whether Musk was talking the stocks down to buy them cheap, he answered that SpaceX isn’t buying any neoclouds. “Pointless.”

I enjoyed this next bit. The Financial Times reported in early October that Apollo was expected to lead SpaceX’s $40 billion financing to buy Nvidia chips. So the same firm whose chief economist says AI has paid nothing outside tech is lining up tens of billions against AI chips. Apollo knows exactly what it’s doing there. It’s lending to a borrower whose revenue doesn’t depend on AI. Musk drew the same line in his post. Building and renting the infrastructure is a dangerous business right now. Using the tool is a different bet.

Two different questions keep getting mashed together. The first is whether investors are overpaying for AI infrastructure, the chips, the data centers, the neocloud landlords borrowing against GPUs that lose value every year. I think a lot of them are, and some of that debt is going to hurt somebody. The second is whether AI creates real value for the businesses that use it. I think it plainly does.

Both can be true at once. The railroads went bankrupt in waves in the nineteenth century, and the railroads still changed everything. Plenty of investors lost their shirts on fiber-optic cable in 2001, and the fiber they laid carried the internet for the next twenty years. The people who build the infrastructure often lose money. The people who use it get rich slowly.

Slok’s note is useful as a warning about the first question. It just doesn’t answer the second.

What should a business do with its AI spending now?

Pick a number before you buy the tool. Calls resolved per hour, proposals out the door per week, days from draft to finished manuscript. If the number doesn’t move in a few months, stop paying for that tool or change how you use it.

Then do the hard part, the part the S&P 493 mostly haven’t done. Change the work. Decide what your people will stop doing now that the tool does it, and what they’ll do with the time instead. If you can’t answer that, you’ve bought a very expensive autocomplete. My piece on adopting AI at work without wrecking what already runs walks through that, and the three horizons of AI helps sort the hype from the value.

And stop reading macro charts for a verdict on your own business. Paul David’s factory owners couldn’t see electricity in the national numbers either. The ones who redrew the floor plan got rich anyway. The ones waiting for the statistics to give permission kept running their belts and shafts until a competitor with a single-story building and a motor on every machine ate their lunch.

If you’re working through what AI means for your writing or your business, the AI & Writing Hub collects everything I’ve written on it, and if you want help putting AI to work properly, my AI services page is where to start.

Pull the steam engine out. Then move the machines.

From the author

The Birth of the Augmented Human
I wrote this one about the other path: the practices that let you use AI hard without losing your own thinking, and what they build over time.

The Death of Thinking
This book follows the risk Slok points at, judgment handed to a machine: how the pattern starts, what it costs, and what’s gone before you notice.

Frequently Asked Questions