No AI payoff outside tech? How to read Apollo's five charts

Plenty of companies are investing heavily in AI, but from the outside it is hard to tell whether that shows up in profits. Torsten Slok, chief economist at the investment firm Apollo Global Management, answered that question in blunt terms. His memo “No Signs of AI Payoff Outside the Tech Sector,” published October 24, opens with this line: despite massive investment and broad expectations, AI has delivered zero financial payoff to US companies outside tech. TechCrunch and Fortune both picked it up.

If you have to justify an AI budget inside your company, that headline is hard to ignore. This post walks through the five charts behind the memo and separates what they show from what they don’t.

The comparison: seven big tech firms versus the other 493#

Slok splits the S&P 500 in two. One group is seven companies: Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla (the so-called Magnificent 7). The other is the remaining 493. He calls the second group the “S&P 493” and compares it with the seven. The memo’s five charts cover the following.

Chart	What it shows	Data
1. Revenue per employee (inflation-adjusted)	Up for the 7, flat for the 493	Bloomberg, Macrobond
2. Profit margins	Up for the 7, flat for the 493	Bloomberg, Macrobond
3. Productivity growth (total factor productivity and output per hour)	No AI-driven acceleration	San Francisco Fed, US Bureau of Labor Statistics
4. Share of spending by the top 10% and bottom 90% of companies, by category	AI-related spending is concentrated in a few companies	Ramp
5. Funding sources of the heaviest AI spenders	Many of the top 1% of spenders are venture-backed	Ramp AI Index

According to a Stocktwits article, revenue per employee is about $780,000 for the 7 and about $525,000 for the 493, and profit margins are close to 30% for the 7 and about 10% for the 493. In the image of chart 1, the line for the 493 has barely moved since around 2022.

The total factor productivity in chart 3 is a series published by the San Francisco Fed, adjusted for how intensively equipment and people are used.

Charts 4 and 5 show how lopsided spending is. The top 10% of companies account for 99.5% of spending on model providers and inference, 95.6% of spending on OpenAI and Anthropic, and 95.7% of cloud spending. Among the top 1% of AI spenders, 80.4% are venture-backed, compared with just 13.1% among median-spending companies.

Slok’s reading: tokens are starting to look like a cost#

From the charts, Slok concludes that mainstream US companies are starting to hold back and to treat token spending not as a source of growth but as a cost that drags on margins. His view is that venture-backed companies are spending more on tokens, while mainstream companies don’t need the latest models and aren’t willing to pay for them. As long as agentic AI and token spending stay confined to part of the economy, he argues, the macroeconomic impact will be limited.

He is cautious about the outlook too. Even as models get cheaper, turning that cheap intelligence into revenue and margin growth “may be harder than you think” for mainstream companies, he writes, because much of the intelligence a business needs is judgment. According to Slok, AI does not understand responsibility, trust, corporate culture, unwritten rules, ethical trade-offs, or fairness. These shape reputational risk and ultimately drive revenue and margins; that is the logic the memo closes on. (The memo ends with a disclaimer that its contents reflect the speaker’s views at the time and are not investment advice.)

What these charts can’t tell you#

“Zero” in the headline is a strong word. What the charts actually show is that the aggregate numbers for the 493 did not rise during a period when AI adoption spread. Three things are worth keeping in mind.

First, this is not an analysis that isolates AI’s effect from other factors. Margins and revenue per employee also move with interest rates, raw material prices, labor costs, and industry mix. The charts put trends from the same period side by side; they do not compare companies that adopted AI with companies that didn’t. For an example of research that compares before and after using staggered adoption timing, see the Harvard paper covered in Understanding debt: more AI code, no clear gain in delivery. There, too, the amount of code went up while the amount of work that reached production showed no significant increase.

Second, these are totals across 493 companies, not the return on investment of any single company. Even if the total is flat, it may still mix companies that are seeing results with companies that aren’t. And if AI spending is small to begin with, the totals will barely move. In Ramp’s June analysis, AI spending per employee was $7,449 a month at the top 1% of companies and $611 a month at the top 10%. The median-spending company spent $11.38 a month, which Ramp describes as roughly the price of one seat of ChatGPT or Claude for business. Slok himself writes that the impact will stay limited as long as spending is confined to part of the economy, and this data can’t distinguish “AI has no effect” from “most companies still use very little of it.”

Finally, some people read the outlook the other way. In the same article, Fortune also carried the view of Jan Hatzius, chief economist at Goldman Sachs. Hatzius expects the AI boom’s lift to growth to peak between late 2026 and early 2027. Even so, his base case is that AI investment eventually pays off in productivity and economy-wide profits. Hatzius also expects one path to that payoff to be labor market disruption, with a negative macroeconomic effect in the short term.

Using this for AI investment decisions at your company#

The memo is useful material for deciding which numbers to use to judge your own company’s AI investment. Slok looks at revenue per employee and profit margins, numbers that show up on the income statement. It seems worth having the same yardstick ready when you report results internally.

Adoption and usage rates only show whether a tool is being used, not whether it shows up in profits. The usage rate discussed in Fujitsu’s 90%+ Copilot usage: lessons for AI adoption is also a measure of how well a tool has taken hold. In software teams, looking only at lines of code or pull request counts makes it easy to overestimate the effect. You need to look at the amount of work that reached production and the rework rate alongside them.

How you treat token costs also connects to Slok’s point. The nature of the cost changes between the stage where you simply hand out per-seat licenses to everyone and the stage where you hand work to agents and token costs grow. If you move to the second stage, it helps to be able to put cost per task in the same table as the labor it saved or the revenue it added, and explain them together.