The News
Torsten Slok, chief economist at Apollo Global Management, wrote over the weekend that profit margins and revenue per worker for the S&P 500, minus the seven biggest tech companies, "appear to be flat," per TechCrunch. His conclusion is that AI "has delivered zero financial payoff to corporate America outside of tech." On Monday Anthropic filed its S-1, which says its revenue from Fortune 500 customers amounts, on average, to less than 0.1% of their total expenses.
The View
Those two sentences landed about a day apart, and I think they're describing the same thing.
Take the Anthropic number seriously for a second. A company with $10 billion of annual expenses that's an average Fortune 500 Claude customer is spending something under $10 million a year on it. That's a rounding error. You can't find a rounding error in a profit margin, in either direction. If the spend is that small, the payoff (if there is one) is going to be small enough to hide too.
So Slok can be right that there's nothing to see in the margins, and Anthropic can still be growing like a weed. Its first-half revenue went from $1.2 billion to $16.3 billion, per the filing. It reports net revenue retention of about 900%, which means the customers it had a year ago are spending roughly nine times as much. Both things are true at once because the starting point was tiny.
Is that bullish or bearish? Depends where you're sitting.
If you sell AI, it's the best kind of news. Your biggest customers have barely started. Anthropic says so itself, with the 0.1% figure right there in the pitch.
If you're running one of the S&P 493, it's a warning about the bill. The payoff is a separate question. Spend at nine times last year's level doesn't stay a rounding error for long. At some point it shows up as a real line in the budget, and the CFO is going to ask where the matching revenue went. Slok's answer is that it went nowhere yet. I'd hedge that a little, since a year or two of flat margins is apparently also what the early web looked like for most companies outside tech.
Some of the small sites I run use these models. They make the work faster. Faster isn't the same as more profitable, and the difference is the whole argument. Revenue per worker only moves if you sell more or employ fewer people, and most companies are still doing the same work a bit quicker with the same headcount.
Slok's other point is the one I'd actually argue with. He says AI "may be a great calculator, but in business, intelligence is mostly judgment." Fair for strategy. Most of the work inside a big company is tickets, contracts, reconciliations and code reviews, though. That's the work these models are being sold for.
Room for Disagreement
Slok's measure is the right one for investors. Margins and revenue per worker are what pay dividends, and "the payoff is coming" has been said about plenty of technologies that cost more than they returned. If AI spend grows ninefold and margins stay flat, that's not a lag. That's a cost.
Bottom Line
If you run a company outside tech, track two numbers side by side every quarter: what you spend on AI, and revenue per employee. The day the first one stops being a rounding error, the second one had better have moved.
— Hank