Are We at the Top of the S-Curve?

Yesterday I wrote that the best model is a portfolio — companies don’t marry a model anymore, they keep a bench of them and switch by the job. That post assumed something worth checking: that businesses are done adding AI and have started working out how much AI they actually want. Ramp’s October AI Index, published this morning, is the best evidence I’ve seen that the check is coming due.¹

Ramp runs a corporate card, so it sees real spend, not survey intentions. Its lead economist, Ara Kharazian, publishes the index monthly. This is the third straight month he has titled his letter “Cracks in the AI Thesis,” and the October numbers show why.

The stutter in per-employee spend

The number to watch is spend per employee. Adoption is easy — every company in America now pays for some AI. Intensity is the question that matters, and intensity is where the data turns.

Among the top 1% of firms by AI spend — the companies that drive about 80% of all spending on OpenAI and Anthropic — per-employee spend hit $9,568 per employee per month in September. That is down 3.1% from the July high of $9,873. Overall per-employee spend rose 11.7% month over month, but the biggest spenders pulled back from their peak and stayed there. Two straight months of virtually no growth in token spend, while token volumes kept rising.

Kharazian writes that the slowdown “has been unusually stubborn,” and that it is not seasonal: token volumes rose over the same period while token spend did not.

I stutter. I don’t love that Ramp’s own word for this is “stutters.” But the word is doing honest work: per-employee spend is not falling off a cliff, it is catching, restarting, and catching again — and what it is catching on is worth understanding.

Companies have started to realize unlimited AI use is a mistake

For two years the implicit policy at a lot of companies was: give everyone AI, don’t ask what it costs, because the productivity story will pay for itself. Ramp’s data says that policy is ending. The heaviest users — the 1% whose bills underwrite everyone else’s free trials — are spending less per head than they were in July.

Ramp’s own explanation is gentler than mine: frontier prices keep falling, and cheaper tokens pull the blended price per million down. But cheaper per unit is supposed to mean more total spend, not less — that is what elasticity promises. When unit prices drop and the bill still stalls, somebody upstream is deciding they have enough² — the book’s chapter on elasticity covers exactly this gap between what cheap units promise and what the bill does. The contracts question in the letter points the same way: Kharazian reports that the enterprise lock-in deals OpenAI and Anthropic are signing are “not particularly restrictive,” with the customer holding the favorable position. The labs are discounting to defend volume. That is what you do when growth stops arriving on its own.

Read the report

The letter itself is short and worth your ten minutes: it is the Ramp Economics Lab’s October AI Index update, written by Ara Kharazian, Ramp’s lead economist, from Ramp’s own card spend data — one company’s data, disclosed as such, with the volatility caveats stated plainly. You can read it directly at ramp.com/data/ai-index-oct-2026.

So are we at the top of the S-curve?

Adoption itself: 55.9% of Ramp businesses pay for AI, up just 0.62 percentage points on the month. Anthropic’s share crept up 0.25 points to 43.8%; OpenAI jumped 1.12 points to 40.6%, but that gain came mostly from share-shifting, not new buyers. Kharazian’s own read: “It feels like we’re at the top of the S-curve, where the growth in the market will have to come from more intense adoption rather than more customers.”

That is exactly the squeeze. The first half of the S-curve — everyone signs up — is done. The second half was supposed to be everyone uses more. Instead the heaviest users are spending less per head than in July, and October’s new peak in total spend sits just 1.5% above August’s.

More intense adoption was the load-bearing assumption underneath a lot of infrastructure commitments. If the top 1% has found its ceiling at $9,568 per employee per month, the growth story has to come from somewhere — and “prices are falling” is a reason the bill stalls, not a reason the bill doubles. The book’s argument has always been that demand you cannot verify is the most expensive thing you can plan around. October’s index is the first month the per-employee line lets you verify it.

Sources

---

1. Ara Kharazian, “October 2026 Ramp AI Index: Cracks in the AI thesis, part 3,” Ramp Economics Lab, November 2026. https://ramp.com/data/ai-index-oct-2026 — all figures in this post (adoption shares, per-employee spend, token volume and spend, contract terms) are from this letter.↩︎

2. The Ramp AI Index methodology page describes the underlying spend data and its limits: https://ramp.com/data/ai-index↩︎