One Percent of Firms Pay 80% of the AI Bill. That's a Fragile Boom.

How fragile is the AI spending boom? Ramp just handed us an uncomfortable answer.

Pretty fragile, it turns out. The corporate card company looked at business spending on OpenAI and Anthropic and found that 1% of firms account for 80% of the dollars. That's not a typo, and it's not a rounding error either.

The Raw Numbers

The top 1% of spenders paid $9,568 per employee in September. That's up 11.7% from August, which sounds great until you notice it's still 3.1% below July's peak. Growth in that cohort has now nearly stalled for two months.

Read that again. The customers carrying four-fifths of the revenue are, collectively, spending less per head than they did in midsummer.

Ramp's data isn't the whole market. It's a sample pulled from corporate cards, which skews toward companies that put AI subscriptions on a card instead of an enterprise contract. But it's a real-time sample, and it's the closest thing we've to a live read on how businesses actually pay for this stuff.

Why Concentration Matters

Enterprise software always skews toward whales. Granted. A handful of accounts usually carry more weight than the long tail. But an 80 to 1 ratio is a different animal. It means the entire revenue story for two of the most valuable private companies on earth rests on a small enough group of buyers that you could fit them in a mid-size conference hall.

And here's the part that should worry proponents: falling prices. Ramp lead economist Ara Kharazian pins the slowdown mainly on cheaper top-tier models. That's genuinely good news for anyone using AI. It's less good if you're the one selling it.

Usage can climb while dollar spending flatlines. The volume story and the revenue story are diverging, and only one of them tends to appear in a fundraising deck. I'm not entirely convinced the market has priced that in.

What the Economists Say

To be fair, Kharazian's read is the reasonable one. Price cuts are how every computing market matures. Cheaper tokens bring in smaller firms, and smaller firms eventually become bigger ones.

But the timing is awkward. The question worth asking: does cheaper inference expand the pie fast enough to offset the deflation? Price drops have routinely outrun volume growth in hardware markets for years, and the gap can persist longer than anyone expects.

Skeptics have been making this point since 2023. They've mostly been wrong so far, and the counterargument deserves airtime. Enterprise adoption is still early, and the mid-market hasn't shown up in the numbers yet.

What to Watch

Three things. First, the next two Ramp prints. If per-employee spending among the top 1% lands below $9,568 in October and again in November, that's three straight months of flat-to-down, and the pricing explanation starts to look like a demand explanation.

Second, Q4 earnings commentary from the hyperscalers. Microsoft, Amazon and Google all break out AI revenue to varying degrees, and any softening in the language around enterprise commitments will move the whole trade.

Third, contract announcements. Watch for large multi-year deals showing up in the filings and press releases of OpenAI and Anthropic customers. Concentration cuts both ways. A single churned whale becomes a visible hole in the quarter.

Time will tell, though. For now the boom is real, the revenue is real, and the customer base is thinner than the headline numbers suggest.