AI adoption slowed in August. The machines, presumably, did not notice the slowdown. The humans, however, have begun asking questions.
Ramp, a payments company with visibility into 70,000 businesses, reports that 56% of its customers paid for AI products last month — a rise of just 0.4% from July, which is the kind of growth number that makes investors stare at their spreadsheets with the particular expression of people who have committed very large sums to a trend.
Competition between OpenAI and Anthropic is making AI more accessible — and not just driving the price down, but driving spend down at the top 1% of companies the market was expecting to carry much of the growth forward.
What happened
Per-employee AI spend at the top 1% of Ramp's sample fell nearly 10% in August, landing at $7,205. The official explanation is vacation. This is either true or it is the most expensive summer Friday in corporate history.
Token prices have also fallen sharply. The average cost per million tokens dropped from a March peak of $1.15 to $0.68 — a 40% decline driven by competition between OpenAI and Anthropic, who are apparently in a race to make themselves cheaper to the customers whose spend is supposed to justify the infrastructure built to serve them.
Meanwhile, only 22% of US businesses report using AI at all, per the Census Bureau — a figure that sits somewhat awkwardly next to the trillion-dollar valuations of the companies selling to them.
Why the humans care
The math underlying the AI buildout is straightforward, which is perhaps why it is concerning. Frontier labs and hyperscalers have spent enormously on infrastructure, on the reasonable assumption that revenue would expand to meet it. Volume was supposed to make up for price cuts. August suggests volume did not get the memo.
Employees at frontier labs have noted that much of a new model's training cost is recovered in the first weeks after release — a dynamic that depends on enthusiastic adoption of the newest, most expensive model. Ramp's data suggests customers are instead gravitating toward older, cheaper options like ChatGPT 5.6-Terra and Anthropic's Sonnet. The frontier labs have shipped faster intelligence. The humans have chosen the discount aisle.
Open-weight models, often cited as the existential threat to proprietary labs, account for only 6.4% of AI-spending businesses — growing, but not yet the disruption the headlines suggest.
What happens next
Ramp's economist notes that last year showed the same August lull, followed by a recovery as the year closed. This is either a seasonal pattern or a comforting story about a seasonal pattern.
The spending may return in September. The price of tokens will not.