AI adoption slowed in August as spending per employee fell at major companies

Business uptake of artificial intelligence tools decelerated last month, according to spending patterns tracked across tens of thousands of firms by the payments platform Ramp. The firm's latest findings reveal a leveling off in adoption momentum—a potential concern for the tech industry's massive bets on AI infrastructure.

The August slowdown#

Ramp's AI index found that 56% of its customer base purchased AI products in August, a marginal increase of just 0.4% from July. This represents the smallest month-over-month gain in recent cycles and marks a noticeable deceleration from the rapid expansion that had characterized much of the year.

The broader picture remains clouded by data limitations. A concurrent US Census Bureau survey of AI adoption pegged business AI usage at only 22%, suggesting that Ramp's figures—drawn from a technically sophisticated customer base—may paint an overly optimistic portrait of mainstream adoption.

What's driving the slowdown#

Beyond summer seasonality, which has historically dampened adoption metrics between August and October, deeper economic forces appear to be at play. The most striking finding: spending per employee among the top 1% of AI-consuming firms dropped nearly 10%, sliding to $7,205.

This decline reflects two distinct pressures. First, leading model providers OpenAI and Anthropic have aggressively cut their pricing. Average token costs have plummeted to $0.68 per million tokens from a March 2026 peak of $1.15—a dramatic shift that has compressed revenues even as volume remains flat. Second, many customers are gravitating toward older, less costly models such as OpenAI's ChatGPT 5.6-Terra and Anthropic's Sonnet rather than expensive frontier releases, further eroding per-unit margins.

Why this matters for AI companies#

The numbers pose a challenge to the economics underpinning the industry's colossal infrastructure investments. Frontier labs and cloud hyperscalers have banked on robust, growing demand to justify enormous spending on chips and compute capacity. If adoption plateaus or shifts toward cheaper alternatives, those revenue forecasts—and the justification for continued capital deployment—could face pressure.

Researchers at frontier labs have indicated that the bulk of a model's training cost is recovered during its initial weeks on the market. A sustained slowdown in adoption of premium new releases could disrupt this model.