AI spend per employee fell 10% at the top spenders

The companies spending the most on AI spent less in August. Ramp's AI Index put median spend among the top 1% of AI buyers at $7,205 per employee, down nearly 10% in a single month. Adoption barely moved: 56% of Ramp customers paid for an AI product in August, up 0.4 points. August is a vacation month and some of this is seasonal noise. But the second number in the report is not noise, and it is the one that changes your budget: the average cost of a million tokens is down 41% since March.

What actually happened

TechCrunch's read of the August index is that price competition, not retreat, is doing most of the work. Average token cost sat at $0.68 per million in August against a 2026 peak of $1.15 per million in March. Ramp economist Ara Kharazian framed the OpenAI-Anthropic fight as "making AI more accessible" — and pushing spend down at exactly the accounts investors expected to carry growth.

Two other details are worth pulling out. Customers are drifting toward older, cheaper models rather than each new frontier release. And open-weight and inference-serving platforms reached 6.4% of AI-spending businesses — still small, still climbing.

For context on the base rate, the U.S. Census Bureau survey puts AI use across all businesses at 22%. The top 1% of spenders are a different species from the median firm, which is why their spend curve gets watched so closely.

Why falling token prices matter for your business

If you priced an automation in March and shelved it because the math was thin, the math is different now. A 41% cut in unit cost moves things that sat just over the line — document processing, transcript summarization, catalog cleanup, anything where volume was the problem rather than capability.

Two practical moves. First, re-run the numbers on the jobs you rejected, using current prices and current models rather than the note you wrote six months ago. Second, follow the top spenders and stop defaulting to the newest release. Most production work — classification, extraction, routing, drafting — runs fine on a cheaper tier, and the frontier model is worth its premium on a small slice of hard tasks. Pin the model per job, measure the quality difference, and pay up only where it shows.

The trap is treating a price cut as permission to stop measuring. Cheaper tokens plus unmeasured usage is how a $200 line item becomes $4,000 without anyone deciding it should. Set a per-job cost ceiling and an alert before you scale anything up.

Sources: TechCrunch, Ramp AI Index.