The effective price U.S. businesses pay for AI tokens has fallen 41% from $1.15 to $0.68 per million tokens since March, according to Ramp data, even as paid AI adoption among its customers reached 56% in August. Usage is shifting away from frontier models, whose share fell from about 53% in early August to 45% by September, while the top 1% of AI spenders cut per-employee spending by nearly 10% in August. Ramp chief economist Ara Khazarian said the figures represent a crack in the assumption that increasingly capable models will generate sustained pricing power and enough revenue to support the industry’s infrastructure buildout. OpenAI’s effective price fell 38% since Aug. 1 to $0.48, while Anthropic’s declined 22% to $0.90. Businesses are increasingly adopting cheaper mid-tier models such as Terra and Sonnet, while price competition from open-source and Chinese models is adding pressure. Morgan Stanley has flagged as much as $300 billion in bonds financing neocloud data centers as vulnerable if token prices fail to keep pace with infrastructure costs. Ramp’s September 2026 AI Index covers more than 70,000 U.S. businesses, though its technology-focused customer base may not represent the broader economy.