Ramp data shows AI bills crack at the top

No prediction market venue, odds or volume were cited on ARK Invest Brainstorm 156, which instead parsed business card spend, and no contract on AI spend concentration was priced on air. The discussion moved on token price deflation, not odds movement. That leaves no market-implied probability to quote. Not financial advice. Markets move fast.

Is AI spend still exploding or starting to crack? It is both, and the split is stark, Ara Kurazzian told ARK Invest.

Kurazzian, lead economist at Ramp, said Ramp tracks more than 80,000 businesses and over $275 billion in annual business spend and that AI has grown faster than any other category Ramp tracks over the past two years. He said Anthropic is now the most popular model by business spend, ahead of OpenAI, with rising enterprise use of xAI Grok and Meta models, and he flagged what he calls cracks in the AI thesis. The first is concentration, meaning 1% of businesses drive 80% of AI spend, and the second is that this top 1%, largely VC-backed tech firms, has started to slow spend in September and early October even as token volume, the metered unit for model usage, rises.

Usage is up, bills are down.

Kurazzian attributed the slowdown to cheaper AI, not less AI. He said competition between OpenAI and Anthropic, new discounts in large enterprise contracts, and more performant standard and light tier models have pushed prices down by almost half over the same period, while Brett on the show noted Anthropic has been aggressive on pricing ahead of a reported IPO push and that Opus 5.5 has taken share in Ramp data. Kurazzian said the median firm on Ramp that spends on AI pays about $15 per employee per month, up from $12 a month earlier, and the top 10% continues to grow, but the top 1% pays $7,000 to $10,000 per employee per month and is now optimizing spend with controls against frontier model use on automated tasks, better defaults and routers that steer a prompt to the cheapest model capable of the job. He agreed with the critique that 70% to 80% of AI vendor revenue comes from other AI companies funded by venture capital, which creates a revenue quality problem if funding tightens and cuts cascade.

Sam asked why OpenAI and Google Gemini do not lead in enterprise if consumer adoption predicts enterprise wins, and Kurazzian pointed to where value is proven today. Coding agents, software that writes, tests and edits code, are the most commercially advanced use case and pay off fastest for tech firms, he said, but he finds it hard to see how the median business extracts the same value from a coding agent as a software engineer at a tech company. He predicted the median firm will spend under $1,000 per employee per month within five years, framing AI as table stakes priced closer to a computer per person, not per month, and noting that real ROI does not require massive spend. Ramp also notes that August declines in its data tend to echo the November to December seasonal lull, so a 9.7% one-month dip is a data point, not a confirmed trend, and that open-source models appeared at 6.4% of AI-spending businesses, according to FourWeekMBA coverage of the September index, a breadth signal not detailed in the discussion.

Watch price per token and router adoption, not just logos. If top 1% spend keeps falling while tokens rise, budgets will shift from frontier calls to orchestrated, cheaper inference, and procurement teams will gate the most expensive models by workflow.