Ramp data: AI spend stalled while token volume kept rising

Businesses are using more AI and paying about the same for it. New Ramp data, reported by TechCrunch on November 6, shows AI token spend grew only 1.5% across August and September, even as token volume climbed. In July, spend had jumped 42% month over month. That's bad news for model companies' revenue curves. For an operator, it's the clearest sign yet that AI spend is now a buyer's market, and the buyers who manage it pay less.

What actually happened

Ramp lead economist Ara Kharazian told TechCrunch the slowdown is not a demand problem. Volume is up. Price is down. Per TechCrunch:

• Price cuts did the work. OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna in September. Anthropic cut pricing on Fable 5.1 and Opus 5.5, and said in a filing that the cuts reduced revenue growth by about $600 million.
• The top 1% stalled. The heaviest AI buyers on Ramp slipped just under the roughly $10,000 per employee they hit at their summer peak.
• Everyone else spends far less. Median firms spend $12–$15 per employee per month. The top 10% spend $600–$1,000.
• October bounced a bit. Token spend rose 8.9% from September 23 to October 23.

Kharazian's explanation for the flat spend is the part worth stealing. Mature AI users set defaults that keep expensive models off long-running tasks, put budgets on employees, and push light work to even cheaper models through routers. His summary: companies "use more AI, but use it for less."

We covered the August version of this trend. Two more months of data say it wasn't summer noise.

Why AI spend discipline matters for your business

A doctor's office or a car dealership will not spend $10,000 per seat on AI. Kharazian said so himself. That's fine. The lesson from the top spenders isn't "spend more." It's that the teams with the most AI experience got cheaper per task, on purpose.

You can copy the three moves without a finance team:

1. Pick a model per job, not per company. Classification, extraction, and yes/no decisions run fine on light models. Save the frontier tier for the hard 10%.
2. Route. A router that sends easy calls to a cheap model and escalates only on low confidence is a few hundred lines of code, not a platform purchase.
3. Budget per seat and per workflow. Set a monthly cap and an alert. Cheap tokens plus no ceiling is how a small line item quietly triples.

And if you shelved an automation because the numbers didn't work in spring, run them again with today's prices. The price war is your subsidy. Use it while it lasts.

Sources: TechCrunch, Ramp AI Index.