Token usage keeps rising, but AI spending barely moved this fall, Ramp finds

Businesses are running far more AI workloads than they were a few months ago, but the amount they’re paying for that usage has barely moved, according to data from the expense-management platform Ramp shared with TechCrunch.

Ramp found that combined token spend — the usage-based cost of running AI models — rose just 1.5% across August and September. That marks a sharp break from July, when spend climbed 42% in a single month, even as the actual volume of tokens processed kept rising through the later months.

Ramp’s lead economist, Ara Kharazian, said the stall is concentrated among the heaviest users. The top 1% of spending firms, the most committed adopters of the technology, dipped slightly below the roughly $10,000-per-employee-per-month pace they reached over the summer. Because that small group accounts for nearly all AI spending, Kharazian said their pullback alone was “sufficient to slow down revenue entirely, despite the fact that we’re seeing rising volumes.” Median firms, spending $12 to $15 per employee monthly, and the top 10%, at $600 to $1,000, both kept growing, but not by enough to offset the top tier’s retreat.

Kharazian attributed the pattern to competition rather than weakening demand. OpenAI halved API prices for its GPT-6 Sol and GPT-6 Luna models in September, and Anthropic followed with cuts to Fable 5.1 and Opus 5.5 that the company has said, in a regulatory filing, reduced its own revenue growth by roughly $600 million. Kharazian also pointed to inexpensive enterprise offerings from Meta and Nvidia, open models out of China and the United States, and a newer tool called Jev that routes simple yes-or-no tasks to cheaper models. He added that experienced AI customers are getting thriftier on their own, setting spending budgets, defaulting away from costly models, and routing routine tasks to lighter ones.

Token spend did pick back up in October, rising 8.9% between September 23 and October 23, though Kharazian said the structural pressure on growth hasn’t gone away. Separately, Bloomberg reported on November 6 that Anthropic’s targeted valuation ahead of its planned IPO has fallen to $1.2 trillion to $1.3 trillion, down from an estimated $2 trillion earlier in October, with credit agencies reportedly set to give the company a high “junk” rating. OpenAI has cut its own proposed valuation to $1.25 trillion from $1.4 trillion and trimmed its latest fundraising target by about $10 billion.

What this means

The numbers matter because so little of the market is actually generating them: Kharazian said almost all AI spending comes from a narrow band of venture-backed, high-growth tech companies, so a pullback at the top can drag down industry-wide revenue even while everyday usage keeps climbing. That concentration is also why Kharazian warned of “cascading effects” if spending among that cohort slows further, since there’s no broad base of other industries yet spending enough to pick up the slack.

What isn’t yet clear is whether a new category of AI product can reset the growth trajectory the way coding agents once did for Anthropic and OpenAI. Investors are watching personal AI agents, already released in some form by OpenAI, SpaceXAI, and Meta, as a possible next driver, but Kharazian is skeptical, since he considers them less economically productive than coding tools.

“The projection is that there will be new products being introduced that will drive further growth,” Kharazian said, “but based on what we can observe today and the products available and what is commercially advanced, AI spend has started to slow down.”