Ramp data shows AI spending stalling just as Anthropic nears its IPO

Ramp, the corporate spend-management platform, published data showing that business spending on AI tokens nearly flatlined across August and September, growing only 1.5% over the two months. That is a sharp reversal from July, when token spend rose 42% from the prior month. The finding arrives as Anthropic prepares for an initial public offering, and it undercuts the assumption that leading model providers can sustain hypergrowth in revenue.
The spending figures come from Ramp's lead economist Ara Kharazian, who has written a series of blog posts titled "Cracks in the AI Thesis." According to the Ramp data, growth did continue among median companies and the top 10% of spenders, but not by enough to offset a pullback at the very top. Median firms spend roughly $12 to $15 per employee per month; the top 10% spend $600 to $1,000. The top 1% — which Kharazian describes as the most "AI-pilled" firms — had been spending just under $10,000 per employee at their summer peaks, and slipped slightly below that level in August and September. Because that small cohort accounts for so much of total outlay, even a modest dip there is enough to flatten overall revenue growth while volumes keep climbing.
Kharazian attributes the stall to supply-side competition and customer maturation rather than weak demand. OpenAI cut API prices in half in September for its GPT-6 Sol and GPT-6 Luna models, and Anthropic responded with significant reductions for Fable 5.1 and Opus 5.5. Anthropic said in a recent filing that the pricing moves cut its revenue growth by about $600 million. Both vendors have also been offering enterprise discounts meant to lock in spending but which effectively trim revenue against targets. The result, in Kharazian's telling, is a buyer's market.
He argues the usual platform playbook — subsidize usage to win customers, then raise prices once rivals thin out, as ride-hailing did — does not apply here because competition is too broad. Meta and Nvidia offer inexpensive or open models to enterprises, model makers in the U.S. and China keep entering, and a newer company called Jev is climbing Ramp's rankings. Jev is a lightweight tool for picking among a few predetermined outputs; Kharazian's point is that many workloads only need a simple yes or no, so routing an already cheap model through it makes the cost cheaper still.
Customers are also getting better at thrift. Kharazian describes experienced buyers imposing budgets on employees, setting defaults that block expensive models from long-running tasks, and using routers to shift load from standard or light models down to cheaper ones. Notion co-founder Akshay Kothari posted internal data that echoed the Ramp picture, summarizing it as token volumes rising while costs fall in line.
October showed some re-acceleration: token spend rose 8.9% between September 23 and October 23. Kharazian still characterizes the headwinds as strong. His broader claim is that assumptions about the trajectory of AI spending are breaking down — organizations can consume more AI while paying less, which benefits buyers and complicates revenue forecasts for the model companies.
Valuation signals have moved in the same direction. Bloomberg reported that Anthropic's potential valuation fell from roughly $2 trillion estimated in early October to $1.2 trillion to $1.3 trillion, and that the company is reportedly expected to receive a high "junk" rating from credit agencies. OpenAI reportedly lowered its proposed valuation to $1.25 trillion from $1.4 trillion and cut its latest fundraising target by about $10 billion.
The bull case for continued hypergrowth rests on new use cases expanding the market. AI coding emerged about a year and a half ago and became the commercially dominant product, propelling Anthropic and OpenAI toward trillion-dollar valuations; investors hope something similar comes next. Some point to personal agents, with OpenAI, SpaceXAI, Meta and the startup Instinct all shipping their own. Kharazian is skeptical that personal agents, which he considers less economically productive than coding, can deliver comparable spending growth.
For developers and product teams, the practical reading is that model pricing has become a negotiating lever rather than a fixed input cost. Teams that route workloads to the cheapest adequate model, cap spend per employee and reserve frontier models for hard tasks are already benefiting, and vendors competing on price will likely keep offering discounts to hold accounts. This suggests procurement and cost-governance tooling will matter as much as raw model quality when teams pick a provider.
A likely trade-off is durability. Deep price cuts and enterprise discounts support adoption but compress provider revenue, and Anthropic's own filing illustrates the arithmetic. If spending growth stays modest, model vendors may lean harder on differentiated premium tiers or bundling to protect margins, which could narrow the cheap options developers rely on. That is inference, not a reported plan.
Why it matters: The data suggests the revenue growth investors expect from AI model providers may be weaker than assumed heading into Anthropic's IPO, and the reported valuation revisions show markets already adjusting. For developers and product teams, cheap tokens and buyer-friendly terms look likely to persist while competition holds, but the same pressure may push vendors toward premium pricing and less generous tiers over time. Teams building on proprietary APIs should treat pricing as volatile and keep model routing flexible.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
AI usage keeps climbing, but spending has stalled. Price cuts and smarter cost controls among the biggest spenders could challenge investor assumptions about AI hypergrowth.