Ramp data shows AI spending slowing as Anthropic nears an IPO

Companies’ AI spending has slowed despite rising usage

Spending is falling among the most active companies

According to Ramp chief economist Ara Kharazian, almost all AI spending comes from fast-growing technology companies backed by venture investors. If this group’s spending slows, it could affect many companies at once and trigger further consequences.

In August and September, spending by the heaviest AI users—the top 1% of companies—edged down. At the summer peak, they were spending nearly $10,000 per employee per month, then fell slightly below that level. Kharazian estimates that the decline is enough to slow revenue growth, despite increased AI usage.

Other companies spend far less. The median company spends $12–15 per employee per month on AI, while companies in the top 10% spend $600–1,000. Kharazian says it is still unclear whether current AI use cases will lead to spending of $10,000 per employee per month. Demand for AI coding agents exists in fields such as software development and finance, but a doctor’s office or car dealership, for example, may not need to spend as much.

The era of exceptionally rapid spending growth may therefore be drawing to a close: the biggest users may have limited room to increase their spending further.

Prices are falling and customers have more options

Kharazian does not believe the slowdown is caused by falling demand: AI usage is increasing. He points to competition among providers as one factor. Price cuts from OpenAI and Anthropic have made AI cheaper to use.

In September, OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna. Anthropic responded with substantial price cuts for Fable 5.1 and Opus 5.5. In a recent financial disclosure, the company said the move reduced revenue growth by about $600 million. Both companies also offer enterprise discounts designed to lock in customer spending; Kharazian says these discounts also affect their ability to meet revenue targets.

The market has become more favorable to buyers. Subsidizing expensive technology to attract customers is a longstanding strategy among large tech companies. Kharazian compares the model to Uber and Lyft rides: once there were fewer options for ordering a ride, the price of getting from one place to another rose significantly. He believes this approach works less well in AI because the market has so many competitors.

Meta and Nvidia also offer enterprise products with low-cost or open-weight models. Other providers of open models are based in China and the US. Among companies whose products are increasingly popular with Ramp customers, Kharazian singles out Jev, a lightweight tool that selects from several predefined answers. Many tasks need only a simple “yes” or “no.” In those cases, a customer can route a request through Jev to an already inexpensive model and cut costs further.

Customers are learning to economize on models

Spending is affected not only by falling prices but also by companies getting better at controlling AI usage. According to Kharazian, experienced users are more likely to choose default settings that prevent the most expensive models from being used for long tasks, set employee budgets, and reroute some requests from standard and lightweight models to even cheaper ones.

Internal data from Notion, published by company co-founder Akshay Kothari, matched Ramp’s observations. His brief takeaway: token usage is rising while costs are falling.

tokens keep flowing, costs keep falling

As a result, companies can use more AI while spending less on it. That benefits customers, but for model developers, increased usage does not necessarily mean a comparable increase in revenue. It also calls into question earlier forecasts for the growth of AI companies’ revenues.

What this means for Anthropic and OpenAI valuations

According to Bloomberg, which reported the figures on Friday, Anthropic’s potential valuation has fallen from the $2 trillion expected in early October to $1.2–1.3 trillion. The company may also receive a high speculative rating from credit agencies, according to reports. OpenAI lowered its proposed valuation from $1.4 trillion to $1.25 trillion and cut its latest fundraising plans by about $10 billion.

Kharazian believes Ramp’s data likely helps explain investors’ concerns. At the same time, hopes for further growth are tied to new AI use cases. About a year and a half ago, AI-assisted coding emerged as a commercially promising field, helping Anthropic and OpenAI reach trillion-dollar valuations. Investors are now looking for a new product that can give the market another boost.

Personal AI agents are one possible next area. OpenAI, SpaceXAI and Meta have released such products, as have startups including Instinct. Kharazian doubts they will deliver continuous growth for the market: in his assessment, personal agents are less productive economically than coding agents. Given the products available today and their commercial maturity, AI spending has already begun to slow.

Dataist’s take: Rising AI usage no longer guarantees comparable revenue growth for providers, as customers increasingly get more for less. That matters especially for Anthropic and OpenAI amid investor expectations and fundraising plans. Potential new demand for personal agents remains an assumption, not a proven source of growth.