AI Usage Climbs as Spending Growth Slows Ahead of Anthropic’s Potential IPO

Preface

Businesses are using more artificial intelligence than ever, yet recent spending data suggests that higher usage is no longer translating into equally rapid revenue growth. Ramp’s data shows AI token volumes continuing to rise while spending growth slowed in late summer, highlighting the impact of falling prices, intense competition, and more cost-conscious customers. The trend comes as Anthropic approaches a potential public offering and investors assess the durability of AI companies’ growth forecasts. This article examines what the spending data indicates, why companies may be paying less even as they use more AI, and how those forces could affect leading model providers. The central distinction is between AI adoption and the revenue generated from that adoption: growing usage can coexist with flat or slower spending when models become cheaper and customers learn to control costs.

Lazy bag

Ramp’s figures suggest that AI usage is still increasing, but spending growth has cooled as prices fall and buyers become more efficient. Token spend grew only 1.5% across August and September, compared with 42% in July, before rising 8.9% from September 23 to October 23. Lower API prices, enterprise discounts, cheaper and open models, and smarter workload management are allowing companies to do more with less. That shift benefits customers but raises questions about whether model providers can sustain the revenue growth investors expect, particularly as Anthropic and OpenAI face valuation and fundraising pressures.

Main Body

Ramp’s recent data describes an AI market in which adoption remains strong but spending is showing signs of strain. Companies are processing more tokens, a measure of activity across AI models, without increasing their spending at the same pace. This divergence matters because many expectations for leading AI companies rely on the assumption that greater usage will consistently produce rapid revenue growth. The numbers suggest that assumption may be less certain when customers can access more computing at lower prices.

Across August and September, token spend grew 1.5%, according to Ramp. In July, it had increased 42% from the previous month. The contrast indicates a marked slowdown in spending growth during late summer, even as usage continued to rise. October data showed some renewed momentum: token spend increased 8.9% from September 23 to October 23. Still, Ramp lead economist Ara Kharazian said the forces weighing on spending remain strong.

Kharazian has argued that investor concerns are not primarily rooted in a lack of demand. Instead, the market is becoming more competitive, and customers are finding ways to meet their needs for less. In comments to TechCrunch, he noted that nearly all AI spending in Ramp’s data comes from venture-backed, high-growth technology companies. That concentration creates a potential vulnerability: if spending among those businesses slows, many could decelerate at the same time, with possible knock-on effects elsewhere.

The spending patterns among firms vary substantially. Ramp’s top 1% of companies, described as the most “AI-pilled,” had been spending close to $10,000 per employee at their summer peaks. Kharazian said that group’s spending stuttered slightly in August and September, leaving it just below that level. Because this small cohort accounts for such intensive usage, even a modest pullback can matter. In his view, a slowdown among these firms could be enough to stall revenue growth despite rising token volumes.

Spending by other groups has grown, but not enough to make up the difference. Median firms spend $12-$15 per employee per month, while firms in the top 10% spend $600-$1,000. Kharazian questioned whether common AI applications outside the most intensive technology settings can support spending near $10,000 per employee per month. Software and finance businesses may have substantial demand for coding agents, but a doctor’s office or car dealership may not need to spend at the same level. This difference makes it difficult to assume that the usage patterns of the most AI-intensive companies will spread uniformly across the economy.

Price competition is a major reason spending can lag behind usage. Kharazian said price cuts by OpenAI and Anthropic have made AI cheaper to use. In September, OpenAI cut API prices in half for GPT-6 Sol and GPT-6 Luna. Anthropic responded with significant reductions to pricing for Fable 5.1 and Opus 5.5. The company said in a recent filing that the pricing decision reduced revenue growth by about $600 million. Both firms have also offered enterprise discounts intended to secure customer spending. While these offers may encourage adoption and help retain accounts, they can also reduce the revenue earned for each unit of use.

These conditions have shifted negotiating power toward buyers. Companies can compare providers, switch between models, or choose lower-cost alternatives as their requirements allow. Kharazian contrasted this with earlier technology markets in which major companies subsidized products to gain customers and later benefited when competition diminished. In AI, he argued, the large number of competing options makes it harder for any one provider to rely on that pattern.

The alternatives include offerings from Meta and Nvidia, which Kharazian said are bringing relatively inexpensive or open-source models to enterprise customers. Buyers can also choose from open-model companies in China and the United States. Another option mentioned in Ramp’s data is Jev, a lightweight tool for selecting among a few predetermined outputs. Kharazian said many tasks require only a simple “Yes/No” response. In those cases, businesses can route work to a cheaper model or tool instead of paying for a more capable, expensive system.

Companies are also becoming more skilled at managing AI usage. As employees and technical teams gain experience, they can better choose which tasks need advanced models and which can be handled by less expensive ones. Kharazian described several cost-control practices: using default settings that limit expensive models on long-running tasks, setting employee budgets, and routing work from standard or light models to even cheaper alternatives. These changes make it possible to increase the amount of AI work performed without increasing expenditure proportionally.

This pattern was echoed by Notion co-founder Akshay Kothari, who recently posted internal data that aligned with Ramp’s observations. He summarized the contrast with the phrase, “tokens go brr. Costs go ahhh.” The comment captures the broader shift: token use can rise while the cost of each task, or the total bill for a given level of output, falls. For customers, that is a welcome development. For model companies, it raises questions about how much revenue growth can be extracted from rising demand.

Kharazian, who has published a series of blog posts called “Cracks in the AI Thesis,” said assumptions about the future path of AI spending are beginning to weaken. In his words, businesses can use more AI while spending less on it. That dynamic does not mean that AI is becoming less useful; rather, it suggests that usage growth alone may not provide a reliable guide to providers’ revenue growth. The distinction is particularly relevant for investors evaluating companies whose valuations depend on sustained expansion.

Reports about Anthropic and OpenAI have added to the scrutiny. Bloomberg reported Friday that Anthropic’s potential valuation had fallen from the $2 trillion estimated in early October to $1.2 to $1.3 trillion. The company was also reportedly expected to receive a high “junk” rating from credit agencies. OpenAI lowered its proposed valuation to $1.25 trillion from $1.4 trillion and reduced its latest fundraising goals by about $10 billion. These reported changes do not establish that the spending slowdown alone caused the valuation or fundraising adjustments, but they illustrate the broader investor reassessment taking place as revenue expectations meet market realities.

There is also a possible route back to faster spending growth: new products could create use cases that do not yet exist at scale. A central argument behind expectations of continued AI hypergrowth is that new applications will make the technology valuable to more businesses and individuals. AI coding emerged only a year and a half ago as a commercially viable product and helped drive the prominence of Anthropic and OpenAI. Investors hope another product category could generate a comparable wave of demand.

Personal agents are one candidate. OpenAI, SpaceXAI, and Meta have released personal agents, while startups such as Instinct are also developing products in the category. However, Kharazian is skeptical that personal agents will provide the same economic boost as coding tools. He described personal agents as “not as economically productive as coding,” suggesting that convenience alone may not create the level of business value needed to accelerate spending significantly.

For now, Kharazian’s assessment is based on observable products and spending behavior rather than on a prediction that growth must stop. The market may yet develop applications that justify larger budgets. But current evidence points to slower spending growth, more competitive pricing, and increasingly sophisticated purchasing behavior. The key question is whether future products can generate enough new, high-value work to offset the downward pressure on prices and the efficiency gains customers are already achieving. Until that becomes clearer, higher AI usage should not automatically be treated as proof of equally fast revenue growth for model providers.

Key Insights Table