Artificial intelligence adoption continues to rise, but the corporate spending curve is no longer moving at the same pace. That is the takeaway from data released by Ramp, a corporate spend management platform: between August and September, spending on tokens grew by just 1.5%, following the 42% surge recorded in July compared to the previous month.
The figure does not signal a sudden loss of interest in generative models. On the contrary, usage volumes are still increasing. However, the gap between rising model requests and a barely growing bill reflects a different market phase: companies are securing more compute capacity and more output without proportionally increasing their AI budget.
For OpenAI, Anthropic, and other model providers, this distinction is significant. Market expansion has often been evaluated on the premise that the widespread adoption of these tools would yield steadily accelerating revenue. Instead, the numbers observed by Ramp suggest that price cuts, negotiated enterprise discounts, and a sharper focus on efficiency could dampen monetization, even as users turn to AI more frequently.
The slowdown starts with top spenders
According to Ara Kharazian, lead economist at Ramp, the decisive share of spending remains concentrated within a very small slice of high-growth, venture capital-backed technology companies. Over the summer, the most intensive cohort of users—the top 1% of analyzed companies—reached spending close to $10,000 per month per employee. Between August and September, that figure moved slightly below its peak.
A pullback in this bracket is all it takes to shift the overall total. Indeed, median companies spend between $12 and $15 per month per employee; in the top 10%, that range is between $600 and $1,000. While this represents significant growth compared to the past, it is not large enough to offset the slowdown among the biggest buyers of AI capacity.
Concentration is also a factor to consider in the valuation debate. A substantial share of vendor revenues does not come from an even base of companies across every sector, but from software, financial, and digital firms that have deployed models and agents across many internal processes. If these customers slow their purchasing pace, the impact on revenues can be swift.
Moreover, not all organizations have the same incentives to reach high spending levels. Coding agents and certain financial applications can justify frequent, costly model usage. For businesses such as a medical practice or a car dealership, however, the available use cases may remain more limited. Growth in horizontal adoption does not, therefore, inherently imply spending on par with that of the narrow group of companies most exposed to software transformation.
Competition is making AI cheaper
Ramp attributes the phenomenon primarily to supply-side dynamics rather than a drop in demand. In September, OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna. Anthropic, in turn, significantly cut prices for Fable 5.1 and Opus 5.5; in a recent corporate disclosure, the company indicated that this decision reduced revenue growth by approximately $600 million.
List prices are not the only lever. Vendors are also offering tailored commercial terms to large enterprises, with discounts designed to secure contracts and ongoing usage. In the short term, however, this strategy can compress revenue per unit of compute sold. In a market with numerous alternative models, buyers have greater room to negotiate and shift workloads toward less expensive options.
Competitive pressure is not coming from OpenAI and Anthropic alone. Meta and NVIDIA are bringing enterprise offerings and low-cost or open-approach models to market. Added to this are models developed by US and Chinese companies. For simple tasks, businesses can also choose lighter tools without resorting to a frontier model for every request.
Ramp also cites Jev, a service gaining ground among the monitored companies that aims to handle predetermined outputs. In many workflows, such as content classification or a binary decision, the highest reasoning or generation capacity isn't necessarily required. Offloading these workloads to less sophisticated solutions lowers overall costs and frees up more advanced models for tasks where they provide a tangible advantage.
An October rebound, but not a definitive answer
The most recent data does not indicate a complete stall. From September 23 to October 23, token spending grew again by 8.9%. It is a recovery that signals a still-dynamic market, but it does not erase the slowdown recorded in late summer, nor does it resolve the issue of revenue sustainability. Determining whether this is a temporary pause or the onset of a more mature phase will require longer historical time series and data that extends beyond the corporate customer base tracked by Ramp.
The scope of the survey itself must be kept in mind. Ramp monitors the spending of companies using its platform, not the entire global market for generative models. While the sample offers a useful indicator of purchasing behavior in the corporate segment, it does not automatically represent all use cases, including major direct enterprise contracts, consumer usage, or self-hosted deployments of open-source models.
This phase comes as Anthropic edges closer to an IPO. For investors, the evaluation will focus not only on AI demand—which remains high—but on providers' ability to defend margins and revenue as model capabilities become more commoditized and accessible. The race to slash prices may accelerate adoption and expand the user base, but it shifts the economic debate to different ground: how much value can be captured by the companies funding increasingly expensive infrastructure, research, and compute capacity.
For enterprise clients, the outlook is favorable, at least in the short term. The presence of competing offerings makes it possible to experiment with multiple tools, combine different models, and negotiate better contracts. For providers, on the other hand, volume growth will no longer be enough as proof of business viability. They will have to prove that rising usage can translate into defensible revenue even in a market where the cost per request continues to fall.



