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

Ramp found that AI activity kept rising while spending nearly stalled. Token spend grew just 1.5% across August and September. This matters because revenue for model companies depends heavily on how much customers pay for processing, not simply on how many tokens they use.

The contrast with July is striking. Token spending had grown 42% from the previous month in July, but growth slowed sharply afterward. Ramp also saw the most AI-intensive 1% of firms spend slightly below their summer peak of nearly $10,000 per employee per month.

Spending did increase among median companies and the top 10%, but not enough to offset the slowdown among the biggest users. October brought some improvement, with token spend up 8.9% from September 23 to October 23. Still, Ramp’s lead economist said strong competitive and cost pressures remain.

The slowdown was large relative to July’s performance. Ramp reported that token spending grew 42% from the previous month in July. Across August and September, it grew only 1.5%. The figures point to a sharp change from rapid expansion to near stagnation.

The comparison is not just about slower customer interest. Token volumes continued to rise, meaning companies were using more AI. The gap between rising activity and almost flat spending suggests that customers were getting more output for each dollar, through lower prices and more efficient usage.

October showed a partial rebound. Token spend increased 8.9% from September 23 to October 23. However, Ramp’s Ara Kharazian said spending growth still faced strong headwinds. The slowdown could challenge expectations that AI companies will continue growing at breakneck speed.

AI tokens are units into which text is divided before a language model processes it. A token may represent a whole short word, part of a longer word, punctuation, or another text fragment. Models count tokens in both the prompt sent by a customer and the response they generate.

Token counts provide a practical measure of AI activity. If a company sends more prompts, requests longer answers, or runs more tasks, its token volume rises. AI providers commonly use those counts in usage-based pricing, charging different rates for input and output tokens or for different models.

In the article, Ramp uses token spending and token volumes to compare AI use with revenue. Companies can increase token volumes while spending less if providers cut prices or if customers route tasks to cheaper models. Thus, higher usage does not automatically produce higher revenue for model companies.

More AI use does not necessarily mean more spending when the price of each task falls. Axiom and Paradox cut prices, while customers learned to control how and when models run. Together, those changes let companies increase activity without increasing their bills at the same rate.

For example, companies can set budgets, use defaults that avoid expensive models for long tasks, and route simple work to cheaper models. Ramp’s Kharazian also described shifting standard or light workloads to even cheaper light models through routers. A simple Yes/No task may not need a powerful model at all.

This creates a favorable result for customers but a harder one for model providers. Ramp found rising token volumes alongside weak spending growth. October showed an 8.9% increase in token spend, but Kharazian said competitive and efficiency pressures still made future revenue growth less certain.

Axiom and Paradox are competing for customers by lowering prices and offering enterprise discounts. That competition has made AI a buyer’s market. Customers can negotiate better terms, switch between providers, or select cheaper models for routine work.

The article gives specific examples. Axiom cut API prices in half in September for Axiom Sol and Axiom Luna. Paradox made significant reductions for Fable 5.1 and Opus 5.5. Paradox said in a recent filing that those reductions lowered revenue growth by about $600 million. Both companies also offered enterprise discounts intended to lock in spending.

The strategy may increase adoption and protect customer relationships, but it reduces revenue earned per unit of usage. Cheaper offerings from Meta, Nvidia, and open-model companies add further pressure. As customers use routers and lightweight models, rising AI volumes may translate into weaker revenue growth for Axiom and Paradox.

The AI market is vulnerable because its spending is concentrated among a relatively small group of companies. Ramp’s lead economist said almost the entire AI spend in its data comes from venture-backed, high-growth technology firms. These businesses are also among the most enthusiastic AI users.

The top 1% illustrate the risk. They reached nearly $10,000 in AI spending per employee during their summer peaks, then slipped slightly below that level in August and September. Even though median companies and the top 10% increased spending, their growth was not enough to replace the top group’s contribution.

If the leading spenders slow simultaneously, model companies may see revenue growth weaken across the board. Kharazian warned that this could create cascading effects. The risk is greater because many AI companies rely on the same fast-growing customer base rather than on broad spending across many industries.

An IPO, or initial public offering, is the process by which a privately held company begins selling shares on a public stock market. It gives investors a way to buy ownership in the company and gives the company access to public-market capital.

Revenue expectations strongly influence valuation because investors are pricing the company’s future business, not only its current sales. If they expect revenue to grow rapidly, they may value the company more highly. If growth appears likely to slow, they may pay less for the same current revenue.

That context matters for Paradox as it prepares for an IPO. Ramp’s data shows rising AI volumes but almost flat token spending during August and September. Price cuts, discounts, cheaper rivals, and more efficient customer usage could limit revenue growth. Those trends challenge assumptions that AI companies will maintain breakneck expansion.