Ramp data shows slowing AI spending despite rising token usage

Article Summary

Ramp data indicates that while AI token usage is increasing, enterprise spending growth has stalled, with token spend rising only 1.5% between August and September compared to a 42% increase in July. This trend is driven by aggressive price competition and cost-cutting measures from major providers like OpenAI and Anthropic, alongside the emergence of cheaper alternatives. Lead economist Ara Kharazian suggests that the era of hypergrowth may be ending, as high-spending firms reduce outlays and broader market adoption fails to offset the revenue impact of lower pricing, despite an 8.9% rebound in token spend recorded between September 23 and October 23.

AI Analysis

## Core Assessment

The spreadsheet reality has finally caught up with the model providers: enterprise token consumption is expanding, but monthly corporate checks are hitting a wall. Ramp's corporate card data shows token expenditure crawling at 1.5% growth across August and September—a stark deceleration from a 42% leap in July—even with a modest 8.9% rebound into late October. The root cause is not falling demand, but aggressive commoditization and pricing discipline. As OpenAI and Anthropic slashed API fees and handed out enterprise lock-in discounts, Anthropic reportedly forfeited around $600 million in top-line growth. Compounding this, cheaper alternatives from Meta, Nvidia, open-source developers, and lightweight decision engines like Jev are steadily capturing simpler yes-or-no workloads.

Engineering teams have stopped treating foundational model calls as free capital. Companies are routing prompts to budget models, setting strict employee caps, and preventing top-tier flagships from handling routine pipelines. Furthermore, the market remains dangerously top-heavy: spending is overwhelmingly driven by venture-backed tech firms, where the top 1% peaked near $10,000 per employee each month. With median firms paying merely $12 to $15 and mainstream industries like healthcare or retail showing no justification for four-figure per-seat budgets, the top tier has capped out while the broader base cannot offset the price cuts. The end result is a buyer's paradise that compresses foundation model margins, directly visible in Anthropic's valuation dropping toward $1.2 to $1.3 trillion and OpenAI curbing fundraising targets by $10 billion.

## Impact and Advice

### Engineering and Infrastructure Directors
Halt unilateral commitments to premium foundation models and deploy routing layers immediately. Configure workflows to default to lightweight models or open-source inference for binary decisions and classification tasks, reserving frontier models strictly for complex generation or high-stakes reasoning. Enforce programmatic token budgets and strict per-employee quotas at the API gateway to prevent internal runaway spend.

### Finance and Procurement Heads
Renegotiate current provider contracts using competitor pricing cuts and the flood of alternative models as immediate leverage. Refuse long-term, fixed-commit spending minimums based on peak summer projections. Instead, demand tiered enterprise discounting with flexible volume bands that capture ongoing industry-wide API price reductions without penalizing your unit economics as internal routing gets more efficient.

## Contrarian View

The spending stall may represent a temporary pricing plateau rather than structural saturation. If research labs release genuinely novel capabilities—such as autonomous task agents that match the disruptive utility and enterprise monetization achieved by AI coding tools eighteen months ago—workload value could surge. In that scenario, expanded operational autonomy would drive enterprise spend back into exponential territory regardless of baseline per-token pricing.

## Bottom Line

Enterprise customers finally learned how to build a sensible model routing pipeline, which is terrific news for your runway and terrible news for the venture math funding frontier training clusters.

Explore this story further