Enterprises are using AI more than ever—but spending on AI hasn’t meaningfully increased. This is an enterprise spend management platform. Ramp That’s the conclusion from the latest report, and the most robust data evidence yet on the AI bubble this year.
Here are the specific numbers:Token spending rose by just 1.5% combined in August and SeptemberIn July, this figure rose 42% month-on-month. Meanwhile, usage volume continued climbing. Spending rose while costs stayed flat—the gap between them was absorbed by price wars and smarter user behavior.
Three data points reveal exactly where the money is going
Ramp segments enterprises by “average monthly AI spend per employee”—a more informative metric than total spend:
- Top 1%At its summer peak, this group spent nearly $10,000 per employee per month, dropping below that threshold in August and September. According to Ara Kharazian, Ramp’s Chief Economist, even this small cohort slowing downis enough to stall overall revenue growth—even as usage volume continues rising.
- Top 10%$600–$1,000 per employee per month, still growing.
- Median enterprise$12–$15 per employee per month. This figure is critical—most companies remain two or three orders of magnitude away from being “heavy AI users.”
A rebound occurred in October: token spending rose 8.9% month-on-month from September 23 to October 23. Yet Kharazian’s stance is clear—headwinds remain strong, and a single rebound does not signal a trend reversal.
It’s not that no one is using AI—it’s that prices have collapsed
The most commonly misinterpreted aspect of this slowdownis not demand-relatedUsage volume is rising, and the number of companies adopting AI is rising—only the price per token is falling.
In September, OpenAI halved API pricing for GPT-6 Sol and GPT-6 Luna. Anthropic followed suit with steep cuts to Fable 5.1 and Opus 5.5—its recent filing acknowledged this move reduced revenue growth by approximately $600 millionBoth firms are also locking in enterprise customers via discounts, effectively trading revenue targets for retention.
Supply-side pressure is even fiercer: Meta and Nvidia are pushing low-cost enterprise models and open-source models, while open-source model vendors in China and the U.S. have driven prices down to just a few cents per million tokens. Several products we previously reviewed are direct outcomes of this price war—Tencent Hunyuan Hy4 at $0.42 per million tokens,Particle-1 at just $0.25Both are fractions of the price for flagship capabilities.
Another new product worth noting is a lightweight tool called Jev that selects the most suitable output from several preset options; it has been climbing steadily on Ramp’s spending leaderboard. Kharazian’s explanation is straightforward: “In many AI use cases, you only need a yes/no decision. So if you run an already inexpensive model on Jev, cheap becomes even cheaper.”
Enterprises are cutting costs themselves—and doing so with increasing sophistication
The other reason is user maturity. Kharazian puts it plainly: “When you become highly proficient with AI, you learn how to control it better.” This manifests in three concrete actions—now widely adopted practices:
- Default disabling expensive models for long-running tasksDowngrading backend batch processing and scheduled jobs—tasks that don’t require top-tier intelligence
- Setting budget caps for employeesCapping allocations per person or per project, rather than allowing unlimited usage
- Routing upstreamSending simple requests to lightweight models, reserving flagship models only for complex ones
Notion co-founder Akshay Kothari’s internal data aligns completely with Ramp’s observations; his summary is one sentence: “Tokens surge, costs plunge.”
What this means for AI companies: The Uber playbook no longer works
Big tech’s classic user-subsidy strategy—low prices first to capture market share, then price hikes after competition clears out—worked for Uber and Lyft. But Kharazian argues this script fails in AI, for one reason only:Competition is too intenseIf you cut prices, Meta counters with open-source models; if you cut further, Chinese open-source models undercut you.
This is an unwelcome signal for Anthropic, which is preparing for IPO. If investors still value AI companies on the assumption of “sustained hypergrowth,” Ramp’s data undermines that premise. More troubling is structural risk: Kharazian points out that nearly all AI spending comes from venture-backed, high-growth tech companies—meaningonce market conditions shift, this entire cohort of customers will contract simultaneouslycreating a chain reaction.
There’s another ceiling: traditional industries like doctor’s offices and car dealerships genuinely need AI, but they’re unlikely to reach $10,000 per person per month in spending. “The era of hypergrowth may be over,” says Kharazian. “Now we know the revenue we can extract from the top 1% may be capped.”
What this means for individual users and teams
The conclusion has flipped—and it favors buyers:
- Now is the window for negotiation.When renewing enterprise AI service contracts, buyers can directly leverage competitor quotes; vendors’ discount headroom is significantly larger than it was a year ago.
- Stop defaulting to flagship models.Tier tasks by difficulty and route them accordingly—costs for the same work can drop by over 50%. See reference: GLM 5.3 FlashX Rely on these high-speed, low-cost models as your primary stack; reserve flagships only for requests that truly demand advanced reasoning.
- Open-source solutions are regaining competitiveness.Qwen3.8 2.4T,Wenxin 5.5 Flash Base The cost structure of such open-source weights remains entirely unaffected by API price hikes.
- Treat “AI spend per person” as a management metric.Ramp’s data essentially gives you a ruler: Is your team spending at the median ($12–$15 per person per month) or in the top 10% ($600–$1,000)? Overspending doesn’t necessarily signal efficiency—it may simply reflect a lack of routing.
Frequently Asked Questions (FAQ)
Has enterprise AI spending truly plateaued?
Slowing spend growth does not mean usage has stalled. Ramp’s data shows token spend rose only 1.5% between August and September, then rebounded 8.9% in October—while call volume continued rising steadily. The underlying driver is a sharp decline in unit pricing, decoupling the “usage growth” and “spend growth” curves.
Why are AI prices falling so rapidly?
A three-way squeeze: OpenAI and Anthropic are slashing prices (API price adjustments at roughly 50% levels since September), open-source models have driven marginal costs down to mere cents, and enterprises deploying routing and budget controls are actively filtering out expensive requests. With oversupply and increasingly sophisticated buyers, prices can only go down.
Does this affect valuations of AI companies?
Yes. Anthropic itself disclosed that its price cuts reduced revenue growth by approximately $600 million, confirming that the price war hits revenue directly. If IPO pricing still assumes hypergrowth, sustained flattening of spend growth poses significant valuation correction risk.
How should our team reduce AI costs?
Three steps: tier tasks by difficulty, route simple requests to cheaper models, and set per-user budget caps. Most teams cut over half their costs with just the first two steps—without any noticeable quality degradation—because expensive models spend most of their time handling tasks that don’t warrant their cost.
Summarize
“Token prices rise, but costs stay flat”—this signals a disrupted growth story for model providers, yet delivers the best possible news for users. AI is shifting from a token-priced luxury to an on-demand utility like electricity or water—only when prices fall can true large-scale adoption begin.
Want to pick a more cost-effective model combination? View our AI Model Library and Tool Comparison Engineor continue reading:GLM 5.3 FlashX benchmarking · Document your initial hypothesis and metric selection with a timestamp. Experiments show this does not fully prevent influence, but at least lets you see afterward what changed in your thinking · Lock in your own judgment before reviewing the Agent’s output.