Ramp Data Shows AI Spending Stalling: Notes for Technology Leaders on Cheaper Tokens and Smarter Budgets

Investors spent the week asking whether the AI boom is about to stall. A selloff in AI-linked shares after the US midterms grabbed the headlines. A quieter data point, published by TechCrunch on Friday, may matter more to people who actually run technology budgets: spending on AI has flattened even though people are using AI more than ever.

What the Ramp data shows

Ramp, the corporate card and spend-management company, analysed how its customers spend on AI tokens. TechCrunch reported the following picture, based on Ramp’s figures:

• Token spend grew 42% in July compared with June.
• Across August and September combined, growth was only 1.5%.
• From 23 September to 23 October, spend rose 8.9%, a modest re-acceleration.
• The median company spends roughly $12–$15 per employee per month on AI tokens, while the top 10% spend $600–$1,000 per employee.
• The top 1% of firms dipped slightly in August and September, landing just below the roughly $10,000 per employee they reached at summer peaks.

Ramp’s lead economist, Ara Kharazian, was careful about interpretation. He told TechCrunch that the slowdown is not a demand problem, since AI volumes are up. Instead, he pointed to price cuts: “OpenAI and Anthropic price cuts have… just made it cheaper to use AI.” He also noted that almost the entire spend on AI comes from venture-backed, high-growth technology companies, which is a narrow base for a technology that is supposed to transform every industry.

Why price cuts reduce spend and still signal growth

The apparent paradox, rising use but flat spending, is easier to understand with a simple model. Spend equals volume multiplied by price. If a provider halves its price and customers double their usage, spend is unchanged but customers receive twice the value. That is a healthy market, even if it is a worrying one for the provider’s revenue line.

TechCrunch reported that OpenAI cut API prices in half for two of its GPT-6 models in September, and that Anthropic cut prices for its Fable 5.1 and Opus 5.5 models. According to the report, Anthropic said in a recent filing that the cuts slashed its revenue growth by about $600 million. Those are the sorts of numbers that matter to investors weighing the large private valuations attached to these companies. The article noted reports that Anthropic’s potential valuation has fallen from around $2 trillion in early October to $1.2–$1.3 trillion, and that OpenAI trimmed its proposed valuation and fundraising goal. (Separately, TechCrunch reported on Friday that OpenAI had reportedly raised an additional $11 billion at a $1.25 trillion valuation, in a round that had not yet closed.)

Taken together, the data suggests a transition from a land-grab phase, where companies spent freely to experiment, to an optimisation phase, where they ask what each token buys.

Smarter cost control is the real story

Kharazian highlighted another driver: customers are getting better at using AI economically. His comment, “You can use more AI, but use it for less,” captures it well. For many tasks you do not need the most powerful model. As he put it, “For a lot of AI use cases, you just need a simple Yes/No.”

The article mentioned that lightweight tools designed to choose among predetermined outputs are among the top performers on Ramp, and that Meta and Nvidia are introducing enterprise offerings with cheap or open-source models. Notion co-founder Akshay Kothari captured the mood in an October post: “tokens go brrr. costs go ahhh.” The industry is moving from awe to accounting.

This is how most technologies mature. Cloud computing went through the same arc: early over-provisioning, a bill shock, then a FinOps discipline that cut waste. AI is now entering its FinOps era.

What this means for technology leaders

Measure cost per outcome, not tokens per day

Token counts are an input. The useful metric is the cost of a resolved support ticket, a reviewed contract, a merged pull request or a qualified lead. Establish baselines now. When vendors cut prices again, you can see whether the benefit reaches your unit economics.

Route work to the right model

Not every request deserves a frontier model. Classification, extraction and routing tasks often run well on smaller or cheaper models. Build a routing layer that sends simple queries to inexpensive models and reserves expensive ones for hard reasoning. Many organisations find they can cut spend substantially with little loss of quality.

Cap and monitor

The gap between the median company and the top 10% in the Ramp data, tens of dollars versus hundreds or thousands per employee, shows how uneven usage can be. Set budgets by team, alert on anomalies and review the biggest consumers monthly. A single runaway agent loop can burn through a month’s allocation overnight.

Avoid long single-vendor commitments

Prices are falling and new models arrive every few weeks. Multi-year prepaid commitments priced at last quarter’s rates can be a liability. Prefer flexible terms, and architect your applications so that swapping a model is a configuration change, not a rewrite.

Pressure-test vendor viability

If providers are cutting prices while raising enormous sums at falling valuations, procurement and risk teams should ask what happens if a vendor’s financing tightens. Reports that credit agencies may assign a high-yield rating to one major lab are a reminder to maintain a fallback provider and to keep data portable.

The investor angle, and why executives shouldn’t over-read it

The Ramp finding arrived in a week when AI-linked equities were under pressure. A BBC analysis noted that fears of tighter regulation after the Democratic midterm victories appeared to spark the selloff, while concerns about the industry had been building for some time. TechCrunch also reported that voters approved most of the 41 data-centre-related ballot measures put to them on election day, reflecting local scepticism about the build-out. Meanwhile, Crusoe sold a troubled Wyoming data-centre project to Google, a reminder that the physical infrastructure behind AI is hard to finance and build.

Kharazian’s own series is titled “Cracks in the AI Thesis”, but his reading of the data is nuanced. Flat spending is not collapsing demand. It is a market learning to be efficient. For a company buying AI, falling unit prices are good news. For an investor who priced in hypergrowth in revenue, they are not. Technology leaders should keep the two perspectives separate: your job is to extract value per dollar, not to defend a valuation.

There is one more caution. Kharazian is sceptical that personal agents, which he says are less economically productive than coding, will deliver the growth the industry needs. Enterprise adopters should therefore anchor investments on use cases with measurable returns, such as software development, customer support and document processing, rather than on novelty.

Limits of the data

Ramp’s customers are not the whole economy. As Kharazian noted, the spend is concentrated in venture-backed technology firms, so the numbers describe early adopters, not the average company. Mid-market and traditional enterprises may still be ramping up. Also, a single quarter’s data can be distorted by pricing changes and by seasonal patterns. The recent 8.9% month-on-month rise shows that the trend can turn quickly.

A 30-day action plan

1. Week 1: Inventory every AI workload, owner and monthly cost. Include shadow usage on corporate cards.
2. Week 2: Define a unit metric for the top five workloads and capture a baseline.
3. Week 3: Pilot cheaper models or routing on at least two workloads; measure quality and cost.
4. Week 4: Set team budgets and alerts, review contract terms, and document a fallback vendor plan.

The bottom line

AI usage keeps climbing while spending levels off, because the technology is getting cheaper and buyers are getting smarter. That is a sign of a maturing market, not necessarily a dying one. Technology leaders should welcome falling prices, manage to outcomes and keep their options open. The winners will not be those who spend the most tokens, but those who turn each token into the most value.

Sources and further reading: TechCrunch – Ramp data shows AI spending stalling just as Anthropic nears its IPO; TechCrunch – After the midterms, is the AI industry in for a rough two years?; BBC News – Why the US midterms sparked an AI stock selloff. Figures and quotes are as reported by TechCrunch from Ramp data and company filings. This is general information, not investment advice.