Ramp AI Spending Slumped at Top Firms, Challenging the Hyperscaler Growth Bet
Ramp AI spending among the heaviest business users fell 9.7% in August, despite years of rapidly expanding enterprise AI consumption. The decline took monthly spending per employee at the top 1% of AI-using firms from $7,976 to $7,205. A summer slowdown offers one explanation. Falling token costs and growing preference for cheaper models suggest a more consequential one.
The same data showed business adoption still rising, but barely. Ramp found that 56% of its customers purchased AI products during August, only 0.4 percentage points above July. That combination, slower adoption and lower spending among the biggest customers, challenges assumptions behind the infrastructure boom.
OpenAI, Anthropic, and their hyperscaler partners need expanding usage to outrun declining unit prices. Enterprise buyers increasingly want the least expensive model that completes each task reliably. That difference between selling more intelligence and delivering cheaper outcomes now sits at the center of AI economics.
Ramp AI Spending Lost Momentum in August
The August decline matters because it appeared among the customers expected to drive a disproportionate share of future AI revenue.
Ramp based its latest analysis on anonymized transaction data from more than 70,000 businesses. The financial platform tracks payments for AI subscriptions and token usage, providing a direct view of corporate purchases.
The September AI Index found that adoption continued growing during August. However, the pace slowed enough to raise questions about demand after several months of intense expansion.
Anthropic reached 43.8% adoption among the businesses measured by Ramp. That represented a monthly gain of 0.34 percentage points. OpenAI rose only 0.09 percentage points, reaching 39.8%.
Those percentages measure companies paying each provider, not exclusive market share. Businesses can purchase products from several AI vendors simultaneously. Ramp previously found that 52% of customers using either OpenAI or Anthropic used both.
The more surprising change appeared in spending intensity. Monthly AI expenditure per employee among the top 1% of adopters dropped from a revised July figure of $7,976 to $7,205.
Ramp cautions that this group contains relatively few businesses. Its estimate is therefore more volatile than figures for the median company or top 10%. Late transactions also caused Ramp to revise July’s result upward.
That methodological warning prevents one monthly movement from becoming proof of a lasting retreat. It does not make the decline irrelevant. The top 1% contributes an outsized share of the revenue available to model providers.
These heavy adopters differ sharply from the typical business. Ramp’s summer spending data placed median monthly AI spending at $11.95 per employee in July. The top 10% spent about $650, while the top 1% exceeded $7,400.
The gap shows why aggregate adoption can mislead. Adding another lightly engaged customer does not replace reduced spending by a company running agents, coding systems, and automated workflows throughout its operations.
August also has a history of weak activity. Ramp observed little adoption growth between August and October during 2025, followed by renewed expansion near year-end. Engineering teams often take vacations during the summer, reducing interactive and automated usage.
Seasonality therefore remains a credible explanation. Yet the spending decline arrived alongside structural changes that do not disappear when employees return. Tokens are becoming cheaper, companies are setting model defaults, and buyers are scrutinizing returns.
The slowdown is best treated as an early signal rather than a settled verdict. It shows that enterprise AI revenue no longer rises automatically whenever adoption increases.
Cheaper Tokens Are Changing What Growth Means
Lower token costs benefit customers, but they force model providers to generate more usage merely to preserve the same revenue.
Ramp’s effective price index fell to $0.68 per million tokens in early September. That was 41% below the 2026 peak of $1.15 recorded in March. Competition and provider price reductions helped drive the change.
A token is a small unit of text processed or generated by a language model. Providers commonly bill API customers according to the number of input and output tokens consumed.
Lower unit costs do not necessarily indicate lower AI activity. A business might process more documents, generate more code, or operate additional agents while paying less. Spend can decline even as useful output grows.
That distinction makes Ramp AI spending an imperfect proxy for adoption intensity. Transaction records show money moving to vendors. They do not directly reveal completed tasks, generated tokens, employee time saved, or business value.
Still, revenue ultimately matters to model developers and infrastructure suppliers. If prices fall faster than volume rises, greater technical usage produces weaker commercial growth.
This is the central reversal in the August data. AI services are becoming more accessible, just as competition was supposed to encourage. That success weakens the connection between adoption and revenue.
The model mix compounds the pressure. Ramp said businesses increasingly use standard models such as GPT-5.6 Terra and Anthropic’s Sonnet family. Companies told Ramp that these systems remained capable while costing less than frontier alternatives.
Frontier models, including Opus, Fable, and Sol, represented 45% of measured token volume. Their share had previously reached 53% during August before falling back, according to Ramp’s account.
The exact model labels and shares will continue changing. The customer behavior behind them is more durable. Companies are learning that email drafting, classification, extraction, and routine coding rarely require the most capable model.
Model routing formalizes that decision. A router evaluates a request and directs it toward a system offering an acceptable balance of capability, latency, and cost.
The enterprise routing shift was already visible before Ramp released its August figures. Bain consultant Jue Wang described companies reserving expensive systems for tasks that actually require them.
This approach changes the competitive unit. Providers are no longer competing only for a companywide contract or preferred chatbot position. They compete for each category of work inside the customer’s routing layer.
A frontier lab can win difficult research and coding requests while losing summaries, classifications, and routine queries to cheaper systems. That customer may remain an active account while producing less revenue.
This is welcome efficiency for enterprise buyers. It resembles cloud optimization, where teams match workloads with suitable compute rather than choosing the largest instance for every task.
The transition can also expose weak AI projects. Cheap inference cannot rescue a workflow that produces unreliable results, requires extensive review, or solves an unimportant problem.
Businesses must therefore measure outcomes, not token totals. A searchable AI knowledge base creates value when employees retrieve useful information faster. Its token consumption alone says little about that result.
Falling prices can expand demand over time. However, Ramp’s August reading suggests that this expansion had not yet fully compensated providers for lower effective prices.
The Hyperscaler Bet Needs Volume to Outrun Efficiency
The infrastructure buildout assumes that cheaper intelligence will create enough additional work to offset falling revenue from each unit of computation.
OpenAI, Anthropic, and other model companies spend heavily on training and serving increasingly capable systems. Cloud partners build data centers, secure electricity, and purchase accelerators based on expectations of durable demand.
That investment case does not require every business to spend like Ramp’s top 1%. It does require sustained growth across enough customers, workflows, and automated agents to support the installed capacity.
The August data pressures that expectation from both directions. Adoption growth decelerated, while per-employee spending fell among the most committed users. Neither change alone invalidates the thesis.
Together, they reveal a vulnerability. The market expects heavy adopters to keep discovering valuable workloads faster than providers reduce prices.
The relationship resembles a volume business with rapidly improving production efficiency. Lower costs can attract customers and stimulate consumption. They can also compress revenue when customers capture most of the savings.
AI companies want the first effect to dominate. Enterprise finance teams want the second. Competition decides how the benefit gets divided.
OpenAI has responded by encouraging customers to measure value rather than raw token costs. Its proposed concept, useful intelligence per dollar, asks whether AI completes meaningful work accurately and economically.
That framework has merit. The cheapest model is not the best choice when repeated failures, human corrections, or escalations erase its apparent savings.
Yet the framework also serves the commercial interests of frontier providers. They need buyers to consider task success, not simply compare token prices across models.
Many enterprises appear to accept the outcome-based argument while rejecting the idea that one premium model should handle everything. They route easy work cheaply and reserve frontier systems for complex requests.
Research into agent economics supports that caution. A workflow economics analysis found that model selection, workflow design, retries, and human oversight can radically alter total costs.
Agents amplify these variables because they complete multistep tasks. One user request can trigger planning, tool calls, searches, validation, and repeated model interactions.
This creates a plausible path back to growth. Even if every token becomes cheaper, autonomous workflows can consume far more tokens than a conversational prompt. More agents could therefore restore volume growth.
However, increased consumption does not guarantee proportionate value. A poorly designed agent can repeatedly inspect the same material, call unnecessary tools, or retry a failing step without improving its result.
Companies are starting to distinguish productive automation from token activity. That change threatens business strategies built around usage as a success metric.
The stakes extend beyond model labs. Hyperscalers earn money from the compute, storage, networking, and managed services behind AI applications. They also invest directly in frontier developers and reserve capacity for expected demand.
If customers obtain the same output with smaller models, better caching, and more selective routing, infrastructure demand can remain high while growing below ambitious forecasts.
The distinction between usage and spending becomes essential here. More tokens do not always require proportionately more compute. Quantization, optimized inference, and specialized hardware can lower the resources needed for each response.
Providers can respond by releasing stronger models, improving developer tools, or building applications that reach nontechnical workers. Each strategy seeks new demand beyond early engineering use.
That helps explain the growing focus on AI workspaces, research assistants, and collaborative agents. Technical teams adopted AI early. The next revenue pool must include operations, finance, marketing, legal work, and other knowledge-heavy functions.
The bet now depends on repeated, valuable use across those roles. A purchased subscription that employees rarely open supports adoption statistics but contributes little expansion revenue.
What the Ramp Numbers Cannot Prove
One volatile month cannot distinguish a seasonal pause from a lasting change in enterprise demand.
Ramp’s dataset offers unusually direct evidence because it observes real payments. It still represents Ramp customers, whose industry mix and technology preferences differ from the broader economy.
Technical sectors such as information, finance, and professional services lead adoption within the sample. That orientation can make the index an early indicator, but it can also overstate adoption across American businesses.
The business adoption survey from the U.S. Census Bureau provides a useful counterweight. TechCrunch reported that an August update placed reported AI use near 22% of businesses, far below Ramp’s 56%.
The surveys measure different populations and behaviors. Ramp detects purchases classified as AI spending. The Census Bureau asks businesses about AI use in producing goods and services.
Neither measure captures the entire market. Employees can use free tools or personally purchased subscriptions outside corporate systems. Companies can also pay for AI products that see little meaningful use.
Ramp’s top 1% estimate deserves particular caution. A small number of firms, billing dates, annual contracts, or delayed transactions can shift the monthly median for that group.
Ramp revised July spending from roughly $7,400 to almost $8,000 after receiving additional transactions. August results remain subject to similar changes.
Vacation patterns introduce another source of noise. If engineers generated fewer tokens during August, spending could rebound when projects restart. Previous year-end recoveries support that possibility.
Even a rebound would not eliminate the pricing challenge. Employees returning from vacation will encounter cheaper models and more cost-conscious default settings than they had earlier in 2026.
The available data also cannot establish whether the decline reflects weaker output. A company paying less might be accomplishing the same work more efficiently. That outcome represents successful adoption, not retreat.
Conversely, a high-spending company might generate large volumes of low-value output. Token totals cannot separate productive automation from waste.
This measurement gap matters for buyers. Finance teams need spending by workflow, team, model, and completed outcome. Product leaders need quality, cycle time, correction rates, and user retention.
Developers need a similarly complete view. A cheap model that fails repeatedly can cost more after retries and debugging. An expensive model can be economical when it completes a difficult task correctly.
For model providers, the ambiguity offers both comfort and danger. Falling spend could indicate healthier customer economics that later accelerates deployment. It could also reveal a ceiling on willingness to pay.
Open-weight models do not yet explain the broad shift. Ramp found that only 6.4% of AI-spending businesses used model-serving or inference platforms during August.
That figure is an upper estimate for open-model use because routing platforms can also provide closed models. Ramp placed adoption across all businesses at 3.6%.
Open models therefore remain supporting context, not the main opponent in this story. The immediate competition comes from cheaper standard offerings sold by the same major providers.
This dynamic can reduce average revenue without requiring customers to switch vendors. An Anthropic customer can move work from Opus to Sonnet. An OpenAI customer can select a less expensive model for routine tasks.
Vendor adoption can rise while revenue quality weakens. That is precisely why the August combination deserves attention even before anyone calls it a trend.
Three Signals Will Decide Whether This Was a Warning
The next test is whether spending recovers while customers continue routing more work toward economical models.
The first signal is September and October spending among Ramp’s top 1%. A sharp rebound toward July’s level would strengthen the seasonal explanation.
That recovery must be evaluated beside transaction revisions. A preliminary increase or decline can change when Ramp receives additional payments. Several months will provide a firmer trend than one observation.
A continued decline would strengthen the structural interpretation. It would suggest that cheaper tokens and tighter model selection are reducing revenue faster than new workloads expand it.
The second signal is frontier models’ share of token volume. Ramp placed that share at 45% after an earlier August peak of 53%.
A sustained recovery would show that businesses still find enough difficult tasks to justify premium systems. That would support provider claims that greater capability creates better economics.
Another decline would show standard models capturing more everyday work. It would pressure frontier labs to improve differentiation, lower prices, or package advanced capabilities into broader products.
Raw model share still requires context. A frontier model can generate substantial revenue from fewer tokens if its price remains higher. Providers could also change packaging, discounts, or subscription structures.
The third signal is adoption outside technical teams. AI labs are increasingly targeting collaborative work performed by analysts, managers, marketers, and other knowledge workers.
Strong repeat use in these groups would enlarge the customer base beyond early engineering adopters. It would also create workflows less tied to software release cycles and developer headcount.
Purchases alone will not settle the issue. Companies must show that nontechnical users return, complete meaningful tasks, and expand their use after initial trials.
Readers should watch for product retention, workflow volume, and reported task outcomes. Those indicators reveal more than launch announcements or account counts.
Enterprise buyers should also examine where their own savings originate. Lower spending caused by better routing is healthy. Lower spending caused by abandoned projects signals a failed deployment.
Teams can support that distinction by connecting AI activity with stored decisions, project records, and completed work. A well-maintained second brain can preserve the context needed to evaluate those results over time.
The wider market does not need spending per employee to rise forever. Businesses benefit when capabilities improve and costs fall. Efficient adoption can produce more value with less expenditure.
Model developers and hyperscalers face a different equation. Their investment plans require revenue that eventually justifies enormous infrastructure commitments.
Ramp AI spending has exposed the conflict between those outcomes. What helps customers can squeeze suppliers, even when more people use AI.
August might still prove to be a vacation-driven pause. The next several readings will show whether heavy users resumed spending or permanently learned to do more with cheaper models.
For developers and business leaders, the practical question is already clear: Is AI spending producing completed, valuable work, or merely accumulating tokens? Measure that answer before the next budget cycle.
