Ramp processes corporate card transactions for more than 70,000 American businesses. When the company published its June 2026 AI Index, one number demanded attention: the top 1% of AI adopters spend $7,449 per employee per month on artificial intelligence. The median company spends $11.38.
That is a 680x gap.
For FinOps practitioners who have spent years benchmarking cloud infrastructure costs, this distribution looks nothing like the cloud spending data. Cloud costs follow a roughly predictable curve across industries and company sizes. AI spending follows a power law, with a small number of organizations accounting for a disproportionate share of total expenditure. And the concentration at the top is accelerating at 14.1% month over month.
What a 680x Gap Looks Like in Practice
A company paying $11 per employee per month is covering a handful of ChatGPT Enterprise or Claude Team subscriptions. That spending shows up as a line item in SaaS procurement. Nobody in finance loses sleep over it.
A company paying $7,500 per employee per month is operating a fundamentally different technology stack. Ramp’s data shows these top-tier adopters combine multiple frontier models with cheaper open-source alternatives through aggregator platforms, layer verticalized solutions like voice agents and coding assistants on top, and consume API tokens at volumes that would have seemed implausible eighteen months ago.
The vendor data tells part of the story. Anthropic now leads OpenAI in business adoption across Ramp’s sample, reaching 41% of tracked businesses compared to OpenAI’s 39.5%. Anthropic gained 2.5 percentage points in the most recent month while OpenAI held essentially flat. But the real signal is not the leaderboard. It is that high-spending companies are running multi-model architectures, routing requests across providers based on cost and performance requirements. This is the same workload placement strategy FinOps teams apply to cloud infrastructure, just without a decade of tooling behind it.
These spending tiers are not two points on the same spectrum. They represent different categories of AI adoption, and the financial management infrastructure needed for each is completely different.
Three Quarters of Companies Cannot Track What They Spend
The spending gap would be less concerning if companies could at least identify where they fall. Most cannot.
A KPMG survey published in June 2026 found that only 26% of companies can fully track their AI costs. Fifty percent have partial visibility. Twenty-two percent have essentially no transparency, discovering their AI consumption only when invoices arrive.
Steve Chase, KPMG’s global head of AI, noted that tokens are being consumed at a speed far exceeding corporate expectations. Several KPMG clients burned through annual AI budgets within months, with one client experiencing a sixfold increase in token usage that finance did not detect until well after the fact.
The pattern repeats across industries. A Deloitte engagement with a healthcare client showed token usage growing 8 to 10% monthly over six months, adding $6 million in annualized unplanned costs before anyone in finance could trace the source. Affirm’s finance chief integrated token spending into annual budgeting only after agent-written code pushed March quarter consumption past all forecasts. Corning took the opposite approach, restricting employee AI tool access across the company while concentrating budget on a small number of large-scale projects.
Three companies, three governance responses, all reacting to the same underlying problem: AI costs that outpaced every existing financial model.
The Salary Crossover That Keeps Getting Closer
Executives from Nvidia and Mercor (an AI recruitment startup) have separately stated that compute costs now approach or exceed employee compensation at some organizations. Mercor’s CEO has said his company spends more on internal agent tokens than on payroll.
Ramp’s data adds nuance to that claim. The $7,449 monthly figure for the top 1% is less than half the roughly $16,000 monthly salary of an average U.S. software engineer. Even at the most aggressive tier of adoption, per-employee AI costs remain below human capital costs for most organizations.
But the growth trajectory changes the math. At 14.1% month-over-month growth, a company spending $7,500 per employee today would cross $28,000 per employee within twelve months. The point where AI spend exceeds engineer salary is not a theoretical exercise for top-tier adopters. It is a planning horizon that finance teams should be modeling now.
The median company has time. The top 1% does not.
Per-Employee AI Cost as a FinOps Benchmark
Cloud FinOps has standardized around unit economics: cost per transaction, cost per customer, cost per revenue dollar. AI spending needs an equivalent, and per-employee cost is the first metric backed by a data set large enough (70,000 businesses) to serve as a real benchmark.
It is an imperfect proxy. A company spending $2,000 per employee on AI that generates measurable productivity gains is in a different position than one spending $2,000 with nothing to show for it. But until token-level cost attribution matures, per-employee spend gives CFOs and FinOps leads three practical reference points.
Benchmarking. Knowing the median is $11 and the 90th percentile is $611 provides a starting frame. If your organization is at $3,000 per employee with no articulated AI strategy and no governance structure, the gap between spending and oversight is the risk.
Forecasting. The 14.1% monthly growth rate at the top tier provides a forward curve that finance teams can use. Organizations modeling 2027 AI budgets should stress-test at 10 to 15% monthly growth, not the linear projections most are using today.
Governance triggers. New tooling is starting to close the gap between metric and operational control. On June 5, Cloudflare launched dollar-based spend caps in its AI Gateway, letting teams set budget limits by model, provider, user, or team. When a budget is hit, requests are blocked or routed to a cheaper fallback model automatically. That is the kind of infrastructure control that turns a per-employee benchmark from a reporting metric into an enforceable constraint.
Where the 680x Gap Goes From Here
The spending gap will compress. It always does in technology adoption cycles. Companies spending $11 per employee today will move to $100 or more within a year as AI assistants become embedded in standard workflows and SaaS vendors bundle AI features into existing subscriptions. Companies at $7,500 will either find efficiency gains through multi-model routing and semantic caching, or discover they were overbuying capacity they did not need.
The Tokenomics Foundation, launched by the Linux Foundation on June 3, exists to standardize AI infrastructure economics so that benchmarking at scale becomes feasible. The FOCUS 1.4 specification, ratified the following day, adds invoice reconciliation and commitment detail columns that bring AI billing data closer to the same standards cloud costs have followed for years.
The infrastructure for tracking per-employee AI cost, setting dollar-denominated budgets, and benchmarking against industry peers is being assembled right now. The Ramp data, KPMG’s visibility survey, and Cloudflare’s spend caps all point in the same direction. The question for FinOps teams is whether they will adopt these tools proactively or discover their per-employee number the same way 22% of companies currently discover their AI costs: when the invoice arrives and the budget is already gone.