Ramp’s AI Index shows compute is rising fast — but labor still sets the ceiling

When the newest AI spending data arrives, the headline is not that enterprises are finally outspending humans. It is that the gap is narrowing in a very uneven way.

Ramp’s AI Index, as reported by TechCrunch, shows that the top 1% of what the company calls “AI-pilled” firms are spending about $7,500 per employee each month on AI compute. That is a striking number on its own. But it still does not surpass the roughly $16,000 a month earned by the average software engineer, which is the comparison that matters for procurement teams, CTOs, and finance leaders trying to decide whether AI is a capacity multiplier or just another expanding cost center.

The more important feature of the data is the distribution. This is not a broad-based rise in spend across the market. It is a highly skewed pattern in which a small elite is pushing hard on usage while the median remains tiny by comparison. TechCrunch cites Ramp’s figures showing the top 10% of firms spend about $611 per employee per month, while the median firm spends only around $11.38 per employee — roughly the price of a single enterprise software seat. That spread says as much about organizational maturity and appetite for deployment as it does about absolute demand for model calls.

For operators, the immediate implication is that AI budgeting has entered a new phase: the question is no longer whether AI experiments are cheap, but whether scaled deployments can be made predictable.

Ramp’s index says AI-pilled firms increased spend by 14.1% per employee in the last month. That is fast enough to matter in planning cycles, especially because compute consumption tends to scale with usage, and usage tends to scale with enthusiasm, internal adoption, and product surface area. In practical terms, this pushes teams to think in token budgets, rate limits, workflow design, caching, model selection, and where inference belongs in the architecture. A company can tolerate a pilot that burns cash. It cannot tolerate a production system whose unit economics drift every time usage grows.

That is where the contrast with labor costs becomes strategically useful. A salary is fixed enough to budget around. Compute is elastic. It can jump as teams add agents, expose more users to copilots, or route more workflows through higher-cost models. The current data suggests that even the heaviest AI users are not yet in a world where compute dominates labor across the market. But they are in a world where compute is becoming visible enough to compete with labor in internal debates about ROI.

That changes rollout strategy. Product teams may still want to ship AI features broadly, but finance will increasingly ask which workflows actually move cycle time, conversion, support deflection, or output quality enough to justify ongoing token spend. A feature that looks inexpensive in a demo can become expensive once it is embedded in a daily workflow for hundreds of employees. That is especially true in functions where AI is not replacing headcount but amplifying throughput — a setup that can improve productivity while also making cost accounting harder.

The skew in Ramp’s data also hints at a coming split in vendor strategy. The top 1% of spenders are the customers most likely to exert leverage over major model platforms, cloud providers, and token services because they create meaningful consumption at scale. Those companies are also the ones most exposed to price changes, because even small shifts in per-token economics can compound quickly when applied across large workloads.

That tends to create two procurement behaviors at once. First, heavy users push for negotiating power: committed spend discounts, enterprise controls, usage reporting, and stronger contractual guardrails. Second, they diversify. If compute is now a core operating expense rather than an occasional API bill, buyers will want more than one model path, better observability, and the ability to move workloads when a vendor’s pricing or latency shifts. In other words, the market may reward the vendors that make cost control easier, not just the ones with the flashiest model demos.

The median numbers matter here too. When the typical firm is still spending about $11.38 per employee per month, there is a lot of room for AI adoption to rise before enterprise economics meaningfully resemble the AI-heavy edge cases. That gives vendors a runway, but it also warns against treating the top 1% as the market average. Pricing power may emerge at the margin, yet most firms are still operating far below the level where compute begins to look like a major line item next to compensation.

For planning teams, the next question is not whether AI spend is climbing — it is. The question is how concentrated that spend remains, how quickly the median moves, and whether deployment value can keep pace with usage growth. The most useful watchlist is straightforward: percentile-level spend shares, month-over-month growth, and time-to-value on the workflows consuming the most tokens.

If the upper tail keeps rising while the median barely moves, the industry will continue to split between firms that treat AI as infrastructure and firms that treat it as software with a few extra features. If the spread begins to compress, budgeting assumptions will need a reset.

For now, the evidence is clear enough to shape action but not dramatic enough to justify a victory lap. Compute spend is climbing quickly, and the heaviest users are beginning to feel it. But labor still sets the broader ceiling, and the distribution of spend suggests that the bigger story is not universal AI inflation. It is divergence: a small set of firms is building genuinely large AI bills while most of the market remains in the shallow end.