Who is paying for all these tokens — and what does that buy?
Median firms spend about the price of a chat seat. The top 1% spend thousands per employee. Cost discipline is rising even as the invoice stays.
Token dashboards flatter the idea that AI spend is a single curve. Ramp’s June 2026 intensity cut shows three economies living under one buzzword.
The median firm on Ramp spent about $11.38 per employee per month on AI — roughly an enterprise chat seat. The top 10% spent $611. The top 1% spent about $7,450 per employee per month. That is not “a licence”. That is a material operating cost line — still less than half a typical engineer’s monthly salary, Ramp notes, but nowhere near a novelty subscription.
Concentration is the cost story
On the live token-spend board for late September, Anthropic and OpenAI again absorb nearly all maker share. So the invoice is concentrated twice: a thin tip of firms spend most of the money, and two labs collect most of the metered dollars those firms send.
By September’s letter, even the tip blinked. Top-1% spend-per-employee fell about 9.7% to roughly $7,205. Ramp flags summer seasonality and sample volatility — fair — but also documents structural pressure: effective price per million tokens down about 41% to $0.68 from a March 2026 peak of $1.15, and more volume flowing through cheaper “standard” models as companies set defaults that throttle frontier use.
Who pays inside the firm?
Corporate cards and bill pay catch the sanctioned path: central IT, engineering platforms, innovation budgets. They miss the employee who buys ChatGPT Plus on a personal card and pastes customer text into it. Ramp’s methodology says so plainly. That gap matters for CFOs who think the AI line item is the whole cost of AI. It is the visible cost. The invisible cost is risk, rework, and duplicated seats.
A practical cost frame for 2026
- Seat economy — median behaviour: cheap, broad, lightly measured.
- Platform economy — APIs, coding agents, multi-vendor routing; where intensity (and hiring gains) cluster.
- Frontier tax — optional until a workflow truly needs it; defaults should make the expensive path intentional.
Positive reading: prices falling and firms imposing model defaults is not the bubble bursting. It is the market growing up. The firms that will win are the ones who can say, without theatre, what each dollar of tokens returned in cycle time, quality, or revenue — and who refuse to confuse a chat seat with a transformation programme.
Sources: How much does it cost to be AI-pilled? · Cracks in the AI thesis, part 2 · Token spend leaderboard