The companies pulling ahead on AI did not get there by approving a bigger budget, and the number that makes them look like they did is the most misread figure of the year. They spend far more because they changed how the work runs, and the spend is what that change looks like once it is metered. Ramp's June 2026 AI Index, built from card and bill-pay activity across more than 70,000 US businesses, gives the figure a sharp edge. The top 1% of companies are running about $7,500 per employee per month through AI, the top 10% about $611, and the median lands at $11.38. Read straight off the page, that is a gap of roughly 680-fold, and the obvious conclusion is that the leaders simply opened the wallet wider. The obvious conclusion is wrong.
What the number says on the surface
Take the headline at face value and the story tells itself in one line: the leaders just spend far more on AI, so to catch up you need a much bigger AI budget. It is a comfortable read, because it turns a hard problem into a procurement question. Approve a larger line item, push tool access out to more teams, and the gap should narrow on its own. Boards like this version because it has a lever they already know how to pull. The trouble is that the figure measures something other than appetite for spending: it tracks how deeply AI has been wired into the way these companies actually operate, with a very different price tag attached. The arithmetic is worth holding lightly too: Ramp frames the spread as roughly 680-fold, though dividing the rounded headline figures lands closer to 660x, and these are Ramp's numbers from its own customer base rather than an official census. The shape of the distribution is what matters, and that shape is not subtle.
Spend is the symptom, the rebuilt work is the cause
A company spending $7,500 per employee per month is not buying a more expensive chat subscription. That level of spend only shows up when AI sits inside the core of how work gets done: drafting and checking at every step, running in the background on tasks that used to wait in someone's queue, woven through the operating model rather than bolted to the side of it. The bill is high because the usage is constant, and the usage is constant because the work was rebuilt to lean on it. Put the figure next to a salary and the budget framing falls apart. A US software engineer costs roughly $16,000 a month fully loaded, more than twice the top-1% per-employee AI spend, so the companies at the frontier run their heaviest AI usage at under half the cost of a single hire. The question they answered was which work to change so the spend earns its keep, rather than how much to spend, and once that is settled the invoice takes care of itself.
The median's $11.38 tells the opposite story with the same clarity. Eleven dollars a month is the signature of a few licences handed out and the work left exactly as it was. People dip into a tool when they remember it exists, the rest of the day runs the way it always has, and the meter barely moves because almost nothing depends on it. The gap between $11 and $7,500 is the gap between a company that bought access and a company that changed how it operates, and only one of those can compound from one quarter to the next.
The distribution above is the clearest way to see it. Drawn to scale, the top-1% bar runs the full width, the top-10% bar is a short stub, and the median is a sliver you have to look for. The leaders are operating on a different plane entirely, several orders past a field that is only a step or two into adoption, and the height of the bar reads how much of the work now runs through AI rather than how generous the finance team has been.
The leaders refuse to rent a single model
There is a second habit hiding inside the top-1% spend that the headline never mentions, and it matters as much as the size of the number. The heaviest spenders deliberately run several models at once. They route the hard reasoning to frontier models from Anthropic and OpenAI, and they push the high-volume, lower-stakes work to cheaper open-source models served through inference platforms like Fireworks AI, fal, and DeepInfra. The mix does two jobs at once. It keeps the cost curve sane, because not every task needs the most expensive model in the building, and it keeps any one vendor from owning the operation, so a price change or a quiet model retirement upstream becomes an inconvenience rather than a crisis.
This is the part the budget framing cannot see. A company that thinks of AI as a line item picks a vendor, signs the contract, and treats the relationship the way it treats any other piece of software it rents, which is comfortable right up until the day the terms change. A company that thinks of AI as part of how it runs treats models as interchangeable parts, keeps two or three live, and moves work between them as prices and capabilities shift, so that day never has the power to stop the work.
Why copying the budget fails
The temptation, once the 680x figure is on the table, is to treat it as a target. If the leaders spend $7,500 and we spend $11, the fix looks like a budget that climbs toward theirs. A company can do exactly that, and the only thing it reliably produces is a larger invoice attached to the same flat results. Spend is the output of a rebuilt operating model, not the input that creates one, so pushing money through unchanged work moves the meter without moving anything that matters. The tools get bought, the licences get handed out, and the day still runs the way it always ran, only now at a higher cost basis.
The order of operations is the whole game. The companies at the top found the few places where AI genuinely compounds in their business, rebuilt the work around those places, and let the spend rise to meet the new way of operating, so the spend was the consequence. Reverse that sequence, lead with the budget and hope the work catches up, and the money simply funds more of what already failed to pay off. A bigger number on the same operating model is the most expensive way to stand still.
The uncomfortable question worth sitting with
If your AI spend tripled tomorrow, do you know which work would actually change, or would you just have a bigger invoice? Most leaders, asked it plainly, find the honest answer harder to reach than they expected, because the spend that moved anything last year came from a specific person rebuilding a specific piece of how the company runs, and the rest was access that mostly sat idle. The 680x gap is real, and it will keep widening, but it is a readout of who rebuilt the work and who only bought the tools. What the leaders won was the question of what to change, and the budget followed them there.