Lunch Money vs Real Money: What AI Spend Charts Actually Tell Us

There’s a chart doing the rounds, built off Ramp’s spend data, showing AI expenditure per employee across companies. The top 1% are burning something like $660 a month per head. The median company is sitting somewhere near lunch money, a few hundred dollars, if that. I’ve seen this chart three times this week in different threads, which either means it’s genuinely interesting or I need to touch grass more often. Possibly both.

What struck me wasn’t the top line. Everyone expected a handful of companies to go feral on GPU clusters and custom pipelines. What’s more interesting is the median, because it tells you most organisations are still poking the thing with a stick to see what it does. That tracks with what I see at work. We’ve got a coding assistant, someone on another team has a different one, and nobody’s entirely sure who owns the decision to consolidate. It’s not malicious, it’s just how these things go. Somebody in finance will eventually ask why we’re paying for four tools that do roughly the same job, and the answer will be “it seemed like a good idea eighteen months ago.”

One comment on the thread nailed it: this is basically the cloud spend story again, before finance teams got serious about tagging and approval flows. A few teams go wild first, then everyone else spends the next year working out why the invoice exists. I lived through that exact pattern with early cloud migrations. Nobody wants to be the person who says no to innovation, so the spend accumulates quietly until renewal season turns it into somebody’s actual problem.

There’s also a decent argument that per-employee numbers are a bit of a con. A logistics company with ten thousand warehouse staff and a dozen engineers is going to look like it barely uses AI at all, while a twenty-person startup where everyone’s an engineer looks like it’s spending like a merchant bank. Same technology, wildly different shape of workforce, and the chart flattens all of that into one misleading number. Someone in the thread made the Amazon comparison and it’s a fair one. Averages are seductive because they’re simple, and simple is usually wrong in interesting ways.

What I keep coming back to is the European commenter who said most companies over there don’t even have a dedicated AI budget yet, it’s still filed under “experimental.” Meanwhile someone else in the same thread mentioned personally spending ten grand a month keeping up with model releases. That gap is the whole story, really. We’re watching two different economies form around the same technology at the same time, one cautious and under-resourced, one throwing serious money at staying current because the ground shifts every few months.

I don’t know where this settles. Genuinely don’t. Part of me thinks the top 1% are doing something the rest of us will catch up to in two years, the way cloud went from experimental to load-bearing infrastructure. Part of me thinks a chunk of that spend is teams brute-forcing bad workflows with more tokens instead of fixing the actual problem, which is a very human way to solve things badly but expensively. Both can be true. The data doesn’t tell you which, it just tells you the money’s moving.

What I do know is that somewhere in Melbourne right now there’s a mid-sized company with four different AI subscriptions, no one accountable for any of them, and a renewal date coming up that’s going to make for an uncomfortable meeting. I’ve sat in that meeting before, just with a different acronym on the invoice. The technology changes. The finance meeting stays exactly the same.