Remember when getting your cloud bill under control felt like an accomplishment? Tagging every resource, right-sizing every instance, catching the intern's forgotten test cluster before finance did. FinOps grew up around exactly that fight, and cloud teams eventually won it. Cloud spend became boring. Boring is the goal.
AI just un-boringed it.
Finout's CEO, Roi Ravhon, put it well in a recent conversation with Google Cloud's FinOps lead: cloud got roughly a decade to mature into a discipline. AI is getting about a year, because the economics underneath it refuse to sit still.
Here's the part that should make you nervous if you haven't budgeted for it yet. Token prices keep dropping, and your AI bill keeps climbing anyway. Why? Because the newer reasoning models think, quite literally, about three times as much to answer the same question. More thinking, more tokens, same task. And it gets worse: ask the identical prompt twice and you'll get two different token counts back. Nothing about this is fixed. Nothing about this is predictable. Try building a budget forecast on top of that.
Fine, you might say, cloud spend used to be unpredictable too, until we built the tooling for it. Isn't this just history repeating?
Not quite. Cloud waste, at its worst, is an oversized instance running when it shouldn't be. AI waste is routing a one-line email summary through the same frontier model you'd use for a legal contract review. Google's own FinOps lead has a line he uses with customers for this: don't reach for Thor's hammer when you don't need it. He's watched teams point every request, big or small, at their most expensive model, because nobody built the layer underneath that routes cheap questions to cheap models.
And the token invoice is only one slice of what you're actually paying for. Add GPU and TPU capacity that's still scarce, training compute, inference compute, the storage feeding all of it, and the very real organizational cost of retraining people to work differently. Stanford's own research backs up the shape of this: for every dollar of tech spend, companies are sinking up to ten dollars into the process redesign and reskilling around it. That ratio should sound familiar to anyone who's ever sold a platform migration to a CFO.
Speaking of CFOs: their patience is thinning fast. The "unlimited budget, just be innovative" era is over. We're back to plain old ROI, except now nobody agrees on how to measure it.
So what actually works? Treat the deterministic parts as deterministic. Right-sizing has hard thresholds. Anomaly detection has math behind it. Don't hand that to an LLM and hope it behaves, because a model will talk itself into being right when it wants to be right. Save the agent for the parts that genuinely benefit from judgment: enrichment, context, summarizing what changed and why. And before anything destructive happens, a server termination, a budget override, put a deterministic check or a human in the loop first. Every time.
If you want a mental model for this, stop thinking of your AI spend like a cloud bill and start thinking of it like onboarding a new SRE. You wouldn't hand a brand new hire root access and tell them to go fix production. You'd give them scoped permissions, the same signals a senior engineer would use, and a review step before anything ships. Do the same for your agents. Feed them golden signals, utilization data, the boring telemetry you already have, and have them propose the fix as a pull request, not an executed change.
We spent years teaching cloud to hold still long enough to budget for it. Tokens aren't going to do us that favor. The FinOps Foundation didn't build its discipline around a tool, and neither should you build yours around one now. Get the deterministic guardrails in first. The agent layer can wait.
Source: The New Stack, "The new FinOps problem isn't cloud bills," featuring Finout CEO Roi Ravhon and Google Cloud's Pathik Sharma (May 12, 2026).
Source article: https://thenewstack.io/finops-ai-token-economics/
