Your AI budget is written in tokens

The line item that’s growing fastest in your technology budget is priced in a unit your budgeting process has never seen. Deloitte’s research on AI spend dynamics puts it plainly: AI is the fastest-growing line in corporate technology budgets, and its economics are token-denominated, with costs that swing on usage patterns, model choice, and per-token pricing rather than on anything you provisioned. Gartner forecasts worldwide AI spending up 47% in 2026 and calls this the inflection year for enterprise AI budgets. The companies I work with are discovering the same thing at smaller scale: the AI line moved from a rounding error to a number the CFO circles, in about four quarters.

Cloud spend trained everyone on a comfortable assumption: the bill is a consequence of decisions you made. You chose the instances, the storage class, the database size. Spend was provisioned, so forecasting was mostly “what did we provision, and what will we change.” Fifteen years of FinOps practice is built on that assumption.

Token-denominated spend breaks it. The bill is now a consequence of decisions your users and your software make at runtime, and the multipliers stack in ways provisioned spend never did. Deloitte’s analysis documents the mechanics: reasoning models can consume several times the compute of a standard model on the same question, and agentic workflows fire many model calls to complete a single task. A product manager toggling a feature to a more capable model tier changes the unit price of every interaction with no deploy and no procurement step. Nothing in the instance-hour playbook covers any of that.

How big this gets, and how fast

EY’s December 2025 AI Pulse survey of 500 US firms found 3% already spending half or more of their IT budget on AI, and respondents expecting that share to hit 19% in 2026. Most companies are nowhere near that, and the headline-grabbing “AI is half our IT spend” firms are a small tail. But a line item that goes from invisible to circled-by-the-CFO inside a year is exactly the kind that produces budget surprises, and the surveys say the whole distribution is shifting right.

The waste data backs this up from the other side. Estimated cloud waste rose in 2026 for the first time in five years, and Flexera’s report points at cost complexity from AI as the driver. Spend that can’t be forecast is spend that can’t be caught drifting.

Forecasting spend that moves at runtime

The forecast has to be built from unit economics, because the total is no longer decided anywhere you can look it up.

Start with tokens per task, per feature. Every AI-backed feature has a measurable cost shape: how many model calls a typical interaction makes, how many tokens each consumes, at what model tier. That’s the atomic unit the budget is actually made of. If you can’t state it for your three most used features, the forecast is fiction regardless of how the spreadsheet looks.

Multiply by usage, and present bands instead of a number. Provisioned spend justified point estimates. Usage-driven spend justifies scenarios: expected, high-adoption, and runaway. The runaway band matters most, because it’s the one that turns into an incident review; an agent stuck retrying, or a prompt change that doubles token consumption, multiplies cost with no deploy going out.

Instrument usage, and alert on it, before spend. A spend alert fires after the money is gone, and for a fast-moving usage bill that can be weeks of budget. Token and request metering per feature is the leading indicator, and it’s the same per-request attribution you’d build for a shared cluster, pointed at a model endpoint. This is the shift-left costing discipline I’ve written about for AI workloads: the estimate goes in the design review, the meter confirms it in production.

Reprice quarterly. Per-token prices, model tiers, and routing behavior all change on timescales shorter than a budget cycle. A forecast built in January is describing a different product by June. The fix is cheap: revisit the unit economics quarterly and re-run the bands, the way you’d re-check commitment coverage.

None of this is exotic. It’s the same discipline cloud cost work already taught us, moved one level down, from instances to requests, and run faster. The teams that get surprised by the token bill aren’t the ones without FinOps tooling. They’re the ones still forecasting a usage-driven line item with a provisioned-spend playbook, and the gap between those two models is exactly where the CFO’s circled number comes from.