AI spend forecasting for the 2026 planning cycle asks finance to project a number in a category where prices are falling, usage is climbing, and the products themselves are changing every few months. A forecast built like a traditional software line, last year plus a growth rate, will be wrong in both directions at once: too high on unit price and too low on volume. Getting it right requires forecasting the drivers separately rather than the total directly.

The organizations that forecast AI spend well in 2026 will not be the ones with the most sophisticated models. They will be the ones that decompose the problem into pieces they can actually reason about, and that revisit the pieces often enough to stay honest.

Forecast the drivers, not the bill

A single top line AI number hides everything that matters. The useful forecast separates the components that move independently:

  • Volume, meaning the number of AI outcomes the business expects to produce, driven by product roadmap and adoption rather than by cost.
  • Unit price, meaning the cost per outcome, which generally falls over time as models get cheaper and as your own routing improves.
  • New workloads, meaning the features and use cases that do not exist yet but are on the roadmap for 2026.
  • Migration and transition cost, meaning the one-time spend of adopting better models as they ship.

Forecasting these separately lets you tell a coherent story. Volume up forty percent, unit cost down twenty five percent, two new workloads coming online in the second half, plus a transition reserve. That is a forecast a board can interrogate. A single number with a growth rate assumption is not.

The two forces that pull in opposite directions

The defining feature of AI spend forecasting is that the two largest drivers move against each other. Prices fall while usage rises. Miss either one and the forecast is useless. Forecast only the price decline and you underfund a business that is scaling its AI usage fast. Forecast only the usage growth and you overfund, holding capital against a bill that never arrives because unit costs dropped.

Early cloud adopters faced the identical dynamic. Compute got cheaper every year, yet cloud bills went up, because falling unit prices unlocked far more usage than the savings removed. The finance teams that understood this budgeted for growth funded by efficiency. The ones that assumed cheaper compute meant lower bills were wrong every single year.

Building the 2026 number

A defensible 2026 AI spend forecast comes together in a sequence. Start from current cost per outcome by workload, using real allocated data rather than list prices. Project volume from the product roadmap, not from last year's spend. Apply a conservative unit cost decline based on the recent trend rather than on hope. Add the new workloads with their own volume and cost assumptions. Then add a transition reserve for the model releases that will certainly come even though you cannot name them yet.

Keeping the forecast alive

A 2026 forecast set in the fourth quarter of 2025 and never touched again will be obsolete by spring. The discipline that makes forecasting work is the re-forecast. Review actuals against the driver assumptions every quarter, adjust the ones that missed, and carry the corrections forward. The point of AI spend forecasting is not to be right in October. It is to be less wrong every quarter, so that AI spend stays a managed investment rather than a surprise that shows up in the consolidated statement.