Tracking current LLM spend tells you where you have been; forecasting tells you where you are headed, which matters more for budget planning, pricing decisions, and deciding when a cost optimization effort needs to happen before it becomes urgent. A good forecast is built from the same underlying drivers as a pre-launch estimate, but grounded in real, observed usage data rather than assumptions.
Building the Forecast From Real Drivers
Start from your current, actual cost per active user or cost per unit of core product usage, measured from real logs rather than estimated, since this grounds the forecast in reality rather than compounding early assumptions that may no longer hold once the product is live.
Project this per-user or per-unit cost forward against your expected growth trajectory, whether that is user growth, usage-per-user growth as features mature and engagement deepens, or both simultaneously, since these two growth vectors can move independently and both need to be modeled explicitly rather than collapsed into a single blended growth rate.
Accounting for Feature Roadmap Impact
New features that add LLM calls, extend existing conversations, or introduce agentic multi-step behavior each shift the underlying cost-per-user baseline, sometimes substantially, and a forecast built purely on extrapolating current per-user cost forward will systematically undershoot once new LLM-touching features ship.
For any roadmap item known to add or meaningfully change LLM usage, estimate its incremental cost per user separately using the pre-launch estimation framework, and layer that onto the baseline forecast at its expected launch date, rather than waiting until after launch to discover the forecast was too optimistic.
Modeling Optimization Effects
A forecast should also account for planned or in-progress cost optimizations, such as a caching improvement, a tiering rollout, or a prompt trim, since these can meaningfully offset growth-driven cost increases and a forecast that ignores them will overstate future spend relative to what actually materializes once optimizations land.
Track optimization impact retrospectively once deployed, comparing forecasted cost against actual cost in the following period, both to validate the forecast model itself and to quantify the actual return on the optimization effort in a way that is useful for prioritizing future optimization work.
Scenario Planning
Build the forecast with at least a base case and an upside case reflecting faster-than-expected growth, since LLM costs scale close to linearly with usage and a growth surprise translates almost directly into a cost surprise, unlike some other infrastructure costs that benefit from economies of scale at higher volume.
Revisit the forecast on a regular cadence, such as monthly or quarterly, comparing actual spend against the prior forecast and adjusting the model's underlying assumptions based on what the comparison reveals, treating the forecast as a living model rather than a document produced once and left unrevised.
Key takeaways
- Ground the forecast in real, measured cost per active user, not an estimate carried over from before launch.
- Model user growth and usage-per-user growth as independent variables, not a single blended growth rate.
- Layer in the incremental cost of roadmap features known to add or change LLM usage before they ship.
- Account for planned cost optimizations so the forecast does not systematically overstate future spend.
- Build base and upside growth scenarios, since LLM cost scales close to linearly with usage.
Bottom line
A living cost forecast, grounded in real usage data and updated regularly against actual outcomes, turns budget planning from a reactive exercise into a proactive one. Teams that forecast well rarely get caught flat-footed by growth, because they already modeled what growth would do to their bill before it happened.