Forecasting cloud spend when usage won't sit still
In brief
Split committed, usage-driven and AI cloud spend, tie usage to business drivers, forecast in ranges and measure variance by line, with a worked example.
Steve Elliott · · 7 min read
On this page
- Why do provider forecasts miss when usage moves?
- What should you split out before you forecast?
- How do you tie usage to a business driver?
- A worked example: one month, forecast and actual
- How should the forecast shape commitment purchases?
- How should AI usage be forecast?
- How should you measure forecast variance?
- Sources
Forecast volatile cloud spend in layers, not as one trend line. Treat committed spend as a known schedule, forecast usage-driven spend as a business driver times a unit cost, give new workloads and AI usage a range, then measure the forecast against actuals line by line.
Provider forecasts are a sound baseline when usage is steady, but they extrapolate history and can't see next quarter's launch, migration or model change. The method below ties each line to a driver and is worked through one illustrative month. Reforecast cadence is covered in annual budget vs rolling forecast.
Why do provider forecasts miss when usage moves?
Because they project the past forward. AWS says a Cost Explorer forecast is "based on your past usage." Forecasts carry an 80% prediction interval, and none appears without enough history, which AWS says is common for accounts with less than one full billing cycle.
Azure Cost Management uses a "time series linear regression" model and projects up to a year ahead. For forecasts longer than 90 days it looks back 90 days, and Microsoft recommends at least 90 days of history for an annual forecast.
Microsoft's FinOps guidance says that for consistent usage without anomalies, the built-in forecast "might be all you need." We agree. But a regression doesn't know that a product launches in March, that a data warehouse is migrating in, or that the support team is switching to a costlier model. In our view those changes break forecasts, and someone in the business knows about them first.
In the FinOps Foundation's forecasting capability, Crawl-level forecasts are built manually and based mainly on historical spend, and engineering isn't involved. At Run, forecast models combine rolling, trend-based and driver-based approaches, tracked against discount-adjusted, amortized usage.
What should you split out before you forecast?
In our view, four layers, split by what moves them, cover most estates.
| Layer | What moves it | How to forecast it | Who supplies the input |
|---|---|---|---|
| Committed | Contract terms | A schedule from the contracts | FinOps, procurement |
| Usage-driven | Business volume and efficiency | Driver volume times unit cost | Business owner (volume), engineering (unit cost) |
| New or changing workloads | Migrations, launches | Estimate with a range and start date | Engineering |
| AI services | Requests, tokens, model, price | Volume times cost per unit of work, wide range | Product, engineering |
Two choices come first. Microsoft's FinOps guidance says to decide whether to forecast costs as billed or effective costs after commitment discounts, and to filter out purchases and forecast them separately. The FOCUS specification draws the same line. In version 1.4, EffectiveCost must be zero on a purchase charge that is intended to cover related eligible charges, and the covered usage charges carry the recognized portion of that purchase.
A forecast on EffectiveCost spreads a commitment across the months that use it; one on BilledCost shows when cash goes out. We think finance needs both, built from the same lines. How your company accounts for prepaid commitments is a question for your controllers and auditors.
How do you tie usage to a business driver?
Pick a volume the business already forecasts, such as orders or active customers. Divide the usage-driven cost of the services behind it by that volume to get a unit cost, and trend it. The forecast is planned volume times expected unit cost.
Volume comes from the business plan, so FinOps isn't guessing at demand. Unit cost belongs to engineering, which can move it through rightsizing or architecture changes. A miss then shows which half missed.
In our view a good driver explains most of the cost movement, is forecast by the business at least monthly and keeps a stable definition. The test for a decision-ready unit cost in why most IT unit costs never change a decision applies to cloud unit economics too.
A worked example: one month, forecast and actual
The numbers below are made up to show the mechanics; they don't come from any company.
In February, a digital retailer forecasts March cloud spend on an effective-cost basis.
- Committed: $420,000. Savings plans and reservations amortize to $380,000 a month, plus a $40,000 monthly support and marketplace contract. It doesn't move with volume.
- Usage-driven: $462,000. Usage not covered by commitments ran at $45 per 1,000 orders last quarter. Engineering commits to $44 after a rightsizing round. Sales plans 10.5 million orders, with a range of 9.8 million to 11.5 million.
- Migration: $60,000. A data warehouse moves in on March 1. Engineering estimates $45,000 to $90,000 a month.
- AI support assistant: $48,000. 400,000 conversations at $0.12 each. The range runs from 300,000 at $0.08 to 600,000 at $0.18.
| Line | Low | Base | High |
|---|---|---|---|
| Committed | $420,000 | $420,000 | $420,000 |
| Usage-driven | $431,200 | $462,000 | $506,000 |
| Migration | $45,000 | $60,000 | $90,000 |
| AI assistant | $24,000 | $48,000 | $108,000 |
| Total (Low and High are outer bounds) | $920,200 | $990,000 | $1,124,000 |
Lines rarely all miss the same way, so summed lows and highs are outer bounds.
| Line | Forecast | Actual | Variance | Volume or timing effect | Rate effect |
|---|---|---|---|---|---|
| Committed | $420,000 | $420,000 | $0 | none | none |
| Usage-driven | $462,000 | $480,600 | +$18,600 | +$13,200 (10.8M orders) | +$5,400 ($44.50 per 1,000) |
| Migration | $60,000 | $30,000 | -$30,000 | -$30,000 (cutover slipped two weeks) | none |
| AI assistant | $48,000 | $96,000 | +$48,000 | +$9,600 (480,000 conversations) | +$38,400 ($0.20 each after a model change) |
| Total | $990,000 | $1,026,600 | +$36,600 |
On the Foundation's formula, (forecast minus actual) divided by forecast, accuracy is -3.7%: spend came in above forecast. That is inside the under-5% variance the Foundation's maturity model lists as a sample Run goal.
Underneath, the AI line missed by 100% and the migration by 50%, partly cancelling out. The migration landed below its own low case because its range ignored timing, and the usage range flexed only volume, so its rate miss went unmodeled. In our view ranges should flex every driver, including start dates. The AI rate effect is the one to chase: a model change raised cost per conversation by two-thirds and nobody updated the forecast. Volume misses are usually business news; rate misses usually belong to whoever owns the architecture or model choice.
How should the forecast shape commitment purchases?
An AWS Savings Plan is a commitment measured in dollars per hour, typically for one or three years, and usage beyond it is charged at On-Demand rates. AWS allows a return only for plans with an hourly commitment of $100 or less, bought in the past 7 days and in the same calendar month.
We think new savings plans and reservations should be sized against the low case of the usage-driven line, not the base. If volume lands high, the excess costs On-Demand rates; if it lands low, the commitment still runs its term.
As of November 2026, Microsoft says Azure reservations purchased after February 1, 2027 aren't eligible for exchange if savings plans support the service, such as Virtual Machines, App Service and SQL Database. Reservations for those services bought before then keep one final exchange. Cancellations are capped at $50,000 of commitment in a rolling 12-month window for a billing profile or single enrollment. In our view that makes a forecast that runs high costlier, because an oversized reservation is harder to undo.
How should AI usage be forecast?
The FinOps Foundation's paper on forecasting AI services costs, last updated July 7, 2025, recommends applying quantity times rate to each cost domain, and "a weekly/monthly forecasting cadence due to the high variability of AI costs." One outcome it sets is AI cost forecasts within 5% of actuals per month. A March 2026 Foundation paper adds that AI prices have tended to fall while quality and performance jump every 6 to 9 months, which it says matters for projecting costs and for long-term commitments.
In our view three practices follow:
- Forecast two drivers. Volume, and cost per unit of work (tokens per request times price). A model change moves the second, not the first.
- Treat a model or prompt change as a forecast event. The team submits a new cost-per-unit estimate before it ships. The Foundation's forecasting capability calls this shifting left: the estimate made before a decision is tracked for accuracy like any other forecast.
- Start wide and narrow. We read the 5% target as a goal for established AI workloads; new ones deserve a range until they have some history.
How should you measure forecast variance?
The Foundation's forecasting capability defines KPIs including forecast accuracy rate for spend, the same rate for usage by service or SKU, and forecast drift rate: the new forecast minus the previous one, divided by the previous one. On thresholds it says: "Each organization defines its acceptable variance." The maturity model page lists sample community goals of below 20% variance at Crawl, less than 10% at Walk and less than 5% at Run. Microsoft's guidance targets under 12% variance for the forecast algorithm's accuracy when there are no anomalies.
We would add four habits:
- Measure by line, not only in total. The example's -3.7% hid two large misses.
- Split each miss into volume and rate. Each half has a different owner.
- Set thresholds by layer. Near zero for committed spend, tight for usage-driven lines, looser for new workloads and AI at first.
- Track drift with a reason. If the March forecast moved from $950,000 in January to $990,000 in February, drift is 4.2%. Drift with a recorded cause is the forecast working; drift without one is noise.
A driver-based forecast works only when the volume, the unit cost and the commitment each have a named owner, which in our view makes it an operating-model question before it is a modeling one.
Sources
- FinOps Foundation, FinOps Framework capability: Forecasting, accessed November 21, 2026
- FinOps Foundation, FinOps Maturity Model, accessed November 21, 2026
- FinOps Foundation, How to Forecast AI Services Costs in Cloud, last updated July 7, 2025
- FinOps Foundation, Effect of Optimization on AI Forecasting, last updated March 17, 2026
- FOCUS, FOCUS Specification v1.4: Effective Cost, v1.4 ratified June 4, 2026 per FinOps Foundation, Introducing FOCUS 1.4: Invoice Reconciliation, Commitment Details, and Specification Maturity, June 10, 2026
- AWS, Forecasting with Cost Explorer, accessed November 21, 2026
- AWS, What are Savings Plans?, accessed November 21, 2026
- AWS, Savings Plans FAQs, accessed November 21, 2026
- Microsoft Learn, Quickstart: Start using Cost analysis, last updated July 1, 2025
- Microsoft Learn, Forecasting (FinOps Framework guidance), last updated April 2, 2025
- Microsoft Learn, Self-service exchanges and refunds for Azure Reservations, last updated July 29, 2026
Frequently asked questions
How is a driver-based cloud forecast different from a trend forecast?
A trend forecast projects past spend forward, which is how AWS Cost Explorer and Azure Cost Management build theirs. A driver-based forecast multiplies a business volume, such as orders or active customers, by a unit cost that engineering owns. It can reflect a planned launch, migration date or efficiency gain before any of them shows up in the bill, and each miss can be traced to volume or rate.
What forecast accuracy should a FinOps team target?
The FinOps Foundation says each organization defines its acceptable variance. Its maturity model lists sample community goals of below 20% at Crawl, under 10% at Walk and under 5% at Run. Microsoft's FinOps guidance targets under 12% variance for its built-in forecast algorithm when there are no anomalies. In our view, set thresholds by line: near zero for commitments, wider for new workloads and AI.
Should a cloud forecast use amortized or billed cost?
In our view, usually both. Amortized, or effective, cost spreads commitment purchases across the months that use them, which suits team budgets and unit costs. Billed cost shows when cash leaves, which treasury needs. Microsoft's guidance says to choose the basis first and forecast commitment purchases separately. In FOCUS 1.4, EffectiveCost and BilledCost carry the two views. Confirm accounting treatment with your controllers.
How often should AI spend be reforecast?
The FinOps Foundation's AI forecasting paper recommends a weekly or monthly cadence because AI costs vary so much, and sets an outcome of forecasts within 5% of actuals per month. We would also reforecast whenever a team changes model, prompt design or routing, since those changes move cost per unit of work without any change in volume.
Steve Elliott
Published November 22, 2026
More on FinOps & Cloud
FOCUS, commitments, shared cost.
Keep reading
Negotiating cloud commitments when capacity is scarce
AWS and Google say demand outruns supply, and all three report backlogs over $550B. What that does to your bargaining power and which terms still move.
Cloud repatriation: when the numbers support moving back
A TCO frame for cloud repatriation: what to count, what the exit-fee waivers cover, what 37signals' numbers show, and when moving back doesn't pay.
Who owns SaaS spend: FinOps, software asset management or procurement?
SaaS is moving into FinOps, but ITAM and procurement already hold contracts and entitlements. Who should own what, where FOCUS stands, and a working split.
Should product teams own their cloud budgets?
Product teams should hold cloud budgets once allocation puts spend in their name and shared costs have an owner. Prerequisites, guardrails and failure modes.