The most common cloud cost forecast is an extrapolation: take last month’s spend, apply a growth rate, and call it a forecast. This approach has a success rate roughly proportional to how stable the underlying infrastructure is — which, for most technology companies, is not very stable at all.
Cloud costs have properties that make naive extrapolation unreliable. They’re usage-based, so they vary with product traffic and growth. They’re subject to step-function changes when new workloads are launched. They’re affected by optimization actions that reduce spend in non-linear ways. And they’re subject to discount mechanisms — Reserved Instances, Savings Plans, committed use discounts — whose amortization schedules need to be modeled separately.
Building a cloud cost forecast that finance can rely on requires decomposing spend into components with different forecasting characteristics and modeling each component appropriately.
The three components of a cloud cost forecast
Committed spend is the most predictable component. Reserved Instances and Savings Plans have fixed amortization schedules that are known at purchase time. If you’ve purchased a 3-year all-upfront RI for an m5.xlarge in us-east-1, you know exactly what the monthly amortization is through the end of the commitment term. Committed spend should be forecast from the commitment schedule with near-zero variance — the only uncertainty is early termination or modification.
Baseline variable spend is the component that scales with existing workloads. This includes on-demand compute for workloads that aren’t fully covered by commitments, storage costs that grow with data volume, data transfer costs that scale with traffic, and managed service costs (RDS, ElastiCache) that scale with usage. Baseline variable spend is best forecast by multiplying the trailing 3-month average by a growth rate derived from business volume projections. If customer count is expected to grow 20% next quarter, and compute costs scale roughly with customer count, baseline variable spend should grow 15–20%.
Planned changes are the hardest to forecast and the most important to get right for long-range planning. New product launches add infrastructure. Migrations can either increase (during the migration period) or decrease (after optimization) spend. Cost optimization actions — committing unused on-demand to Reserved Instances, cleaning up zombie resources, right-sizing oversized instances — reduce spend but are hard to time precisely. Planned changes should be forecast as point estimates with an explicit range, not hidden inside the baseline.
Building the forecast model
A practical forecast model for quarterly cloud costs:
Step 1: Extract committed spend from the commitment schedule. Export all active Reserved Instances and Savings Plans from Cost Explorer. Calculate the monthly amortization for each. Sum them. This number is your committed spend forecast for each month — it’s highly accurate.
Step 2: Calculate baseline variable spend. Take the past 3 months of on-demand spend by service (EC2 on-demand, RDS, data transfer, S3, etc.). Calculate the month-over-month growth rate. Apply the growth rate from your business projections (or the trailing average if no business projection is available).
Step 3: Build a planned change register. List every planned infrastructure event for the forecast period: new workloads, migrations, planned optimization projects, committed spend purchases. For each, estimate the monthly cost impact and a confidence percentage. Apply confidence-weighted cost impacts to each forecast period.
Step 4: Sum the components with uncertainty ranges. The total forecast is the sum of committed spend (near-certain), baseline variable (medium confidence, ±10%), and planned changes (lower confidence, ±20–40%). Communicate the forecast as a range, not a point estimate.
The accuracy benchmark
Most organizations that haven’t built a structured forecast have monthly variance of 25–40% between forecast and actual. With a decomposed model applied consistently, this typically drops to 8–15% within the first two quarters, and can reach 5% or better with good planned change visibility.
A 5% variance on a $500K/month cloud bill is $25,000 — acceptable as a budget contingency. A 25% variance is $125,000 — difficult to absorb without a buffer that represents an inefficient capital allocation.
The variance benchmark matters because it determines how much budget contingency finance needs to hold. Better forecasts free up contingency budget for productive use.
What to tell the board when the forecast is wrong
Even good forecasting models miss. When they do, the communication matters as much as the number.
The structure that works: actuals vs. forecast, the primary driver of variance, whether the variance is expected to continue, and what the updated outlook is. For example: “Q2 cloud costs were $1.8M against a $1.6M forecast. The $200K variance was primarily driven by a new customer data ingestion pipeline that was launched in June ahead of schedule. This cost is expected to continue at approximately $80K/month; we’ve updated the Q3 forecast accordingly.”
This tells the board what happened, why, and what comes next. It demonstrates control over the cost structure even when a specific forecast was off. That’s the standard finance needs to hold cloud cost management to.
CostDefender provides the historical spend data, trend analysis, and anomaly identification that makes accurate cloud cost forecasting possible — replacing extrapolation with an evidence-based model.