Commitment strategy: savings plans vs reserved instances
Commitments are the fastest lever on a cloud bill and the easiest to get wrong. Here is how to size coverage without locking yourself into last year's architecture.
Discounted commitments are the single largest lever most teams have on a hyperscaler bill — 20-40% off list on steady workloads. They are also the lever teams postpone longest, because a three-year decision feels heavier than a rightsizing ticket.
Start from the floor, not the average
Look at 12 months of hourly usage and find the baseline that never disappears: the compute that runs at 3am on a Sunday. That floor is what you commit to. Everything above it stays on-demand until it proves it is permanent.
Coverage targets that survive change
- 60-80% coverage of the stable floor is a healthy target for most scaleups.
- Prefer one-year terms while your architecture is still moving; three-year only for workloads you would bet on.
- Flexible savings plans over instance-family reservations when engineering is still migrating instance types.
Track utilisation weekly, not quarterly
An unused commitment is worse than no commitment. Watch utilisation and coverage as two separate numbers: utilisation below 95% means you over-bought, coverage below target means you left discount on the table.
Who owns the decision
Commitments are a finance decision informed by engineering data. The failure mode is finance buying blind or engineering deferring forever. A monthly 30-minute review with both sides and one shared dashboard is usually enough.
What good looks like
Predictable discount, no idle commitments, and a renewal calendar nobody is surprised by. If you can answer "what happens to our bill if we cut 20% of compute next quarter" in a minute, your commitment strategy is working.