Reserved Capacity
Commit to capacity you'll use for 1+ years to cut 30-70% off the compute bill.
Why
On-demand cloud pricing has a ~40% premium for flexibility you rarely use. Reserved Instances / Savings Plans commit to a quantity for 1-3 years in exchange for 30-70% discount. For a steady-state workload (which most production compute is), paying on-demand is burning money. The risk of reservation is over-committing — if you reserve 100 instances and use 60, you pay for 40 idle. The discipline is to reserve your steady-state baseline (the floor you'll always need) and leave the burst on-demand.
When
Once you have 3 months of steady-state usage data. Takes 1 hour to analyze + purchase. Re-evaluate quarterly.
How
1. Pull 90 days of compute usage; find the floor (min instances always running) 2. Reserve 70-80% of the floor for 1 year (conservative start) - Use Savings Plans (flexible across instance families) over RIs where possible 3. Leave the burst (floor → peak) on on-demand 4. Set a monthly reservation-utilization alert; target > 90% utilization 5. Re-evaluate at month 9: did usage grow? Buy more reservations for year 2 6. Never reserve 100% of peak — you pay for idle capacity in the trough
Example
90-day usage: floor 40 instances, peak 65, avg 48 Reservation plan: - Reserve 32 instances (80% of floor) via 1-yr Compute Savings Plan - Discount: 52% → 32 × $100 × 0.48 = $1,536/mo (was $3,200) - Savings: $1,664/mo = $20k/year - Burst (33 instances) stays on-demand: $3,300/mo Total: $4,836/mo (was $6,500) — 26% reduction, zero operational change.
Common Problems
- ⚠Reserving 100% of peak usage — you save on the busy months and pay for idle reservations in the quiet ones, netting less than expected.
Best Practices
- ✓Start with 1-year commitments; 3-year locks in a price but also locks in an instance family you may want to leave.
- ✓Use Savings Plans over Reserved Instances for flexibility — they apply across instance sizes and families.