Reserved Instances

TL;DR: A Reserved Instance is a gym membership instead of drop-in fees — commit to a specific instance type and region for 1 or 3 years for up to 72% off On-Demand. It's not a physical server you reserve; it's an automatic billing discount ("a coupon") that activates whenever you run the matching instance type, no config changes. The classic play for steady-state production: cover your baseline with RIs, burst on On-Demand or Spot. Modern teams increasingly prefer the more flexible Savings Plans, but RIs still win for stable, specialized fleets.

The numbers

• Up to 72% off (Standard RI); payment tiers All Upfront (deepest) → Partial → No Upfront (still 30–40% off); longer term = deeper discount.
• Regional vs Zonal scope: Regional applies across AZs and gives instance-size flexibility (two m5.large ↔ one m5.xlarge) — the usual choice; Zonal also reserves capacity in one AZ but loses cross-AZ flexibility.
• Convertible RIs trade some discount (~54% vs 72%) for the ability to exchange instance families if your needs change.
• RIs apply before Savings Plans and can be sold on the Reserved Instance Marketplace if you over-commit.
• Field examples: a stable 50-instance m5.xlarge SaaS went ~$73K → ~$28K/yr ($45K saved) on 3-yr All Upfront Regional RIs; an e-commerce shop reserved its 20-instance baseline and burst to 80+ on On-Demand/Spot for holidays — the classic hybrid.

Do this

1. Use Cost Explorer's Reserved Instance Recommendations first — it analyzes real usage and tells you exactly what to buy.
2. Cover the baseline, not the fleet — start at 50–70% of steady usage; add more as confidence grows (over-buying is the expensive mistake).
3. Choose Regional scope for cross-AZ flexibility and instance-size flexibility, unless you specifically need guaranteed capacity in one AZ.
4. Pick term by confidence — 3-year for workloads stable 6+ months with no change planned; 1-year (or Convertible) while architecture is still evolving.
5. Track RI utilization in Cost Explorer — aim for 95%+; anything lower is money left on the table.

Gotchas

• Newer generations are a different SKU — an m5 RI doesn't cover m6i; instance-size flexibility works within a family, not across generations.
• Non-refundable commitment — you can resell on the Marketplace at a loss, but plan the term to real commitment.
• A 72% discount you underuse loses to a fully-utilized smaller one — right-size before committing so you don't reserve oversized capacity.
• Region-locked — expanding to a new region means On-Demand there.

Skip this if

• The workload is dev/test that stops nightly (EC2 Instance Scheduling is the lever), experimental/variable (Spot Instances), or an evolving multi-service architecture where family/region may shift — Compute Savings Plans protect against change.
• You want the deepest EC2-only discount without per-instance-type rigidity — compare EC2 Instance Savings Plans (~72%, family+region locked but size/OS-flexible) via the Savings Plans overview.