Amazon EC2 Cost and Capacity Optimization: Save Without Losing Headroom

Balance EC2 savings and capacity: what Savings Plans, Reserved Instances, capacity reservations and Spot really guarantee, and how to plan.

Cost optimization asks you to run less. Capacity planning asks you to keep enough in reserve. On EC2 the two conflict, and teams that push only on cost can find that an instance type is unavailable on the day they need it. This guide shows how to balance both.

Understand what each discount actually gives you

Pricing models reduce the price, but they differ in what they guarantee:

  • Savings Plans lower the hourly price in return for a spending commitment. They do not reserve capacity.
  • Reserved Instances also lower the price. A zonal Reserved Instance reserves capacity in one Availability Zone, but a regional one does not.
  • On-Demand Capacity Reservations hold capacity in a zone for you, and you pay for it whether you use it or not.
  • Spot Instances are cheap spare capacity that AWS can take back with short notice.

Check AWS documentation for the current rules before you rely on any of these, since terms can change.

Plan capacity for the moments that matter

The need for reserved capacity depends on the workload’s availability requirement and permitted alternatives. Plan for events where a failure is costly: a product launch, a seasonal peak or a recovery after a zone problem. Review the reservation type, lifecycle terms and matching requirements before choosing coverage for those windows. For other workloads, compare flexibility, assurance and the actual cost of each alternative.

Be flexible about instance types and zones

Flexibility is the best protection against capacity shortages and also the way to save money:

  • Allow several instance types and generations in Auto Scaling groups, using mixed instance policies.
  • Spread workloads across Availability Zones.
  • Avoid depending on one rare instance type.

A correctly configured group may use an allowed alternative, but this still depends on available capacity, quotas and compatible application configuration.

Use Spot where interruption is acceptable

Batch jobs, build servers, stateless web tiers and many data processing jobs can handle interruption. Combine Spot with On-Demand for a base that must stay up, and design the work so a stopped instance does not lose data. Do not run databases or single points of failure on Spot.

Remember that scaling needs room

Autoscaling only helps if there is capacity to scale into. Check service quotas, subnet IP space and the capacity of the instance types you pick. In Kubernetes, a disruption budget does not create capacity, as we explain in PodDisruptionBudgets do not create capacity. If you run EKS, our Kubernetes consulting team can review node groups and scaling.

Measure both sides

Track cost per workload and also track headroom: peak utilisation, failed launches and scaling delays. Report a cost reduction alongside any increased launch risk so the business can assess the trade-off. A scaling strategy and bottleneck analysis shows where the real limits are.

A balanced approach

  • Commit to the steady base with Savings Plans.
  • Reserve capacity only for critical windows.
  • Use Spot for interruptible work.
  • Review each quarter.

For a joint review of cost and capacity, see our cloud cost optimization service or contact Ampity.

Worked example: launch capacity and recovery capacity

Assume a fictional API normally needs six instances across two Availability Zones. A scheduled registration event is expected to need ten under an agreed test workload. The business also wants to know what happens if one zone becomes unavailable. These are three different questions: ordinary demand, event demand and degraded-operation demand. A monthly cost estimate alone cannot establish whether any of them can be met.

Begin with tested per-instance throughput at the required latency and error boundary. Do not multiply theoretical virtual CPU by a generic requests-per-second estimate. Include application startup, health checks, connection establishment and cache warming in the time to useful capacity. An instance marked running is not necessarily ready to serve the critical transaction. Preserve the tested artifact and configuration so the capacity assumption can be reproduced.

Suppose ten instances can handle the event in the illustrative test, with five placed in each zone. Losing one zone leaves five, not ten. If the remaining zone is expected to launch five more, recovery depends on sufficient subnet addresses, quota, compatible instance availability and a scaling process that completes before the service objective is exceeded. Spreading instances across zones does not automatically provide full event capacity during a zone loss.

Decide whether the degraded service must preserve every feature, retain only the critical registration path or temporarily queue some work. That business choice changes the compute requirement. Record the allowed degradation and recovery time explicitly. A plan based on full operation cannot quietly become a reduced-service plan during an incident simply because the latter uses fewer instances.

Compare purchasing and assurance separately

MechanismPrimary question it addressesWhat to check separately
Savings PlansPrice of eligible committed usageAvailability of launch capacity
Regional Reserved InstanceBilling discount within applicable rulesCapacity in a particular zone
Zonal Reserved InstanceMatching zonal capacity and discount rulesExact attributes and failover-zone requirements
On-Demand Capacity ReservationMatching capacity in a particular zoneBilling, targeting, quotas and lifecycle conditions
SpotLower-price interruptible capacityInterruption recovery and alternate pools

The Capacity Reservations documentation distinguishes reservations from discount mechanisms. Immediate-use and future-dated reservations have different lifecycle conditions, so do not assume every reservation can be released immediately without a remaining commitment. Confirm the reservation type, matching attributes and effective state before relying on it for an event or treating its cost as optional.

A Savings Plan is a spending commitment for eligible usage, not a promise that an unavailable instance type will launch. The Reserved Instances overview describes the discount model and the capacity distinction associated with scope. Preserve the actual purchase attributes rather than labeling every discounted instance reserved capacity in an architecture review.

Build a capacity evidence pack

Record the Region and allowed zones, instance attributes, architecture, tenancy, application artifact, required service quota and subnet address headroom. Add the tested time from launch request to ready application, the assumed demand profile and the fallback if the preferred pool is unavailable. A quota increase request is not an approved quota, and a reservation request is not proof that capacity is active and usable.

Test the fallback using the same acceptance workload. An alternative family may have different network or storage characteristics, and an Arm alternative requires compatible software. A long list of allowed instance types does not create meaningful flexibility if only one has passed deployment and application testing. Retain a small, genuinely usable set and document when each alternative is acceptable.

When using Spot Instances, design retries and state handling for interruption. Use a durable task identity, checkpoint work where appropriate and prevent duplicate externally visible effects. If an interrupted job already charged a customer or published a report, retrying the whole job without reconciliation can be incorrect even when compute replacement is fast. Cheap replacement capacity does not solve business-level duplication.

Evaluate cost under normal and adverse scenarios

Use separate scenarios instead of one optimistic total. The normal scenario includes ordinary compute and existing commitments. The event scenario adds necessary temporary capacity and supporting services. The adverse scenario includes the agreed degraded-operation plan, potential unused reservation cost and recovery work. These are planning views, not a prediction that an incident will occur or a quoted AWS insurance price.

For illustrative arithmetic, assume six ordinary units cost $60 per day and four extra event units add $40 per day under invented rates. Two event days would add $80 of compute to this simplified baseline. If matching capacity is retained for a longer period, its chargeable hours must be modeled for that period, not only for the event's traffic hours. Include the relevant pricing terms before comparing this with an alternative deployment plan.

Do not report the same eligible usage as an independent saving under multiple purchasing mechanisms. Discounts and reservations interact under documented matching and billing rules. Finance should reconcile the modeled scenario to existing purchases and the chosen cost basis. A change can improve assurance while increasing cost; that can be an appropriate business decision when its purpose and boundary are explicit.

Limitations and next action

This planning method does not guarantee capacity across an AWS service disruption or replace a tested recovery procedure. An unavailable dependency, failed deployment or constrained database can prevent useful service even when EC2 launches succeed. Include those dependencies in the recovery rehearsal. If the workload cannot tolerate interruption, a Spot-only plan is unsuitable until the application and business process have been redesigned for that failure mode.

The next action is a capacity review record for one critical event: demand assumptions, tested throughput, ordinary and degraded capacity, permitted pools, active reservation evidence, cost basis and a named release owner. Resolve any dependency on unapproved quota or untested alternatives before making a customer availability promise. Review the record after significant application or demand changes rather than keeping a stale capacity diagram as permanent proof.

Related services