AWS Cost Management Concepts
Use AWS Cost Management for the canonical lessons, workflows and examples. This note is a short exam decision view, not a separate pricing catalog.
Match the requirement
| Requirement | Starting point | Important limit |
|---|---|---|
| Investigate recorded spend | Cost Explorer | Billing data is delayed; select the correct cost metric |
| Analyze detailed billing data | CUR | Schema, resource detail and delivery vary |
| Allocate ownership | Tags and allocation | Activate supported billing tags; not every charge has a resource |
| Notify at a target | Budgets | Not a guaranteed hard spending cap |
| Detect unusual changes | Anomaly Detection | Learns from delayed billing history |
| Discount a steady baseline | Savings Plans / RIs | Eligibility and unused commitments matter |
| Reduce waste safely | Approved automation | Owner/dependency/recovery checks before changes |
Exam reasoning
Right-size before buying commitments. Compare utilization with coverage; neither alone proves efficient spending. Preserve availability and recovery requirements when reducing AZs, replicas or retention. Treat data transfer as a path-specific calculation, not one global rate. Stopping compute can leave storage and commitment charges intact.
A three-year commitment may be cheaper for a stable workload but inappropriate for an imminent migration. An anomaly monitor is not the best answer to a fixed monthly threshold; a budget alert is not the best answer to unexpected daily behavior. Neither guarantees immediate shutdown.
Concept map: foundations through operations
Foundations and visibility
Separate spend reduction from cost avoidance and business efficiency. A workload that doubles useful transactions while spending 20% more can improve cost per transaction without reducing the invoice. Choose a consistent cost metric and measurement period.
Use pricing models for payment/commitment choices, Cost Explorer for recorded trends and detailed exports for reproducible analytics. Invoice cash timing, amortized commitment cost, credits and net costs answer different questions. A forecast is uncertain, not a guaranteed ceiling.
Allocation and Organizations
Distinguish account ownership, resource tags, activated billing tags, Cost Categories and shared-cost rules. A tag on a resource is not automatic historical allocation. Track missing/unallocated spend rather than hide it.
Consolidated billing can share eligible discounts under preferences; it does not grant another account zonal capacity or create per-OU payer accounts. SCPs restrict supported operations, not dollar totals. Showback informs; chargeback follows an agreed internal accounting rule.
Compute and commitments
Rightsize using CPU, memory, I/O, network and latency evidence, then evaluate the hourly floor. Scheduling and Spot require availability, restart and interruption design. Stopped resources can retain storage/IPv4 charges.
RI utilization measures purchased benefit use, not guest CPU. Savings Plans commit discounted-rate dollars/hour with no rollover. High coverage can coexist with idle commitments. Capacity Reservations address availability separately from discount choice.
Storage and databases
Object-storage economics depends on object size, class, retrieval, minimum duration, versions and replication. Block storage bills provisioned size/performance, not files used. File storage adds throughput/access and availability choices.
Relational databases need query, engine/license, replica and storage evidence. DynamoDB needs item-size, consistency, index and request-unit modeling; fewer returned fields do not automatically mean fewer billed reads. A cheaper managed target can fail compatibility or increase migration effort.
Network, serverless and monitoring
Network cost is a directed path calculation with service-specific exemptions. NAT, TGW, PrivateLink and internet transfer are distinct charges. Multi-AZ resilience is not disposable simply because same-AZ communication is cheaper.
Lambda charges memory-duration and requests, with feature-specific allocation/execution costs. VPC ENIs do not have a separate Lambda hourly fee; local private RDS access is not inherently NAT traffic. Benchmark memory rather than always minimizing it.
Budgets address objectives, CAD unexpected patterns, operational alarms workload behavior. All billing-based tools have latency. Automation needs owner approval, dependencies, protection, bounded scope and a recovery plan.
Scenario reasoning and calculations
A development team proposes a commitment for resources running only 220 of 720 monthly hours. For hypothetical us-east-1 study inputs on 2026-10-12, assume $1/hour On-Demand and $0.70/hour committed: $220 usage versus $504 commitment. The nominal discount loses unless other eligible hourly usage can consume the commitment. Scheduling does not cancel the fixed obligation.
A migration proposes cheaper database storage but requires a new engine, dual-running capacity and conversion of stored procedures. Compare one-time conversion/parallel cost with recurring savings over a realistic horizon. For an illustrative $6,000 change and $500/month avoidable net savings, simple payback is 12 months, before financing and risk. Reject the change if the application will retire in six months or compatibility is unproven.
A large bill increase requires attribution before cleanup: determine account/service/usage type, selected metric, purchase timing and owner. Preserve backups and data, choose a reversible optimization and measure subsequent billing against an equivalent baseline.
Practice
For a supplied bill, identify one ownership gap, one avoidable usage item and one commitment risk. Explain the evidence, rejected alternative and safe validation step. Use the service cost notes for item-size, GB-second and traffic-path calculations.
References
Across the wiki
- GCP Cost Optimization, Committed Use Discounts (CUDs), and FinOps — cost management (GCP)
- Cost Optimization — cost management (Kubernetes)