Compute costs

Key points

• Steady baseline: cover it with Savings Plans or Reserved Instances, then let Auto Scaling add On-Demand or Spot capacity for the peaks.
• On-Demand billing: per second with a 60-second minimum for Linux and Windows. SUSE Linux Enterprise Server is billed per full hour.
• Payment options: All Upfront, Partial Upfront or No Upfront. The more you pay upfront, the lower the rate.
• Capacity is not a discount: zonal Reserved Instances and Capacity Reservations hold capacity in one AZ. Regional Reserved Instances and Savings Plans give a discount but reserve no capacity.
• Stacking: a Capacity Reservation has no discount of its own, but Savings Plans or regional Reserved Instances apply to it.

Key points

• Compute Savings Plans: the most flexible. Apply to EC2 regardless of family, size, AZ, Region, OS or tenancy, and also to Fargate and Lambda. Up to 66%.
• EC2 Instance Savings Plans: tied to one family in one Region, free to change size, OS and tenancy. Up to 72%.
• Convertible vs Standard RIs: Convertible can be exchanged during the term but cannot be sold. Standard cannot be exchanged but can be sold in the Reserved Instance Marketplace.
• Database Savings Plans: 1-year, No Upfront commitment across Aurora, RDS, DynamoDB, ElastiCache, DocumentDB, Neptune, Keyspaces, Timestream, DMS and OpenSearch Service. Up to 35%.
• Sharing: by default the discount flows to every account in the organization. The management account can restrict it to groups or turn it off per account.

Key points

• Interruption notice: a 2-minute warning arrives as an EventBridge event and in instance metadata. The behavior can be terminate, stop or hibernate.
• Rebalance recommendation: an earlier signal of elevated risk. Capacity Rebalancing in Auto Scaling groups and EC2 Fleet replaces the instance before the notice.
• Diversify: be flexible across instance types (at least 10 per workload is the rule of thumb) and every AZ, or use attribute-based instance type selection.
• Allocation strategy: price-capacity-optimized picks the pools with the most spare capacity at the lowest price, which lowers the chance of interruption.
• Mixed instances group: one Auto Scaling group with an On-Demand base and Spot above it, across several types. EC2 Fleet does the same when you manage the lifecycle yourself. Spot Fleet (RequestSpotFleet) is a legacy API.
• Fit: stateless web tiers, batch, CI/CD, big data, HPC, rendering and containers. Not for stateful, tightly coupled or interruption-intolerant work.

Key points

• Dedicated Hosts: a whole physical server with visibility of sockets and physical cores, host affinity and per-host billing. Bring per-socket, per-core or per-VM licenses.
• Dedicated Instances: instances on single-tenant hardware, billed per instance, without control over placement on the host.
• On-Demand Capacity Reservations: capacity in one AZ for any duration, no commitment for immediate use, cancel at any time. Billed whether instances run or not.
• Future-dated reservations: request capacity for a future date with a commitment period, for planned events and launches.
• Capacity Blocks: reserve a cluster of GPU instances for machine learning.

Key points

• Compute Optimizer: analyzes the last 14 days of CloudWatch metrics and recommends sizes and idle resources for EC2, Auto Scaling groups, EBS, Lambda, ECS on Fargate, RDS, Aurora and more.
• Enhanced infrastructure metrics: a paid option that extends the lookback to 93 days, so monthly or quarterly peaks count.
• Instance family: pick by the bottleneck. General purpose (M, T), compute optimized (C), memory optimized (R, X), storage optimized (I).
• Graviton: Arm-based instances cost up to 20% less than comparable x86 instances and use up to 60% less energy for the same performance. Also offered in RDS, Aurora and EKS.
• Order of work: rightsize, then buy Savings Plans or Reserved Instances, so the commitment is not sized to waste.

Key points

• Lambda: billed by requests and duration in GB-seconds, so memory size drives cost. Nothing is billed while idle.
• Lambda on Arm: the arm64 architecture (Graviton2) offers up to 34% better price performance than x86.
• Provisioned Concurrency: removes cold starts, but you pay for the configured concurrency for as long as it is set, even with no traffic.
• Fargate: billed per second for vCPU, memory and storage above the 20 GB included, with a 1-minute minimum (5 minutes for Windows). No instances to patch or bin-pack.
• Fargate Spot and Savings Plans: Fargate Spot is up to 70% off for interruption-tolerant ECS tasks. Compute Savings Plans cut steady Fargate use by up to 50%.
• Trade-off: EC2 with Auto Scaling for constant high utilization, Lambda for spiky or event-driven loads, Fargate for containers without cluster management.

Key points

• Instance Scheduler on AWS: an AWS Solution that starts and stops EC2 instances, Auto Scaling groups and RDS instances on tag-based schedules, across accounts and Regions.
• Stopped EC2: no instance charge, but EBS volumes and public IPv4 or Elastic IP addresses are still billed.
• Hibernation: saves RAM to the encrypted EBS root volume and resumes the processes with the same instance ID. Enabled at launch, Linux RAM under 150 GiB. Pre-warms apps with slow startup.
• Stopped RDS: no instance-hour charge, but storage and backups are billed, and RDS starts the instance again after 7 consecutive days.
• Availability by class: production gets Multi-AZ and headroom. Dev and test can run in one AZ, smaller and only in working hours.