01
Workloads in private subnets read large volumes from Amazon S3: a data lake, model training, log processing. By default that traffic leaves through the NAT Gateway, which bills every gigabyte it processes.
Default configuration
Private subnet → NAT Gateway → S3. Every GB pays the NAT Gateway data processing fee.
$5,000 / month
Optimised architecture
An S3 gateway VPC endpoint and one route in the subnet route tables: S3 traffic goes straight to S3 inside the Region. AWS charges nothing extra for gateway endpoints.
$0 / month
Your figures
Price used: $0.05 / GB (NAT Gateway)
Estimated savings
$5,000 / month
$60,000 / year
How we implement it
Trade-offs to plan for
02
In a Kubernetes (EKS) cluster spread over several Availability Zones, services call each other without regard to zone. AWS bills traffic between instances in different zones $0.01 per GB in each direction, so $0.02 per GB for a call that crosses.
Default configuration
Pods spread evenly over N zones: about (N−1)/N of internal calls cross a zone.
$6,667 / month
Optimised architecture
Kubernetes Services with trafficDistribution: PreferSameZone (or Istio locality load balancing) keep calls in the caller's zone whenever a healthy endpoint exists there.
$1,333 / month
Your figures
Price used: $0.01 / GB × 2 (out + in)
Estimated savings
$5,333 / month
$64,000 / year
How we implement it
Trade-offs to plan for
03
CI runners or batch jobs of several hours keep state on disk, so teams avoid Spot Instances, which can be interrupted with a two-minute notice, and pay On-Demand prices around the clock.
Default configuration
A fleet of m5.2xlarge On-Demand instances (Linux), running all month.
$32,704 / month
Optimised architecture
The same fleet on Spot, with state on EBS or checkpoints on S3. On the two-minute Spot interruption notice, an EventBridge rule triggers automation that saves the job, moves the volume to a replacement instance and resumes.
$13,033 / month
Your figures
Price used: m5.2xlarge: $0.448/h On-Demand · $0.1486/h Spot
Estimated savings
$19,671 / month
$236,047 / year
How we implement it
Trade-offs to plan for
04
A leaderboard or ad-tech table writes thousands of items per second under one key per day (for example 2026-11-18). A single DynamoDB partition accepts at most 1,000 write units per second: past that, writes are throttled whatever the capacity you pay for. Raising provisioned capacity costs more and does not fix the hot key.
Default configuration
One partition key per day, and capacity over-provisioned to try to absorb the throttling.
$22,542 / month
Optimised architecture
Write sharding: a suffix 0…N−1 added to the key spreads writes over at least peak ÷ 1,000 partitions, and you provision only the capacity you need (or use on-demand mode).
$11,271 / month
Your figures
Price used: $0.000772 / WCU-hour
Estimated savings
$11,271 / month
$135,254 / year
Minimum shards needed: 20
How we implement it
Trade-offs to plan for
In two weeks, we audit your AWS account (network, compute, storage, data) and hand you a prioritised list of savings with the effort each one takes.
AWS public list prices in USD, excluding tax, Free Tier and negotiated discounts (Savings Plans, EDP), checked on 2026-11-18 in the AWS Price List and pricing pages. Spot: lowest price in the Region published by AWS that day; it changes by Availability Zone and over time. Illustrative scenarios, not a client system.
Sources
Describe your context, even in two lines. We come back with precise questions, then a proposal.