How to reduce your AWS or Azure bill: a practical guide for Australian businesses

Where does the money go in an SMB cloud bill?

In most small and mid sized accounts, compute, databases, storage and network charges make up the bulk of the bill, and a surprising share of each is waste. Before cutting anything, find out what you are paying for and who owns it.

Open AWS Cost Explorer or Azure Cost Management, group by service for the last three months, then by resource or tag. Typical findings in an account nobody has reviewed for a year:

• Development and test environments running 24 hours a day, seven days a week.
• Servers and databases sized for a launch peak that never returned.
• Snapshots, backups and disk volumes left behind when servers were deleted.
• Logs kept forever because no retention period was set.
• NAT gateways and load balancers for environments nobody uses.
• AI API spend growing quietly inside a feature that ships every request to the largest model.

This guide is about running costs. If you are still planning a move, our cloud migration cost guide covers the project budget.

Which quick wins should you do first?

Switch off what isn’t needed, shrink what is oversized, and delete what is left over. These changes take days and carry little risk when done with care.

Quick win	How	Typical saving	Risk
Schedule non production environments	Stop dev and test outside business hours	Up to about 70% of their compute cost if they run only 50 hours of a 168 hour week	Low
Rightsize instances and databases	Use 30 days of CPU and memory data; drop a size where usage stays low	Dropping one size within a family roughly halves that resource’s hourly price	Low to medium; test first
Delete unattached volumes and old snapshots	Review by age; keep what your backup policy requires	Varies; can be large on old accounts	Low if checked against backup policy
Set log retention	Choose 30 to 90 days for most application logs; archive what must be kept	Often substantial on busy apps	Low
Move old objects to cheaper storage tiers	Lifecycle rules for files rarely read	Depends on volume	Low
Release unused public IPv4 addresses	AWS charges for every public IPv4 address, in use or idle	US$0.005 an hour each, about US$3.65 a month	Low

The scheduling figure is simple arithmetic: 50 running hours out of 168 is about 30%, so the saving on those resources approaches 70%. Instance sizes within a family step up in doubles, which is why one size down saves about half; whether you can drop a size depends on measured usage.

Why are NAT gateway and data transfer charges so high?

Network charges are billed by the hour and by the gigabyte, so a design that routes traffic the long way pays twice. They rarely show up as one big line, which is why they are missed.

At the time of writing, an AWS NAT gateway in Sydney costs US$0.059 an hour plus US$0.059 per GB processed (US East is US$0.045 for each). One NAT gateway left running for a month is about US$43 before any data. An architecture with one per availability zone across three environments can carry several hundred US dollars a month in NAT charges alone.

Ways to reduce them:

• Use VPC gateway endpoints for S3 and DynamoDB, so that traffic doesn’t pass through the NAT gateway.
• Remove NAT gateways from environments that don’t need outbound internet access.
• Keep chatty services in the same availability zone where resilience allows.
• Put a CDN in front of public files and images, so fewer requests reach your servers.
• Check that backups and replication aren’t crossing regions unless that is the intent.

How do you use savings plans and reservations without locking in mistakes?

Commit only to the baseline you are confident will run for the whole term, after rightsizing, never before. Committing first locks in the waste.

AWS says Compute Savings Plans reduce costs by up to 66% and apply regardless of instance family, size, region or operating system, and also to Fargate and Lambda. EC2 Instance Savings Plans go up to 72% but tie you to one instance family in one region. Microsoft says Azure Reservations can reduce resource costs by up to 72% from pay as you go prices, and allows exchanges and refunds up to US$50,000 in a rolling 12 month window.

A sensible sequence:

1. Rightsize first and let usage settle for a month.
2. Look at the lowest steady hourly spend over the last 60 to 90 days. That is your baseline.
3. Commit to most of the baseline, not all of it, with the flexible option and a one year term.
4. Review every quarter and add commitments as the baseline grows.

Serverless or containers: which costs less?

Serverless is usually cheaper for spiky, low or unpredictable traffic; containers or instances are usually cheaper for steady, busy workloads. The crossover depends on traffic shape more than technology.

Workload	Usually cheapest	Why
Internal tool used in business hours	Serverless functions, or containers scaled to zero	Pay nothing when nobody is using it
Webhooks, scheduled jobs, file processing	Serverless functions	Billed per request and duration
Customer facing app with steady daytime traffic	Containers (for example Fargate) with a savings plan	Predictable usage benefits from commitment
Busy API running all day	Containers or instances with a savings plan	Per request pricing adds up at volume
Interruptible batch work	Fargate Spot or spot instances	AWS says Fargate Spot runs at up to 70% off the regular Fargate price

Lambda is billed per request and by duration and memory. Fargate is billed for the vCPU, memory and storage your tasks use from the time the image starts downloading until the task stops. Both can be covered by Compute Savings Plans. Architecture affects cost far more than this choice, so measure before rebuilding anything to switch.

How do you control AI and GPU costs?

AI spend is driven by tokens, model choice and how often you call the model, and it grows with usage in a way fixed servers don’t. Treat it as a separate line in the budget with its own alerts.

The main levers: route simple steps to smaller, cheaper models; cache repeated prompts and answers; trim the context you send; batch work that doesn’t need an instant answer; and set per feature spending limits. Self hosting an open weight model on GPUs only pays off at sustained high volume, because GPU instances cost money every hour whether or not they are busy. Our guide to LLM running costs has the worked numbers.

Is the Sydney premium worth paying?

Australian regions cost more per hour, but for most Australian businesses the premium is worth it. At the time of writing, AWS on demand Linux prices compare like this:

Instance	Sydney (ap-southeast-2)	Melbourne (ap-southeast-4)	US East (us-east-1)	Sydney premium
t3.medium	US$0.0528/hr	US$0.0528/hr	US$0.0416/hr	27%
m7i.large	US$0.126/hr	US$0.126/hr	US$0.1008/hr	25%
c7i.large	US$0.11655/hr	US$0.11655/hr	US$0.08925/hr	31%

For one m7i.large running all month, that is about US$92 in Sydney against US$74 in US East: a difference of roughly US$18. Against that, hosting offshore adds latency for every Australian user and moves data out of the country, which can create privacy, contract and customer trust problems. Our guide to data residency vs data sovereignty explains what onshore hosting does and doesn’t achieve. Keep production and personal data onshore; consider cheaper regions only for workloads with no Australian data or latency need, such as some batch processing.

What about paying in US dollars and GST?

AWS publishes its prices in US dollars, so an Australian business carries exchange rate risk on every invoice. When the Australian dollar falls, your cloud bill rises with no change in usage.

AWS says accounts based in Australia have bought AWS services from AWS Australia since 1 November 2021, are charged 10% GST, and receive Australian GST compliant tax invoices. For Azure, check the currency and billing entity in your Microsoft agreement or partner arrangement. Build a currency buffer into budgets, and model commitments in AUD at a conservative exchange rate. Your accountant can advise on GST credits.

How do you keep costs down once they are down?

Costs creep back unless someone owns the bill, budgets have alerts, and changes are reviewed. Tools help, but ownership is what works.

• Tag every resource with application, environment and owner, and enforce tags on creation.
• Set monthly budgets per account or project, with alerts on actual and forecast spend. AWS Budgets supports both; Microsoft notes Azure budgets alert but don’t stop resources or consumption.
• Add anomaly alerts so a runaway job is caught in a day, not at month end.
• Name one person who reviews the bill monthly with whoever owns the architecture.
• Include a cost estimate in every significant change or new feature.
• Review commitments and rightsizing every quarter.

Cloud cost is one part of total cost of ownership; our software maintenance cost guide covers the rest.

How AWLabs approaches cloud cost reduction

We start with a Development Consult, then a paid Discovery Audit, credited in full against the build, that reviews your bill, architecture and usage and ranks savings by effort and risk. Changes are priced as a fixed price Schedule of Work with three options: The Minimum, The Recommended and The Complete. Where your data needs to stay in Australia, we keep it there, and we won’t recommend moving it offshore to save a few dollars. Senior engineers lead every build. See cloud infrastructure.