Why cloud spend drifts in multi-cloud estates
Most enterprises did not plan to run three clouds. A data team chose Google Cloud for analytics, the Microsoft agreement pulled identity and productivity workloads into Azure, and product engineering grew up on AWS. Each estate is reasonable on its own. Together they produce three invoices, three discount models, three sets of native tools and no single answer to the question "what does this product cost us to run?"
That fragmentation is where waste hides. Commitments are bought per provider and rarely reconciled. Tagging standards differ, so attribution breaks at the cloud boundary. Data moves between providers and pays transfer charges nobody designed for. Cloud cost optimization in this setting is less about finding a clever discount and more about building one consistent way to see, own and reduce spend everywhere.
This guide covers the architecture-level strategies that work across AWS, Azure and Google Cloud, then the provider-specific levers worth knowing. For a shorter, provider-neutral list, see our cloud cost optimization checklist.
Step one: one view of spend across all three clouds
Every later decision depends on comparable data. Each provider can export detailed billing records, and those exports are the foundation:
- AWS: Cost and Usage Reports (or Data Exports) to S3, with Cost Explorer for interactive analysis.
- Azure: scheduled exports from Cost Management to a storage account.
- Google Cloud: Cloud Billing export to BigQuery.
Normalise the three into a single model — service, environment, owner, cost centre, amortised cost — so a virtual machine in one cloud can be compared honestly with its equivalent in another. The FinOps Foundation's FOCUS specification defines a common schema for exactly this, and the major providers have been adding support for it. Whether you use a FinOps platform or a warehouse you build yourself matters less than the discipline of treating amortised cost, not the invoice total, as the number you manage.
Step two: attribution that survives the cloud boundary
AWS and Azure call them tags; Google Cloud calls them labels. Pick one small standard — owner, environment, application, cost centre — and apply the same keys and values in every provider. Then enforce it at creation:
- AWS: tag policies and Service Control Policies through AWS Organizations.
- Azure: Azure Policy with required-tag and inherit-tag rules on subscriptions.
- Google Cloud: organisation policies and label requirements in your provisioning code.
Use structural boundaries too — AWS accounts, Azure subscriptions and Google Cloud projects per team or product — because they attribute cost even when a tag is missed. Decide explicitly how shared costs such as networking, security tooling and platform teams are allocated, and publish showback per team before attempting chargeback.
Step three: buy commitments against a proven baseline
Commitment discounts are the largest single lever in most cloud bills, and the easiest to get wrong. Each provider offers its own model:
AWS
- Compute Savings Plans apply flexibly across instance families, regions, Fargate and Lambda; EC2 Instance Savings Plans and Reserved Instances trade flexibility for depth.
- Reserved capacity also exists for RDS, ElastiCache, OpenSearch and Redshift.
Azure
- Azure Reservations for virtual machines, SQL Database and other services, and the Azure savings plan for compute for more flexible coverage.
- Azure Hybrid Benefit lets eligible existing Windows Server and SQL Server licences be applied in Azure.
Google Cloud
- Sustained use discounts apply automatically to eligible Compute Engine usage.
- Committed use discounts are available as resource-based or spend-based commitments.
The rule is the same everywhere: rightsize first, observe steady-state usage over months, then commit to the stable floor — not the forecast. Prefer flexible instruments for workloads still changing, diarise every expiry, and review coverage and utilisation monthly. An unused commitment is pure waste with a contract attached.
Working through this in your own environment?
Send us the context and we'll tell you plainly what we'd do first — and whether it's a fit for us at all.
Talk to our teamStep four: rightsizing and elastic capacity
Each cloud publishes rightsizing recommendations — AWS Compute Optimizer and Trusted Advisor, Azure Advisor, and the Google Cloud Recommender. Treat them as a starting list, validated against a full business cycle rather than one quiet week.
- Schedule non-production off outside working hours. Dev and test environments running around the clock are among the most common sources of waste.
- Make autoscaling scale in, and test that it does.
- Use spot capacity — EC2 Spot, Azure Spot VMs, Google Cloud Spot VMs — for batch, CI and fault-tolerant workloads.
- Check instance generation and architecture. Newer families, including Arm-based options such as AWS Graviton, Azure Cobalt and Google Axion, are often more cost-efficient for compatible workloads.
- Right-size Kubernetes requests. In EKS, AKS and GKE, over-requested pods reserve capacity nobody uses; cluster autoscaling only helps once requests reflect reality.
Step five: storage tiers and data transfer
Storage
- Apply lifecycle rules: S3 Intelligent-Tiering or lifecycle transitions, Azure Blob access tiers with lifecycle management, Google Cloud Storage Autoclass or lifecycle rules.
- Delete unattached disks and expired snapshots on a schedule.
- Set log retention deliberately — verbose logs with long retention cost quietly.
- Match backup retention and replication to real recovery requirements.
Data transfer
Egress and cross-region traffic is where multi-cloud designs pay their hidden tax. Map the flows between providers and regions, keep chatty services together, put a CDN in front of repeatedly served content, use private endpoints where they avoid paid routes, and move data incrementally rather than in full scheduled copies. Sometimes the cheapest fix is architectural: placing an analytics workload in the same cloud as the data it reads.
Architecture decisions that set the cost curve
The biggest savings are designed in, not trimmed later. When reviewing a multi-cloud estate, we look for patterns that make cost grow faster than usage:
- Always-on clusters sized for peak that could be serverless or scheduled.
- Per-tenant infrastructure that could be shared safely.
- Duplicate platforms — two observability stacks, two secret stores — one per cloud.
- Managed services chosen by habit where a simpler tier would meet the SLA.
- Environments created by hand instead of as code, so nothing gets destroyed.
Building cost visibility into the DevOps workflow — cost estimates in pull requests, budgets per environment, anomaly alerts — means engineers see the price of a design while they are choosing it.
Running it as a FinOps practice
Savings last when the loop runs every month: data refreshed, anomalies reviewed, commitments checked, unit costs reported — cost per order, per user, per tenant — and a named owner for each action. Set budgets and anomaly detection in each provider (AWS Budgets and Cost Anomaly Detection, Azure budgets and cost alerts, Google Cloud budgets) and route the alerts to the team that owns the spend, not only to finance.
Archonova runs this as a cloud cost optimization engagement across AWS, Azure and Google Cloud: quick wins first, then architectural fixes, then the attribution and review routine — on its own or alongside our managed services.
A practical action plan for your team
1. Export billing data from every provider
Turn on AWS Cost and Usage Reports, Azure Cost Management exports and Google Cloud billing export to BigQuery, then normalise them into one view so spend can be compared like for like.
2. Make spend attributable
Agree a short tagging and labelling standard across all three clouds, enforce it with policy at creation, and track the unattributed share as its own metric.
3. Take the one-time savings
Rightsize against a full business cycle, remove idle and orphaned resources, and schedule non-production environments off outside working hours.
4. Commit against a proven baseline
Buy Savings Plans, reservations or committed use discounts only for steady usage you have already observed, and diarise every renewal.
5. Fix the architecture that keeps costing
Reduce cross-region and egress traffic, tier storage automatically and move suitable workloads to spot or serverless capacity so savings hold as usage grows.
6. Run it as a monthly practice
Publish showback per team, review anomalies and unit costs each month, and give a named owner the mandate to act.
What to do next
Archonova reviews your AWS, Azure and Google Cloud billing, identifies the one-time and structural savings, and helps your team keep them. Request a cloud cost review to get a prioritised plan.