Cloud Cost Optimisation Strategies: 2026 Playbook with Savings Percentages

10 strategies ranked by effort and impact. Start with the low-effort wins and stack strategies for compound savings of 30-45%.

Impact Priority Matrix

Start Here (High Impact / Low Effort)

  • 1. Idle Resource Cleanup (5-15%, hours)
  • 2. Storage Tiering (40-70%, days)
  • 3. Rightsizing (10-30%, days)
  • 4. RI / Savings Plans (20-40%, immediate)

High Value (High Impact / Medium Effort)

  • 5. Spot Instances (60-90%, weeks)
  • 6. Egress Optimisation (10-40%, weeks)
  • 7. K8s Bin-Packing (15-35%, weeks)
  • 8. AI/GPU Scheduling (40-70%, weeks)
  • 9. Database Optimisation (20-50%, weeks)
  • 10. Serverless Migration (30-70%, months)

1.Reserved Instances / Savings Plans

20 - 40% savingsLow effortImmediate

Commit to 1-3 year usage for predictable workloads. AWS Savings Plans are more flexible than RIs. Azure Reservations cover VMs, SQL, Cosmos DB. GCP Committed Use Discounts cover compute and memory.

Tip: Start with 70% coverage on stable workloads. Increase to 80-90% as confidence grows. Never commit 100% of spend.

2.Rightsizing

10 - 30% savingsLow-Medium effortDays

Match instance size to actual utilisation. Most workloads run at 10-30% CPU utilisation on oversized instances. Downsizing to the right fit delivers immediate savings with minimal risk.

Tip: Target instances running below 40% average CPU for 14+ days. Every cloud provider offers free rightsizing recommendations.

3.Spot / Preemptible Instances

60 - 90% savingsMedium effortWeeks

Use spot instances for fault-tolerant workloads: batch processing, CI/CD runners, data pipelines, training jobs. AWS Spot saves 60-90%, Azure Spot up to 90%, GCP Preemptible up to 91%.

Tip: Diversify across instance types and availability zones. Use Spot Fleet (AWS) or Spot.io for automated management.

4.Storage Tiering

40 - 70% savingsLow effortDays

Move infrequently accessed data to cheaper storage tiers. S3 Intelligent Tiering, Azure Cool/Archive, GCP Nearline/Coldline. Most organisations have 40-60% of storage in the wrong tier.

Tip: Enable S3 Intelligent Tiering for new buckets by default. Set lifecycle policies for objects older than 90 days.

5.Idle Resource Cleanup

5 - 15% savingsVery Low effortHours

Delete unattached EBS volumes, unused Elastic IPs, idle load balancers, orphaned snapshots, and stopped instances with attached storage. The lowest-effort, lowest-risk saving.

Tip: Run a sweep every month. Automate with Lambda/Functions or use your FinOps tool to flag idle resources automatically.

6.Egress Optimisation

10 - 40% savingsMedium effortWeeks

Data transfer out is the hidden cloud tax. Use CloudFront/CDN for static content, VPC endpoints for S3/DynamoDB, and cross-region replication only when needed. See egresscost.com for deep dive.

Tip: Audit your data transfer bill. Most organisations do not realise egress is 5-15% of total spend.

EgressCost.com

7.Serverless Migration

30 - 70% savingsHigh effortMonths

Move event-driven and variable workloads to Lambda/Functions/Cloud Run. Pay per invocation instead of per hour. Ideal for APIs with variable traffic, cron jobs, and data processing pipelines.

Tip: Start with new workloads, not legacy migrations. Serverless saves most when utilisation is below 20% on current instances.

8.Database Optimisation

20 - 50% savingsMedium-High effortWeeks

RDS/Aurora rightsizing, read replica consolidation, DynamoDB on-demand vs provisioned switching, and managed service vs self-hosted comparison. Database spend is typically 15-25% of total cloud.

Tip: Check for over-provisioned IOPS, multi-AZ deployments on non-critical databases, and unused read replicas.

9.Container / K8s Bin-Packing

15 - 35% savingsMedium effortWeeks

Optimise pod resource requests and limits. Most K8s clusters run at 30-50% utilisation because pods request more CPU/memory than they use. Bin-packing increases node utilisation and reduces node count.

Tip: Use Kubecost or OpenCost to identify over-provisioned pods. Set requests to P95 actual usage, not peak.

10.AI / GPU Workload Scheduling

40 - 70% savingsMedium effortWeeks

Schedule training jobs during off-peak hours, use spot GPU instances with checkpointing, right-size GPU SKUs (L4/T4 for inference, H100 for large training only), and batch inference requests.

Tip: A single H100 instance costs $30+/hr on-demand. Spot pricing cuts this by 60-70%. Checkpointing makes interruptions manageable.

Full AI/GPU cost guide

Compound Savings Model

No single strategy delivers 30-45% savings. Stacking strategies compounds the effect. Here is a realistic progression for a $1M/month cloud spend:

Strategy AddedIncremental SavingsRunning TotalAnnual $ (at $1M/mo)
Idle cleanup8%8%$960k
+ Storage tiering6%14%$1680k
+ Rightsizing12%24%$2880k
+ RI/Savings Plans15%35%$4200k
+ Spot instances8%40%$4800k

At $1M/month, stacking five strategies delivers approximately $4.8M in annual savings. The entire FinOps programme at Walk phase costs $230k-$490k/yr. The ROI is 10-20x.

Savings by Cloud Provider

MechanismAWSAzureGCP
Commitment DiscountSavings Plans (flexible)ReservationsCUDs (committed use)
Spot / Preemptible60-90% discountUp to 90% discountUp to 91% discount
Automatic DiscountNone (manual)None (manual)SUDs (5-17% auto)
Free Tier Duration12 months12 monthsAlways free + trial

GCP's Sustained Use Discounts (SUDs) are unique: automatic 5-17% discounts on instances running 25%+ of the month. No commitment required. AWS and Azure require explicit reservation purchases for similar savings.

Updated May 2026

Updated 2026-05-11