Discover 100 proven techniques to optimize Azure spending, improve governance, and maximize cloud efficiency
Top 100 Best Ways to Reduce Azure Cloud Costs
Organizations waste an average of 32% of their Azure spend. That is $265 billion wasted globally in 2026. This guide covers 100 specific, actionable ways to recover that money — from five-minute quick wins that save 15–25% this week, to architectural changes that compound savings quarter after quarter. Every tactic includes a real savings percentage and complexity rating.
32%
Average cloud spend wasted — Flexera 2026 State of the Cloud Report. Over $265B wasted globally per year.
72%
Maximum savings from Azure Reserved Instances on a 3-year commitment for eligible workloads
90%
Maximum savings from Azure Spot VMs for interruptible batch workloads vs pay-as-you-go
30–50%
Typical overall Azure spend reduction achievable within 90 days by implementing the strategies in this guide
The Azure bill is not high because Azure is expensive. It is high because cloud infrastructure defaults to waste. VMs are provisioned for peak load and never right-sized. Dev environments run weekends they never touch. Snapshots from deleted VMs accumulate invisibly. Storage sits in Hot tier years after anyone last accessed it. Reserved Instances go unbought on workloads that have run continuously for three years. Every one of these is recoverable money — and this guide shows you exactly how to recover it, with real savings percentages drawn from 2026 industry data and specific implementation steps for each tactic.
Figure 1 — Where Azure cost waste typically hides: breakdown of wasted spend by category
Most Azure waste concentrates in just three categories. The fastest ROI comes from addressing compute and non-production first — these two categories alone account for 63% of wasted spend.
Category 1Compute and VM OptimisationSave 20–90%Tips 1–12
01
Right-size overprovisioned VMs using Azure Advisor
73% of Azure VMs are at least one size larger than their workload requires — a finding from 2026 Azure Advisor telemetry. Open Azure Advisor → Cost → right-sizing recommendations. Azure analyses 7 days of CPU, memory, and network utilisation data and recommends specific target VM sizes. Downsizing 10 D4s v3 VMs (4 vCPU) to D2s v3 (2 vCPU) saves approximately $2,400/month. Accept recommendations in batches during a low-traffic window, test for 48 hours, then commit. Repeat monthly as workloads evolve.
Typical saving: $200–$5,000+/month depending on fleet size. Priority: do this first.
02
Auto-shutdown dev, test, and staging VMs on a schedule
A development VM running 24/7 is used roughly 10 hours per weekday — that is 250 hours used out of 720 hours billed monthly (35% utilisation). Scheduling auto-shutdown at 18:00 and auto-start at 07:30 Monday–Friday reduces VM runtime to about 207 hours/month — a 71% cost reduction. Enable via the VM blade → Auto-shutdown, or use Azure Automation runbooks for fleet-wide scheduling. Azure DevTest Labs automates this across entire lab environments with one policy.
Typical saving: 50–70% on all non-production compute. One of the highest-ROI actions available.
03
Use Azure Spot VMs for fault-tolerant batch workloads
Spot VMs use unused Azure capacity at up to 90% discount vs pay-as-you-go. They can be evicted with 30-second notice when Azure needs the capacity. Ideal for: CI/CD pipelines, ML training jobs, video encoding, batch data processing, simulation workloads, and load testing. Not suitable for: production databases, domain controllers, or any workload that cannot tolerate interruption. Use Azure VM Scale Sets (VMSS) with mixed instance policy to combine Spot VMs for scalable capacity with a small number of regular VMs as a fallback minimum.
Typical saving: 60–90% vs pay-as-you-go on eligible batch workloads.
04
Switch to B-series burstable VMs for low-average-CPU workloads
B-series VMs (B1s, B2s, B4ms, etc.) are priced 40–60% below equivalent D-series VMs. They accumulate CPU credits when idle and spend them during bursts. Perfect for: web servers with traffic spikes, small databases, development VMs, microservices, and workloads with average CPU under 40%. Monitor CPU credit balance after switching — if credits are consistently depleted, the B-series is under-sized for that workload and you need a D-series instead. The B1s at approximately $7.60/month is one of the cheapest VMs in Azure.
Typical saving: 40–60% vs equivalent D-series VMs for low-average-CPU workloads.
05
Delete orphaned managed disks from deleted VMs
When a VM is deleted, its managed disks are not automatically deleted unless you explicitly selected "Delete with VM." Over time, subscriptions accumulate hundreds of unattached managed disks billing silently. Premium SSD P30 (1TB) costs $135.17/month. A subscription with 20 orphaned P30 disks is wasting $2,700/month. Find them: Azure Portal → Disks → filter "Unattached." Or use CLI: az disk list --query "[?diskState=='Unattached']". Review each disk for any data that needs preserving, then delete or snapshot-and-delete.
Typical saving: $500–$5,000+/month at enterprise scale. Audit monthly.
06
Enable VM auto-scaling with Azure Virtual Machine Scale Sets
Applications with variable traffic patterns — e-commerce, SaaS platforms, API backends — provision VMs for peak load, then run those VMs at 10–20% utilisation during off-peak hours. Azure VMSS scales the instance count dynamically based on CPU, memory, HTTP queue depth, or custom metrics. Configure scale-out thresholds (e.g., add instances when CPU > 70% for 5 minutes) and scale-in thresholds (e.g., remove instances when CPU < 30% for 10 minutes). The difference between 20 VMs running 24/7 and scaling between 3 and 20 based on demand is roughly 60% cost reduction.
Typical saving: 30–60% on variable-traffic workloads. Benefit grows with traffic variability.
07
Use ARM-based VMs (Ampere Altra) for Linux workloads
Azure's Dplds v5 and Dplsds v5 VM series use Ampere Altra ARM processors and are priced 15–30% lower than equivalent Intel/AMD x86 VMs at the same vCPU/memory configuration. Linux workloads that are ARM-compatible (most containerised applications, Node.js, Python, Java, Go) can move to ARM VMs with minimal effort. Most Docker images have multi-architecture builds. Test with a non-production workload first, then migrate production if results are equivalent.
Typical saving: 15–30% on Linux VM costs for ARM-compatible workloads.
08
Move workloads to newer VM generation (v5 over v3/v4)
Azure regularly introduces new VM generations (Dv5, Ev5, Fsv2) at the same price point as older generations but with better price-performance ratios — or at lower price points for equivalent performance. Migrating from older generation D3 v2 VMs to newer D2s v5 VMs at the same workload tier often produces 15–25% better price-performance. Azure Advisor flags opportunities to upgrade VM generations. This is particularly impactful for workloads running 3+ years on the same VM family.
Typical saving: 15–25% price-performance improvement, or same performance at lower cost.
09
Delete idle and zombie VMs identified by Azure Advisor
Azure Advisor identifies VMs with less than 5% average CPU utilisation over the past 7 days and network throughput below thresholds — these are effectively idle. Before deleting, verify with the resource owner: some monitoring VMs, jump boxes, or standby instances have legitimate low-utilisation patterns. For confirmed zombie VMs (orphaned test instances, forgotten proof-of-concept deployments, deprecated services), deletion is 100% savings. Implement a resource owner tagging policy so every VM has an identified owner who receives cost accountability reports.
Saving: 100% of each idle VM's cost. Run this audit every month.
10
Use Azure DevTest Labs pricing for development subscriptions
Azure DevTest Labs provides discounted VM pricing (no Windows OS licensing cost) for development and testing scenarios. Visual Studio subscribers get additional monthly Azure credits ($50–$150/month depending on subscription tier). DevTest Labs enforces auto-shutdown policies, caps maximum VMs per user, and tracks costs per lab. For a development team of 10 engineers running Windows Server VMs, shifting them to DevTest Labs typically saves 30–55% compared to standard subscription pricing.
Typical saving: 30–55% on Windows development VM costs. No additional setup cost.
11
Right-size managed disk tiers to match actual IOPS requirements
Managed disk tiers (Standard HDD, Standard SSD, Premium SSD, Ultra Disk) have significant price differences. Many workloads default to Premium SSD during VM creation without actually needing the higher IOPS. Check actual IOPS usage in Azure Monitor → Metrics → Disk Read/Write Operations/sec. If a Premium SSD P30 (5,000 IOPS, $135/month) is only using 500 IOPS, downgrading to Standard SSD E30 ($40/month) or enabling on-demand disk bursting saves $95/month per disk. Additionally, enable on-demand bursting instead of buying a higher tier for occasional spikes.
Typical saving: $30–$100/disk/month for over-tiered disks. High impact at scale.
12
Deallocate (stop) VMs rather than just shutting down the OS
A VM that is "stopped" at the OS level (e.g., shutdown /s in Windows or shutdown now in Linux) is still in the Stopped state in Azure — which still charges for the VM's compute cost. Only "deallocating" a VM (Azure Portal → Stop button, or az vm deallocate) stops compute billing. The VM retains its managed disks (disk billing continues) and its private IP address if you configured a static one. This distinction is one of the most common sources of unexpected Azure charges for teams new to Azure.
Saving: 100% of the VM's compute cost while deallocated. Disks still bill.
Category 2Commitment-Based Pricing — Reserved Instances, Savings Plans, and Hybrid BenefitSave 40–85%Tips 13–23
13
Purchase Reserved VM Instances for stable production workloads
Azure Reserved Instances (RIs) provide up to 72% discount (3-year term) or up to 40% (1-year term) vs pay-as-you-go pricing on the same VM — with zero performance or configuration difference. The VM operates identically; only the billing rate changes. Best candidates: production database servers, application servers, backend APIs, and any VM running more than 700 hours per month continuously. Before purchasing, run 30–60 days of pay-as-you-go to confirm stable utilisation. Purchase at the subscription scope to maximise discount application flexibility. RIs can be exchanged for a different size within the same VM family if needs change.
Saving: up to 40% (1-year) or 72% (3-year) vs pay-as-you-go. Break-even typically within 2 months.
14
Use Azure Savings Plans for flexible compute commitments
Azure Savings Plans commit you to a fixed hourly spend amount (e.g., $10/hour) rather than a specific VM configuration. In exchange, you receive up to 65% discount (3-year term) on any covered compute service — including VMs, AKS nodes, Azure Functions Premium, Azure Container Instances, Azure Virtual Desktop, and Azure Databricks. Savings Plans are better than Reserved Instances when: your VM family changes over the commitment period, your workloads span multiple services, or you want a single commitment covering a heterogeneous compute footprint. Layering: use RIs for your stable, known configuration baseline, and a Savings Plan for the flexible remainder.
Saving: 25–65% vs pay-as-you-go. Best for diverse or shifting compute portfolios.
15
Enable Azure Hybrid Benefit for Windows Server VMs
Azure Hybrid Benefit lets you bring your existing on-premises Windows Server licences with active Software Assurance to Azure, eliminating the Windows licensing cost embedded in Azure VM pricing. On a D4s v3 Windows VM, this saves approximately 36–40% vs pay-as-you-go. It can be enabled on existing VMs without stopping them: Portal → VM → Configuration → Azure Hybrid Benefit → Enable. Use Azure Policy to enforce Hybrid Benefit on all eligible VMs and audit compliance. Combined with Reserved Instances, the total saving can reach 70–80% vs PAYG.
Saving: 36–40% on Windows VM costs. Immediate — no VM restart required. Enable today.
16
Enable Azure Hybrid Benefit for SQL Server workloads
SQL Server licensing is one of the most expensive components of running SQL workloads in Azure. Azure Hybrid Benefit for SQL Server lets you apply your on-premises SQL Server licences with Software Assurance to Azure SQL Database, Azure SQL Managed Instance, and SQL Server VMs — saving up to 55% on Azure SQL Database costs. Stacking Hybrid Benefit with Reserved Instances achieves up to 80–85% savings vs pay-as-you-go. Enable via the Azure portal on each SQL resource or set as default via Azure Policy for the subscription.
Saving: up to 55% (SQL Database) or up to 85% when stacked with RIs.
17
Purchase Azure Reservations for managed databases
Azure Reservations are available for Azure SQL Database, Azure Cosmos DB, Azure Database for PostgreSQL, Azure Database for MySQL, and other managed database services — providing 30–65% savings vs pay-as-you-go vCore pricing on 1 or 3-year terms. Managed databases typically have more predictable resource consumption than VMs, making them excellent candidates for reserved capacity. Purchase at the same scope level as your database deployments (resource group or subscription) to maximise discount application.
Saving: 30–65% on managed database vCore costs.
18
Layer Reserved Instances + Hybrid Benefit for maximum discount
Reserved Instances and Azure Hybrid Benefit are stackable — both discounts apply simultaneously. A Windows Server VM running SQL Server on D4s v3 at PAYG costs approximately $450/month. Adding a 3-year RI reduces it to $180/month (60% saving). Adding Hybrid Benefit for both Windows (36%) and SQL Server (55%) reduces it further to approximately $80–$90/month — an 80%+ saving vs the original PAYG price. The compound saving from stacking is the single most powerful cost reduction available for Windows/SQL workloads on Azure.
Combined saving: up to 80% vs pay-as-you-go for Windows + SQL Server workloads.
19
Use the RI exchange feature to avoid commitment lock-in
A common objection to Reserved Instances is fear of commitment lock-in. Azure allows RI exchanges within the same VM family (e.g., D-series) at no additional cost — you can exchange a D4s v3 1-year RI for a D8s v3 1-year RI if your workload grows, or exchange from v3 to v5 generation. Cancellations are also possible subject to a 12% early termination fee. This significantly reduces the risk of 3-year commitments. Purchase 1-year RIs first for new or uncertain workloads, then upgrade to 3-year when utilisation is confirmed stable.
Risk reduction: exchange flexibility makes 3-year commitments much safer than commonly perceived.
20
Monitor RI utilisation and address under-utilised reservations
A Reserved Instance that is 100% utilised delivers full savings. One that is 0% utilised costs the same as pay-as-you-go (you pay the RI price but receive no discount benefit). Monitor RI utilisation in Azure Cost Management → Reservations → Utilisation. Target 90%+ utilisation — anything below 70% indicates over-commitment. If a reservation is under-utilised, exchange it for a smaller VM size that matches your actual fleet, or widen the scope from a single resource group to the entire subscription so more VMs can benefit from the discount.
Saving: ensures every purchased RI delivers its intended discount. Review monthly.
21
Purchase Azure Storage Reservations for large Blob Storage accounts
Azure offers reserved capacity for Blob Storage — committing to a specific amount of storage (e.g., 100TB) for 1 or 3 years at a discounted rate. For organisations with stable, large blob storage footprints (data lakes, backup repositories, media archives), storage reservations reduce the per-GB rate by 18–38% compared to pay-as-you-go. Applicable to all three redundancy types (LRS, ZRS, GRS). Evaluate for any storage account consistently above 10TB of active data.
Saving: 18–38% on blob storage costs for stable high-volume storage workloads.
22
Use Azure Database Savings Plans for flexible database discounts
The Azure Savings Plan for Databases (1-year term) applies to Azure SQL Database, Azure SQL Managed Instance, and Azure Database for PostgreSQL, MySQL, and MariaDB. Like compute Savings Plans, it commits you to an hourly spend amount rather than a specific configuration, providing flexibility when database tiers change. The maximum 35% saving applies specifically to SQL Database serverless tier — production vCore tiers typically see 12–18% savings. Still worthwhile if your database configuration changes frequently enough that Reserved Instances are impractical.
Saving: 12–35% on database compute costs with more flexibility than RIs.
23
Stack AHB + RI for on-premises SQL Server to Azure SQL MI migration
Azure SQL Managed Instance with Hybrid Benefit applied using existing SQL Server Enterprise licences with Software Assurance reduces the per-vCore cost dramatically. Applying a 3-year Reserved Instance on top achieves an 85% discount vs the full pay-as-you-go price of SQL MI. For organisations migrating SQL Server from on-premises to Azure, this combination makes SQL MI competitive in cost with continuing to run on-premises — which fundamentally changes the ROI calculation of migration projects.
Combined saving: up to 85% vs PAYG. One of the highest-value individual optimisations in Azure.
Figure 2 — The discount stacking model: how Azure pricing layers compound to 80%+ savings
Reserved Instances and Azure Hybrid Benefit apply simultaneously on the same resource — not one or the other. For Windows/SQL workloads, stacking both is mandatory, not optional.
Category 3Storage Cost ReductionSave 40–80%Tips 24–34
24
Enable Blob Storage lifecycle management policies
Azure Blob Storage has four tiers: Hot ($0.018/GB), Cool ($0.01/GB), Cold ($0.004/GB), and Archive ($0.00099/GB). Most organisations store everything in Hot tier — including data nobody has accessed in years. Lifecycle management policies automatically move blobs to cheaper tiers based on age or last-access date. Example policy: move to Cool after 30 days untouched, Cold after 90 days, Archive after 365 days. Moving 100TB from Hot to Cool saves $800/month. Moving it to Archive saves $1,700/month. Lifecycle policies cost nothing to enable and run automatically. This is mandatory for any storage account over 1TB.
Saving: $800–$1,700/month per 100TB moved from Hot to Cool or Archive. Enable immediately.
25
Delete old and redundant snapshots on a retention schedule
Managed disk snapshots bill at $0.05/GB/month on LRS. VM snapshots accumulate silently as teams create them for migrations, OS upgrades, or "just in case" checkpoints and forget to delete them. An organisation running 200 VMs can easily accumulate 50TB of snapshots billing $2,500/month with no retention policy in place. Implement a snapshot retention policy: keep the last 7 daily, 4 weekly, 12 monthly. Automate deletion using Azure Backup's built-in retention rules or Azure Automation runbooks. Audit current snapshot inventory immediately.
Typical saving: $500–$5,000+/month at scale. Automate monthly cleanup.
26
Choose the right storage redundancy for each use case
Azure storage redundancy options range from LRS ($0.018/GB, 3 copies in one datacenter) to GRS ($0.036/GB, 6 copies across two regions) — a 2× price difference. ZRS ($0.023/GB) is the best choice for most production scenarios (3 copies across 3 availability zones in the same region). RA-GRS ($0.045/GB) is 2.5× more expensive than LRS. Audit each storage account: if data is already backed up by Azure Backup or replicated by another mechanism, LRS suffices. Dev/test, temporary, and non-critical storage never needs GRS or RA-GRS.
Saving: 30–60% vs GRS/RA-GRS for workloads where LRS or ZRS suffices.
27
Enable storage account access tier analytics to find cold data
Last-access time tracking for blobs is a feature that must be explicitly enabled (Azure Portal → Storage Account → Lifecycle Management → Enable last access time tracking). Once enabled, lifecycle policies can tier blobs based on actual last-access date — much more accurate than creation-date-only tiering. Before enabling lifecycle policies, turn on last-access tracking and wait 7–14 days to collect data. Then build the policy around actual access patterns. This prevents moving frequently-accessed data to cheaper tiers (which would increase retrieval costs and latency).
Enables precise lifecycle tiering — prevents over-tiering that increases retrieval costs.
28
Delete unused Azure File Shares and empty containers
Azure File Shares bill based on provisioned capacity (Premium tier) or used capacity (Standard tier). Premium file shares with unused provisioned capacity are particularly wasteful — you pay for the provisioned size regardless of how much data is stored. Audit all file shares: check the "used capacity" vs "provisioned capacity" ratio in Azure Monitor. Reduce provisioned capacity to match actual usage plus a 20% buffer. For file shares associated with deleted workloads, delete the share entirely. Standard tier file shares only bill for used data but unused Standard shares left open still show confusion in billing reports.
Saving: 100% of provisioned-but-unused Premium file share capacity.
29
Use Azure Backup with appropriate retention policies
Azure Backup costs scale with backup retention duration and frequency. Many organisations default to maximum retention (365 days or more) and daily backups for every resource without evaluating business requirements. Align backup policies to actual recovery point objectives: dev environments may only need 7-day retention, staging may need 30 days, production may need 365. Reducing retention from 365 to 90 days on non-production workloads typically reduces Azure Backup costs by 40–50% for those workloads. Additionally, archive the long-term backup data (annual snapshots) to Azure Backup Vault Archive tier at 75% lower cost.
Typical saving: 30–50% on backup costs by right-sizing retention policies.
30
Compress and deduplicate data before storing in Azure Blob
Storage costs are directly proportional to volume. Compressing data before uploading to Azure Blob Storage reduces the billable GB by 20–70% depending on data type (text compresses very well; already-compressed video or images compress minimally). Use gzip compression for logs, JSON, CSV, and structured data. Enable Azure Blob's built-in server-side deduplication for block blobs where applicable. For log analytics, structured log formats (Parquet, ORC, Avro) are significantly smaller than raw JSON, which also speeds up query times in Azure Synapse or Azure Data Lake.
Saving: 20–70% on storage volume for compressible data types like logs and structured data.
31
Delete log data after retention period with diagnostic settings
Azure diagnostic settings send resource logs to storage accounts or Log Analytics. Without a retention policy, logs accumulate indefinitely in storage accounts. Set a retention period on every Log Analytics workspace (go to the workspace → Usage and estimated costs → Data Retention). Set the retention period to match your compliance requirements — typically 30–90 days for most log types. Data beyond the interactive retention period can be moved to archive retention (still searchable but at 7× lower cost). Log storage that compounds without retention is one of the most common sources of unexplained storage cost growth.
Saving: 30–60% on Log Analytics and storage costs. Set retention on every workspace.
32
Use Azure Data Lake Storage Gen2 for big data workloads
Azure Data Lake Storage Gen2 (ADLS Gen2) is built on Blob Storage but with hierarchical namespace enabled, making it significantly cheaper and faster for big data analytics scenarios compared to running Azure Blob Storage accessed via Azure Synapse or Azure Databricks. The hierarchical namespace enables efficient directory operations that reduce the number of API calls and compute time in analytics pipelines — indirectly reducing Databricks and Synapse runtime costs by 20–40% on data-intensive jobs.
Saving: 20–40% on analytics compute costs through more efficient data access patterns.
33
Delete unattached public IP addresses
Azure charges $3.65/month for each reserved (static) public IP address that is not attached to a running resource. When VMs or load balancers are deleted, their associated public IPs often remain as orphaned resources. CLI audit: az network public-ip list --query "[?ipConfiguration==null]". Unattached public IPs have no value and should be deleted. In large subscriptions, unattached IPs can number in the hundreds, costing $1,000+/month for nothing. Add this check to your monthly FinOps runbook.
Saving: $3.65/month per orphaned IP. Hundreds across a large subscription add up quickly.
34
Enable intelligent tiering for Azure NetApp Files if using ANF
Azure NetApp Files is priced by service level (Standard, Premium, Ultra). Most teams provision at Premium or Ultra because initial performance testing showed the need — but storage performance requirements decrease after initial migration phases. Use ANF's dynamic service level change feature to move volumes from Ultra ($0.000542/GB/hour) to Standard ($0.000203/GB/hour) for cold or archival volumes. Standard ANF is 63% cheaper than Ultra. Review ANF volume IOPS utilisation monthly and downgrade tiers where usage permits.
Saving: 40–63% by right-sizing ANF service levels to actual IOPS requirements.
Category 4Networking and Data Transfer CostsSave 30–70%Tips 35–45
35
Route traffic via Azure Private Link instead of public internet
Data egress from Azure to the public internet costs $0.087/GB after the first 100GB/month. Azure Private Link routes traffic between services entirely within the Azure backbone — at no egress cost. For applications calling Azure SQL, Storage, Key Vault, Service Bus, or other PaaS services from within the same region, Private Link eliminates the egress cost entirely while also improving security by removing public endpoint exposure. For architectures that make millions of service-to-service API calls per month, this change can save thousands per month.
Saving: eliminates per-GB egress charges on service-to-service traffic within Azure.
36
Consolidate resources into the same Azure region
Data transfer within the same Azure region is free. Data transfer between Azure regions costs $0.02–$0.08/GB depending on regions. Applications where components are distributed across multiple regions (e.g., web servers in East US calling a database in West Europe) pay inter-region transfer fees on every API call, query, and data return. Collocating application tiers in the same region eliminates these costs entirely. If multi-region is required for availability, design the data flow to minimise cross-region calls (read replicas, data synchronisation batches rather than real-time cross-region reads).
Saving: 100% of inter-region transfer costs for collocated workloads.
37
Use Azure CDN to reduce origin egress for high-traffic content
Azure CDN caches content at edge nodes globally. Once content is cached, subsequent requests are served from the edge — no origin egress cost. CDN pricing is cheaper than origin storage egress pricing, and for highly repeated content (web assets, images, videos, public API responses) the cache hit rate is typically 80–95%. CDN egress costs approximately $0.008–$0.025/GB at global tiers vs $0.087/GB origin egress. A website delivering 100TB/month of the same content could save $6,200/month by serving 85% of it from CDN cache instead of origin.
Saving: 60–80% on egress costs for cacheable content at scale.
38
Audit and eliminate idle load balancers and application gateways
Azure Standard Load Balancer costs $0.025/hour (approximately $18/month). Azure Application Gateway costs $0.246/gateway-hour for Small ($178/month). These resources often persist after the backend VMs they were load balancing are deleted. Audit load balancers for zero-backend-count configurations. Application Gateways with no backend pool members or no active listeners are pure waste. CLI: az network lb list and review backend pool sizes. Delete any LB or AG with empty backend pools or zero connection metrics for 30+ days.
Saving: $18–$400/month per unused load balancer or application gateway.
39
Right-size ExpressRoute circuits to actual bandwidth usage
ExpressRoute circuits are provisioned at fixed bandwidth tiers (50Mbps to 10Gbps) and billed regardless of utilisation. Many organisations provision for peak projected bandwidth that is never actually reached. Monitor actual ExpressRoute circuit utilisation in Azure Monitor over a 30-day period. If P95 bandwidth utilisation is consistently below 40% of the provisioned tier, downsizing to the next lower tier reduces the monthly circuit cost by 30–60%. A 1Gbps circuit costs approximately $1,500/month; a 200Mbps circuit costs approximately $700/month — $800/month savings for the same workload.
Saving: 30–60% on ExpressRoute monthly fees by right-sizing to actual bandwidth usage.
40
Use VNet service endpoints instead of NAT Gateway for internal traffic
Azure NAT Gateway costs $0.045/hour ($32/month) plus $0.045/GB processed. For traffic that could use VNet service endpoints (e.g., Azure Storage, SQL, Service Bus traffic from within a VNet) VNet service endpoints route traffic directly over the Azure backbone at no additional cost. NAT Gateway is only necessary for outbound internet access — internal Azure service traffic can use service endpoints free of charge. Review which services in your NAT Gateway traffic path could use service endpoints instead.
Saving: $32/month base cost + $0.045/GB processing for traffic routable via service endpoints.
41
Optimise VNet peering data transfer costs
VNet peering data transfer costs $0.01/GB (intra-region) and $0.02–$0.06/GB (cross-region). In hub-spoke network architectures, traffic flowing through the hub VNet is charged peering rates twice (spoke-to-hub + hub-to-spoke). For high-bandwidth spoke-to-spoke traffic, configuring direct spoke-to-spoke peering (bypassing the hub) eliminates one leg of the peering charge. For cross-region data synchronisation, batching updates rather than streaming real-time reduces the billable GB by compressing the data and reducing API overhead per GB transferred.
Saving: 20–50% on inter-VNet transfer costs through topology and data flow optimisation.
42
Delete unused VPN Gateway connections
Azure VPN Gateways bill hourly regardless of connection state or traffic volume. A Basic VPN Gateway costs approximately $26/month; a VpnGw3 AZ costs approximately $1,230/month. VPN Gateways created for migrations, PoCs, or vendor access that have since concluded often persist indefinitely. Audit VPN Gateways for connection count and bytes transferred in Azure Monitor. A VPN Gateway with zero connections and zero bytes transferred for 30+ days is a zombie resource. Delete the gateway and recreate it if connection is needed again.
Saving: $26–$1,230/month per unused VPN Gateway depending on SKU.
43
Implement connection pooling to reduce API call volume
Many Azure services bill per-request (Azure Storage transactions, Cosmos DB RUs, Service Bus messages, API Management calls). Applications that open a new connection for each operation instead of pooling connections generate significantly more billable transactions due to connection overhead. Implement connection pooling for Azure SQL, Cosmos DB SDK connection reuse, Storage SDK batch operations, and Service Bus session management. This reduces billable transaction counts without changing business logic or reducing actual work done.
Saving: 10–30% on per-transaction billed Azure services through efficient connection management.
44
Use Azure Front Door's built-in WAF instead of separate WAF + CDN
Running a separate Azure Application Gateway with WAF + Azure CDN + Azure Traffic Manager to achieve global load balancing, WAF protection, and CDN caching costs significantly more than Azure Front Door Premium (which includes WAF, CDN, and global load balancing in a single service at a unified price). If your architecture currently uses separate Application Gateway, CDN, and Traffic Manager services for a global application, consolidating into Azure Front Door Premium typically reduces the combined monthly cost by 20–40% while simplifying the architecture.
Saving: 20–40% on global traffic management costs through service consolidation.
45
Choose the lowest-cost region that meets latency and compliance requirements
Azure VM prices vary by region. East US and West US are typically among the cheapest regions, while regions like Brazil South, South Africa North, or Australia East run 10–25% higher for equivalent VM sizes. For workloads without strict geographic requirements (back-end processing, data transformation, ML training), choosing a cheaper region saves immediately without any architectural change. Use the Azure Pricing Calculator to compare specific VM sizes across multiple regions before deploying new workloads.
Saving: 10–25% on compute costs by selecting a cost-optimised region for non-latency-sensitive workloads.
Category 5Database and Managed Service OptimisationSave 30–85%Tips 46–56
46
Use Azure SQL Database serverless tier for sporadic dev/test databases
Azure SQL Database Serverless tier automatically scales compute between a minimum and maximum vCore count and pauses the database after a configurable inactivity period (as short as 1 hour). You are only billed for compute while the database is active. A development database that is queried 8 hours per weekday bills for approximately 174 vCore-hours/month instead of 720 hours — a 76% cost reduction. Enable auto-pause on all development and test SQL databases that are not required 24/7. The database resumes automatically (with a ~30-second cold-start delay) when a query arrives.
Saving: 60–90% on compute costs for databases used fewer than 12 hours per day.
47
Right-size Azure SQL databases by monitoring DTU/vCore utilisation
Like VMs, managed databases are frequently over-provisioned. Monitor average DTU or vCore utilisation over 7–30 days. If average CPU is below 20% and maximum is below 60%, the database is a strong candidate for downsizing. Azure SQL Database allows instant elastic pool scaling without downtime. Moving from Business Critical to General Purpose tier if the workload does not require BC's in-memory OLTP or high IOPS delivers 30–50% cost reduction. Elastic Pools consolidate multiple databases sharing a DTU or vCore pool — ideal for multi-tenant SaaS applications where databases have variable, asynchronous peak patterns.
Saving: 30–60% for over-provisioned databases; additional 20–30% from Elastic Pools for SaaS patterns.
48
Optimise Cosmos DB Request Units by reviewing partition key design
Cosmos DB charges per Request Unit (RU) consumed. A poorly chosen partition key causes "hot partition" problems — one partition receives disproportionate traffic, requiring high provisioned RUs to prevent throttling even though average utilisation is low. Redesigning to a high-cardinality partition key (e.g., user ID + date composite) distributes load evenly, reducing the provisioned RUs needed for the same throughput by 30–60%. Additionally, switch from provisioned throughput to autoscale mode for Cosmos DB containers with spiky usage — autoscale scales down to 10% of max RUs during quiet periods, reducing baseline cost.
Saving: 30–60% on RU costs with optimal partition design + autoscale mode.
49
Enable Cosmos DB free tier (one per account, 1,000 RUs + 25GB forever)
Cosmos DB offers a free tier that provides the first 1,000 RU/s and 25GB of storage at no cost, forever — on one account per subscription. This does not reduce costs on existing Cosmos accounts, but any new Cosmos DB deployment should be the free-tier account. Organisations with multiple Cosmos DB accounts paying for low-throughput databases should consolidate them into containers within a single free-tier account to maximise the benefit. The 1,000 RU/s and 25GB are enough to power a production application for a small SaaS or internal tool.
Saving: $24+/month per account that benefits from the free RU/GB allocation.
50
Use Azure Database for PostgreSQL Flexible Server over General Purpose
Azure Database for PostgreSQL Flexible Server provides more granular compute options (Burstable, General Purpose, Memory Optimised) and stops billing for compute when the server is stopped — unlike the legacy Single Server offering. Migrate from Single Server (which is retiring) to Flexible Server and right-size to Burstable tier for non-production instances. A Burstable B1ms (1 vCore) Flexible Server costs approximately $14.60/month vs the cheapest General Purpose at $90+/month — a 84% difference for low-utilisation workloads. Stop development Flexible Servers after hours to eliminate compute billing entirely.
Saving: 20–84% depending on right-tier selection for Flexible Server.
51
Implement Azure Redis Cache TTL aggressively to reduce database load
Azure Cache for Redis is cheaper than the database queries it avoids. A C1 Redis cache costs $54.75/month. Caching frequently read data (product catalogues, user profiles, configuration data, session data) with appropriate TTLs reduces the number of database reads — directly reducing DTU consumption for SQL databases or RU consumption for Cosmos DB. Applications making 10M database reads per day for the same 1,000 product records could serve 90% of those reads from Redis — reducing database costs by 3–4× the Redis cost.
Saving: 30–60% on database costs for read-heavy workloads. Redis ROI is typically 5–10×.
52
Archive historical data from hot SQL databases to cold storage
Azure SQL Database pricing scales with database size in the DTU model and with IOPS requirements for large data sets. Historical data (orders from 3+ years ago, archived audit logs, historical analytics tables) that is rarely queried but retained for compliance inflates the database's required tier. Partition historical data by date and archive anything older than the active query window to Azure Blob Storage (Cool or Archive tier) or Azure Synapse. The database can reference historical data via external tables or linked services. This reduces the active database size and allows downgrading to a smaller, cheaper tier.
Saving: 40–70% on database tier costs after archiving cold historical data.
53
Delete read replicas that are not actively serving read traffic
Azure SQL Hyperscale read replicas, Cosmos DB multi-region replicas, and Azure Database for PostgreSQL read replicas each cost the same as or more than the primary instance. Replicas created for load testing, migration validation, or capacity planning are frequently left running. Check Azure Monitor for replica connection counts and query throughput — a replica with zero connections for 14+ days is not being used. Delete it immediately. Re-provision when legitimately needed again. The cost is typically $300–$3,000/month per unnecessary replica.
Saving: 100% of replica cost for unused read replicas. Audit monthly.
54
Use geo-replication only where RPO genuinely requires it
Geo-replication for Azure SQL Database creates a continuously synchronised copy in a secondary region and bills at the same rate as the primary database. For most non-critical databases, Point-in-Time Restore (free, included in database pricing) with an acceptable 1–5 hour RPO is sufficient. Only enable geo-replication where the business genuinely requires <30-minute RPO. Removing unnecessary geo-replication on 5 databases at $200/month each saves $1,000/month. This is one of the most common cases of "insurance we pay for but never actually need at this tier."
Saving: 100% of geo-replication charges for databases where PITR recovery is sufficient.
55
Migrate single-tenant databases to Elastic Pools for multi-tenant SaaS
SaaS applications that provision a dedicated Azure SQL Database per customer pay for compute capacity for each database separately. Customers do not all peak at the same time, so the aggregate provisioned capacity vastly exceeds the aggregate peak load. Azure SQL Elastic Pools share a DTU or vCore pool across multiple databases — each database draws from the shared pool as needed. 100 databases at 10 DTUs each individually costs the same as 1,000 DTUs in an Elastic Pool that could sustain the entire tenant fleet. The savings are typically 40–70% for SaaS applications with more than 10 tenants.
Saving: 40–70% on multi-tenant SaaS database costs through Elastic Pool consolidation.
56
Enable Azure Defender for Cloud free recommendations for database cost waste
Azure Advisor's database cost recommendations surface right-sizing opportunities, idle database identification, and SKU optimisation suggestions for all managed database services at no cost. Run an Advisor review specifically on database recommendations monthly. The free-tier Advisor analysis identifies the highest-value database cost changes without requiring a third-party tool. Filter recommendations by potential annual savings and prioritise accordingly. Many organisations find $5,000–$50,000/year in identified database savings during their first Advisor review.
Free to run — identifies database savings proportional to the size of your database estate.
Category 6Kubernetes and Container Cost Reduction (AKS)Save 40–90%Tips 57–67
57
Enable Cluster Autoscaler to match node count to actual pod demand
AKS clusters with fixed node counts pay for every node 24/7 regardless of pod utilisation. AKS Cluster Autoscaler automatically adds nodes when pods cannot be scheduled (scale-out) and removes underutilised nodes when pods are redistributed (scale-in). Enable on every AKS node pool: az aks update --enable-cluster-autoscaler --min-count 2 --max-count 20. Configure scale-in delay (default 10 minutes) to avoid premature scale-in during traffic bursts. For clusters with variable pod demand (CI/CD agents, batch processing, event-driven workloads), the autoscaler typically saves 30–60% vs a fixed-size cluster.
Saving: 30–60% on AKS node costs for variable-demand workloads. Enable on every cluster today.
58
Implement Vertical Pod Autoscaler (VPA) to right-size pod requests
Kubernetes pod resource requests determine how many pods fit on a node — and therefore how many nodes you need. Developers consistently over-request CPU and memory for safety, inflating the total resource requests beyond actual consumption. VPA analyses actual pod utilisation over time and recommends (or automatically applies) optimal resource requests. If pod requests are 2× actual usage, the effective node density can be doubled — halving the number of nodes needed for the same workload. Install the VPA controller from the AKS add-on marketplace and start in "Off" (recommendation-only) mode before enabling "Auto" mode.
Saving: 20–40% reduction in required nodes for clusters with over-requested pods.
59
Use Spot VM node pools for dev/test and batch AKS workloads
Create AKS node pools using Spot VMs (--priority Spot) for workloads that can tolerate interruption: CI/CD runners, ML training jobs, batch data processing, load testing, and development environments. Spot node pools cost 60–90% less than equivalent on-demand node pools. Configure the workloads on Spot pools with nodeSelector and toleration to ensure only interruptible workloads are scheduled there. Use KEDA (Kubernetes Event-driven Autoscaling) to scale Spot node pools to zero when no batch work is queued, eliminating idle costs entirely for batch scenarios.
Saving: 60–90% on AKS node costs for batch and dev/test workloads.
60
Enable AKS automatic cluster tier for free management
AKS control plane management is always free. AKS Automatic (GA 2026) goes further — it automatically manages node provisioning, OS patching, scaling, cost-efficient VM selection, and kubeconfig management. Clusters on AKS Automatic use Node Auto-Provisioning (NAP) which selects the most cost-efficient VM SKU for each workload based on actual resource requests — potentially choosing Spot, ARM, or burstable instance types that a human operator might not select manually. For new clusters, starting on AKS Automatic rather than manually managed clusters reduces both operational cost and compute cost from day one.
Saving: reduced operational overhead plus auto-selected optimal VM SKUs.
61
Scale AKS node pools to zero during nights and weekends for dev clusters
Development and staging AKS clusters used only during business hours (approximately 220 hours/month vs 720 hours/month total) can be scaled to zero nodes outside business hours. AKS user node pools support scaling to 0 nodes. Use an Azure Automation runbook or Azure Kubernetes Service's start/stop feature (az aks stop / az aks start) to stop the entire cluster at 18:00 UTC and restart at 07:30 UTC Monday–Friday. AKS stop/start only bills for the managed disks (persistent volumes) while stopped — compute billing is eliminated entirely. This saves approximately 70% of compute costs for dev/staging clusters.
Saving: 65–75% of AKS node compute costs for clusters not needed outside business hours.
62
Set namespace-level resource quotas to prevent cost overruns
Without Kubernetes ResourceQuota objects, a single misconfigured deployment can consume all available cluster capacity (or trigger unlimited cluster autoscaler expansion). Apply ResourceQuota to every Kubernetes namespace with team-appropriate limits on total CPU requests, total memory requests, and maximum replica counts. Apply LimitRange objects to set default resource requests and limits on pods that do not specify them. This prevents "request 0, no limit" pods (which effectively claim all available node resources) from inflating node requirements unnecessarily.
Prevention: avoids runaway scaling events that can multiply AKS costs 10× overnight.
63
Use KEDA (Kubernetes Event-Driven Autoscaling) for queue-based workloads
KEDA scales Kubernetes deployments to zero (0 replicas) when there is no work in the queue, and back up as work arrives. For workloads driven by Azure Service Bus queues, Event Hubs, Azure Storage queues, or HTTP request rates, KEDA enables true scale-to-zero — eliminating the minimum baseline of running pods that Horizontal Pod Autoscaling requires. KEDA is natively supported in AKS and is installed as an add-on. A batch processing fleet that previously ran 20 pods 24/7 "just in case" can scale to 0 overnight, saving 65% of compute costs immediately.
Saving: 40–80% on AKS pod costs for event-driven, queue-based processing workloads.
64
Use COST tool or Kubecost for per-namespace / per-team AKS cost visibility
Azure Cost Management shows AKS node costs at the cluster level but cannot show how much each team's namespace is spending. Kubecost (free OSS tier available) or the AKS Cost Analysis add-on (free, built into the Azure portal for AKS clusters with Azure Monitor enabled) allocates AKS compute and storage costs to namespaces, pods, and labels. Showing teams their actual daily spend in dashboards changes behaviour — teams with cost visibility consistently reduce their requests and implement proper autoscaling within weeks. This is governance infrastructure, not a direct cost tool, but it enables all other optimisations.
Enabler: teams with cost visibility reduce AKS spending by 20–35% within 90 days on average.
65
Consolidate multiple small AKS clusters into one with namespace isolation
Organisations that provision a separate AKS cluster per team or per application pay control-plane-adjacent costs (system node pool overhead, monitoring agents, networking) per cluster. A system node pool requires a minimum of 1–2 nodes (approximately $100–$300/month) regardless of workload size. Consolidating 10 small clusters (each with a $200/month system node pool overhead) into one well-governed cluster with Kubernetes namespace isolation, network policies, and RBAC saves $1,800/month in system overhead alone. Use Azure Policy and Gatekeeper to enforce governance at the namespace level as a control substitute for cluster isolation.
Saving: $100–$300/month per eliminated cluster in system node pool overhead.
66
Apply Reserved Instances to AKS node pools for production clusters
AKS node pool VMs are standard Azure VMs and are eligible for Reserved Instance discounts and Azure Savings Plans — the control plane management is free but the node VMs bill at standard VM rates. For production AKS clusters with a stable minimum node count, purchase RIs for the baseline node fleet. Apply a Savings Plan for the autoscaling component (since the autoscaler may use varying VM sizes as it scales). Stack RIs with Azure Hybrid Benefit on Windows node pools where applicable for maximum discount.
Saving: 30–72% on production AKS node costs via Reserved Instances or Savings Plans.
67
Use Azure Container Apps instead of AKS for stateless microservices
Azure Container Apps is a serverless container platform built on AKS that handles cluster management, autoscaling (including to zero), KEDA integration, and ingress automatically. For stateless microservices, event-driven processors, and API backends with variable traffic, Container Apps is typically 40–70% cheaper than managing AKS clusters manually because you eliminate the always-running system node pool, the per-cluster monitoring overhead, and the operational complexity. Container Apps bills per vCPU-second and per GB-second of active consumption — zero cost at zero load.
Saving: 40–70% for scale-to-zero workloads vs equivalent AKS cluster management overhead.
Category 7Serverless and PaaS Cost OptimisationSave 20–80%Tips 68–76
68
Use Azure Functions Consumption Plan over Premium Plan for variable workloads
Azure Functions Consumption Plan bills per-execution (1M executions free, then $0.20/million) and per GB-second of execution time. It scales to zero — you pay nothing when no functions are running. Functions Premium Plan charges an always-on cost of approximately $140/month minimum for the pre-warmed instances. Premium Plan is justified only for: functions requiring VNet integration, execution times over 5 minutes, or workloads needing consistent sub-second cold-start latency. Anything else — infrequent processing jobs, webhook handlers, scheduled tasks — belongs on Consumption Plan and saves 50–90% vs Premium.
Saving: 50–90% for sporadic serverless workloads by choosing Consumption over Premium Plan.
69
Downgrade App Service plans during non-business hours
Azure App Service plans bill hourly at the plan's tier regardless of traffic. An App Service plan running at P2v3 for production peak capacity is massively over-provisioned at 3am. Use Azure Automation to scale the plan down to a cheaper tier at night (e.g., P1v3 or B2) and back up before business hours. Alternatively, enable App Service auto-scale rules to scale out/in worker instances within the plan based on CPU/request rate — the plan's hourly cost stays fixed, but fewer instances per plan means less per-instance cost if using instance-based billing plans. For development App Services, use the Free or Shared tiers.
Saving: 30–60% on non-production App Service plans via scheduled tier scaling.
70
Move from App Service Standard to Premium v3 (better price-performance)
App Service Premium v3 plans offer better compute performance per dollar than Standard S3 plans and are required for Savings Plan discounts. The P1v3 plan ($135/month) provides 2 vCPU and 8GB RAM vs the S2 plan ($100/month) with 2 vCPU and 3.5GB RAM — significantly more memory at 35% higher cost that is often offset by needing fewer instances. For applications migrating from Standard to Premium v3, apply an Azure Savings Plan for Compute for additional 25–65% savings on top. Check Advisor recommendations for App Service premium v3 migration candidates in your subscription.
Saving: 15–30% on App Service costs when moving to Premium v3 with Savings Plans applied.
71
Use Azure Logic Apps Standard over Consumption for high-volume workflows
Logic Apps Consumption plan bills per-action execution ($0.000025/action). At high volumes (millions of executions), this accumulates rapidly. A workflow with 20 actions running 100,000 times per day generates 2 million action executions per day × $0.000025 = $50/day = $1,500/month. Logic Apps Standard plan (single-tenant, runs on App Service) costs approximately $300–$600/month and has unlimited action executions within the plan. Above approximately 30 million monthly action executions, Standard plan is cheaper. Evaluate migration at 10M+ monthly executions.
Saving: 30–70% at high workflow volumes when migrating from Consumption to Standard Logic Apps.
72
Use Service Bus Basic tier for simple queue scenarios
Azure Service Bus Standard tier costs $9.81/month + $0.10/million operations. Service Bus Premium tier costs $668/month per Messaging Unit. Service Bus Basic tier (for simple queue-only scenarios without topics/subscriptions) costs approximately $0.05/million operations with no monthly base fee. If your use case is a simple queue (no pub/sub topics, no sessions, no large messages), Basic tier is 95% cheaper than Standard. Audit your Service Bus namespaces — many were provisioned at Standard tier by default when Basic would suffice.
Saving: 80–95% for simple queue-only scenarios by selecting Basic over Standard Service Bus tier.
73
Use Azure Event Grid over Service Bus for fan-out event routing
Azure Event Grid charges $0.60/million events and is designed for fan-out (one publisher, many subscribers) event routing scenarios. Azure Service Bus Standard charges $0.10/million operations but is better suited for reliable point-to-point or competitive consumer queue patterns. If you are using Service Bus with many subscribers reading the same messages (a pub/sub pattern using topics), migrating to Event Grid for the fan-out portion reduces messaging cost and simplifies the architecture for event-driven scenarios that do not require delivery guarantees beyond best-effort.
Saving: 40–70% on messaging costs for fan-out / event notification scenarios suited to Event Grid.
74
Consolidate API Management instances across environments
Azure API Management Developer tier costs approximately $49/month and is suitable for non-production. Standard tier costs $730/month. Premium tier costs $2,800+/month. Teams that provision separate APIM instances per environment (dev/test/staging/prod) pay $730/month × 4 = $2,920/month. Consolidating all non-production environments into a single APIM Developer or Standard instance (with separate APIs and products per environment, using revisions or environments feature) saves $1,460–$2,190/month. Only production needs Premium tier — if Premium features (multi-region, VNet) are even required.
Saving: $1,000–$2,500/month by consolidating dev/test APIM instances to a single Developer tier.
75
Use Durable Functions for long-running workflows instead of premium compute
Long-running orchestrations (multi-step approval workflows, data processing pipelines, saga patterns) are sometimes implemented as always-on VMs or App Service instances. Durable Functions extend Azure Functions with built-in state management for long-running, stateful workflows — and they run on the Consumption Plan (scale-to-zero, pay per execution). An approval workflow that a VM runs for 24 hours costs VM compute for 24 hours. The same workflow in Durable Functions costs a few cents in function executions and free orchestration storage wait time.
Saving: 40–70% for orchestration-style workloads currently running on always-on compute.
76
Delete orphaned API Management subscriptions and unused products
Azure API Management does not charge per subscription or product, but orphaned subscriptions (for decommissioned services, former developers, or sunset products) create security risk and complicate billing analysis. Quarterly cleanup of unused APIM subscriptions, products, and APIs keeps the instance lean and ensures that the cost of the APIM instance is genuinely allocated to active services. More importantly, this often surfaces unused APIM instances that could be deleted entirely — which saves the full $730/month+ instance cost.
Governance: enables accurate cost allocation and identifies deletable APIM instances.
Category 8AI, GPU, and Azure OpenAI Cost ManagementSave 30–90%Tips 77–85
77
Right-size AI model selection — use GPT-5-nano or GPT-4.1-mini for simple tasks
The price difference between Azure OpenAI models spans 200×: from GPT-5-nano at $0.05/M input tokens to o3 at $10.00/M input tokens. Teams that default to GPT-5 or GPT-4.1 for all tasks — including classification, extraction, FAQ answering, and simple summarisation — overpay dramatically. Benchmark your specific use case on GPT-5-nano or GPT-4.1-mini first. If quality is acceptable (it typically is for structured tasks), the cost reduction is 95–98%. Reserve flagship models (GPT-5, o3) for genuinely complex reasoning and multi-step analysis tasks.
Saving: 95–98% on AI token costs for tasks that do not require frontier model capability.
78
Use Azure OpenAI Batch API for non-real-time AI processing
The Azure OpenAI Batch API provides a flat 50% discount on all token costs in exchange for a 24-hour turnaround time. The model is identical — same quality, same accuracy, just processed asynchronously. Any AI workload that is not user-facing in real time belongs on the Batch API: nightly document summarisation, bulk classification, corpus re-embedding, data extraction from uploaded documents, overnight report generation. Deploying a Batch endpoint alongside your Standard endpoint and routing non-urgent jobs to Batch halves the cost of the offline portion of your AI spend.
Saving: 50% on Azure OpenAI token costs for all non-real-time AI processing jobs.
79
Delete idle fine-tuned model deployments
Fine-tuned Azure OpenAI model deployments bill $1.70–$3.00/hour regardless of whether they receive any traffic — an idle GPT-4o fine-tune costs $40–$70/day ($1,200–$2,100/month). Fine-tuned deployments created for experiments, seasonal features, or demos frequently persist indefinitely after they stop being used. Audit your AI Foundry deployments monthly. Check request counts over the past 30 days in Azure Monitor. If a fine-tuned deployment shows zero or near-zero requests, delete it. You can re-deploy from the saved weights if needed without retraining.
Saving: $1,200–$2,100/month per idle fine-tuned model deployment. Audit monthly.
80
Enable prompt caching and engineer system prompts for cache hits
Azure OpenAI automatically caches input token prefixes and applies a 50–90% discount on cached tokens when the beginning of a prompt exactly matches a recent prior request. The cache hit rate is maximised when your system prompt (which is typically long and stable) is placed first and user content last — ensuring the cacheable prefix is as long as possible. Normalise system prompt formatting: consistent whitespace, field ordering, and phrasing ensures semantically equivalent prompts produce the same cache key. Monitor cache hit rate in Azure Monitor and investigate if it drops below 60% for repeated-prompt workloads.
Saving: 50–90% on input token costs for workloads with repeated system prompts.
81
Set max_tokens limits on all Azure OpenAI API calls
Output tokens are priced 4–8× higher than input tokens on most Azure OpenAI models. An uncapped max_tokens parameter on a verbose use case can produce responses 10–50× longer than necessary — at proportionally higher cost. Always set an explicit max_tokens limit appropriate to the use case: FAQ answers rarely need more than 500 tokens; classification labels need 20; structured data extraction needs 1,000–3,000. The single most common cause of unexpected Azure OpenAI bills is uncapped output on a few queries that generated extremely long responses.
Prevention: a single uncapped query can cost as much as 1,000 normal queries. Always set max_tokens.
82
Migrate to Provisioned Throughput Units (PTU) at high volume
Azure OpenAI PTU deployments reserve dedicated GPU compute at a flat hourly rate (~$2,448+/month minimum). Above approximately 150–500 million tokens per month of consistent traffic, PTU is cheaper per effective token than pay-as-you-go Standard and also provides a 99% latency SLA on token generation. Before purchasing PTUs, validate actual token volume over 30 days. Configure Spillover from the PTU deployment to a Standard fallback to handle traffic spikes without over-provisioning PTU capacity. Annual PTU reservations save an additional 35% vs monthly pricing.
Saving: 40–70% on AI token costs at high volume. Validate usage before committing.
83
Use GPU Spot Instances for ML model training jobs
GPU instances (NC, ND, NV series) are among the most expensive Azure VMs — an ND96asr v4 (8× A100 GPUs) costs approximately $36/hour pay-as-you-go. Azure Spot pricing on GPU instances can reduce this to $7–15/hour — a 60–80% saving. Machine learning training jobs that can checkpoint progress and restart are ideal Spot candidates: if the VM is evicted, the job resumes from the last checkpoint. Azure Machine Learning's built-in Spot VM support automatically checkpoints training runs and handles eviction/restart, making Spot training viable without custom engineering.
Saving: 60–80% on GPU compute for ML training using Spot Instances with checkpointing.
84
Truncate context windows for multi-turn conversations
In multi-turn chat applications, every message sent to Azure OpenAI includes the full conversation history. A 50-turn conversation on GPT-4.1 (1M token context) where each turn is 500 tokens accumulates 25,000 tokens of context by turn 50 — sending 25,000 input tokens where the first turn only sent 500. Without context pruning, the per-request input token cost grows linearly with conversation length, making long conversations 50× more expensive than the first turn. Implement sliding window context pruning: keep only the last N turns (typically 10–20) plus the system prompt and any pinned critical context.
Prevention: avoids 10–50× cost growth in multi-turn conversations. Implement context pruning.
85
Implement exponential back-off instead of aggressive retry for 429 errors
Azure OpenAI rate-limit responses (HTTP 429) indicate the deployment's tokens-per-minute or requests-per-minute quota has been reached. Aggressive immediate retry loops resend the full prompt (including all input tokens) repeatedly — billing input tokens for each attempt that is processed before hitting the limit. Implement exponential back-off: wait 1 second, then 2, then 4, then 8 — before retrying. Input tokens consumed before a 429 response was generated may or may not be billed (depends on where in processing the rate limit was hit). Monitoring your retry rate in Application Insights or Azure Monitor identifies if aggressive retries are materially inflating your input token costs.
Prevention: reduces duplicate API spend from aggressive retry patterns by 50–80%.
Category 9Governance, Visibility, and FinOps ToolingPrevent 20–40% wasteTips 86–93
86
Enable mandatory resource tagging via Azure Policy
Cost optimisation without tagging is like investigating a crime with no suspect names. Tags enable: cost allocation by team, project, or cost centre; automated reports to resource owners; policy enforcement for automatic shutdown or deletion. Implement a mandatory tagging policy at the subscription level via Azure Policy with required tags: Environment (prod/staging/dev), CostCentre, Team, and Application. Use the "deny" effect to reject resource creation without required tags. Backfill tags on existing resources using Azure Resource Graph queries and the Azure CLI tagging command. Cost reports become actionable the moment every resource has an owner.
Foundation: enables all cost accountability and automated clean-up workflows.
87
Set budget alerts at 50%, 75%, 90%, and 100% thresholds
Azure Cost Management + Billing's Budget feature sends email or Action Group alerts when spending reaches configured thresholds. Create budgets at the subscription and resource group level with alerts at 50%, 75%, 90%, and 100% of monthly target. Add a forecast-based alert at 100% for early warning of overspend before the end of the month. Action Group alerts can trigger an Azure Automation runbook that automatically shuts down non-critical resources when a budget alert fires — creating an automated spending brake. This prevents the most common cause of large, unexpected Azure bills: a runaway workload running unnoticed for weeks.
Prevention: stops cloud cost overruns before they become major incidents. Set up in 10 minutes.
88
Run Azure Advisor cost recommendations monthly
Azure Advisor continuously analyses your subscription and surfaces cost recommendations including right-sizing opportunities, unattached disks, idle VMs, Reserved Instance purchase opportunities, and Hybrid Benefit gaps. It quantifies each recommendation with an estimated annual saving and complexity rating. Azure Advisor is free and built into the portal. Schedule a monthly review of the Cost recommendations tab with your engineering and finance teams. Assign each recommendation to a resource owner for implementation within 14 days. Organisations that run Advisor monthly achieve 15–30% lower Azure bills vs organisations that do not, with no additional tool cost.
Saving: Advisor typically surfaces 10–30% of subscription spend as actionable cost savings. Free to run.
89
Implement showback / chargeback reporting to resource owners
When engineering teams see their own Azure spend in a weekly report that their manager also receives, their resource consumption decreases by 20–35% on average (FinOps Foundation 2025 State of FinOps data). Use Azure Cost Management's cost allocation rules and Power BI integration to build team-level cost dashboards showing: daily spend trend, top 5 most expensive resources, budget vs actual, and week-on-week change. Send automated weekly summaries to team leads. Chargeback (actually billing the business unit) has the strongest behaviour change effect; showback (reporting only) is sufficient in most organisations to produce significant savings.
Saving: 20–35% from behaviour change when teams see and own their cloud spending.
90
Implement Azure Policy to prevent expensive resource creation without approval
Azure Policy can "deny" the creation of specific resource types or SKUs that are expensive and require explicit approval: GPU VM creation (require approval for any NV/NC/ND series), ExpressRoute circuits above a certain bandwidth, Cosmos DB accounts without free tier, API Management Premium tier, or any resource in an expensive region. This creates a managed governance control that prevents impulsive or accidental high-cost resource provisioning. Pair with an approval workflow: a pull request to a Terraform repository reviewed by a FinOps engineer before applying expensive resource changes.
Prevention: stops the most common sources of unexpected cost spikes before they happen.
91
Use Azure Cost Management + Billing anomaly detection
Azure Cost Management includes built-in anomaly detection that uses ML to identify unusual spending patterns and alerts you when current spending deviates significantly from historical patterns. Enable anomaly alerts in Cost Management → Alerts → Anomaly alerts. These alerts fire before the end of the month billing cycle, giving you time to investigate and address the root cause. Common anomaly triggers: a runaway autoscale event, an accidentally public-internet-exposed storage account generating massive egress, or a script loop creating duplicate resources.
Prevention: catches cost anomalies hours or days after they begin rather than weeks later on the bill.
92
Adopt the FinOps framework and establish a cloud cost review cadence
One-time cost optimisation sprints deliver temporary savings — costs creep back as new deployments accumulate waste. The FinOps Foundation framework (Inform → Optimise → Operate) is the discipline that sustains savings over time. Establish: a monthly cost review meeting with engineering leads and finance, a quarterly commitment review (RI utilisation, Savings Plan coverage), a weekly anomaly review, and a FinOps champion per team who owns their team's cost recommendations. Organisations with an active FinOps practice sustain 30–40% lower cloud costs than comparable organisations without one.
Sustained saving: 30–40% lower cloud costs vs organisations without active FinOps governance.
93
Negotiate an Enterprise Agreement or Azure Commitment Discount
For organisations spending above $100,000/year on Azure, a Microsoft Enterprise Agreement or Azure Commit-to-Consume commitment provides negotiated discounts on Azure services beyond the standard pricing. The discount percentage increases with commitment size — typically 10–20% for $500K/year commitments, rising to 25–30%+ for multi-million-dollar annual commitments. Contact your Microsoft account team or a Microsoft Cloud Solution Provider (CSP) to explore commitment discounts. CSP partners also provide additional benefits including consolidated billing, technical support, and sometimes additional partner-funded discounts on top of Microsoft's standard rates.
Saving: 10–30% across all Azure spend through enterprise commitment negotiation.
Category 10Automation and Continuous OptimisationSustain all savingsTips 94–100
94
Build an Azure Automation runbook for orphaned resource clean-up
Orphaned resources (unattached disks, idle public IPs, empty NICs, unused NSGs, empty resource groups, stale snapshots) accumulate continuously as teams deploy and delete workloads. A manual quarterly audit recovers some savings but misses the continuous accumulation. Build an Azure Automation PowerShell runbook that runs weekly, identifies orphaned resources based on resource state and tags, and emails a report to resource owners with a 7-day self-service deletion deadline — then auto-deletes anything still orphaned after 7 days. This converts a recurring manual task into a sustainable automated process.
Sample check:
az disk list --query "[?diskState=='Unattached' && timeCreated < '2026-01-01']" -o table
Sample check:
az disk list --query "[?diskState=='Unattached' && timeCreated < '2026-01-01']" -o table
Sustained saving: catches orphaned resource accumulation as it happens, not quarterly.
95
Automate RI purchase recommendations with Azure Cost Management APIs
Azure Cost Management provides a REST API for RI recommendations — the same data that powers the Advisor UI but accessible programmatically. Build a monthly automation that: pulls the top RI recommendations, filters for recommendations with >40% projected savings and >85% projected utilisation, generates a report for FinOps review, and (optionally) auto-purchases approved recommendations via the Reservations API. This eliminates the human delay between Advisor identifying an RI opportunity and the organisation acting on it — in many organisations this delay is 30–90 days per recommendation.
Sustained saving: eliminates 1–3 month delay in capturing RI savings identified by Advisor.
96
Tag resources at deployment time with IaC (Terraform or Bicep)
Manual tagging after deployment is always incomplete — some resources are missed, tags are misspelled, and they are not updated when ownership changes. Embedding tags in Terraform or Bicep templates ensures every deployed resource is tagged correctly from creation: environment, team, cost centre, application, and created-by. Use a Terraform locals block to define standard tags centrally: local.common_tags = { Environment = var.environment, Team = var.team_name, CostCentre = var.cost_centre } and reference this block in every resource. Azure Policy's Deny effect becomes viable when tagging is enforced at the IaC level.
Governance: ensures 100% tag coverage, enabling all cost allocation and automation workflows.
97
Use Azure Bicep or Terraform to enforce cost-optimised defaults
If your Terraform or Bicep modules default to large VM sizes, Premium SSD, and high storage redundancy (GRS), every new deployment starts over-provisioned. Update your IaC module defaults to cost-optimised configurations: default VM size to Standard_D2s_v5 (not D8s_v3), default storage SKU to ZRS (not GRS), default database tier to General Purpose (not Business Critical), and require explicit override with justification comment to provision more expensive configurations. The most powerful cost governance is preventing over-provisioning at the point of deployment, not cleaning it up later.
Prevention: stops over-provisioning at source. All future deployments start cost-optimised by default.
98
Implement cost estimation in your CI/CD pipeline (Infracost)
Infracost is an open-source tool (infracost.io) that estimates Azure infrastructure costs from Terraform plans and posts the cost estimate as a pull request comment — before the infrastructure is deployed. When an engineer's PR that proposes adding a new VM or storage account shows "+$1,200/month" as a PR comment visible to reviewers and team leads, it changes the conversation about whether that resource is justified. Integrating cost estimation into your IaC review workflow shifts cost accountability to the design phase, where changes are free, rather than the billing phase, where they are expensive.
Prevention: catches expensive infrastructure decisions before deployment rather than after billing.
99
Use Azure Carbon Optimisation (GA 2026) to combine cost and carbon reduction
Azure Carbon Optimisation reached General Availability in 2026, allowing teams to view carbon emissions data alongside cost data in Azure Cost Management. The key insight: carbon emissions and compute waste are correlated — an idle VM costs money and emits carbon. Azure Carbon Optimisation surfaces "carbon hotspots" that often align perfectly with cost optimisation candidates. For organisations with ESG reporting requirements, co-optimising for cost and carbon is more efficient than separate initiatives. Enable Carbon Optimisation in Azure Cost Management → Carbon and review alongside cost recommendations.
Dual benefit: aligns cost optimisation with carbon reduction for ESG-driven organisations.
100
Continuously measure unit economics — cost per customer, per transaction, per API call
The FinOps Foundation identifies the shift from "managing cost" to "managing unit economics" as the maturity step that separates organisations achieving 30–40% ongoing efficiency improvement from those achieving one-time savings then regressing. Unit economics (cost per customer served, cost per transaction processed, cost per API call served) connect Azure spending to business value. When engineering teams can see that their service costs $0.023 per customer-served and the business target is $0.015, they have a specific, measurable optimisation goal rather than a vague instruction to "reduce costs." Build unit economics dashboards using Azure Cost Management + Application Insights + Power BI to track spending efficiency over time, not just absolute spend.
Long-term efficiency: the maturity shift that sustains 30–50% cost-per-unit improvement year-over-year.
Figure 3 — Where to start: savings potential vs implementation complexity matrix
Start top-left (high savings, low complexity). Move to top-right once the quick wins are exhausted. The bottom-right quadrant (high complexity, lower savings) should only be tackled after the other three are complete.
The Seven Actions That Deliver 30% Savings in 30 Days
Right-size VMs using Azure Advisor (Tip 1). Review all recommendations, downsize anything with <20% average CPU. Typical saving: 20–40% on VM costs.
Auto-shutdown dev/test VMs at 18:00 weekdays, 07:30 restart Monday morning (Tip 2). Typical saving: 50–70% of non-production compute immediately.
Enable Azure Hybrid Benefit on all eligible Windows and SQL Server resources (Tips 15 & 16). Can be enabled without stopping VMs. Saving: 36–55% on eligible resources.
Enable Blob Storage lifecycle management to automatically tier data from Hot → Cool → Archive (Tip 24). Free to enable. Moving 10TB from Hot to Archive saves $170/month.
Delete orphaned disks, snapshots, and public IPs (Tips 5, 25, 33). Find them in Azure Portal or via CLI audit. Pure waste — 100% savings on deletion.
Purchase Reserved Instances for any VM or database running more than 700 hours/month continuously (Tip 13). Saving: 40–72% vs pay-as-you-go.
Set Budget Alerts at 50%, 75%, 90%, 100% of monthly budget (Tip 87). Takes 5 minutes. Prevents the most common cause of large unexpected bills: a runaway workload running unnoticed.
Organisations waste 32% of their Azure spend on average (Flexera 2026). This is not bad luck — it is the natural default state of unmanaged cloud. The difference between a 32%-wasted Azure bill and a well-optimised one is consistent application of the tactics in this guide, governed by monthly review and FinOps accountability.
Related FAVRITE Articles
- How Much Does Azure OpenAI Cost? Complete Pricing Guide (2026)
- Azure vs AWS: Which Is Better? Complete Comparison Guide (2026)
- What Is Azure OpenAI and How Does It Work? Complete Guide (2026)
- Top 100 Ways to Learn Azure for Free (2026)