Azure VM Size: How to Choose the Right Size Azure offers dozens of VM families and hundreds of SKUs. Pick one based only on vCPU count or a familiar name, and you get a bottleneck, a bloated bill, or a workload that can't scale.

The stakes are real. Flexera's 2026 State of the Cloud survey of 753 cloud decision-makers puts wasted cloud spend at 29%. That's a cloud-wide figure, not an Azure VM rate, but oversized compute is a familiar contributor.

The right size depends on CPU, memory, storage I/O, network, GPU needs, availability, and growth. This guide walks through a practical framework: classify the workload, pick a family, then verify the exact SKU.

Key Takeaways

  • A VM size sets vCPUs, memory, I/O limits, network capacity, and special hardware.
  • Classify the workload first, choose a family second, compare SKUs last.
  • Validate with telemetry, load tests, regional availability, and total cost.
  • Re-check sizing as demand changes.

What is an Azure VM Size?

Microsoft defines a size as a "specific VM configuration, including vCPUs, memory, and accelerators." It is separate from three other choices: the OS image, the managed disks, and the number of instances. In the Azure CLI, --image picks the OS and --size picks the VM size.

The main VM families

Microsoft's VM sizes overview groups sizes by purpose:

  • General purpose (including burstable B-series): enterprise apps, web and app servers, small databases
  • Compute optimized: high CPU performance, such as analytics and web servers
  • Memory optimized: large in-memory databases and data warehousing
  • Storage optimized: high disk throughput and I/O
  • GPU accelerated: deep learning, AI research, graphics-intensive apps
  • High-performance compute (HPC): fluid dynamics, molecular dynamics, weather simulation, financial-risk analysis
  • FPGA accelerated: still listed, but NP-series retires May 31, 2027

Family names and availability change, so confirm them in Microsoft's documentation before committing.

What a size actually includes

Compute. vCPU count, processor generation, and CPU performance drive throughput, concurrency, and response time. B-series VMs earn credits below baseline and spend them above it. Run out, and you're throttled to baseline.

Constrained-vCPU sizes expose half or a quarter of the cores while keeping memory, storage, and I/O, and the VM price stays the same. They exist mainly for per-core licensing.

Memory and storage. RAM matters most for databases, caches, and analytics. Don't confuse it with disk.

Temporary disks are not managed disks, and data on them can be lost during maintenance, redeployment, or stop events. Some sizes, like Dsv5, have no local disk at all. Persistent capacity comes from managed disks, which you provision separately.

Network and specialized hardware. Check network bandwidth, accelerated networking (support depends on VM type and OS), GPUs, and HPC interconnects. Any of these can decide whether a workload performs.

Azure VM size components including compute memory storage network and hardware

Why the right size pays off

Accurate sizing pays off in several ways:

  • Predictable latency and higher utilization
  • Controlled spend with smoother scaling
  • Fewer capacity incidents in production

It also gives FinOps and engineering a shared set of numbers to work from.

What to Consider When Choosing the Right Azure VM Size

No single size fits every workload. Narrow by family first, then pick the SKU. Microsoft's Well-Architected Framework focuses on cores, memory, storage, and network bandwidth, and recommends pilots or proofs of concept for workload-specific testing.

Workload profile

Classify the behavior before looking at specs:

  • Steady-state production apps: general purpose
  • Bursty or dev/test: burstable B-series
  • CPU-bound: compute optimized
  • Memory-heavy databases and caches: memory optimized
  • High-I/O databases: storage optimized
  • Fault-tolerant batch: Spot VMs
  • ML, rendering, simulation: GPU or HPC

CPU and memory

vCPU count alone misleads. Look at average and peak CPU, thread count, memory working set, and cache needs. Watch for throttling and memory pressure. Azure Advisor gives useful benchmarks:

Advisor recommendation Criteria
Shut down unused VM P95 max CPU under 3%; outbound network under 2% over 7 days
Resize user-facing workload On the recommended SKU: P95 CPU and network at or below 40%; P99 memory at or below 60%
Resize non-user-facing workload On the new SKU: P95 CPU and network at or below 80%; P99 memory at or below 80%
High-utilization resize CPU and memory both at or above 90% (P50 of 30-minute averages, 7 days)

Azure Advisor VM right-sizing thresholds for CPU memory and network utilization

Note that the resize thresholds apply to the recommended SKU, not the current one. Advisor's savings estimates also exclude discounts and existing commitments.

Storage performance

Both the VM and each disk have IOPS and throughput caps, and the VM cap wins. According to Microsoft's performance guidance, a fast disk can't lift a VM's ceiling.

Evaluate read/write patterns, latency, disk count, and disk type separately from compute. Reach for a storage-optimized size only when the VM-level limits are what's holding you back.

Network, GPU, and constraints

Check inbound and outbound throughput, east-west traffic, GPU memory and compatibility, and any licensing rules that limit your eligible families.

Region, availability, and scale

A SKU that exists on paper may not be available to your subscription. Check:

  • Azure portal: Usage + quotas shows regional and per-family vCPU quotas
  • Azure CLI: az vm list-skus with location filters; inspect Zones and Restrictions
  • PowerShell: Get-AzComputeResourceSku -Location
  • Products by region: Microsoft's regional catalog

Quota doesn't guarantee capacity. For zonal deployments, confirm both the VM size and disk types exist in your zones. Plan VM Scale Sets for horizontal scaling and keep a fallback SKU.

Azure VM regional availability and scaling validation workflow

Cost and testing

Price the whole stack: VM, OS license, managed disks, network traffic, and backups. Use the Azure Pricing Calculator (default 730 hours) and the Retail Prices API. Then weigh the models:

  • Reserved VM Instances: depend on family, size, and region
  • Savings plans: 1- or 3-year hourly commitment, billed whether used or not
  • Spot VMs: variable pricing, 30 seconds' eviction notice, no availability guarantee

Test candidates under representative load and track CPU, memory, disk, network, latency, and errors. Revisit as usage shifts.

Reading a SKU name: Standard_D4s_v5

  • D: the D family
  • 4: 4 vCPUs
  • s: Premium Storage capable (not an included disk)
  • v5: series version 5

Standard_D4s_v5 has 4 vCPUs, 16 GiB of memory, and no local disk. It's still documented, though Microsoft's modernization guidance points to v6 and v7. Naming rules aren't universal, so check each suffix in the naming conventions page.

How Lucidity Can Help

Choosing a VM size settles compute sizing. The attached block storage still needs sizing. Managed disks bill on provisioned capacity, not used space, and Microsoft states that "you can't shrink or downsize a managed disk." So an oversized disk keeps costing you.

Lucidity covers that storage layer. It is not a compute-sizing tool. The platform:

  • Monitors every volume across Azure, AWS, and Google Cloud in real time
  • Expands, shrinks, and optimizes disks with zero downtime through AutoScaler
  • Flags four types of idle disks: unattached, reserved, unmounted, and zero-I/O
  • Runs as a NoOps, application-agnostic layer with no code changes
  • Is SOC 2 compliant and built for ITOps, DevOps, and FinOps teams

Lucidity reduces cloud storage spend by up to 70%. It also surfaces disks that native dashboards and standard advisor recommendations miss. It complements Azure Monitor, Azure Advisor, and your own benchmarking rather than replacing them.

Lucidity cloud storage optimization platform monitoring Azure disks

Once your VM sizing is settled, run Lucidity's free, self-serve Assessment to see how much storage you're paying for but not using. Setup takes about five minutes.

Conclusion

Choose an Azure VM size by matching workload behavior to a family, then validating the exact SKU with telemetry and testing. Prioritize the performance, resilience, and scale you need without paying for idle capacity, not the largest or most familiar SKU.

Treat sizing as ongoing. Review compute, attached storage, regional availability, and cost as workloads evolve.

Frequently Asked Questions

Which Azure VM to choose?

Match the family to the workload: general purpose for balanced apps, compute optimized for CPU-heavy work, memory optimized for databases, storage optimized for high I/O, and GPU or HPC for specialized processing. Validate the SKU with telemetry and Microsoft documentation.

What is the recommended VM size for AVD?

It depends on user count, persona, session type, apps, and profile storage design. Microsoft examples include D4s_v5 for medium single-session users and up to 4 users per vCPU for medium multi-session loads. Treat these as starting estimates and pilot before rollout.

How much storage does an Azure VM have?

VM size defines only local or temporary disk characteristics, and some sizes, like Dsv5, have none. Persistent capacity comes from separately provisioned managed disks. You set capacity, IOPS, throughput, and redundancy independently.

What are the available Azure virtual machine SKUs?

Major families include general purpose, compute, memory, and storage optimized, plus GPU, FPGA, and HPC. Specific SKUs vary by region, quota, generation, and availability. Check the Azure portal, Microsoft's regional availability page, Azure CLI, or PowerShell.

What is VM size in Azure?

It is the resource configuration that sets a VM's vCPUs, memory, network capability, storage-related limits, and specialized hardware. These in turn drive performance, scalability, and cost.