
Cloud waste makes the stakes real. Flexera's 2026 survey of more than 750 cloud decision-makers found wasted cloud spend rose to 29%, the first increase in five years. That figure covers cloud broadly, not just Azure VMs, but sizing mistakes feed it.
The right SKU balances compute, memory, disk performance, network capacity, workload variability, regional availability, and resilience. This guide walks through each.
Key Takeaways
- A VM SKU bundles vCPUs, memory, storage limits, network, and hardware generation.
- Start from workload behavior, then pick a family.
- Check IOPS, throughput, quota, and regional capacity before deploying.
- Benchmark first, monitor in production, and right-size as demand changes.
What is an Azure VM SKU?
A VM SKU is the specific size and configuration you select for a virtual machine. Microsoft defines a size as a fixed configuration of vCPUs, memory, and accelerators.
A series groups sizes with similar hardware and features. Take the D family: the Dsv6 series includes Standard_D4s_v6, with 4 vCPUs and 16 GiB of memory.
Two clarifications save confusion:
- Gen1 and Gen2 describe BIOS or UEFI boot architecture. The
v5orv6suffix is the series version, not the boot generation. - Managed disks, storage transactions, and bandwidth are billed separately from VM compute.
Microsoft groups sizes by workload:
| Category | Typical workloads |
|---|---|
| General purpose | Dev/test, small to medium databases, web servers |
| Compute optimized | Web and app servers, batch, gaming, analytics |
| Memory optimized | Relational databases, caches, in-memory analytics |
| Storage optimized | Big data, database and file servers, high local-disk I/O |
| GPU | Graphics, visualization, deep learning and AI |
| HPC | Fluid dynamics, finite element analysis, weather and risk modeling |
B-series burstable VMs sit within general purpose and suit web servers, proofs of concept, and small databases.
Core components of an Azure VM SKU
Microsoft's naming pattern follows this order: family, optional subfamily, vCPU count, optional constrained count, additive feature letters, optional accelerator or memory capacity, then version. Feature letters include:
aandpfor AMD and Arm processorsdfor local temporary disksfor Premium Storage capablelfor low memoryifor isolated
Check the latest Microsoft naming documentation before relying on any pattern, since attributes evolve.
Take Standard_D4ds_v5 as an example:
| Attribute | Specification |
|---|---|
| vCPU / RAM | 4 / 16 GiB |
| Max remote data disks | 8 |
| Uncached Premium SSD baseline | 6,400 IOPS; 145 MBps |
| Remote storage burst | 20,000 IOPS; 1,200 MBps |
| Temporary disk | 1 x 150 GiB |
| Network | 12,500 Mbps expected maximum, not guaranteed |

vCPU and RAM are only part of the picture. Disk limits and network ceilings often bind first.
Attributes that affect compatibility
These feature letters affect compatibility:
- Temporary disk: Dsv6 has none, unlike D4ds_v5. Scripts that assume a temp drive will break.
- Premium SSD v2 and Ultra Disk: Dsv6 supports both, subject to regional availability. They are data-only and lack host caching.
- Processor architecture: Intel, AMD, and Arm (Ampere Altra, Microsoft Cobalt 100) differ in software compatibility.
- NVMe: NVMe-only generations need compatible OS images.
- Constrained vCPU:
Standard_E32-8s_v5exposes 8 active vCPUs while keeping 32-size memory and I/O. SQL Server licensing counts available vCPUs, though VM pricing stays the same.
Burstable behavior matters too. B-series VMs earn CPU credits below baseline and spend them above it; once exhausted, the VM throttles to baseline.
Storage bursting is credit-based and can sustain maximum burst for up to 30 minutes on a best-effort basis.
Benefits of choosing an appropriate SKU
A well-matched SKU delivers predictable latency, sufficient throughput, fewer bottlenecks, and no payment for idle capacity. The details are specific:
- Exceeding VM or disk IOPS and throughput caps raises latency, and the two limits are independent.
- Each SKU behaves differently by workload, so Microsoft recommends workload-specific pilots.
- Resizing a VM leaves OS and data disks untouched. Compute rightsizing is not disk rightsizing.
That last point is easy to miss. Shrinking a VM does nothing about an oversized Premium SSD attached to it.
What to Consider When Choosing the Right Azure VM SKU
Workload profile and resource bottleneck
Classify the workload first: steady, bursty, seasonal, batch, interactive, or highly variable. Then map CPU, memory pressure, disk I/O, network throughput, latency, concurrency, and peak demand to whichever resource will constrain performance.
A steady API server and a nightly batch job need different answers, even at identical average CPU.
Narrowing by family and size
Choose a family before comparing sizes within it:
- General purpose for balanced web and app tiers
- Compute optimized for CPU-bound services and batch
- Memory optimized for databases and caches
- Storage optimized for high local I/O
- GPU or HPC for AI, rendering, and simulation
Don't pick a family because it has more vCPUs. A memory-optimized size with the right RAM beats a bigger general-purpose one.
Performance testing and observability
Establish baselines with representative load tests, measuring response time, transactions per second, and concurrency. Azure Monitor exposes the metrics you need:
Percentage CPUandAvailable Memory PercentageNetwork In TotalandNetwork Out TotalVM Cached/Uncached IOPS Consumed Percentage- Cached/uncached
Bandwidth Consumed Percentage
Compare peaks, not just averages. Azure Advisor flags resizing using a default 7-day lookback (adjustable to 90 days). For user-facing workloads, its targets are P95 CPU and outbound network at or below 40%, and P99 memory at or below 60%.

Region, quota, capacity, and resilience
Three separate conditions must all hold in your US region: the SKU is listed, your subscription has quota, and capacity exists. Quota does not guarantee capacity. Subscriptions carry both Total Regional vCPU and VM-family vCPU quotas, and deallocated cores still count.
Plan for failure:
- SKU restrictions can apply by location, zone, and subscription
- VM Scale Sets in Flexible mode can mix VM types and spread across zones and fault domains
- East US and West US are paired, but pairing does not provide automatic DR or failover
- Keep fallback SKUs ready when your preferred size can't allocate

Capacity reservations lock in a size, region, and optional zone, but unused capacity is billed.
Storage, networking, and total cost
A VM's price is a fraction of its total cost. Add managed disk type and size, snapshots, backup, egress, and OS licensing. Snapshots bill on used data, and Azure Backup charges per protected instance plus storage.
Use the Azure Pricing Calculator and Microsoft's current pricing pages rather than static numbers, which go stale quickly.
Scaling, purchasing model, and rightsizing
Your purchasing choice should follow workload shape:
| Option | Fit |
|---|---|
| Pay-As-You-Go | Unknown or short-lived demand |
| Reservations | Stable usage, 1- or 3-year commitment; no capacity guarantee |
| Savings plan | Flexible hourly commitment across eligible usage |
| Spot VMs | Interruptible batch and dev/test; 30-second eviction notice, no SLA |
| Hybrid Benefit | Qualifying Windows Server licenses with Software Assurance |
Microsoft advertises up to 80% savings when combining Hybrid Benefit with reservations versus standard Windows PAYG, varying by region and usage. Commit only after monitoring demand over time.

The FinOps Foundation advises a review schedule tied to workload peaks rather than a fixed monthly interval.
How Lucidity Can Help
VM sizing covers compute. The disks attached to those VMs are a separate cost and performance question, and that's where Lucidity fits. It is an autonomous, multi-cloud block storage management platform supporting Azure, AWS, and Google Cloud.
Lucidity gives real-time visibility into every volume. Its AutoScaler expands, shrinks, and optimizes disks with zero downtime and no code changes. That helps teams cut over-provisioned capacity and eliminate manual provisioning and ticket triage.
Other relevant capabilities:
- Lumen identifies four idle disk types: unattached, reserved, unmounted, and zero-I/O
- AI-powered environment mapping separates production from non-production
- Conversational AI returns savings reports in natural language
- Enterprise controls with SOC 2 compliance and FinOps certification

Lucidity optimizes storage. It doesn't replace workload testing or Microsoft's VM SKU documentation.
Conclusion
Choosing an Azure VM SKU is about matching size and family to real workload behavior. Work through the decision in order:
- Define workload behavior
- Choose a VM family
- Compare limits and attributes
- Confirm regional capacity and total cost
- Test under realistic load
- Monitor after launch
Traffic, data volume, and storage patterns shift over time. Revisit both the SKU and its attached disks when they do.
Frequently Asked Questions
What is a VM SKU in Azure?
A VM SKU is the specific configuration that sets compute, memory, storage, network, and related capabilities. A family is the broader group that contains many such sizes.
What are the available SKUs for Azure virtual machines?
Azure offers general-purpose, burstable, compute-optimized, memory-optimized, storage-optimized, GPU, and HPC families, each with multiple sizes. The live catalog varies by region, so check current Microsoft resources.
How do I choose the right VM size in Azure?
Profile CPU, memory, disk, network, latency, and variability, then check region, availability, and cost. Benchmark the choice, monitor it in production, and right-size as needed.
What is the difference between an Azure VM size and SKU?
Azure documentation often uses them interchangeably. SKU identifies the precise purchasable configuration, while size usually describes its resource profile.
How can I check whether an Azure VM SKU is available in my region?
Use the portal's Size selector, the Products available by region page, az vm list-skus -l eastus2 --zone, or Get-AzComputeResourceSku. az vm list-sizes is deprecated. Validate quota and capacity separately.
How often should I right-size an Azure VM?
Base reviews on workload change rather than a calendar alone. Revisit after major releases, traffic shifts, migrations, or sustained changes in CPU, memory, disk, network, latency, or cost.