
The stakes are real for enterprise teams. An undersized family causes CPU, memory, disk, or network bottlenecks. An oversized one burns budget: Flexera's 2026 survey of more than 750 cloud decision-makers estimated 29% cloud waste across clouds, not just Azure.
This guide walks through a practical process: learn the main families, baseline your workload, compare technical needs and pricing, then validate and revisit the choice with Microsoft's own tooling.
Key Takeaways
- B-series suits burstable workloads; D-series is the usual starting point for balanced applications.
- E and M favor memory, F favors CPU, L favors storage, N and H cover GPU and HPC.
- vCPU and RAM are only part of the picture; check disk, network, quota, region, and generation.
- Choose from measured behavior, then review continuously.
What Is an Azure VM Family?
A VM family (also called a series) is a group of predefined configurations built around a particular CPU-to-memory, storage, networking, or accelerator profile.
Microsoft defines a series as a group of sizes with similar hardware and features, and a size as a specific configuration of vCPUs, memory, and accelerators. An instance is the VM you actually deploy.
In short: the family sets the resource profile, the size sets the scale within it.
| Family | Category | Typical purpose |
|---|---|---|
| B | General purpose | Burstable baseline workloads |
| D | General purpose | Balanced production workloads |
| F | Compute optimized | CPU-intensive processing |
| E | Memory optimized | Memory-intensive workloads |
| M | Memory optimized | Very large memory workloads |
| L | Storage optimized | Low latency, high throughput, local storage |
| N | GPU accelerated | GPU compute, AI, graphics |
| H | HPC | Compute- and memory-bandwidth-sensitive jobs |
These families span multiple generations, and retirement is generation-specific rather than family-wide. Bv1 retires November 15, 2028, and Dv3/DSv3 on November 15, 2029. Check Microsoft's lifecycle pages before standardizing on any generation.
Core components of an Azure VM family
CPU. vCPU count, generation, and CPU-to-memory ratio shape concurrency, throughput, response time, and licensing. Ratios aren't fixed per family: the F family spans 2, 4, and 8 GiB per vCPU across its v7 series.
Memory. RAM capacity matters most for databases, caches, analytics, and in-memory applications, especially when working sets grow.
Storage and network. Disk limits are set at both the VM and disk level. Microsoft's own example: three P30 disks supply 15,000 aggregate IOPS, but a D8s_v3 caps uncached IOPS at 12,800. The VM throttles before the disks do.
Network is similar. A D4ds_v5 has an expected maximum outbound of 12,500 Mbps, which isn't guaranteed throughput.
Decoding a VM name
Take Standard_D4ds_v5:
- D: family
- 4: vCPUs
- d: local temporary disk
- s: Premium SSD compatibility
- v5: generation
Other letters include a (AMD), p (Arm), l (low memory), and r (RDMA-capable network). Don't assume a suffix means the same thing everywhere; confirm in the size's own documentation.
Why families matter for enterprise workloads
The right family shows up in outcomes you can measure: latency, query performance, batch completion time, storage response, and total cost of ownership. Choosing a family first narrows the field before you compare exact sizes, disk tiers, pricing models, and regional availability.

What to Consider When Choosing the Best Azure VM Series
Treat selection as a bridge between technical requirements and business outcomes. Validate specs, quotas, pricing, and regional availability against Microsoft sources before committing.
Workload behavior and baseline
Capture normal, peak, and seasonal CPU, memory, disk, network, latency, and throughput. Without that, you'll either throttle or pay for idle capacity.
Don't copy the on-premises server spec. Use Azure Monitor, application telemetry, and database monitoring. Azure Monitor collects host CPU, disk, and network metrics automatically; VM Insights adds guest memory and disk views.
Azure Advisor gives a useful benchmark. Its default lookback is 7 days (configurable up to 90), and it targets:
- User-facing workloads: P95 CPU and outbound network at or below 40%, P99 memory at or below 60%
- Non-user-facing workloads: all three at or below 80%
CPU-to-memory ratio and application profile
Decide whether the workload is CPU-bound, memory-bound, or balanced. Watch CPU utilization, memory pressure, cache hit rate, response time, and throughput.
- D-series: general-purpose application tiers
- F-series: sustained CPU-heavy processing
- E- or M-series: memory-intensive databases
For SQL Server, Microsoft recommends Ebdsv5 for the best price-performance in most production cases, with E and M options for heavier memory needs. Test before you commit.
Licensing can tip the balance. Constrained vCPU sizes keep the parent's memory, storage, and I/O while reducing active vCPUs. Microsoft's M64-16ms example carries one-quarter the SQL Server licensing cost of M64ms, with compute and OS costs unchanged.

Storage performance and disk requirements
Compare VM-level disk count, IOPS, throughput, latency, and temporary storage, not just capacity. These drive database latency, transaction throughput, and batch duration.
Two paths exist:
- L-series: Lsv3 offers directly attached NVMe (1 x 1.92 TB per 8 vCPUs) for throughput and latency, but it isn't durable storage.
- Faster managed disk on another family: Ultra Disk or Premium SSD v2 are data-disk options with region, zone, and VM-size restrictions.
Test L-series when you need local NVMe speed and can tolerate non-durable data. Test a faster managed disk when you need persistence and the VM's limits leave headroom.
Network, region, availability, and quota
Bandwidth, latency needs, zone support, scale-set support, and quotas all affect whether a series is practical.
- Zone-supported US regions: Central US, East US, East US 2, South Central US, West US 2, and West US 3. Zone support in a region doesn't guarantee every SKU is available.
- vCPU quotas apply per subscription and region, with separate regional and VM-family limits. Default is 20 total cores and 20 per series per region, though defaults vary by offer.
- Internet egress in North America starts with 100 GB free per month, then $0.087/GB for the next 10 TB.
Choose a US region on latency, data residency, resiliency, and service availability, not compute price alone.
Pricing, licensing, and purchasing model
Family and size set the compute rate, but the bill also includes OS licensing, managed disks, monitoring, backup, and outbound transfer.
| Model | Microsoft's stated savings |
|---|---|
| Reserved VM Instances | 36%-72% vs. pay-as-you-go, 1 or 3 years |
| Compute savings plan | 11%-65%, hourly commitment billed regardless of use |
| Spot VMs | Up to 90%, but eviction is possible |
| Hybrid Benefit + reservation | Up to 80% combined |
These are maximums tied to specific SKUs and conditions. Use the Azure Pricing Calculator for real numbers. Note that Hybrid Benefit requires at least eight core licenses per VM. A low hourly rate doesn't mean the lowest total cost if performance, availability, or commitment terms don't fit.

Growth, scaling, and continuous rightsizing
Expected growth and volatility should shape your first choice. Resizing works up or down, but a running VM restarts, and some targets need deallocation. A SCSI-based VM can't directly resize to a remote-NVMe size.
VM Scale Sets add or remove instances by metric or schedule. After deployment, review Azure Advisor and telemetry, test alternative families, and reassess when utilization, architecture, or demand shifts.
How Lucidity Can Help
Once you've settled on a compute family, the storage layer still needs attention. Lucidity is an autonomous cloud block storage optimization platform for enterprise ITOps, DevOps, and FinOps teams. It manages the block storage attached to your VMs, while VM family selection stays with you.
What it offers for Azure environments:
- Real-time visibility into block storage volumes
- AutoScaler that expands, shrinks, and optimizes storage with zero downtime and no code changes
- Lumen that finds unattached, reserved, unmounted, and zero-I/O disks, then analyzes IOPS, throughput, latency, and cost
- Reduced manual provisioning and ticket triage
- SOC 2 compliance
Lucidity's Assessment scans an Azure environment in about 5 minutes with no installation. Across 600+ assessments, average enterprise disk utilization was 30%. Run it alongside your VM rightsizing review to see how much disk capacity sits idle.

Conclusion
The right Azure VM series starts with workload evidence, not popularity, a single vCPU target, or a default like D-series.
Balance CPU, memory, storage, network, availability, regional limits, licensing, and purchasing model against real performance and business needs.
After you deploy, keep validating the fit:
- Monitor the deployed workload
- Validate Azure Advisor recommendations
- Review storage and compute utilization
- Revisit the family as demand and architecture change
Frequently Asked Questions
How much does Azure VM cost per month?
It depends on family, size, operating system, US region, runtime, storage, networking, licensing, and purchasing model. Use the Azure Pricing Calculator or Microsoft's ratecards for current figures.
Which VM is free in Azure?
Microsoft's Learn page lists B1s, B2pts v2 (Arm), and B2ats v2 (AMD), up to 750 hours per month each for 12 months. The Windows pricing page lists B2ts v2 instead of B1s, so confirm on Microsoft's free services page. It applies to new customers.
How many VMs can I run with 64GB RAM?
There's no universal number. It depends on each VM's assigned memory, workload needs, subscription and regional quotas, and capacity. Dividing RAM evenly isn't a production sizing method.
What is the difference between an Azure VM family, series, and size?
The family is the workload category, the series is the specific hardware variant, and the size defines vCPUs, RAM, storage, and network limits.
Which Azure VM series is best for SQL Server or other memory-intensive databases?
E-series (such as Ebdsv5) or M-series often fit. Storage latency, IOPS, throughput, licensing, and database telemetry should drive the final choice, confirmed by testing.
How do I know when to resize an Azure VM?
Look for sustained CPU or memory pressure, disk or network throttling, rising latency, or persistent underutilization. Check Azure Advisor recommendations, then validate against real workload behavior.