
The trap starts with names. A "DS3" is not a 3-vCPU machine, and the VM rate is only one line on the bill. Flexera's 2025 report estimates 27% of cloud spend is wasted.
This guide covers decoding D-series names, estimating total cost in US Azure regions, comparing alternatives, and validating sizing with telemetry.
Key Takeaways
- D-series is general-purpose; variants and generations change processor, storage support, and price-performance.
- The number in a size name doesn't always equal vCPUs, so check the spec table.
- Compute rate is one part of cost; disks, licensing, networking, and backup add more.
- Right-sizing is ongoing and should rest on CPU, memory, disk, and network data.
What Are Azure D-Series VMs?
D-series is Microsoft's general-purpose VM family. Documented workloads include enterprise applications, relational databases, in-memory caching, and analytics. Think web apps, APIs, application servers, dev environments, and moderately sized databases.
Compute and storage are separate. Picking a D-series size doesn't set your disk tier. Persistent managed disks are billed as their own resources, while included local temporary disks carry no separate storage charge.
Generations and variants
- Dsv5: Intel processors
- Dsv6: fifth-generation Intel Xeon
- Dasv5 / Dasv6: AMD (Dasv6 uses fourth-generation EPYC)
- Dpdsv6: Azure Cobalt 100 Arm64
Availability changes, so confirm current specs in Microsoft's documentation before deploying.
When to move away from D-series
- Sustained memory pressure points to memory-optimized families.
- CPU-bound processing may fit compute-optimized sizes better.
- Mostly idle, bursty workloads may suit burstable VMs.
- GPU needs or very large local storage requirements call for other families.
How to Decode Azure D-Series VM Names and Specifications
Microsoft's naming convention has these fields: family, optional subfamily, vCPU count, optional constrained-vCPU count, feature letters, and version.
Feature letters:
- a: AMD processor
- p: Arm processor
- s: Premium Storage capable
- d: local temporary disk
- l: low memory
Worked example: Standard_DS3_v2
Standard_DS3_v2 breaks down as Standard tier, D family, size 3, s for Premium Storage, and version 2. It has 4 vCPUs and 14 GiB of memory, plus 28 GiB local temp storage. Compare that with D2s_v5, which has 2 vCPUs.
Its other limits:
- 16 remote disks
- Up to 12,800 uncached IOPS and 192 MB/s on Premium SSD
- 4 NICs and 3,000 Mbps aggregate bandwidth (an upper limit, not a guarantee)
Never infer resources from the numeral. Check Microsoft's size page.
Reference table
| Size | vCPUs | RAM (GiB) | Local temp (GiB) | Premium Storage | Version |
|---|---|---|---|---|---|
| Standard_DS3_v2 | 4 | 14 | 28 | Yes | v2 |
| Standard_D2s_v3 | 2 | 8 | 16 | Yes | v3 |
| Standard_D2ds_v4 | 2 | 8 | 75 | Yes | v4 |
| Standard_D2s_v5 | 2 | 8 | None | Yes | v5 |
| Standard_D2as_v5 | 2 | 8 | None | Yes | v5 |
| Standard_D2ads_v5 | 2 | 8 | 75 | Yes | v5 |
| Standard_D2pds_v6 | 2 | 8 | 110 | Yes | v6 |
Older generations and retirement
| Series | Status |
|---|---|
| Dv2/Dsv2 | Retires May 1, 2028; capacity-growth restrictions may begin July 2026 |
| Dv3/Dsv3 | Retires November 15, 2029 |
| Dv4/Dsv4 | No retirement date stated on the cited spec page |
From July 1, 2026, purchase and renewal restrictions cover one-year Dv2/Dsv2 and one- or three-year Dv3/Dsv3 reservations. Existing reservations stay valid for their original terms, and discounts don't transfer automatically across series. Check Microsoft's notices at the time you buy.

Azure D-Series VM Cost Guide
The core formula: verified hourly rate × billable runtime. Pay-as-you-go compute bills per second. Run it separately for Linux and Windows, since Windows Server adds licensing unless Azure Hybrid Benefit applies.
For current prices, use the Azure Pricing Calculator or the Retail Prices API, which exposes USD prices by SKU, region, and price type. Record the date, currency, OS, region, and pricing model with every number. Agreement pricing can differ from public retail.
Costs beyond the VM rate
| Resource | Billing driver |
|---|---|
| Managed disks | Provisioned size and tier; transactions where applicable |
| Snapshots | Used data; incremental snapshots store changes |
| Bandwidth | Ingress free; first 100 GB/month egress free |
| Public IP | Static IPv4 bills until the resource is deleted |
| Azure Monitor | Log ingestion, retention, and alerts; platform metrics free |
| Azure Backup | Used data plus separate backup storage |
A deallocated VM stops compute billing, but disks and some networking charges continue. A stopped-but-allocated VM keeps billing.
Purchase models compared
| Model | Terms | Notes |
|---|---|---|
| Pay-As-You-Go | No commitment | Most flexible |
| Savings Plan for Compute | 1 or 3 years, hourly commitment | Advertised 11%-65% |
| Reserved VM Instances | 1 or 3 years | Advertised 36%-72%; compute only |
| Spot | No SLA | Up to 90%; 30-second eviction notice |
| Azure Hybrid Benefit | Needs eligible Windows licenses | Covers licensing, not base compute |
These maxima come from benchmark configurations, not every D size. Reservations also don't guarantee capacity.
A calculation framework
Estimate each piece separately, then add:
- Monthly compute: hourly rate × expected hours
- Attached disks: tier and size, plus snapshots
- Licensing: Windows or publisher charges, less Hybrid Benefit if eligible
- Supporting services: egress, public IPs, monitoring, backup
- Total, then apply any commitment discount to compute only

Cheapest hourly isn't cheapest overall
An undersized VM with too little memory can force scale-out. A disk bottleneck can waste CPU you're paying for. A commitment that no longer matches the workload locks in cost. Price total cost per unit of work, not per hour.
How to Choose and Validate the Right Azure D-Series Size
Profile the workload
Capture the metrics that drive size selection:
- Sustained and peak CPU
- Memory working set
- Disk IOPS and throughput
- Network traffic and latency targets
- Concurrency and uptime needs
Note whether demand is steady, seasonal, or bursty.
VM insights in Azure Monitor reports CPU, guest available memory, disk, and network metrics. Pair them with application telemetry and load tests. vCPU count alone never tells you where the bottleneck is.
Pick generation and variant
Weigh these factors against your workload profile:
- Processor compatibility
- Memory and premium disk needs
- Dependence on local temp disk
- Network throughput
- Availability in your target US region
Test a small range
Try two or three candidate sizes, record performance and cost, and keep only evidence-based headroom.
Azure Advisor's documented resize logic is a useful reference for headroom targets:
- User-facing workloads: P95 CPU and outbound network at or below 40%, P99 memory at or below 60%
- Non-user-facing workloads: those same limits rise to 80%

Treat these as Advisor criteria, not universal guarantees.
Pre-deployment checks
- Regional capacity: allocation can fail even with quota
- Quota: total regional vCPU and family-specific limits both apply
- Resize: a running VM restarts, and deallocation may be needed
- IPs: deallocation releases dynamic public IPs; static ones remain
- Temporary disk: data may be lost during maintenance, redeployment, or stopping, so keep nothing persistent there
Validate after launch
Compare predicted and observed CPU, memory, disk, network, latency, error rate, and monthly cost. Schedule a review whenever traffic or application behavior shifts.
D-Series Alternatives and Common Selection Mistakes
| Example | Class | vCPU:memory |
|---|---|---|
| D2s_v5 | General-purpose | 2:8 GiB |
| F2s_v2 | Compute-optimized | 2:4 GiB |
| E2s_v5 | Memory-optimized | 2:16 GiB |
| B2s_v2 | Burstable | 2:8 GiB |
These are examples, not family-wide ratios.
B-series: CPU use below baseline accumulates credits, and above-baseline use spends them. Exhausted credits throttle CPU back to baseline. Check credit limits before using B-series in production.
F-series: Less memory per vCPU suits compute-heavy processing but hurts memory-bound applications.
Common mistakes
- Choosing by vCPU count alone
- Reading the size number as vCPUs
- Ignoring disk and network limits
- Picking an old generation for a long-lived workload
- Storing persistent data on a temporary disk
Decision path
Start with D-series for balanced workloads. Reassess toward B-series for idle, bursty workloads, F-series for CPU-bound ones, and memory-optimized families when memory is the constraint.

How to Reduce and Manage D-Series VM Costs
Build a routine around a few controls:
- Review rightsizing on a schedule, using utilization and application performance
- Schedule auto-shutdown for dev/test VMs
- Tag resources so Cost Analysis can group spend
- Set budgets that send alerts (they don't stop consumption)
- Enable anomaly alerts, which use 60 days of history
- Use Azure Policy's allowed-SKU rule to restrict deployable sizes
Review disks separately
Deleting a VM doesn't necessarily delete its disks, and those keep billing. Filter for disks in the Unattached state, then check dependencies and recovery needs before removing anything.
Lucidity gives real-time, multi-cloud visibility into block storage and autonomously optimizes it without infrastructure changes.
Its Lumen product identifies four idle-disk types (unattached, reserved, unmounted, and zero-I/O), which together can represent up to 70% of unused block storage spend. Lucidity addresses storage waste, not VM compute rightsizing, so it complements the sizing work above.

Make it continuous
Engineering, ITOps, and FinOps should review performance, availability, utilization, and cost together. The FinOps Foundation's 2025 report ranks workload optimization and waste reduction as the leading priority. A one-time resize based on a billing report misses that picture.
Conclusion
The best D-series choice matches observed workload behavior, storage needs, performance goals, regional availability, and your pricing commitment. The right fit is rarely the cheapest or largest option.
Work through these steps in order:
- Decode the exact SKU
- Verify its official specs
- Calculate full monthly cost
- Test against telemetry
- Keep reviewing as demand changes
Frequently Asked Questions
What is D series in Azure?
D-series is Azure's general-purpose VM family, used for web apps, application servers, databases, and caching. Capabilities vary by generation and variant.
What is Standard_DS3_v2?
It's a D-family, Premium Storage-capable, version 2 size with 4 vCPUs, 14 GiB memory, and 28 GiB local temp disk. Get current pricing from the Azure Pricing Calculator.
What is the F series in Azure?
F-series is compute-optimized for CPU-bound workloads with modest memory needs. Compared with D-series, it offers less memory per vCPU.
What is the Azure B series?
B-series is burstable: it banks CPU credits when idle and spends them during spikes. Sustained demand that exhausts credits throttles CPU to baseline.
What are type 1 and type 2 hypervisors?
Type 1 hypervisors run directly on hardware, while Type 2 run on a host OS. Microsoft's hypervisor documentation describes Azure's Hyper-V-based hypervisor, which separates guest VMs from each other and the host.