Azure GPU VM Pricing: Sizes and Costs Explained

Introduction

Azure GPU virtual machines give you on-demand access to accelerated compute for AI training, inference, rendering, visualization, and high-performance workloads. The catch: there is no single price.

Cost shifts with GPU model, VRAM, GPU count, vCPU and memory, region, operating system, pricing model, and runtime.

US Linux pay-as-you-go rates run from $0.333/hour for a fractional graphics VM to $98.32/hour for an eight-GPU H100 node. That is a 295x spread.

This guide covers US price bands, NC, ND, and NV sizes, hidden cost components, the NC24ads_A100_v4 and H100 options, and a method for building a realistic monthly budget.

Key Takeaways

  • US Linux rates range from about $0.33/hour (fractional GPU) to $98.32/hour (eight H100s).
  • GPU model, GPU count, region, runtime, commitment type, storage, and networking drive the bill.
  • Cheap sizes suit development, visualization, and interruptible batch jobs; ND nodes suit large training.
  • Match the SKU to workload fit and utilization, not to the most powerful GPU.

How Much Does an Azure GPU VM Cost? (Pricing Overview)

Azure GPU VMs have no fixed price. Check any figure here against the Azure Pricing Calculator or Retail Prices API for your region, SKU, operating system, and billing model.

Typical Cost Range

All figures below are US pay-as-you-go list prices as of November 18, 2026, in USD. Monthly equivalents use 730 hours and cover compute only.

  • Entry-level: fractional or older GPUs, such as NV4ads_V710_v5 ($0.333/hour, about $243/month) and NC4as_T4_v3 ($0.526/hour, about $384/month). Good for development, visualization, testing, and light inference.
  • Mid-range: single-GPU production VMs. NC24ads_A100_v4 costs $3.673/hour (about $2,681/month). NC40ads_H100_v5 costs $6.98/hour (about $5,095/month).
  • High-end: eight-GPU ND nodes. ND96isr_H100_v5 costs $98.32/hour (about $71,774/month).

Azure GPU VM pricing tiers from entry-level to high-end nodes

Representative Azure GPU VM Prices

Linux prices, East US unless noted (ND rows use South Central or East US 2).

Size GPU GPUs VRAM (GB) vCPU / RAM (GiB) $/VM-hour $/GPU-hour $/730 hrs
NV6ads_A10_v5 A10 1/6 4 6 / 55 0.454 2.724* 331
NC4as_T4_v3 T4 1 16 4 / 28 0.526 0.526 384
NV36ads_A10_v5 A10 1 24 36 / 440 3.20 3.20 2,336
NC24ads_A100_v4 A100 PCIe 1 80 24 / 220 3.673 3.673 2,681
NC40ads_H100_v5 H100 NVL 1 94 40 / 320 6.98 6.98 5,095
NC96ads_A100_v4 A100 PCIe 4 320 96 / 880 14.692 3.673 10,725
ND96amsr_A100_v4 A100 8 8x80 96 / 1900 32.77 4.096 23,922
ND96isr_H200_v5 H200 8 8x141 96 / 1850 84.80 10.60 61,904
ND96isr_H100_v5 H100 8 8x80 96 / 1900 98.32 12.29 71,774

*Fractional rows show a full-GPU-equivalent allocation price, not measured performance.

A multi-GPU hourly price is the whole node, not one GPU. And none of these numbers include disks, networking, or licensing.

Pricing Models

Three billing options matter for GPU work:

  • Pay-as-you-go: no commitment, full flexibility, highest rate.
  • Reservations and savings plans: one- or three-year commitments. For NC24ads_A100_v4, rates drop from $3.673 to $3.0574/hour (1-year) or $2.2659/hour (3-year), billed every hour whether the VM runs or not.
  • Spot: NC24ads_A100_v4 Spot is $0.6788/hour. Spot has no SLA and roughly 30 seconds' eviction notice. Discounts vary by SKU; the ND H200 Spot meter matched pay-as-you-go as of November 18, 2026.

Azure GPU VM pricing model comparison for pay-as-you-go reservations and Spot

Microsoft advertises savings of up to 90% on Spot, but that is a ceiling, not a promise for every SKU.

Monthly budget formula:

  1. Multiply the hourly VM rate by expected billed hours.
  2. Add storage, networking, licensing, and monitoring.
  3. Add any managed-service or orchestration costs.

Azure GPU VM Sizes: NC, ND, and NV Explained

Azure GPU VMs live in the N-family. Generations change, so confirm the current list of active series on Microsoft Learn before you commit. NVv3 retires on September 30, 2026, and NVv4 is on the end-of-life list.

How to Read an Azure GPU VM Name

Take NC24ads_A100_v4:

  • NC: the series (compute-oriented)
  • 24: vCPU count
  • a: AMD processor
  • d: local temporary disk
  • s: Premium Storage capable
  • A100: the GPU model
  • v4: the generation

Letters like i (isolated) and r (RDMA/InfiniBand) appear on larger sizes. None of these letters tell you GPU count or VRAM, so read the spec sheet.

NC Series

NC targets machine learning, fine-tuning, inference, and applied AI. Current options include:

  • NC A100 v4: A100 PCIe 80GB, one to four GPUs
  • NC H100 v5: H100 NVL 94GB, one or two GPUs
  • NCas T4 v3: 16GB T4, one to four GPUs, the budget choice

A single-GPU NC VM usually wins when the model fits in one GPU's VRAM, scaling is modest, and budget matters. Moving to a multi-GPU node only helps if your software actually uses the extra GPUs. Otherwise you pay double for idle silicon.

ND Series

ND is built for large-scale and distributed training. The premium buys interconnect, not just GPU count:

  • ND A100 v4: eight A100s, NVLink 3.0 inside the VM, HDR InfiniBand between VMs
  • ND H100 v5: NVLink 4.0, 400Gbps InfiniBand per GPU (3.2Tbps per VM)
  • ND H200 v5: eight H200s, 900GB/s NVLink

Azure ND A100 H100 and H200 GPU infrastructure comparison

Judge the whole node: expected utilization, scaling efficiency, and checkpointing. Comparing only the GPU name misses the point.

NV Series

NV (and the related NG) suits virtual desktops, remote workstations, rendering, and light inference. Key options and licensing notes:

  • Fractional A10 sizes: 4GB to 12GB of VRAM at $0.454 to $1.60/hour
  • A10 v5 sizes: include a GRID license (one virtual workstation or 25 concurrent application users)
  • V710 sizes: no additional GPU licensing

For graphics-heavy but small workloads, NV often beats paying for an NC or ND card.

Match GPU Size to the Workload

Workload Practical starting point Key criteria
Development, testing NC T4 or fractional NV Low cost, Spot-tolerant
Fine-tuning NC A100 / NC H100 VRAM fit
Batch inference NC on Spot Interruption tolerance
Production inference NC A100 / H100, reserved Latency, concurrency, capacity
Large-model training ND A100/H100/H200 Interconnect, checkpointing
VDI, visualization, rendering NV / NG Users per GPU, licensing

Key Factors That Affect Azure GPU VM Cost and Full Cost Breakdown

The hourly GPU rate is one line on the bill. Configuration, usage habits, and supporting services decide the rest.

GPU Model, VRAM, and VM Configuration

GPU generation, VRAM, GPU count, vCPUs, RAM, local storage, and interconnect all move the price. Size the VM to the memory and throughput you actually need, then compare hourly rates. A 94GB H100 NVL costs about 90% more per hour than an 80GB A100 ($6.98 vs $3.673), which only pays off if your job is faster or needs the memory.

Region, Operating System, and Licensing

Region meters vary, but capacity, latency, and compliance often override the cheapest option.

Operating system matters too. On NC24ads_A100_v4, Windows adds $1.104/hour, about $806/month. That is an observed SKU difference, not a universal surcharge. Azure Hybrid Benefit can offset Windows licensing if you hold eligible Software Assurance.

Azure Machine Learning adds no extra platform fee, though you still pay for the underlying compute and related resources.

Runtime, Utilization, and Billing State

Running and stopped-but-allocated VMs both bill for compute. Only a deallocated VM stops instance charges. Disks and some networking resources keep billing.

Waste hides here:

  • Long-running notebooks left allocated overnight
  • Multi-GPU nodes used for single-GPU jobs
  • Spot interruptions that restart work without checkpoints

Flexera's 2025 report found respondents estimate 27% of IaaS and PaaS spend is wasted. That is a general cloud figure, not a GPU one, but it fits the pattern.

Storage, Networking, and Supporting Services

GPU jobs move a lot of data. Budget for these East US meters:

  • Premium SSD P10: $19.71/disk-month
  • Standard SSD E10: $9.60/disk-month
  • Blob Hot LRS: $0.0208/GB-month
  • Inter-region transfer: $0.02/GB
  • Internet egress: first 100GB free, then $0.087/GB
  • Standard public IPv4: $0.005/hour

Azure GPU workload supporting cost meters for storage and networking

Datasets, checkpoints, model weights, and snapshots add up. Disks also keep billing after a VM is deallocated.

This is where Lucidity fits. Its autonomous block storage platform gives visibility into Azure block storage and helps right-size over-provisioned or idle disks attached to GPU workloads.

Lucidity's research indicates unattached, reserved, unmounted, and zero-I/O disks can represent up to 70% of unused block-storage spend. It does not lower the GPU compute rate, only the storage around it.

Cost Breakdown

Category Billed by Stops when deallocated?
Recurring compute Runtime Yes
Managed disks, snapshots Provisioned capacity No
Blob storage, data transfer Data volume No
Software and licensing Subscription or runtime Varies
Deployment engineering One-time effort n/a
Monitoring and operations Subscription or usage No
Resizing, migration, replacement Periodic decisions n/a

Reservations and savings plans are commitments. You pay them regardless of whether the VM runs.

Low-Cost vs High-Cost Azure GPU VMs: How to Choose and Estimate the Right Budget

A cheaper hourly rate can cost more overall. If a $0.53/hour T4 takes ten times longer than an A100 at $3.673/hour, you lose on both time and money. Conversely, an eight-GPU node at 30% utilization wastes most of what you pay.

Azure GPU VM cost versus performance tradeoff between T4 and A100

Match the SKU to the job:

  • Lower-cost sizes: experimentation, graphics, and interruptible batch work
  • Higher-cost nodes: large training, high-memory inference, and distributed jobs where interconnect and VRAM are the bottleneck

Budgeting checklist:

  • Workload type and model size
  • VRAM requirement and concurrency
  • Expected billed runtime per month
  • Region and operating system
  • Storage footprint and data movement
  • Availability needs and quota
  • Whether Spot or committed pricing is viable

What Most People Miss on Azure GPU VM Cost

  • Comparing per-GPU prices against per-node prices
  • Forgetting persistent storage and egress
  • Leaving allocated VMs running while idle
  • Picking a multi-GPU node for a single-GPU job
  • Ignoring quota (Azure ML defaults to 0 GPU; a listed meter does not guarantee capacity)
  • Using stale price snapshots

Validate every estimate in the current calculator or Retail Prices API. Set budgets and alerts, but remember alerts do not stop spending. Test your chosen SKU, track utilization, and revisit the choice as requirements change.

Conclusion

Azure GPU VM costs depend on GPU model, size, region, pricing model, runtime, and the resources around the VM. The right budget balances performance, VRAM, availability, utilization, and total cost of ownership, not just the advertised hourly rate.

Frequently Asked Questions

What is the pricing for Azure GPU instances?

US Linux pay-as-you-go rates run from $0.333/hour to $98.32/hour, depending on series, SKU, region, OS, and pricing model. That is compute only; storage, networking, and licensing come on top.

What is the current pricing for Azure GPU H100?

NC40ads_H100_v5 (one 94GB H100 NVL) is $6.98/hour pay-as-you-go in East US. ND96isr_H100_v5 (eight H100s) is $98.32/hour per VM, or $12.29 per GPU. Spot pricing for that SKU was $18.17/hour.

What is the cost of the NC24ads_A100_v4 instance?

In East US it costs $3.673/hour on Linux (about $2,681 per 730 hours) and $4.777/hour on Windows. Disks, networking, and region changes alter the final bill.

Which Azure VMs have GPU?

The N-family: NC for compute and ML, ND for large-scale training, and NV/NG for visualization and VDI. Microsoft also lists newer series such as ND MI300X v5 and ND GB200 v6; check availability at purchase time.

What are the different sizes of Azure VMs?

Sizes vary by vCPU, memory, GPU count, VRAM, storage, networking, and generation. For GPU workloads, pick from the NC, ND, or NV series based on your VRAM and scaling needs.