
Plenty of teams provision for the ceiling, attach the disk to a VM that tops out at 180,000 uncached IOPS, and pay for performance they never see. Others benchmark once, treat the result as a guarantee, and get surprised at peak load.
This article covers the current limits, the rules that govern them, how to measure real performance, and the misreadings that cost money. Verify every figure against Microsoft's current documentation before sizing, because these specs have changed recently.
Key Takeaways
- Ultra Disk allows up to 400,000 IOPS and 10,000 MB/s per disk.
- The VM's uncached limits often bind first.
- IOPS, throughput, latency, block size and queue depth must be judged together.
- Capacity, IOPS and throughput are provisioned separately, and each is billed.
- Keep headroom below the ceiling in production.
What Azure Ultra Disk Max IOPS Represents in Azure Storage
IOPS is input/output operations per second. Microsoft defines it as the number of I/O requests issued to the disk each second. Throughput is the data moved per second, and latency is the time to complete one request.
Max IOPS is a service-level ceiling for one managed disk. It is one ceiling along the path from your application's requests through the disk, the attached VM, and the storage stack. It is not a measured result, and it says nothing about how much data moves per second.
The relationship is simple: IOPS × I/O size = throughput. Small I/Os favor IOPS. Large I/Os favor throughput.
Ultra Disk does not burst. The provisioned value is the sustained ceiling for the disk, and Ultra falls outside Microsoft's supported bursting types. Exceeding disk or VM limits causes capping and can raise latency.
Factors That Influence Max IOPS in Real-World Operation
A published limit describes supported conditions, not your production workload. Rated values come from documentation. Measured values come from your own testing.
Platform constraints:
- Ultra Disk works only as a data disk. No OS-disk use, host caching, availability sets, or ZRS.
- Supported infrastructure is single VM or availability zone, using LRS.
- Not every VM size, region, and zone combination is supported. Check with
az vm list-skusorGet-AzComputeResourceSku. - NVMe-based sizes require Generation 2 VMs.
Workload constraints:
- Request size, random versus sequential access, and read/write mix.
- Queue depth and concurrency. An application that can't issue enough parallel requests never reaches the ceiling.
- CPU saturation, driver configuration, and database settings.
Range of Azure Ultra Disk Max IOPS
The effective range combines three things: the disk's documented maximum, the IOPS value you provision, and the lower limit imposed by the VM or workload.
Nominal Operating Range
Per Microsoft's disk-types documentation (updated September 2026), the current rules are:
| Property | Rule |
|---|---|
| Capacity | 4 GiB to 64 TiB, in 1-GiB increments |
| Minimum IOPS | 100 per disk |
| Maximum IOPS | 1,000 IOPS per GiB, capped at 400,000 |
| Minimum throughput | 1 MB/s |
| Maximum throughput | 0.25 MB/s per provisioned IOPS, capped at 10,000 MB/s |

The density rose from 300 to 1,000 IOPS per GiB in Microsoft's November 2025 announcement. Older articles and some Microsoft pages, including shared-disk documentation, still show the 300 figure. Don't mix the two.
Per-region quotas are separate: 100 TiB of capacity and 1,000 Ultra disks per subscription. These are not IOPS limits.
Allowable Tolerance and Boundary Limits
The effective ceiling is the most restrictive applicable limit. Compare two memory-optimized VMs:
| VM size | Controller | Uncached IOPS | Uncached throughput |
|---|---|---|---|
| Standard_E96bds_v5 | SCSI | 180,000 | 6,500 MB/s |
| Standard_E112ibds_v5 | NVMe | 400,000 | 10,000 MB/s |
These are VM aggregates across all attached disks, not per-disk values. On the first VM, a 400,000 IOPS disk delivers at most 180,000.

The provisioned value is a configurable ceiling, not a burst allowance or a guarantee of observed performance.
Safe Operating Margin
Microsoft's Well-Architected guidance gives no universal headroom percentage. It says to monitor throttling on disk and VM and raise Ultra performance as demand grows. Pick your own margin based on observed peaks.
Running near the limit invites:
- Throttling and rising latency
- Growing queues
- Unpredictable transaction times
- Little room for backups, failover, maintenance or seasonal spikes
Key Technical Properties and How Max IOPS Is Specified, Measured, and Validated
Ultra Disk performance comes from IOPS, throughput, latency, request size, queue depth, and concurrency working together. IOPS alone tells you little.
Independent Performance Provisioning
Ultra Disk treats capacity, IOPS, and throughput as separate dimensions. Hourly billing covers all three, regardless of how much data or how many I/Os the workload actually produces. Each dimension still has minimums, maximums, and coupling rules, such as the 1,000 IOPS per GiB density.
Over-provisioning raises the bill without improving the application. Microsoft's pricing page and scalability page still show tier-rounding examples (200 GiB billed as 256 GiB), while the newer billing article describes per-GiB billing. Confirm the current rule before building a cost model.
Workload Variability and Adjacent Limits
One disk produces different numbers under different workloads:
- 4 KB random reads at deep queue depth: high IOPS, modest throughput
- 64 KB sequential I/O: lower IOPS, higher throughput
- Read/write mix and concurrency: shift the balance again
Disk IOPS also interacts with VM IOPS caps, VM throughput caps, database limits, and host configuration. Ultra Disk has no host caching, so cache-hit results from other disk types don't predict Ultra behavior.
Specification, Measurement, and Validation
Keep rated, provisioned, and tested numbers apart. Cross-check:
- Disk types and scalability docs for per-disk limits
- VM size pages for uncached IOPS and throughput
- The Ultra configuration guide for regions and compatible series
- Pricing and billing articles for cost
Use Azure Monitor with the exact metric names:
- Data Disk IOPS Consumed Percentage and Data Disk Bandwidth Consumed Percentage (per LUN). At 100%, the disk caps the application.
- VM Uncached IOPS Consumed Percentage and VM Uncached Bandwidth Consumed Percentage. At 100%, the VM is the bottleneck.
- Data Disk Queue Depth and Data Disk Latency, which help spot saturation and lag.
Compare provisioned against consumed values across normal and peak periods. Test with production-like request sizes and read/write patterns. Microsoft's benchmark guide uses 4 KB random reads with high queue depth for IOPS and 64 KB blocks for throughput. Those are methodology examples, not Ultra-specific results.

For large estates, manual checks stop scaling. Lucidity's platform monitors every volume across AWS, Azure, and Google Cloud, giving continuous visibility into utilization and provisioning. It doesn't change Azure's underlying limits.
Implications of Operating Outside the Recommended Range and Common Misinterpretations
When demand exceeds a disk or VM limit, requests queue, latency rises, operations get throttled, and application behavior turns erratic.
Those same limits drive costly sizing mistakes:
- Paying for IOPS the VM can't consume
- Buying a larger VM than necessary
- Skipping headroom for peaks
Microsoft's enterprise example is light on hard numbers. BlackRock's Aladdin platform runs on Ultra Disk with M-series VMs and reportedly handled record volumes, but the case study discloses no IOPS, latency, block-size, or queue-depth figures.
Common misreadings compound the risk:
- Confusing IOPS with throughput
- Assuming 400,000 IOPS applies to every disk
- Ignoring VM-level limits
- Treating one test as a service guarantee
- Using preview-era or pre-November 2025 figures
Conclusion
Max IOPS is a hard ceiling on what the disk can deliver. Sizing reliably means checking the disk limit, VM limit, provisioned IOPS, throughput, latency, and workload behavior together. Then measure, and leave headroom. A resilient, cost-conscious design depends on observed utilization as much as on the published spec.
Frequently Asked Questions
What is the maximum IOPS for an Azure Ultra Disk?
Currently 400,000 IOPS per disk, at up to 1,000 IOPS per GiB. VM uncached limits can be lower, and Microsoft can change specs, so verify before sizing.
What is Azure IOPS?
IOPS is input/output operations per second. It counts how often operations occur, not how much data each one moves.
What is disk throughput?
Throughput is data transferred per second. It equals IOPS times I/O size, so high IOPS with small blocks can still mean low throughput.
What is the difference between Azure Ultra Disk IOPS and VM IOPS limits?
The disk has its own provisioned and maximum IOPS. The VM adds per-disk and aggregate uncached limits that can cap effective performance below the disk's setting.
Can you change Azure Ultra Disk IOPS without downtime?
Yes. Change IOPS and throughput while attached, with no VM restart, under Size + performance in the portal. Microsoft allows four changes per 24 hours, and each can take up to an hour to apply.
How do you monitor Azure Ultra Disk IOPS?
Use Azure Monitor's disk and VM consumed-percentage metrics, plus queue depth and latency. Compare provisioned against consumed IOPS over normal and peak periods.