What general purpose VMs are not capacity constrained for azure machine learning managed online endpoints?

Question

What general purpose VMs are not capacity constrained for azure machine learning managed online endpoints?

Avery Jordan 20

My team uses Azure Machine Learning Managed Online Endpoints to serve inferences for a variety of ML workloads. We have always selected our VM SKUs from this list https://learn.microsoft.com/en-us/azure/machine-learning/reference-managed-online-endpoints-vm-sku-list?view=azureml-api-2, and use a mixture of Dsv2 and Dasv4 VMs, both of which are general purpose.

Recently there have been several announcements about capacity constraints on older VM types, including the Dsv2 and Dasv4 MV SKUs, as outlined here: https://learn.microsoft.com/en-us/azure/virtual-machines/migration/sizes/previous-gen-series-capacity-limitations.

This capacity constraint list overlaps with all of the general purpose VMs that are listed for the AML online endpoints. I want to know what type of VM I should migrate my workloads to for AML online endpoints, and preferably that is a general purpose VM of a newer generation.

0 comments

Answer accepted by question author

0 additional answers

Your answer

Answer 1

Allan Solomon Mejia 3,755

Hello @Avery Jordan

The VM SKU list for Azure Machine Learning Managed Online Endpoints indicates which VM families are supported by the service, but it doesn't guarantee capacity availability in every region. With the retirement and capacity constraints affecting older generations (such as Dsv2 and Dasv4), it's a good idea to plan a migration to newer VM families where possible.

For general-purpose inference workloads, consider evaluating newer VM series such as:

Dsv5 / Dasv5 – Current-generation general-purpose VMs and the recommended successor to Dsv2/Dasv4 for many workloads.
Ddsv5 – If your inference workload benefits from local NVMe temporary storage.
Dplsv6 or other newer v6-series VMs (where supported by Azure ML in your target region).

Before migrating, I recommend verifying:

That the target VM SKU is supported for Managed Online Endpoints in your region.
Regional quota availability for the new SKU.
Any performance differences (CPU generation, memory bandwidth, and storage) by running a small-scale benchmark before switching production traffic.

If capacity remains a concern, you may also want to deploy across multiple regions or configure multiple deployments behind the same endpoint to improve resilience against regional capacity shortages.

Could you share:

Which Azure region you're deploying to?
Whether your endpoint uses CPU-only inference or requires GPU acceleration?
The approximate VM size you're using today (for example, Standard_D4s_v2 or Standard_D8as_v4)?

That will help identify the closest supported replacement with the least operational impact.

Please "Accept the Answer" if this information helped you. This will help us and others in the community as well.

Avery Jordan 20 Reputation points

2026-08-05T22:54:33.87+00:00

Hi there, thanks for your response. I am deploying in West-US-2, I only need CPU inference, and I am typically using the D8as_v4 or the D4s_v2 sizes. Sometimes we also use the D3s_v2 or the D4as_v4 size. I think from reviewing the list, the Dadsv5 could be a good replacement for us. I don't think there are any v6 series supporting in my target region at the time (as far as I can tell).

How can I verify regional quota availability for the new SKUs?
Allan Solomon Mejia 3,755 Reputation points

2026-08-05T23:22:56.1833333+00:00

Hi @Avery Jordan

Since you're deploying in West US 2 and using CPU-only Managed Online Endpoints, Dasv5 is a reasonable candidate to evaluate if it's available in your region and supported by Azure Machine Learning.

To check regional quota availability, you have a few options:

1. Azure portal

Go to Subscriptions → Usage + quotas.

Filter by West US 2 and the DASv5 VM family (or the applicable vCPU family).

This shows your current quota, usage, and whether additional quota is available. You can also submit a quota increase request directly from this page if needed.

2. Azure Quotas REST API or Azure CLI

If you automate deployments, you can query your vCPU quota programmatically using the Azure Quotas API or the Azure CLI.

3. Test deployment

Even if quota is sufficient, it's worth deploying a small test endpoint with the target SKU. This confirms both SKU availability and Azure ML Managed Online Endpoint support in West US 2 before migrating production workloads.

Since you're moving from D3s_v2/D4s_v2 to Dasv5, I'd also recommend running a small inference benchmark to compare latency and throughput before switching production traffic.

Please don't forget to "Accept the Answer" if this information helped you. This will help us and others in the community as well.