Microsoft announced general availability of Azure Cobalt 200 VMs at Ignite, bringing its latest Arm-based custom CPU to customers running demanding cloud and agentic AI workloads. The chip is built on Arm Neoverse Compute Subsystems V3 (CSS V3). Microsoft reports up to 50% better CPU performance than Cobalt 100, and Arm says its own testing shows similar gains.
Where you can run it
The VMs are available in 14 Azure regions and are expected to expand to 23 by the end of the year. The 14 regions, from Microsoft's Azure blog:
- Australia East
- Canada Central
- Central US
- East US
- East US 2
- Indonesia Central
- Italy North
- Korea Central
- South Central US
- Spain Central
- Sweden Central
- UK South
- West US 2
- West US 3
Microsoft's materials are not fully consistent on the year-end target. A WindowsForum keynote roundup notes that the Book of News says 24 regions by year-end, while Microsoft's Ignite summary and Azure's Cobalt blog say 23. Treat 23 as the Azure blog figure. Check regional and size availability for your own subscription before you design anything around it.
VM families and sizes
Arm's blog names the Dpsv7, Dplsv7 and Epsv7 families, with configurations scaling to 128 vCPUs and 512 GiB of memory. Microsoft's own announcement is broader. It lists five series across four categories:
| Category | Series | Memory per vCPU | Max vCPUs | Max memory |
|---|---|---|---|---|
| General purpose | Dplsv7 / Dpldsv7 | 2 GiB | 128 | 256 GiB |
| General purpose | Dpsv7 / Dpdsv7 | 4 GiB | 128 | 512 GiB |
| Memory optimized | Epsv7 / Epdsv7 | 8 GiB | 128 | 1,024 GiB |
| High memory (new) | Mpsv4 / Mpdsv4 | 16 GiB | 84 | 1,388 GiB |
| Storage optimized (new) | Lpsv5 | 8 GiB | 128 | 1,024 GiB |
Lpsv5 includes local NVMe on every size, up to 23 TB. Microsoft lists maximum network bandwidth of up to 85 Gbps and remote storage throughput of up to 70 Gbps across the table. A third-party guide, KloudSchool, says the Mpsv4/Mpdsv4 series is slightly lower on both. That guide describes the preview period, so confirm the exact limits in Azure documentation for the size you want.
Developers can deploy through existing Azure tools and use the VMs in Virtual Machine Scale Sets or as AKS node pools. Microsoft's listed operating systems are AlmaLinux, Azure Linux, Debian, Oracle Linux, Red Hat Enterprise Linux, Rocky Linux, SUSE Linux Enterprise Server, Ubuntu and Windows 11 Client. Windows Server is not on that list. That matters to a lot of WindowsForum readers. Canonical separately says Ubuntu 22.04, 24.04 and 26.04 LTS are available on Cobalt 200.
Silicon and controls
Microsoft's blog describes a TSMC 3nm SoC with 132 Neoverse V3 cores. Each vCPU is a full physical core, with 3 MB of private L2 and 192 MB of system-level cache. Memory is 12-channel DDR5, with encryption on by default. Azure Boost is included, and Microsoft claims these gains over Cobalt 100:
- Up to 15% higher network bandwidth.
- 20% higher remote-storage IOPS.
- 10% higher remote-storage throughput.
Every server also has an Azure Integrated HSM. Microsoft says on-chip accelerators will cut Azure SQL backup times by up to 2.5 times.
One practical control is the processor mode setting:
- Deterministic Mode runs cores at a steady 3.2 GHz. Microsoft suggests it for distributed data and analytics clusters that run at the pace of their slowest node, and for services with tight performance targets.
- Opportunistic Mode can boost to 3.6 GHz when power and thermal headroom allow. Microsoft says that gives up to 12% more performance, with some frequency variability.
You choose the mode with the ProcessorMode setting at VM creation. You can also change it later through the Azure portal, Azure CLI, PowerShell, or Resource Manager and Bicep templates.
The benchmarks: who compared what
Don't blend these numbers. They use different baselines and different testers.
Microsoft's comparison against Cobalt 100 (internal benchmarking):
- Up to 1.5× for web serving.
- Up to 1.6× for Java.
- Up to 1.5× for AI-agent tool execution.
- 1.9× for databases.
Arm's comparison against AMD EPYC Turin-based Azure instances:
- Up to 139% higher MySQL performance.
- 56% higher NGINX performance.
- 65% higher Redis performance.
- Up to 56% faster x265 4K encoding.
- Up to 56% higher Milvus HNSW vector-search performance.
- Up to 57% higher fastText performance.
- Up to 46% higher SWE-bench task completion at 144 concurrent workers.
- An average of 1.48× across three EDA simulation workloads.
Arm's separate DiskANN figure is 1.45× higher geometric-mean vector-search QPS on Cobalt 200 than on Cobalt 100. Arm also reports an 8× average and up to 14× QPS boost from Arm-optimized software, comparing DiskANN v0.61.0 with v0.56.0. That second figure compares two software versions, so it isn't a hardware-only gain.
Arm describes its results as internal tests. Its methodology document lists Ubuntu 26.04 and specific instance sizes. For example, the MySQL tests used E16ps_v7, E16ps_v6 and E16as_v7 instances. Treat all of these as vendor numbers until you've reproduced them with your own workload.
Customer testimonials support the direction but not a universal result. Amadeus reported more than 30% higher flight-search throughput versus Cobalt 100. OpenAI said "we can serve more users with the same number of cores." Microsoft also says Teams media processing saw 1.65× the performance of Cobalt 100, and that GitHub Actions larger Arm runners finish jobs faster.
The market and capacity claims
Microsoft plans for Cobalt 200 to represent a significant portion of new Azure Compute capacity in 2027 and beyond. Arm adds that its server share at leading hyperscalers rose from 15% in 2024 to over 50% in 2025, and it expects more than 60% by the end of 2026. That is Arm's own forecast, not an independent market result. Microsoft says it will keep investing in x86 chips. So this is an expansion of choice, not an x86 exit.
Practical guidance before migrating
- Check Arm builds. The WindowsForum keynote roundup offers this advice: check that your workloads and container images have Arm builds before you migrate.
- Start with Cobalt 100 workloads. Microsoft says workloads already on Cobalt 100 can move to Cobalt 200 where they need more performance, larger sizes or the new families.
- Match the family to the bottleneck. Use Mpsv4 for large in-memory databases and Lpsv5 for local-storage-heavy staging and indexing.
- Pick a processor mode deliberately. Use Deterministic for latency-consistent clusters and Opportunistic for web, API and transaction workloads.
- Benchmark your own stack. Arm's tests used Ubuntu 26.04 with pinned versions of the tested software. Your results may differ.
- Check Windows needs. If you need Windows Server on Arm, Microsoft's launch list doesn't include it.
Neither source offers a universal migration procedure or a guarantee that every x86 workload moves over unchanged. Arm points to migration learning paths, a cloud migration program and a software ecosystem dashboard as starting points.
Analysis
The headline gain is credible in direction. Microsoft and Arm both report it, and customers like OpenAI and Amadeus report similar improvements on their own workloads. The AMD comparisons come from Arm, which sells the architecture, so read them as marketing-grade evidence rather than neutral lab results. The most useful new facts for admins are the two new families, the selectable processor modes, the supported OS list, and the regional footprint. Those determine whether Cobalt 200 fits your estate.
References
- Microsoft Azure Cobalt 200 VMs reach general availability, bringing Arm-powered performance to cloud and agentic AI workloads - Arm Newsroom Arm Newsroom · 2026-11-17T18:35:22+00:00
- Your Privacy Choices Opt-Out Icon azure.microsoft.com
- How Arm Neoverse CSS-based Microsoft Cobalt 200 virtual machines accelerate cloud and agentic AI performance - Arm Newsroom Arm Newsroom · 2026-11-17T18:35:00+00:00