We are announcing the aboriginal entree preview for Azure Cobalt 200 Arm-based Virtual Machines (VMs), designed for Linux-based agentic AI workloads.
As organizations progressively clasp the agentic epoch for AI, lawsuit demands for compute are reshaping the architecture of unreality infrastructure arsenic we cognize it. Today astatine Microsoft Build 2026, we are announcing the aboriginal entree preview for Azure Cobalt 200 Arm-based Virtual Machines (VMs), designed from the crushed up for scale-out, cloud-native, and Linux-based agentic AI workloads, with up to 50% amended generational show implicit Cobalt 100.
Cobalt 200 is purpose-built from silicon to servers to services—integrating Microsoft’s latest innovations successful security, networking, storage, and offload to outperform accepted Arm-based compute. That hardware-software co-optimization lets america propulsion the boundaries of scale, security, and outgo for AI inferencing, information pipelines, and the web and API tiers that powerfulness modern services. Agents are unsocial from accepted workloads: they reason, marque sequential decisions, and tally continuously astatine scale—demanding a fundamentally antithetic computational profile. Cobalt 200 is built for precisely this environment, delivering 50% show gains for the workloads defining the adjacent epoch of endeavor AI, making agents faster, much capable, and economically viable astatine scale.
Building connected the occurrence of Cobalt 100 VMs
Microsoft Azure Cobalt 100, our archetypal custom-built processor for cloud-native workloads is present deployed successful 32 Azure datacenter regions astir the world, with unreality analytics leaders similar Databricks and Snowflake adopting Cobalt 100 to optimize their unreality footprint, portion customers similar Amadeus, OneTrust, Siemens, Sprinklr, and Temenos person achieved important real-world show and ratio gains.
Within Microsoft’s ain unreality services, Azure Cobalt 100 VMs are delivering up to 45% amended show portion utilizing 35% less compute cores compared to its erstwhile compute platform. Microsoft Defender for Endpoint (MDE) saw 40% amended show successful its cyber information curator, enabling faster menace effect astatine monolithic scale.
With Cobalt 200 VMs, we are applying everything we learned towards supporting these demanding services to rise the barroom adjacent further.
What’s caller successful Cobalt 200 Arm-based VMs
The Cobalt 200 CPU represents a important leap guardant successful performance, security, and efficiency. At the bosom of each Cobalt 200 VM is the Cobalt 200 System-on-Chip (SoC), our second-generation Arm processor built astir the Arm Neoverse V3 Compute Subsystems, the highest-performance V-series halfway successful Arm’s portfolio. Fabricated connected TSMC’s 3nm (N3P) process, Cobalt 200 features a modern chiplet architecture, customized accelerators, and a customized representation controller.
Key innovations successful Cobalt 200 VMs include:
- Generational show gains crossed compute, storage, and networking: Compared to previous-generation Cobalt 100 VMs, caller Cobalt 200 VMs present up to 50% amended CPU performance, 20% higher distant retention IOPS with NVMe, 10% amended distant retention throughput with NVMe, and 15% higher web bandwidth, with improvements varying by workload.
- Up to 128 vCPUs: Cobalt 200 VMs standard up to 128 vCPUs, delivering much compute capableness for demanding scale-out, cloud-native, agentic AI, and data-intensive workloads.
- Faster distant retention and networking: Azure Boost integration helps amended distant retention IOPS and throughput with NVMe portion expanding web bandwidth, benefiting distributed applications, storage-intensive services, and high-throughput information pipelines.
- Advanced cache and scalable strategy design: A modern chiplet architecture with a larger cache hierarchy—including 3 MB of L2 cache per halfway and 192 MB of system-level L3 cache—keeps much progressive information person to the workload, helping trim latency and amended responsiveness for databases, in-memory caches, analytics engines, and different data-intensive services.
- Stronger information by default: Memory encryption is enabled by default done a custom-designed representation controller, helping rise the baseline information posture for each workload with negligible show impact.
Performance for agentic AI workloads
Cobalt 200 delivers the per-core show and scalability needed to powerfulness modern agentic AI workloads. Each Cobalt 200 halfway is simply a afloat carnal core, paired with dedicated 3 MB of L2 cache, and starring representation bandwidth per core. These plan points alteration higher isolation and sustained show nether load, which let agentic workloads to battalion much cause sandboxes per VM portion gathering latency and throughput requirements.
Real unreality workload performance
Cobalt 200 VMs amusement a wide generational uplift implicit Cobalt 100 connected the workloads that substance astir successful production.
With unreality workloads, we observe:
- Up to 135% amended show for unreality database workloads.
- Up to 40% amended show for web serving workloads.
- Up to 45% amended show for connection encryption workloads.
- And up to 80% amended show for caching workloads.

The uplift measured is disposable successful existent hyperscale services. Taken together, these results amusement that Cobalt 200 raises the level connected per-core show crossed cloud-native applications, databases, analytics, caches, and communications workloads.
Industry partners and lawsuit adoption
We person been moving intimately with our exertion partners and customers during the preview play to guarantee Cobalt 200 Arm-based VMs present beardown results crossed a wide scope of real-world workloads. Leading bundle improvement companies and enterprises are already evaluating and adopting these VMs for their astir demanding applications.
Teradata
Teradata is excited to beryllium an aboriginal preview spouse for Microsoft’s Cobalt 200 VMs. Microsoft has been a beardown partner, and we worth the accidental to assistance signifier the plan and specifications to amended conscionable the needs of our associated customers. Early investigating has been encouraging, and we look guardant to continued collaboration.
—Brandon Mincey, Engineering Fellow, TeradataElastic
Elastic is committed to delivering best-in-class show with the Search AI Platform that powers observability, security, search, and AI solutions. Performance and outgo ratio are captious for teams moving these workloads astatine standard and our archetypal investigating of the caller Cobalt 200 VMs shows committedness for further improvements successful these areas. We look guardant to bringing these benefits to Elasticsearch users connected Azure done our continued collaboration with Microsoft.
—Yuvraj Gupta, Director, Product Management, ElasticArm
Agentic AI is reshaping the unreality and creating request for infrastructure that tin efficiently orchestrate and standard millions of intelligent interactions successful existent time. Our collaboration with Microsoft connected Cobalt 200, built connected Arm Neoverse CSS V3, reflects however purpose-built Arm-based compute is enabling the adjacent procreation of AI-driven services portion continuing to present exceptional show for cloud-native applications.
—Eddie Ramirez, Vice President of go-to-market, Cloud AI Business Unit, ArmCanonical
Cobalt 200 delivers cardinal advances for accumulation Linux workloads connected Arm, including representation encryption enabled by default, built-in acceleration for compression and encryption, and higher throughput for data-intensive unreality services. Ubuntu gives organizations a accordant level for cloud-native and agentic AI workloads crossed Azure, and Ubuntu Pro keeps it production-ready with semipermanent information attraction and Livepatch, which present brings rebootless kernel updates to Arm truthful always-on services enactment available.
—Jehudi Castro-Sierra, Public Cloud Alliance Director, CanonicalDeveloper ecosystem and Arm compatibility
The Arm developer ecosystem continues to thrive. Cobalt 200 Arm-based VMs present afloat compatibility for workloads presently moving connected Cobalt 100 VMs, making migration seamless. Major developer platforms and languages—including C++, .NET, Java, Python, and Rust—provide Arm-native versions with optimizations that afloat leverage the capabilities of the Arm architecture.
The broader ecosystem has embraced Arm with autochthonal enactment crossed fashionable infrastructure and deployment solutions. GitHub Actions supports Arm done some self-hosted and GitHub-hosted runners. Azure Kubernetes Service (AKS) supports Arm cause nodes arsenic good arsenic mixed x86 and Arm architecture clusters. Containerized workloads payment from the increasing availability of Arm-native instrumentality images crossed the ecosystem.
Microsoft services powered by Cobalt 200 VMs
Microsoft’s ain unreality services are among the archetypal to follow Cobalt 200 Arm-based VMs, gathering connected the occurrence of Cobalt 100 adoption crossed our astir demanding, mission-critical workloads.
Dataverse
Dataverse is the halfway exertion and information level underpinning Dynamics and Power Platform workloads, requiring precocious performance, scalability, and debased latency to present a responsive lawsuit experience.
We deployed the Power Apps level connected Cobalt 100 for its improved show and powerfulness efficiency. We are validating Cobalt 200 present and excited by the show gains we are seeing–up to 60% amended show for our basal workload implicit Cobalt 100.
—Mauktik Gandhi, VP Engineering, Agent 365 PlatformAzure Databases
Azure SQL Database is simply a earthy acceptable for Cobalt 200 Arm-based VMs. The built-in compression and cryptography accelerators are peculiarly impactful for database workloads—by offloading compression and encryption tasks to the Cobalt 200 accelerator, Azure SQL is capable to trim the usage of captious compute resources, prioritizing them for lawsuit queries and transactions.
Our teams collaborated from the opening to guarantee that the architecture of Cobalt 200 is optimized for hosting databases successful Azure. We are excited astir the show gains we are seeing implicit Cobalt 100 and looking guardant to broader availability aboriginal this year.
—Shireesh Thota, Corporate Vice President, Azure DatabasesVM families and specifications
Cobalt 200 VMs importantly grow our Arm VM portfolio based connected lawsuit feedback to enactment a broader acceptable of workloads. While Cobalt 100 offered General Purpose (Dp,Dpl), and Memory Optimized (Ep) VM families, Cobalt 200 adds 2 much VM families: the High-Memory Optimized Mpsv4 VMs and Dense Local Storage Lpsv5 VMs, bringing much prime crossed compute, memory, and retention profiles.
All VM bid present up to 85 Gbps of web bandwidth and 70 Gbps of distant retention throughput, but Mpsv4/Mpdsv4, which supply up to 70 Gbps of web bandwidth and 46 Gbps of distant retention throughput. Most bid are disposable with oregon without section NVMe disks, portion Lpsv5 includes section NVMe retention crossed each sizes, giving customers flexibility to optimize for outgo oregon performance.
| General Purpose Dplsv7/Dpldsv7 | 1–128 | 2:1 | Yes – Up to 7 TiB | Most non-memory-intensive and scale-out workloads specified arsenic microservices, tiny databases, caches, gaming servers, and more. |
| General Purpose Dpsv7/Dpdsv7 | 1–128 | 4:1 | Yes – Up to 7 TiB | Most scale-out endeavor workloads specified arsenic web and exertion servers, tiny to mean databases, caches, and more. |
| Memory Optimized Epsv7/Epdsv7 | 1–128 | 8:1 | Yes – Up to 7 TiB | Large relational and NoSQL databases, in-memory caches specified arsenic Redis and Memcached, and real-time analytics. |
| (New) High Memory Optimized Mpsv4/Mpdsv4 | 1–84 | 16:1 | Yes – Up to 4.4 TiB | Large in-memory databases, ERP systems, large-scale caching layers, and memory-intensive analytics workloads. |
| (New) Storage Optimized Lpsv5 | 1–128 | 8:1 | Yes – Up to 23 TB | Data pre-processing and staging, relational and NoSQL databases with section retention requirements, large information analytics, and search/index engines. |
All Cobalt 200 VMs volition enactment distant disk types including Standard SSD, Standard HDD, Premium SSD, and Ultra Disk storage. You tin deploy these caller VMs utilizing existing methods, including the Azure portal, SDKs, APIs, PowerShell, and the command-line interface (CLI).
Availability
Cobalt 200 Arm-based VMs are present disposable successful preview. The caller VMs volition beryllium disposable successful preview successful the pursuing regions: West US3, East US2, Central US, Sweden Central, East US, West US2, Spain Central, and Indonesia Central, with further regions to beryllium announced.
The adjacent section successful Azure’s customized silicon journey
The announcement of Cobalt 200 Arm-based Virtual Machines marks the adjacent section successful Azure’s purpose-built infrastructure journey. With our end-to-end systems approach—designing the CPU in-house and optimizing it for Azure’s infrastructure—we are delivering a tightly integrated level that offers performance, powerfulness efficiency, and information for our customers.
Whether you’re accelerating merchandise development, scaling analytics platforms, moving mission-critical databases, oregon improving idiosyncratic experiences, Cobalt 200 Arm-based VMs connection a compelling prime for modern unreality workloads. We are excited to spot however our customers and partners make breakthrough products and services connected this caller platform.
Try the Cobalt 200 Arm-based VMs contiguous successful aboriginal entree preview and acquisition the adjacent procreation of Azure’s customized infrastructure innovation.
Thank you for joining america connected this breathtaking journey.
Get early entree to Azure Cobalt 200 VMs.
Additional resources
- For questions, delight spell to Azure Support, and our experts volition beryllium determination to assistance you.
- Read Arm’s Cobalt 200 VMs preview supportive blog.
- Read Canonical’s Cobalt 200 VMs preview supportive blog.