<-- Back to All News

Microsoft Announces availability of Azure Cobalt 200, proven on real workloads and built for the agentic era

Publish Date: November 18, 2026

Executive Overview

As enterprise artificial intelligence evolves from simple chat completions into continuous, multi-step autonomous agents, infrastructure attention is rapidly expanding beyond GPU accelerators to the general-purpose CPU layer that orchestrates everything around them. Autonomous agents rely heavily on low-latency CPU execution for agent sandboxes, file search, real-time code compilation, database querying, microservice routing, and external tool execution. Without substantial generational leaps in per-core CPU performance, memory bandwidth, and energy efficiency, the control plane and data plane supporting modern AI applications become severe latency and cost bottlenecks.

At Microsoft Ignite 2026, Rani Borkar and Arun Kishan announced the general availability of Azure Cobalt 200-based virtual machines (VMs), Microsoft’s second-generation custom Arm64 cloud silicon. Delivering up to 50 percent better performance than Cobalt 100, Cobalt 200 VMs launch immediately across 14 Azure regions with expansion to 23 regions by the end of 2026, and are projected to represent a significant share of new Azure Compute capacity beginning in 2027. Engineered through a full systems approach co-optimizing silicon, software, networking, power, and rack architecture, Cobalt 200 introduces dedicated VM families for high-memory (Mpsv4) and storage-intensive (Lpsv5) workloads alongside hardware-enforced security and on-chip offload accelerators.

For Chief Information Officers, cloud infrastructure architects, and FinOps leaders, the general availability of Cobalt 200 marks a decisive tipping point for Arm-based enterprise computing. Proven across massive internal Microsoft services and flagship external customers including OpenAI, Databricks, Snowflake, Amadeus, and InMobi, Cobalt 200 demonstrates that custom Arm silicon is no longer a niche cost-saving experiment for stateless web tiers, but a premier, high-performance compute foundation for databases, electronic design automation, and agentic AI infrastructure.

Features

Azure Cobalt 200 introduces a ground-up custom silicon and systems architecture engineered to maximize per-core throughput, memory efficiency, I/O acceleration, and hardware security. All technical specifications below are extracted directly from the official announcement.

TSMC 3nm Cobalt 200 SoC and Arm Neoverse V3 Core Architecture

At the heart of every Cobalt 200 VM lies the Cobalt 200 system-on-chip (SoC), fabricated on TSMC’s advanced 3nm process technology using a multi-chipset architecture and advanced packaging. The SoC integrates 132 Arm Neoverse V3 cores (co-developed with Arm Neoverse CSS V3). Each vCPU maps to a full physical core equipped with 3 MB of private L2 cache and 192 MB of shared system-level cache, ensuring consistent performance under heavy load while keeping active working sets close to execution cores.

Custom Memory Subsystem, Azure Boost, and Hardware Security

Cobalt 200 incorporates a custom memory controller supporting 12 channels of DDR5 memory with memory encryption enabled by default at negligible latency overhead. Every Cobalt 200 VM includes the latest generation of Microsoft Azure Boost, offloading networking and storage processing from the host CPU to deliver up to 15 percent higher network bandwidth, 20 percent higher remote storage IOPS, and 10 percent higher remote storage throughput compared with Cobalt 100. For hardware-rooted protection, every Cobalt 200 server includes the Azure Integrated HSM to safeguard server hardware against unauthorized intrusion and protect encryption keys for hosted workloads.

On-Chip Compression and Cryptography Accelerators

To reclaim valuable CPU cycles from repetitive infrastructure tasks, Cobalt 200 embeds dedicated, low-power compression and cryptography accelerators directly on the silicon. Offloading these operations reduces host CPU consumption and delivers dramatic workload improvements—for example, reducing Azure SQL backup times by up to 2.5 times while minimizing performance disruption to live transactional workloads.

Dual Processor Modes: Deterministic vs. Opportunistic

Recognizing that different enterprise workloads prioritize either strict predictability or peak burst speed, Cobalt 200 VMs introduce two selectable processor modes configurable via the Azure portal, Azure CLI, PowerShell, ARM templates, or Bicep through the ProcessorMode setting. Deterministic Mode locks cores at a consistent 3.2 GHz for distributed data clusters and tight SLA services where straggler nodes hurt performance. Opportunistic Mode dynamically boosts cores up to 3.6 GHz when power and thermal headroom permit, delivering up to 12 percent additional performance for web services, APIs, transaction processing, and media workloads.

Expanded VM Families and Mature Arm Ecosystem Support

Cobalt 200 VMs span five VM series across general-purpose, memory-optimized, high-memory, and storage-optimized categories, adding two new Arm-based VM families to Azure: Mpsv4 for high-memory workloads and Lpsv5 for workloads requiring fast access to large local datasets. Cobalt 200 supports major operating systems including Ubuntu (with Canonical Ubuntu 26.04 LTS Kernel Livepatch for rebootless Arm kernel fixes), AlmaLinux, Oracle Linux, Debian, Rocky Linux, and Windows 11 Client, and deploys natively across Virtual Machine Scale Sets and Azure Kubernetes Service (AKS) node pools.

Benefits

Migrating cloud-native, database, and agentic workloads onto Azure Cobalt 200 VMs delivers compelling performance, operational, and economic advantages across the enterprise IT estate.

First, Cobalt 200 unlocks substantial generational performance leaps that translate directly into fleet consolidation and lower unit costs. In Microsoft’s benchmarks over Cobalt 100, Cobalt 200 achieved up to 1.5x performance for web serving, up to 1.6x for Java-based workloads, up to 1.5x for AI agent tool execution, and 1.9x for database workloads. Because each vCPU is a full physical core with generous private L2 and system-level cache, organizations achieve higher throughput and lower tail latency using fewer cores, reducing cloud compute spend and software licensing exposure while supporting higher traffic density.

Second, the introduction of Deterministic and Opportunistic processor modes gives performance engineers unprecedented control over compute behavior without changing VM SKUs. Distributed analytics and database clusters—which historically suffered when variable turbo frequencies created uneven node execution speeds—can lock into Deterministic Mode at 3.2 GHz for predictable scaling. Conversely, latency-sensitive API gateways, web tiers, and agent tool sandboxes can activate Opportunistic Mode up to 3.6 GHz to capture up to 12 percent extra throughput whenever thermal headroom allows.

Third, integrating Azure Boost, default DDR5 memory encryption, Azure Integrated HSM, and on-chip compression and cryptography accelerators directly into the baseline server architecture removes traditional trade-offs between security and performance. Enterprises gain hardware-rooted key protection, encrypted memory, faster database backups, and accelerated network and storage I/O automatically across every Cobalt 200 VM, strengthening enterprise security posture without incurring the heavy CPU tax typically associated with software-based encryption and compression.

Use Cases

Proven across both Microsoft’s internal hyperscale services and leading global enterprises, Azure Cobalt 200 VMs excel across high-concurrency AI, data analytics, software development, and real-time transaction workloads.

High-Concurrency AI Serving and Agentic Workflows

OpenAI utilizes Azure Cobalt VMs to sustain the massive global scale of ChatGPT. Having already transitioned services to Arm on Cobalt 100, OpenAI tested identical ChatGPT workloads on Cobalt 200 and recorded higher throughput and lower latency, enabling the platform to serve more concurrent users with the same core count while reducing time-to-first-response. Similarly, ad-tech leader InMobi runs high-throughput local neural network inference using SIMD on Cobalt 200, processing approximately 50,000 inferences per second on each 32-core machine, creating compounding fleet-wide efficiency gains.

Serverless Data Analytics and Enterprise Databases

In the data platform ecosystem, Azure Databricks runs serverless compute for its Genie AI coworker, SQL queries, and data pipelines on Cobalt 200, achieving faster cold starts and higher execution efficiency. Snowflake has likewise progressed across Arm generations on Azure from v5 to Cobalt 100 and now Cobalt 200 to accelerate customer analytics. Internally, Microsoft powers Azure Database for PostgreSQL (announced on Cobalt 200 at Build 2026) and Azure SQL on Cobalt 200, leveraging high memory capacity, local storage, and on-chip accelerators to accelerate queries and cut backup times by up to 2.5x.

High-Throughput Travel Search and Real-Time Media Processing

In travel technology, Amadeus benchmarked flight search workloads on Cobalt 200 and achieved more than a 30 percent improvement in flight search throughput compared with Cobalt 100, exceeding internal engineering expectations. Within Microsoft, the Microsoft Teams video media processing service observed 1.65x performance over Cobalt 100 with linear scaling across physical cores up to 90 percent vCPU utilization, while Xbox improved sign-in, social presence, and caching-layer efficiency on Arm.

Silicon Engineering (EDA) and Developer CI/CD Pipelines

Microsoft’s own silicon engineering teams run Electronic Design Automation (EDA) workloads for design verification, physical implementation, and sign-off on Cobalt 200 VMs, recording 1.3x to 4.15x performance gains over Cobalt 100 and accelerating development of future Maia and Cobalt chips. Simultaneously, GitHub Actions larger Arm runners on Cobalt 200 complete developer builds noticeably faster across millions of daily CI/CD workflows, shortening feedback loops and lowering compute consumption.

Alternatives

Infrastructure leaders evaluating Azure Cobalt 200 VMs should compare custom Azure Arm64 silicon against traditional x86 instances and competing hyperscaler Arm architectures.

  • x86-64 Cloud Virtual Machines (Intel Xeon and AMD EPYC on Azure): Enterprises running legacy proprietary binaries, older Windows Server applications, or specialized x86-only commercial off-the-shelf software must continue relying on Intel or AMD x86 VMs. While x86 instances offer universal backward compatibility without recompilation or container image rebuilding, they typically carry higher per-core licensing and compute costs and lack Cobalt 200’s bespoke co-optimization with Azure’s custom on-chip compression and cryptography accelerators.
  • Competing Hyperscaler Custom Arm Silicon (AWS Graviton and Google Cloud Axion): Organizations operating multicloud estates can evaluate AWS Graviton4 or Google Cloud Axion Arm64 VMs, both of which deliver strong price-performance and mature Linux ecosystem support. However, enterprises deeply anchored in Azure Database for PostgreSQL, Azure SQL, AKS, GitHub Actions, and Microsoft Foundry benefit from Cobalt 200’s native Azure Boost I/O integration, Azure Integrated HSM security boundary, and zero-egress proximity to their existing Azure data and AI estates.
  • Dedicated GPU Instances for Lightweight Inference: Teams executing high-volume, small-model neural network inference (such as recommendation engines or ad-serving classifiers) traditionally defaulted to discrete GPU VMs. As demonstrated by InMobi’s 50,000 SIMD inferences per second per 32-core Cobalt 200 VM, modern Arm64 CPUs with wide vector pipelines can handle lightweight local inference at dramatically lower cost and higher regional availability than scarce GPU clusters.
An Alternative Perspective

Despite Cobalt 200’s impressive benchmark numbers and flagship customer endorsements, infrastructure and operations leaders must carefully evaluate migration friction, regional availability, and architectural caveats before committing wholesale to an Arm64 fleet transition.

First, migrating legacy enterprise estates from x86 to Arm64 is rarely a zero-effort lift-and-shift exercise. While modern interpreted runtimes, Java, .NET, containerized Linux microservices, and popular open-source databases transition smoothly, real-world corporate portfolios frequently contain third-party security agents, legacy monitoring daemons, compiled C/C++ native libraries, or vendor-supplied binaries that still lack certified Arm64 builds. Platform teams must invest upfront engineering effort into multi-architecture CI/CD build pipelines, dependency audits, and performance regression testing before realizing Cobalt 200’s promised price-performance savings.

Second, geographic coverage and capacity ramp require disciplined capacity planning. Launching in 14 Azure regions today and expanding to 23 regions by the end of 2026 covers major commercial hubs, but falls short of Azure’s full global regional footprint. Multinational enterprises operating active-active disaster recovery architectures across secondary or sovereign regions must verify paired-region Cobalt 200 availability—particularly for the newly introduced Mpsv4 high-memory and Lpsv5 storage-optimized families—so failover topologies are not forced to fall back onto heterogeneous x86 SKUs with different performance profiles.

Third, organizations must test Opportunistic Mode carefully under sustained multi-tenant production loads. Because Opportunistic Mode boosts cores from 3.2 GHz up to 3.6 GHz only when power and thermal headroom allow, performance gains of up to 12 percent are inherently non-deterministic. Workloads with strict P99 latency SLAs or tightly synchronized distributed barriers should default to Deterministic Mode rather than building capacity assumptions around peak opportunistic boost clocks.

Final Thoughts

The general availability of Azure Cobalt 200 VMs represents a major milestone in Microsoft’s silicon-to-systems strategy. Combining TSMC 3nm manufacturing, 132 Arm Neoverse V3 physical cores, 12-channel encrypted DDR5 memory, Azure Boost networking and storage offload, Azure Integrated HSM hardware security, and on-chip compression and cryptography accelerators, Cobalt 200 delivers up to 50 percent generational gains across the exact workloads that power modern enterprises—from PostgreSQL and Databricks to ChatGPT serving and agentic tool sandboxes.

With Cobalt 200 set to comprise a major share of new Azure Compute capacity starting in 2027, enterprise IT leaders should treat Arm64 readiness as a core 2027 architectural priority. By prioritizing greenfield AKS node pools, managed PostgreSQL and SQL tiers, Java microservices, and GitHub Actions runners for early Cobalt 200 adoption—and leveraging Deterministic versus Opportunistic processor modes intentionally—organizations can capture immediate price-performance advantages while modernizing their compute foundation for the agentic era.

Source