Microsoft has made Azure Cobalt 200 virtual machines generally available, bringing its second-generation custom Arm server processor to customers in 14 cloud regions. Announced November 17, 2026, at Microsoft Ignite, the rollout follows an earlier preview and represents a commercial service launch rather than a future chip announcement. Microsoft reports up to 50% higher CPU performance than its Cobalt 100 generation, depending on workload.
The company plans to expand availability to 23 regions by the end of 2026 and expects Cobalt 200 to supply a significant portion of new Azure Compute capacity beginning in 2027. Early users include OpenAI, Databricks, Snowflake, InMobi and Amadeus, although their operating environments and performance results differ.
Microsoft Azure, August 18, 2026: “Deploying cloud native apps to Cobalt 200 VMs.” Official demonstration of ARM64 migration, containers and AKS deployment, recorded before general availability.
What is inside Cobalt 200?
The processor is manufactured on TSMC’s 3-nanometer process and integrates 132 Arm Neoverse V3 CPU cores. Each exposed Azure vCPU maps to a physical core, with 3 MB of private L2 cache per core and 192 MB of shared system-level cache. A 12-channel DDR5 memory controller, hardware encryption, compression acceleration and Azure Boost networking complete the infrastructure design.
Microsoft offers two operating modes. Deterministic Mode holds cores at a consistent 3.2 GHz, while Opportunistic Mode can boost frequency up to 3.6 GHz when thermal and power headroom permit. Predictable clocks may matter to synchronized analytical clusters; faster variable clocks can benefit web and transaction services. Neither option guarantees identical application performance across every workload.
The official Azure announcement reports up to 1.6 times the Java workload performance, 1.5 times the agent-tool execution performance and 1.9 times the database workload performance of Cobalt 100. Those values come from Microsoft testing, not a universal independent benchmark or a promise of lower customer bills.
Microsoft Azure’s original November 17 X announcement confirms Cobalt 200 VM general availability. It documents the launch, not an independent performance measurement.
Why cloud CPUs still matter in the GPU era
AI systems do not perform every action on GPUs. Agents often execute code, query databases, access storage, call services and run isolated development environments. Those operations depend heavily on CPU performance, networking and memory. Cobalt 200 targets the infrastructure surrounding model inference, not replacement of accelerators used for large-scale neural network computation.
Microsoft says Cobalt 200 virtual machines include Azure Boost, which offloads networking and remote storage work. Compared with the previous generation, Microsoft cites up to 15% higher network bandwidth, 20% higher remote-storage IOPS and 10% higher storage throughput. Such gains may reduce infrastructure overhead, but an application limited by another bottleneck could benefit far less.
New memory and storage choices
The rollout includes general-purpose, memory-optimized and storage-oriented VM families. High-memory Mpsv4 provides 16 GiB per vCPU, while storage-optimized Lpsv5 offers up to 23 TB of local NVMe capacity. Microsoft lists deployment support for Ubuntu, Debian, Red Hat Enterprise Linux and other operating systems, with common Azure Kubernetes Service and virtual machine management tools. Customers should still verify ARM64 binary compatibility, licensing and regional SKU availability before migrating.
Microsoft Azure, January 27, 2026: “Silicon to Systems: What powers Microsoft’s AI Infrastructure.” Official archival overview of Cobalt, Maia accelerators and data center design, rather than a November benchmark.
How the launch changes infrastructure planning
The milestone builds on Microsoft’s broader silicon strategy. BitcoinVersus.Tech previously reported on Microsoft’s Maia accelerator, explained how virtual machines isolate workloads and covered virtual CPU allocation. The site also examined NVIDIA Vera agent sandbox tests, another example of CPUs becoming more prominent in AI infrastructure.
Cobalt 200 is available today, but better price-performance must be established using comparable VM configurations, actual cloud rates and representative workloads. The decisive metric for customers remains useful work completed per dollar and per watt, not a peak benchmark alone.
Sources: Microsoft Azure, November 17, 2026; Arm, November 17, 2026; Microsoft Ignite infrastructure announcement; Microsoft Azure original X post.

Leave a Reply