Microsoft releases Cobalt 200 cloud VMs, targeting the CPU work behind AI agents
The release brings Microsoft’s second-generation Arm-based processor to general availability. Arm’s benchmarks span databases, search and coding-agent workflows—but measure different kinds of gains.
Available November 17, 2026, Cobalt 200 gives Azure customers a new Arm-based option for CPU-heavy cloud work around model calls—code execution, retrieval, and tool coordination—rather than model generation itself. Microsoft says it delivers up to 50% better CPU performance than Cobalt 100, while Arm’s separate tests against AMD-based Azure instances report gains on selected services and agent-related workloads. These are vendor-run, workload-specific measurements; Microsoft’s expectation that Cobalt will account for a significant share of new Azure capacity in 2027 and beyond remains a plan.
01
The Dpsv7, Dplsv7, and Epsv7 VM families scale to 128 virtual CPUs and 512 GiB of memory.
02
Against AMD EPYC Turin-based Azure instances, Arm reports up to 139% higher MySQL, 56% higher NGINX, and 65% higher Redis performance.
03
Arm reports up to 56% higher Milvus HNSW vector-search performance and up to 46% higher SWE-bench task completion with 144 concurrent workers.
Microsoft made Azure Cobalt 200 virtual machines generally available on November 17, 2026, giving cloud customers its next-generation Arm-based computing option. Announced at Microsoft Ignite, the release targets ordinary business applications and the work behind AI agents: running code, retrieving information and coordinating tools. Arm accompanied the launch with workload-specific performance claims.
Microsoft reports up to 50% better CPU performance than Cobalt 100 VMs. That is the generational comparison. Arm also published tests against AMD EPYC Turin-based Azure instances, plus separate results for optimized search software. Those comparisons answer different questions and should not be read as one universal speed claim.
The processor work behind an agent
Arm’s case starts with what happens around a model call. An agent may execute code, fetch enterprise records, invoke a service and coordinate several applications. CPUs handle those operations alongside the databases, caches and other services that support business software. Cobalt 200 is aimed at that interconnected workload, not just a model’s response generation.
The processor uses Arm Neoverse Compute Subsystems V3, a pre-integrated computing framework that Microsoft customizes for Azure. Those customizations include compression, cryptographic accelerators and power management for individual cores. The available VM families—Dpsv7, Dplsv7 and Epsv7—scale to 128 virtual CPUs and 512 GiB of memory.
Arm illustrates the arrangement with a customer-service demonstration. A small language model hosted on Cobalt 200 classifies requests and routes them to specialized agents for scheduling, complaints, product inquiries and refunds. The agents combine CPU-based model execution, data retrieval, external model APIs and tools to complete their tasks. It is a demonstration of the workflow, rather than a customer deployment result.
What Arm measured against AMD
For its competitive tests, Arm compared Cobalt 200 with Azure instances powered by AMD EPYC Turin processors. Its published benchmarks cover established cloud applications as well as components used in agent workflows. These are vendor-run measurements, not independent verification; the results below describe the particular workloads Arm tested.
The agent-related tests span information retrieval and software-engineering tasks. They are not interchangeable measures: one tests a search component, while another measures task completion in a workflow with many workers running at once. Arm reports:
Up to 56% higher Milvus HNSW vector-search performance, for finding similar information in a vector database.
Up to 46% higher SWE-bench task-completion performance in end-to-end software-engineering workflows with 144 concurrent workers.
Search gains also depend on software
Arm and Microsoft have also been optimizing DiskANN, Microsoft’s technology for searching large collections of vector representations. The work includes processor-level optimizations for similarity calculations using compact, 4-bit data. That targets a computational building block of retrieval, which agents can use to find relevant information.
Here, the distinction between hardware and software is especially important. Arm reports 1.45 times the geometric-mean search queries per second on Cobalt 200 versus Cobalt 100. Separately, it reports an average eightfold gain, reaching 14-fold on specified 4-bit distance-calculation workloads, when comparing DiskANN version 0.61.0 with 0.56.0. The larger multiplier is a software-version comparison, not a claimed eightfold generational CPU improvement.
Arm names Amadeus, Databricks, InMobi, OpenAI, SAP and Snowflake among customers already benefiting from Cobalt 200 workloads. It also says Microsoft plans for the processor to account for a significant portion of new Azure compute capacity in 2027 and beyond. General availability opens access now; that larger capacity commitment remains a plan.
Sources
blogs.microsoft.comIgnite 2026: Building the frontier ecosystem - The Official Microsoft Blog
newsroom.arm.comHow Arm Neoverse CSS-based Microsoft Cobalt 200 virtual machines accelerate cloud and agentic AI performance
newsroom.arm.comMicrosoft Azure Cobalt 200 VMs reach general availability, bringing Arm-powered performance to cloud and agentic AI workloads
Reader comments
Newest comments first. Replies stay oldest first.