"This is an AI-powered audio summary and responses are AI-generated."
The rise of agentic AI is placing new demands on cloud infrastructure, driving the need for greater performance, efficiency, and scalability. As outlined by Mohamed Awad, EVP of Arm’s Cloud AI Business Unit, the availability of Microsoft Cobalt 200 virtual machines (VMs), powered by the Arm Neoverse Compute Subsystems (CSS) V3 platform, is designed to address these challenges, and accelerate cloud and agentic AI performance.
Building on the success of Microsoft Azure Cobalt 100, Cobalt 200, Microsoft’s second-generation Arm-based silicon, delivers substantial generational performance improvements across databases, web infrastructure, and compute-intensive applications. Cobalt 200 also highlights the value of Neoverse CSS in helping cloud providers accelerate custom silicon development and bring differentiated infrastructure to market faster.
Cobalt 200: The next generation of Arm-based Azure compute infrastructure #
Built on Arm Neoverse CSS N2, Cobalt 100 established a strong foundation for Arm-based computing in Azure, with significant performance and price-performance advantages across widely deployed cloud-native workloads. Customers such as Databricks, OneTrust, Snowflake, Sprinklr and others are running critical applications on Cobalt 100 and benefiting from significant price- performance advantages.
Through the adoption of Neoverse CSS V3, Cobalt 200 brings a new generation of compute performance to Azure. The new Dpsv7, Dplsv7, and Epsv7 VM families are designed to support a broad range of workloads, from distributed applications and databases to emerging AI infrastructure. With configurations scaling up to 128 vCPUs and 512 GiB of memory, Cobalt 200 provides additional flexibility for enterprises deploying demanding cloud applications.
Leading software partners such as Elastic and GitHub are among the customers that have reported significant performance efficiency improvements with Cobalt 200 and rolling out these benefits to their end customers.
Cobalt 200 delivers significant performance improvements across cloud-native workloads #
To evaluate the performance of Cobalt 200, Arm benchmarked commonly deployed workloads against contemporary x86-based Azure instances powered by AMD EPYC Turin instances.
As shown in the graphic above, the Arm-powered Microsoft Cobalt 200 delivers strong performance gains versus x86. These include:
- Up to 139% higher MySQL performance vs. AMD EPYC Turin for transactional databases supporting agent tools and backend services.
- 56% higher NGINX performance vs. AMD EPYC Turin and for web serving, secure networking, and API front ends.
- 65% higher Redis performance vs. AMD EPYC Turin for in-memory caching and data retrieval.
- Up to 56% faster x265 4K video encoding vs. AMD EPYC Turin for video processing and compression.
Together, the results demonstrate substantial generational improvements and strong competitive performance across database, web infrastructure, and compute-intensive applications.
Cobalt 200 performance gains across agentic AI benchmarks #
The Arm-powered Microsoft Cobalt 200 also delivers strong AI reasoning, planning, and agentic AI workflow performance versus AMD EPYC Turin. These include:
- Up to 56% higher Milvus HNSW vector search performance vs. AMD EPYC Turin for vector database retrieval and similarity search supporting agentic AI applications.
- Up to 57% higher fastText performance vs. AMD EPYC Turin for word-representation.
- Up to 46% higher SWE-bench task completion performance vs. AMD EPYC Turin for end-to-end agentic software engineering workflows at 144 concurrent workers.
Cobalt 200 performance for EDA workloads #
EDA tools are used to simulate and verify digital designs. RTL simulation is central to functional verification and can consume significant compute capacity as teams run large regression suites. Arm’s tests show that Cobalt 200 delivers, on average, 1.48x faster performance than AMD EPYC Turin across three EDA simulation workloads, as shown in the graphic below.
Agentic AI in action: Customer service application on Cobalt 200 #
To demonstrate how Cobalt 200 can support emerging agentic AI applications, Arm developed an application showcasing an agentic workflow running on Cobalt 200 VMs (as shown in the video below).
The application uses a Cobalt 200-hosted small language model to classify incoming customer requests and route them to specialized agents handling meeting scheduling, customer complaints, product inquiries, and refunds. These agents combine CPU-based inference, enterprise data retrieval, external model APIs, and tool execution to complete multi-step workflows, including scheduling meetings, generating personalized product brochures, and initiating refund processing through enterprise systems.
As organizations move from experimentation toward production deployment of agentic AI, infrastructure capable of efficiently supporting these interconnected services will become increasingly important. Cobalt 200 provides a compelling foundation for these emerging applications.
Accelerating vector search for agentic AI with Microsoft DiskANN #
One of the key requirements for agentic AI is efficient retrieval of relevant information from large datasets. Retrieval-augmented generation (RAG), semantic search, and agent memory increasingly depend on vector search technologies capable of efficiently processing high-dimensional embeddings. Microsoft’s DiskANN technology provides a foundation for scalable approximate nearest-neighbor search and supports vector retrieval capabilities across multiple Microsoft data services.
Arm and Microsoft have been collaborating to optimize DiskANN for Arm-based infrastructure, including improvements to vector similarity computations. Recent engineering work introduced Arm NEON and dot-product optimizations for 4-bit quantized vector operations, accelerating an important computational component of vector search.
Microsoft DiskANN delivers 1.45x higher geometric-mean vector-search QPS on Cobalt 200 versus Cobalt 100. Arm-optimized software boosts QPS by 8X on average and up to 14x on u4xu4 squared-L2 workloads, comparing DiskANN v0.61.0 with v0.56.0.
This collaboration demonstrates how Arm’s hardware architecture and software optimization capabilities can work together to accelerate important building blocks of agentic AI infrastructure.
Arm at Microsoft Ignite 2026 #
Arm will be at Microsoft Ignite in San Francisco (booth 6247R ) showcasing the performance benefits of Cobalt 200 VMs, as well as the Arm MCP server and Arm Performix in action through live demos.
We also have the following resources to help you get started on the journey to Cobalt 200:
-
Migration to Cobalt 200 with Arm Azure specific Learning Paths: Streamline the transition to Cobalt 200 instances with comprehensive guides and best practices.
-
Arm Cloud Migration : Organizations can work with Arm and Microsoft to assess existing applications, identify suitable workloads for migration, and evaluate performance and efficiency improvements on Azure Cobalt virtual machines.
-
Arm Software Ecosystem Dashboard : Stay up-to-date on the latest software support for Arm. Whether migrating existing applications or developing new AI-powered workloads, Arm and Microsoft are working together to help customers take advantage of the performance, scalability, and efficiency of Arm-powered custom silicon.
For all details about the Cobalt 200 GA benchmark testing methodology download the PDF file below. Any re-use permitted for informational and non-commercial or personal use only.