Microsoft makes Cobalt 200 virtual machines generally available


The Azure machines use Microsoft’s second-generation Arm-based processor for workloads including the tool execution and isolated environments used by superintelligence (SI) agents.
Microsoft made its Cobalt 200 virtual machines generally available at its Ignite conference on 17 November 2026. The machines give Azure customers access to computing capacity powered by the company’s second-generation Arm-based processor, including for data-intensive applications and the work behind SI agents.
A virtual machine is a software-based computer running on shared physical hardware. Microsoft announced an early-access preview of the Cobalt 200 machines on 2 June 2026, after introducing the processor in 2025.
Microsoft says the machines are available in 14 Azure regions and can offer up to 50% better performance than Cobalt 100, depending on the workload. Its announcement does not list prices.
Microsoft positions Cobalt 200 for tasks such as file searches, compiling code and running tools in sandboxes—isolated environments where programs can execute. The processor handles that computing work rather than serving as a dedicated SI accelerator.
Pat Stemen, Microsoft’s vice president for Azure Cobalt, told The New Stack that Cobalt 200 uses a performance-optimized core design instead of Cobalt 100’s cost-optimized design, with larger caches and improved memory bandwidth. He said Microsoft built internal tests for common agent operations and tuned the product against them.
Microsoft says its Teams video-processing service saw 1.65 times the performance of Cobalt 100. Separately, Amadeus said its benchmark lab measured more than 30% higher flight-search throughput than on Cobalt 100. These are tests reported by Microsoft and a customer, respectively.
Customers can choose a consistent processor speed of 3.2 GHz or a variable mode that can run at up to 3.6 GHz when power and temperature allow. Microsoft says the variable mode can provide up to 12% more performance. Pat Stemen says the extra performance is typically available until the machine’s cores are heavily loaded. It recommends the consistent mode for distributed data and analytics, where the slowest machine can set the pace, and the variable mode for uses such as web services and media processing.
Alongside general-purpose and memory-optimized machines, Cobalt 200 adds a high-memory Mpsv4 family and a storage-optimized Lpsv5 family. Mpsv4 offers 16 GiB of memory per virtual CPU for uses including large in-memory databases; Lpsv5 offers up to 23 TB of local NVMe storage for tasks such as search and databases.
Developers can deploy the machines through the Azure portal, SDKs, APIs, PowerShell or Azure CLI, and use them in Virtual Machine Scale Sets or as Azure Kubernetes Service node pools. They can use the ProcessorMode setting to choose or change the processor mode at creation or later through the Azure portal, Azure CLI, PowerShell, or Azure Resource Manager and Bicep templates. Microsoft says workloads already running on Cobalt 100 are compatible with Cobalt 200. It points to Arm-native language platforms and container images, and says Azure Kubernetes Service supports Arm nodes and mixed x86 and Arm clusters.
The launch regions are Australia East, Canada Central, Central US, East US, East US 2, Indonesia Central, Italy North, Korea Central, South Central US, Spain Central, Sweden Central, UK South, West US 2 and West US 3.
Microsoft plans to expand availability to 23 regions by the end of 2026. It also expects Cobalt 200 to make up a significant portion of new Azure Compute capacity beginning in 2027, while continuing to invest in x86 chips.
The project says the denial-of-service attack has affected most relay operators and overwhelmed relays. It urges operators to install the update as soon as possible.
2 min

Cognition says the Bots can delegate software projects to Devin agents over days or weeks, while teams decide which changes need human review and which can be approved automatically.
3 min

The researchers attribute 21–31 million metric tons of carbon dioxide to electricity used by 403 hyperscale facilities between May 2024 and April 2025, using regional power mixes and assumptions about operating loads.
3 min