Azure Cobalt 200 virtual machines are now generally available, with Microsoft saying they deliver up to 50% better performance than the previous Cobalt 100 generation. The company announced the launch at Ignite on November 17, and the VMs are live in 14 regions today, expanding to 23 regions by the end of 2026.
Cobalt is Microsoft’s own Arm-based server processor. This generation is aimed at cloud-native, data-intensive and agentic AI work, such as the sandboxes and tool execution that AI agents rely on. Microsoft says Cobalt 200 VMs will make up a significant share of new Azure compute capacity beginning in 2027.
What is inside Cobalt 200?
The Cobalt 200 chip is built on TSMC’s 3nm process and combines 132 Arm Neoverse V3 cores. Each vCPU is a full physical core with its own private L2 cache, backed by 192 MB of system-level cache. Memory runs through 12 channels of DDR5, and memory encryption is on by default.
Every VM includes the latest Azure Boost, which moves networking and storage work off the host. Against Cobalt 100, Microsoft lists up to 15% more network bandwidth, 20% more remote storage IOPS and 10% more remote storage throughput. Every server also carries the Azure Integrated HSM for protected key storage.
Built-in accelerators for compression and cryptography take common tasks off the CPU. Microsoft says they can cut Azure SQL backup times by up to 2.5 times.
How much faster is it?
These are Microsoft’s own benchmark figures against Cobalt 100:
- Web serving: up to 1.5 times
- Java workloads: up to 1.6 times
- AI agent tool execution: up to 1.5 times
- Databases: 1.9 times
Customers quoted in the announcement include Amadeus, Databricks, InMobi, OpenAI and Snowflake. Amadeus said flight search throughput improved by more than 30% in its benchmark lab. Microsoft’s own Teams media processing service saw 1.65 times the performance.
What can you choose?
Cobalt 200 VMs come in five series across general purpose, memory optimized, high memory and storage optimized families. Two families are new for Arm on Azure:
- Mpsv4 is the high-memory family, with 16 GiB of memory per vCPU, aimed at large in-memory databases, ERP systems and caches.
- Lpsv5 is storage optimized, with up to 23 TB of local NVMe storage per VM for data staging, search and indexing.
There are also two processor modes. Deterministic Mode holds cores at 3.2 GHz for steady performance. Opportunistic Mode can boost to 3.6 GHz when power and thermal headroom allow, which Microsoft says gives up to 12% more performance with some variation in frequency. You pick the mode with the ProcessorMode setting in the Azure portal, Azure CLI, PowerShell or templates.
Prices
Microsoft’s announcement lists no prices for Cobalt 200 VMs, in any currency. Pricing depends on the VM size and the region, so check Azure’s pricing calculator or your Microsoft account team before planning a migration.
Who it suits
Teams running web and API services. Opportunistic Mode and the web serving gains suit services where speed matters more than steady timing.
Database and analytics teams. The Mpsv4 and Lpsv5 families, plus the database result of 1.9 times, are aimed at memory-hungry and storage-heavy workloads. Azure Database for PostgreSQL on Cobalt 200 was announced at Build 2026.
Builders of AI agents. Microsoft pitches the chips for agent orchestration, tool sandboxes and vector search, where per-core speed can cut latency. Note that these VMs are CPUs, not AI accelerators.
Who should wait. Anyone tied to x86-only software needs to confirm Arm support first, and customers outside the 14 launch regions will have to wait for the rollout to reach them.
What is still unknown
The announcement does not list which regions are live today or the pricing per VM size, and all performance numbers are Microsoft’s own tests with results that vary by workload. The next step is the VM size and deployment guidance linked from the Azure blog post, which has the full series table.






