Microsoft launches Cobalt 200 ARM chips for faster AI agent sandboxes
Microsoft’s new Cobalt 200 ARM-based virtual machines are now generally available on Azure, offering up to 50% better performance for AI agent workloads.
Microsoft has made its second-generation Arm-based Cobalt 200 virtual machines generally available on Azure. Announced at the Ignite conference in San Francisco, these chips are specifically tuned to accelerate the sandboxes and code execution environments that power modern AI agents.
What happened
The Cobalt 200 VMs are currently live in 14 Azure regions, with plans to expand to 23 regions by the end of 2026. This launch follows a six-month preview period and comes almost exactly one year after Microsoft first revealed the architecture. The company states that the new hardware delivers up to 50% better performance compared to the previous Cobalt 100 generation.
Pat Stemen, Microsoft’s vice president for Azure Cobalt, explained that the engineering team spent the last year refining the infrastructure for specific operational patterns. They focused on workloads common to AI agents, such as searching files, compiling code, and executing tasks within isolated sandboxes. Internal benchmarks indicate that tool execution for these agents is up to 1.5 times faster than on the prior generation.
This release also introduces two new VM families alongside the general-purpose options. The high-memory Mpsv4 family offers 16 GiB of memory per vCPU, catering to large in-memory databases. The storage-optimized Lpsv5 family provides up to 23 TB of local NVMe storage, addressing the needs of enterprise workloads that require significant local disk throughput.
How it works
The Cobalt 200 chip is built on TSMC’s 3nm process and features 132 Arm Neoverse V3 cores. Unlike previous designs that prioritized cost efficiency, this generation uses performance-optimized cores. Each virtual CPU maps directly to a full physical core, equipped with 3 MB of private L2 cache. The system also includes a substantial 192 MB of system-level cache and 12 channels of DDR5 memory to boost bandwidth.
A key feature of the Cobalt 200 is the ability to choose between two processor modes. Deterministic mode locks cores at a consistent 3.2 GHz, providing predictable performance essential for distributed data clusters where the slowest node dictates speed. Opportunistic mode allows cores to reach 3.6 GHz when thermal and power conditions permit, offering up to 12% more performance for web services and media workloads. Developers can select or switch these modes via the Azure portal or command-line tools.
The architecture also integrates Azure Boost, a technology that offloads networking and storage processing from the main host CPU. This separation results in up to 15% higher network bandwidth and 20% higher remote storage IOPS compared to Cobalt 100. By removing these overheads from the compute cores, more processing power remains available for the primary application logic.
Key details
- Cobalt 200 VMs are generally available in 14 Azure regions, expanding to 23 by late 2026.
- Performance is up to 50% better than Cobalt 100, with 1.5x speedup for AI agent tool execution.
- The chip uses 132 Arm Neoverse V3 cores on TSMC’s 3nm process, with each vCPU mapping to a physical core.
- Two operating modes are available: deterministic (fixed 3.2 GHz) and opportunistic (up to 3.6 GHz).
- Azure Boost offloading delivers 15% higher network bandwidth and 20% higher remote storage IOPS.
- New VM families include Mpsv4 for high memory and Lpsv5 for up to 23 TB local NVMe storage.
Why it matters
For engineers building AI agents, the bottleneck often lies not in the large language model itself, but in the surrounding tooling. Agents frequently need to search file systems, compile code snippets, or run scripts in secure sandboxes. These tasks are bursty and sensitive to latency. The Cobalt 200’s focus on large caches and fast memory bandwidth directly addresses these pain points, allowing agents to react more quickly and handle complex multi-step workflows without stalling.
The shift toward Arm-based infrastructure in the cloud is accelerating. Microsoft expects Cobalt 200 to constitute a significant portion of new Azure compute capacity in 2027. For development teams, this means that optimizing applications for Arm architectures is no longer a niche concern but a mainstream requirement. The improved tooling and proven performance metrics from services like Azure SQL and Microsoft 365 reduce the risk associated with migrating existing enterprise workloads.
Furthermore, the flexibility to choose between deterministic and opportunistic modes gives architects finer control over cost and performance trade-offs. Teams running distributed analytics can ensure consistent timing, while customer-facing APIs can leverage burstable speed during peak traffic. This level of granularity helps in designing systems that are both efficient and resilient.
What you can do
- Deploy Cobalt 200 VMs in any of the 14 currently supported Azure regions using standard Azure tools.
- Benchmark your AI agent sandboxes against Cobalt 100 to quantify the performance gains in tool execution.
- Select deterministic mode for distributed data clusters to avoid straggler nodes slowing down the group.
- Use opportunistic mode for web APIs and media services to capitalize on extra headroom during low load.
- Evaluate the new Mpsv4 and Lpsv5 VM families if your workload requires high memory or local NVMe storage.
- Plan for Arm compatibility in your CI/CD pipelines as Cobalt 200 becomes a larger part of Azure’s capacity.


