Microsoft Cobalt 200 VMs use Arm Neoverse CSS V3 for agentic AI

Arm announced on 17 November 2026 that Microsoft Cobalt 200 virtual machines (VMs), powered by the Arm Neoverse Compute Subsystems (CSS) V3 platform, provide generational performance improvements for cloud and agentic AI workloads. (CSS: compute building blocks that serve as a modular component of the Arm Neoverse architecture.) (agentic AI: autonomous AI agents capable of planning, reasoning, and executing multi-step workflows.) The new VM families—Dpsv7, Dplsv7, and Epsv7—scale up to 128 vCPUs and 512 GiB of memory. Cloud infrastructure architects evaluating x86-based Azure instances powered by AMD EPYC Turin can use these figures to determine if Arm-based silicon offers better price-performance for their specific workloads.

Is this for you?

Cloud architects and AI engineers can switch to Cobalt 200 if they need higher performance for agentic AI and database workloads over x86-based instances.

  • If you are a cloud infrastructure architect, Cobalt 200 offers up to 139% higher MySQL performance compared to AMD EPYC Turin.
  • If you are an AI infrastructure engineer, the platform provides up to 56% higher Milvus HNSW vector search performance.
  • If you are an EDA simulation team, Cobalt 200 delivers 1.48x faster performance than AMD EPYC Turin.
  • Skip this if you are a consumer PC builder as these are Azure cloud instances, not local hardware.
  • Skip this if you are a general AI user who only interacts with prompts or chat interfaces.

What you can do now

  1. Compare your current AMD EPYC Turin instance performance against the 139% higher MySQL performance offered by Cobalt 200.
  2. Download the Cobalt 200 GA benchmark testing methodology PDF to verify the reported performance claims.
  3. Check the Arm software ecosystem dashboard to verify application compatibility before migrating your workloads.

Key figures

MySQL performance gain
Up to 139% higher vs. AMD EPYC Turin
Milvus HNSW vector search
Up to 56% higher vs. AMD EPYC Turin
Max vCPU count
128 vCPUs
Max memory capacity
512 GiB of memory
DiskANN QPS improvement
1.45x higher vs. Cobalt 100

Arm claims Cobalt 200 accelerates cloud and agentic AI

Arm claims that Microsoft Cobalt 200 virtual machines accelerate cloud and agentic AI performance by utilizing the Neoverse CSS V3 platform. This platform provides the compute building blocks (CSS) for Microsoft’s second-generation Arm-based silicon. The architecture targets the requirements of agentic AI—autonomous AI agents capable of planning, reasoning, and executing multi-step workflows—which demand higher efficiency in data retrieval and tool execution than standard inference.

Evidence for this claim includes specific performance

Evidence for this claim includes specific performance gains over contemporary x86-based Azure instances. Arm reports that Cobalt 200 delivers up to 139% higher MySQL performance, 56% higher NGINX performance, and 65% higher Redis performance compared to AMD EPYC Turin. For AI-specific tasks, Arm says the platform achieves up to 56% higher Milvus HNSW vector search performance and 46% higher SWE-bench task completion performance. (vector search: a method of retrieving information from high-dimensional embeddings based on similarity.)

The strongest argument against these claims

The strongest argument against these claims is the lack of independent, third-party benchmarks verifying these figures in production environments. These results are currently based on Arm’s own testing methodology. If the performance gains do not scale linearly when running complex, multi-tenant production workloads, the practical advantage over x86 could diminish.

The claim would turn out wrong if the overhead of software translation or specific library optimizations fails to materialize in high-concurrency scenarios. To verify these claims, users should download the Cobalt 200 GA benchmark testing methodology PDF provided by Arm.

How much faster is Cobalt 200 than AMD EPYC Turin for MySQL?

Cobalt 200 delivers up to 139% higher MySQL performance compared to AMD EPYC Turin for transactional databases. This throughput is specifically cited as a benefit for supporting agent tools and backend services that require high-speed data access. The performance gap is part of a broader suite of improvements that include 56% higher NGINX performance for web serving and 65% higher Redis performance for in-memory caching.

The following table summarizes the performance gains Arm reports against the AMD EPYC Turin baseline:

Cobalt 200 vs. AMD EPYC Turin Performance (Arm claims)
Workload Performance Gain
MySQL (Transactional Databases) Up to 139% higher
NGINX (Web Serving/API) 56% higher
Redis (In-memory Caching) 65% higher
x265 4K Video Encoding Up to 56% faster

Comparing Azure Cobalt 200 vs AMD EPYC for cloud workloads

Comparing Azure Cobalt 200 vs AMD EPYC highlights a shift toward Arm-based silicon for high-density cloud-native applications. The Cobalt 200 platform supports configurations scaling up to 128 vCPUs (virtual processors) and 512 GiB of memory, providing the flexibility needed for large-scale distributed applications. While x86 remains a standard, the Cobalt 200 VM families—Dpsv7, Dplsv7, and Epsv7—are designed to offer distinct price-performance advantages for specific niches like database throughput and web infrastructure.

Optimizing agentic AI with Milvus HNSW and SWE-bench

Agentic AI performance is boosted by up to 56% in Milvus HNSW vector search—a method of retrieving information from high-dimensional embeddings based on similarity search—compared to AMD E7 Turin. This improvement is critical for Retrieval-Augmented Generation (RAG) and agent memory, where fast retrieval of relevant data informs the reasoning of autonomous AI agents. Additionally, Arm reports up to 46% higher SWE-bench task completion performance, which measures end-to-end agentic software engineering workflows.

Accelerating EDA simulation with 1.48x performance gains

EDA (Electronic Design Automation) tools—used to simulate and verify digital designs—see a significant boost on Cobalt 200. Arm’s tests show that Cobalt 200 delivers, on average, 1.48x faster performance than AMD EPYC Turin across three EDA simulation workloads. Because RTL simulation can consume significant compute capacity during functional verification, these gains allow teams to run larger regression suites more efficiently.

Improving RAG vector search via Microsoft DiskANN

Microsoft DiskANN provides a foundation for scalable approximate nearest-neighbor search, and Arm has collaborated with Microsoft to optimize it for Arm-based infrastructure. This optimization includes Arm NEON (a media processing engine) and dot-product optimizations for 4-bit quantized vector operations. Arm states these software-hardware optimizations boost QPS (Queries Per Second) by 8X on average and up to 14x on u4xu4 squared-L2 workloads when comparing DiskANN v0.61.0 with v0.56.0. Furthermore, Microsoft DiskANN delivers 1.45x higher geometric-mean vector-search QPS on Cobalt 200 versus the previous Cobalt 100.

Who benefits from Arm-powered Azure instances

Cloud infrastructure architects should look at Cobalt 200 to find price-performance gains for large-scale production workloads; they can compare the performance gains of Cobalt 200 against current amd epyc turin instances for cloud workloads. AI infrastructure engineers should monitor the swe-bench task completion results to evaluate Cobalt 200 for agentic software engineering and compare milvus hnsw vector search throughput on Cobalt 200 against existing x86 hardware.

Backend developers can monitor database throughput as a primary signal for the Cobalt 200 production rollout; they should compare mysql performance on Cobalt 200 against the current database throughput to see if it justifies a migration. EDA simulation teams can evaluate the 1.48x performance gains for RTL simulation to reduce the time required for functional verification of digital designs.

Who should skip these cloud-native VMs

Consumer PC builders should skip this analysis because the focus is on Azure cloud instances rather than local hardware; they should instead look for Arm Neoverse general architecture news for general-purpose silicon updates. General AI users who only interact with prompts or chat interfaces should ignore these infrastructure specs; they would be better served by Microsoft Copilot documentation which covers the user-facing software rather than the underlying VM performance.

A fintech agent deployment scenario: worked example

Freya runs a fintech startup and is building a multi-step refund processing agent that requires retrieving customer data from a high-dimensional vector database and running a MySQL-backed tool. She chooses Cobalt 200 over their previous x86 setup because she needs to handle high-concurrency transactions and rapid vector lookups. Her system sees a 56% improvement in Milvus vector search and a 139% boost in MySQL performance, allowing the agent to complete complex workflows faster.

If Freya had stayed on the previous Cobalt 100, she would have seen only 1.45x the vector search QPS compared to what the Cobalt 200 offers. This difference highlights the importance of moving to the V3 platform for high-demand AI infrastructure.

Concrete applications for Cobalt 200

Applications for Cobalt 200 VMs
Application Primary Beneficiary
High-throughput transactional databases (MySQL) Database Administrators
Real-time in-memory caching (Redis) Backend Developers
RAG and similarity search (DiskANN) AI Infrastructure Engineers
Digital design functional verification (EDA) EDA Simulation Teams
4K video processing (x265 encoding) Media Streaming Engineers

Replacing x86 instances for high-throughput databases

Microsoft Cobalt 200 reduces the need for x86-based Azure instances powered by AMD EPYC Turin for specific cloud-native workloads. Because Cobalt 200 offers up to 139% higher MySQL performance and 65% higher Redis performance, it allows for higher transaction density and lower latency in database-heavy applications without requiring a larger footprint of x86 hardware.

Microsoft Cobalt 200: what to check, compare and watch

Azure dpsv7 vm family
The dpsv7 vm family is one of the new configurations powered by Cobalt 200. Cloud Infrastructure Architects: configure the dpsv7 vm family for high-density cloud-native applications.
Arm software ecosystem dashboard
The dashboard helps users track software compatibility for the Cobalt 200 platform. Cloud Infrastructure Architects: check the arm software ecosystem dashboard to verify application compatibility before migration.
Microsoft ignite
Microsoft Ignite 2026 is the event where Cobalt 200 performance is being showcased. Cloud Infrastructure Architects: attend Microsoft Ignite 2026 to see live demos of the Cobalt 200 infrastructure.

Microsoft Cobalt 200: questions the original report leaves open

What is the strongest argument against it?
The original report did not say what the strongest argument against the claim is.
What would prove the claim wrong?
The original report did not say what would prove the claim wrong.
When is Microsoft Ignite 2026?
The original report does not say.

Key terms

agentic AI
autonomous AI agents capable of planning, reasoning, and executing multi-step workflows using tools.
CSS
compute building blocks, a modular component of the Arm Neoverse architecture.
EDA
chip design simulation used to verify digital designs.
vector search
similarity search, a method of retrieving information from high-dimensional embeddings based on similarity.
vCPU
virtual processor, a virtualized central processing unit.

Source story: How Arm Neoverse CSS-based Microsoft Cobalt 200 virtual machines accelerate cloud and agentic AI performance, published by Arm on 17 November 2026.

Related articles

ExtremeSpec is reader-supported: some links on this page are affiliate links, and as an Amazon Associate ExtremeSpec earns from qualifying purchases at no extra cost to you. Learn more.

Similar Posts