Azure Maia and Azure Cobalt
Azure Maia and Azure Cobalt are Microsoft-designed chips in Azure datacenters. Maia accelerators run AI inference for Microsoft services, with no direct customer rental stated. Cobalt Arm CPUs power VM sizes customers can deploy: Cobalt 100 sizes such as Dpsv6, and Cobalt 200 sizes in early access preview.12
Also known as Microsoft Azure Maia, Microsoft Azure Cobalt, Maia 200, Cobalt 100, Cobalt 200, Microsoft custom silicon
- Intended for
- Cloud architects, Infrastructure and FinOps teams, Developers porting workloads to Arm, Technology buyers tracking AI infrastructure
- Runs on
- Azure datacenters (Maia 200 in the United States, Europe and Australia, for Microsoft services), Azure regions offering Cobalt 100 VM sizes, Eight preview regions for Cobalt 2002
At a glance
- What is it?
- Microsoft designs two families of silicon for Azure. Maia is an AI accelerator line. Maia 200, announced on 26 January 2026, is built on TSMC's 3-nanometer process for inference, and Microsoft's 12 October 2026 post says it is in production in the United States, Europe and Australia, serving Microsoft workloads including Microsoft Copilot. Cobalt is an Arm-based CPU line. Cobalt 100 powers general-purpose and memory-optimized VM sizes marked with a p in the size name, and Cobalt 200 VMs entered early access preview at Microsoft Build 2026. Microsoft also states that Maia 300 production deployments begin in 2027.134
- What does it do?
- Maia runs model inference for Microsoft services. Microsoft says it serves OpenAI models, its MAI models and open-source model families, with GitHub Copilot and Microsoft Foundry expected to follow by the end of 2026. A Maia SDK with PyTorch integration and a Triton compiler is in preview for developers who sign up. Cobalt-based VMs run Linux and Windows client workloads such as web and application servers, small to medium databases and caches; Dpsv6 sizes offer 2 to 96 vCPUs with 4 GiB of memory per vCPU and a full physical core per vCPU.135
- Who needs it?
- Cobalt: teams running scale-out workloads that can run on Arm64 and want an alternative to x86 VM sizes. Maia: mainly architects and buyers who want to understand the infrastructure behind Microsoft Copilot and, later, Foundry, plus developers interested in the Maia SDK preview. The Microsoft posts checked do not describe a way for customers to rent Maia capacity directly.
- What does it need?
- For Cobalt VMs: Arm64 builds of your operating system image and application
- VM-family vCPU quota in a region that offers the Cobalt size
- For Cobalt 200: access to the early access preview and one of the eight preview regions
- For Maia: nothing for service users; the Maia SDK preview requires sign-up
- What it is not
- Maia is not a VM size or GPU alternative you can deploy today; the posts checked describe it serving Microsoft's own products and do not offer it as a customer SKU. Cobalt 200 is not generally available; it is an early access preview in eight named regions. Microsoft's performance comparisons for both chips are vendor claims without independent verification in these sources. Cobalt sizes need Arm64 builds of your operating system and software.2
Availability and licensing. Mixed. Cobalt 100 VM sizes such as Dpsv6 are listed on Learn without a preview label (checked 26 Oct 2026). Cobalt 200 VMs are in early access preview in eight regions, announced at Build 2026. Maia 200 is in production for Microsoft services; the posts checked do not offer it for direct customer rental. The Maia SDK is in preview by sign-up. Maia 300 production deployments begin in 2027.1
The problem it solves
AI demand raises the cost and power needed for each token served, and cloud-native CPU workloads face similar pressure on price and energy. Microsoft's stated response is to design chips, racks and software together for its own datacenters.
For customers this shows up in two ways: Arm-based Cobalt VM sizes they can choose for suitable workloads, and Maia capacity working behind Microsoft services such as Microsoft Copilot.
How it works
Maia 200 (Microsoft's figures): TSMC 3 nm, 216 GB of HBM3e, native FP8 and FP4 tensor cores, and Ethernet-based scale-up networking with a custom transport layer, supporting clusters of up to 6,144 accelerators. First deployed in the US Central region near Des Moines, Iowa, followed by US West 3 near Phoenix, Arizona.
Maia 300 (announced 12 October 2026): production deployments from 2027, racks of 96 accelerators, compute chiplets on TSMC N2P.
Cobalt 100: Arm64 VM sizes such as Dpsv6 with up to 96 vCPUs and 384 GiB of memory; a p in the size name marks an Arm CPU (Microsoft Cobalt or Ampere Altra).
Cobalt 200: early access preview VM families Dplsv7, Dpsv7, Epsv7, Mpsv4 and Lpsv5 (with local-disk variants).1235
Diagram as a list
Copilot & business apps
- Microsoft services (Microsoft Copilot; Foundry and GitHub Copilot planned)Consume Maia inference capacityConnects to Azure Maia 200 accelerators
- Customer Arm64 workloadsWeb, application, database and cache serversConnects to Cobalt VM sizes (Dpsv6; v7 families in preview)
Developer tools
- Maia SDK (preview)PyTorch integration and Triton compilerConnects to Azure Maia 200 accelerators
Cloud & hybrid infrastructure
- Azure Maia 200 acceleratorsRun model inferenceConnects to Ethernet scale-up network with custom transport
- Ethernet scale-up network with custom transportConnects accelerators in clusters
- Azure Cobalt 100 and 200 Arm CPUsGeneral-purpose cloud compute
- Cobalt VM sizes (Dpsv6; v7 families in preview)VM sizes customers deployConnects to Azure Cobalt 100 and 200 Arm CPUs
Capabilities
Maia 200 inference accelerator3
Microsoft's second-generation AI accelerator, built for inference, with HBM3e memory and FP8 and FP4 tensor cores.
Why it matters: Microsoft says it lowers the cost of token generation for its services.
Limits: Performance and cost figures are Microsoft claims; no customer SKU is described.
Maia in production for Microsoft services1
Maia 200 serves Microsoft workloads including Microsoft Copilot and runs OpenAI, MAI and open-source models.
Why it matters: Explains part of the infrastructure behind Microsoft AI services.
Limits: GitHub Copilot and Microsoft Foundry are stated to follow by end of 2026, not yet confirmed as running on Maia.
Maia SDK (preview)3
An SDK with PyTorch integration, a Triton compiler, an optimized kernel library and a low-level programming language, plus a simulator and cost calculator.
Why it matters: Lets developers, startups and academics experiment with the chip.
Limits: Preview by sign-up; no general availability or pricing timeline stated.
Cobalt 100 VM sizes5
Arm64 VMs such as Dpsv6, with 2 to 96 vCPUs, 4 GiB of memory per vCPU and a full physical core per vCPU.
Why it matters: An Arm option for web servers, application servers, small to medium databases and caches.
Limits: Requires Arm64 software; Dpsv6 has no local temporary disk.
Cobalt 200 early access preview2
New Arm VM families (Dplsv7, Dpsv7, Epsv7, Mpsv4, Lpsv5) announced at Build 2026, which Microsoft says offer up to 50% better CPU performance than Cobalt 100.
Why it matters: Next-generation Arm option for general-purpose, memory and storage workloads.
Limits: Preview in eight named regions; performance claims are Microsoft figures that vary by workload.
Maia 300 roadmap1
Next-generation accelerator with production deployments starting in 2027 and racks of 96 accelerators.
Why it matters: Signals where Microsoft plans to run inference for its products from 2027.
Limits: A roadmap statement; details may change.
Practical use cases
A web and API tier runs on many x86 VMs and the team wants to test a different price-performance option.
- Approach
- Move a stateless tier to Cobalt 100 Dpsv6 VMs.
- Role of Azure Maia and Azure Cobalt
- Cobalt provides the Arm CPU capacity.
- Data, infrastructure and skills
- Arm64 builds of the application and its dependencies.
- Type of benefit
- Price-performance choice
- Caveats
- Porting effort; compare cost and performance on your own workload.
- First step
- Deploy one Dpsv6 VM behind a test load balancer.
A database or caching team wants to evaluate the next Arm generation early.
- Approach
- Join the Cobalt 200 early access preview in a preview region.
- Role of Azure Maia and Azure Cobalt
- Cobalt 200 VMs provide the test capacity.
- Data, infrastructure and skills
- Preview access and a workload in one of the eight preview regions.
- Type of benefit
- Early hardware evaluation
- Caveats
- Preview status; Microsoft performance claims vary by workload.
- First step
- Benchmark a representative database workload on a Dpsv7 VM.
Sources 2
A kernel or compiler team wants to understand how models run on Maia.
- Approach
- Sign up for the Maia SDK preview and use the simulator and PyTorch integration.
- Role of Azure Maia and Azure Cobalt
- The SDK exposes the Maia programming model.
- Data, infrastructure and skills
- Sign-up approval and PyTorch or Triton skills.
- Type of benefit
- Early hardware evaluation
- Caveats
- Preview only; no customer Maia capacity described.
- First step
- Apply for the Maia SDK preview.
Works with
Optional integration
- Microsoft 365 CopilotMicrosoft states Maia 200 serves Microsoft Copilot and built-in AI features in Teams, Copilot and Cowork.
- Microsoft FoundryMicrosoft states Foundry is expected to run on Maia 200 by the end of 2026 (planned).
- GitHub CopilotMicrosoft states GitHub Copilot is expected to run on Maia 200 by the end of 2026 (planned).
Complementary tools
- Azure GPU virtual machinesMicrosoft describes Maia as part of a heterogeneous AI infrastructure that also includes GPUs.
Relationship labels follow Microsoft's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Check Arm readiness
Confirm your OS image, runtimes and dependencies have Arm64 builds.
Check: The application builds and passes tests on Arm64.
Pick a Cobalt 100 size
Start with a Dpsv6 size that matches your vCPU and memory needs.
Check: The size offers the vCPU and memory you need.
Get quota and deploy
Request VM-family quota if needed and deploy a test VM.
Check: The VM runs your workload.
Compare cost and performance
Run the same workload on your current x86 size and on Cobalt and compare results and cost.
Check: You have your own measured comparison rather than vendor figures.
Evaluate Cobalt 200 if relevant
Request early access preview in one of the eight preview regions.
Check: Preview access is granted.
Track Maia
Follow Microsoft announcements for Foundry and GitHub Copilot on Maia, and sign up for the Maia SDK preview if you develop kernels.
Check: You know which of your services run on Maia, if any.
Official resources
Deciding whether you need it?
Three free tools apply Microsoft's own published guidance to your situation, and say when Azure Maia and Azure Cobalt or any new technology is not needed.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- Microsoft to expand Azure Maia system to multi-gigawatt deployments (opens in a new tab)
- New Azure Cobalt 200 VMs (Azure blog, Build 2026) (opens in a new tab)
- Maia 200: The AI accelerator built for inference (opens in a new tab)
- Virtual machine sizes overview (opens in a new tab)
- Dpsv6 size series (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.