The 50% Generational Jump on Microsoft's In-House CPU
Microsoft has officially launched Cobalt 200, its second-generation Arm-based server processor, and the first Cobalt 200 servers are already live in Microsoft datacenters. Wider availability in Azure virtual-machine families is scheduled across 2026, with select SKUs landing in early 2026 and a broader rollout later in the year. The headline number — a "up to 50% performance improvement over Cobalt 100 across a blended set of real-world workloads," as Microsoft's infrastructure team puts it — is the kind of step that moves custom silicon from "interesting side project" to "serious lever on hyperscaler economics."
Cobalt 200 is built on TSMC's 3nm process. It packs 132 active cores arranged as two 66-core chiplets based on Arm's Neoverse Compute Subsystems V3 (CSS V3) platform — the same generation Arm has been shipping to other hyperscalers. Each core gets 3 MB of L2 cache; the chip backs them with 192 MB of system-level L3 cache. Per-core dynamic voltage and frequency scaling (DVFS) is the quieter headline: every one of the 132 cores can run at a different performance level, which means better energy efficiency when workloads are mixed.
Why This Matters Beyond Microsoft
Hyperscaler custom silicon used to be a curiosity. Today it is the operational reality: AWS Graviton has been the default for many workloads for years, Google has its Axion Arm chip, and Microsoft's Cobalt line is closing the gap. The common thread is that Arm-on-cloud is no longer a niche for embedded or developer-preview workloads — it is the cost-per-core and performance-per-watt baseline that x86 must meet.
A concrete Microsoft example: the team reports Teams media-processing workloads ran 45% faster on an earlier Cobalt test versus the previous compute platform, and needed 35% fewer compute cores to deliver the same capacity. Multiplied across production Microsoft services and customer VMs, the unit-economics shift is significant. Combined with the fact that more than 30% of typical cloud workloads spend measurable CPU time on compression and encryption, Cobalt 200's on-die compression and cryptography accelerators plus Azure Integrated HSM are a real TCO lever, not a marketing slide.
Advertisement
Security Is the Quietly Interesting Part
Cobalt 200 has memory encryption on by default with what Microsoft calls "negligible performance impact." It also supports Arm's Confidential Compute Architecture (CCA), extending confidential-computing enclaves to Arm servers in Azure. For customers running regulated workloads — financial services, healthcare, public sector — always-on memory encryption plus attested enclaves is the kind of feature that reshapes compliance conversations. For multi-tenant clouds, it is the foundation that lets hyperscalers tell regulators a coherent story about data-at-rest, data-in-use and data-in-transit protection without asking customers to re-architect.
What Customers Should Actually Do
Three things are worth acting on. First, if you already run production workloads on Azure D-series or similar VMs, ask your account team which workloads will benefit from migration to Cobalt 200 SKUs as they become generally available through 2026. Application-level testing is the only way to confirm; in general, Arm-friendly workloads (web serving, microservices, Java and Go applications, data-intensive streaming, in-memory caches) tend to see the biggest wins.
Second, treat memory encryption and CCA as platform capabilities you can actually use, not just check-box items. Architects building confidential-computing pipelines in Azure have more options in 2026 than they did last year.
Third, negotiate. Custom-silicon generations typically come with favourable pricing to drive adoption. Azure's public-sector and enterprise customers in 2026 should be asking what the Cobalt-200 list and committed-use discount look like versus comparable x86 SKUs, not taking the first quote.
Frequently Asked Questions
What exactly is Azure Cobalt 200?
Cobalt 200 is Microsoft's second-generation in-house server CPU: a 132-core chip built on TSMC's 3nm process using two 66-core chiplets based on Arm's Neoverse Compute Subsystems V3 platform. Each core has 3 MB of L2 cache, the chip provides 192 MB of L3 system cache, and per-core dynamic voltage and frequency scaling lets each core run independently for better efficiency. It succeeds Cobalt 100 and delivers up to 50% higher performance on blended real-world workloads according to Microsoft's internal telemetry.
When can customers actually use it?
Microsoft reports that the first production Cobalt 200 servers are already live in its datacenters. Select Azure VM families running on Cobalt 200 are scheduled to become generally available in early 2026, with a broader rollout across Azure regions continuing through the year. Customers in Azure European regions (Frankfurt, Amsterdam, Paris) will be among the earliest to access the new SKUs.
How does this compare to AWS Graviton and Google Axion?
Cobalt 200 is Microsoft's answer to what AWS has been doing with Graviton for years and Google has started with Axion. All three are Arm-based, targeted at cloud-native workloads, and designed to deliver better performance-per-watt than general-purpose x86 SKUs. The notable Cobalt 200 specifics are the 132-core density and the on-die encryption/compression accelerators combined with always-on memory encryption — the security posture is arguably the strongest among current hyperscaler Arm chips.
Sources & Further Reading
- Announcing Cobalt 200: Azure's next cloud-native CPU — Microsoft Tech Community
- Microsoft Azure Cobalt 200 Launched with 132 Arm Neoverse V3 Cores — ServeTheHome
- Microsoft unveils Azure Cobalt 200 CPU — Tom's Hardware
- Microsoft reveals next generation Cobalt CPU and Azure Boost — Data Center Dynamics
- Cobalt 200: Microsoft's next-gen Arm CPU targets lower TCO — Network World














