FR
live

Azure Cobalt 200 reaches general availability and puts Arm at the center of agentic cloud compute

On November 17, 2026, Microsoft announced general availability for Azure Cobalt 200 VMs, its second-generation in-house Arm processor, up to 50% faster than Cobalt 100. If you still run cloud-native or database workloads on x86, price out the Arm migration before the end of the year.

A dark server motherboard resting on a workbench, a single amber processor lit at its center.

November 17, 2026. At Ignite 2026, Microsoft announced general availability for Azure Cobalt 200 VMs, its second-generation in-house Arm processor. 14 regions. The number of Azure regions where these VMs can be deployed today, rising to 23 by the end of 2026. 132 Arm Neoverse V3 cores. The SoC built on TSMC’s 3nm process at the heart of every machine. Why it matters: OpenAI, Snowflake, and Databricks already run production workloads on it — the shift from x86 to Arm in the cloud has stopped being a bet and become a cost decision.

Arm leaves the niche and becomes the default compute

Three years ago, choosing an Arm VM in the cloud was still an experiment: container images were missing, some binaries only existed for x86, and infrastructure teams treated the architecture as a risk. Cobalt 200 closes that era. Cobalt 100, the first generation, is already deployed in 32 Azure regions and underpins Databricks, Snowflake, Amadeus, Siemens, and Temenos. Inside Microsoft itself, Cobalt 100 delivers up to 45% more performance while using 35% fewer compute cores than the previous platform, and Microsoft Defender for Endpoint gained 40% on its cyber-data pipeline.

What changes with the second generation is the scale. Cobalt 200 doubles down on every axis that matters for production: up to 50% better CPU performance generation over generation, 20% higher remote storage IOPS with NVMe, 10% better remote storage throughput, and 15% higher network bandwidth. The numbers are not a marketing average — they vary by workload, but the direction is unanimous across databases, web serving, and encryption.

What the silicon actually changes

The Cobalt 200 SoC is not a simple refresh. It packs 132 Arm Neoverse V3 cores — the highest-performance V-series core in Arm’s portfolio — in a multi-chiplet design fabricated on TSMC’s 3nm N3P process. Each core is a full physical core with 3 MB of private L2 cache, and the chip carries 192 MB of system-level L3 cache. The custom memory controller runs 12 channels of DDR5 and enables memory encryption by default with negligible latency impact.

Two points deserve an infrastructure engineer’s attention. First, memory encryption by default: unlike x86 VMs, where protecting memory in the clear often requires a paid option or AMD SEV / Intel TDX configuration, every Cobalt 200 VM boots with its memory encrypted. Second, an Azure Integrated HSM ships on every server, making hardware key protection and intrusion detection properties of the machine rather than add-ons.

The integrated compression and cryptography accelerators change the economics of databases: Azure SQL uses them to cut backup times by up to 2.5× while preserving the live workload. This is the least visible — and most decisive — part of Microsoft’s bet: moving repetitive work onto specialized low-power silicon instead of paying for it in general-purpose CPU cores.

Real-workload numbers, not synthetic benchmarks

The gains announced on real workloads speak louder than the headline “+50%”: 1.9× on databases, 1.5× on web serving, 1.6× on Java workloads, 1.5× on AI-agent tool execution, 1.65× on Microsoft Teams media processing. Microsoft’s silicon teams even report between 1.3× and 4.15× on their EDA design-and-verification workloads — a figure that matters, because those same workloads design the next generation of Maia and Cobalt chips.

Customer quotes confirm the operational translation. OpenAI runs ChatGPT on Cobalt: “we saw higher throughput and lower latency on Cobalt 200. That means we can serve more users with the same number of cores.” Amadeus measures 30% more flight-search throughput than on Cobalt 100. InMobi runs roughly 50,000 inferences per second on 32-core machines using SIMD, describing per-machine efficiency gains that compound across a fleet. Snowflake and Databricks report the same trajectory from their v5 VMs through Cobalt 100 to Cobalt 200.

The x86-to-Arm switch is now a cost decision

The question is no longer “can Arm hold the workload?” but “what does staying on x86 cost?”. Cobalt 200 is fully compatible with workloads already running on Cobalt 100, making generation-to-generation migration transparent — and migration from x86 remains the only real project. The major development platforms (C++, .NET, Java, Python, Rust) ship Arm-native versions, GitHub Actions supports hosted Arm runners, and AKS accepts Arm nodes as well as mixed x86/Arm clusters.

The new VM families now cover most of the spectrum. Dpsv7/Dpdsv7 (memory-to-vCPU ratio 4:1) for web and application servers, Epsv7/Epdsv7 (8:1) for databases, Dplsv7 (2:1) for microservices, and two additions: Mpsv4 (16:1, up to 84 vCPU) for in-memory databases and ERP, and Lpsv5 with up to 23 TB of local NVMe for data staging and search. All scale to 128 vCPU, with 85 Gbps of network bandwidth and 70 Gbps of remote storage throughput through Azure Boost.

Cost is the argument that tips the decision. At equal performance, Arm remains structurally cheaper than x86 when buying capacity — that is why AWS Graviton and then Azure Cobalt have eaten into the fleet, and why Microsoft announces that Cobalt 200 will represent “a significant share of new Azure Compute capacity” starting in 2027. When the cloud provider itself steers its capacity expansion toward Arm, the price differential mechanically flows into VM pricing.

The verdict

If you are deploying new cloud-native, database, or data-pipeline workloads on Azure, test Cobalt 200 now — migration from Cobalt 100 is transparent and the 50% gain translates directly into capacity or into your bill. The Dpsv7, Epsv7, and Lpsv5 families cover nearly every common use case, and memory encryption by default is a security bonus you do not have to provision.

If your workload is a Windows application or a closed binary compiled only for x86, stay on x86 until the vendor ships an Arm-native build — compatibility exists, but a legacy component without an Arm build remains a point of friction. If you are already on AWS Graviton, the logic is identical: the custom-silicon duel already happened, and the winner is the architecture, not a vendor. The general availability of Cobalt 200 on November 17, 2026 does not create the shift — it makes it irreversible.

References

The cyber brief, every Tuesday

The flaws that matter and the patches to apply, in a ten-minute read.

No spam. One-click unsubscribe.
read next

On the same topic

AWS Lambda adds GPU instances and brings inference into serverless

On November 18, 2026, AWS let you run Lambda functions on EC2 G-series instances, with no driver, framework, or scaling management, using ready-made vLLM, SGLang, and PyTorch images. For intermittent inference and batch media processing, serverless GPU is finally workable.

Google Cloud and Azure link their clouds in minutes through Cross-Cloud Interconnect

On November 18, 2026, Google Cloud extended its partner Cross-Cloud Interconnect to Microsoft Azure in preview, offering a private encrypted link provisioned in minutes across three regions. For a team running both GCP and Azure, physical interconnection is now a choice rather than an obligation.

← Back to the feed

Type at least two characters.

↑ ↓ navigate ↵ open esc dismiss