Azure Cobalt 200 VMs vs Cobalt 100: Which Fits Small Teams
Short answer for small B2B software teams
Microsoft made Azure Cobalt 200 virtual machines generally available on November 17, 2026, announced at Microsoft Ignite 2026 (launch post). For a small B2B software company, the decision is narrow: keep your current VMs, or plan a move for Arm-ready workloads that are limited by per-core throughput, starting with a compatibility check rather than a purchase.
TL;DR: Cobalt 200 is Azure's second-generation Arm VM family, with Microsoft-reported gains of up to 50% over Cobalt 100. No Cobalt 200 list price appears on the pages reviewed, and a June 2026 Azure notice still described a private preview, so confirm your region and pricing before you commit budget.
Key Takeaways
Microsoft says Cobalt 200 VMs deliver up to 50% better performance than Cobalt 100. Treat that as a best-case, vendor-reported figure, not a typical result.
The launch post says the VMs are in 14 regions now and will expand to 23 by the end of 2026, so confirm your region is live before you plan.
Five series cover the range: Dpsv7 and Dplsv7 for general purpose, Epsv7 for memory-optimized, Mpsv4 for high memory, and Lpsv5 for local NVMe storage. Dplsv7 and Epsv7 are named in the June 2026 Azure Message Center notice (AZ564451); we did not find Learn size pages for them, so the size details in this post cover Dpsv7, Mpsv4 and Lpsv5 only.
The real migration work is in the constraints: Arm64 builds, NVMe-capable OS images, Generation 2 images, required accelerated networking, and no live migration on these series.
No Cobalt 200 list price appears on the pages reviewed, so any cost comparison stays Quote-based until Azure publishes SKU prices.
A June 2026 Azure Message Center entry described a private preview, which conflicts with the November launch post. Check availability in your region before you commit.
What Cobalt 200 is and what Azure launched
Azure Cobalt 200 is Microsoft's custom Arm-based processor for Azure virtual machines, and the Cobalt 200 series are the Azure VM sizes that run on it. Microsoft's Learn reference describes the Dpsv7 series as powered by "Azure's second-generation Cobalt 200 processor," based on Arm Neoverse V3 cores, with an entire physical core for each vCPU (Dpsv7 size reference).
The launch post says Azure Boost is included in every Cobalt 200 VM. It names two new families: Mpsv4 for high-memory workloads and Lpsv5 for workloads that need fast access to large datasets. It also says Azure Database for PostgreSQL became available on Cobalt 200 at Microsoft Build 2026. The post frames the chip around agentic AI workloads such as agent sandboxes, file search, code compilation and tool execution.
Headline claim: up to 50% better than Cobalt 100 according to Azure blog (2026). The same page says the VMs are in 14 regions now and will expand to 23 by the end of 2026.
For a small team, the change shows up in three places: which VM size you pick for a given load, which OS image you boot, and which regions you can deploy into. Those three decisions also set how much work a move takes, which is why this guide weighs them above the headline benchmark.
How we evaluated these options
This guide weights six criteria by what a small B2B software team actually decides on. The weights are this guide's judgment, not a vendor or analyst figure, and they are shown so you can change them.
| Criterion | Weight | Why it matters to a small team |
|---|---|---|
| Vendor performance evidence | 20% | Shows whether fewer vCPUs can serve the same load |
| Regional availability | 20% | Decides whether your users and data can be served from the region |
| Arm64 software compatibility | 20% | Sets how much code, container and driver work a move needs |
| Price transparency | 15% | Lets you budget; no Cobalt 200 list price was on the pages reviewed |
| Operational constraints | 15% | Covers live migration, Generation 1 images and NVMe image support |
| Migration and rollback effort | 10% | Decides whether a move can happen this month or should wait |
The weights sum to 100%. The two comparison targets are Cobalt 100, the previous generation and the comparison Microsoft draws itself, and your current VMs. The sources reviewed did not give x86 or current-VM benchmark numbers, so the current-VM side is handled as a checklist of compatibility and cost questions rather than as a number.
Cobalt 200 against Cobalt 100, feature by feature
Cobalt 100 is the closest comparison because Microsoft's launch post measures Cobalt 200 against it. The Register's report on Cobalt 100's general availability in October 2024 described the largest Cobalt 100 VMs as having 96 vCPUs and 384 GiB of memory, with 14 broadly available regions, according to The Register (2024).
General-purpose ceiling: 128 vCPUs and 512 GiB according to Microsoft Learn (2026). The Dpsv7 page also says these VMs have no local disk and suit web and application servers, small to medium databases and caches.
Lpsv5 is the local-storage option: it offers up to 23,040 GB of local NVMe temporary storage at 128 vCPUs (Lpsv5 size reference).
| Feature | Cobalt 200 | Cobalt 100 |
|---|---|---|
| Largest general-purpose VM (vCPUs) | 128 | 96 |
| Largest general-purpose memory (GiB) | 512 | 384 |
| Regions at general availability | 14 (November 2026) | 14 (October 2024) |
| Regions later | 23 planned by end of 2026 | 32 (see Signal vs Speculation) |
| Largest local NVMe capacity (GB) | 23,040 (Lpsv5, 128 vCPUs) | Not in sources reviewed |
| Network bandwidth ceiling (Mbps) | 85,000 (Dpsv7, 128 vCPUs) | Not in sources reviewed |
| Processor speed (GHz, deterministic / opportunistic) | 3.2 / 3.6 | Not in sources reviewed |
| Physical core per vCPU | 1 | Not in sources reviewed |
| Live migration | Not supported | Not in sources reviewed |
| Generation 1 images | Not supported | Not in sources reviewed |
Cobalt 200 figures come from the Learn size references for Dpsv7, Lpsv5 and Mpsv4. Cobalt 100 figures come from the 2024 Register report.
Two rows decide most of the migration work. The Learn pages say the Dpsv7, Mpsv4 and Lpsv5 series "will only work on OS images that support NVMe," so an older image needs checking before any resize. They also list Generation 1 VMs as not supported, so a Generation 1 image must be rebuilt or converted. Accelerated networking is required on all three series and is on by default.
Live migration is not supported on those three series, but memory-preserving updates are. Maintenance plans that assume live migration will not hold on these sizes. The high-memory Mpsv4 series reaches 1,388 GiB at 128 vCPUs and is aimed at large in-memory databases and ERP systems (Mpsv4 size reference).
Benchmarks: what the vendor measured and what it did not
Microsoft's benchmark multipliers against Cobalt 100 are up to 1.5x for web serving, up to 1.6x for Java-based workloads, up to 1.5x for AI agent tool execution and 1.9x for database workloads.
| Workload or measure | Cobalt 200 vs Cobalt 100 | Basis (2026 launch post) |
|---|---|---|
| Web serving | Up to 1.5x | Microsoft benchmark (2026) |
| Java-based workloads | Up to 1.6x | Microsoft benchmark (2026) |
| AI agent tool execution | Up to 1.5x | Microsoft benchmark (2026) |
| Database workloads | 1.9x | Microsoft benchmark (2026) |
| Flight search throughput at Amadeus | More than 30% | Customer-reported (2026) |
| Network bandwidth with Azure Boost | Up to 15% higher | Microsoft figure (2026) |
| Remote storage IOPS with Azure Boost | Up to 20% higher | Microsoft figure (2026) |
Every row comes from the same launch post. The Amadeus row is a customer's own result, quoted by Microsoft.
Read these figures with four limits in mind. They are "up to" figures, so they describe a best case. The post does not give test configurations, instance sizes or methodology. The 1.9x database figure is one workload, not a general promise. And the post gives no Cobalt 200 price-performance number, which is the figure a budget needs. The sources reviewed include no independent benchmark of Cobalt 200, so every number in this section is vendor-supplied or customer-supplied.
Clock behavior matters for planning too. On the Dpsv7 reference, Deterministic mode holds 3.2 GHz, and Opportunistic mode runs up to 3.6 GHz (Dpsv7 size reference).
Signal vs Speculation
Confirmed by Azure's own dated pages:
The launch post, dated November 17, 2026, announces general availability and names 14 regions.
The Learn size references for Dpsv7, Mpsv4 and Lpsv5 carry a November 17, 2026 update stamp and give the sizes, memory and limits used above.
Conflicting signal:
An Azure Message Center entry, AZ564451, published June 2, 2026, lists the same five series with a Private Preview timeline and no region list, according to Azure Message Center mirror (2026).
Our read: the November post is later and states general availability, so the June entry should be treated as superseded for the GA claim. The June notice still shows that series may roll out on different timelines, so confirm each series in your region before you plan.
Footprint trend:
Cobalt 100 footprint: 32 Azure datacenter regions (Tech Community announcement).
Our read: Cobalt 200 launched in 14 regions, the same count Cobalt 100 had at its general availability. The broader footprint for Cobalt 200 is still ahead, and this is interpretation, not vendor data.
Forward-looking claims are plans, not facts yet. These include 23 regions by the end of 2026 and Cobalt 200 representing "a significant share of new Azure Compute capacity beginning in 2027."
Market context, as opinion: an opinion column in Semiconductor Engineering in July 2025 estimated Arm's 2025 share of data-center CPUs at about 15%, against an Arm expectation that half of top hyperscalers' compute would be Arm-based, according to Semiconductor Engineering (2025). The author calls the figures "rough numbers" built on shipment assumptions, so use them as direction, not as a benchmark.
Our opinion: for a small team, the rollout footprint and image checks matter more than the 50% claim. A per-core gain on a workload you do not run is worth nothing. A region you can deploy into and an image that boots are what make the gain real.
Pricing and total cost: what is published
The launch post does not state a Cobalt 200 price. It points readers to general Azure Compute VM pricing. The Linux VM pricing page we opened (Azure Linux VM pricing) did not show Cobalt 200 or Arm rows in the text we could read. The Dpsv7 reference says disk storage is billed separately from virtual machines, so a disk-heavy design costs more than the VM line alone suggests.
Pricing checked November 17, 2026.
| Cost line | Cobalt 200 | Cobalt 100 | What decides it |
|---|---|---|---|
| Compute list price per hour | Quote-based (not on pages reviewed) | Quote-based (not on pages reviewed) | SKU rows on the pricing page or calculator for your region |
| Managed disks | Billed separately from the VM | Not restated in sources reviewed | Disk type and size |
| Reservations or commitments | Not published in sources reviewed | Not published in sources reviewed | Commitment length and region |
| Image and compatibility work | Your estimate | Your estimate | Arm64 builds, NVMe and Generation 2 images |
| Dual-running period for rollback | Your estimate | Your estimate | How long current VMs stay up during cutover |
Until Azure publishes SKU rows, mark any Cobalt 200 compute cost as Quote-based in your budget and compare it with a calculator run for your own region and size. For a broader view of trimming recurring SaaS tool spend, see how SaaS teams save on workspace tools.
Who this is for and who should wait
Who this is for: a small B2B software team that runs an Arm64-ready service, such as a container image built for arm64 or code with no x86-only libraries. It should have a capacity or latency problem on current VMs, deploy in a region on the current list, and have an engineer who can run the compatibility checks.
Red flags: a binary or driver with no Arm64 build; a required region that is not yet live; a workload that depends on live migration to stay up.
Azure Cobalt 200 VMs
Best fit: Arm64-ready web and application tiers, Java services, databases and agent tool execution, which are the workload types the launch post benchmarks. The Dpsv7 reference describes the same range for web and application servers, small to medium databases and caches (Dpsv7 size reference).
Limitations: Arm64 only, NVMe-capable OS images, Generation 2 images only, no live migration, 14 regions at launch with 23 planned, and no published Cobalt 200 list price on the pages reviewed.
Implementation: confirm the series and region on Azure's current size reference, validate Arm64 images in a non-production subscription, then cut over one service at a time with a rollback path to current VMs. Primary evidence: the launch post, the Dpsv7 size reference and the Lpsv5 size reference.
Azure Cobalt 100 VMs
Best fit: teams that have already validated Arm64 on the earlier generation and want to avoid a second move. Cobalt 100 also had the wider footprint reported so far, at 32 regions by late 2025.
Limitations: the largest general-purpose size is 96 vCPUs, and the sources reviewed give no Cobalt 100 price. Cobalt 200 adds requirements, including NVMe-capable and Generation 2 images, that an existing Cobalt 100 image may not meet, so check before any resize.
Implementation: an existing Cobalt 100 estate can stay as it is. A move to Cobalt 200 is a resize plus image change, not a like-for-like swap. Primary evidence: the Microsoft Tech Community announcement and the 2024 Register report.
Staying on your current VMs
Best fit: services that are not Arm-ready or are not capacity-bound. Limitations: you do not get the per-core gains the launch post describes. Implementation: revisit the decision once Azure publishes Cobalt 200 SKU prices and your region is confirmed.
A proposed workflow for the migration review
A proposed, configurable workflow built on US Tech Automations could handle the review work that slows a move down, without changing any VM by itself. The trigger is a weekly inventory export of your VMs, with name, size, OS image, region and tags, from your Azure subscription. The action flags every VM whose size belongs to a Cobalt series and creates one review task per VM. Each task lists the candidate Cobalt 200 size, the image generation check, the NVMe image check and the accelerated networking check. The output is a task list in your ticketing system, with each task assigned to a named engineer. Prerequisites: read access to the subscription inventory, an export or API path into your ticket tool, and a tag schema that names the owning team.
The human review points stay where they belong. An engineer confirms Arm64 compatibility for each application, decides whether a resize is worth it, and approves each change through your own change process. The workflow only creates tasks and records decisions. A second proposed workflow starts from a CPU alert: when a Cobalt VM's utilization stays above a threshold you set for a week, the workflow posts a summary with the metric history to the on-call channel and opens a sizing review. Its output is a decision record, not a resize. Prerequisites: a metric export from Azure Monitor and a webhook to your chat tool. The threshold and the review owner are your decisions.
Worked example, using illustrative counts rather than measured results: suppose a web tier runs eight 4-vCPU VMs on Cobalt 100, which is 32 vCPUs in total. If the up-to 1.5x web-serving figure held for your service, the same throughput would need about 21 vCPUs. That rounds up to six Standard_D4ps_v7 VMs (24 vCPUs), down from eight, a 25% cut in VM count (2 of 8). The math assumes the best case and ignores disks, networking and image work, so a real service may show a smaller gain.
Automation layer or DIY: the trade-off
Many small teams would build this review in Zapier, Make or n8n, or script it in-house. Those routes can support run histories, retries, error branches and audit evidence when they are configured that way. The buyer still designs and owns idempotency, so a re-run does not open duplicate tasks, along with escalation paths, access controls and maintenance of every step. A proposed US Tech Automations design would configure the same trigger, action and output chain with those controls named up front. It still depends on your export access, your reviewer and your change-approval rules, so none of those responsibilities goes away.
When NOT to use US Tech Automations: skip an automation layer for three or four VMs that one engineer can review in the Azure portal in an afternoon, because a checklist and a spreadsheet are cheaper to keep. Skip it when the blocker is an x86-only binary, since no workflow fixes a missing Arm64 build; the work there is the compatibility fix itself. And skip it when Azure's own alerts and cost views already cover your trigger, because an extra layer is one more thing to maintain. For a related post on Azure-hosted AI model choices, see this Azure comparison. For a SaaS-focused comparison, see this SaaS post.
Common mistakes to avoid
Treating the "up to 50%" figure as your expected gain instead of a best case.
Resizing before checking the OS image for NVMe and Generation 2 support.
Assuming live migration will protect uptime during host maintenance on these series.
Budgeting from the VM line alone, when disk storage is billed separately.
Assuming a region is live because the launch post lists it as planned.
Skipping a rollback path to current VMs during cutover.
Frequently asked questions
Is Azure Cobalt 200 generally available now?
Yes: Microsoft's November 17, 2026 launch post announces general availability, but a June 2026 Azure notice described a private preview, so confirm your region and series before you plan.
Does Cobalt 200 cost more than Cobalt 100?
The sources reviewed cannot answer that yet, because no Cobalt 200 list price appears on the pricing pages reviewed. Treat both as Quote-based until Azure publishes SKU rows.
Can I move a workload to Cobalt 200 without changes?
Not without checks: the Cobalt 200 series run on Arm64 and need NVMe-capable, Generation 2 images, so most moves need image and compatibility work first.
Which series should a small team look at first?
Start with Dpsv7 for general-purpose services, because its Learn reference covers the widest range of sizes for that workload type. Memory-heavy or local-storage workloads point to Mpsv4 or Lpsv5 instead.
Does live migration work on these VMs?
No: the Learn references list live migration as not supported on Dpsv7, Mpsv4 and Lpsv5, while memory-preserving updates are supported.
Do I need to act this month?
Only if one workload is capacity-bound, already Arm64-ready and deployable in a region you have confirmed. Otherwise, prepare the checks now and wait for SKU pricing.
What to do this month
This month the useful steps are narrow. Confirm the regions and series you need on Azure's current size references, run an Arm64 and image check on one non-critical service, and keep a rollback path to your current VMs. Keep the price question open as Quote-based until Azure publishes SKU rows. If you want the inventory review turned into a configurable task flow, see how US Tech Automations configures this, and bring your own export and reviewer rules.
About the Author

Helping businesses leverage automation for operational efficiency.