What this is. You run NKP, or you are about to inherit it, and you need to know what actually changed in 2.19 before you plan an upgrade or a new build. This walks the release in the order a platform team cares about: the baseline versions, the new bare metal story, the AI stack, the operational model AI Navigator sits on, upgrades, and the things that break or disappear when you move up. Every feature here is tied to the official release notes and its tracking ID, so you can look it up.
What this is not. This is not a complete NKP guide, an install walkthrough, or a licensing breakdown. It does not cover Nutanix Enterprise AI, which ships on its own cadence and is mentioned only where it touches the platform. Where a feature has prerequisites or limits, this post names that they exist and points you at the documentation rather than reproducing every condition.
Version scope. NKP 2.19, general availability, release notes dated 5 November 2026. NKP Insights 1.9 ships with it and gets its own short section. Deliberately excluded: Nutanix Enterprise AI internals, the step by step upgrade procedure, and anything not stated in the 2.19 release notes. Where earlier public previews of 2.19 described features that the final release notes do not list, the release notes win and those items are left out.
On this page
- 1. The short version
- 2. Platform baseline: Kubernetes, OS, Prism Central
- 3. Dual native infrastructure
- 4. NKP Metal
- 5. The AI platform, now packaged
- 6. AI Navigator and the MCP Server
- 7. One action upgrades
- 8. What is deprecated or removed
- 9. Known issues worth reading before you upgrade
- 10. NKP Insights 1.9
- 11. If you read nothing else
1. The short version
NKP 2.19 is the release where Nutanix puts bare metal Kubernetes and virtualized Kubernetes under one control plane, ships a curated AI stack instead of a parts list, and makes workload cluster upgrades a single action. If you own a Kubernetes estate on Nutanix, the headline items are NKP Metal, the AI Catalog, GPU scheduling and slicing for both NVIDIA and AMD, and one action per cluster upgrades. The rest is version bumps, an MCP based operations layer, and a set of deprecations you need to read before you upgrade.
| Area | What landed in 2.19 | Why you care |
|---|---|---|
| Baseline | Kubernetes 1.36.2, Rocky Linux 9.8, RHEL 9.8, Prism Central Projects 2.0 | New floor for new clusters, and a consolidation step before you upgrade |
| Bare metal | NKP Metal, via Foundation Central and host inventory | HCI style lifecycle for physical Kubernetes, including GPU and edge |
| AI stack | AI Catalog, GPU scheduling and slicing, CNCF AI Conformance | Validated AI apps and GPU handling instead of a DIY assembly |
| Operations | MCP Server on by default, AI Navigator resource access | Programmatic, RBAC scoped access to cluster and platform state |
| Upgrades | One action per cluster upgrades, air gapped release metadata bundle | Platform apps, Kubernetes, and infrastructure move in one step |
| Removed | Kubernetes Dashboard, GitRepository catalogs, nkp upgrade workspace | Things that will break on upgrade if you depend on them |
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · New Features and Enhancements, all IDs cited in the sections below.
2. Platform baseline: Kubernetes, OS, Prism Central
Start here, because the baseline decides what a new cluster looks like and what you have to clean up before an upgrade.
| Item | 2.19 value | Tracking ID |
|---|---|---|
| Default Kubernetes | 1.36.2 (management and attached clusters support 1.36.x and 1.35.x) | NCN-110591 |
| Rocky Linux | 9.8 | NCN-115946 |
| Red Hat Enterprise Linux | 9.8 | NCN-115133 |
| Compatible Prism Central | pc.7.5 and pc.7.6 | Release Notes, Table 1 |
| Compatible AOS | AOS 7.5 and AOS 7.6 | Release Notes, Table 1 |
Prism Central Projects 2.0. NKP clusters can now be created and operated inside Prism Central Projects 2.0. A cluster uses either the Prism Central default project or a single user defined project, and the choice applies to the whole cluster: the control plane and every worker node pool must sit in the same project. That constraint is the part that bites on upgrade.
Nutanix DockerHub registry integration. NKP 2.19 can pull container images from the Nutanix DockerHub registry. You generate a token on the Nutanix portal and supply it when you create a management cluster. Workload clusters inherit the management cluster token by default, and you can override it per cluster from the CLI, the UI, or a YAML manifest, and change it after provisioning. One thing to catch: when you upgrade a management cluster that does not already have a token, you have to supply one during the upgrade.
Default cluster onboarding in Prism Central. During deployment and upgrade, NKP now installs and registers NKP clusters with Prism Central automatically, so the infrastructure team gets visibility without a separate registration step.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Kubernetes 1.36 Support (NCN-110591) · Rocky Linux 9.8 Support (NCN-115946) · Red Hat Enterprise Linux 9.8 Support (NCN-115133) · Prism Central Projects 2.0 Support (NCN-109515) · Nutanix DockerHub Registry Integration (NCN-103153) · Default Cluster Onboarding in Prism Central (NCN-109556) · Prism Central and AOS Version Compatibility, Table 1 · Supported Kubernetes Versions, Table 2.
3. Dual native infrastructure
The framing for 2.19 is that containers and virtual machines are both first class, and you should not have to run two separate management stacks to get there. Virtualized Kubernetes clusters on AHV and bare metal Kubernetes clusters are managed through one NKP control plane. NKP Metal, covered next, is what makes the bare metal half real.
The practical point is the one that costs you time today. If your GPU heavy AI work wants bare metal but the rest of your estate is virtualized, the usual outcome is two operating models, two upgrade cadences, and two on call runbooks. A single control plane over both substrates is aimed squarely at that split.
One control plane manages both substrates, so the operating model does not fork when you add bare metal for GPU workloads.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Bare-metal Deployment Support (NCN-103015), which states NKP Metal adds physical worker node pools to VM based clusters under NKP lifecycle management.
4. NKP Metal
NKP Metal is the flagship of the release. It brings deploying and managing Kubernetes workloads on physical servers into NKP itself, using Foundation Central and an NKP host inventory. The workflow is: image and register hosts, assign them to workspaces, then either create full bare metal clusters or add physical worker node pools to existing VM based clusters. NKP Metal gives those workloads direct access to compute, network, storage, and accelerator resources, and it handles host and node pool lifecycle through NKP rather than through a pile of scripts.
The reason this matters is that bare metal Kubernetes usually shows up exactly where hand built clusters hurt most: GPU dense AI and machine learning nodes where you do not want a hypervisor tax on the accelerators, and edge sites that need local processing. Being able to add a physical node pool to an otherwise virtualized cluster is the detail that keeps you from standing up a parallel stack just to get a few GPU nodes.
NKP Metal images and registers physical hosts, then builds either a full bare metal cluster or a physical node pool inside an existing cluster. The point is that no layer is left as a manual step.
nkp create cluster metal CLI command. Worth knowing before you hand Metal to a team that only has Workspace Admin.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Bare-metal Deployment Support (NCN-103015) · Known Issues: Metal Host Management and Cluster Creation Unsupported for the Workspace Admin Role (NCN-118452).
5. The AI platform, now packaged
The second theme is that the AI plumbing is now part of the platform instead of a project you staff. Three pieces matter for a platform team.
AI Catalog applications. NKP 2.19 adds a curated set of AI and machine learning applications to the workspace catalog for Pro and Ultimate. The named set is KAI Scheduler, Slurm Operator, Kueue, Kubeflow Pipelines, JupyterHub, and Milvus Operator. Between them they cover GPU aware scheduling, queue based workload management, AI development environments, and vector data storage. These are curated and validated, not a marketplace, which is the difference between a supported starting point and a wiki page from two years ago. Deployment prerequisites vary by application and can include GPU operators, monitoring, ingress, and certificate management, so read each application’s requirements before you promise a data science team a one click install.
| Catalog application | What it covers |
|---|---|
| KAI Scheduler | GPU aware scheduling |
| Kueue | Queue based workload management |
| Slurm Operator | Batch and HPC style job management |
| Kubeflow Pipelines | AI and ML pipeline orchestration |
| JupyterHub | Notebook development environments |
| Milvus Operator | Vector database for AI workloads |
GPU scheduling and slicing. NKP 2.19 supports GPU scheduling and slicing through Kubernetes dynamic resource allocation, for both NVIDIA and AMD GPUs. That AMD support is the notable part, because it means your AI scheduling story is not tied to a single accelerator vendor. One explicit limit: NVIDIA multi instance GPU partitioning is not included in this release.
CNCF Kubernetes AI Conformance. NKP 2.19 meets CNCF Kubernetes AI conformance by combining validated platform applications, GPU capabilities, and operational controls across the cluster life cycle. The conformance sets a vendor neutral baseline for AI and ML workloads: a conformant platform provides accelerator management, gang scheduling, autoscaling, inference networking, observability, and security on top of standard Kubernetes conformance. The practical payoff is portability. Workloads behave the same across public cloud, on premises, and edge, so you can move them and keep using standard AI tooling without re platforming or vendor lock in.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · AI Catalog Applications (NCN-111354) · Advanced GPU Scheduling and Slicing (NCN-112517) · CNCF Kubernetes AI Conformance (NCN-113222).
6. AI Navigator and the MCP Server
NKP 2.19 turns its Model Context Protocol server into a standard part of the management cluster, and wires AI Navigator on top of it. If you have been ignoring the AI Navigator branding as marketing, the thing underneath is a read mostly, RBAC scoped API over your cluster and platform state, which is more interesting to a platform team than the chat window on top.
MCP Server on by default. The NKP MCP Server is enabled by default on the management cluster for Pro and Ultimate licenses in internet connected environments. Three scope limits to note: air gapped deployments do not include the MCP Server, and it is not installed on attached clusters or workload clusters. So this is a management cluster capability, not a fleet wide one.
What the MCP Server exposes, and how it is fenced. The server builds its resource catalog from the connected cluster and can list resource kinds and custom resources on demand. It provides read only access to storage, scheduling, policy, autoscaling, networking, and Flux resources. Two controls matter for security review: you can allow or deny which Kubernetes resource kinds the server exposes, and field filters limit what list and get operations return. The server does not expose Secrets. Because the allow and deny rules and the catalog can differ per cluster, review them after an upgrade.
| Capability | Detail | Tracking ID |
|---|---|---|
| Platform resource access | Kubernetes, Flux GitOps, and NKP platform resources (clusters, workspaces, licenses), RBAC scoped to your role | NCN-113830 |
| Session metrics | Prometheus metrics for chat completions and embeddings, with authenticated session ID; dashboard shows unique sessions over time | NCN-115932 |
| Doc and KB indexing | Indexes Nutanix documentation and knowledge base articles through product APIs | NCN-111788 |
| Resource discovery | discover_resource_types lists resource kinds and custom resources; serves last good catalog during refresh | NCN-115284 |
| Configurable exposure | Allow or deny resource kinds, field filters on returns, no Secrets exposed | NCN-115308 |
| Database recovery | Restores a lost PostgreSQL connection without restarting the pod; readiness probe at /health/ready | NCN-115857 |
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Model Context Protocol Server Enabled by Default (NCN-115786) · AI Navigator Access to Platform Resources (NCN-113830) · AI Navigator Session Metrics (NCN-115932) · Documentation and Knowledge-Base Indexing (NCN-111788) · Dynamic Kubernetes Resource Discovery (NCN-115284) · Configurable MCP Resource Exposure (NCN-115308) · Additional Read-Only MCP Resources (NCN-113836) · AI Navigator Database Connection Recovery (NCN-115857).
7. One action upgrades
NKP 2.19 introduces simplified, per cluster upgrades for workload clusters. After you upgrade the management cluster to 2.19, you can upgrade supported ClusterClass workload clusters, such as Nutanix and EKS clusters, from the NKP UI or with a single CLI command. The workflow upgrades platform applications first, then Kubernetes, then cluster infrastructure, then the related add ons, in that order.
A single upgrade action walks the stack in a fixed order. You trigger it once instead of sequencing four upgrades by hand.
Non ClusterClass clusters. For clusters that do not use ClusterClass, the new nkp upgrade platform-apps command upgrades platform applications independently, before you run the existing Kubernetes and infrastructure upgrade procedures. This is also the replacement path for the deprecated nkp upgrade workspace command, covered in the next section.
Air gapped version discovery. NKP 2.19 adds an nkp-release-specs-nkp-version.tar bundle to the air gapped and private registry image bundles. It carries release version metadata, including supported Kubernetes versions, provider compatibility, and supported upgrade paths, which NKP uses for version discovery. Push it alongside your other image bundles into your private registry or mirror, or air gapped version discovery has nothing to read.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Simplified One-click Cluster Upgrades (NCN-97132) · Release Metadata Bundle for Air-gapped Version Discovery (NCN-112250).
8. What is deprecated or removed
This is the section to read before you upgrade, because every item here is something that changes or stops working when you move to 2.19.
| Change | What it means for you | Tracking ID |
|---|---|---|
| Kubernetes Dashboard removed, replaced by Headlamp | NKP installs and manages Headlamp. A cluster you upgrade leaves Kubernetes Dashboard running; migrate to Headlamp. The upstream Kubernetes Dashboard project is archived. If you run your own Dashboard for application workloads, manage it separately. | NCN-111720 |
| GitRepository based catalog applications removed | NKP no longer supports GitRepository based catalogs from the controller or the CLI. Migrate any remaining AppDeployment resources to an OCIRepository based catalog. | NCN-115673 |
| nkp upgrade workspace deprecated | For clusters that do not use ClusterClass, run nkp upgrade platform-apps on each cluster instead. For ClusterClass clusters, nkp upgrade cluster upgrades platform applications and infrastructure in one command. | NCN-115808 |
| –kommander-applications-repository flag deprecated | No longer supported on nkp install kommander and nkp upgrade kommander. Installation and upgrade no longer use the kommander-applications tar file in the air gapped or full bundle. | NCN-116066 |
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · NKP 2.19 Deprecations: Kubernetes Dashboard (NCN-111720), –kommander-applications-repository Flag (NCN-116066), nkp upgrade workspace Command (NCN-115808), GitRepository-based Catalog Applications (NCN-115673).
9. Known issues worth reading before you upgrade
These are the known issues most likely to affect a real upgrade or build. The release notes carry more, with workarounds. Read the full list before a production move.
| Issue | Impact and workaround | Tracking ID |
|---|---|---|
| Worker node pool upgrade stall with a single instance Nutanix AI PostgreSQL database | An upgrade can stall when Nutanix AI uses a single instance gateway, NAI-API, and PostgreSQL with PodDisruptionBudget enabled, because those pods cannot be evicted. Configure multi instance gateway and nai-api and disable PodDisruptionBudget for PostgreSQL before upgrading. | NCN-118363 |
| Workload cluster Kubernetes version during a partial upgrade | During a partially completed management cluster upgrade, the UI offers only 1.36.2 for new workload clusters, and you cannot select an earlier compatible version. Complete the management cluster Kubernetes upgrade first. | NCN-117879 |
| Workload cluster creation fails when NO_PROXY contains a CIDR | With a proxy and CIDR ranges in NO_PROXY, UI creation fails because it cannot list the AOS cluster, subnet, or OS image. Create the cluster with the NKP CLI. | NCN-108961 |
| Platform application install fails with long workload cluster names | A name over 58 characters pushes the derived Flux Kustomization label past the 63 character Kubernetes limit. Use a name of 58 characters or fewer. | NCN-115563 |
| Case sensitive matching of Prism Central resource names | Since 2.17, CAPX and CCM match Prism Central resource names case sensitively. A case mismatch fails cluster lifecycle operations. Match the exact case in your cluster configuration. | NCN-117267 |
Two limitations are also worth flagging for air gapped and private registry environments: the Harbor Trivy scan has no vulnerability database in an air gapped environment until you fetch and push one (NCN-105807), and the NCR private registry uses a self signed certificate by default, which you have to configure trust for on both management and workload clusters (NCN-105470).
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · Known Issues: NCN-118363, NCN-117879, NCN-108961, NCN-115563, NCN-117267 · NKP 2.19 Limitations: Harbor Trivy Scan (NCN-105807), Self-signed CA Certificate Trust for the NCR Private Registry (NCN-105470).
10. NKP Insights 1.9
NKP Insights 1.9 ships with 2.19 and is upgraded as part of the NKP upgrade, so you do not uninstall and reinstall to move it. It supports the same Kubernetes version as the platform. Most of the 1.9 changes are reliability and database work rather than headline features: configurable recovery strictness during startup, Helm chart defaults tuned for high load that retain the CloudNativePG database objects through upgrade and uninstall, a 24 hour image metadata refresh to cut calls to external registries, idle PostgreSQL session timeouts, and CVE reports generated per container image tag.
Two operational notes carry over. You can route Insights alerts through Alertmanager, with documented examples for Slack and Microsoft Teams. And after a fresh deploy or an upgrade, Insights focuses on your workloads first, so you may see no alerts for the underlying NKP and Kubernetes components until you enable NKP related alerts explicitly.
Sources: NKP 2.19 Release Notes, 5 Nov 2026 · NKP Insights Release Notes, Insights 1.9: Configurable Recovery Failure Threshold (NCN-118675), Helm Chart Defaults for High Load (NCN-118676), 24-Hour Image Metadata Refresh (NCN-118678), Idle PostgreSQL Session Timeout (NCN-118680), CVE Reports for Image Tags (NCN-118682) · Alert Notifications with Alertmanager · NKP Insights Alerts for Customer Workloads · Upgrade Support.
11. If you read nothing else
| # | Take this away |
|---|---|
| 1 | 2.19 runs bare metal and virtualized Kubernetes under one control plane. NKP Metal, via Foundation Central, images physical hosts and builds full clusters or adds physical node pools to VM based clusters. |
| 2 | The AI stack is packaged: a curated catalog (KAI Scheduler, Kueue, Slurm, Kubeflow Pipelines, JupyterHub, Milvus), GPU scheduling and slicing for NVIDIA and AMD, and CNCF AI conformance. NVIDIA MIG partitioning is not included. |
| 3 | The MCP Server is on by default on the management cluster for Pro and Ultimate, internet connected only. It is an RBAC scoped, Secret free read surface you should review on upgrade. |
| 4 | Workload cluster upgrades are one action: platform apps, then Kubernetes, then infrastructure, then add ons. |
| 5 | Before upgrading: consolidate Prism Central projects, migrate GitRepository catalogs to OCIRepository, move off nkp upgrade workspace, and plan the Kubernetes Dashboard to Headlamp move. |
Sourcing note. This post is based on the official Nutanix Kubernetes Platform 2.19 Release Notes, dated 5 November 2026, including the NKP Insights 1.9 release notes published in the same document. Each section cites the specific features and their NCN tracking IDs. Feature names, version numbers, and tracking IDs are taken directly from that document. Where an earlier public preview of 2.19 described capabilities that the final release notes do not list, this post follows the release notes and leaves those items out. Confirm version numbers and prerequisites against the release notes and the product documentation on the Nutanix Support Portal before you rely on them in a design.
Nutanix, AHV, Prism, Prism Central, Flow, NKP, AOS, and Nutanix Kubernetes Platform are trademarks of Nutanix, Inc. Kubernetes is a registered trademark of The Linux Foundation. All other marks are the property of their respective owners. This is a personal blog. It is not affiliated with, endorsed by, or speaking on behalf of Nutanix or any employer.
Leave a Reply