Nutanix Ships NKP Metal: A Hyperconverged Giant Bets on Bare-Metal Kubernetes

For over a decade, Nutanix's entire pitch was "run everything on AHV, our hypervisor, and we'll make the storage and networking disappear." At .NEXT 2026 in Chicago on April 7, the company that built a multibillion-dollar business on that pitch announced NKP Metal — Kubernetes running directly on bare metal, no hypervisor required, shipped as a flagship SKU rather than a niche one. A vendor whose whole existence depended on you virtualizing everything just told its own customers that some workloads are better off without the layer it sells.

What NKP Metal Actually Ships​

NKP Metal extends the existing Nutanix Kubernetes Platform to provision Kubernetes clusters straight onto physical servers instead of onto AHV-hosted VMs. The mechanics reuse Nutanix's existing lifecycle tooling rather than bolting on something new:

• Node provisioning through Nutanix Foundation, the same imaging tool Nutanix has used for years to bring up hyperconverged nodes — now pointed at bare servers instead of hypervisor hosts.
• OS and firmware lifecycle through Nutanix Lifecycle Manager (LCM), so patching a bare-metal Kubernetes fleet uses the same operational muscle memory as patching an AHV cluster.
• Storage either through the standard Kubernetes CSI driver against existing Nutanix storage, or through a new purpose-built option called Cloud Native AOS, aimed at clusters with no hypervisor layer to hang traditional Nutanix storage services off of.

The feature is in early access now for customers already licensed on NKP Pro or NKP Ultimate, with general availability targeted for the second half of 2026. It's explicitly positioned at GPU-dense AI training and inference, plus edge deployments where a hypervisor host has nothing useful to abstract.

Why Bare Metal, Why Now — and Why the Obvious Reason Is Wrong​

The intuitive story is "hypervisors slow down GPUs, so ripping one out makes AI workloads faster." That story is mostly wrong for Nutanix specifically, and it's worth being precise about why, because the imprecise version overstates the case.

Nutanix AHV has supported direct GPU passthrough — dedicating a physical GPU to a single VM — for years, and passthrough is not where the hypervisor tax lives. Independent benchmarking puts direct PCIe GPU passthrough at 98-100% of bare-metal throughput on KVM-family hypervisors (AHV is KVM-based), with the gap closing further on modern platforms — NVIDIA's own testing on GB200 NVL4 shows a 0.21% delta between virtualized-with-passthrough and bare metal. If a shop is already running dedicated GPUs through AHV passthrough, moving those exact VMs to bare metal buys close to nothing on raw compute.

The overhead that's real — 65-76% of peak throughput on virtualized instances, a 24-35% penalty in some measurements — shows up specifically in mediated GPU sharing: vGPU setups where a hypervisor slices one physical GPU across multiple VMs, which is also a mode AHV supports via NVIDIA vGPU/vCS. That penalty is a fractional-sharing tax, not a "hypervisors are slow" tax, and dense training/inference workloads that want a whole GPU (or eight, over NVLink) per job were never the ones paying it.

So what is NKP Metal actually removing? Not GPU throughput — the operational and licensing layer wrapped around it. A hypervisor still means a VM to provision, patch, and eventually migrate; scheduling jitter and memory-ballooning behavior tuned for general-purpose workloads, not a single long-running training job; and — the part enterprise vendors don't lead with — a hypervisor entitlement licensed per core, sitting underneath a Kubernetes entitlement also licensed per core, for a VM that exists purely to host one container runtime. Bare metal collapses two license layers and two lifecycle tools into one. That's a real, sellable simplification. It's just a different one than "your GPUs go faster."

What It Costs: NKP's Per-Core Model Against an Operator-Run Fleet​

NKP's list pricing is per vCPU or per physical core across all worker nodes, sold as an annual support subscription: NKP Pro runs roughly $1,849 per core per year, NKP Ultimate roughly $3,082 per core per year (L3 support included; Ultimate adds the deeper support tier). Those are support-contract prices, not raw software license fees — Nutanix's Kubernetes distribution itself builds on the open-source Kubernetes and CAPI ecosystem, same as everyone else's — but the meter that matters to a buyer is cores under contract, and that meter doesn't turn off.

Run the arithmetic on a modest 64-core bare-metal Kubernetes fleet — a couple of dense multi-GPU training boxes plus a handful of CPU nodes, nothing exotic:

That six-figure delta at 64 cores isn't a rounding error, and it isn't really an apples-to-apples "which is cheaper" argument either — Nutanix is selling an enterprise support contract with a phone number to call when Foundation misbehaves at 3 a.m., and an operator running Cluster API's Metal3 or Tinkerbell providers directly is accepting that they own that reconciliation loop themselves, in exchange for a fleet that costs exactly what the hardware costs. Neither is wrong. They're different bets on who absorbs operational risk, and NKP's per-core meter is the concrete price of Nutanix absorbing it for you.

What NKP Metal Doesn't Replace​

It's worth being explicit about the boundary, because "Nutanix now does bare metal" can read as more convergence with a self-hosted, Cluster-API-run fleet than actually exists. NKP Metal is still gated behind an NKP Pro or Ultimate license, still provisioned through Nutanix's proprietary Foundation tooling, and still lifecycle-managed through Nutanix LCM — an enterprise support product with Nutanix as the vendor of record, not open tooling an operator can fork, audit, or run without a support contract.

A Cluster-API-managed fleet — the kind a git-push PaaS like bex runs underneath a service — provisions the same category of bare-metal Kubernetes cluster, but through Metal3 or Tinkerbell providers that are Apache-2.0, have no per-core meter, and hand the operator (or an AI agent operating on the operator's behalf) the same declarative Cluster/Machine/MachineDeployment object graph regardless of which infrastructure provider is underneath. There's no license key to renew, no core count to true up against an annual invoice, and no vendor relationship required to read the reconciliation logic when something breaks. NKP Metal proves bare-metal Kubernetes is enterprise-ready. It doesn't make the license meter optional.

The Signal That Actually Matters​

The pricing table is the concrete artifact; the more durable point is what it means that this table exists at all. Nutanix closed its most recent fiscal quarter at $703 million in quarterly revenue and $2.435 billion in annual recurring revenue, growing 15% year over year — this is not a startup chasing a trend, it's an incumbent with a hypervisor business to protect. AHV licensing is a real and profitable line item for Nutanix, and every core that runs bare-metal Kubernetes instead of an AHV-hosted VM is a core that isn't paying the hypervisor tax Nutanix itself collects.

A vendor doesn't cannibalize its own highest-margin product speculatively. Nutanix shipped NKP Metal as a flagship SKU — not a beta flag, not a niche listing three products down the page — because enough of its own enterprise customer base was already asking "why do I need a hypervisor under my GPU cluster" that saying no to the request was more expensive than answering it. That's a different kind of evidence than another self-hosting blog or a Hacker News thread arguing PaaS-over-Kubernetes: it's a company with a hypervisor to sell agreeing, in a press release aimed at its own customers, that bare metal is where a specific and growing category of workload belongs.

For a skeptical buyer evaluating whether a self-hosted, Cluster-API-run platform is credible infrastructure or a hobbyist's bet, that's the citation worth having on hand. The technology underneath a Cluster-API-provisioned fleet — Metal3, Tinkerbell, kubeadm — was never in question among people who'd read the Cluster API docs. What NKP Metal adds is a name a procurement committee already trusts, on a product line proving the same thesis for reasons entirely its own.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, backed by Cluster API's declarative fleet management instead of a per-core enterprise license. Star the repo on GitHub or deploy your first app today.

Sources​

• Nutanix Introduces NKP Metal, Bringing Bare-Metal Kubernetes to its Platform — Nutanix press release
• .NEXT 2026: Nutanix Introduces NKP Metal — StorageNewsletter
• Nutanix pushes agentic AI, bare-metal Kubernetes at .NEXT 2026 — Blocks and Files
• No More Compromises: Nutanix NKP Metal Brings Bare-Metal Kubernetes Into the Enterprise Operating Model — VMblog
• Nutanix Kubernetes Platform (NKP) datasheet — Nutanix
• Virtualizing NVIDIA GB200 NVL4: Zero Performance Overhead with Full Isolation — OpenNebula
• Accelerating Private Clouds with NVIDIA GPUs on Nutanix AHV — Nutanix
• Nutanix Posts Q3 FY2026: Revenue $703.1M, ARR $2.43B — TradingView
• Cluster API Provider Metal3 — GitHub
• Cluster API Provider Tinkerbell — Tinkerbell docs