Hyperconverged infrastructure (HCI) is a cluster of industry-standard x86 servers in which every node runs the hypervisor and contributes its local drives to a shared, software-defined storage pool, and the cluster, often including virtual networking, is managed as one system. Instead of a separate SAN, the storage software keeps copies or erasure-coded fragments of data on other nodes, so data stays online when a drive or server fails, and capacity grows by adding nodes. As of October 2026 the platforms built this way include Microsoft Azure Local, Nutanix, Proxmox VE with Ceph, Scale Computing HyperCore and VMware vSAN.
More virtualization and storage coverage is listed in Enterprise Techie’s topic index.
In this guide
Key facts
| Item | What the documentation says |
|---|---|
| Definition | Server, storage, network and management software combined in a single unit (SNIA) |
| Data protection | Full copies on other nodes, or erasure coding across nodes |
| Capacity overhead | 2x or 3x for copies; 1.25x to 1.5x for erasure coding in vSAN ESA |
| Smallest HA cluster | Two nodes plus a witness; Proxmox VE with Ceph needs three servers |
| Separate scaling | SAN-backed, storage-only or external-array designs |
What is HCI?
The SNIA Dictionary defines a hyper-converged system as “A product that combines server, client, storage, network, and management software in a single unit.” Broadcom says vSAN “aggregates local or direct-attached capacity devices of all ESX hosts in a cluster and creates a single storage pool,” so HA, vMotion and DRS work without external shared storage. Proxmox says that with Ceph on the hypervisor nodes, “Separate storage networks (SANs) and connections via network attached storage (NAS) disappear.”
The Nutanix Cloud Bible adds that a hyperconverged system “Must appear and provide the same capabilities as centralized storage (e.g. HA, live-migration, etc.).” The same nodes run the VMs and the storage software, and one console manages both.
How HCI storage works
In a multi-node cluster, each write is stored on more than one node. Nutanix describes data being “synchronously replicated to another one or two Nutanix CVM’s OpLog (dependent on RF) before being acknowledged.” The protection method sets usable capacity:
- Mirroring or replicas. In Storage Spaces Direct, two-way mirroring has 50 percent storage efficiency and three-way mirroring 33.3 percent. Proxmox creates Ceph pools by default with “a size of 3 replicas and a min_size of 2 replicas.”
- Erasure coding. Data is split into data and parity fragments on different nodes. Broadcom’s vSAN ESA table lists RAID-5 (2+1) at 1.5x on 3 hosts, and RAID-5 (4+1) at 1.25x and RAID-6 (4+2) at 1.5x on 6 hosts, against 2x for RAID-1. Storage Spaces Direct dual parity needs four servers, with efficiency rising “from 50 percent up to 80 percent” as servers are added.
- The trade-off. Microsoft recommends mirroring for most performance-sensitive workloads, and Nutanix calls its erasure coding “Not suitable for workloads that are write or overwrite intensive, such as VDI.”
Read paths differ. Nutanix documents data locality: the Controller VM (CVM) on a VM’s host serves its reads and writes, and after the VM moves, AOS migrates its data locally as it is read. Scale Computing’s SCRIBE layer wide-stripes data across the disks in the cluster, and Scale says SCRIBE “does not run in a virtual machine as a virtual storage appliance (VSA) or a controller VM.” Broadcom calls vSAN “a distributed layer of software that runs natively as a part of the ESXi hypervisor.”
Scale-out and small sites
Microsoft says Storage Spaces Direct scales “by adding more drives or adding more servers,” then “automatically onboards the new drives and rebalances the storage pool.” Azure Local hyperconverged deployments run from one to 16 machines, or up to eight in a rack aware cluster.
Two-node clusters can’t keep a majority vote alone after one node fails, so a third vote lives elsewhere:
- Azure Local: “For two-node systems, a witness is required,” per the deployment prerequisites, and portal deployment stores a cloud witness in an Azure Storage account.
- VMware vSAN: a two-node cluster is two hosts at one site plus a witness host, usually at the main site. Broadcom says these “are often used for remote office/branch office environments,” and one witness host can serve up to 64 of them.
- Proxmox VE: a QDevice can add a third vote to a 2-node cluster, but hyper-converged Ceph requires “at least three (preferably) identical servers.” See the Proxmox cluster guide.
- Scale Computing: supports single-node systems, replicated pairs and “storage-only mode for local witness.”
HCI vs converged vs traditional 3-tier
Traditional three-tier infrastructure separates servers, a storage network and shared arrays. SNIA defines converged infrastructure as “The pooling of compute, storage, and networking resources using either common management tools or common (shared) physical resources.” That usually means a tested design: NetApp describes FlexPod as “a predesigned, integrated, and validated architecture” of Cisco UCS servers, Cisco switches and NetApp arrays. HCI removes the array tier; Nutanix’s 3-tier comparison shows the storage controllers moving onto the hosts. The array side is covered in NAS vs SAN.
| Traditional 3-tier | Converged (CI) | Hyperconverged (HCI) | |
|---|---|---|---|
| Building blocks | Servers, SAN or NAS arrays, storage fabric | Servers, switches and arrays in a validated design | Nodes with local drives |
| Storage logic | Array controllers | Array controllers | Software on every node |
| Management | Per tier | Common tools across tiers | One console |
| Growth | Servers and arrays separately | By tier, within the design | Whole nodes; some add storage-only nodes |
| Data protection | RAID and replication in the array | RAID and replication in the array | Copies or erasure coding across nodes |
Disaggregated HCI and external storage
The term has no single definition. Microsoft’s Storage Spaces Direct docs call separate storage and compute clusters a converged deployment, “also known as ‘disaggregated,’” and a panelist in a 2021 SNIA Q&A said “The solutions marketed as disaggregated HCI are not HCI.” What vendors document:
- Azure Local: disaggregated deployments run one to 64 machines on SAN storage, and hyperconverged clusters can attach Fibre Channel or iSCSI arrays from qualified vendors, though rack aware clusters can’t.
- Nutanix: in a compute cluster the CVM “does not manage any local disk drives”; the array handles compression, deduplication and encryption. The design covers Dell PowerFlex, Dell PowerStore and Everpure, and says NetApp “Will utilize NFSv3.” See Nutanix external storage.
- Proxmox VE: RBD storage can use an external Ceph cluster.
- VMware vSAN: a vSAN storage cluster “provides storage resources, but not compute resources,” needs at least four ESX hosts (Broadcom recommends six) and serves compute clusters. See vSAN alternatives.
HCI platforms compared from their own documentation
Listed alphabetically, not ranked. Microsoft Learn says Azure Stack HCI “is now part of Azure Local.”
| Platform | Hypervisor | Storage layer | Node minimums documented | Resilience settings |
|---|---|---|---|---|
| Microsoft Azure Local | Hyper-V | Storage Spaces Direct | 1 machine; 2 need a witness | Two-way or three-way mirror, dual parity |
| Nutanix (AOS) | AHV, ESXi or Hyper-V | Distributed storage via a CVM per node | 5 nodes for RF3 | RF2 or RF3; erasure coding |
| Proxmox VE | KVM, plus LXC containers | Ceph on the cluster nodes | 3 servers for Ceph | Replicated or erasure-coded pools |
| Scale Computing HyperCore | KVM-based | SCRIBE, inside the hypervisor | Single node, replicated pair, 3+ nodes | Two or more copies of every chunk |
| VMware vSAN | ESX (ESXi in vSAN 8) | vSAN, part of the hypervisor | 3 hosts; 2 plus a witness host | RAID-1, RAID-5 or RAID-6 per policy |
Where HCI fits and where it doesn’t
HCI suits branch offices, edge sites and VM estates where compute and storage grow together and nobody wants to run a separate array. The limits:
- Uneven growth. A node added for capacity also adds CPU, memory and licensed cores or sockets. Workarounds include Nutanix “storage-only” nodes, vSAN storage clusters and Azure Local SAN attachment.
- Licensing unit. Azure Local “is priced per physical core” plus any Azure services, with guest OS licenses separate. Nutanix applies NCI licenses “on the number of physical CPU cores.” Proxmox VE subscriptions are per occupied CPU socket; see Proxmox vs VMware.
- Failure domains. A node failure removes compute and a share of storage at once, and maintenance lowers redundancy while it lasts. Microsoft notes three-way mirroring survives another failure while one server reboots; single parity does not.
- Rebuild times. Lost copies are rebuilt into free space on surviving nodes. In a three-host vSAN cluster with one failure to tolerate, vSAN does not rebuild after a host failure, and Broadcom advises keeping consumption below 80 percent and reserving capacity for rebuilds. Proxmox warns “Especially with small clusters, recovery might take long.”
- One management plane. NIST SP 800-125 (January 2011) says “The security of the entire virtual infrastructure relies on the security of the virtualization management system that controls the hypervisor.” In HCI that console controls storage too.
HCI evaluation checklist
- Minimum nodes and witness: where the witness runs and what network it needs. An Azure Local cloud witness needs an Azure Storage account.
- Resilience and usable capacity: each policy’s overhead from the vendor’s table, plus rebuild reserve. Nutanix calls capacity that still allows a full rebuild “resilient capacity.”
- Hardware compatibility: Microsoft support “may only be provided” for hardware in the Azure Local Catalog; vSAN requires certified devices, drivers and firmware.
- External storage: supported arrays and protocols, and which data services move to the array.
- Hypervisor choice: Nutanix runs AHV, ESXi or Hyper-V; the others ship one. See KVM vs VMware and Hyper-V vs VMware.
- Upgrade process: whether hypervisor, storage, firmware and drivers update as one rolling bundle. Nutanix’s upgrade token ensures “only one CVM reboots at a time”; Azure Local adds OEM Solution Builder Extension packages.
- Licensing unit: per core or per socket, plus capacity, guest OS or cloud charges.
- Backup integration: Veeam Backup & Replication 13 lists Azure Local, Nutanix AHV, Proxmox VE, Scale Computing HyperCore and VMware vSphere, but not Trusted Launch VMs on Azure Local. Proxmox also offers Proxmox Backup Server.
Bottom line
HCI replaces the array with software on every node: fewer components to buy and manage, in exchange for growing compute and storage together and rebuilding across the cluster after failures. Shortlist on documented minimums, overheads and licensing units, then test a node failure and a rolling upgrade. Wider options are in the VMware alternatives shortlist.
Frequently asked questions
What is hyperconverged infrastructure (HCI)?
A cluster of standard servers in which each node runs the hypervisor and contributes local drives to a shared software-defined storage pool, managed as one system. Data is protected by copies or erasure coding across nodes, and the cluster grows by adding nodes.
HCI vs traditional infrastructure: what is the difference?
Traditional three-tier infrastructure uses separate servers, a storage network and shared arrays, each managed and scaled on its own. HCI moves storage into the servers, so compute and storage scale together under one console, at the cost of less freedom to add storage alone.
Converged vs hyperconverged: what is the difference?
Converged infrastructure packages servers, networking and a separate array as one tested design, such as FlexPod. Hyperconverged infrastructure removes the array and runs storage as software on the server nodes.
What is disaggregated HCI?
Usually an HCI management model with storage that scales separately, on a SAN or a storage-only cluster, such as Azure Local disaggregated deployments or vSAN storage clusters. A SNIA panelist argued such designs are converged rather than hyperconverged.
How many nodes does HCI need?
Azure Local runs on one machine and needs a witness for two; vSAN needs three hosts, or two plus a witness host; Proxmox VE with Ceph needs three servers.
Sources
- SNIA Dictionary
- Microsoft Learn: Azure Local hyperconverged deployments
- Microsoft Learn: Storage Spaces Direct fault tolerance
- Nutanix Cloud Bible
- Proxmox VE: Hyper-Converged Ceph
- Scale Computing: HyperCore Theory of Operations
- Broadcom: vSAN RAID configuration
- NIST SP 800-125
Last checked: October 2026. Guidance and product features change; check each source before you rely on it.
Buying? Start from RFPhub’s free Virtualization Platform RFP template.