How Do You Size a Server for Proxmox, Hyper-V or VMware in 2026?

Most teams sizing a virtualization server in 2026 are doing it for one of two reasons: the current hosts are due for refresh, or the hypervisor itself is changing. Many organizations are re-evaluating VMware after its licensing changes and looking at Proxmox VE or Microsoft Hyper-V as alternatives, and a platform change is the natural moment to revisit the hardware as well. The good news is that sizing works the same way for all three. Start from the VMs you actually run, size memory first, then cores, storage and network, and design the cluster so it keeps running when a host fails.

How do I size a server for virtualization?

Begin with an inventory, not a spec sheet. For every VM, record allocated vCPUs, allocated memory and disk capacity, plus measured peak CPU, memory and storage I/O over a representative period that includes month-end and backup windows. Your current hypervisor's performance data gives you most of this. Then work through four steps:

• Total the demand: allocated memory, measured CPU, storage capacity and IOPS.
• Add headroom for growth over the period you plan to keep the hardware.
• Decide the failure tolerance. With N+1, the cluster must carry the whole load with one host offline; with N+2, with two offline.
• Divide the load across hosts, then choose a server that meets the per-host figure with memory channels populated evenly.

How much RAM does a VM host need?

Memory is usually the first resource to run out, and the one to overcommit least. Memory reclamation features such as ballooning, page sharing and dynamic memory exist in all three hypervisors in some form, but relying on them for production sizing invites unpredictable performance. Plan for allocated VM memory to fit in physical RAM, plus overhead for the hypervisor itself. On Proxmox with ZFS, the ARC cache also uses host memory, so cap it deliberately rather than leaving it to compete with VMs.

A worked example: 40 VMs averaging 16GB allocated come to 640GB. On a three-host cluster sized for N+1, two hosts must hold all 640GB, so each host needs at least 320GB for VMs. Add hypervisor overhead and growth, and 512GB per host is a sensible build. Populate every memory channel evenly. Current Intel Xeon Scalable processors have eight DDR5 channels per socket, while AMD EPYC 9004 and 9005 have twelve, so balanced totals differ by platform: 512GB is sixteen 32GB DIMMs on the R760's 32 slots, while a 24-DIMM EPYC build lands at 384GB or 768GB. Browse registered ECC modules in our server memory range, and match DIMM type and speed across the whole host.

How many cores per VM host?

Size cores from measured utilization. vCPUs are scheduled onto physical cores or hardware threads, and a host can run more vCPUs than it has cores because most VMs are idle most of the time. How far you can push that depends on the workload: general-purpose server VMs tolerate a relatively high ratio, while databases, VDI and latency-sensitive applications need much less overcommit. Watch CPU ready time on VMware, steal and wait metrics on KVM-based platforms such as Proxmox, and the equivalent scheduling counters on Hyper-V. Rising values mean the host is oversubscribed.

Core count is also a licensing question. Several operating system and virtualization licenses are counted per physical core or per socket, so a host with far more cores than the workload needs can raise software cost without adding useful capacity. High-core-count platforms like the DL385, which supports 4th Gen EPYC 9004 with up to 128 cores or 5th Gen EPYC 9005 with up to 160 cores, make sense where consolidation is the goal and the licensing model allows it. Fewer, faster cores can be the better answer for per-core licensed databases.

What hardware works well for Proxmox?

Proxmox VE is built on Debian Linux, so it runs on a wide range of standard x86 server hardware using in-kernel drivers, and mainstream Dell, HPE and Lenovo rack servers are common choices. The details that matter most are storage and networking:

• Storage controller: ZFS and Ceph want direct access to the disks. Use an HBA, or a controller set to HBA or pass-through mode, not a hardware RAID volume. Our guide to RAID versus HBA explains the difference.
• SSDs: use enterprise SSDs with power-loss protection, especially for ZFS sync writes and Ceph, where consumer drives can perform poorly and wear quickly.
• Boot: a mirrored pair of small drives or M.2 boot devices, kept separate from VM storage.
• Network: separate cluster, storage and VM traffic. Ceph and live migration benefit from 25GbE or faster network adapters, and cluster heartbeat traffic wants a dedicated low-latency link.
• Cluster size: Proxmox HA relies on quorum, so plan three or more nodes, or add an external quorum device to a two-node cluster.

What changes for Hyper-V and VMware?

The hardware logic is the same, with platform-specific rules. VMware ESXi supports hardware listed on its compatibility guide, so check the exact server, storage controller and NIC before ordering, and use pass-through or HBA mode for vSAN. Hyper-V runs on hardware certified for the Windows Server version you deploy, and Storage Spaces Direct likewise needs disks presented directly rather than behind RAID. Hyper-V failover clusters and vSAN can both run as two nodes plus a witness, while Ceph-based Proxmox clusters are typically three nodes or more. If you are moving away from VMware, size the new cluster on current utilization rather than copying the old host count. Many estates were sized years ago, and consolidating onto current DDR5 platforms often means fewer, larger hosts, which also means fewer machines to license, patch and support.

Size memory for the cluster with a host down, not for the cluster at full strength.

Which servers make good virtualization hosts?

The 2U dual-socket rack server is the default VM host: enough DIMM slots for large memory, enough drive bays for local or hyperconverged storage, and enough PCIe slots for redundant networking. The Dell PowerEdge R760 takes 4th or 5th Gen Xeon Scalable processors with 32 DDR5 slots and up to 24 front 2.5-inch drives. The HPE ProLiant DL380 offers 32 DDR5 slots and up to 30 SFF bays, and the DL385 brings AMD EPYC core density on 24 DDR5 slots. Lenovo's ThinkSystem SR650 family fills the same role; see our Lenovo ThinkSystem servers page for current configurations.

If some VMs need GPUs for AI inference, the PowerEdge R760xa with four H100 NVL is a 2U host with up to 4TB of DDR5 and four PCIe Gen5 GPUs that can be assigned to VMs by PCIe pass-through. Treat GPU hosts as a separate pool from general-purpose VM hosts, because their power, cooling and scheduling needs are different.

Nexus Compute supplies new Dell PowerEdge, HPE ProLiant and Lenovo ThinkSystem servers, plus memory, SSDs and network adapters, sourced through authorized distribution with the full manufacturer warranty. Send us a VM inventory or utilization export and your target hypervisor, and we will size the hosts, check compatibility before dispatch and quote within 48 business hours.