How the calculation works
Each row is a group of similar servers or VMs: how many, cores (or vCPUs) each, peak CPU %, RAM, RAM in use and storage in use. Memory is in GB and storage in decimal TB (1 TB = 1,000 GB); vendor figures given in GiB are used as GB.
1. Size each VM. With right-sizing, a VM gets enough vCPUs for its peak to land at your target utilization (tu), and the memory it uses:
- vCPUs = cores × peak ÷ tu, rounded up, at least 1 and no more than the source cores.
- vRAM = RAM in use, rounded up to a whole GB, no more than the source RAM. The headroom for memory comes from the host limit below, not from each VM.
- With “keep current cores and RAM”, each VM gets the source sizes unchanged.
- CPU demand = cores × peak, in core-equivalents. One source core counts as one target core. Newer CPUs usually do more per core, so this leans conservative.
2. Add growth. vCPUs, vRAM, CPU demand, storage and IOPS grow by your annual rate, compounded over the years you plan for.
3. What one host offers.
- Cores for VMs = sockets × cores per socket, less the Nutanix CVM on Nutanix AHV. Nutanix sizes the CVM in logical cores; with hyperthreading on, the calculator counts its vCPUs as half as many physical cores. That conversion is ITBroker.com’s planning assumption, not a Nutanix sizing rule.
- Memory for VMs = host RAM − hypervisor reserve − CVM memory (Nutanix only) − 8 GB per Ceph OSD (Proxmox VE with Ceph; Proxmox recommends at least 8 GiB per OSD).
- vCPU limit = cores for VMs × your vCPU:pCPU ratio.
- CPU limit = cores for VMs × tu: peak demand stays below your target.
- Memory limit = memory for VMs × tu × memory overcommit (100% means none).
- Storage (in the cluster only) = raw TB × (1 − free space reserve) ÷ the protection multiplier, against storage in use ÷ your data reduction ratio. External SAN or NAS storage doesn’t count toward hosts.
Target utilization is applied once on each path: to vCPU counts when sizing each VM (the ratio limit has no utilization term), and to CPU demand and memory on the host.
4. Count hosts. Each limit needs working hosts = demand ÷ the per-host limit, rounded up. The failover reserve (N+1 or N+2) is added, so the VMs still fit with that many hosts down; for storage, the data has to fit on the remaining hosts with the protection layout those hosts allow (vSAN ESA RAID-5 uses 4+1 only from 6 hosts). The total is the largest of these, and at least the minimum for the platform and protection scheme. For storage in the cluster, that minimum applies to the hosts left after the reserve, so the layout can be rebuilt with those hosts down: RF3 with N+1 needs 6 nodes, not 5. The binding constraint is the limit that needs the most hosts.
5. Ratios. The vCPU:pCPU ratio is vCPUs ÷ cores for VMs across all hosts, and again with one host (or two) down.
Protection multipliers and minimums (from the vendor pages listed under the calculator): vSAN RAID-1 with one failure to tolerate 2×, 3 hosts; vSAN ESA RAID-5 1.5× (2+1) from 3 hosts and 1.25× (4+1) from 6; vSAN OSA RAID-5 1.33×, 4 hosts; vSAN RAID-6 1.5×, 6 hosts; vSAN RAID-1 with two failures to tolerate 3×, 5 hosts; Storage Spaces Direct two-way mirror 2×, 2 servers, three-way mirror 3×, 3 servers; Ceph pool size 3 3×, 3 servers; Nutanix RF2 2×, 3 nodes, RF3 3×, 5 nodes. Proxmox VE HA needs at least three nodes for quorum, so a failover reserve on Proxmox VE means at least 3 hosts.
Assumptions to change for your platform. The 8 GB hypervisor reserve is a placeholder, not a vendor figure: Hyper-V calculates its own host reserve, VMware’s per-VM memory overhead depends on each VM’s size, and ZFS on Proxmox VE needs memory of its own. The 30% free space reserve follows Broadcom’s advice of about 30% slack for vSAN (vSAN starts rebalancing at 80% device use); on Storage Spaces Direct, Ceph and Nutanix it is a planning assumption. The report also checks the largest VM’s storage against one virtual disk: 62 TB per VMDK on vSphere, 64 TB per VHDX on Hyper-V, 64 TB tested by Nutanix on AHV; Proxmox VE states no single limit. Data reduction stays at 1:1 unless you have measured it on your own data. Nothing here models CPU generation differences, GPU passthrough, licensing costs or prices.
Worked example
36 physical servers moving to Proxmox VE with Ceph, sized for 3 years at 10% growth a year:
| Group | Servers | Cores | Peak CPU | RAM / in use | Storage | Each VM |
|---|---|---|---|---|---|---|
| 1 | 12 | 16 | 35% | 128 / 70 GB | 900 GB | 7 vCPUs, 70 GB |
| 2 | 4 | 24 | 60% | 384 / 300 GB | 4,000 GB | 18 vCPUs, 300 GB |
| 3 | 20 | 8 | 15% | 32 / 14 GB | 200 GB | 2 vCPUs, 14 GB |
At 80% target utilization, group 1 gets 16 × 35% ÷ 80% = 7 vCPUs and the 70 GB it uses. Today that adds up to 196 vCPUs, 2,320 GB of vRAM, 148.8 cores of peak demand and 30.8 TB of storage. Three years at 10% multiplies each by 1.331: 261 vCPUs, 3,088 GB, 198.1 cores and 41 TB.
The hosts are 2 × 32 cores with 1,024 GB, keeping back 40 GB each (an 8 GB hypervisor reserve and 32 GB for four Ceph OSDs at 8 GB), with 30 TB of raw storage, Ceph pool size 3, the 30% free space reserve, a 4:1 ratio and N+1:
| Limit | Per host | Working hosts | With N+1 |
|---|---|---|---|
| vCPU ratio | 64 cores × 4 = 256 vCPUs | 261 ÷ 256 → 2 | 3 |
| CPU peak | 64 × 80% = 51.2 cores | 198.1 ÷ 51.2 → 4 | 5 |
| Memory | 984 GB × 80% = 787.2 GB | 3,088 ÷ 787.2 → 4 | 5 |
| Storage | 30 TB × 70% ÷ 3 = 7 TB | 41 ÷ 7 → 6 | 7 |
Plan for 7 hosts (6 + 1 for failover): storage sets the count. CPU peak and memory are next at 5. The vCPU:pCPU ratio is 261 ÷ 448 cores = 0.6:1 with all hosts up and 0.7:1 with one down, well under the 4:1 limit, because right-sizing cut the vCPUs. By year: 6 hosts today and in year 1, 7 in years 2 and 3. Adding raw storage per host, or choosing a protection scheme with less overhead, is what would bring the count down here, not more cores.
Reading the report
The page shows the totals, the host count and the binding constraint. The emailed report adds the load on each host with every host up and with the reserve in use, the hosts each limit needs on its own, raw and usable storage, the count year by year, a check of the largest VM against one socket’s cores and memory (a NUMA node is one socket or smaller, so vendors advise fitting a VM within one physical NUMA node where possible), one virtual disk and the platform’s per-VM maximum, and the cores or sockets each vendor counts for licensing. Use it to compare host configurations before asking for quotes: a different core count, more RAM or more drives per host can change which limit binds.

