The first question in most sizing conversations is how many hosts. The one that matters comes before it: which of these workloads have to be kept apart from the others, and by what?

In VCF the answer decides whether you need another cluster or another workload domain. They look similar on a diagram. They aren't similar to operate.

What a Domain Buys You

The hierarchy in VCF 9 runs from the fleet to a VCF instance, then to the management domain and workload domains inside it, then to the clusters inside each domain. A domain is where several things get fixed that a cluster can't change on its own.

It gets its own vCenter. The documented limits per instance are 25 vCenters (the management domain's plus 24 workload domains), 2,500 ESX hosts, and 25 NSX Manager instances, so the number of domains is a bounded resource, not a free one.

It carries the principal storage decision. In the material I work from, principal storage (vSAN in either architecture, VMFS on Fibre Channel, or NFSv3) is fixed when the domain is created and can't be changed afterward. Supplemental storage can be added later. Which of those can differ between clusters in the same domain is worth confirming for your target build before you rely on it.

It decides whether NSX Manager is shared. A dedicated NSX Manager instance is recommended for the management domain, and workload domains can share one. That choice has consequences further down the road: Avi Load Balancer has a strict one-to-one relationship with an NSX Manager and isn't supported in workload domains that share one. A documented reference pattern gives each domain its own Avi Controller cluster, all hosted in the management domain, which keeps one tenant's load from affecting another's.

It also tends to become its own lifecycle track. Upgrades and verification run per domain (the upgrade verification material has a separate suite for workload domains from 9.0.x onward), so six domains means six rounds of the same work.

What a Cluster Buys You

Capacity, mostly, and a way to vary the hardware or the failure zone without moving the boundary. Add hosts when the same security, tenancy, and lifecycle rules apply and the workloads just need more room. Add a cluster when they also need a different hardware profile (uniform hardware per cluster is the standing recommendation), a dedicated Edge cluster, or another failure zone.

That last one comes up more than people expect. The three-zone Supervisor model needs three vSphere Zones, one control plane node in each, and a zone assigned to a Supervisor can't be reused by another Supervisor in the same vCenter. In other words the topology you want for resilient Kubernetes can force a cluster count before anyone has sized a single VM. A stretched cluster has its own host-count floor too: eight nodes for the management domain, six for a non-management cluster.

Smaller numbers matter as well. An additional vSAN cluster needs three hosts at minimum, but a three-node cluster won't rebuild components automatically after a failure, and four is the recommendation.

Adding a host to a cluster is a Tuesday. Rearranging domains is a project.

The Cost Nobody Puts in the Estimate

A new domain isn't a line item for hosts. It's another vCenter, plus a three-node NSX Manager cluster if it gets a dedicated one, and those appliances live in the management domain. The management cluster has to be sized for the domain count you plan to reach, not the one you start with. An HA NSX Manager cluster already needs a minimum of four physical hosts to survive one failure. Multiply that by the number of dedicated managers and see whether the management domain's design still holds.

Reasons I'd Add a Domain

  • A different principal storage model. Fibre Channel for one workload set and vSAN for the rest is a domain boundary, because that choice is fixed at creation.
  • A separate network administration boundary. If another team owns the NSX configuration and nobody wants their changes landing on a shared manager, give them their own. Load balancing plans may force the same answer.
  • Hard isolation requirements. The design decision list for tenancy has three tiers: soft separation with projects, hard separation with organizations, and physical separation. PCI scope or external customers can push toward the harder two.
  • A different release cadence. A set of workloads that has to stay on its own upgrade schedule is easier to protect in its own domain.
  • A blast radius you won't accept. One vCenter and one NSX Manager failing should cost you one domain, and you decide how much that is.

Migrations Tend to Decide This for You

If you're bringing an existing estate into VCF, the boundaries often get inherited. Import brings a separate vCenter and its cluster in as a workload domain, while Convert brings the existing stack in as the management domain. Every vCenter that was standing before the project becomes a candidate domain, whether or not anyone thought it deserved to be one.

The choice in front of the team is real: import the current cluster as a domain, or deploy a new domain and migrate onto it. Old vCenter sprawl is a poor reason to keep five domains. Consolidating on purpose costs an effort now; keeping them by accident costs an upgrade track each, indefinitely.

Some of the limits above come from vSphere 8-era and earlier delivery material, and the domain counts are per 9.0. Pull the figures for your target build before they go into a design document.

Questions I Ask Before Anyone Draws a Domain Box

Which workloads need a different principal storage model, and is that reason enough by itself?

Who administers NSX for each workload set, and would they share a manager with another team's changes?

What upgrade cadence can each group of workloads tolerate?

Does any compliance or tenancy requirement demand more than soft isolation?

Has the management domain been sized for the vCenters and NSX Managers the final domain count implies?

If you're staring at a domain diagram and unsure which of the boxes are boundaries and which are just capacity, send me a DM. Fifteen minutes on the whiteboard, remote, is usually enough to tell.