GPU servers in a vSphere 8 or VCF 9 cluster: host releases, compatibility, licensing, CPUs

What a GPU host needs before it joins the cluster

A GPU server joins a vSphere 8 or VCF 9 cluster like any other host once it meets three conditions: an ESX release NVIDIA lists for the card (for the RTX PRO 6000 and RTX PRO 4500 Blackwell Server Editions ESXi 8.0 Update 3g or later, VCF 9.0.1 or later, or VCF 9.1), the vSphere Foundation edition or an Enterprise Plus licence for NVIDIA vGPU, and the NVIDIA vGPU Manager of a current branch. The processors then decide the VMware licence, because Broadcom counts the CPU cores per host; GPUs do not count.

This page covers the host side: releases, lists, licences and where the GPU hosts sit in the cluster. How a VM gets its GPU, passthrough or vGPU, with vMotion and DRS for each, is in our guide to GPUs in VMware vSphere, and the full Private AI stack on VCF is in VMware Private AI Foundation with NVIDIA.

Requirements as of October 2026. NVIDIA vGPU product support matrix for VMware vSphere (29 September 2026), NVIDIA vGPU release table, NVIDIA AI Enterprise 8.2 support matrix (2 September 2026), Broadcom ESX 9.0 hardware requirements (19 October 2026) and KB 313548, all read 24 October 2026; CVE status per NVIDIA's security bulletin of 30 September 2026 as cited in our driver-branch guide.

ESX releases per card

NVIDIA's support matrix lists three hypervisor rows for vSphere: VCF 9.1, VCF 9.0 and ESXi 8.0. For the 8.0 row its footnote reads: "Requires VMware vSphere Hypervisor (ESXi) 8.0u3 P06 and later updates to release 8.0 unless explicitly stated otherwise." The Blackwell cards carry their own footnotes on top: VCF 9.0.1 or later on the 9.0 line, and ESXi 8.0 Update 3g or a later 8.0 update. On the 8.0 line the two floors are the same build, because 8.0 Update 3 P06 was released as 8.0 Update 3g (build 24859861), as our guide to GPUs in VMware vSphere notes; the footnote that separates the Blackwell cards from the L40S and L4 is the VCF 9.0.1 one.

NVIDIA vGPU product support matrix for VMware vSphere, last updated 29 September 2026, read 24 October 2026. The RTX PRO 6000 Workstation and Max-Q editions and the RTX PRO 5000 are not on the matrix either.

The H200 NVL needs a separate word. It is in NVIDIA AI Enterprise 8.2's GPU list, which supports vSphere 8.0 and 9.0 and later, but not in NVIDIA's vGPU matrix for vSphere as read on 24 October 2026. Plan an H200 NVL host on vSphere with DirectPath I/O, whole cards per VM and the NVLink-bridged pairs assigned together, and ask NVIDIA to confirm in writing before a design depends on compute vGPU for this card. Our H200 NVL MIG profiles cover what the card offers on bare metal.

On vSphere, the MIG features differ by release. Time-sliced vGPUs inside a MIG instance and several MIG-backed vGPUs in one VM need VCF 9.1, so a host on ESXi 8.0 or VCF 9.0 plans its RTX PRO 6000 Server Edition VMs on profiles that fill a whole GPU instance, as our RTX PRO 6000 Server Edition vGPU profiles set out per release.

Three lists to check before you order

Three lists, kept by two vendors, decide whether a GPU server is a supported vSphere host.

The Broadcom Compatibility Guide comes first. Broadcom's ESX 9.0 hardware requirements, updated on 19 October 2026, send you there for "a list of supported platforms", for supported processors and for network adapters, and our guide to ESX 9 hardware requirements shows how to check each host. A GPU server is often a different model, from a different maker, than the existing hosts, and its listing for the ESX release you run is the first thing to confirm. For Private AI Foundation, Broadcom also asks you to verify the GPUs in its Compatibility Guide for AI/ML compute.

NVIDIA's vGPU support matrix comes second: the card and the hypervisor release, as in the table above. NVIDIA's list of NVIDIA-Certified Systems comes third, if NVIDIA AI Enterprise is part of the plan: the 8.2 support matrix states that its GPUs "are supported with servers listed on the NVIDIA-Certified Systems". A server that passes the first two lists but not the third runs vGPU, and has no NVIDIA support path for the AI Enterprise subscription on it.

The configurations behind these lists matter as much as the model name. Server makers list a card for a given riser, slot, power cable and fan kit, and our list of RTX PRO 6000 Server Edition compatible servers gives those details for the common two-socket platforms. We supply GPU servers for vSphere hosts and size them to the release you run; send us your vSphere or VCF release and the card you plan, and our team replies with a configuration and the listings it relies on.

Licensing a GPU host: the cores count, the GPUs do not

Broadcom's KB 313548 sets the rule: license a "minimum of 16 physical cores for each CPU (physical processor)" in every ESXi host. The article counts processor cores only and does not mention GPUs, so four cards add nothing to the VMware licence of a host, while the processors behind them do. The same article gives the bundled vSAN capacity, 1 TiB per VCF core and 0.25 TiB per vSphere Foundation core, rounded up to the next TiB, so a GPU host in a vSAN cluster brings its share of capacity entitlement with its cores.

The second licence question on a GPU host is the edition. NVIDIA's matrix states that "Support for NVIDIA vGPU software requires the vSphere Foundation edition" of ESXi, or a vSphere Enterprise Plus licence. Broadcom's product line comparison for vSphere 9.1 lists Dynamic DirectPath I/O in all three vSphere editions and NVIDIA GRID vGPU in two of them, and states that vSphere Standard and Enterprise Plus are "only available as versions up to the 8 Update 3 release". A vGPU host on vSphere 9 therefore runs vSphere Foundation or VCF. Which tier fits an estate is the subject of our renewal guide to Standard, VVF and VCF.

Our arithmetic. VMware cores: physical cores with at least 16 per CPU (Broadcom KB 313548). NVIDIA minimum: six cores per GPU, NVIDIA-Certified Systems configuration guide (30 September 2026). Last column: seven cores per GPU plus two per GPU for the hypervisor and, for the RTX PRO 6000 Server Edition, two per MIG instance with every card in four instances, per NVIDIA's reference architectures, as worked out in our CPU guide.

The entry host sits exactly on the minimum: two 16-core processors bill 32 cores, and a processor with fewer cores would bill the same. Above that, every core is licensed for the whole subscription term. A four-card host for one inference service needs 24 cores by NVIDIA's guide, and two EPYC 9355 give 64; the headroom pays off once the host runs many VMs, because each vGPU VM brings its own vCPUs, and NVIDIA's reference architecture for MIG adds two cores per instance. Size the cores on the VMs the host will carry, then read the licence from the result. Our guide to the CPU for a GPU server has the per-GPU rules in full.

Two traps from the licence side. Broadcom's KB warns that cores disabled in the BIOS may give inaccurate script results and asks for "all physical CPU cores" to be active when counting, so switching cores off is no way to plan a smaller count; our guide to the Broadcom core count script shows how the count is taken. And the cores on the renewal quote should match what the hosts present: check the figure per host before the GPU servers are added to the subscription.

A dedicated GPU cluster or GPU hosts in the existing cluster

Several host-level rules push GPU servers into a cluster of their own.

1. CPU vendor. vMotion requires processors of the same vendor class, and Broadcom's KB 313545 states that "An EVC-enabled cluster only allows CPUs from a single vendor". AMD EPYC GPU hosts cannot join an Intel cluster with live migration between them, as our VMware host refresh guide explains.
2. Cluster image. With vSphere Lifecycle Manager images, the vGPU Manager becomes part of the cluster image, and Broadcom's KB 424294, cited in the same guide, supports images for heterogeneous hardware clusters only from ESX 9. On ESXi 8, a GPU server from a different maker than the other hosts therefore needs a cluster with its own image.
3. Mobility. A vGPU VM moves only to a host with the same GPU type, and the DRS settings for vGPU migration are cluster options that apply to every VM in the cluster.
4. HA. We found nothing about GPUs in Broadcom's HA admission control, which calculates on CPU and memory, so the spare GPU capacity for a failed or patched host has to be planned by hand.
5. Private AI Foundation. Broadcom's requirements ask for at least three GPU-enabled ESX hosts in the initial cluster of the workload domain, as our Private AI Foundation guide quotes.

The result for most estates: a separate cluster of at least three identical GPU hosts when vGPU is used or Private AI Foundation is planned, one card type per cluster, and one host's worth of free GPUs kept for maintenance. A single passthrough host for one team can live in an existing cluster of the same CPU vendor; its VMs are shut down before the host enters maintenance mode, because DirectPath I/O VMs do not migrate. Cluster sizing beyond GPUs follows our guide to host consolidation under per-core licensing.

Host driver branches over the life of the host

NVIDIA's vGPU release table, read on 24 October 2026, gives two active branches: vGPU 19 on the R580 driver branch, long-term support until July 2028, latest release 19.6; and vGPU 20 on R595, a production branch until March 2027, latest release 20.2. Release 20.2 pairs vGPU Manager 595.91.04 with guest drivers 595.91.07 for Linux and 596.86 for Windows. The RTX PRO 4500 Server Edition and the liquid-cooled RTX PRO 6000 Server Edition arrived with 20.0 and are not in vGPU 19.

The VMware dates run alongside. Broadcom ends general support for vSphere 8 on 11 October 2027 (KB 459135), as our article on vSphere 8 end of general support details. A GPU host bought now on ESXi 8.0 Update 3g reaches that date with vGPU 19 still supported, so the host plan should include the step to VCF 9 within the hardware's life; the procedure is in our vSphere 8 to VCF 9 upgrade steps. Upgrade the host driver before the guest drivers, and move vGPU VMs only to hosts on the same or a newer release, as NVIDIA's migration rules require.

Power and rack positions for GPU hosts

GPU hosts draw more than the virtualisation hosts around them. By our estimate, cards plus processors plus 15 per cent for the rest, a two-card RTX PRO 6000 Server Edition host draws about 1.8 kW, a four-card host about 3.4 kW at 600 W per card or 2.7 kW at the 450 W cap, and an eight-card host about 6.3 kW; Dell gives about 8 kW per eight-card node as an indicative figure, and feeds should be planned on the maker's figure. At 80 per cent of a 230 V 16 A feed, 2.94 kW, a four-card host at full power already needs 32 A or three-phase. A measurement on the delivered system confirms or replaces these estimates. Our guide to GPU rack power and cooling covers feeds and airflow per rack.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with RTX PRO 6000 Server Edition, RTX PRO 4500 Server Edition, L40S, L4 or H200 NVL cards, assembled and tested before delivery, with manufacturer warranty on every component on one EU contract and invoice, and put NVIDIA AI Enterprise and vGPU licences on the same quote. VMware and Broadcom licensing is part of what we supply and implement, as our software page states, and for the cluster itself our engineering partner Vixen.UNO delivers VMware optimisation: an audit of the estate and its licences, editions matched to the real workloads and modernisation with a rollback plan. Send us your current host count, release and renewal date, and we propose GPU hosts that fit the cluster and the licence.

FAQ