NVIDIA AI Enterprise on VMware vSphere: the licence for vGPU and what passthrough changes

What NVIDIA AI Enterprise licenses on VMware vSphere

On VMware vSphere, NVIDIA AI Enterprise is the licence for compute vGPU, the C-series profiles that give a VM a share of a GPU for CUDA workloads, and for the supported NVIDIA AI software that runs on those GPUs. NVIDIA licenses it per GPU. Graphics vGPU for virtual desktops is a separate product with licences of its own, RTX vWS, vPC and vApps, counted per concurrent user. A GPU handed whole to a VM by DirectPath I/O on ESXi needs no vGPU host driver, and Broadcom documents it without an NVIDIA AI Enterprise licence, yet NVIDIA's licensing guide still counts every GPU in a server that hosts NVIDIA AI Enterprise software.

This guide covers the vSphere side of the licence. Prices and terms are in our guide to NVIDIA AI Enterprise licensing, the per-user editions in NVIDIA vGPU licensing for RTX vWS, vPC and vApps, and the host side of a cluster in GPU servers in a vSphere cluster. All documents cited were read on 12 November 2026.

NVIDIA AI Enterprise Infra 8, licensing page of vGPU for Compute, and NVIDIA Enterprise Licensing Guide (both 2 September 2026); NVIDIA vGPU Client Licensing User Guide 21.0 (5 November 2026); NVIDIA vGPU packaging, pricing and licensing guide (24 March 2026); read 12 November 2026. The passthrough rows are explained in the section on DirectPath I/O.

Two product lines with separate releases

NVIDIA documents the two families separately. Graphics vGPU is "NVIDIA vGPU software", with the release families 19, 20 and 21. Compute vGPU is "NVIDIA vGPU for Compute" in the infrastructure software of NVIDIA AI Enterprise, releases 7.x and 8.x, with host and guest drivers of its own. The vSphere pages of vGPU 21.0 do not mention NVIDIA AI Enterprise, so a compute host is planned from the NVIDIA AI Enterprise support matrix.

NVIDIA AI Enterprise documentation hub and support matrices 8.2 (2 September 2026) and 7.8 (12 August 2026); NVIDIA vGPU documentation index and the vSphere release notes of vGPU 21.0 (5 November 2026), 20 (26 October 2026) and 19 (23 October 2026); read 12 November 2026. NVIDIA's servers returned two copies of the vGPU index on that day, and the older copy still listed 20.2 and 19.6.

NVIDIA AI Enterprise has no release on the R620 driver branch yet: its documentation hub marks 8.2 as the latest and lists no 8.3 and no release 9. A compute host therefore runs an R595 or R580 host driver today, even where desktop hosts move to vGPU 21.0.

The two document sets word the VMware support differently. The NVIDIA AI Enterprise matrix says "8.0 and later, 9.0 and later". The vGPU pages require ESXi "8.0u3 P06 and later updates to release 8.0 unless explicitly stated otherwise" and, for the RTX PRO 6000 and RTX PRO 4500 Blackwell Server Editions, VCF 9.0.1 or later on the 9.0 line and ESXi 8 Update 3g or a later update on the 8.0 line. Plan the hosts on the stricter figures.

The vGPU 21.0 page adds two conditions. ESXi 9.x, it says, "is distributed only as part of VMware Cloud Foundation (VCF)", and NVIDIA vGPU software requires the vSphere Foundation edition of ESXi or a vSphere Enterprise Plus licence. Broadcom ends general support for vSphere 8 on 11 October 2027 (KB 459135, read on 5 November 2026); the details are in our article on the end of general support for vSphere 8.

Host software on ESXi and drivers in the guest

Compute vGPU needs NVIDIA software on both sides of the hypervisor. On the host it is the vGPU Manager, which NVIDIA's VMware deployment guide for AI Enterprise (31 July 2026) calls the "NVIDIA AI Enterprise Host Software (VIB)". It is installed with esxcli while all VMs are powered off and the host is in maintenance mode. Afterwards "the default graphics type is Shared, which provides vSGA functionality"; for vGPU it has to be changed to Shared Direct. The versions to install are in the support matrix; the file names in the guide's examples are demonstrations.

NVIDIA AI Enterprise support matrices 8.2 (2 September 2026) and 7.8 (12 August 2026), rows for the vGPU Manager and the vGPU for Compute guest driver; NVIDIA vGPU 21.0 release notes for VMware vSphere (5 November 2026); read 12 November 2026.

Host and guest releases have to be compatible, which is not the same as identical. NVIDIA states that the 21.0 vGPU Manager works with guest drivers of 21.0 and of all 20.x and 19.x releases, while 21.0 guest drivers are incompatible with every vGPU Manager release of 20.x and earlier. An upgrade therefore starts on the hosts.

The server counts as well: the 8.2 matrix states that its GPUs "are supported with servers listed on the NVIDIA-Certified Systems", a list explained in our article on NVIDIA-Certified Systems.

How the compute vGPU licence is counted

Two NVIDIA pages of 2 September 2026 describe the count, and they describe different things. The licensing guide sets the rule: "NVIDIA AI Enterprise software is licensed on a per-GPU basis", with a licence for every GPU installed in a server or workstation that hosts any of its software. The licensing page of vGPU for Compute describes what the driver checks: "For vGPU for Compute, software enforces one license per vGPU assigned to a VM." It continues: "That license covers up to 16 vGPU instances on a single GPU, or one vGPU that uses the entire physical GPU framebuffer."

NVIDIA does not reconcile the two statements. By our reading, a card split into up to 16 compute vGPUs stays within one licence. The RTX PRO 6000 Blackwell Server Edition ends at 12 vGPUs per card with the DC-8C type. The H200 NVL reaches 20 and 32 with the H200-7C and H200-4C types; for such a layout, have the count confirmed in the quote.

A VM requests its licence at boot and returns it at shutdown; without one, a compute VM runs at full capability for 20 minutes and then at an idle level of compute. The licences come from the NVIDIA License System, at version 4.0.0 since October 2026, as the cloud-hosted CLS or as a DLS instance on your own network, both described in our guide to the NVIDIA licence server.

We supply NVIDIA AI Enterprise and vGPU licences with the cards and servers they run on; send us the card models, the number of hosts and your ESXi or VCF release and we answer with the licence lines in a written quote within one business day.

Cards for compute vGPU on vSphere

A card runs compute vGPU on vSphere when the NVIDIA AI Enterprise matrix lists it and NVIDIA documents C-series types for it. NVIDIA's general vGPU list usually agrees; the H200 NVL is the exception.

NVIDIA, "GPUs supported by vGPU" (copies of 2 and 29 October 2026 with the same rows for these cards); vGPU 21.0 release notes for VMware vSphere (5 November 2026); NVIDIA AI Enterprise 8.2 support matrix and the vGPU for Compute reference pages for Blackwell, H200 and Ada Lovelace (2 September 2026); read 12 November 2026. "Per card" is the number of vGPUs of one type on one GPU.

The H200 NVL is absent from the vGPU list because it is a compute-only card: NVIDIA's list notes that "Support for GPUs that support only C-series vGPUs is withdrawn in vGPU software 16.0" and that "These GPUs continue to be supported with NVIDIA AI Enterprise." Its C-series types are documented in release 8.2, where four MIG-backed types carry the footnote "on ESXi, starting with vSphere 8.0 update 3". The accurate wording for the H200 NVL is compute vGPU through NVIDIA AI Enterprise. NVIDIA's licensing guide also states that each H200 NVL includes a five-year NVIDIA AI Enterprise subscription, which has to be activated; it names no such inclusion for RTX PRO cards.

Two cards of the table have their own pages here: the NVIDIA RTX PRO 4500 Blackwell Server Edition (buy page, specifications) and the NVIDIA RTX 5000 Ada Generation (buy page, specifications). The Server Edition is in NVIDIA AI Enterprise 8.2 and in vGPU 20 and 21; release 7.8 and vGPU 19 do not list it. The RTX 5000 Ada is a workstation card whose display ports must be turned off for vGPU, as NVIDIA notes. Of the eleven professional cards with card pages here, these two are the only ones on NVIDIA's vGPU list: the RTX PRO 4500 Workstation Edition, the RTX PRO 4000 and RTX PRO 2000, the RTX 4500 Ada, RTX 2000 Ada and RTX 2000E Ada, the RTX A1000 and the RTX A400 are not on it, and neither are the RTX PRO 6000 Workstation Edition and Max-Q. The RTX PRO 5000 Blackwell is on it only as the 72 GB card, from release 20.2, and the vSphere lists of 20.x and 21.0 do not contain that card; the 21.0 documents list it for Red Hat Enterprise Linux with KVM.

In vGPU 21.0, time-sliced vGPUs inside a MIG instance need VCF 9.1 or later. Profile tables are in our guides to RTX PRO 6000 Server Edition vGPU profiles and RTX PRO 4500 Server Edition vGPU profiles, the VM counts in MIG and vGPU: how many VMs one card holds.

GPU passthrough with DirectPath I/O, without vGPU

With DirectPath I/O a VM owns the whole GPU, and no NVIDIA software runs in the hypervisor. Two vendors describe the licence consequence, each for its own layer.

Broadcom's documentation of VMware Private AI Foundation with NVIDIA 9.1 sets the two modes side by side on its page "Preparing VCF for Private AI Workload Deployment", last updated on 4 November 2026. For vGPU: "You need an NVIDIA AI Enterprise license." For GPU passthrough: "You do not need an NVIDIA AI Enterprise license." It adds that the cluster image in vSphere Lifecycle Manager then needs no NVIDIA VIB file for the vGPU host manager driver, that live workloads cannot be migrated across hosts with vSphere vMotion, and that deep-learning workloads from NVIDIA then run from the feature branch instead of the production branch.

NVIDIA's licensing guide requires a licence "for every GPU installed on the server or workstation that will host any software" of NVIDIA AI Enterprise, and its software licence agreement, last modified on 7 May 2026, counts, for on-premises deployments, the physical GPUs in the computing environment that the product accesses. The 8.2 support matrix lets a passthrough guest run the NVIDIA data-centre driver or the vGPU guest driver. NVIDIA support for the software comes only with a subscription.

By our reading, the two statements answer different questions. Broadcom's sentence concerns the path of the GPU into the VM: a compute VM on passthrough uses no vGPU host driver and no licence server. NVIDIA's rule concerns the software in the VM: a server whose GPUs run NVIDIA AI Enterprise software counts those GPUs, on vSphere as on bare metal. Which NVIDIA components carry no licence fee is set out in CUDA, TensorRT, NIM: free or NVIDIA AI Enterprise.

Graphics is a separate case. NVIDIA's client licensing guide for vGPU 21.0 treats GPU pass-through, VMware vDGA and bare metal alike: for workstation or professional 3D graphics they need an RTX vWS licence, enforced through software, and for PC-level applications they fall under vApps, enforced only through the licence agreement. A CAD workstation VM on a passthrough card is therefore an RTX vWS seat.

Passthrough is also the only GPU mode NVIDIA Run:ai accepts in virtual machines. Its system requirements say: "vGPU is not supported. NVIDIA Run:ai currently supports GPU passthrough only." The details are in NVIDIA Run:ai licensing and hardware requirements. What passthrough costs a VM in vMotion, snapshots and DRS is covered in GPUs in VMware vSphere: passthrough or vGPU.

The NVIDIA licence in VMware Private AI Foundation with NVIDIA

Where the GPU hosts belong to VMware Private AI Foundation with NVIDIA, Broadcom's requirements name three licences: a VCF subscription, a VMware Private AI Foundation with NVIDIA add-on licence and, for AI workloads on vGPU, an NVIDIA AI Enterprise licence for the host driver VIB on the ESX hosts and for the guest drivers. Release 9.1 runs on VCF 9.1 and asks for at least three GPU-enabled ESX hosts in the initial cluster of a workload domain.

The add-on licence is still needed where the product is part of VCF for a region, for functions such as the Private AI setup in VCF Automation, the deep learning VM image and Private AI Services. The deep learning VM of release 9.1.1 supports the vGPU host drivers of NVIDIA AI Enterprise 8.0 and later (595.x.x), the image of 9.1 those of 7.x (580.x.x) and of 8.0. See our guide to VMware Private AI Foundation with NVIDIA.

Before you order licences for a vSphere GPU host

1. Decide for each kind of VM how it gets its GPU: compute vGPU, graphics vGPU or passthrough. That fixes the licence family.
2. Check the ESXi or VCF release against the vGPU floor, and check the vSphere edition.
3. Choose the NVIDIA AI Enterprise release, 8.2 or 7.8, by its end of support. The RTX PRO 4500 Server Edition needs 8.2.
4. Count the GPUs of the compute hosts. Above 16 compute vGPUs per card, ask for the count in writing.
5. Count the peak of concurrent users for the graphics editions.
6. Check the server model against the NVIDIA-Certified Systems list, and name every card with its edition; our article on NVIDIA part numbers shows what to state.

These six points are also what a quote needs. Tell us which VMs need a GPU, in which mode, and how many cards each host holds, and the written quote places the NVIDIA licence lines beside the hardware.

FAQ