VCF 9.1 Deployment · Part 13 of 17

Status in a VCF 9.1 deployment: Optional. Private AI Foundation with NVIDIA is an add-on solution, not a base component.
What it does: It turns a GPU accelerated workload domain into a self service platform for deep learning VMs, GPU backed Kubernetes clusters, and generative AI services.
Depends on: A healthy VCF 9.1 management domain, a GPU enabled VI workload domain with vSphere Supervisor, VCF Operations, VCF Automation, and a VMware Private AI Foundation with NVIDIA add-on license.

TL;DR

• Prerequisites: VCF 9.1, a VMware Private AI Foundation with NVIDIA add-on license, at least 3 GPU enabled ESX hosts, and for vGPU an NVIDIA AI Enterprise (NVAIE) license plus an NGC API key.
• Components: a GPU enabled VI workload domain, vSphere Supervisor, VM classes with a vGPU profile, a content library, and the VCF Automation Private AI Foundation Quickstart.
• The deep learning VM image ships with NVIDIA data center GPU driver 580.95.05 and reads the gpu-driver-url OVF property for disconnected sites.
• Assign the add-on license to the management domain to activate the guided setup UI, and to the GPU workload domain to activate the solution.
• Run the Quickstart wizard once and it publishes 5 catalog items: AI Workstation, RAG Workstation, Triton Inference Server, AI Kubernetes Cluster, and AI Kubernetes RAG Cluster.
• Verify in the Private AI Foundation panel in the vSphere Client and by deploying a test deep learning VM.

This part shows you how to deploy VMware Private AI Foundation with NVIDIA on top of a running VCF 9.1 instance. You will assign the add-on license, prepare a GPU accelerated workload domain, configure vGPU on the ESX hosts, create GPU backed VM classes, run the guided setup in the vSphere Client, and finish in VCF Automation by running the Private AI Foundation Quickstart so your users get ready to deploy catalog items.

Private AI Foundation is optional in a VCF 9.1 deployment. You add it only when you plan to run AI and machine learning workloads. Before you start, confirm that your management domain is healthy, that you have already created a VI workload domain for the GPU hosts, and that vSphere Supervisor and VKS are available in that domain. You will also need VCF Automation deployed, because the self service catalog items are created there.

Confirm every item below before you assign the license. A missing driver VIB or an undersized cluster stops the guided setup partway through.

Requirement	Value or version	Notes
VCF release	VCF 9.1	Management domain healthy, SDDC lifecycle green.
Add-on license	VMware Private AI Foundation with NVIDIA	Activates catalog items, DSM pgvector, the DLVM image, and the guided UI.
GPU hosts	3 or more GPU enabled ESX hosts	Included in the initial cluster of the workload domain.
vGPU license	NVIDIA AI Enterprise (NVAIE)	Required only for vGPU mode, not for passthrough.
Host driver	NVIDIA vGPU host driver VIB	Download from nvid.nvidia.com and add to a vSphere Lifecycle Manager image.
Guest driver	Data center GPU driver 580.95.05	Default in the deep learning VM image; overridable with gpu-driver-url.
Registry access	NGC API key	Needed to pull production branch NGC containers such as nvcr.io/nvidia/pytorch-pb25h1.
Platform services	Supervisor, VCF Automation, VCF Operations	Supervisor enabled on the GPU cluster; VCF Automation reachable.

Note: Check your GPU model against the Broadcom Compatibility Guide for GPUs and Accelerators for AI/ML Compute before you buy or rack hardware. A GPU that is not listed for AI and machine learning use will not pass validation.

You assign the add-on license in two places. Assigning it to the management domain activates the guided setup UI in the vSphere Client. Assigning it to the GPU workload domain activates the solution features on that domain.

1. Log in to the management vCenter with the vSphere Client at https://<management_vcenter>/ui as a Single Sign-On administrator.
2. Open the main menu and click Administration.
3. Under Licensing, click Licenses.
4. Click the Add button and paste the VMware Private AI Foundation with NVIDIA add-on license key.
5. Click Next, give the key a recognizable name, then click Finish.
6. Open the Assets tab, select the management domain vCenter, then click Assign License.
7. Select the add-on license key and click OK.
8. Repeat the assignment against the GPU workload domain vCenter so the solution features are active where the AI workloads run.

Private AI Foundation runs in a VI workload domain, not in the management domain. Use an existing GPU domain or create a new one with the GPU hosts in the initial cluster.

1. Log in to SDDC Manager at https://<sddc_manager_fqdn> as an administrator.
2. In the left navigation, click Workload Domains.
3. Confirm the target GPU domain is present and its status reads Active, or click Add Workload Domain to create one.
4. If you are creating a domain, select VI as the type, choose vSAN or the storage you prepared, and add at least 3 GPU enabled hosts to the initial cluster.
5. Enter the cluster name, the vCenter FQDN and IP, the NSX details, and click Next through each screen.
6. Review the summary and click Finish, then wait for the workflow to report Successful.

For a full walkthrough of the domain wizard fields, see the earlier part on how to create a VI workload domain in VCF 9.1.

Choose vGPU when you want to share a GPU across several workloads, or passthrough when a workload needs the full device. This step covers the vGPU path, which needs the host driver VIB in a vSphere Lifecycle Manager image. For a deeper look at the driver stack, see the walkthrough on the NVIDIA GPU Operator and vGPU drivers.

1. On each ESX host, open the hardware vendor BIOS and enable SR-IOV, then save and reboot.
2. Open the VMware Host Client for each host and enable the graphics devices for AI operations.
3. Download the NVIDIA vGPU host driver VIB that matches your VCF version from nvid.nvidia.com.
4. In SDDC Manager, go to Lifecycle Management, then Image Management, and import the VIB into a vSphere Lifecycle Manager image.
5. In the vSphere Client, select the GPU cluster, open Updates, then Image, and attach the image that contains the vGPU host manager VIB.
6. Click Remediate All and wait for every host to report compliant.
7. If you plan to use Multi-Instance GPU sharing, enable MIG on each host in the cluster.
8. Confirm the driver on one host by running the command below over SSH.

nvidia-smi -q | grep "Driver Version"

A VM class defines the compute, memory, and GPU profile that AI workloads request. Create at least one class that carries a vGPU profile so the catalog items have something to place on.

1. In the vSphere Client, open Workload Management.
2. Click the Services tab, then open VM Classes.
3. Click Create VM Class.
4. Enter a name such as gpu-a100-2 and set the vCPU and memory values that match your GPU host, for example 16 vCPU and 64 GB.
5. Turn on PCI Devices, click Add PCI Device, and select NVIDIA vGPU.
6. Choose the vGPU profile that your driver exposes, for example a full profile or a MIG slice, then set the GPU memory.
7. Click Next, review the class, and click Finish.
8. Open the target namespace, click Add VM Class, select the new class, and click OK so it is available for provisioning.

Catalog items need the VMware deep learning VM image and a vSphere Kubernetes release. A connected site pulls these automatically. A disconnected site needs a local content library that you populate by hand.

1. In the vSphere Client, open Content Libraries and click Create.
2. Enter a name such as paif-images, select the workload domain vCenter, and click Next.
3. For a connected site, choose Subscribed content library and enter the VMware subscription URL. For a disconnected site, choose Local content library.
4. Select the datastore for the library and click Finish.
5. For a disconnected site, upload the deep learning VM image and the Ubuntu VKr manually, then associate the library with the Supervisor.
6. Confirm the Supervisor is enabled on the GPU cluster in Workload Management, because the catalog items provision into a Supervisor namespace.

Tip: For disconnected sites, the deep learning VM image reads the gpu-driver-url OVF property so it can fetch the guest driver from a local web server instead of the internet.

Once the license is on the management domain, a Private AI Foundation panel appears in the vSphere Client. That panel connects to SDDC Manager and walks you through the remaining infrastructure tasks in one place.

1. Log in to the management vCenter with the vSphere Client as a Single Sign-On administrator.
2. In the side panel, click Private AI Foundation.
3. On the guided deployment page, review each task card and its status.
4. Work through the cards in order: workload domain, vGPU on hosts, VM classes, content library, and VCF Automation setup.
5. For any card that is not complete, click its action link, complete the wizard it opens, and return to the panel.
6. Confirm every card reports a green complete state before you move on.

This final step publishes the self service catalog items. You point the Quickstart at a project and namespace that already carry the GPU VM classes and the Private AI images, answer a few questions, and it generates the templates and publishes the items.

1. Log in to VCF Automation as an organization administrator.
2. Open the Private AI Foundation Quickstart from the Automation home page.
3. On the first screen, select the project and the Supervisor namespace that hold the GPU VM classes and Private AI images.
4. Enter the NGC API key so the workloads can pull production branch containers.
5. Select the vGPU VM classes to expose and the content library that holds the deep learning VM image.
6. If you plan to run retrieval workloads, enable the pgvector option so Data Services Manager provisions the PostgreSQL database.
7. Click Next, review the summary, then click Run or Finish to generate the templates.
8. Wait for the wizard to report success, then open Service Broker and confirm the 5 catalog items are present.

Run these checks in order. Each one confirms a different layer, from license to running workload.

Check	Where	Expected result
License active	vSphere Client, Licenses	Add-on key assigned to both the management and GPU domains.
Guided panel	vSphere Client side panel	Private AI Foundation panel visible, all cards complete.
vGPU host driver	ESX host over SSH	nvidia-smi returns a driver version and lists the GPU.
VM class	Workload Management, VM Classes	A class with a vGPU profile is bound to the namespace.
Catalog items	VCF Automation Service Broker	Five items present, including AI Workstation and AI Kubernetes Cluster.
Test deploy	Service Broker	An AI Workstation deep learning VM powers on and reports the GPU.

For the final check, request the AI Workstation catalog item with a small vGPU class, wait for it to finish, then open the VM console and run nvidia-smi inside the guest to confirm the driver loaded and the GPU is visible.

Four failures account for most stalled deployments. Each has a clear cause and fix.

Symptom	Cause	Fix
Private AI Foundation panel missing	Add-on license not on the management domain	Assign the add-on key to the management vCenter, then refresh the vSphere Client.
Hosts fail remediation	Wrong vGPU host driver VIB for the VCF version	Download the matching VIB from nvid.nvidia.com, rebuild the image, and remediate again.
VM class shows no vGPU option	SR-IOV or graphics not enabled on the host	Enable SR-IOV in BIOS, enable the graphics device in the Host Client, and reboot.
Deep learning VM cannot pull containers	Missing or wrong NGC API key	Re-enter a valid NGC API key in the Quickstart and redeploy the item.

Is Private AI Foundation mandatory in VCF 9.1. No. It is an optional add-on that you deploy only when you plan to run AI and machine learning workloads. A base VCF 9.1 instance runs without it.

Do you always need an NVIDIA AI Enterprise license. Only for vGPU mode. GPU passthrough does not need NVAIE, but a passthrough workload takes the whole device and cannot use vMotion.

How many GPU hosts do you need. At least 3 GPU enabled ESX hosts in the initial cluster of the workload domain.

What does the Quickstart actually create. It publishes 5 catalog items: AI Workstation, RAG Workstation, Triton Inference Server, AI Kubernetes Cluster, and AI Kubernetes RAG Cluster.

Can you run this without internet access. Yes. Use a local content library, a private Harbor registry, and the gpu-driver-url OVF property to serve the guest driver from an internal host.

• Preparing VCF for Private AI Workload Deployment, TechDocs 9.1
• Requirements for Deploying VMware Private AI Foundation with NVIDIA, TechDocs 9.1
• Follow the Guided Setup in the vSphere Client, TechDocs 9.1
• Install VMware Private AI Foundation with NVIDIA using VCF Automation, VCF Blog

« Previous: How to Enable vSphere Supervisor and VKS in VCF 9.1 (Part 12)
★ VCF 9.1 Deployment Complete Guide
Next: How to Configure a Stretched Cluster in VCF 9.1 (Part 14) »