TL;DR · Key Takeaways
- Private AI Foundation is not one SKU. You buy three things: VCF (per core), the Private AI Foundation add-on (per core), and NVIDIA AI Enterprise (per GPU), and the last one comes from NVIDIA, not Broadcom.
- The budgeting mistake that bites every first project: treating this as a single line item. Cores and GPUs are two independent billing axes, and they scale differently.
- H100 PCIe/NVL and H200 NVL ship with a 5-year NVAIE subscription already included. Buying NVAIE again for those GPUs is wasted money.
- That bundled subscription clock starts at the GPU ship date plus 90 days, not when you rack it. GPUs sitting in a lab are burning entitlement.
- My verdict: subscription over perpetual for almost everyone, and let the GPU choice decide your NVAIE strategy.
Here is the conversation that derails more Private AI projects than any GPU shortage: finance asks for a single cost-per-server number, an architect hands over a VCF core count, and three months later procurement discovers the NVIDIA software was never in the quote. Private AI Foundation licensing is not hard, but it is layered, and the layers are billed on different units by different vendors. Get the model wrong at design time and you either over-buy by a wide margin or stall the project waiting on a purchase order nobody scoped.
This post breaks down exactly what you are buying, compares the choices that actually carry money (per-core add-on, per-GPU NVAIE, subscription versus perpetual, bundled versus standalone), and ends with a clear recommendation instead of a shrug. If you have not yet read what Private AI Foundation actually is or how the platform stacks together layer by layer, start there. This part assumes you know the components and want to know the bill.
The three licenses you are actually buying
Private AI Foundation with NVIDIA (PAIF) is a solution made of three commercial pieces that you assemble. None of them is optional if you want the full platform, and they do not come from the same place.
- VMware Cloud Foundation (VCF). The base private cloud: vSphere, vSAN, NSX, VCF Operations and Automation. Licensed per physical core, with a vSAN capacity entitlement per core. This is the floor everything else sits on. The mechanics of core counting are covered in the VCF 9 licensing breakdown, and they apply unchanged here.
- The Private AI Foundation add-on. A Broadcom add-on layered on top of VCF, also licensed per core. This turns on the Private AI Package: Model Store, Model Runtime, the API Gateway, Agent Builder, the Data Indexing and Retrieval Service, vector database support, and the Deep Learning VM automation. No add-on, no platform services, just plain VCF with GPUs in it.
- NVIDIA AI Enterprise (NVAIE). The NVIDIA software stack: the vGPU host and guest drivers, NIM microservices, NeMo, CUDA libraries, and enterprise support. Licensed per GPU, and purchased from NVIDIA or an NVIDIA partner, not from Broadcom. Broadcom states this plainly in its own launch material: NVAIE licenses must be purchased separately.
Per core vs per GPU: the two axes that confuse everyone
The single biggest source of licensing surprises is that the Broadcom layers and the NVIDIA layer scale on different units, and those units do not move together. A dual-socket host with 64 cores and two GPUs costs 64 core-units of VCF, 64 core-units of the PAIF add-on, and 2 GPU-units of NVAIE. Add a third GPU to that same host and your NVIDIA cost jumps 50 percent while your Broadcom cost does not change at all. Add a second host to scale cores and the reverse happens.
This matters because it changes how you should pack hosts. Dense GPU hosts (more GPUs per socket) are efficient on Broadcom licensing and expensive on NVIDIA licensing. Sparse hosts are the opposite. There is no universally right answer, but there is a wrong one: sizing the cluster on core count alone and discovering the GPU bill afterward.
| Dimension | VCF + PAIF add-on (Broadcom) | NVIDIA AI Enterprise (NVIDIA) |
|---|---|---|
| Billing unit | Physical CPU core | Physical GPU |
| Who you buy from | Broadcom or VMware partner | NVIDIA or NVIDIA NPN partner |
| Purchase models | Subscription (term), per core | Subscription, cloud consumption, or perpetual + 5yr support |
| Scales with | Cores in the cluster running it | GPU count, not GPU size or vGPU profile |
| Can come bundled? | No, always purchased | Yes, included with select GPUs (H100/H200 NVL) |
| Support tier | Broadcom support per contract | Business Standard included; Business Critical extra |
NVAIE: subscription vs perpetual vs bundled-with-GPU
The NVIDIA layer is where most of the real decisions live, because it has three genuinely different acquisition paths and they do not cost the same over time. NVIDIA AI Enterprise is licensed per GPU and sold as a term subscription, as cloud consumption, or as a perpetual license that still requires five years of support services. Indicative partner list pricing has run around 4,500 USD per GPU per year for a subscription, roughly 18,000 USD for five years (the multi-year deal has effectively given the fifth year for free), and about 22,500 USD per GPU for perpetual with its mandatory five-year support. Cloud marketplace consumption sits near 1 USD per GPU per hour. Treat these as reference points and confirm current numbers with your partner, because they move and they discount.
The path that quietly saves the most money is the one nobody quotes: certain data center GPUs ship with NVAIE already attached. Each NVIDIA H100 PCIe and NVL and each H200 NVL Tensor Core GPU includes a five-year NVAIE subscription, and the A800 40GB Active includes three years. If you spec those cards and then also put NVAIE on the purchase order, you are paying twice for the same entitlement on the same silicon.
The licensing gotchas that actually bite
These are the ones I flag in every design review, because each has cost a real project either money or a delivery date.
- The bundled subscription clock starts before you do. The NVAIE entitlement included with a GPU starts at the board ship date to the OEM plus 90 days, and that start date cannot be changed because it is tied to the card. GPUs that sit in a staging rack or a procurement holding pattern for two quarters have already eaten two quarters of a five-year subscription. Order GPUs close to when you will actually deploy them.
- DGX Blackwell does not include what DGX Hopper did. NVAIE is bundled in the DGX software stack for Hopper-architecture DGX systems, but for Blackwell-architecture DGX systems you must purchase NVAIE separately. Teams upgrading their DGX generation assume the software carries over. It does not.
- The add-on is per core, not per GPU host. The Private AI Foundation add-on is licensed against the cores where it runs, which can be the whole cluster, not only the cores in your GPU-bearing hosts. Validate the licensable scope with your account team before you size, because this assumption alone can double or halve the add-on line.
- NVAIE counts physical GPUs, not vGPU slices. Carving one A100 into seven MIG instances or several vGPU profiles does not multiply your NVAIE count. You license the physical GPU. This is good news, and people routinely over-estimate it.
- GPU-less still needs a license. If you run NVAIE software on a host or instance with no NVIDIA GPU, NVIDIA still requires one subscription per server or instance. Edge and CPU-inference corners of a design get missed here.
What I’d Do
The verdict, because a comparison without one is just a brochure. For almost every enterprise standing up Private AI Foundation, go subscription on all three layers and let the GPU choice drive your NVAIE strategy. Spec H100 NVL or H200 NVL where the workload fits and you fold a five-year NVAIE subscription into the hardware at no extra software line, which is the single cleanest cost move available. Only reach for perpetual NVAIE when both your GPU fleet and your AI workload are genuinely locked for five years or more, which in a field moving this fast is rare. Reserve cloud consumption pricing for burst and proof-of-concept, not steady-state, because at roughly 1 USD per GPU per hour the math turns against you within months of continuous use.
Above all, build the budget on two axes from day one: cores for Broadcom, GPUs for NVIDIA, scoped separately, validated separately. The projects that stall are not the ones that picked the wrong tier. They are the ones that treated a two-vendor, two-unit license model as a single number. What is your GPU mix looking like, and have you checked which of those cards already carry NVAIE before the next PO goes out?
References
- VMware Cloud Foundation Blog: Private AI Foundation with NVIDIA on VCF 9.0 (per-core model, NVAIE purchased separately)
- NVIDIA Enterprise Licensing Guide: per-GPU licensing, bundled GPUs, subscription vs perpetual
- VMware Private AI Foundation with NVIDIA solution brief
« Previous: Part 2 | VMware Private AI Complete Guide | Next: Part 4 »
About The Author
