NVIDIA AI Enterprise: what the licence includes, how it is counted, and when you actually need it

What you are actually buying

NVIDIA AI Enterprise is a per-GPU entitlement to a software platform organised in two layers. The application layer holds NIM microservices, NeMo, Triton Inference Server, TensorRT and TensorRT-LLM, RAPIDS, the PyTorch container, Riva, DeepStream, Holoscan, TAO, Omniverse and NVIDIA’s pre-trained models, released in feature, production and long-term branches. The infrastructure layer holds the data-centre driver, the vGPU manager and guest driver, Fabric Manager, the GPU, Network, DPU and NIM operators for Kubernetes, Base Command Manager, Run:ai for self-hosted deployments and the container toolkit.

Three things are exclusive to licence holders, according to NVIDIA’s own licensing guide: select NIM containers, the production and long-term support branches, and vGPU for Compute. Everything else in the list also exists in open-source or freely downloadable form. The fourth thing the licence buys is enterprise support with response targets, which is what most buyers are really paying for. This is the software layer behind the AI Enterprise page in our catalogue; this article is about how it is counted and when you can skip it.

How it is counted

The rule is short: one licence for every physical GPU in a server or workstation that hosts any AI Enterprise software. Boards with several GPUs count each GPU, so an A16 needs four licences. A CPU-only server or instance running the CPU parts of the stack needs one licence per server, regardless of how many processors it has. MIG slices are not mentioned in the licensing guide at all: partitioning a card into four instances does not change the count, which stays one per physical GPU.

Response targets per NVIDIA’s Enterprise Support and Services guide; support requires an NVIDIA-Certified System.

Two conditions sit underneath. Support is only available on NVIDIA-Certified Systems (Qualified Systems for the NVL72 racks), so a card in an uncertified chassis has a licence but no support path. And a subscription buys time, not longevity of a branch: a five-year subscription does not extend the life of the software branch you deploy.

What comes with the hardware, and what does not

This is where quotes go wrong. Some GPUs include a subscription in the price, others include nothing, and the difference is worth reading before the L40S and H200 NVL are compared on card price alone.

Per NVIDIA’s licensing guide and product pages, September 2026. OEM promotional bundles for the RTX PRO 6000 Server Edition exist; they are a reseller offer, not an NVIDIA entitlement.

The H200 NVL line is the one that changes comparisons. If a project needs AI Enterprise anyway, for vGPU compute profiles or for supported NIM in production, the Hopper card carries five years of it, and the RTX PRO 6000 comparison has to add eight licences to an eight-card server before the totals are honest.

The one part that is enforced: vGPU for Compute

Most of the platform runs whether or not a licence server can be reached: NVIDIA’s own words are that the software “runs with or without a valid license server connection”, and the licence system is only required when vGPU for Compute drivers are in use. That one case is enforced in the driver, and it is the case that catches virtualisation teams.

C-series profiles, the compute vGPU types such as L40S-48C or DC-8C on the RTX PRO 6000 Server Edition, are licensed only through AI Enterprise. The RTX vWS, vPC and vApps licences that virtual-desktop teams already own do not unlock them. Each C-series vGPU assigned to a VM checks out a licence at boot and returns it at shutdown; one per-GPU licence covers up to 16 vGPU instances on that GPU, or one vGPU that takes the whole card. NVIDIA’s H200 NVL profile tables allow 20 or 32 small vGPUs per card, and the guide does not say how those are licensed, so ask before designing more than 16 tenants on one GPU.

The licence server itself comes in two forms. The Cloud License Service is hosted on NVIDIA’s licensing portal and needs outbound HTTPS. The Delegated License Service is a virtual appliance on your network, clusterable up to nine instances, with node-locked licences for air-gapped hosts available since vGPU software 15.0. Clients carry a configuration token that is valid for years; the day-to-day risk is not the token but a firewall change that cuts the VMs off from the server, which you notice seven days later.

What you can run without it

More than the brochure suggests. The CUDA toolkit is free under its EULA. TensorRT-LLM and vLLM are Apache-2.0, Triton Inference Server is BSD-3; NeMo is open source. A single-tenant inference server running vLLM on eight RTX PRO 6000 cards needs no AI Enterprise licence to run, and many do exactly that.

What changes without the licence: NIM containers are available to developers for research and testing on up to 16 GPUs, but NVIDIA states that production use requires AI Enterprise, and some NIMs are licence-only; you get the monthly feature branch rather than a production branch with nine months of stable APIs; and there is no NVIDIA support contract behind a driver bug at two in the morning. A 90-day trial licence exists for the move to production; it needs an NVIDIA-Certified server and excludes Run:ai.

The honest split for a buyer: one or two GPUs, one team, open-source serving stack, no VMs sharing the card, then the licence is optional and mainly buys support. Fractional GPUs in virtual machines for several teams, or NIM in production, then it is not optional, and the count is eight per eight-card server whatever the cards cost.

Per user or per GPU: the two licence families

The same card can carry two different licence models depending on the profile type, and mixing them up is the most common mistake we see on quotes.

The vGPU packaging guide sends compute workloads to the AI Enterprise guide; a Q-series and a C-series profile on the same card are two different licences.

Our guide on MIG and vGPU counts covers how many VMs each card holds; this table is about which budget line each VM lands on.

Branches and dates you have to plan around

AI Enterprise separates the infrastructure layer from the applications, and each has its own branch cadence. Feature branches ship monthly with one month of support and may change APIs. Production branches ship every six months with nine months of support and monthly security patches. Long-term support branches arrive roughly every two and a half years, timed to a new GPU architecture, with three years of support and quarterly patches.

Per NVIDIA’s lifecycle pages, September 2026. Infrastructure 8.0 dropped V100, RTX A4000, RTX 4000 SFF Ada and the Quadro RTX cards; they stay on 7.x.

Two consequences. A regulated environment that wants three-year API stability belongs on the long-term branch on infrastructure 7.x, and both of those end within 2027 and 2028 respectively, so the upgrade is already on the calendar. And the Triton container with the TensorRT-LLM backend is gone from production branch 6; deployments move to the vLLM backend variant.

Where the partner platforms put the licence

VMware Private AI Foundation with NVIDIA on VCF 9.1 needs three entitlements: the VCF subscription, the Private AI Foundation add-on, and AI Enterprise for the vGPU host driver and guest drivers, plus a licence server or portal token; the passthrough path is documented without the AI Enterprise requirement, and the minimum is three GPU hosts in the first cluster. Red Hat OpenShift AI enables NIM with an NGC key from an account that carries the AI Enterprise Viewer role, so the entitlement comes from NVIDIA, not from Red Hat. Nutanix Enterprise AI consumes NIM with your own NIM token and states that NIM is licensed through AI Enterprise. In every case the per-GPU licence is a separate line from the platform, and the VMware core count is a separate line again.

The traps, in the order we meet them

Cards ordered without the licence, then vGPU compute profiles will not start; the fix is a per-GPU purchase, not a per-VM one. Counting VMs instead of GPUs, or assuming the vWS licences already on the estate cover AI workloads. Multi-GPU boards counted as one. Included subscriptions never activated because nobody registered the serial numbers. A DGX Spark bought on the assumption that the department’s AI Enterprise licence covers it; NVIDIA says other entitlements do not. Hypervisors outside the support matrix: version 8.2 lists ESXi, Red Hat and Ubuntu KVM and Nutanix AHV; Proxmox, Hyper-V and XenServer are not in the table. And DOCA on BlueField adapters, which needs its own adapter support entitlement on top of AI Enterprise.

What we supply

Eurokommerz sells NVIDIA AI Enterprise subscriptions and perpetual licences with the GPUs they belong to, across the EU, and we put the entitlement on the quote in words: which GPUs carry an included subscription, which need a purchased one, and which support tier applies. For the H200 NVL we activate the five-year subscription with you; for RTX PRO 6000 servers we size the licence count against the vGPU plan, not against the number of cards someone hoped to share.

FAQ