CUDA, TensorRT, NIM and vGPU: which NVIDIA software is free and which needs AI Enterprise

What runs without a licence fee and what needs a subscription

Most of the NVIDIA software on a GPU server costs nothing to use in production: the CUDA Toolkit, cuDNN and TensorRT under NVIDIA's own licence agreements, and TensorRT-LLM, Triton Inference Server, Dynamo, NCCL, the GPU Operator and DCGM as open-source projects. NVIDIA AI Enterprise becomes necessary for three things: NIM microservices in production, compute vGPU profiles in virtual machines, and NVIDIA's enterprise support with its production software branches. A single-tenant inference server running vLLM on bare metal needs no NVIDIA subscription at all.

"Free" here means no licence fee; the terms still apply. Each component comes with a licence text, and those texts say where the software may run, what you may pass on to others and which versions are meant for production. This page goes through them component by component, as read on 24 October 2026. It is general information about the published licence texts, not a legal assessment; whether a specific use is covered is a question for your legal department.

How AI Enterprise is counted, its terms and its support levels are in our guide to NVIDIA AI Enterprise licensing, and activation is in activating NVIDIA AI Enterprise licences. This page answers the question that comes before both: which part of your stack needs the licence in the first place.

Licence by component, October 2026

CUDA EULA (last updated 26 January 2026, toolkit 13.4), licence pages of the cuDNN documentation (updated 2 September 2026) and the TensorRT 11.4 documentation (updated 5 October 2026), NIM General FAQ (6 August 2026) on NVIDIA's documentation sites; licence files and README of the NVIDIA GitHub repositories for open-gpu-kernel-modules, TensorRT, TensorRT-LLM, Triton, Dynamo, NCCL, GPU Operator, Container Toolkit and DCGM; all read on 24 October 2026. "Listed" means the component appears in NVIDIA's description of the AI Enterprise platform, which then brings NVIDIA support and production branches for it; "not listed by name" means we found no such entry.

vLLM, the inference engine many such servers run, comes from outside NVIDIA; its repository is licensed under Apache-2.0. The model weights are a separate matter again, with licences that range from Apache-2.0 to texts that restrict commercial use or exclude regions; our guide to open LLM licences for company use covers them.

What the CUDA, cuDNN and TensorRT licences allow

The three SDK licences follow the same pattern. The CUDA EULA grants "a non-exclusive, non-transferable license, without the right to sublicense" to install and use the SDK. It names no fee and sets no production restriction for released versions; pre-release versions are "not intended for use in production or business-critical systems". A company may let "employees and contractors of your entity or of your subsidiary(ies)" use the SDK "from your secure network to perform work on your behalf", so one download covers the whole development team.

Four conditions matter on a server:

1. NVIDIA GPUs only. The CUDA supplement states that "the SDK is licensed for you to develop applications only for use in systems with NVIDIA GPUs". The TensorRT licence carries the same condition in its TensorRT supplement, and the cuDNN licence as well.
2. Distribution only inside your own application. The parts you may ship to customers are listed: for CUDA in Attachment A of the EULA (the CUDA runtime, cuBLAS, cuFFT, cuSPARSE, NVRTC and other libraries), for cuDNN and TensorRT the runtime .so and .dll files. Your application "must have material additional functionality, beyond the included portions of the SDK", and the SDK may not be distributed "as a stand-alone product".
3. Developer tools stay internal. "Unless a developer tool is identified in this Agreement as distributable, it is delivered for your internal use only."
4. Critical applications are excluded from NVIDIA's testing and liability. The CUDA EULA names avionics, medical and automotive uses; the TensorRT and cuDNN licences contain similar wording.

For a company that builds its own inference service and runs it on its own servers, none of these conditions costs anything. They matter when software is passed on: a product shipped to customers with CUDA libraries inside, or a container image published outside the company.

The driver has a licence of its own. NVIDIA's driver licence agreement of 25 February 2025 allows the software to be deployed "for your own use" on infrastructure you own or lease, rules out "commercial hosting services with the SOFTWARE" unless the agreement grants it, and excludes GeForce and Titan software from data-centre deployment, as our article on GeForce cards in a server quotes in full. The kernel modules themselves are open source: NVIDIA's open-gpu-kernel-modules repository licenses the files as MIT and the linked kernel module as MIT/GPLv2, while the user-space driver components and the GSP firmware come from the matching NVIDIA driver release.

Open-source NVIDIA projects

A large part of the inference stack is published by NVIDIA under open-source licences, with the source on GitHub.

Licence fields and LICENSE files of the GitHub repositories NVIDIA/TensorRT-LLM, triton-inference-server/server, ai-dynamo/dynamo, NVIDIA/nccl, NVIDIA/gpu-operator, NVIDIA/nvidia-container-toolkit and NVIDIA/DCGM, read on 24 October 2026. Triton's README gives 2.73.0 (container 26.09) as the current release.

Two of these projects state where enterprise support comes from. Triton's README says that Triton is part of NVIDIA AI Enterprise and that, through it, "NVIDIA global support is available for Triton Inference Server". The software is the same with or without a subscription; the subscription buys the support contract and the production branches with stable APIs and security patches. Dynamo's repository names a community forum and Slack for help and does not mention AI Enterprise.

The TensorRT repository on GitHub holds only the open parts: "the sources for TensorRT plugins and ONNX parser, as well as sample applications", described as "a subset of the TensorRT General Availability (GA) release". Building them needs the GA release (the repository's build instructions name 11.2.1.2; the TensorRT documentation was at 11.4.0 on 24 October 2026), which NVIDIA distributes from its developer site under the TensorRT licence above. The question of the TensorRT licence therefore has two answers: Apache-2.0 for what is on GitHub, NVIDIA's licence agreement for the core library.

NIM: no fee for development, AI Enterprise in production

NIM is where "free" and "licensed" sit closest together. NVIDIA's NIM General FAQ, last modified on 6 August 2026, gives members of the NVIDIA Developer Program, which costs nothing to join, two things: access to NIM API endpoints for prototyping, and the right to download NIM and self-host it "for research, application development, and experimentation" on up to 16 GPUs. It defines production as "any use of NIM for purposes other than development, testing, research or evaluation" and states: "Using NIM in production requires an NVIDIA AI Enterprise license." NVIDIA adds that pricing "is based on the number of GPUs, not the number of NIMs", so one licence per GPU covers every NIM on that GPU.

For the move to production, the FAQ points to a 90-day AI Enterprise trial that is "designed for production deployments and includes a license for commercial use". NVIDIA's get-started page states that this trial includes Omniverse and does not include Run:ai.

NVIDIA's NIM offerings page lists the NIM Certified Feature Branch "at no charge under the applicable license terms", which sits awkwardly beside the FAQ's production rule. Our guide to NVIDIA NIM requirements sets out how the two documents differ and why the quote should name the offering and licence a production server rests on. For planning, the safe reading is the FAQ's: a NIM that serves staff or customers needs AI Enterprise on every GPU it runs on.

The open alternative to NIM is to run the same models with vLLM, SGLang or TensorRT-LLM directly, which our comparison of vLLM, SGLang, TensorRT-LLM and Ollama covers. NIM packages an engine, optimised profiles and NVIDIA's validation; the engines underneath carry the open-source licences in the table above. We supply GPU servers for either path and put the licences in the same quote; tell us which models you want to serve and whether NIM is in the plan, and our team replies with the configuration and the licence lines it needs.

Compute vGPU: the licence the driver enforces

For virtual machines the picture changes. NVIDIA's vGPU product covers graphics profiles: Q-series under RTX vWS, B-series under vPC and A-series under vApps, editions that our guide to vGPU licensing with RTX vWS, vPC and vApps explains. Compute profiles, the C-series types such as DC-24C on the RTX PRO 6000 Server Edition, come only with NVIDIA AI Enterprise, licensed per physical GPU. NVIDIA's release notes for vSphere put it in one sentence: "C-series vGPU types are not available. Instead, NVIDIA vGPU for Compute is supported with NVIDIA AI Enterprise."

This is the part that stops working without a licence. A compute vGPU checks out a licence from the NVIDIA License System when the VM boots; without one it runs at full capability for 20 minutes and then drops to an idle level of compute, as our licensing guide describes. Graphics vGPU profiles are enforced in the same way under their own licences. Everything else on this page keeps running whatever the licence status, because on bare metal and in containers there is no licence server to ask. The licence server options, the cloud service and the delegated appliance for closed networks, are in our article on the NVIDIA vGPU licence server.

MIG on bare metal needs no licence: partitioning an RTX PRO 6000 into four 24 GB instances for four containers is a driver feature. MIG instances handed to separate VMs are MIG-backed vGPUs, and then the vGPU licence rules apply again. GPU passthrough of a whole card to one VM uses the normal data-centre driver in the guest; NVIDIA requires a licence only once AI Enterprise software, such as a production NIM, runs in that VM.

Three GPU servers and the licences they need

NVIDIA's licensing guide, last updated on 2 September 2026, counts AI Enterprise per GPU: "A software license is required for every GPU installed on the server or workstation", with one licence per GPU on boards that carry several. The examples use the reference configurations on our AI servers page.

NVIDIA AI Enterprise licensing guide (2 September 2026) and NIM General FAQ (6 August 2026), read 24 October 2026; the assignment of licences to the example stacks is our reading of those documents.

The entry server is the common case. Two cards, an open-source engine and one team: the licence agreements of the driver, CUDA and the engine are all it needs. A subscription would add NVIDIA's support for a driver fault in production and access to production branches, which some IT departments require for any software they run.

For the four-card server, the software decides. The same machine running vLLM needs nothing; running NIM for 300 staff, it needs four licences, and as a virtualisation host with compute vGPUs for twelve data scientists, also four, however many VMs share the cards. If such a host runs Proxmox VE, note that Proxmox documents NVIDIA AI Enterprise as not officially supported on its platform, as our guide to NVIDIA vGPU on Proxmox VE explains.

The eight-card H200 NVL server is the exception that changes comparisons. "Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription", per the licensing guide, which starts the term at "the ship date of the GPU board to the OEM partner, plus 90 days", a date that cannot be changed. The subscription has to be activated and may be installed only on the NVIDIA-Certified System in which the GPUs are installed. The H100 PCIe and H100 NVL include five years as well and the A800 40 GB Active three; the guide names no subscription for the RTX PRO 6000 or RTX PRO 4500 Server Editions, the L40S or the L4.

Support, production branches and certified servers

What AI Enterprise adds beyond NIM and compute vGPU is mostly about operating the stack. The licensing guide includes Business Standard support with subscriptions and offers Business Critical at extra cost. The subscription gives access to production and long-term branches with stable APIs and regular security patches, where the open releases move on a faster cadence; our article on NVIDIA driver branches and CUDA versions shows how those branches line up.

Support is tied to the hardware. NVIDIA's AI Enterprise 8.2 support matrix, updated on 2 September 2026, states that its GPUs "are supported with servers listed on the NVIDIA-Certified Systems", and it lists the RTX PRO 6000 and RTX PRO 4500 Server Editions, the L40S, the L4 and the H200 NVL. A subscription on a server outside that list has no support path, so check the server model against NVIDIA's list before the subscription is ordered.

The decision follows from the stack. With bare metal, open-source engines and your own team handling driver updates, the licence is optional. With NIM in production, compute vGPU in VMs, or an internal rule that production software must carry vendor support, it is required, per GPU, for every GPU in the server. If a project is unclear, describe the GPUs, the hypervisor and the serving stack, and we tell you which lines the quote needs and which it does not.

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. On request the server arrives with the operating system, drivers, CUDA and a container runtime installed. NVIDIA AI Enterprise and vGPU licences go on the same quote as the hardware, and for H200 NVL cards the quote states the included subscription. Models, retrieval and the platform on top are our Private AI/ML service.

FAQ