Status: research record, read 2026-08-07. This memo records what the NVIDIA licensing surfaces permit and forbid for a config-plane inference entry in the AICR catalog. It is an engineering read of public terms, not legal advice. Every claim cites its source and the date it was read; NVIDIA revises these documents, and per-artifact governing terms on NGC catalog pages override any general statement here.
Why this read happened before any build
The AICR catalog brief plans an inference entry sourced from NVIDIA's NIM deployment surfaces, and the catalog overview states the boundary in one line: catalog the deployment shapes, never redistribute the gated artifacts. This memo verifies that the line holds under the actual license text, and turns it into a concrete contract the entry must follow.
What the sources say
The deployment scaffolding carries Apache-2.0
- NVIDIA/nim-deploy is Apache-2.0. The repository describes itself as reference implementations, holding YAML files, Helm charts, operator code, and guides, and warns that it is experimental and carries no enterprise support. It points long-term production users to the official releases on NGC.
- NVIDIA/k8s-nim-operator is Apache-2.0. It provides the NIMCache, NIMService, and NIMPipeline custom resources, plus the NeMo customizer, evaluator, and guardrail resources.
Apache-2.0 permits retention, modification, and redistribution with attribution and license preservation. The whole config plane of a NIM deployment, meaning the charts, the CRDs, the values, and the manifests rendered from them, can be retained in this catalog the same way the H100 training entry retains its rendered Applications.
The runtime artifacts sit behind NVIDIA's enterprise license
- The NVIDIA Software License Agreement governs enterprise software delivered through NGC. Section 8.5 forbids copying, reselling, renting, sublicensing, transferring, or distributing the software. Section 5 makes the customer responsible for securely maintaining log-in information, which covers NGC API keys. Section 8.9 forbids sharing benchmarking, competitive analysis, or performance data without written consent.
- The Product-Specific Terms for AI Products add the NIM-specific layer. Section B.2 separates trial and not-for-resale licenses from production use. Section E permits community NIMs on RTX and GeForce systems without a subscription only outside commercial multi-user serving; commercial deployment requires an NVIDIA AI Enterprise subscription. Section 1.7 attaches notice and usage-reporting duties to any distribution of derivative materials from enterprise products. Section G.1.1 documents NIM telemetry and its
NIM_TELEMETRY_MODEcontrol. - The NVIDIA Deep Learning Container License governs many general NGC containers and likewise forbids redistribution.
- developer.nvidia.com/nim offers free hosted API endpoints for prototyping through the Developer Program and points production users to NVIDIA AI Enterprise. Third-party reports describe per-GPU AI Enterprise pricing and a free development tier for self-hosted NIMs; those figures did not come from an NVIDIA page during this read and stay out of the catalog's claims.
- Model weights carry their own per-model licenses that vary by model. An NGC catalog page can name governing terms per artifact, and those per-artifact terms are the authority for that artifact.
The verdict for a config-plane entry
The planned boundary holds, and it is sharper than expected in our favor. The license line and the config-plane line coincide: everything the catalog wants to retain is Apache-2.0 scaffolding, and everything the licenses gate is runtime material the catalog never touches anyway.
The entry may:
- Retain the nim-deploy charts and the NIM Operator CRDs from their Apache-2.0 GitHub sources, with attribution and the upstream license preserved, pinned by exact commit the way the training entry pins AICR v0.14.0.
- Author and retain values, variants, and rendered manifests derived from that Apache-2.0 scaffolding, and pin them under a digest-bound platform index.
- Record
nvcr.ioimage references and digests, model names, and per-model license names as configuration data. A reference is data about what the user's cluster would pull with the user's own entitlement; recording it redistributes nothing. - Surface
NIM_TELEMETRY_MODEas a first-class control point in variants, since the product terms document it as the telemetry control.
The entry must never:
- Mirror, re-publish, or cache NIM container images, model weights, or any NGC-delivered artifact in the catalog's OCI namespace. Section 8.5 forbids it, and the catalog has no need for it.
- Commit or embed an NGC API key anywhere. The key is the user's credential under Section 5 and enters the shape only as a target fact, the same way the training entry externalizes its secrets.
- Publish performance or benchmarking observations of NIM services. Section 8.9 forbids it without consent. The catalog's config-plane-only boundary already keeps every receipt on this side of that line, because no NIM workload runs in any proof.
- Retain chart artifacts served from the NGC registry itself. The GitHub Apache-2.0 tree is the provenance the catalog uses; anything delivered through NGC is treated as gated unless that artifact's own governing terms say otherwise.
- State which license tier a user needs. The entry links NVIDIA's pages and says that production self-hosted NIM requires an NVIDIA AI Enterprise subscription per NVIDIA's terms, and leaves tier selection between the user and NVIDIA.
What this means for the proof design
Config-plane proofs for the inference entry run exactly like the training entry's proofs. They render the Apache-2.0 charts on kind with placeholder pull secrets, exercise import, variant, promotion, and delivery wiring through ConfigHub, and never pull from nvcr.io. This respects the license boundary and the no-GPU boundary with one design: the gated pull happens only on the user's cluster, with the user's key, under the user's entitlement. Every receipt states that no NIM container ran and no model was fetched.
Open items before the entry is built
- Pick the exact nim-deploy chart and operator version to retain, and record their commits the way the training entry records its release pins.
- Choose the first model profile to describe as data, and record that model's own license name from its NGC page as part of the entry.
- Re-read the per-artifact governing terms on the specific NGC pages the entry references at build time, since per-artifact terms override this general read.
Addendum recorded at build time, 2026-08-07
The build closed the three open items the same day, with one discovery that changed the sourcing decision. The nim-llm Helm chart this read expected to retain no longer exists on GitHub: nim-deploy removed it on 2025-05-09 (pull request 143), and the chart now ships only through NGC, which this read classifies as gated. The inference entry therefore retains the repository's KServe subtree instead, at commit 3ef33472 under the same Apache-2.0 license. The per-artifact re-read happened for the first described profile: on 2026-08-07 the NGC catalog page for the Llama 3.1 8B NIM named four governing terms, the NVIDIA Software License Agreement, the Product-Specific Terms for AI Products, the NVIDIA Open Model Agreement, and the Llama 3.1 Community License Agreement, and the profile record carries those names as dated data.
The rules generalized, 2026-08-08
This read produced four rules that were never NIM-specific, and they are now doctrine rather than a decision recorded in one planning note: reference and never mirror, keep the credential with the cluster that pulls, publish no vendor performance numbers, and name the terms a page listed on a date without telling anyone which entitlement they need.
The third open item above asked for a per-artifact re-read on the specific pages an entry references. One was read by hand and carried as data. The other nine are now enumerated in the gated-artifact register, each with the catalog page a person can read, and a lane refuses a gated reference the register does not list and a listing no document names.
Their terms remain unread, and the reason is recorded rather than left as an absence: those catalog pages are served behind bot detection, and reading them automatically would mean working around it, which this catalog will not do. The gap is measured and published instead of filled with a guess.
The re-read cadence is settled too, and it is not a date. Entries are keyed by the exact image reference including its tag, so a version bump produces a reference nobody enumerated and the lane refuses it until someone records what the new artifact's page says. A calendar cadence goes stale quietly, because nothing fails when a date passes.