SUSE is an open-source infrastructure vendor that turned its Kubernetes into an AI platform by curating what runs on it, and the map reads it strongest where Rancher always was. Layer 2A is strong on multi-cluster Kubernetes with AI-conformant RKE2, virtual clusters that hand out GPU slices under quotas, and an Apache 2.0 AI Factory operator; eight cells are moderate (a hardware-agnostic abstraction with SUSE Linux Enterprise Server (SLES) 16 and Multi-Instance GPU (MIG) capable virtualization at 0A, networking at 0B, Longhorn and MinIO as platform storage at 0C; supported open-source vector stores plus NVIDIA's retrieval blueprint at 1B; NVIDIA's fixed retrieval-augmented generation (RAG) ingest pipeline at 1C; open inference engines plus NVIDIA AI Enterprise through the Factory at 2B; observability and Rancher AI's supervisor and human-validation gates at 2C; SUSE Industrial Edge at Layer 3), and 1A is a gap: no catalog and no data governance on SUSE's paper.
The capture is the lightest on the map, because SUSE sells support for software the enterprise could run without it. Twenty-eight components read Retained 14, Delegated 2, Ceded 12: SLES, Harvester, RKE2, K3s, K3k, Fleet, Longhorn, MinIO, Milvus, Qdrant, OpenSearch, Open WebUI, vLLM, Ollama, LiteLLM, NeuVector, and the AI Factory operator are Apache 2.0 or GPL projects the enterprise operates itself; what's SUSE's is Rancher Prime's extensions, the Application Collection's hardened images and lifecycle, the AI Factory's subscription catalog and blueprints, SUSE Observability, Rancher AI, and Industrial Edge, and what's NVIDIA's through SUSE's paper (NIM, NeMo, Nemotron, the RAG blueprint) is Ceded to NVIDIA except for model access through the standard interface. The decision-authority readings are Delegated at five layers because SUSE's engines document the pins, per-object policies, and confirmation prompts the override rule asks for.
The buyer's trade: a sovereign, air-gappable AI stack on their own hardware with the fewest captive opinions of any row, in exchange for assembling retrieval and pipelines from parts, taking the agent runtime and the retrieval models from NVIDIA, and waiting on the pieces still in preview (Universal Proxy, the Multi-Linux Manager MCP server, the SLES 16 MCP host, OpenShell and NemoClaw, Longhorn v2, Run:ai in the Factory's docs). What would move a cell: a SUSE retrieval engine (1B), Universal Proxy reaching GA with tool-call policy (2C), or the AI Factory's docs naming Run:ai and the NVIDIA agent runtime as supported deployables (2A, 2B).
Layer-by-layer status: Layer 0A (Hardware-Agnostic Substrate: SLES 16, SUSE Virtualization With MIG, Any Vendor's Metal), Layer 0B (RKE2 Container Networking and SUSE Virtualization VM Networks Over Any Fabric), Layer 0C (Longhorn Block Storage and MinIO Objects; General-Purpose, Not Built for AI Throughput), Layer 1A (No Data Catalog or Data Governance on SUSE's Paper), Layer 1B (Supported Open-Source Vector Stores and RAG in the AI Library; NVIDIA's RAG Blueprint Through the Factory), Layer 1C (NVIDIA's Fixed RAG Ingest Pipeline Through the Factory; No General Pipeline Layer), Layer 2A (Rancher Heritage Strength: Multi-Cluster Kubernetes, AI-Conformant RKE2, Virtual Clusters with GPU Quotas, an Apache 2.0 AI Factory Operator), Layer 2B (Supported Open Inference (vLLM, Ollama) + NVIDIA AI Enterprise (NIM, NeMo, Nemotron) Through the Factory; No SUSE Runtime), Layer 2C (Observability for AI Workloads + Rancher AI's Supervisor, Agent Configs, and Human-Validation Gates; MCP Gateway in Preview), Layer 3 (+1) (Liz and the Crew for Infrastructure Operations, SUSE Industrial Edge for the Plant Floor; Platform-Enabled Beyond).
Assessment framework: 4+1 Layer AI Infrastructure Model. Scoring model: Decision Authority Placement Model (DAPM) — Retained, Delegated, or Ceded. Published by The CTO Advisor LLC (DBA The Advisor Bench). Author: Keith Townsend. Date assessed: October 19, 2026. Version: v1.8 - 4+1 v3: Layer 0 Sub-Layers.
SUSE is an open-source infrastructure vendor that turned its Kubernetes into an AI platform by curating what runs on it, and the map reads it strongest where Rancher always was. Layer 2A is strong on multi-cluster Kubernetes with AI-conformant RKE2, virtual clusters that hand out GPU slices under quotas, and an Apache 2.0 AI Factory operator; eight cells are moderate (a hardware-agnostic abstraction with SUSE Linux Enterprise Server (SLES) 16 and Multi-Instance GPU (MIG) capable virtualization at 0A, networking at 0B, Longhorn and MinIO as platform storage at 0C; supported open-source vector stores plus NVIDIA's retrieval blueprint at 1B; NVIDIA's fixed retrieval-augmented generation (RAG) ingest pipeline at 1C; open inference engines plus NVIDIA AI Enterprise through the Factory at 2B; observability and Rancher AI's supervisor and human-validation gates at 2C; SUSE Industrial Edge at Layer 3), and 1A is a gap: no catalog and no data governance on SUSE's paper.
The capture is the lightest on the map, because SUSE sells support for software the enterprise could run without it. Twenty-eight components read Retained 14, Delegated 2, Ceded 12: SLES, Harvester, RKE2, K3s, K3k, Fleet, Longhorn, MinIO, Milvus, Qdrant, OpenSearch, Open WebUI, vLLM, Ollama, LiteLLM, NeuVector, and the AI Factory operator are Apache 2.0 or GPL projects the enterprise operates itself; what's SUSE's is Rancher Prime's extensions, the Application Collection's hardened images and lifecycle, the AI Factory's subscription catalog and blueprints, SUSE Observability, Rancher AI, and Industrial Edge, and what's NVIDIA's through SUSE's paper (NIM, NeMo, Nemotron, the RAG blueprint) is Ceded to NVIDIA except for model access through the standard interface. The decision-authority readings are Delegated at five layers because SUSE's engines document the pins, per-object policies, and confirmation prompts the override rule asks for.
The buyer's trade: a sovereign, air-gappable AI stack on their own hardware with the fewest captive opinions of any row, in exchange for assembling retrieval and pipelines from parts, taking the agent runtime and the retrieval models from NVIDIA, and waiting on the pieces still in preview (Universal Proxy, the Multi-Linux Manager MCP server, the SLES 16 MCP host, OpenShell and NemoClaw, Longhorn v2, Run:ai in the Factory's docs). What would move a cell: a SUSE retrieval engine (1B), Universal Proxy reaching GA with tool-call policy (2C), or the AI Factory's docs naming Run:ai and the NVIDIA agent runtime as supported deployables (2A, 2B).
CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant
The enterprise operating system under every SUSE path, open source the enterprise runs itself with SUSE's subscription for updates and support. Retained on the open-source seam.
The hyperconverged virtualization layer running VMs and containers on one Kubernetes, with GPU multi-tenancy and per-VM placement. Open source the enterprise operates: Retained on the open-source seam.
SUSE runs on any certified server the enterprise buys. The enterprise's choice: Retained.
NVIDIA's partitioning, drivers, and GPU Operator packaged and supported by SUSE. Re-read October 5, 2026 against the Red Hat standard: the GPU Operator is Apache 2.0 open source the enterprise runs on any Kubernetes, so the operator layer reads Retained on the possess-and-operate line. vGPU licensing is NVIDIA's proprietary software, and leaving NVIDIA means rebuilding; that capture sits in the NVIDIA column.
MIG vGPU support is generally available in SUSE Virtualization 1.7 (January 2026, announced March 24, 2026). RKE2 meets the Cloud Native Computing Foundation (CNCF) AI conformance with the NVIDIA GPU Operator installed per NVIDIA's docs, and SUSE AI Factory with NVIDIA embeds NVIDIA AI Enterprise, NIM, NeMo and Nemotron. The GPU Operator is Apache 2.0 and reads Retained; vGPU licensing is NVIDIA's proprietary software, and that capture sits in the NVIDIA column.
SUSE sells the software that makes any vendor's hardware an AI substrate, and nothing below it. SUSE Linux Enterprise Server 16 (SLES 16, generally available November 4, 2025) is the operating system, with the Agama installer, SELinux by default and a sixteen-year lifecycle. SUSE Virtualization, the KubeVirt-based Harvester under Apache 2.0, runs VMs and containers on the same Kubernetes. It adds NVIDIA Multi-Instance GPU (MIG) vGPU multi-tenancy and per-VM node scheduling. Both run on whatever x86 or Arm servers the enterprise buys. The buyer gets an abstraction over its own metal, with the GPU slicing handled. The architect's concern is that SUSE sells none of the hardware, so compute density and accelerator topology come from the OEM and NVIDIA. Platform software at 0A reads moderate by rule (C3). Calibration: Red Hat, VMware and Nutanix read moderate on the same shape; Dell and HPE strong on hardware they sell.
Delegated, on the open-source seam. The scheduler places VMs and GPU slices within the enterprise's policies, and a per-VM nodeSelector or required affinity rule is a pin the engine must honor: vendor decides, visible, overridable, Delegated. Portability: SLES, SUSE Virtualization and the GPU Operator are open source the enterprise runs, Retained.
Watch-list, notes only: VM auto-balancing and live storage migration (early access); the SLES 16 Model Context Protocol (MCP) host (technology preview); the Openchip RISC-V stack (partnership, no product); SUSE as launch partner for the AWS European Sovereign Cloud. Named, not scored: Fsas Technologies (Fujitsu) as the AI Factory's launch hardware partner and Vultr as its co-engineered cloud. Public evidence that moves the cell: SUSE-sold hardware.
Interconnects, switching and routing, network virtualization
RKE2 bundles four CNI plugins, with Canal (Flannel between nodes, Calico for network policy, VXLAN overlay) as the default and Cilium, Calico and Flannel selectable, plus Multus for secondary networks. SUSE packages and supports them in RKE2. Open source the enterprise runs itself: Retained on the open-source seam.
Harvester's VM networking: a Canal-based management network, custom cluster networks that group host NICs for isolation, VLAN and untagged bridge VM networks, and a built-in Layer 4 load balancer for VMs and guest Kubernetes clusters. Open source the enterprise operates: Retained on the open-source seam.
Assessment pending
RKE2 bundles four Container Network Interface (CNI) plugins. Canal is the default: Flannel carries the VXLAN overlay between nodes and Calico enforces network policy. Cilium, Calico and Flannel are selectable, with Multus for secondary networks. SUSE Virtualization adds VM networking. Cluster networks group host network interface cards (NICs) for isolation, beside VLAN and untagged VM networks and a built-in Layer 4 load balancer. The buyer gets software-defined networking that runs on whatever switches it owns. The architect's concern is that it isn't an AI fabric. GPU-to-GPU bandwidth and congestion control come from the switch vendor. Calibration: Red Hat, VMware and Nutanix read moderate on network virtualization; NVIDIA and HPE strong on fabrics they sell.
Retained, on the open-source seam. RKE2's documentation shows no gateway or egress placement controller in Canal, so the engine only executes the networks and policy the enterprise declared: code decides, Retained, where Red Hat reads Ceded because OVN-Kubernetes places egress IPs. Portability: the bundled CNIs and SUSE Virtualization networking are open source the enterprise runs, Retained.
Arrays, storage servers and software, front-end protocols: where the data sits
Rancher-originated distributed block storage for Kubernetes, bundled with Rancher Prime. Open source the enterprise operates: Retained on the open-source seam.
SUSE-built, SUSE-supported MinIO charts installed from the Application Collection or the AI Factory catalog. Self-run open source behind the S3 standard: Retained.
Assessment pending
SUSE's storage is the storage under its Kubernetes. SUSE Storage is Longhorn, the Apache 2.0 Cloud Native Computing Foundation (CNCF) project Rancher Labs created, bundled with Rancher Prime. It's distributed block storage with replication across nodes, snapshots, backups to S3 or NFS, and disaster recovery volumes. Objects come from MinIO, which the SUSE Application Collection builds, tests and distributes under the Rancher Prime and SUSE AI subscriptions. The AI Factory installs from that catalog. The buyer gets storage it trusts under the platform, on the same servers. The architect's concern is depth. Longhorn and MinIO are general-purpose storage for the platform's state, not a tier built to feed accelerators. SUSE has had no native object tier since it retired SUSE Enterprise Storage, its Ceph product. Calibration: Red Hat, VMware and Nutanix read moderate on platform storage; WEKA and VAST strong on storage built for the job.
Delegated where the grade rests. Longhorn places replicas within per-volume node and disk selectors, data locality and replica auto-balance settings, policies the engine must honor, and eviction is a documented, cancelable action: vendor decides, visible, overridable, Delegated. Portability: Longhorn and MinIO are open source the enterprise operates, with MinIO consumed through the S3 standard, Retained.
Watch-list, notes only: the Longhorn v2 data engine (technology preview).
The data itself, organized and governed: the Governance Catalog that Layer 2C queries
Nothing at this layer runs on accelerators.
SUSE's row documents no 1A function. Longhorn and MinIO hold the bytes, a 0C function, and the row cites no access policy, catalog or lineage for either. The row already said where SUSE's data story stops: no catalog, no lineage, no classification and no policy engine over data as an estate. The buyer brings the data foundation. Calibration: Red Hat reads moderate on documented bucket policies; Nutanix moderate on Data Lens compliance governance.
Absent: SUSE ships no data catalog or data governance, and nothing at 1A is inherited from it.
Evidence gap: MinIO's bucket and IAM policies, if sourced on SUSE's paper, would add a 1A access function and lift the cell to moderate. Public evidence that moves the cell: that source, or a data catalog with classification and lineage.
Low-latency retrieval for RAG — vector/hybrid search, context windows
Open-source retrieval components installed from the AI Library and run by the enterprise, OpenSearch's hybrid query included. Retained on the open-source seam.
SUSE's packaging and support opinions around the open components, the subscription's value. SUSE's: Ceded.
NVIDIA's generally available RAG stack under NVIDIA AI Enterprise on the customer's GPUs through SUSE's paper. Proprietary third-party software: Ceded to NVIDIA.
The AI-Q with RAG blueprint deploys NVIDIA's RAG server, ingestor, and the nemotron-vlm-embedding microservice with the Nemotron Super 49B model under NVIDIA AI Enterprise on the customer's GPUs; the SUSE AI Library path runs on any GPU the enterprise has.
SUSE's retrieval layer is curated open source with SUSE support, plus NVIDIA's blueprint. The SUSE AI Library, delivered through the SUSE Application Collection as signed, hardened images with software bills of materials, installs Milvus (the default vector database for the stack), Qdrant, and OpenSearch as vector stores, Open WebUI with its pipelines as the retrieval-augmented generation (RAG) front end, and PyTorch (LiteLLM, the model router, is scored at 2B), all deployable air-gapped, with documented OpenTelemetry integration patterns for Milvus and Open WebUI once the applications are instrumented. SUSE AI Factory with NVIDIA (generally available July 9, 2026, co-engineered with Vultr) ships the NVIDIA AI-Q with RAG blueprint as a pre-built, generally available stack: NVIDIA's RAG server, ingestion service, and NeMo Retriever embedding microservice with the Nemotron Super 49B model, plus a minimal low-GPU variant the docs say isn't for production. SUSE ships no retrieval engine of its own. The buyer gets a supported, secured open-source retrieval stack they'd otherwise assemble, on their own hardware, and NVIDIA's RAG through one contract. Applying the test: the capability is real and generally available on SUSE's paper and general in kind; it's the enterprise's to compose (OpenSearch's own hybrid query included) and there's no SUSE-integrated hybrid or permission-aware retrieval service. Calibration: Kamiwaza reads strong on a context manager and retrieval service of its own; Elastic strong on a native engine; Cloudera gap with its studio in preview; Cloudflare moderate on an index plus open models. SUSE is a curated kit plus a partner blueprint, shipped and supported. Moderate.
Low on the open path, NVIDIA's on the other. Milvus, Qdrant, OpenSearch, and Open WebUI are open projects the enterprise runs itself: Retained; the vectors are captive to whichever embedding model was chosen, under the carve-out. SUSE's contribution, the Application Collection's curated images, SBOMs, lifecycle, and support, is SUSE's and Ceded, and it's what a customer pays for. The NVIDIA RAG blueprint and NeMo Retriever microservices are NVIDIA AI Enterprise software with no source license, Ceded to NVIDIA through SUSE's paper on the channel-substitution rule. The enterprise composes the pipeline: vendor decides at the store, visible, overridable, Delegated, the Elastic reading for a surface the enterprise configures end to end.
Inference flagged: the current SUSE AI version (the documentation set is 1.0; the November 2025 release added components without a new major version). Public evidence that moves the cell: a SUSE retrieval engine, or hybrid and permission-aware retrieval on SUSE's paper.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
NVIDIA's fixed-function ingest pipeline into the RAG vector store, generally available in the Factory's pre-built blueprint. Proprietary third-party software through SUSE's paper: Ceded to NVIDIA.
The AI-Q with RAG blueprint's ingestor-server is NVIDIA's NV-Ingest pipeline (parsing, OCR and table extraction, chunking, embedding, bulk and streaming ingestion into the vector database) under NVIDIA AI Enterprise; the minimal variant disables OCR and table and chart extraction.
SUSE moves data into one place: the vector database under NVIDIA's RAG. SUSE AI Factory with NVIDIA deploys the AI-Q with RAG blueprint as a generally available stack, and its ingestor is NVIDIA's NV-Ingest: document parsing with optical character recognition and table and chart extraction, chunking, embedding, and bulk or streaming ingestion into the vector store, a fixed-function pipeline with one destination. Beyond it, SUSE moves software rather than data: Fleet delivers manifests and Helm charts to thousands of clusters, the Factory promotes AI workloads from development to production, and Kubeflow Pipelines (in the AI Library as a technology preview, with its own ML Metadata lineage) orchestrate training steps. There's no extract-transform-load engine, no change data capture, no generally available lineage on SUSE's paper, no cost-aware movement, and no KV-cache tiering. The buyer gets NVIDIA's ingest for NVIDIA's RAG, and brings their own pipelines. Applying the general-versus-fixed-function test: a fixed-function AI-ingest pipeline that terminates in the vendor's retrieval surface, on SUSE's paper through the NVIDIA bundle. That's the NetApp AI Data Engine shape (moderate), not the absence of a layer. Calibration: NetApp reads moderate on a fixed sync-classify-vectorize pipeline; Nutanix and VMware gap with no pipeline layer; Cloudera and Elastic score real movement products. SUSE is NetApp's shape with NVIDIA's pipeline. Moderate.
NVIDIA's. The ingest pipeline's parsing, chunking, and embedding opinions are NVIDIA's inside a blueprint SUSE ships: Ceded to NVIDIA through the channel. The pipeline runs as configured with no per-document override documented: vendor decides, visible, not overridable, Ceded, the NetApp reading.
Grade moved during review: drafted gap, moved to moderate when the AI Factory docs showed the NVIDIA RAG blueprint's ingestor as a generally available deployable on SUSE's paper. Watch-list, notes only: Kubeflow Pipelines (technology preview in the AI Library; ML Metadata lineage inside). Public evidence that moves the cell: a data pipeline or movement product on SUSE's paper.
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
The open-source Rancher the enterprise operates, whose cluster definitions and GitOps configuration lift to community Rancher. Retained on the open-source seam.
The proprietary extensions and entitlements Prime adds over the open core, whose configuration doesn't lift to community Rancher. SUSE's: Ceded.
The supported Kubernetes distributions the enterprise runs, conformant to the standard interface. Retained on the open-source seam.
Isolated K3s control planes on shared GPU infrastructure with tenant quotas. Open source the enterprise runs; the policy packaging is Prime's. Retained on the open-source seam, flagged.
Continuous delivery of manifests and charts to fleets of clusters. Open source the enterprise runs: Retained.
The open-source control plane that installs, manages, and monitors AI applications and blueprints on Rancher-managed clusters. Self-run open source: Retained.
The subscription-gated application catalog and SUSE's immutable, version-controlled blueprint content. SUSE's: Ceded.
RKE2 meets CNCF AI conformance with the NVIDIA GPU Operator installed per NVIDIA's docs and a gang scheduler (Volcano or Kueue) installed alongside; the AI Factory operator discovers and installs NVIDIA applications and blueprints from the NGC catalog. NVIDIA Run:ai appears in the April announcement and not in the product docs.
This is the layer SUSE bought Rancher for, and it's where the AI additions landed. Rancher Prime manages fleets of Kubernetes clusters (fifty thousand teams run Rancher) with RKE2 and K3s as the supported distributions; RKE2 is generally available as a Cloud Native Computing Foundation (CNCF) Certified Kubernetes AI Conformant distribution, meeting the conformance with the NVIDIA GPU Operator installed per NVIDIA's documentation and a gang scheduler such as Volcano or Kueue installed for distributed training. Virtual Clusters through K3k (Apache 2.0) are generally available: hundreds of isolated K3s control planes on one physical cluster, with Virtual Cluster Policies that hand GPU slices to tenants under quotas, updated live, so teams self-serve GPUs without owning nodes. Fleet does GitOps at scale; SUSE Virtualization schedules VMs beside containers with MIG multi-tenancy. SUSE AI Factory with NVIDIA (generally available July 9, 2026) is a Rancher UI extension plus Kubernetes operator (Apache 2.0) that installs, manages, and monitors AI applications from a subscription catalog and the NGC catalog through immutable, version-controlled blueprints, from ClickOps to GitOps. The buyer gets fleet-scale Kubernetes with GPU multi-tenancy from a vendor whose scheduler is the upstream one. The architect's concern is which opinions lift, and the answer is nearly all of them. RKE2, K3s, K3k, Fleet, Rancher, and the Factory operator are Apache 2.0; what's SUSE's is Prime (support, extensions, the AI assistant, GPU multi-tenancy packaging) and the Factory's subscription catalog and pre-validated blueprints. The scheduling judgment is Kubernetes', bounded by the quotas the enterprise writes. Calibration: Nutanix and VMware read strong on orchestration heritage (Prism, VCF Automation); CoreWeave strong on open-substrate GPU orchestration (Kubernetes plus Slurm); Cloudera and Elastic moderate on platform-scoped orchestration. SUSE is CoreWeave's shape sold as software. Strong.
Bounded, and mostly on the open-source seam. RKE2, K3s, K3k, Fleet, and the AI Factory operator and UI extension are Apache 2.0 the enterprise runs itself: Retained; the Rancher core the enterprise runs is Retained, and Prime's proprietary extensions (the AI assistant, GPU multi-tenancy packaging, entitlements) are Ceded. The Factory's subscription catalog and pre-validated blueprints are SUSE's: Ceded. The scheduler places pods and GPU slices within the quotas and policies the enterprise set; those bounds alone would read Ceded, the Nutanix and VMware reading, but the per-object pins below carry the cell. Decision authority under the override rule (September 4, 2026): the scheduler places workloads, and pod nodeSelector and required affinity are per-object pins it must honor, the same mechanism this row's Layer 0 cites. K3k GPU quotas are bounds and stay configuration. Vendor decides, visible, overridable, Delegated, the Red Hat reading.
Watch-list, notes only: NVIDIA Run:ai for GPU orchestration (named in the April 21, 2026 announcement of SUSE AI Factory with NVIDIA; not named in the product's overview, building blocks, blueprints, or release notes; it would read Ceded to NVIDIA as a proprietary scheduler if scored); SUSE Observability Hosted (KubeCon EU 2026). Inference flagged: whether Virtual Cluster Policies' GPU quotas are Prime-only packaging or in open-source K3k (the product page lists GPU multi-tenancy as a Prime feature; the SUSE Virtual Clusters docs cover the VirtualClusterPolicy resource). Public evidence that moves the cell: nothing upward from strong.
Model serving, agent execution, inference APIs, distributed inference
Supported open-source engines and tools the enterprise runs on its own GPUs. Retained on the open-source seam.
Standard-shaped inference interfaces over the open engines and NVIDIA's NIM microservices. Delegated on the inference-interface ruling.
NVIDIA's inference and agent-building stack on the customer's GPUs through SUSE's paper (GA July 9, 2026), beyond the standard interface. Proprietary third-party software: Ceded to NVIDIA.
Open-weights models the enterprise can serve itself, bundled in the Factory. Delegated on the open-weights ruling.
SUSE's packaging and support opinions around the open components. SUSE's: Ceded.
Tools the enterprise writes and runs in its own processes, called by models served on the stack. The enterprise's: Retained.
The Factory's GA release embeds NVIDIA AI Enterprise with NIM microservices, the NeMo framework, and Nemotron open models; the AI-Q blueprint's deep-research agent runs on them. OpenShell (alpha on NVIDIA's own README) and NemoClaw appear in the April announcement and not in the product docs. The SUSE AI Library path (vLLM, Ollama) runs on any GPU.
SUSE serves models with open engines and takes the rest from NVIDIA. Serving: the SUSE AI Library installs vLLM (added as the inference engine in November 2025) and Ollama as supported, hardened images, with Open WebUI as the chat front end, LiteLLM as the OpenAI-compatible router, MLflow for the model lifecycle, Kubeflow (technology preview) for ML pipelines, PyTorch, and mcpo (an MCP-to-OpenAPI bridge) for tool integration, air-gapped if required, on the customer's GPUs; SUSE AI became a CNCF-conformant AI platform in November 2025. SUSE AI Factory with NVIDIA (generally available July 9, 2026) embeds NVIDIA AI Enterprise: NIM microservices as containerized inference engines running entirely on premises, the NeMo framework, and the Nemotron open models, with the AI-Q with RAG blueprint's deep-research agent as the shipped agentic example. Universal Proxy, SUSE's Apache 2.0 MCP proxy (first release January 30, 2026), fronts MCP servers with discovery, a registry, session management, and authentication, and isn't in the Factory's documentation; NVIDIA's OpenShell agent runtime and NemoClaw were announced with the Factory in April and aren't in its docs either. Liz, Rancher Prime's AI agent, is scored at Layer 3. The buyer gets a supported open inference stack they'd otherwise assemble, and NVIDIA's inference stack through one contract. The architect's concern: SUSE ships no runtime of its own. The serving engines are open source (the enterprise's), the NVIDIA stack is NVIDIA's (through SUSE's paper), and SUSE's value is curation, security, observability, and support. That's a real dependence and a thin one, and the agent runtime is still on the announcement side of the line. Calibration: Nutanix reads moderate on platform-native serving with agents pre-GA; VMware moderate on a model runtime plus an agent builder; Dell's 2B is two components (blueprints and accelerator services) with NVIDIA's runtime at the center, and it watch-lists NemoClaw and OpenShell; Kamiwaza moderate on its own runtime; Cloudera strong on any-model serving plus a GA agent builder of its own. SUSE is Dell's shape sold as software. Moderate.
Low for the engines, NVIDIA's for the models it packages. vLLM, Ollama, Open WebUI, LiteLLM, MLflow, and mcpo are open source the enterprise runs: Retained; model access through vLLM's and LiteLLM's OpenAI-compatible interfaces, and through NIM's, is Delegated on the inference-interface ruling, and the models the enterprise brings are its own. NVIDIA AI Enterprise's blueprints, RAG services, and NeMo containers beyond the interface are NVIDIA's: Ceded to NVIDIA through SUSE's paper; the Nemotron open models are Delegated on the open-weights ruling. The Application Collection's curated images are SUSE's: Ceded. Customer-created tools running in the enterprise's own processes are Retained. The model decides which tool to call; the enterprise's client-side tools gate the effect: model decides, visible, overridable, Delegated, the Nutanix and VMware reading.
Watch-list, notes only: OpenShell and NemoClaw (announced April 21, 2026; OpenShell is alpha per NVIDIA's README; not in the Factory's product docs; the Dell row watch-lists the same runtime), Universal Proxy (Apache 2.0; first release January 30, 2026; no GA badge; not in the Factory docs), the SLES 16 MCP host (technology preview), Kubeflow (technology preview), SUSE Observability Hosted. Public evidence that moves the cell: a SUSE-owned agent runtime, or the Factory's docs naming OpenShell and NemoClaw as supported deployables.
Policy-driven placement and resource coordination — the Autonomy Layer
The OpenTelemetry-native observability platform with AI-workload bindings, in Rancher Prime and as a hosted edition. SUSE's platform: Ceded.
The generally available multi-agent supervisor, agent configuration, confirmation gates, and controlled API gateway inside Rancher Prime. SUSE's: Ceded.
NeMo Guardrails is an Apache 2.0 toolkit not named in the AI Factory's product docs; SUSE Observability, Rancher AI, and SUSE Security run on CPU.
SUSE's reasoning plane has two shipped legs and one in preview. Observability: SUSE Observability (OpenTelemetry-native, integrated with Rancher Prime and offered hosted, with agentic root-cause analysis) ships the AI Observability Dashboards (generally available since May 2025) and, since November 2025, the OpenTelemetry operator and metric bindings and health monitors for vLLM, Ollama, Milvus, OpenSearch, GPU infrastructure, KServe, and Kubeflow, with a GenAI StackPack that turns LLM telemetry (prompts, tokens, response quality) into topology, dashboards, and health monitors, once applications are instrumented. Orchestration and policy inside Rancher AI (generally available March 2026): Liz is a supervisor that routes each request across specialized agents (Rancher, Continuous Delivery, Provisioning, Application Collection, Observability, Security, CloudCasa, and custom MCP agents the administrator adds) selected by AIAgentConfig metadata, each agent reasoning with an LLM the administrator configures (Ollama, OpenAI, Gemini, Bedrock, or any OpenAI-compatible endpoint), acting through a Rancher MCP server that the docs call the secure, controlled gateway to the Rancher and Kubernetes APIs with a read-only mode, and gated by Human Validation Tools, the administrator's list of tools that require explicit user confirmation before Liz executes them. SUSE Security (NeuVector) is container security, a state guardrail scored below this layer; its Vulnerability Scanner MCP server is experimental. The gateway leg for the estate is Universal Proxy, one point of access for agent interactions with MCP servers, with discovery, a curated registry, security hardening, and telemetry, Apache 2.0 since its first release on January 30, 2026 and absent from the Factory's docs, so it's watch-listed. The buyer gets to see what their AI workloads are doing across the estate, and a governed multi-agent operator whose actions a human confirms. Applying the test: Intelligence-2C has an observability leg (GA, cross-workload), an orchestration leg with agent configuration and a confirmation gate (GA, scoped to Rancher AI's own agents); no agent registry across systems, no agent identity distinct from the user, and the estate-wide tool gateway in preview. Infrastructure-2C, placement across model, cost, and compliance tiers, is absent. Calibration: Nutanix reads moderate on a single gateway leg; VMware moderate on AgentMinder's identity, gateway, and observability; Elastic moderate on registry, orchestration, and observability; Cohere gap with North's in-product policies alone; NetApp gap on data guardrails. SUSE has Elastic's observability and orchestration legs (the second inside one product), with the estate-wide gateway pending. Moderate.
Split. SUSE Observability's monitors and the StackPack are SUSE's (Ceded); Rancher AI's supervisor, agent configurations, and validation lists are SUSE's (Ceded). The layer's decision, which agent acts and with which tool, is Liz's LLM routing, visible in the UI, gated by the administrator's Human Validation Tools and the MCP server's read-only mode before the effect: model decides, visible, overridable, Delegated.
Watch-list, notes only, and these are what move the cell: Universal Proxy (Apache 2.0; first release January 30, 2026; not in the Factory docs), the SUSE Security Vulnerability Scanner MCP server (experimental; NeuVector itself is container security below this layer), the Multi-Linux Manager MCP server (technology preview), the SLES 16 MCP host (technology preview), NeMo Guardrails (an Apache 2.0 toolkit the enterprise could run, Retained if it does; not named in the Factory's docs). Public evidence that moves the cell: Universal Proxy GA with tool-call policy, or an agent registry or identity on SUSE's paper.
AI-powered business capabilities — business logic, workflow automation
The first-party industrial IoT application platform for the Tiny Edge. SUSE's engine, with open sourcing announced as intent: Ceded.
Liz runs against the models the enterprise configures; Industrial Edge runs on the Tiny Edge.
SUSE's own applications are an agent for the people who run SUSE and a platform for the people who run plants. Liz, the AI assistant in Rancher Prime, went from tech preview (November 2025) to generally available (March 2026) as a context-aware agent that coordinates a Crew of specialists for Linux, observability, security, provisioning, and fleet management, detects issues, explains them, and, through external MCP servers the enterprise plugs in (Atlassian, a firewall, documentation), creates tickets and takes actions across the operational stack, presenting a confirmation prompt the user must accept before any change is applied; Rancher Prime exposes an MCP server of its own so external agents can drive it, with n8n and Revenium as the documented examples and Amazon Web Services, Fsas Technologies, and Stacklok as named relationships, while the Multi-Linux Manager MCP server and the SLES 16 MCP host are technology previews. SUSE Industrial Edge (available April 22, 2026, on the Losant platform acquired February 19, 2026) is a first-party industrial IoT application: protocol-agnostic device connectivity (Siemens, Beckhoff, OPC UA), a no-code visual workflow engine, multi-level dashboards, pre-built application templates, anomaly alerts, and AI-driven quality checks, with SUSE AI endpoints on the roadmap and the technology slated for open sourcing. Everything else at this layer is what customers build on SUSE AI or buy from certified partners. The buyer gets an agentic SRE for their infrastructure, a workflow platform for their factories, and a platform for the rest. Applying the test: the layer asks for AI-powered business capabilities. Industrial Edge is one, a first-party, generally available business-process platform with AI features inside it, and the grade rests on it. Liz and the Rancher Prime MCP server are an operations assistant and a tool surface over SUSE's own platform, which Layer 3 excludes by name, since it credits first-party business applications, not operations assistants or tool surfaces: real and generally available, named here, but they earn no Layer 3 grade. Neither is the breadth of a security or analytics estate. Calibration: Nutanix and VMware read gap, platform enablement earning no Layer 3 grade; Kamiwaza moderate on a platform plus its Kaizen agent; CoreWeave moderate on Monolith's engineering-test modules in one vertical, its Forge developer plane read at 2B; Elastic strong on first-party security and observability applications; Qlik strong on an analytics estate. SUSE is moderate on one first-party vertical application. Moderate, with the boundary named: Industrial Edge maturing into an AI-driven process platform would argue for strong.
Low. Industrial Edge's workflows and dashboards are the enterprise's designs on SUSE's engine (Ceded until it's open-sourced); the MCP servers the enterprise plugs in are its own. Liz acts on the environment only after a confirmation prompt the user accepts, a persisted gate before the effect, but Layer 3 credits first-party business applications only, which takes Liz out of the grade, so the gate no longer carries the reading. Decision authority, re-read after Liz was removed from the grade (Layer 3 credits first-party business applications only): Industrial Edge carries it alone, and its detection is the enterprise's own design (threshold events the enterprise writes, a TensorFlow Predict node running a model the enterprise loads), so code decides, visible, overridable, Retained. It would read vendor only if SUSE documents a built-in anomaly or quality feature that infers from the data.
Ruled October 5, 2026: moderate holds. Industrial Edge is a domain-scoped process platform, a fixed-function slice, and with Liz excluded (Layer 3 credits first-party business applications only) it carries the grade alone. Watch-list, notes only: the Multi-Linux Manager MCP server (technology preview), the SLES 16 MCP host (technology preview), SUSE AI endpoints on Industrial Edge (roadmap), the open-sourcing of Losant's technology (announced intent), SUSE AI certified partner applications. Public evidence that moves the cell: Industrial Edge's AI features documented as generally available product capabilities. Named, not scored at Layer 3, which credits first-party business applications only: Liz and the Crew (an operations assistant over Rancher Prime, scored at 2C) and the Rancher Prime MCP server (a tool surface); both chips left this cell October 19, 2026.
SUSE documentation (SUSE AI 1.0 deployment and AI Library; SUSE AI Factory 1.0 monitoring; SUSE AI Factory with NVIDIA overview, building blocks, pre-built blueprints, lifecycle appendix, and release notes; Rancher AI architecture, user, and administrator guides; RKE2 GPU operators and AI conformance; SUSE Virtualization VM creation; SUSE Storage StorageClass parameters and settings; Longhorn eviction; SUSE Virtual Clusters; Vulnerability Scanner MCP server; SLES 16.0 release notes; Application Collection MinIO reference and subscriptions) and product pages (Rancher Prime, SUSE AI, SUSE AI Factory, SUSE Observability, SUSE Virtualization); SUSE press releases and blogs (SUSE AI Factory with NVIDIA generally available, July 9, 2026, and announced April 21, 2026; SUSE Industrial Edge, April 22, 2026; Losant acquisition, February 19, 2026; intelligent infrastructure management update, March 24, 2026; KubeCon EU 2026 posts on the agentic ecosystem, MCP plug-and-play, AI crews, virtual clusters and GPU sharing, MIG in SUSE Virtualization 1.7, SUSE Observability Hosted; secure agentic AI for infrastructure management with partners, April 2026; simplicity, security, and observability for AI workloads, November 11, 2025; AI observability dashboards GA, May 28, 2025; Universal Proxy first release, January 30, 2026; SUSE AI GA at KubeCon North America 2024; SLES 16 launch, November 4, 2025); the SUSE/suse-ai-up, SUSE/aif, rancher/k3k, and harvester/harvester repositories; NVIDIA's NIM Operator, NeMo Retriever, RAG reference architecture, and OpenShell documentation for the NVIDIA components. Peer-reviewed cell by cell through the labs claims ledger (suse-<layer>-chatgpt, ChatGPT gpt-5.5) and as a whole row by Antigravity (suse-row-agy); totals and escalated items in labs/reviews/suse-judgment.md. /reconcile (October 5, 2026): Layer 3 calibration citations updated for the Layer 3 re-reads (Nutanix, VAST, VMware to partner; CoreWeave, NVIDIA, AMD to gap) and the MongoDB BI ruling; no grade or authority change on this row. The Layer 3 grade is stated as resting on Industrial Edge; Liz and the Rancher Prime MCP server are named and excluded because Layer 3 credits first-party business applications only, components unchanged. /reconcile (October 5, 2026): Layer 3 escalation ruled: moderate holds; Industrial Edge is a domain-scoped, fixed-function slice and carries the grade alone. No grade, component, or authority change. /reconcile (October 5, 2026): override rule applied to a cell scored before it; layer0, layer1b, layer2a authority Ceded to Delegated on documented pins or per-request overrides. Grades and components unchanged. /reconcile (October 5, 2026): Layer 3 authority vendor / Delegated to code / Retained: Liz's gate no longer carries the grade, since Layer 3 credits first-party business applications only, and Industrial Edge executes the enterprise's designs (Losant docs). Grades and components unchanged. /reconcile (October 5, 2026): Layer 0 NVIDIA GPU Integration Delegated to Retained: the GPU Operator is Apache 2.0 open source the enterprise operates; the old reason ('substitutable at the accelerator') failed the Red Hat standard. vGPU licensing named as NVIDIA's. Grade unchanged. v1.7 (Layer 0 sub-layers, October 16, 2026): Layer 0 split into 0A Compute, 0B Network Fabric and 0C Storage Substrate by runtime decision (Keith, October 16, 2026); storage left 1A for 0C, and 1A, now Data Foundation & Governance, is graded on the catalog that classifies the data. 1A moderate to gap; 0A moderate, 0B moderate, 0C moderate; summary edits (summary worksheet, Keith, October 16, 2026); 2A post-write fix (Item 4); 3 post-write fix (Item 4). v1.8 (October 19, 2026, /reconcile partner ruling): calibration prose updated to the new Layer 3 readings of peer rows; no grade or reading on this row changed. (Same pass) Layer 3: the Liz and Rancher Prime MCP server chips removed, matching the cell's own reading that Layer 3 credits first-party business applications only; Industrial Edge carries the grade alone. Grade and authority unchanged. (Same pass) Plain-language pass: methodology rule numbers in the prose and audit notes replaced with the reasoning they stood for; no grade, chip or authority reading changed.