Red Hat is the instrument's open-source substrate vendor: an operating system, a Kubernetes platform, and an AI serving stack the enterprise can run on any hardware or cloud. It is strong where workloads are orchestrated (2A: OpenShift, Kueue, fleet policy) and run (2B: vLLM, llm-d, Models-as-a-Service), moderate across most of the data layers, a gap at 1B, where OGX's retrieval APIs are still in preview, a single enforcing gateway at 2C, and a gap at Layer 3, where it sells no business applications. Nearly every component reads Retained on the possess-and-operate line; this is the row the methodology's Red Hat standard describes.
The capture is decoupled and mostly not in the code. What accumulates is dependence on Red Hat's supported builds and on the OpenShift operating model (Routes, security context constraints, OLM-managed operators, Kuadrant policies), which lift to OKD and upstream projects without rebuilding but take rework on another Kubernetes distribution; the cost of leaving is support and operational familiarity, not reconstructed opinions. IBM's data, governance, agent, and application estate is scored on the IBM row. On-premises watsonx requires OpenShift, so the two rows compose in the common IBM deployment, and in that stack the authority above the substrate is IBM's.
The buyer's trade: the most portable AI infrastructure stack on the map, in exchange for assembling the data, retrieval, and agent-governance layers from parts, or buying them from IBM. The Responses API on OGX (formerly Llama Stack) reached general availability in 3.5, but the agent-governance surfaces (the MCP Gateway, the MCP catalog, agent runtime identity) are still in preview.
Layer-by-layer status: Layer 0A (Hardware-Agnostic Substrate: RHEL, OpenShift Virtualization, Accelerator Enablement), Layer 0B (OVN-Kubernetes and SR-IOV, Network Virtualization Over Any Fabric), Layer 0C (Ceph Through OpenShift Data Foundation; General-Purpose, Not Built for AI Throughput), Layer 1A (Bucket Policies Over Ceph; No Data Catalog), Layer 1B (Retrieval APIs in Preview: OGX Vector Stores, Vector I/O, and Docling Ingestion Unscored), Layer 1C (Kafka, Debezium, Camel, and Kubeflow Pipelines on Red Hat Subscriptions; No Enterprise Lineage), Layer 2A (One Control Plane for Containers, VMs, and GPUs: OpenShift, OpenShift AI Kueue, ACM, Ansible), Layer 2B (Open Serving Stack: vLLM, llm-d, Models-as-a-Service), Layer 2C (Enforcing Inference Gateway; Agent Governance in Preview), Layer 3 (+1) (No Business Applications; Certification Is Not Procurement).
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 27, 2026. Version: v1.24 - 4+1 v3: Layer 0 Sub-Layers.
Red Hat is the instrument's open-source substrate vendor: an operating system, a Kubernetes platform, and an AI serving stack the enterprise can run on any hardware or cloud. It is strong where workloads are orchestrated (2A: OpenShift, Kueue, fleet policy) and run (2B: vLLM, llm-d, Models-as-a-Service), moderate across most of the data layers, a gap at 1B, where OGX's retrieval APIs are still in preview, a single enforcing gateway at 2C, and a gap at Layer 3, where it sells no business applications. Nearly every component reads Retained on the possess-and-operate line; this is the row the methodology's Red Hat standard describes.
The capture is decoupled and mostly not in the code. What accumulates is dependence on Red Hat's supported builds and on the OpenShift operating model (Routes, security context constraints, OLM-managed operators, Kuadrant policies), which lift to OKD and upstream projects without rebuilding but take rework on another Kubernetes distribution; the cost of leaving is support and operational familiarity, not reconstructed opinions. IBM's data, governance, agent, and application estate is scored on the IBM row. On-premises watsonx requires OpenShift, so the two rows compose in the common IBM deployment, and in that stack the authority above the substrate is IBM's.
The buyer's trade: the most portable AI infrastructure stack on the map, in exchange for assembling the data, retrieval, and agent-governance layers from parts, or buying them from IBM. The Responses API on OGX (formerly Llama Stack) reached general availability in 3.5, but the agent-governance surfaces (the MCP Gateway, the MCP catalog, agent runtime identity) are still in preview.
CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant
The enterprise operating system under every Red Hat path, open source the enterprise runs itself with Red Hat's subscription for updates and support. Retained on the open-source seam.
Virtual machines and containers on the same Kubernetes API, namespaces, and RBAC, with GPU passthrough and vGPU for accelerated VMs. Open-source virtualization the enterprise operates: Retained, the SUSE Harvester reading.
OpenShift runs on Dell, HPE, Lenovo, Cisco, Supermicro, and the major clouds through one control plane. The enterprise's choice: Retained.
Red Hat's own documentation says "Red Hat supports Intel Gaudi 2 and Intel Gaudi 3 devices." The other accelerators' operators run on OpenShift but on their vendors' support: "The NVIDIA GPU Operator is only supported by NVIDIA," AMD's is "the AMD community release of the Operator," and IBM's Spyre Operator points to IBM support. Those are noted, not scored here. The Gaudi operator runs on Kubernetes the enterprise operates, Retained on the possess-and-operate line; what doesn't lift is the accelerator, read on Intel's row.
The most common accelerator path on OpenShift, packaged and supported on Red Hat's paper. AMD ROCm and IBM Spyre use the same operator model, so the dependency is real but not exclusive.
Red Hat sells the software that turns any vendor's hardware into an AI substrate, and nothing below it. RHEL 10 and RHEL AI are the operating system. OpenShift Virtualization runs VMs and containers on one Kubernetes with GPU passthrough and vGPU. Operators enable NVIDIA and AMD accelerators on x86 or Arm servers, on premises or on any major cloud, and IBM Spyre on Power and Z. The buyer gets one abstraction over its own metal. The architect's concern is that Red Hat owns no silicon and no rack, so compute density and accelerator topology come from the OEM and the accelerator vendor. Platform software at 0A reads moderate. Calibration: SUSE, 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 workloads 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: RHEL, OpenShift Virtualization and the open accelerator operators are open source the enterprise runs, Retained.
Resolved October 23, 2026: NVIDIA's OpenShift documentation describes passthrough and vGPU on OpenShift Virtualization as "tested and supported by Red Hat." Named, not scored: Intel Gaudi 3 and IBM Spyre on x86, which Red Hat AI 3 lists as Technology Preview, "not supported with Red Hat production service level agreements."
Interconnects, switching and routing, network virtualization
The default network provider for OpenShift: an overlay built on Open Virtual Network (OVN) and Open vSwitch, with distributed logical switches and routers, network policy, admin network policy, egress IPs and an egress firewall. Open source the enterprise runs itself under Red Hat's subscription: Retained on the open-source seam.
Multus attaches VMs and pods to secondary networks through NetworkAttachmentDefinitions (Linux bridge, OVN-Kubernetes layer2 and localnet), and the SR-IOV Network Operator manages SR-IOV devices and attachments for VMs, including DPDK workloads and a dedicated live-migration network. Open source the enterprise operates: Retained. For pods, the operator also runs SR-IOV devices in RDMA mode and attaches InfiniBand networks: setting the link type to InfiniBand turns RDMA mode on.
Assessment pending
OpenShift's default network plugin, OVN-Kubernetes, carries network policy, admin network policy, egress IPs and egress firewalls. The SR-IOV Network Operator with Multus gives VMs and pods secondary networks at line rate, including RDMA and InfiniBand attachments for pods. 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: SUSE, VMware and Nutanix read moderate on network virtualization; NVIDIA and HPE strong on fabrics they sell.
Ceded where the grade rests. The network policies the enterprise writes are executed as written, but OVN-Kubernetes assigns egress IPs across the nodes the enterprise labels, and a label is a bound, not a pin: vendor decides, visible, not overridable, Ceded. Portability: OVN and the SR-IOV operator are open source the enterprise runs, Retained.
Enablement, not a fabric (October 22, 2026): OpenShift AI 3.5 documents multi-node PyTorch training over NVIDIA GPUDirect RDMA on InfiniBand or RoCE, with no preview label, configured through the NVIDIA Network Operator. The NICs, the driver stack and the fabric are NVIDIA's, so it reads as a note, not a chip.
Arrays, storage servers and software, front-end protocols: where the data sits
Block, file, and S3 object storage for OpenShift workloads. IBM owns the product and sells the full tier as Fusion Data Foundation; Red Hat delivers Essentials through its own paper. Open-source Ceph the enterprise operates: Retained.
Assessment pending
OpenShift Data Foundation Essentials, included in OpenShift Platform Plus, gives OpenShift workloads Ceph-based block, file and S3 object storage through Rook and the NooBaa gateway. The buyer gets storage that lives with the platform, on the same servers. The architect's concern is throughput and roadmap. ODF is general-purpose software-defined storage, not documented for feeding accelerators, and Red Hat's storage portfolio moved to IBM in 2023, which sells the full product as Fusion Data Foundation. Calibration: SUSE and Nutanix read moderate on the same shape; IBM strong on Storage Scale.
Delegated where the grade rests. Ceph places data within the pool and bucket policies the enterprise sets per object, policies the engine must honor: vendor decides, visible, overridable, Delegated. Portability: Ceph is open source the enterprise operates, Retained.
Confidence flag: where ODF Essentials ends and IBM's paid tier begins (external key manager encryption, Regional DR) wasn't confirmed against Red Hat's Platform Plus feature list.
The data itself, organized and governed: the Governance Catalog that Layer 2C queries
Block, file, and S3 object storage for OpenShift workloads. IBM owns the product and sells the full tier as Fusion Data Foundation; Red Hat delivers Essentials through its own paper. Open-source Ceph the enterprise operates: Retained.
Assessment pending
Red Hat's 1A function is access governance over the data its platform stores. OpenShift Data Foundation enforces the bucket policies the enterprise sets on Ceph's S3 object storage, deciding who may use which data. Nothing on Red Hat's paper answers what the data is: sensitivity, classification, lineage or freshness. Advanced Cluster Security governs workloads, not data, so it reads at 2A with the fleet. In the common IBM-plus-Red Hat stack, the catalog a reasoning plane would query is IBM's watsonx.data intelligence, scored on the IBM row. That's governance over data without a queryable classification catalog: moderate. Calibration: Nutanix reads moderate on Data Lens access governance; SUSE gap, with no access function documented on its row.
Ceph enforces the bucket policies the enterprise wrote, as written: code decides, visible, overridable, Retained, the AWS Lake Formation reading. Portability: Ceph is open source the enterprise operates, Retained.
Public evidence that moves the cell: a data catalog with classification and lineage on Red Hat's paper.
Low-latency retrieval for RAG — vector/hybrid search, context windows
GPU-accelerated vector search and retrieval microservices are available on OpenShift AI; neither is required by the open retrieval path.
Red Hat's retrieval path is OGX on OpenShift AI, and its retrieval surfaces haven't left preview. OpenShift AI 3.5's OGX documentation lists the Vector_IO API, which lets the enterprise "manage and query vector embeddings," as Developer Preview, the Vector Stores API as Technology Preview, the Vector Store Files API as Developer Preview, and the File Processors API that runs Docling's parsing and chunking as Developer Preview; the Files and Embeddings APIs are Technology Preview. The Responses API is generally available, but it "orchestrates inference and retrieval but relies on separate providers for file storage (Files API) and vector indexing (Vector Stores API)," so every documented RAG workflow runs through preview surfaces. The vector engines OGX connects to (remote Milvus, PostgreSQL with pgvector, Qdrant) are their owners' software; Red Hat supports the connector, not the engines. What the enterprise can run in production on Red Hat's paper today is the Responses API, with retrieval it builds from engines it deploys and operates itself. Lab 005 (Authority you reclaim is authority you run) shows what that build takes on Kubernetes: moving the CTO Advisor's own RAG application off a managed store, the vectors went to CloudNativePG and pgvector behind the application's own retrieval seam, and the retrieval headend was the application's code. That headend is the part a platform's generally available retrieval service would supply, and on OpenShift AI today it doesn't. Applying the GA-gate: preview surfaces go to notes and never into scored components, and the layer's function has no generally available Red Hat product behind it. The Responses API's orchestration is read at 2B. Calibration: Cloudera reads gap with authority Absent on the same shape, retrieval still in Technical Preview; CoreWeave reads gap on enterprise-provided retrieval; Cloudflare and Cohere read moderate on shipped, generally available primitives. Red Hat has its retrieval primitives in preview. Gap, Absent, with the preview surfaces watch-listed.
None to borrow at the layer's function today: the retrieval the enterprise runs in production on Red Hat's platform is its own assembly of vector engines it deploys and operates. Absent: nothing generally available offered, nothing inherited.
Grade moved October 24, 2026: moderate to gap when OpenShift AI 3.5's OGX documentation showed the retrieval APIs in preview (GA-gate). Watch-list, notes only, and these move the cell: the OGX Vector_IO API (Developer Preview), Vector Stores API (Technology Preview), Vector Store Files API (Developer Preview), File Processors API with Docling (Developer Preview), Files and Embeddings APIs (Technology Preview), AutoRAG (Technology Preview since OpenShift AI 3.4; its 3.5 additions, the visual pipeline, multilingual support, and structural contextualization, are Technology Preview too) and the default vector store for Gen AI Studio (Technology Preview in 3.5). Red Hat's review of v1.2 asked for a status label built on AutoRAG; the docs' Technology Preview badge governs. When the retrieval APIs reach GA, OGX and Docling are open source the enterprise runs, Retained. OpenShift AI's OGX "supports remote Milvus, PostgreSQL with the pgvector extension, and Qdrant as vector store providers"; Red Hat supports the connector, not the engines, and Elasticsearch appears only as a validated pattern. Docling is an IBM-originated Linux Foundation project. Public evidence that moves the cell: a generally available support level on the Vector Stores or Vector_IO API.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
Event streaming with schema registry and consumer-group management under a Red Hat support SLA. Open-source Kafka the enterprise runs: Retained, the open-core reading on the Confluent row.
Row-level change capture from operational databases into Kafka. Open source the enterprise operates: Retained.
Integration and ETL routes across enterprise and SaaS systems, with Red Hat updating connectors when upstream APIs change. Open source the enterprise operates: Retained.
ML data preparation and training pipelines with experiment tracking and ML-artifact lineage. Open source the enterprise operates: Retained on the possess-and-operate line.
Assessment pending
A general, open-source data-movement toolkit on Red Hat's paper. Red Hat Application Foundations carries Streams for Apache Kafka (Strimzi) with the Apicurio schema registry, the Red Hat build of Debezium for change data capture out of operational databases, and the Red Hat build of Apache Camel with more than 300 Red Hat-maintained connectors for SAP, Salesforce, Oracle, and the major enterprise systems. Red Hat AI Enterprise adds OpenShift AI Data Science Pipelines (Kubeflow Pipelines on Argo), Ray, and MLflow, with lineage for ML artifacts: which pipeline step produced which model. All of it is self-run and lifts to any Kubernetes. The architect's concern: the enterprise assembles the pipeline from parts. There is no lineage across enterprise data (which operational record fed which training set), no managed stream-processing engine, and no lakehouse destination of Red Hat's own. Moderate at the top of the band: general (Camel routes, Kafka topologies, and Kubeflow pipelines take whatever shape the problem defines), but strong is pegged to the best shipping implementation, and the strong 1C cohort (Snowflake Openflow, Databricks Lakeflow, Qlik CDC plus Talend, VAST DataEngine, Confluent, IBM) covers batch, streaming, CDC, declarative transformation, and lineage as one integrated surface. Above Elastic's moderate on Logstash and agents. IBM's Confluent and DataStage are scored on the IBM row; two Kafka offerings in one corporate family is a fact worth knowing at procurement.
Low. Kafka, Debezium, Camel, Kubeflow Pipelines, Ray, and MLflow are open source the enterprise operates: Retained on the possess-and-operate line. The engines execute routes and pipelines the enterprise writes, with no vendor judgment in between: code decides, visible, overridable, Retained, the Elastic 1C reading.
Watch-list, notes only: the Feast feature store (entered as Technology Preview in OpenShift AI 2.20; general availability not confirmed). The Fourth Cloud Red Hat assessment reads Kafka and Camel under Fourth Cloud's own rules; the two instruments ask different questions, so the difference isn't a reconcile item. KV cache tiering, named, not scored: the Red Hat AI 3.5 announcement states general availability of CPU KV cache offloading, but no product document carries it as a generally available feature; KV cache offloading to shared storage in Distributed Inference with llm-d (GPU memory to CPU DRAM to CephFS on OpenShift Data Foundation) is Developer Preview in Red Hat AI Inference 3.5.0, as is Hierarchical KV Cache Tiering in OpenShift AI 3.5. Neither would move the grade, which rests on the missing integrated surface and enterprise lineage. Public evidence that moves the cell: enterprise data lineage on Red Hat's paper, or an integrated pipeline surface.
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
Enterprise Kubernetes for containers, VMs, databases, and batch on one API, on premises and on every major cloud. Open source the enterprise runs (OKD upstream): Retained.
Native GPU queue management and multi-tenant quota enforcement across NVIDIA, AMD ROCm, Intel Gaudi, and IBM Spyre. Open source the enterprise operates (Open Data Hub upstream): Retained.
Multi-cluster governance policies applied consistently across the fleet with drift detection. Open source (Open Cluster Management upstream): Retained.
Infrastructure automation across hybrid environments, including OEM firmware lifecycle coordinated with OpenShift maintenance. Open source (AWX upstream): Retained.
Continuous compliance against CIS, NIST, PCI DSS, and HIPAA for Kubernetes workloads: vulnerability scanning, runtime policy, network segmentation, image provenance, blocking non-compliant workloads at deploy time. Workload governance, not data governance. Open source (StackRox) the enterprise runs: Retained.
GPU node management on OpenShift. Kueue provides GPU queueing and fair-share natively, which makes Run:ai optional rather than required.
One Kubernetes control plane for containers, virtual machines, databases, batch, and AI workloads, the same on premises and on every major cloud. OpenShift AI adds Kueue: GPU queueing, fair-share scheduling, and multi-tenant quotas across NVIDIA, AMD ROCm, Intel Gaudi, and IBM Spyre. Advanced Cluster Management (Platform Plus) enforces policy across the whole fleet from one place, and Ansible Automation Platform automates the infrastructure around it. The architect's concern: on premises, scheduling is bounded by the physical nodes the enterprise owns, and deep multi-node GPU scheduling still leans on NVIDIA's operators. Strong: the strong 2A cohort (AWS on EKS, GCP, Azure, Nutanix, CoreWeave, SUSE) reads on managed or self-run Kubernetes with GPU scheduling, and OpenShift with Kueue and fleet policy stands with it. In the common IBM deployment, IBM's own schedulers (Spectrum LSF, Turbonomic) sit beside and above this layer; they are scored on the IBM row.
Low on portability, and pinnable at runtime. OpenShift, OpenShift AI, Advanced Cluster Management, and Ansible are open source the enterprise operates (OKD, Open Data Hub, Open Cluster Management, and AWX upstream): Retained. This is the behavior the methodology's Red Hat standard describes. Quotas, fair-share, and Kueue admission are bounds the scheduler works within, which is configuration; a nodeSelector or required affinity rule is a per-workload pin the engine must honor: vendor decides, visible, overridable, Delegated, the CoreWeave 2A reading.
Instrument follow-up: SUSE, Nutanix, and VMware read Kubernetes pins Ceded at 2A from before the override rule; this cell and CoreWeave read them Delegated. Placement (October 16, 2026): Red Hat Advanced Cluster Security reads here under the function audit and the hardware-management placement rule.
Model serving, agent execution, inference APIs, distributed inference
Production model serving with an OpenAI-compatible API, tool calling, and structured outputs across accelerator vendors. Open source the enterprise runs: Retained on the possess-and-operate line, the SUSE vLLM reading; the standard interface is a further exit, not the basis. Not scored within this chip under the GA evidence table: Intel (Gaudi) ("The following container images are Technology Preview features:").
Kubernetes-native distributed inference with KV-cache-aware routing, and priority-based flow control for mixed workloads (generally available in 3.5). Disaggregated prefill-decode is Technology Preview on OpenShift. Open source the enterprise operates: Retained.
Governed self-service model access for teams, built on Kuadrant policies: subscriptions that grant groups access to specific models with token quotas and rate limits, and self-service API keys, in front of the cluster’s own models through an OpenAI-compatible API. Open-source policy resources the enterprise operates: Retained on the possess-and-operate line.
A validated catalog of open-weight models (Granite, Llama, Qwen, Mistral, gpt-oss) supported on Red Hat AI. Open weights the enterprise runs anywhere: Retained. Granite is also scored on the IBM row as the model's owner.
LLM call tracing, tool execution tracking, and reasoning-step auditability. Open source the enterprise runs: Retained.
Evaluation orchestration across lm-evaluation-harness, RAGAS, Garak, and GuideLLM, with safety benchmarks and compliance reporting before a model, RAG configuration, or agent deploys. It evaluates and doesn't enforce at runtime, so it carries no 2C weight. Open source (Apache 2.0) the enterprise operates: Retained.
The business logic, APIs, and deterministic validators the enterprise invokes through tool calls on the OpenAI-compatible serving API. The enterprise's own code: Retained.
Most production serving runs on NVIDIA GPUs, and NVIDIA NIM is an available path; the Red Hat runtime itself serves across NVIDIA, AMD, Intel, and IBM Spyre.
A complete open serving stack under one support relationship. Red Hat AI Inference is supported vLLM with an OpenAI-compatible API across NVIDIA, AMD, Intel, and IBM Spyre accelerators; llm-d (generally available since OpenShift AI 3.0) adds Kubernetes-native, KV-cache-aware distributed inference, the open counterpart to NVIDIA's Dynamo, and 3.5 adds generally available priority-based flow control for mixed workloads; Models-as-a-Service (generally available in 3.4) adds token quotas, rate limits, self-service API keys, and showback; and a validated model catalog covers Granite, Llama, Qwen, Mistral, and gpt-oss. Tool calling and structured outputs are documented, so the runtime carries the enterprise's own agent loop, and EvalHub (generally available in 3.5) evaluates models, RAG configurations, and agents before they deploy. The architect's concern: Red Hat's hosted agent surfaces lag. The Responses API reached general availability on OGX (formerly Llama Stack) in 3.5, but tool-calling passthrough in Distributed Inference with llm-d and the MCP Gateway are Technology Preview, and OpenShell agent sandboxing is Developer Preview. Strong, under the September 11 agent-execution note: strong is judged by whether the runtime carries the enterprise's own agent loop, and a hosted agent runtime is never the price of it (Cerebras and Groq hold strong with no hosted loop). Calibration: NVIDIA strong on NIM and Dynamo, which Red Hat matches in open source; SUSE moderate on packaged vLLM, Ollama, and NIM without a distributed-inference layer or governed serving of its own; IBM's watsonx.ai, which runs on this engine in the common IBM stack, is scored on the IBM row.
Low, and the lowest on the instrument at this layer. vLLM, llm-d, the Kuadrant-based Models-as-a-Service policies, MLflow, EvalHub, the validated open-weight models, and the enterprise's own tools are all things the enterprise possesses and operates: Retained. Models-as-a-Service policies are Kuadrant resources that lift to any cluster running Kuadrant, not to another gateway product, which is the residual seam. The model decides which tool to call, and the enterprise's own tool code, executing client-side and able to refuse, gates the effect: model decides, visible, overridable, Delegated, the NVIDIA reading.
Watch-list, notes only: tool-calling passthrough in Distributed Inference with llm-d (Technology Preview in 3.5), validated tool-calling configurations in the model catalog (Technology Preview), disaggregated prefill-decode on OpenShift (Technology Preview; generally available only on CoreWeave Kubernetes Service), the MCP Gateway (Technology Preview in 3.5), the MCP catalog administrative interface and the Agent Catalog's agent starter kits (Developer Preview), and OpenShell agent sandboxing (Developer Preview in 3.5). Red Hat's review of v1.2 read tool calling as generally available in both standard and distributed modes; the docs carry the distributed passthrough as Technology Preview. The Responses API on OGX left the watch-list at general availability in 3.5. InstructLab is not scored: its status as a supported product, rather than a community project, wasn't confirmed. Out of the Red Hat AI Inference chip under the GA evidence table (October 24, 2026): Intel (Gaudi) ("The following container images are Technology Preview features:"). Out of the Models-as-a-Service chip under the GA evidence table (October 24, 2026): usage tracking and showback, routing to external model providers, external OIDC authentication and vLLM runtime support, each Technology Preview in OpenShift AI 3.4 ("Usage tracking and showback reporting through an observability dashboard (Technology Preview)").
Policy-driven placement and resource coordination — the Autonomy Layer
Request-time enforcement of the enterprise's authorization, rate-limit, and token-quota policies on inference endpoints. Open source the enterprise operates: Retained.
Assessment pending
The Connectivity Link AI inference gateway (Kuadrant on Envoy and Istio, generally available in OpenShift AI 3.4) enforces access control, rate limits, and token quotas on every inference call under policies the enterprise writes, and Models-as-a-Service adds per-team keys and showback. That is one leg of a plane, and it acts: it authorizes and blocks requests at runtime, which clears the action line. The architect's concern: it governs model calls, not agents. The MCP Gateway (identity-based tool filtering, OAuth token exchange, agent traceability) is still Technology Preview in 3.5, the MCP Lifecycle Operator is Technology Preview, the MCP catalog and agent runtime identity are Developer Preview, and there is no generally available agent identity, agent registry, or cross-agent orchestration. Moderate, at the thin end of the band. Calibration: Nutanix reads moderate on a single generally available gateway (RBAC, per-token rate limits, audit); Cloudflare moderate on a gateway over model calls plus endpoint firewalling; GCP and Azure strong on complete planes; IBM moderate on watsonx Orchestrate's control plane, scored on the IBM row. EvalHub, generally available in 3.5, evaluates agents before they deploy but blocks nothing at runtime, so it's scored at 2B, the layer its function fits. llm-d's KV-aware request routing is routing, not reasoning; the live per-inference placement gap and the missing deterministic outcome-validator are universal across the instrument, noted rather than charged to Red Hat. The Fourth Cloud assessment scores this layer 0 because it grades autonomous placement reasoning; this instrument grades action governance.
Low. The gateway is open source the enterprise operates, and it exercises no judgment of its own: it enforces the access, rate, and token policies the enterprise writes. Under the override rule that reads code decides, visible, overridable, Retained, whoever runs the engine, the Azure, Salesforce, and Elastic reading. Cloudflare's and Nutanix's gateways read vendor and Delegated because they make vendor-side routing and failover decisions; this one executes the enterprise's declarative policy.
Watch-list, notes only: the MCP Gateway (Technology Preview in OpenShift AI 3.5) and its NeMo Guardrails tool-call enforcement (Technology Preview), the MCP Lifecycle Operator (Technology Preview), the MCP catalog (Developer Preview), agent deploy and runtime management with SPIFFE workload identity (Developer Preview), and OGX agents. Red Hat's review of v1.2 read 3.5 as an agentic governance suite; of what it named, EvalHub is generally available but evaluates rather than enforces, priority-based flow control is serving (2B), and the agent templates are the Developer Preview Agent Catalog. Public evidence that moves the cell: the MCP Gateway reaching general availability, or a generally available agent registry or agent identity on Red Hat's paper.
AI-powered business capabilities — business logic, workflow automation
NVIDIA's software and reference blueprints are among the certified partner offerings on OpenShift; context, not a Red Hat application.
Red Hat ships no business applications. Its first-party AI surfaces sit outside the Layer 3 grade, which credits first-party business applications only: Ansible Lightspeed and OpenShift Lightspeed are assistants over Red Hat's own automation and console, Developer Hub and Dev Spaces are developer tools, and OperatorHub and the Ecosystem Catalog are distribution. ISVs certified on OpenShift and OpenShift AI deploy as operators under Red Hat's support matrix, but certification is a support statement about the platform, and the ISV still sells and supports its application. Layer 3's grade rests on business applications the enterprise can buy or get supported on the vendor's paper (as ruled October 19, 2026: business applications bought or supported on the vendor's paper). Certification isn't procurement. Calibration: Cisco, Arista and Nokia read moderate on first-party security applications; HPE reads partner on ISV applications customers buy from HPE on the GreenLake Marketplace; Dell, Nutanix, Red Hat, VMware, VAST and Supermicro read gap on the same certification-only shape.
None on Red Hat's paper. The enterprise borrows application judgment from the certified ISVs it buys from; Red Hat certifies that they run.
Named, not scored: the Red Hat Marketplace, operated by IBM, which sold ISV software until it was sunset on April 1, 2025; Red Hat Certified Partner Ecosystem and the Ecosystem Catalog (certification), Lightspeed assistants (consoles), Developer Hub and Dev Spaces (developer tools). Public evidence that moves the cell: a business application sold or supported on Red Hat's paper.
Red Hat OpenShift self-managed subscription guide; Red Hat Application Foundations subscription guide; OpenShift AI 3.0, 3.3, 3.4, and 3.5 release notes and documentation (Llama Stack, llm-d, Models-as-a-Service, Data Science Pipelines); Red Hat AI 3 announcement (October 2025); Red Hat AI Inference documentation (tool calling, supported accelerators); IBM documentation on Fusion Data Foundation and the 2022 move of Red Hat storage to IBM; Red Hat Developer articles (RAG on OpenShift AI, the Camel Q2 2026 digest, AI pipelines in OpenShift AI 3.3); the Fourth Cloud Red Hat OpenShift assessment v2.6 (May 29, 2026), including Red Hat vendor feedback; published 4+1 model. v1.0 (/reconcile, October 5, 2026): split from the former IBM / Red Hat OpenShift AI row along the vendor support boundary, every cell graded and reconciled against SUSE, VMware, Nutanix, CoreWeave, NVIDIA, Elastic, Confluent, Cloudflare, and the rescoped IBM row. /reconcile (October 5, 2026): the 1C note treating Fourth Cloud's Kafka and Camel readings as a follow-up is settled; Fourth Cloud has its own rules (Keith). Text only. /reconcile (October 5, 2026): Layer 0 Accelerator Enablement Delegated to Retained: the operators and drivers are open source the enterprise operates (Spyre Operator Apache 2.0); the old reason ('substitutable at the accelerator') failed the Red Hat standard. Grade unchanged. v1.3 (October 15, 2026): Red Hat's review of v1.2 adjudicated against product documentation (reviews/redhat-vendor-response-2026-10-15.md), using the Red Hat AI 3.5 announcement (September 9, 2026), the OpenShift AI 3.5 release notes (new features, Technology Preview, and Developer Preview chapters), the Red Hat AI Inference 3.5.0 release notes, and the EvalHub repository (github.com/eval-hub/eval-hub, Apache 2.0). Where the announcement and the docs differ, the docs' status badge governs. Llama Stack renamed OGX; Responses API generally available; EvalHub added at 2B; AutoRAG and KV cache tiering named, not scored. Text only: no grade or DAPM change. v1.4 (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. 0A moderate, 0B moderate, 0C moderate. v1.5 (October 19, 2026, /reconcile partner ruling): Layer 3 partner to gap, authority Absent; the AI ISV program is certification, not procurement on the vendor's paper. (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. v1.6 (October 22, 2026, /reconcile): 0B SR-IOV chip records RDMA mode and InfiniBand attachments for pods (OpenShift 4.20 hardware networks), and the notes record OpenShift AI 3.5's GPUDirect RDMA training as enablement of NVIDIA's NICs and fabric, not a Red Hat fabric. Found by the 0B fabric research. Grade and authority unchanged. v1.7 (October 23, 2026, /reconcile): 0A's enablement chip and gap scope accelerator support to what Red Hat AI 3 documents without a preview label (NVIDIA, AMD with ROCm and its GPU Operator, IBM Spyre on Power and Z); Intel Gaudi 3 and Spyre on x86, Technology Preview, move to a named, unscored note. The vGPU confidence flag closes on NVIDIA's OpenShift Virtualization page. A rule code comes out of the gap. Found by the 0A compute research. Grade and authority unchanged. v1.8 (October 23, 2026, /reconcile): 0A: Red Hat's docs contradict the chip's "packaged and supported by Red Hat" for NVIDIA, AMD and Spyre. Under "does the vendor sell the thing," only Gaudi stays on Red Hat's paper. Found by the Layer 0 data pass after the sniff read. Grade and cell authority unchanged. v1.9 (October 23, 2026, /reconcile): Layer 1 research pass. 1B: "Vector Database Integration (Milvus, pgvector, Elasticsearch" removed (Red Hat supports the OGX connector to remote Milvus, pgvector and Qdrant, not the engines; Elasticsearch is a validated pattern). Found by the Layer 1 slot cut. Grade and cell authority unchanged. v1.10 (October 23, 2026, Layer 1 calibration): The pgvector chip comes off: Red Hat documents pgvector as a remote provider for OGX, not a Red Hat product. Grade unchanged. v1.11 (October 24, 2026, /reconcile): 1B moves from moderate to gap: OpenShift AI 3.5 documents OGX's retrieval APIs as Developer or Technology Preview. Found re-reading the cell after the Layer 1 calibration pass. Grade changed: 1B moderate to gap. v1.12 (October 24, 2026, /reconcile): 1B cites Lab 005 as evidence of what retrieval on the platform requires and the gap Red Hat's retrieval APIs leave. Grade unchanged. v1.13 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 1 Layer 1A chips under the GA evidence table version 2. Grade unchanged. v1.14 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 4 Layer 1C chips under the GA evidence table version 2. Grade unchanged. v1.15 (October 24, 2026, /reconcile): GA evidence recorded on 2 of 2 Layer 0B chips under the GA evidence table version 2. Grade unchanged. v1.16 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 4 Layer 0A chips under the GA evidence table version 2. Grade unchanged. v1.17 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 1 Layer 0C chips under the GA evidence table version 2. Grade unchanged. v1.18 (October 24, 2026, /reconcile): GA evidence recorded on 5 of 5 Layer 2A chips under the GA evidence table version 2. Grade unchanged. v1.19 (October 24, 2026, /reconcile): GA evidence recorded on 6 of 7 Layer 2B chips under the GA evidence table version 2. Grade unchanged. v1.20 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 1 Layer 2C chips under the GA evidence table version 2. Grade unchanged. v1.21 (October 24, 2026, /reconcile): GA evidence table version 2.1 applied: 1 chips recorded or narrowed. Grade unchanged. v1.22 (October 24, 2026, /reconcile): GA evidence quotes re-pointed to the live owner pages (4 chips) after the quote spot-check. Grade unchanged. v1.23 (October 24, 2026, /reconcile): showback and the other Technology Preview MaaS items out of the Models-as-a-Service chip. Grade unchanged. v1.24 (October 27, 2026): each component now carries a durable offering id; nothing else changed.