Red Hat AI subscription guide

This document clarifies the subscription choices for Red Hat® AI offerings and provides step-by-step guidance on how to approximate the necessary entitlements for your organization’s Red Hat AI environment. Red Hat AI is a platform that accelerates model deployment and reduces the operational cost of developing and delivering AI solutions across hybrid cloud environments. The value is a complete, comprehensive AI platform that simplifies procurement and deployment.

Red Hat AI provides a flexible, workload-specific pricing model that shifts away from only central processing unit (CPU) counting, allowing better alignment with the hardware demands of AI. 

Depending on the product selected, the pricing model follows 1 of 3 primary structures:

1. Red Hat AI Enterprise is a bundled AI platform. It uses a flat-rate per-node model that includes unlimited AI accelerators such as graphics processing unit (GPU) entitlements and removes CPU core and socket limitations. The node is restricted to AI use cases only.
2. Red Hat AI Inference and Red Hat Enterprise Linux® AI are accelerator-centric. These products are licensed per physical accelerator, such as GPU and tensor processing unit (TPU), CPU core counts do not affect the subscription cost.
3. Red Hat OpenShift® AI is a layered add-on. It mirrors standard Red Hat OpenShift units (core-pairs or bare-metal physical nodes) but requires separate Red Hat AI Accelerator SKUs for each physical GPU used.

Red Hat AI is a portfolio of products ranging from point-product offerings to an integrated AI platform for developing and deploying efficient and cost-effective AI models, agents, and applications across open hybrid cloud environments. It supports on-site, virtual, and physical infrastructure, and private cloud, public cloud, and edge deployments. Red Hat AI products and SKUs include:

• Red Hat AI Enterprise: A platform used for developing and deploying AI applications at scale. Subscriptions are billed on a per-node basis, bundling Red Hat OpenShift AI, the underlying Red Hat OpenShift Container Platform, and unlimited Red Hat AI Accelerator entitlements into a single SKU.
• Red Hat AI Inference: A specialized offering used to optimize model inference across hybrid cloud environments for faster, more cost-effective deployments. Subscriptions are licensed on a per-GPU basis.
• Red Hat OpenShift AI: A layered platform used for managing the lifecycle of predictive and gen AI models. Subscriptions mirror standard Red Hat OpenShift units per virtual central processing unit (vCPU) or per bare-metal node. While it includes inference components, such as vLLM, it requires the underlying Red Hat OpenShift entitlement.
• Red Hat Enterprise Linux AI: A foundation model platform used for running large language models (LLMs) on individual servers. Red Hat Enterprise Linux AI combines the high-efficiency components of Red Hat AI Inference with an image mode instance of Red Hat Enterprise Linux.
• Red Hat AI Accelerator: A specific SKU that allows certified Kubernetes operators (NVIDIA, AMD, Intel Gaudi, etc.) to function within Red Hat OpenShift and OpenShift AI. This SKU is an entitlement SKU, not a standalone product. It is required for OpenShift AI customers using hardware accelerators but is already included in Red Hat AI Enterprise, Red Hat AI Inference, and Red Hat Enterprise Linux AI.

Red Hat AI products feature a variety of subscription models, ranging from per node, vCPU, and bare metal to per GPU. Refer to the documentation for Red Hat AI to learn more about deployment methods and supported infrastructure types. To get a more detailed description of specific Red Hat AI software product subscriptions, refer to Appendix 2. 

Red Hat AI software product subscription summary:

Red Hat AI Enterprise

• Metric: Measured using node-based, flat-rate pricing.
• Hardware: Unlimited CPU cores and AI accelerators (such as GPUs) allowed per node.
• Virtualization: One bare-metal subscription entitles your organization to unlimited Red Hat AI Enterprise virtual machines (VMs) on Red Hat OpenShift Virtualization Engine.
• Use: Included Red Hat OpenShift entitlements are strictly limited to AI use cases (training, tuning, inference, and agents).

Red Hat AI Inference

• Metric: Licensed per physical accelerator (such as GPU and TPU).
• Entitlements: Standalone Red Hat AI Accelerator SKU is not required.
• Usage: Dedicated to serving LLMs.
• Platform: Deployable on Red Hat Enterprise Linux, Red Hat OpenShift, or third-party Linux and Kubernetes platforms.

Red Hat Enterprise Linux AI 

• Metric: Licensed per physical accelerator (such as GPU and TPU).
• Entitlements: Standalone Red Hat AI Accelerator SKU is not required.
• Environment: Includes the necessary Red Hat Enterprise Linux image mode instance.
• Scope: Intended for single-server environments.

Red Hat OpenShift AI

• Metric: Follows standard Red Hat OpenShift core-based or bare-metal models.
• Requirements: A base Red Hat OpenShift subscription is mandatory.
• Accelerators: Requires an additional Red Hat AI Accelerator SKU for each GPU or TPU used.
• Scope: The specific subset of compute nodes running AI workloads must be entitled, rather than the entire cluster.
• Note: This model typically requires more diligent calculation to ensure sizing accuracy.

Red Hat AI Enterprise worker nodes include OpenShift Container Platform entitlements and are compatible with clusters of several Red Hat OpenShift variants. Red Hat AI Enterprise nodes can be added to clusters running OpenShift Virtualization Engine, Red Hat OpenShift Kubernetes Engine, and others, using the OpenShift Container Platform entitlements specifically on those Red Hat AI Enterprise nodes.

Key use-case and compliance rules:

• Use restrictions: OpenShift Container Platform capabilities included with Red Hat AI Enterprise must be dedicated exclusively to AI use cases.
• The Red Hat AI Enterprise exception to the single-version rule: Standard Red Hat terms require that only a single version of the Red Hat OpenShift platform (e.g., OpenShift Virtualization Engine, OpenShift Container Platform, or Red Hat OpenShift Platform Plus) be deployed within a single cluster. Red Hat AI Enterprise is a specific exception to this rule. Users may deploy Red Hat AI Enterprise nodes within clusters made up of nodes running any of these platform versions.

OpenShift AI subscription types

Mirroring the underlying Red Hat OpenShift product, OpenShift AI offers 2 subscription models, each available with Standard (8x5) or Premium (24x7) support service level agreements (SLAs). Subscriptions are required for compute nodes running OpenShift AI components or AI workloads.

1. Core-pair subscriptions (2 cores or 4 vCPUs)

• Calculation: Count the aggregate number of physical cores or vCPUs across compute nodes in the OpenShift clusters your organization intends to entitle.
• Compatibility: Supported on OpenShift Container Platform, OpenShift Kubernetes Engine, and OpenShift Platform Plus.
• Mixed environments: Core-pair SKUs can be used for OpenShift AI even if the underlying Red Hat OpenShift environment is licensed via bare metal.

2. Bare-metal subscriptions (1 physical node)

• Calculation: 1 subscription covers 1 physical server, regardless of CPU socket or core count.
• Hardware requirements: Available only for x86 and Arm physical nodes where Red Hat OpenShift is installed directly on the hardware.
• Restrictions: 
  • IBM Power and IBM Z are excluded from this model (these require core-pair subscriptions).
  • Virtualization is only permitted with OpenShift Virtualization Engine—an included feature of Red Hat OpenShift.
  • If a third-party hypervisor is being used, then the VMs must be tallied up using the 2-core/4-vCPU SKU.
• Prerequisite: This option requires an underlying subscription to a Red Hat OpenShift bare-metal platform (OpenShift Kubernetes Engine, OpenShift Container Platform, or OpenShift Platform Plus).

Choosing the right Red Hat OpenShift platform subscription

Consult the Red Hat OpenShift subscription guide to determine when it makes sense to choose core-pair subscriptions or bare-metal socket pair subscriptions.

Adding AI Accelerator SKUs to Red Hat OpenShift and OpenShift AI 

When using Red Hat OpenShift with or without OpenShift AI for AI workloads, Red Hat AI Accelerator SKU subscriptions must be added for the hardware accelerators in your organization’s environment. Accelerators are only counted when they are used to execute a compute workload. A workload is considered a compute workload when its primary purpose is not actively drawing pixels on a user’s screen in near real time or moving data across a network.

Examples of AI compute workloads (SKU required):

• General AI software: Applications written in Python, Java, or Perl for AI tasks.
• High-intensity models: Running LLMs or other compute-heavy software.
• Data science: Used for model training and fine-tuning.
• Scientific simulation: Physical modeling capabilities such as fluid dynamics or protein folding are available.

AI accelerator (1 accelerator unit)

• Requirement: This subscription is mandatory for discrete hardware—such as GPUs, TPUs, neural processing units (NPUs), field-programmable gate arrays (FPGAs), or data processing units (DPUs)—that provides compute acceleration for AI workloads and is not part of the integrated CPU package.
• Unified entitlement: The same SKU is used for each physical accelerator, regardless of the Red Hat OpenShift edition. A single accelerator subscription covers both Red Hat OpenShift and OpenShift AI if both are installed on the same cluster.
• Support SLAs: Available with Standard (8x5) or Premium (24x7) support. The SLA must match the support level of the underlying core-pair or bare-metal subscription.

How to count Red Hat AI Accelerator SKU subscriptions for OpenShift AI

A wide range of hardware—including GPUs, TPUs, application-specific integrated circuits (ASICs), NPUs, and FPGAs—is classified as AI acceleration. These typically consist of a physical card or board occupying a peripheral component interconnect (PCI) slot. The subscription count is almost always based on the unit quantity purchased from the hardware vendor (e.g., a server with 8 GPUs requires 8 Red Hat AI Accelerator SKU subscriptions).

Counting logic and examples

Each subscription covers 1 physical accelerator device.

• Physical nodes: A server with 4 physical GPUs requires 4 Red Hat AI Accelerator SKU subscriptions (in addition to the node's CPU core-pair or socket-pair entitlements).
• Virtualization (virtual GPU): A virtual node using 1 physical GPU—even if partitioned into multiple virtual graphics processing units (vGPUs)—requires only 1 subscription. Counting is based on physical hardware, not virtual slices.
• Public cloud environments: In environments with limited physical visibility, the count is determined by the number of virtual accelerator devices assigned and visible to each worker node.

When is this SKU not required?

A Red Hat AI Accelerator SKU subscription is not required if the hardware is not being used for AI compute acceleration. Common examples include:

• Networking: DPUs used as smart network interface cards (SmartNICs) for network acceleration only (even if they possess unused Arm cores).
• Graphics and streaming: GPUs dedicated to rendering graphics—that is, drawing pixels on a screen—rather than AI acceleration.
• Red Hat AI Enterprise: Includes unlimited accelerator entitlements per subscribed node.
• Red Hat Enterprise Linux AI and Red Hat AI Inference: These products are already priced and entitled based on individual accelerator counts.

Red Hat AI Enterprise

For Red Hat AI Enterprise, individual CPU cores, sockets, or AI accelerators do not need to be counted.

• Core density: There are no price increases or limitations based on CPU core counts within a subscribed node.
• Hardware accelerators: There is no requirement to purchase separate standalone SKUs for GPUs or TPUs. Red Hat AI Enterprise provides a unified entitlement for unlimited hardware accelerators per node.
• Control plane and infrastructure: As with other self-managed OpenShift offerings, the control plane and infrastructure nodes that manage the platform are included and do not count toward your organization’s subscription total.
• Virtualization: When a physical server is fully dedicated to AI workloads, a single bare-metal Red Hat AI Enterprise subscription covers unlimited Red Hat AI Enterprise VMs running on that physical host.

Red Hat OpenShift AI

Because OpenShift AI is a layered product, the underlying Red Hat OpenShift subscription logic is relied on for any self-managed components.

• Control plane and infrastructure: Entitlements for the Red Hat OpenShift control plane and infrastructure nodes are covered by the base Red Hat OpenShift subscription. Your organization does not need to count them for OpenShift AI.
• Partial cluster coverage: Unlike some platform licenses, OpenShift AI does not require full cluster coverage. Your organization will need to purchase subscriptions for the specific worker nodes where AI workloads or OpenShift AI components are actively running.

OpenShift Platform Plus includes advanced layered products that manage, protect, and scale your organization’s IT environment across multiple clusters and cloud environments. These components include:

• Red Hat Advanced Cluster Management for Kubernetes
• Red Hat Advanced Cluster Security for Kubernetes
• Red Hat Quay
• Red Hat OpenShift Data Foundation Essentials (Note: OpenShift Data Foundation Essentials is available exclusively as part of OpenShift Platform Plus)

Purchasing and integration options with Red Hat AI Enterprise

OpenShift Platform Plus is available in both core-pair and bare-metal socket-pair subscription models (subject to standard platform limitations). If your organization requires these advanced features—either to scale beyond the base OpenShift Container Platform or to integrate them into a Red Hat AI Enterprise environment—your organization has 2 primary purchasing paths:

1. OpenShift Platform Plus add-on subscription: Purchase the specific OpenShift Platform Plus add-on SKU, which bundles these capabilities but excludes the base OpenShift Container Platform entitlement.
2. Individual component SKUs: Purchase individual add-on licenses only for the specific tools needed (Red Hat Advanced Cluster Management, Red Hat Advanced Cluster Security, Quay, or OpenShift Data Foundation Essentials) alongside the base platform or Red Hat AI Enterprise SKU.

Compliance note for OpenShift Platform Plus on Red Hat AI Enterprise nodes

Consistent with the restricted-use entitlement of the underlying platform, any OpenShift Platform Plus components integrated into a Red Hat AI Enterprise environment must be used exclusively for AI workloads. These add-ons do not serve as unrestricted subscriptions for general-purpose or non-AI application use. For full licensing details, please refer to the Red Hat OpenShift subscription guide.

Use the following guidelines to determine the number of subscriptions required for Red Hat AI Enterprise, Red Hat AI Inference, OpenShift AI, or Red Hat Enterprise Linux AI.

Red Hat AI Enterprise sizing guidance:

While the self-managed OpenShift sizing framework provides a helpful foundation, Red Hat AI Enterprise requires a different approach. Unlike standard models based on core-pairs or socket-pairs, Red Hat AI Enterprise uses a per-node pricing model. This simplifies procurement by focusing solely on the total number of dedicated AI compute nodes—whether physical or virtual.

Follow these steps to estimate your organization’s Red Hat AI Enterprise entitlement requirements:

Step 1: Identify AI workload instances

Identify the number of AI pods or instances required for training, tuning, inference, and agent management.

Note: Red Hat AI Enterprise entitlements are strictly limited to AI-related workloads; standalone Red Hat OpenShift use for non-AI applications is not permitted under this SKU.

Step 2: Determine hardware requirements

Calculate the memory and compute power needed to support AI pods. Unlike standard sizing, Red Hat AI Enterprise costs are not tied to CPU core limits. Select high-performance hardware with high core density to maximize throughput without increasing the subscription count.

Step 3: Select AI compute nodes

Choose the physical servers or VMs that will serve as compute nodes. To maximize the value of the Red Hat AI Enterprise entitlement, these should ideally be equipped with hardware accelerators.

Step 4: Calculate total node subscriptions

Your organization’s total Red Hat AI Enterprise requirement is the sum of the nodes dedicated to AI. One subscription covers 1 node, regardless of its CPU or socket count.

• Bare metal: If Red Hat OpenShift is installed directly on a server, count the physical nodes.
• Virtual or cloud: If Red Hat OpenShift is running in a VM (e.g., in a public cloud or shared hypervisor where the underlying hardware is not entitled), count the virtual nodes.

Key inclusion: Red Hat AI Enterprise includes unlimited accelerator entitlements per subscribed node. Your organization can install as many GPUs as the hardware supports without needing additional SKUs.

Step 5: Optimize for virtualization

When deploying on bare-metal hosts fully dedicated to AI workloads, a single Red Hat AI Enterprise physical subscription entitles you to run unlimited Red Hat AI Enterprise VMs on that specific host. This allows you to scale the environment based on the hardware's total capacity rather than individual VM counts.

Red Hat AI Inference sizing guidance

To estimate entitlements for Red Hat AI Inference, shift your IT team’s focus from CPU core-pairs to the number of physical hardware accelerators required to serve your organization’s models.

Follow these steps to estimate your organization’s Red Hat AI Inference requirements:

Step 1: Identify inference workloads

Determine the number of LLM instances your organization intends to serve.

Note: Red Hat AI Inference is a containerized engine purpose-built for gen AI model inference; it does not support model training, tuning, or predictive AI.

Step 2: Identify the underlying platform

Determine which platform will host the Red Hat AI Inference container. Note the following requirements:

• Red Hat OpenShift: Requires a separate Red Hat OpenShift subscription. Distributed inferencing and the llm-d component are only available on Red Hat OpenShift.
• Red Hat Enterprise Linux: Requires a separate Red Hat Enterprise Linux subscription.
• Non-Red Hat platforms: Supported under Red Hat’s third-party policy. Red Hat supports Red Hat AI Inference, but the customer is responsible for platform issues that cannot be reproduced on Red Hat Enterprise Linux or Red Hat OpenShift.

Step 3: Determine accelerator requirements

Calculate the number of accelerators needed based on model size, memory consumption, and request volume. Because large models often span multiple GPUs to fit into memory, sizing is determined by the physical hardware footprint rather than CPU utilization.

Step 4: Choose compute nodes

Select the physical servers or vms that will host the accelerators. While Red Hat AI Inference can run on virtual nodes, the subscription is always calculated based on the count of underlying physical accelerator devices.

Step 5: Calculate total Red Hat AI Inference subscriptions

The total number of Red Hat AI Inference subscriptions equals the total count of physical AI accelerators used for inference.

• Key advantage: The Red Hat AI Inference SKU includes Red Hat AI Accelerator entitlements; your organization does not need to purchase a standalone Red Hat AI Accelerator SKU.

Step 6: Strategy for coverage

Red Hat recommends a full-coverage approach for accelerators that may serve inference workloads. Because inference tasks are often dynamic and ephemeral, entitling every accelerator simplifies management and prepares your organization for fluctuating demand. This removes the burden of tracking specific use and aligns with Red Hat's volume-based pricing value.

OpenShift AI sizing guidance

When sizing entitlements for OpenShift AI, keep in mind its role as a layered add-on. It mirrors the underlying Red Hat OpenShift unit of measure (either core-pairs or bare-metal nodes), providing a consistent framework for scaling.

Follow these steps to estimate OpenShift AI requirements:

Step 1: Identify AI workload instances

Determine the number of AI-specific pods or instances required, including workbenches, training or tuning pipelines, and model deployments.

Note: Unlike Red Hat AI Inference, OpenShift AI supports both generative and predictive AI workloads.

Step 2: Determine aggregate resources

Calculate the total compute (cores/vCPUs) and memory required to support these AI pods. Because OpenShift AI is an add-on, it uses the same sizing units as the OpenShift Container Platform to keep procurement simple.

Step 3: Select target compute nodes

Identify the physical or virtual worker nodes where the AI software will be deployed.

• Partial coverage: Organizations are permitted to subscribe to only a specific subset of nodes within a cluster for OpenShift AI, rather than entitling the entire environment.

Step 4: Calculate Red Hat OpenShift AI add-on subscriptions

An organization may choose a unit of measure (core-pair versus bare-metal) that differs from its underlying Red Hat OpenShift installation, though matching them is often simpler for administration:

• Core-based nodes: Count the aggregate cores or vCPUs across all nodes running AI workloads. Divide total cores by 2 (or vCPUs by 4) to determine the subscription count.
• Bare-metal nodes: Count the physical servers dedicated to AI. One subscription covers 1 server, regardless of CPU socket or core density.
• Prerequisite check: A corresponding base Red Hat OpenShift subscription (OpenShift Container Platform, OpenShift Kubernetes Environment, or OpenShift Platform Plus) is mandatory for each node receiving the OpenShift AI add-on.

Step 5: Calculate AI Accelerator subscriptions

If the nodes identified in step 3 use GPUs, TPUs, or other hardware accelerators, the Red Hat AI Accelerator SKU must be added.

• Calculation: One subscription per physical accelerator device (e.g., a node with 4 GPUs requires 4 Accelerator SKUs).
• Requirement: This is mandatory because AI Accelerator entitlements are not included in the base OpenShift AI subscription.

Red Hat Enterprise Linux AI sizing guidance

To determine the entitlements for Red Hat Enterprise Linux AI, shift your organization’s focus from CPU core-pairs to counting physical hardware accelerators (GPUs) within a single-server environment.

Red Hat Enterprise Linux AI is a foundation gen AI platform designed for organizations that want to serve LLMs on a dedicated, standalone machine. It is helpful to view Red Hat Enterprise Linux AI as a software appliance: it is delivered as a self-contained, bootable image that includes the Red Hat Enterprise Linux operating environment.

Key advantage: No separate operating system subscription

Unlike other offerings, Red Hat Enterprise Linux AI does not require a separate Red Hat Enterprise Linux subscription. The instance is tailored specifically to provide an isolated foundation for gen AI models on a single machine.

Follow these steps to estimate Red Hat Enterprise Linux AI requirements:

Step 1: Identify AI workload requirements

Verify that your use case involves serving LLMs using tools like vLLM.

Note: Red Hat Enterprise Linux AI is dedicated exclusively to gen AI and does not support predictive AI workloads.

Step 2: Choose the dedicated compute node

Select the physical hardware that will host the Red Hat Enterprise Linux AI image.

• Scaling limit: Red Hat Enterprise Linux AI supports single-machine deployment. If your organization’s workload requires a multinode distributed setup or cluster orchestration, you must transition to Red Hat AI Enterprise or Red Hat OpenShift AI.

Step 3: Calculate subscriptions based on accelerators

The primary unit of measure for Red Hat Enterprise Linux AI is the physical AI accelerator.

• Calculation: Count the total number of physical GPUs (or other accelerators) installed in a dedicated server.
• Entitlement: Purchase a corresponding number of Red Hat Enterprise Linux AI SKUs for each physical device. This is required.

Step 4: Account for included accelerator entitlements

When sizing the IT environment, your organization does not need to purchase the standalone Red Hat AI Accelerator SKU. Because Red Hat Enterprise Linux AI is inherently licensed based on GPU usage, the entitlements for those accelerators are already covered by the base Red Hat Enterprise Linux AI subscription.

Alternative architectures (Arm, IBM Z, and IBM Power)

Note: References to IBM Z in this section also apply to IBM LinuxONE.

Like Red Hat OpenShift, OpenShift AI supports Arm, IBM Z, and IBM Power. These platforms are ideal for organizations using these architectures as their standard for cloud-native applications and microservices.

Subscription models by architecture

• IBM Z and IBM Power: These platforms support only the core-based subscription model. Socket-based bare-metal subscriptions are not available for these architectures.
• Arm clusters: Arm environments follow the same entitlement rules as x86 (supporting both core-based and bare-metal models).

IBM Z: Subcapacity entitlement

For IBM Z, your organization is not required to entitle the entire physical node, but can choose to entitle only the specific capacity used by OpenShift AI. This is known as subcapacity entitlement.

• Subscriptions are required for the subset of cores designated for compute nodes.
• This applies regardless of the partitioning method, including CPU pooling, capping, or separate logical partitions (LPARs).

IBM Power: Core-pair logic

IBM Power uses the core-pair subscription model. Note that the concept of a vCPU does not apply to the Power architecture; entitlements are calculated based on physical core counts.

NVIDIA NVL hardware

NVIDIA NVL72 is a high-density architecture featuring 18 CPU trays and 72 GPUs per rack, designed for extreme AI computational workloads. For Red Hat AI Enterprise subscriptions, sizing is based on Red Hat's definition of a compute node. Although the NVL72 operates as a cohesive unified environment, each individual CPU tray runs a distinct instance of the integrated Red Hat AI Enterprise platform. Consequently, Red Hat requires 1 compute node subscription per CPU tray. For example, an 18-tray NVL72 system requires eighteen 1-node Red Hat AI Enterprise subscriptions.

Example 1: Dedicated gen AI environment on high-density bare metal

Scenario: A gen AI project requires a dedicated environment for training, tuning, and serving LLMs. The data science team uses high-performance hardware for these compute-intensive tasks.

Configuration: 3 high-density bare-metal nodes, each equipped with 128 CPU cores and 8 hardware accelerators.

Solution: Red Hat AI Enterprise provides a simplified, cost-effective procurement path for all-in-one AI capabilities without the burden of high core counts or multiple accelerator SKUs.

Step 1: Identify AI workload instances

• Compute nodes: 3 dedicated nodes.
• Workload type: Gen AI (training, tuning, and vLLM inference).
• Usage restriction: Red Hat AI Enterprise entitlements are exclusively for AI workloads; standard application pods are not permitted on these specific nodes.

Step 2: Determine node requirements

Unlike standard Red Hat OpenShift sizing, Red Hat AI Enterprise pricing is not tied to aggregate core counts. Sizing is determined solely by the number of physical or virtual hardware nodes.

• Total physical servers: 3.

Step 3: Define node configuration

• Hardware profile: 128 cores, 2 sockets, and 8 GPUs per node.
• Deployment type: On-site bare metal.

Step 4: Calculate total subscriptions

The Red Hat AI Enterprise SKU entitles 1 physical node or 1 virtual node. Because Red Hat AI Enterprise has no price increases based on core density, a 128-core server requires only 1 subscription—the same as a 16-core server.

• Calculation: 3 physical servers times 1 subscription per node = 3 Red Hat AI Enterprise subscriptions.

Step 5: Calculate Red Hat AI Accelerator SKU subscriptions

In this model, AI Accelerator entitlements are already included in the base Red Hat AI Enterprise SKU.

• Calculation: There are no limits on the number of GPUs per node; therefore, the 24 total GPUs (8 per node) require 0 additional standalone SKUs.

Result

To support this high-density environment, only 3 Red Hat AI Enterprise subscriptions are required.

Note: In a standard core-based model, 384 cores (128 times 3) would typically require 192 core-pair subscriptions. By using Red Hat AI Enterprise, the organization streamlines its entitlement management to just 3 units while gaining unlimited GPU support.

Example 2: Gen AI model serving on bare-metal infrastructure

Scenario: A generative AI project focused on serving LLMs at scale for internal applications. Low latency and high throughput are the primary requirements.

Configuration: 3 physical bare-metal servers, each equipped with 8 GPUs (24 GPUs total).

Solution: Red Hat AI Inference running on Red Hat OpenShift. This architecture separates the platform layer from the specialized inference engine.

Step 1: Identify AI workload instances

• Workload type: Gen AI model inference (serving LLMs).
• Capacity: Total of 24 physical hardware accelerators across 3 nodes.
• Usage restriction: Red Hat AI Inference is purpose-built for inference; it does not support training or tuning.

Step 2: Identify the underlying platform

Red Hat AI Inference is a containerized engine and must sit on a host platform.

• Selected platform: OpenShift Container Platform.
• Platform entitlements: To cover the 3 physical host servers, 3 subscriptions of OpenShift Container Platform (bare-metal node) are required.

Step 3: Determine accelerator hardware footprint

Red Hat AI Inference pricing is independent of CPU cores or vCPUs. Sizing is determined strictly by the number of physical accelerators.

• Calculation: 3 nodes times 8 GPUs per node = 24 physical GPUs.

Step 4: Calculate total Red Hat AI Inference subscriptions

The Red Hat AI Inference SKU is entitled per physical accelerator device.

• Calculation: 24 physical GPUs times 1 subscription per accelerator = 24 Red Hat AI Inference subscriptions.

Step 5: Calculate Red Hat AI Accelerator SKU subscriptions

Because the Red Hat AI Inference SKU is inherently accelerator-based, the hardware entitlements are already included.

• Requirement: No additional Red Hat AI Accelerator SKUs are necessary.

Result

To support this high-performance inference environment, the organization requires:

• 3 OpenShift Container Platform (bare-metal node) subscriptions.
• 24 Red Hat AI Inference subscriptions.

Sizing tip: This hybrid approach allows the organization to scale Red Hat OpenShift based on the number of physical boxes, while scaling the AI Inference capabilities based on the number of GPUs—providing a precise match between cost and hardware use.

Example 3: Mixed gen AI and predictive AI (virtualized)

Scenario: An internal data science platform running within a corporate datacenter. The environment supports a mix of predictive AI (fraud detection and/or forecasting) and gen AI (internal documentation search).

Configuration: 20 virtual worker nodes on a supported hypervisor. Each VM has 8 vCPUs and 1 physical GPU.

Solution: OpenShift AI. This solution allows the organization to manage both traditional and generative AI lifecycles with Premium 24x7 support.

Step 1: Identify AI workload instances

• Workload type: Mixed (gen AI and predictive).
• Infrastructure: 20 virtual worker nodes.
• Requirements: Access to distributed training pipelines (predictive) and vLLM-based serving (gen AI).

Step 2: Determine aggregate resources

Based on the architectural design:

• vCPUs: 8 vCPUs times 20 nodes = 160 vCPUs.
• Accelerators: 1 physical GPU times 20 nodes = 20 physical GPUs.

Step 3: Define compute node profile

• Deployment: Virtualized on-site (e.g., Red Hat OpenShift Virtualization).
• Node profile: 8 vCPUs with 1 physical GPU passed through to each VM.

Step 4: Calculate total platform subscriptions

Because the environment is virtualized, the core-pair model applies (where 1 unit entitles 2 cores or 4 vCPUs). As a layered add-on, OpenShift AI requires a matching underlying OpenShift entitlement.

• Red Hat OpenShift subscriptions: 160 vCPUs divided 4 = 40 core-pair subscriptions.
• OpenShift AI subscriptions: 160 vCPUs divided 4 = 40 core-pair subscriptions.

Step 5: Calculate Red Hat AI Accelerator SKU subscriptions

Unlike Red Hat AI Enterprise, OpenShift AI does not include accelerator entitlements. A separate SKU is required for each physical accelerator card used.

• Calculation: 20 nodes times 1 GPU per node = 20 Red Hat AI Accelerator SKU subscriptions.

Result

To support this virtualized mixed-workload environment with Premium support, the following are required:

• 40 Red Hat OpenShift Container Platform Premium (core-pair) subscriptions.
• 40 Red Hat OpenShift AI Premium (core-pair) subscriptions.
• 20 Red Hat AI Accelerator (1 Accelerator) subscriptions.

Sizing tip: When calculating vCPUs for core-pair SKUs, remember the "divide by 4" rule. If your nodes had 10 vCPUs instead of 8, round up your organization’s subscription count to ensure each vCPU is covered by an entitlement unit.

Example 4: Mixed AI workloads on third-party virtualization versus OpenShift Virtualization 

Scenario: An enterprise is deploying an AI platform for mixed training and inference workloads across its existing on-site infrastructure. The infrastructure team prefers to use a third-party hypervisor (such as VMware ESXi or Nutanix AHV) to carve up high-density hardware into smaller, distinct virtual nodes for different data science teams.

Configuration: The physical footprint consists of 16 bare-metal servers. Each server contains 2 sockets (48 cores per socket/96 total physical cores) and 4 AI accelerators. Using its third-party hypervisor, the customer configures 2 VMs (virtual nodes) per physical socket. Each VM is assigned 24 cores (48 vCPUs) and 2 AI accelerators.

Solution: Red Hat AI Enterprise. In standard Red Hat OpenShift models on third-party hypervisors, customers must buy 2-core SKUs for each allocated vCPU. Red Hat AI Enterprise instead requires subscriptions strictly based on the number of virtual nodes managed.

Step 1: Identify AI workload instances

• Compute nodes: 2 virtual machines per socket times 2 sockets per server times 16 physical servers = 64 virtual nodes.
• Workload type: Mixed gen AI and predictive AI (distributed training and model serving).
• Use restriction: Red Hat AI Enterprise entitlements are exclusively for AI workloads; general corporate IT applications cannot run on these virtual instances.

Step 2: Determine node requirements

Because Red Hat OpenShift is not running directly on the physical hardware, customers cannot license the environment based on the 16 physical bare-metal servers. Instead, under Red Hat AI Enterprise rules for third-party virtualization, sizing is determined by counting the total number of VMs.

• Total virtual nodes: 64.

Step 3: Define node configuration

• Virtual profile: 48 vCPUs and 2 AI accelerators per virtual node.
• Deployment type: Virtualized via third-party hypervisor.

Step 4: Calculate total subscriptions

The Red Hat AI Enterprise SKU entitles 1 physical node or 1 virtual node. Because the hypervisor abstraction forces the software to see 64 individual virtual host operating systems, 1 subscription is required per VM.

• Calculation: 64 virtual nodes times 1 subscription per node = 64 Red Hat AI Enterprise subscriptions.
• Note: If the infrastructure team changes the configuration later to add more VMs (e.g., splitting the hardware into smaller 6-core instances), customers must purchase additional Red Hat AI Enterprise subscriptions for those new virtual nodes.

Step 5: Calculate AI Accelerator subscriptions

AI Accelerator entitlements are inherently bundled into the Red Hat AI Enterprise SKU regardless of the deployment topology.

• Calculation: The 128 total virtual GPUs (1 per VM) require 0 additional standalone SKUs.

Result

To support this third-party virtualized architecture, 64 Red Hat AI Enterprise subscriptions are required.

Architecture recommendation: move to OpenShift Virtualization

Red Hat highly recommends that customers deploy Red Hat AI Enterprise directly on the bare-metal servers and use OpenShift Virtualization to handle their VM requirements. Running OpenShift Virtualization natively on Red Hat OpenShift bare metal provides a fully integrated virtualization layer.

Instead of counting 64 virtual nodes, customers need to license the 16 physical servers (16 Red Hat AI Enterprise subscriptions). The data science team could spin up, tear down, or reconfigure the virtual nodes and virtual GPUs as they choose without additional subscription costs.

Terms that users need to understand:

• AI accelerator: AI accelerators are hardware computing devices, such as GPUs or ASICs, that are engineered to enhance and expedite AI computations. These devices provide processing capabilities and optimized architectures for AI tasks including machine learning, model training, and inference. These devices can include, but are not limited to, discrete GPUs, TPUs, DPUs, FPGAs, ASICs, and NPUs that are installed as add-ons. Red Hat AI Accelerator SKU subscription is based on the number of physical devices, even if they are virtualized or have many cores. This definition excludes accelerators integrated with the main CPU (like integrated GPU or NPUs). Accelerators are counted when they are used to execute a compute workload. 
• AI compute workload: A workload is considered a compute workload when the primary purpose of the workload is not actively drawing pixels on a user’s screen in near real time or moving data across a network.
• AI use cases: The use of the included OpenShift Container Platform entitlement in Red Hat AI Enterprise is subject to restricted-use language. This entitlement is limited to hosting Red Hat AI Enterprise for the execution of AI use cases only. These include AI model training, tuning, inference, and AI agent deployment, along with management running on OpenShift AI components hosted on OpenShift Container Platform.
• Core-pair: One of the bases for self-managed Red Hat OpenShift and OpenShift AI subscriptions. It is defined as 2 physical CPU cores or 4 vCPUs. For hyperscaler deployments (e.g., Amazon Web Services, Microsoft Azure, or Google Cloud Platform), 4 vCPUs are always considered a single core-pair.
• Bare-metal node (1 physical node): Another basis for self-managed OpenShift and OpenShift AI subscriptions. A physical node is 1 server, regardless of the number of CPU sockets in the server or cores in the CPUs. The product must be installed directly on the bare-metal server.
• Virtual node: A virtual machine (VM) or container running an instance of the software.
• Service-level agreement (SLA): A choice of support service-level agreement is required for Red Hat subscriptions. The 2 choices available are Premium 24x7 and Standard 8x5.

• Red Hat AI Enterprise: To determine the correct quantity for Red Hat AI Enterprise subscriptions, users must count each physical node or virtual node where the integrated AI platform is deployed, as it is priced on a flat-rate basis per node. A significant advantage of this unified SKU is that it includes unlimited AI Accelerator entitlements, meaning there is no need to count individual GPUs or other physical accelerator devices installed within those entitled nodes. Furthermore, there are no limitations or price increases based on the number of CPU cores or sockets in a node. The included OpenShift Container Platform entitlement is restricted to AI use cases and explicitly may not be used for standalone, non-AI workloads. In scenarios where a physical server is fully dedicated to running AI virtual machines, a single Red Hat AI Enterprise subscription for that physical node can cover unlimited virtual machines running Red Hat AI Enterprise on top.
• Red Hat AI Inference: To determine the correct quantity for Red Hat AI Inference subscriptions, users must count the total number of physical hardware accelerators (such as GPUs or TPUs) that will be used to serve their models. Unlike typical server software, AI Inference is entitled purely per accelerator, meaning that the number of CPU cores or sockets in the node does not affect the subscription count. This applies regardless of whether the product is deployed on a physical or virtual node, requiring 1 subscription for each physical device being used. Crucially, the AI Accelerator entitlement is already included within the AI Inference SKU, so there is no need to purchase the standalone Red Hat AI Accelerator SKU separately for these specific workloads. This product is specifically designed for customers whose primary goal is serving LLMs efficiently at scale across Red Hat or third-party platforms. AI Inference runs on Red Hat Enterprise Linux, Red Hat OpenShift, non-Red Hat Linux platforms and non-Red Hat Kubernetes-based platforms. 
• Red Hat OpenShift AI: To determine the correct subscription count for OpenShift AI, users must count the entitlements for the specific OpenShift worker nodes where OpenShift AI components and AI workloads are running. For self-managed deployments, this follows the standard OpenShift units of measure: either core-pairs (2 physical cores or 4 vCPUs) or bare-metal nodes. A key distinction of Red Hat OpenShift is that it does not require full cluster coverage; customers only need to subscribe to the specific subset of compute nodes used by the AI platform. It is mandatory that these nodes also have a corresponding Red Hat OpenShift subscription, as OpenShift AI is a layered product that runs on top of the container platform. OpenShift AI subscriptions do not include accelerator entitlements by default; therefore, users must count and purchase a separate Red Hat AI Accelerator SKU for each physical hardware device (such as a GPU or TPU) installed in the worker nodes executing AI compute workloads.
• Red Hat Enterprise Linux AI: To determine the correct subscription quantity for Red Hat Enterprise Linux AI, users must count the total number of physical hardware accelerators (such as GPUs) installed in the host machine. This product is entitled purely per accelerator, regardless of whether it is deployed on a physical or virtual node, and it is explicitly not entitled based on CPU cores or sockets. Because Red Hat Enterprise Linux AI is delivered as a self-contained software appliance, it includes the necessary operating environment and does not require a separate Red Hat Enterprise Linux subscription. And, because the product's pricing model is based on accelerators, customers need not purchase the standalone Red Hat AI Accelerator SKU separately for these workloads. Red Hat Enterprise Linux AI is restricted to single-machine deployments for tuning and serving LLMs and cannot be scaled across a distributed cluster.