> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/ai-enterprise/ai-factory/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/ai-enterprise/ai-factory/_mcp/server.

# Enterprise Hardware Reference Architecture Foundation

> The boundary between the software reference design and authoritative NVIDIA Enterprise Hardware Reference Architectures.

The NVIDIA Enterprise AI Factory Reference Design is built on [NVIDIA Enterprise Hardware Reference Architectures](https://www.nvidia.com/en-us/technologies/enterprise-reference-architecture/). Use that portfolio to identify the applicable architecture family. The linked technical publications remain the source of truth for its infrastructure design points.

After you select an architecture family, use [OEM AI Factory Offerings](/ai-enterprise/ai-factory/ecosystem-integrations/oem-ai-factory-offerings) to discover publisher-owned full-stack options. Confirm orderable configurations, lifecycle, compatibility, and support directly with the applicable OEM.

Explore [storage and data infrastructure ecosystem solutions](/ai-enterprise/ai-factory/ecosystem-integrations/solution-catalog?layer=hardware-foundation).

## What the Software Design Consumes

The software design consumes the following infrastructure capabilities:

| Capability                        | Interface Assumption                                                                                                                                                                                                                                                                     |
| --------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Accelerated compute               | Sufficient GPU, CPU, memory, local storage, interconnect, and isolation for the workload envelope. Refer to [Enterprise Hardware Reference Architectures](https://docs.nvidia.com/enterprise-reference-architectures/index.html).                                                        |
| Scale-up and scale-out networking | Bandwidth, latency, loss, topology, segmentation, and RDMA capabilities required by the selected workloads and services. Refer to [NVIDIA networking documentation](https://docs.nvidia.com/networking/).                                                                                |
| Separated infrastructure services | Supported NIC or DPU resources for offloaded networking, storage, isolation, monitoring, and containment when the approved design requires a control domain independent of the host. Refer to [NVIDIA DOCA Platform Framework](https://docs.nvidia.com/networking/display/dpf25071).     |
| Data infrastructure               | Capacity, throughput, metadata behavior, protection, locality, retention, and recovery required by datasets, models, indexes, checkpoints, and telemetry. Refer to [Enterprise Hardware Reference Architectures](https://docs.nvidia.com/enterprise-reference-architectures/index.html). |
| Confidential computing            | Supported hardware and firmware roots of trust, measured execution, attestation evidence, protection in use, and separable infrastructure control domains required by the approved threat model. Refer to [NVIDIA Trusted Computing Solutions](https://docs.nvidia.com/nvtrust/).        |
| Facilities                        | Power, cooling, rack, cabling, environmental, physical-security, serviceability, and resilience envelope. Refer to [Enterprise Hardware Reference Architectures](https://docs.nvidia.com/enterprise-reference-architectures/index.html).                                                 |
| Lifecycle                         | Current firmware, driver, platform support, diagnostics, replacement, upgrade, and escalation paths. Refer to [NVIDIA AI Enterprise support matrix](https://docs.nvidia.com/ai-enterprise/support-matrix/latest/index.html).                                                             |

## What Remains Outside This Site

Refer to authoritative hardware and OEM sources for the following details:

* Server, GPU, NIC, DPU, switch, and storage bills of materials.
* Physical and logical fabric topology.
* Scale-unit composition and maximum scale.
* Rack, power, cooling, cabling, and facility design.
* Infrastructure endorsement or certification criteria.
* Time-sensitive product compatibility and support matrices.

## Confidential-Computing Boundary

The hardware foundation supplies trust capabilities and evidence. It does not decide which workloads require them or accept the resulting risk. [Governance and Assurance](/ai-enterprise/ai-factory/core-reference-design/security-and-governance) defines the approved policy and evidence requirements. Platform Identity and Trust Services appraise evidence and condition access, credential, or key release. Use the [NVIDIA Trusted Computing Solutions library](https://docs.nvidia.com/nvtrust/) to confirm supported hardware, firmware, attestation, and confidential-computing combinations.

For published confidential inference deployment patterns, refer to the following architectures:

* [Deploying Proprietary Models Securely with NVIDIA Confidential Computing on Self-Hosted Virtual Machines](https://docs.nvidia.com/enterprise-reference-architectures/deploying-proprietary-models-confidential-compute-self-hosted-vms/latest/index.html).
* [Deploying Proprietary Models Securely with NVIDIA Confidential Computing on Self-Hosted Kubernetes](https://docs.nvidia.com/enterprise-reference-architectures/deploying-proprietary-models-confidential-compute-self-hosted-kubernetes/latest/index.html).

## DPU and Separated-Control Boundary

The hardware foundation identifies supported DPU and network infrastructure. [Isolation and Runtime Services](/ai-enterprise/ai-factory/core-reference-design/cloud-native-platform#isolation-and-runtime-security-boundary) defines when an implementation uses a separately managed enforcement, monitoring, or containment domain. [Accelerator Operators](/ai-enterprise/ai-factory/core-reference-design/cloud-native-platform#nvidia-operators) manage supported platform software; they do not replace runtime policy or accept workload risk.

Use the [NVIDIA BlueField platform](https://www.nvidia.com/en-us/networking/products/data-processing-unit/) and [NVIDIA DOCA](https://www.nvidia.com/en-us/networking/products/software/doca/) pages for current product capabilities. Use the selected Enterprise Hardware Reference Architecture and OEM guidance for supported system placement, topology, lifecycle, and configuration.

## Selection Workflow

#### Use the Workload Envelope

Carry demand, data, availability, security, growth, and facility constraints from the core design.

#### Select the Architecture Family

Evaluate the current Enterprise Reference Architecture patterns and their supported scale points.

#### Confirm Implementation Support

Check current system, platform, network, storage, and software support with authoritative owners.

#### Return the Infrastructure Contract

Record the capacity, topology, service, lifecycle, and availability assumptions consumed by the software layers.

## Decision Inputs and Authoritative Owners

Use the following authorities to make and record infrastructure decisions:

| Decision Input                                                                                   | Use This Authority                                                                                                                                                          | Required Handoff Output                                                                    |
| ------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| Workload type, service objectives, data paths, isolation, and growth                             | [Workload requirements record](/ai-enterprise/ai-factory/core-reference-design/workloads) and applied Capability Reference Designs or Industry Profile                                               | Approved requirement set and unresolved risks                                              |
| GPU, system, network, storage, topology, scale, and facilities                                   | [NVIDIA Enterprise Reference Architecture portfolio](https://www.nvidia.com/en-us/technologies/enterprise-reference-architecture/) and its linked technical publications    | Selected reference architecture family, scale assumptions, and infrastructure design owner |
| Operating system, hypervisor or cloud, Kubernetes, drivers, operators, and release compatibility | [NVIDIA AI Enterprise Infrastructure Support Matrix](https://docs.nvidia.com/ai-enterprise/support-matrix/latest/index.html)                                                | Release-specific compatibility evidence and exception list                                 |
| OEM platform, configuration, lifecycle, and commercial support                                   | [OEM AI Factory Offerings](/ai-enterprise/ai-factory/ecosystem-integrations/oem-ai-factory-offerings) and OEM configuration and support documentation aligned to the selected reference architecture | Named OEM offer, configuration owner, support path, and assumptions                        |
| Ecosystem software candidates                                                                    | [Ecosystem Solution Catalog](/ai-enterprise/ai-factory/ecosystem-integrations/solution-catalog) and publisher documentation                                                                          | Candidate list and publisher-owned support evidence                                        |

## Handoff Milestones

Complete the following handoffs before you approve the deployment design:

1. **NVIDIA architecture guidance:** Agree on the reference architecture family, capability envelope, and support-matrix constraints.
2. **OEM system configuration:** Translate those constraints into an OEM orderable configuration and lifecycle plan.
3. **Customer design approval:** Confirm facility fit, operational ownership, support paths, exceptions, and sign-off evidence in the deployment decision record.

> **Info**
>
> This page intentionally does not restate hardware architecture details. Refer to the current Enterprise Reference Architecture publication for those decisions.