Infrastructure

NCP

ActivePublished: 5 August 2026Updated: 5 August 2026Published

Key innovation

An NVIDIA partner program in which cloud providers build and operate GPU infrastructure to NVIDIA's validated reference architecture - standardizing the design of AI data centers (compute, networking, storage, security).

How it works

1) A cloud provider joins the NVIDIA Cloud Partner program. 2) It designs and builds a data center to the NCP reference architecture (GPU clusters, high-bandwidth networking, storage, security). 3) The deployment is validated against NVIDIA's specifications, ensuring performance and interoperability. 4) The partner operates a GPU cloud service for customers (AI training/inference), using technical support, go-to-market and priority GPU allocation.

Problem solved

Building large GPU data centers for AI is complex, costly and error-prone. NCP provides a proven blueprint and NVIDIA support, reducing risk, shortening deployment time and ensuring interoperability and access to GPUs.

Components

Reference architectureDesign standard

A validated blueprint for an AI data center (compute, networking, storage, security).

GPU clustersCompute

Standardized configurations with the latest NVIDIA GPUs (from ~128 to 16,000+ nodes).

Official

Networking and storageConnectivity and data

High-bandwidth networking (e.g. InfiniBand/Spectrum-X) and storage per spec.

Official

NVIDIA support and allocationPartner program

Technical support, go-to-market and priority GPU allocation for partners.

Official

Implementation

Reference implementations

Cisco - NCP-compliant reference architecture networking

Cisco

Implementation pitfalls

Vendor lock-in (NVIDIA)Medium

An architecture built on NVIDIA GPUs and stack increases dependence on a single vendor.

Fix:Deliberate cost/exit planning, hybrid approaches where feasible.

Capital, power and GPU supplyHigh

Building clusters requires huge capital, power and contracted GPUs (constrained supply).

Fix:Infrastructure financing, proprietary power contracting, priority allocation via NCP.

Compliance and validationLow

Maintaining compliance with an evolving reference architecture requires continual updates.

Fix:Working with NVIDIA, regular architecture reviews.

Evolution

2024

Launch of the NCP program and reference architecture

Inflection point

NVIDIA formalizes the Cloud Partner program and publishes a reference architecture for AI data centers, standardizing GPU cloud build-outs.

2025

Neocloud and partner ecosystem

Growth of the 'neocloud' and partner ecosystem (e.g. Cisco with NCP-compliant networking) building 'AI factories' to NVIDIA's architecture.

Hyperparameters (configurable axes)

Cluster scaleHigh

From ~128 to over 16,000 nodes per the reference architecture.

Network fabricHigh

InfiniBand or Ethernet (e.g. Spectrum-X) to interconnect GPUs.

GPU generationMedium

NVIDIA accelerators used (e.g. Hopper, Blackwell, Rubin).

Computational complexity

Time complexity: Nie dotyczy (program/architektura referencyjna, nie algorytm). Space complexity: Nie dotyczy.

Compute bottleneck

GPU supply, power and networking

A partner cloud's scale is determined by GPU availability (NVIDIA allocation), the data center's power and scale-up/scale-out networking bandwidth - not a single algorithm.

Depends on

Alokacja GPUMoc i chlodzenie