NVIDIA AI Factory Explained: Reference Architecture vs Reality

NVIDIA uses the term "NVIDIA AI factory" to describe several different designs, products and operating layers. There are at least three layers to distinguish: the Enterprise AI Factory Validated Design, NVIDIA Enterprise Reference Architectures implemented by partners, and NVIDIA DSX for designing and operating larger AI factories.

An AI factory also needs compute, networking, storage, power, cooling and capacity planning infrastructure, as described in the NVIDIA Enterprise AI Factory documentation.

What's the difference between the layers of the NVIDIA AI Factory concept?

These layers serve different purposes. The Enterprise AI Factory design specifies a recommended infrastructure and software stack. NVIDIA Enterprise Reference Architectures provide configurations that OEMs can implement. NVIDIA DSX addresses factory-level design and operations, including infrastructure lifecycle management, power management and IT/OT integration.

What's included in the NVIDIA Enterprise AI Factory?

The Enterprise AI Factory Design Guide includes designs for accelerated computing platforms using Blackwell-based systems, including RTX PRO Blackwell Server Edition and HGX B200/B300 systems. The design also includes networking and the NVIDIA AI Enterprise software stack.

NVIDIA AI Enterprise provides monthly security updates for high and critical vulnerabilities on production branches, which currently have a nine-month support window, and quarterly updates on long-term support branches, which currently have three years of support. Enterprise production support is backed by a service-level agreement (SLA).

NVIDIA has also published information about Northrop Grumman's AI factory, which uses NVIDIA infrastructure and partners to deploy AI services across a workforce of nearly 100,000 employees.

What's included in an NVIDIA DGX SuperPOD?

NVIDIA describes DGX SuperPOD as a "full-stack data center platform" that includes compute, storage, networking, software and infrastructure management.

NVIDIA Enterprise Reference Architectures offer another procurement route. These can be implemented by systems partners using NVIDIA-Certified Systems. Hardware support is generally provided by the OEM or channel partner supplying the systems, while NVIDIA separately supports its software stack.

The distinction is which organization supplies, integrates and supports each part of the infrastructure.

What do NVIDIA partners provide in terms of AI factories?

Systems vendors including Dell and Supermicro offer NVIDIA-endorsed AI factory designs, while companies such as Red Hat provide co-engineered software platforms based on Red Hat AI Enterprise and NVIDIA AI Enterprise.

These offerings differ in scope. An OEM hardware implementation and a co-engineered software platform do not provide the same infrastructure or support responsibilities.

What infrastructure is required for an NVIDIA AI Factory?

NVIDIA's "Enterprise AI Factory Considerations" identify several infrastructure requirements:

• Power: Very high power-density racks, with multiple kilowatts per node and tens of kilowatts per rack, substantially exceeding typical CPU-based rack densities. This is one of the key differences between an AI factory and a traditional data center.
• Cooling: Advanced cooling, including liquid or high-efficiency air cooling, to dissipate heat from high-density accelerator systems operating at high utilization for extended periods.
• Storage: Parallel, scale-out storage with tiered capacity, including fast NVMe storage for hot data, such as training datasets and models, and object storage for colder data.
• Network fabric: Low-latency, high-bandwidth fabrics, including 100–400 GbE, RDMA-capable networking and dedicated GPU interconnects, to handle high-volume east-west traffic.
• Capacity planning: Scalable infrastructure with dynamic resource allocation, particularly for bursty agentic workloads where tokens per second and queries per second are important capacity measures.

These requirements are not automatically covered simply because a proposal includes NVIDIA GPUs and networking. Unless explicitly included in the contract, the buyer still needs an owner responsible for facility power delivery, cooling, storage integration and capacity planning.

These requirements also affect the build sequence for power, compute and network infrastructure.

Why the capacity math has shifted

NVIDIA identifies three AI scaling laws that help explain the growth in compute demand:

• Pre-training scaling: Training foundation models requires increasingly large datasets, models and computing systems. NVIDIA reports that pre-training compute requirements increased by roughly 50 million times over five years.
• Post-training scaling: Creating an ecosystem of customized models from a foundation model can require substantial additional compute. NVIDIA estimates that developing such an ecosystem could consume approximately 30x the compute needed to pre-train the original model.
• Test-time scaling: Complex reasoning workloads can require multiple inference passes before producing an answer. NVIDIA reports that challenging reasoning queries can consume 100x or more compute than a traditional single-pass inference request.

These figures describe NVIDIA's scaling observations and estimates, not fixed compute requirements for every deployment. They show why AI factory capacity planning must account for workload types and inference behavior rather than GPU count alone.

DSX: Software and infrastructure tooling for large AI factories

NVIDIA DSX is more than an operating system. It is a collection of reference designs, application programming interfaces (APIs), software libraries and technologies that AI factory architects and operators can use to design, simulate, build and operate AI factories.

Five components are particularly relevant to AI factory operations:

• DSX Exchange: The integration layer between IT and operational technology (OT) systems. It provides an event bus for exchanging power, cooling, grid and other operational information between building management systems and AI factory software.
• DSX MaxLPS: A framework for designing and operating AI factories at high performance per watt, combining site design, NVIDIA Dynamic Power Software and workload-specific performance optimizations.
• DSX Flex: Defines power targets, timing and reported operating states, allowing AI factory systems to coordinate responses to grid requirements or facility power constraints.
• NVIDIA Infra Controller: Provides API-driven bare-metal infrastructure lifecycle management. It supports multi-tenant GPU infrastructure, including hardware-based tenant isolation using NVIDIA BlueField Data Processing Units (DPUs).
• NVSentinel: Provides Kubernetes-native GPU fault detection and automated remediation as part of NVIDIA's DSX operating stack.

These five components do not represent the entirety of DSX. The broader platform also includes simulation, accelerated Kubernetes, observability, security, resilience and infrastructure management technologies.

How DSX MaxLPS increases capacity within a fixed power budget

NVIDIA claims that DSX MaxLPS can enable up to 40% more GPU capacity within the same power envelope by reallocating unused power headroom rather than relying solely on static power provisioning.

In a September 2026 NVIDIA and Nscale evaluation, MaxLPS increased managed GPU capacity from 140 to 192 GPUs, a 37.1% increase, while maintaining the same 264.4 kW provisioned power budget. Aggregate throughput increased by 49.2%, although P99 time to first token also increased by 17%.

The achievable capacity gain depends on workload mix, hardware configurations, available power headroom, redundancy requirements and latency constraints. The results should therefore be validated against the workloads and operating limits of the proposed facility.

At larger scales, the distinction between facility operations and AI infrastructure operations becomes less clear. Power and cooling telemetry need to be available to AI factory management software so that compute resources can be operated within facility constraints.

Looking to Build Your Own AI Factory?

At Zettabyte, we offer two approaches to building and operating AI factory infrastructure:

• Provider-run dedicated infrastructure: Our Enterprise AI Data Centers service provides dedicated compute, storage and networking infrastructure that companies can use to run and manage their AI workloads.
• Build-Operate-Train-and-Transfer (BOTT): Our BOTT model for Sovereign AI Data Centers enables organizations to develop sovereign AI infrastructure with end-to-end ownership of hardware, software, data and supply chain.