Design and delivery of the factory, through to commissioning.

What we mean by an AI factory

NVIDIA describes an AI factory as specialised computing infrastructure that manages the AI life cycle, from data ingestion through training, fine-tuning and high-volume inference, with intelligence as the product. The public definition is on the NVIDIA glossary. The same idea appears in NVIDIA’s account of five layers that have to work as one system: energy, chips, infrastructure, models and applications.

We use that language because it matches the work. A factory here is a designed production system: a workload, a data path, a compute domain, a fabric, a facility and an operating model. The output is usable inference and training capacity, under constraints the organisation can name.

Strategy and design

The design starts from questions that are expensive to reverse later.

• Is the workload training, fine-tuning, inference, or a mix, and what latency and throughput matter?
• Where must the data and the inference stay, and which systems of record are in scope?
• What power, cooling, weight and access does the building actually have?
• Who will run the platform after handover, and with which tools?

From those answers we write a reference architecture and a capacity view. That covers the GPU domain, the fabric (InfiniBand or Ethernet, chosen for the workload), the storage performance profile, and the facility interfaces. Vendor options come after that picture is stable.

Build and integration

Build covers the physical and logical assembly of the factory.

• Accelerated compute, from a first rack through to large NVLink and multi-rack domains.
• High performance fabric and the storage path the workload will actually saturate.
• Power distribution, liquid or air heat rejection, and the containment the room requires.
• Integration with identity, networks and the enterprise systems the factory must reach.

We treat procurement as part of design, with a vendor-neutral options review. We do not publish partnership badges. If a platform is named on this site, it is because it is a normal part of the architecture, and the choice still belongs to the workload.

Operations

Day-two work is specified before the racks arrive. Scheduling (Slurm, Kubernetes, or the platform you will actually staff), observability, firmware and driver currency, capacity headroom, and a maintenance window that the facility can support.

Commissioning is a written test, not a handover meeting. It covers power draw against the design, thermal behaviour, fabric health, storage throughput, and the failure cases you agree in advance. The operating notes and the acceptance record go with the system.

1. 01

   Assess

   Workload, data, site, and the constraints that bound the design.

2. 02

   Design

   Reference architecture, capacity, fabric, storage, and facility interfaces.

3. 03

   Build

   Compute, network, storage, power and cooling, integrated as one system.

4. 04

   Integrate

   Identity, enterprise systems, models and agents, and the paths data may take.

5. 05

   Operate

   Commissioning, acceptance, observability, handover, and the run model.

What you leave with

A typical AI factory engagement is structured to produce:

• A workload and constraint brief
• A reference architecture, capacity view and site readiness note
• A vendor-neutral options paper and, when you proceed, support through procurement and integration
• Commissioning tests, an acceptance record and an operating handover

Power and cooling realities are covered in the briefing Power, cooling and the building in front of you. The planning sequence is in What an AI factory is, and how to plan one.