What is an AI factory in artificial intelligence?
NVIDIA defines an AI factory as “a specialized computing infrastructure designed to create value from data by managing the entire AI life cycle, from data ingestion to training, fine-tuning, and high-volume AI inference. The primary product is intelligence, measured by token throughput.”
The comparison with a manufacturing plant is deliberate. An AI factory transforms raw data and electricity into AI models and tokens, turning data into actionable output at industrial scale. Jensen Huang put it plainly at Computex 2025: AI data centers “are in fact AI factories. You apply energy to it, and it produces something incredibly valuable, and these things are called tokens.”
NVIDIA also notes that an AI factory need not be an on-premises data center. It can be an AI-dedicated cloud, a hybrid deployment or a telecom edge site. What makes it an AI factory is the purpose and the metric, not the building type: output is judged by tokens produced and, increasingly, by cost per token and tokens per watt.
Sources: NVIDIA glossary: AI factory; NVIDIA, Computex 2025 keynote; NVIDIA, How AI Factories Generate Revenue. Reviewed November 2, 2026.
AI factory vs traditional data centers
Is an AI factory the same as a data center? Physically it is a kind of data center, and much of it uses the same AI infrastructure found in high performance computing, but it is designed around different assumptions.
| Aspect | Traditional data centers | AI factory |
|---|---|---|
| Main job | Store and process data for many applications. | Produce AI output: trained models and inference tokens. |
| Key metric | Uptime, space and power cost. | Token throughput, cost per token, tokens per watt. |
| Compute | Mostly CPU servers, mixed tenants. | Dense GPU systems operating as large clusters. |
| Rack power | Often low enough for air cooling. | Rack-scale systems over 100 kW, liquid cooled. |
| Network | General-purpose Ethernet. | High-bandwidth, low-latency fabrics between GPUs. |
Because the workload defines the facility, the AI data center and the AI factory are best treated as the same physical infrastructure viewed from two angles: the building and its power, and the production system running inside it.
How AI factory work flows across the AI lifecycle
AI factories operate the full AI lifecycle as a continuous pipeline rather than a set of isolated AI experiments. The main AI factory functions are:
- Data intake and preparation. Raw data, including unstructured data, sensor data and external data from existing systems, enters through data pipelines. Data management, quality control, data preprocessing and feature engineering turn large datasets into training-ready inputs. Transforming data at this stage sets the data readiness of everything downstream.
- Model training. GPU clusters on AI servers run machine learning algorithms to train new artificial intelligence models or continue training existing AI models. Model training is the most compute-intensive stage for many AI workloads.
- Fine-tuning and evaluation. Data scientists refine models on domain data and test them before release.
- Inference. Trained models are deployed into production environments, serving real time inference for generative AI applications, retrieval augmented generation and other AI workflows.
- Feedback. Production results and new data flow back through the data pipelines to retrain and refine models, so the model lifecycle keeps improving.
The point of the AI factory model is that these stages share one platform. AI factories enable teams to move AI models from experiment to production without rebuilding infrastructure for each project.
NVIDIA’s materials emphasize inference performance and energy efficiency. For infrastructure, that shifts design attention toward sustained throughput and tokens per watt, not just peak training speed.
Core components of AI infrastructure in an AI factory
Building an AI factory requires every layer to work together. A practical way to group the core components:
- Compute: AI servers built on GPU platforms such as NVIDIA GB200 and GB300 NVL72, which operate as rack-scale systems for training and inference of AI models.
- Networking: high-bandwidth fabrics within and between racks; network performance often limits cluster efficiency.
- Data layer and storage: fast storage and data platforms feeding training and inference.
- Software: orchestration, scheduling, AI lifecycle management and inference serving.
- Physical infrastructure: power delivery, liquid cooling, heat rejection, monitoring and the site itself.
The physical layer is where many AI initiatives slip. GPU systems can be ordered faster than powered, liquid-cooled space can be built to house them.
NVIDIA AI factory designs: Omniverse DSX and Vera Rubin DSX
Beyond GPU systems, NVIDIA publishes reference designs for whole AI factories.
- Omniverse DSX Blueprint (October 2025). Introduced at GTC Washington, D.C., as an open blueprint for designing and operating gigawatt-scale AI factories, validated at the NVIDIA AI Factory Research Center at Digital Realty in Manassas, Virginia. Named partners include Eaton, GE Vernova, Hitachi, Schneider Electric, Siemens Energy, Trane and Vertiv.
- Vera Rubin DSX AI Factory reference design (March 2026). Released at GTC 2026 with compute, Spectrum-X networking and storage designs plus power, cooling and controls best practices. The Omniverse DSX Blueprint became generally available at the same time for physically accurate AI factory digital twins. Vertiv announced a prefabricated “OneCore Rubin DSX” solution built with the blueprint.
- DSX framework. NVIDIA describes DSX as co-designed infrastructure that integrates compute, power, cooling, networking and operations to maximize tokens per watt.
For site developers the practical value is a shared design language: a digital twin lets the facility, power and cooling be simulated against a specific GPU platform before construction. See our NVIDIA DSX POD guide for the pod-level view.
Sources: NVIDIA, October 28, 2025; NVIDIA, March 16, 2026; NVIDIA DSX; NVIDIA AI factories. Reviewed November 2, 2026.
Power and cooling: what an AI factory requires per rack
Rack-scale GPU systems drive the physical design. Published figures from NVIDIA and its partners:
- GB200 NVL72: NVIDIA’s MGX platform supports up to 120 kW per rack for Blackwell GB200 NVL72.
- GB300 NVL72: NVIDIA’s Enterprise Reference Architecture describes a liquid-cooled rack with eight 33 kW power shelves, with the full rack requiring up to 142 kW. Vertiv published a 142 kW per rack cooling and power reference architecture for GB300 NVL72. In NVIDIA’s DGX SuperPOD reference architecture, each GB300 scalable unit requires 1.2 MW.
- Next generations: NVIDIA has described an 800 VDC power architecture for future AI factories, designed for racks from 100 kW to over 1 MW, with full-scale production planned alongside its Kyber rack-scale systems in 2027.
We do not publish per-rack figures for Vera Rubin systems here; plan with the platform documentation current at the time of order.
At these densities, air cooling alone is not enough. Direct-to-chip liquid cooling, coolant distribution units and site heat rejection become part of the AI factory’s core design, not an add-on.
Sources: NVIDIA, MGX technical blog; NVIDIA Enterprise Reference Architecture, GB300 NVL72; Vertiv, June 11, 2025; NVIDIA DGX SuperPOD GB300 reference architecture; NVIDIA, 800 VDC architecture. Reviewed November 2, 2026.
Examples of AI factory building
The term is now used for projects of very different sizes:
- Lambda announced the conversion of an unoccupied, 2009-built facility in Kansas City into an AI factory for more than 10,000 NVIDIA GPUs.
- Firmus describes Project Southgate in Tasmania as a campus of modular, 100% liquid-cooled AI factories, with a 90 MW first stage.
- Nebius reported a 1.2 GW AI factory site in Pennsylvania among its new owned sites in its Q1 2026 shareholder letter.
These are industry examples, not QIT projects. They show that an AI factory can be a converted building, a modular campus or a gigawatt greenfield site.
Sources: Lambda, Kansas City; Firmus, Project Southgate; Nebius Q1 2026 shareholder letter. Reviewed November 2, 2026.
Modular containers as AI factory building blocks
NVIDIA’s own approach leans on modularity: its MGX blog describes about 80% factory pre-integration cutting deployment from 12 months to under 90 days at the rack level, and the DSX designs treat the facility as repeatable units. The same logic applies one level up, to the enclosure.
QIT develops liquid-cooled modular GPU data centers planned around roughly 1 MW of IT load per block, phased into 5–10 MW sites, mainly in Texas. Used as AI factory building blocks, modules offer:
- A repeatable unit. A defined block of power, cooling and racks that can be matched to a GPU platform and replicated.
- Phasing. Capacity can follow demand and confirmed power, one block at a time.
- Liquid cooling from the start. Coolant distribution is designed into the module for rack-scale systems; see the liquid-cooled GPU container guide.
- Faster sites. Existing powered properties can host blocks without a new building. QIT installs a factory-integrated container on a prepared site in 15 hours.
An illustrative calculation: at the 142 kW per rack published for GB300 NVL72, a 1 MW IT block holds about seven such racks before any design margin. A QIT container delivers 1 MW of IT capacity in a 20 ft footprint, so each container is one such block; the rack mix depends on the GPU platform.
Benefits of an AI factory for AI solutions
For organizations running AI at scale, AI factories provide practical benefits over building separate infrastructure for each of their AI solutions:
- Operational efficiency: one shared platform for training, fine-tuning and inference instead of isolated clusters.
- A clear metric: cost per token and tokens per watt make infrastructure choices comparable.
- Faster AI development: production-ready AI moves from experiment to deployment on the same stack.
- Control of sensitive data: an AI factory on dedicated infrastructure can keep customer data, patient data or other regulated data within defined boundaries.
Use cases range from drug discovery and fraud detection to predictive maintenance, customer behavior analysis and supply chain optimization. The common thread is a steady, high-volume demand for GPU compute for AI workloads. For infrastructure developers, building AI factories means providing that compute with predictable power and cooling.
Plan an AI factory site
If you are planning AI factory capacity in Texas, QIT can discuss a phased modular design around your site, power and GPU platform.
Read the NVIDIA DSX POD guide, the modular data center guide, capacity for GPU clouds and the behind-the-meter power guide. Then contact QIT.
Prepared by QIT · Updated November 2, 2026
Explore all field notes