There is no standards body that defines an AI factory. There is a vendor that coined the phrase, a European funding programme that uses the same two words for something else, and a large number of facilities that have adopted the label because it reads better than colocation. If you are buying compute in Australia, it is worth separating the three.
Where the term comes from
NVIDIA is the origin. Jensen Huang used the framing from GTC 2024 onward, and NVIDIA’s glossary now defines an AI factory as specialised computing infrastructure designed for the entire AI life cycle, from data ingestion and preparation through training, fine-tuning and high-volume inference. The accompanying idea, set out in NVIDIA’s own AI factory vision, is a unit of output: the factory produces tokens, and its performance is measured as throughput and cost per token rather than as racks, cores or floor space.
That is a genuinely useful reframing for anyone running production inference, and it is also a vendor’s reframing of a category in which that vendor sells the most expensive component. Both things are true. Treat it as an argument, attributed to NVIDIA, rather than as an industry definition.
The European Union uses the same phrase for something quite different. Under EuroHPC, AI Factories are a separate and unrelated funded programme built around AI-optimised supercomputers, providing compute and support services to European industry, startups and researchers. The EU has established 19 AI Factories and 13 AI Factory Antennas, with an EU contribution estimated at EUR 400 million in 2024 and a total budget of up to EUR 800 million from Digital Europe Programme funds. In that context “AI factory” names a public capability programme with an access model, not a facility archetype.
So when a supplier says AI factory, the first question is which of those two they mean, and the usual answer is neither.
| AI factory, NVIDIA framing | Conventional data centre | AI Factories, EuroHPC programme | |
|---|---|---|---|
| Origin of the term | NVIDIA, from GTC 2024 and its glossary | Established industry category | EuroHPC funded programme, 19 sites and 13 antennas |
| Unit of purchase | A rack-scale system such as GB200 NVL72, bought as a coherent compute domain | Space, power and cooling, which you fill | Access to AI-optimised supercomputers, with support services |
| Cooling and density | Around 120 kW per rack, direct-to-chip liquid, no air-cooled configuration | Roughly 5 kW to 10 kW per rack, ASHRAE air-cooled classes A1 to A4 | Not a facility archetype |
| What the label signals | Output measured as throughput and cost per token | Space, power and cooling | A public capability programme with an access model |
Where the distinction is real engineering
Strip the branding and there is a substantive shift underneath, and it is about rack-scale co-design rather than about the building.
A traditional facility sells you space, power and cooling, and you fill the rack. The current generation of AI hardware inverts that. NVIDIA’s GB200 NVL72 arrives as a rack, not as servers you assemble. Published specifications put it at around 120 kW per rack, distributed over a 48 V DC busbar, with direct-to-chip cold plates on GPUs, CPUs and NVLink switches fed by coolant distribution units and a facility water loop. There is no air-cooled configuration. A facility either has the water loop, the floor loading and the electrical capacity, or it cannot take the rack at all.
Compare that with the general enterprise baseline of roughly 5 kW to 10 kW per rack and it is clear why retrofitting is rarely a small job. ASHRAE’s TC 9.9 thermal guidance reflects the same shift: alongside the long-standing air-cooled classes A1 to A4, the fifth edition added an H1 class for high-density liquid-cooled equipment with a much narrower coolant supply band, in the region of 18 to 22 degrees. Facility water temperature is now a design constraint on your hardware choice.
Two further things are real. First, the interconnect is part of the unit of purchase. In a rack-scale system, the NVLink domain, the scale-out fabric and the scheduler are co-designed, so you are buying a coherent compute domain rather than a count of GPUs. Second, the operating profile is different. Training runs are long, synchronous and checkpoint-sensitive, so a brief thermal or power event costs hours of work rather than a retry. That changes what redundancy and maintenance procedures need to look like.
Whether you call the result an AI factory or a high-density colocation hall, those constraints are physical and worth paying for.
Where it is framing
Three claims deserve a follow-up question.
Tokens per megawatt as a facility metric. Token throughput per unit of energy depends on the model, the quantisation, the batch size, the context length and the serving stack. A facility contributes power, cooling and network, and it can absolutely make throughput worse, but it cannot be credited with a token figure on its own. Ask which model and which serving configuration produced the number.
“Purpose-built for AI” without a rack density figure. The useful version of that claim is a number and a date: kW per rack available today, cooling method, and whether the capacity is installed or planned. Everything else is a brochure.
The factory metaphor implying capacity on demand. A factory implies you can order more output. GPU capacity is contracted, and the meaningful commercial terms are reservation length, whether capacity is dedicated or burstable, what happens on a hardware failure mid-run, and the currency the invoice arrives in. AUD-denominated pricing is not a detail when a training programme runs across a volatile exchange rate.
What an Australian buyer should evaluate
Ignore the label and check six things.
Power available now. kW per rack today, the redundancy configuration, and the utility capacity behind it. Ask what a second row of the same density would require.
Cooling method and water loop. Direct liquid cooling or immersion, coolant supply temperature, leak detection, and the maintenance procedure for pulling a node without risk to adjacent racks.
The compute unit you are actually buying. Whether you are contracting a coherent rack-scale GPU cluster or a set of individually scheduled GPUs. For distributed training the difference is significant; for inference it often is not, and paying for it anyway is a common error.
Network path for your dominant workload. Training wants bandwidth and low variance between nodes. Inference wants a short, consistent path to your users. A site can be strong at one and mediocre at the other.
Commercial structure. Reservation term, billing currency, egress treatment, and what a failed node or a failed run costs you.
Jurisdiction and control plane. Who operates the control plane, who supports it, where billing sits, and whether the operator or its parent is subject to a foreign disclosure regime. For Commonwealth-adjacent work, check certification under the Hosting Certification Framework, where Certified Strategic is the highest level of assurance and sensitive government data, whole-of-government systems and systems rated PROTECTED must be hosted using certified services.
The part no definition covers
Neither NVIDIA’s framing nor the EuroHPC programme definition says anything about legal jurisdiction. A facility can meet every density and thermal criterion described above and still sit inside a corporate structure that allows a foreign authority to compel access to what runs on it. Density is an engineering property. Jurisdiction is a corporate one. They are set independently, and only one of them is visible from a specification sheet.
Amaze is ISO 27001 certified and operates as an Australian company, with data residency in Australia and no CLOUD Act exposure. The facilities are in Sydney and Melbourne with direct connectivity to Australian internet exchange infrastructure. Whether that constitutes an AI factory depends entirely on whose definition you are using, which is a reasonable indication of how much the term is worth in a procurement document.
Related reading: What makes AI data centres different and Choosing an AI infrastructure partner.
Frequently asked questions
Is “AI factory” an industry standard term? No. It is NVIDIA’s framing, used from GTC 2024 and set out in the company’s glossary, and separately it is the name of a EuroHPC funding programme in Europe. No standards body defines it, so two suppliers using the phrase may be describing very different facilities.
What is the single most useful question to ask a facility claiming to be an AI factory? How many kW per rack are available today, with what cooling method, and is that capacity installed or planned. A rack-scale system such as the GB200 NVL72 needs roughly 120 kW and a facility water loop, so the answer to that one question rules most sites in or out immediately.
Do we need rack-scale systems for inference workloads? Often not. Rack-scale co-design exists to make large distributed training efficient by keeping GPUs tightly coupled. Many production inference workloads run well on smaller, independently scheduled GPU nodes, and contracting a coherent training domain you do not use is an expensive way to buy inference capacity.
Does a high-density facility guarantee better AI economics? No. Cost per unit of output depends on the model, the serving configuration, utilisation and the commercial terms as much as on the facility. A well-engineered site with a poor reservation structure and low utilisation can be more expensive per token than a modest one that is fully used.