"AI factory" is in danger of going the way of "AI-ready". Every new hall with a few GPU racks in it is starting to wear the label.
The difference is real, though, and you can measure it.
A conventional data centre is built to host other people's workloads. Lots of different ones, each with its own owner and its own pattern of load. In the Uptime Institute's 2026 survey, the average rack draws under 8 kW. Air cooling copes. Success is measured in uptime.
An AI factory is built to do one thing, which is turn power into tokens. NVIDIA's own definition measures the product in token throughput. A single GB200 NVL72 rack draws around 120 kW, roughly fifteen times that average, and it has to be liquid cooled. The building, the power train, the network and the software are all designed around keeping thousands of GPUs busy on the same kind of work.
So the questions change with it. How many tokens per megawatt. What share of the GPUs are doing useful work at any given hour. An idle GPU is about the most expensive thing in the building.
That's the test for the label. If the site was designed around turning power into tokens and keeping the GPUs busy, it's a factory. If it's an existing hall with some GPU racks and the same air cooling, it's a data centre with GPUs in it.
Useful, and perfectly fine. Just not the same thing.