INFRASTRUCTURE are advancing so rapidly that a facility designed today may be partially obsolete before it opens.
Jean-Francois describes the challenge through the lens of how the industry used to think about data centre lifespans. A traditional commercial cloud facility – what the industry calls an availability zone – might be financed over 15 to 20 years, with hardware refreshed once during that period. AI infrastructure operates on a very different timescale.
“ We’ re not building things where the hardware lasts more than three, potentially five years,” Jean-Francois asserts.“ You have to look at the data centre in terms of Lego bricks. What can you break down into components that, at some point, you pull out and replace?”
Training versus inference The distinction between different types of AI workloads adds another layer of complexity. Training and inference have very different infrastructure requirements.
From inside the building, Jamie notes, a training facility and an inference facility look broadly similar: both feature highdensity racks, liquid cooling and dense internal connectivity. The differences become apparent when you step outside.
“ Inference is below one millisecond from a latency perspective,” explains Jamie.“ You’ re going to be in the metro, 30 km out. With ML training, you’ re up to 20 milliseconds of latency and you can be 150 km from a metro.”
In practical terms, training campuses can be located in more remote, lower-
56 June 2026