Filed under market.
Every answer an AI model gives comes out of a building on a piece of land that somebody had to find, connect and permit. We use these models in our own work every day, and it is a useful discipline: each query is a small reminder that the demand in the connection queues is real, and that it has a physical address.
The workloads are not all alike, and the difference decides where they can go. Training a model is a long, heavy computation that does not much care where it runs, as long as power is abundant, affordable and firm for years. Inference, the answering, is lighter per task but sits in the path of every user, and some of it has to be close to them. Training follows power north. Latency-sensitive inference pays for proximity in the south.
Both change the site itself. AI racks draw several times the power of the cloud racks that came before them, and new deployments are increasingly cooled with liquid rather than air. That puts more capacity on less land, brings the substation and the cooling plant closer to the center of the layout, and returns heat at higher temperatures, which makes it easier to use in a district network. A site laid out for the last generation of data centers can be the wrong shape for this one.
This is why the portfolio is weighted the way it is: deep capacity in the north for the workloads that can travel, and southern positions only where a connection path is already in hand. The demand is not in question. What is in question is whether a site was chosen for the workload that will actually arrive.
