Data center demand is often discussed as a single growing quantity. It is more useful to treat it as two different requirements that happen to share a building type, because they do not want the same things from a site.
Conventional cloud and enterprise workloads are latency-sensitive and relatively moderate in density. They want to be near users and near network. AI training workloads are different: they are far less sensitive to end-user latency, considerably denser per rack, and dominated by the availability and cost of power. Inference sits between the two and increasingly pulls back toward the user.
Why the distinction changes siting
If the binding requirement is proximity to users and networks, the viable sites are metropolitan, and the constraint is land and connectivity. If the binding requirement is large, firm, affordable power, the viable sites are wherever that power is, and the constraint is grid position and generation.
These two logics rarely point at the same place. A market that reads its opportunity through the first lens while capital is arriving with the second will mis-price its own sites - and developers will optimise for the wrong attributes.
The density problem is a building problem
Higher rack densities are not simply more of the same load. They change the cooling approach, the mechanical and electrical design, the structural requirements and the practical retrofit path for existing stock. A facility designed for conventional densities cannot always be upgraded economically, which means some existing capacity is not substitutable for the new demand however well located it is.
For investors this matters more than it first appears. Capacity is frequently discussed as fungible. It is not. The question is not how much capacity exists, but how much of it can serve the workload that is actually growing.
What this means in Indonesia
Indonesia has the demand fundamentals that make digital infrastructure attractive: a large, young, connected population and a rapidly digitising economy. The constraints are physical and institutional rather than commercial.
The questions that decide a project here are consistent: is there firm power at the right point on the grid, on a timeline the project can commit to; is the land status clean and the environmental pathway workable; is there a local structure that a hyperscale tenant and an infrastructure investor will both accept; and is there a partner capable of holding the development together through the period before revenue.
Projects that can answer those four questions with evidence tend to attract capital. Projects that answer them with intention tend to stall - not because the demand thesis was wrong, but because the demand thesis was never the constraint.
The strategic read
The useful conclusion is not that AI will drive data center demand; that is already widely held. It is that the shape of the demand determines which sites are viable, and that the differentiator is shifting from market access toward power position and deliverability.
Markets and developers that understand which kind of demand they are actually competing for - and build the power and regulatory position to serve it - will be able to transact. Those that compete on narrative will find the diligence unforgiving.