AI & Data Center Stocks

The buildout has layers, and they do not move together

Artificial intelligence is bought as a single story and paid for in several. The processor is one purchase, the building that houses it is another, the electricity that feeds it is a third, and the cooling that keeps it alive is a fourth. Each has its own suppliers, its own lead time and its own constraint, and money arrives at them in sequence rather than all at once.

That sequence is the useful part. Capital enters at the chip, because that is what gets announced. Then it has to be installed somewhere, so it moves to the buildings and the contractors. Then those buildings need power on a scale that utilities did not plan for, so it moves to generation and to the grid. Then the density of the racks makes air cooling insufficient, so it moves to liquid cooling and thermal management. Each layer becomes the bottleneck in turn, and the bottleneck is where the pricing power sits.

The compute layer is the most visible and the most crowded. The infrastructure layer — servers, networking, the physical assembly — is a contract business with thinner margins that has been repriced by demand arriving faster than anyone had planned capacity for. The power layer is the slowest to respond, because a turbine or a substation takes years, and it is where the constraint has moved most recently.

Robotics and industrial automation belong here for a reason that is not obvious: they are the same capability applied to a physical process rather than a digital one, and they are bought by the same customers with the same justification.

The risk shared by the whole family is that the spending is discretionary and concentrated. A handful of buyers account for most of it, and a decision taken in one boardroom repricing the group is not a hypothetical.