The traditional chain converts incoming medium-voltage supply down, passes it through an uninterruptible supply that can ride through a loss of source until generators start, distributes it around the hall, and converts it again inside each server. Every conversion loses a few percent, and every stage is equipment with a lead time.
AI racks have broken the assumptions this chain was designed around. A rack drawing over a hundred kilowatts cannot sensibly be fed by many separate cords with individual power supplies; the design moves toward higher-voltage distribution, rack-level power shelves feeding a shared direct-current busbar, and in the largest designs, distribution at higher voltage all the way to the rack.
The chain divides into equipment bought from different industries. What rides through a loss of supply and what carries the load until it returns are two separate purchases, and both are separate again from the distribution running across the hall. The medium-voltage switchgear and transformers at the head of the chain are the same products the grid buys, and are covered with them; the conversion inside the rack is designed by the server maker and is covered with the server.