A transformer steps grid voltage down to the levels servers, cooling plant and switchgear can actually use [8]. Nothing downstream energises until one is installed and the utility connects it. Every build now runs two clocks in series: the equipment lead time and the interconnection wait. The equipment lead time is a purchase order. The interconnection wait is not.
Sixteen six-month build cycles fit inside an eight-year connection wait [1]. Four fit inside the 24-month emerging-market case [2]. An executive the Ukrainian-language report names, in transliteration, as Vin Kin Chung puts it directly: outside the industry the conversation is about GPUs, inside it starts with generator and transformer delivery dates [9]. The reporting is Reuters', relayed by mezha.net [22].
Run the 23% backwards on HD Hyundai Electric's book and the starting point was about $6.9bn, so roughly $1.6bn net was added in six months [3]. A growing backlog is a longer line, not more throughput; whoever orders next stands behind everyone already in it. The mix is moving too. Data centres go from 6.3% of the company's new power equipment orders to an expected 16% [6], about 2.5x in a year [4], and HD Hyundai attributes the European part of that to US hyperscaler spending in Finland, Germany and the UK [11].
Two forecast numbers deserve the usual treatment. Bank of America puts a single AI rack above 1.5 MW by end-2030, nearly 100 times a conventional server rack [12]. Divide and the implied comparison rack is about 15 kW [5], which the report never names, so the multiple rests on a baseline you cannot inspect. McKinsey's claim that liquid cooling cuts electricity by more than 27% [19] has the same gap: no baseline PUE, no climate, no rack density. For that 27% to land in your hall, your air-cooled starting point has to be as inefficient as whatever theirs was. Bank of America expects liquid cooling in 70% of new AI data centres by 2030 against roughly 30% now [18], which is a swing of about 40 percentage points [6].
The solid-state transformer is the smaller prize and the bigger project. It replaces large magnetic coils and copper windings with semiconductors [13], and UBS estimates roughly 4% better energy efficiency plus lower costs, with penetration possibly reaching 40% by 2030 and Chinese makers gaining on cost and capability [14][15]. A second executive, transliterated as Ping Cheng, describes it as an energy gateway requiring a different overall design and power architecture, so uptake takes time [17]. Delta Electronics has reported using them in a small data centre [16]. "Small" is carrying weight in that sentence.
Set the two efficiency figures side by side and the cooling lever is roughly seven times the transformer lever [7], on different bases, which is why the cooling retrofit is the argument teams are actually having. Power and cooling form one system. More watts in means more heat out, as Matti Zhao notes in the same report [20]. For a capacity plan, the binding dates are the energisation slot and the ship date on the switchgear. Both sit outside the buyer's org chart, and both are longer than the building they gate [2][3].
