A 180kW rack, an 18-month transformer, and a shortfall nobody has shown the math for

A single Nvidia GB300 rack reportedly draws more than 180 kilowatts, which cryptobriefing.com puts at roughly the load of 60 average American homes in one cabinet ^[1][2]. That is the unit of account now: the question for anyone financing AI capacity is no longer whether the chips ship, but whether the interconnect, the transformer and the switchgear arrive ^[8].

The demand side is well rehearsed. US data centers consumed about 176 terawatt-hours in 2025, roughly 4.4 percent of national generation, with projections of 466 to 580 TWh by 2030, or 9 to 12 percent of US electricity ^[4][5]. That upper case is a 3.3x increase in five years ^[4]. Goldman Sachs forecasts US data center power demand rising from 31 gigawatts in 2025 to 66 gigawatts by 2027 ^[6], a 2.1x move ^[3]. A cluster above 100 megawatts rivals a small city ^[3], and at 180 kilowatts a rack, 100 megawatts is fewer than 560 racks ^[5]. The physical footprint of a grid-scale load is now small enough to fit in one building, which is exactly why utilities keep getting surprised.

Here is where the published arithmetic stops working. The same piece says utilities can realistically deliver about 93 GW of additional practical supply, and then describes a substantial shortfall against AI data center demand ^[7]. But the demand figures it cites imply 35 GW of incremental data center load between 2025 and 2027 ^[1], against 93 GW of stated additional supply, roughly 2.7 times the increment ^[2]. No timeframe is given for the 93 GW, and no reconciliation is offered. The shortfall may well be real, but on the numbers presented it is asserted rather than shown. Treat any specific gigawatt gap you are quoted in a pitch deck the same way, and ask what period and what geography it covers.

The constraint that is documented is not aggregate megawatts, it is queueing. Transformers, switchgear and related grid equipment are in critically short supply, and lead times for large power transformers have stretched enough to cascade delays through the data center pipeline ^[8][9]. The source's own illustration is the useful one: you can order servers, lease land and sign contracts, and none of it matters if the transformer is 18 months out ^[9]. In some regions, power requests have been denied or delayed outright because local grids cannot absorb the load without risking reliability for existing customers ^[10]. National supply averages do not help you when the substation says no.

That reprices whoever already cleared the queue. Utilities with spare capacity can command premium rates and terms, equipment makers face a demand surge that could hold revenues up for years, and independent power producers near data center corridors hold more valuable positions than they did ^[12]. Bitcoin miners are an accidental beneficiary: years of power procurement left some of them holding purchase agreements and interconnections that AI buyers want, and deals have already appeared in which miners pivot or lease capacity to AI workloads ^[11]. Nvidia sits on the other side of that trade. Its chips create the demand, but if energy constraints slow buildouts, the orders slow with them ^[13].

What to watch: transformer and switchgear lead times, since that number moves before any generation number does ^[9]; the ratio of announced megawatts to energized megawatts at large sites; the count of regional denials and deferrals ^[10]; and the terms miners extract when they sell or lease interconnection rights ^[11], which is the cleanest available price for a queue position.