The AI grid shortfall thesis, checked against its own arithmetic

Crypto Briefing reports that a single Nvidia GB300 rack draws more than 180 kilowatts and concludes that utilities are falling behind, leaving a substantial shortfall relative to AI data center demand ^[1][7]. The two headline numbers the piece cites do not support that conclusion, which matters because the same framing is being used to argue that utilities, grid equipment makers and independent power producers hold increasingly valuable positions ^[12].

Start with the arithmetic. Goldman Sachs, as cited in the piece, sees US data center power demand rising from 31 gigawatts in 2025 to 66 gigawatts by 2027 ^[6], a 2.1x increase ^[3] and roughly 35 GW of incremental load ^[1]. Against that, the piece states utilities can realistically deliver about 93 GW of additional practical supply ^[7]. That is about 2.7 times the incremental demand implied for the same period ^[2]. Set side by side, those two figures describe headroom, not a deficit. A shortfall exists only if the 93 GW lands on a slower clock than the 35 GW of demand, and no timeframe for the 93 GW is given ^[7].

The longer-dated figures are larger and softer. US data centers consumed about 176 TWh in 2025, some 4.4% of national generation ^[4], with projections of 466 to 580 TWh by 2030, or 9 to 12% ^[5]. The top of that range is about 3.3 times the 2025 level ^[4]. A band that wide, five years out, is a scenario set, not a measurement.

The scale rhetoric is doing work the numbers do not. Over 180 kW is presented as the load of 60 average American homes ^[2] and a 100 MW cluster as rivalling a small city ^[3]. That cluster is fewer than about 560 racks ^[5]. Same physical fact, two very different registers.

What is more credible is the qualitative part, and it is unquantified. The piece argues the binding constraint has moved from chips to physical infrastructure, with transformers, switchgear and other grid equipment in critically short supply and GPU clusters sitting dark without them ^[8]. Lead times for large power transformers have stretched significantly, and an 18-month wait is offered as the illustration of why servers, land and contracts do not matter ^[9]. In some regions, power requests have been denied or delayed because local grids cannot absorb the load without risking reliability for existing customers ^[10]. No region is named, no queue is measured, no lead time is stated as observed.

Adoption is the thinnest section. The evidence offered is that some bitcoin miners now hold power purchase agreements and interconnections that AI companies covet, and that deals have already occurred in which mining companies pivot or lease capacity to AI workloads ^[11]. No counterparty, no date, no megawatts, no terms.

Note where the piece sits. It is a crypto publication, and its conclusion is that miner-held grid access is the coveted asset ^[11], followed by a list of beneficiaries ^[12]. Its one bearish observation runs the other way: Nvidia's chips drive the demand, so energy constraints that slow buildouts mean fewer GPU orders ^[13].

The operating consequence: do not underwrite on the TWh chart. Underwrite on equipment delivery dates and interconnection status, which is the constraint the piece asserts and does not measure ^[8][9].