Morgan Stanley says a 32 GW power shortfall hits memory and optics suppliers harder than Nvidia

If the 34% is a share of what developers need, a 32 GW gap implies a requirement of about 94 GW through 2028, against roughly 62 GW that can actually be connected ^[12][13]. A hyperscale rack of Nvidia training chips can draw 120 to 140 kilowatts ^[10]. At those ratings, the missing power would have run between about 229,000 and 267,000 racks ^[14]. The real number would be lower if the 32 GW also has to cover cooling and other site overhead.

Morgan Stanley's case for the two chipmakers is that they can move. According to Benzinga, as cited by Quartz, Nvidia can send chips to a project that has secured power when another facility is delayed, while suppliers tied to the delayed project have fewer options ^[7]. The bank called that flexibility a cushion that smaller, lower-tier suppliers do not have ^[6].

The bank's own estimate limits that flexibility. In a US market about a third short of power ^[2], moving a chip from a stalled site to a connected one shifts it around inside the same 62 GW or so of capacity ^[13]. That adds no supply. So the protection leans heavily on one item from the bank's list, growth across multiple geographies ^[6].

The waits are measured in years. Transmission for a new facility takes five to ten years to plan and energize, while the building goes up in 18 to 36 months ^[11]. Even the fastest grid connection arrives two years after the slowest build is finished ^[15]. Nvidia chief executive Jensen Huang has said that real estate, electricity and the structure must be lined up before any computing gear goes in, and he put that sequence at two to three years, according to Benzinga ^[9].

Who pays depends on what a delay turns into. If stalled projects connect late, component makers hold inventory until they do. If those projects never connect, the shortfall becomes lost demand, and the chipmakers' cushion is only as big as the powered capacity left to take rerouted chips. Morgan Stanley has argued a third case, per Benzinga: customers under electricity pressure buy the most computationally efficient hardware, so scarcity raises Nvidia's share of whatever does get built ^[8].

I think the bank has the direction right. A slipped project hurts the supplier that sold into that one site the most, and Nvidia sells into many ^[7][6]. The counter-case is cancellation. If enough projects die outright, a 32 GW gap is more than rerouting can absorb, and the chipmakers' order books shrink with everyone else's ^[2]. The view is wrong if Nvidia or Broadcom cut numbers over power-related delays before memory and optics makers report pushouts ^[1][4].