Nvidia guides to $108B while the edge it is told to chase caps out at 30 kW a rack

The guide is where the tension sits. $108 billion against $96.2 billion works out to about $11.8 billion of sequential growth, near 12.3 percent ^[15], and the money behind it is going into monolithic liquid-cooled mega-clusters that hyperscalers are already committed to ^[13]. A vendor absorbing demand at that rate has little commercial reason to redesign for buildings it cannot yet fill, which is why the distributed-edge case in SiliconANGLE's piece ^[14] should be read as a forecast about where the next market has to come from, not as a disclosed roadmap.

The physics under it is firmer than the market sizing. NVLink-scale racks draw up to 140 kW or more in a single cabinet ^[5]. A typical telco central office or enterprise hall is capped at roughly 30 kW to 50 kW per rack ^[6]. Even the most generous edge rack in that range is short by a factor of 2.8, and against a 30 kW cap the gap is about 4.7 times ^[16]. Attempting the drop-in breaks power delivery and overloads the cooling ^[12], so the choice is an electrical retrofit or splitting the compute envelope across four or five racks joined by a high-speed scale-up fabric, with optical interconnects carrying the bandwidth that copper cannot at that distance and power budget ^[9][10].

That is the part worth putting on a roadmap, and it is also the part the source leaves unpriced. There is no latency figure, no cost per optical port, and no statement from Nvidia about edge disaggregation anywhere in the material; the quoted remarks concern the buildout and Vera Rubin reaching full production ^[17]. The 30 GW of aggregated edge capacity is explicitly the author's own estimate from a year of conversations and data gathering ^[7], and the two source records supplied here are the same SiliconANGLE article under two URLs, so every figure in the thesis traces to one writer ^[18]. Distrust the gigawatts before the kilowatts. The 140 kW is a specification; the 30 GW is a guess, and it is fragmented by the author's own account ^[8].

Scale still tells you something even if the estimate is loose. Thirty gigawatts expressed in 140 kW units is roughly 214,000 rack-equivalents, but delivered in 30 kW increments it is on the order of a million rack positions ^[19], and the disaggregation trick needs four or five of them adjacent and free in the same facility ^[9]. Fragmentation is not a footnote on the opportunity, it is the thing that decides how much of the 30 GW is reachable at all.

What survives regardless of the sizing is the ordering of constraints. Once the scale-up domain crosses cabinet boundaries, the fabric stops being a cabling line item and becomes the architecture decision, sitting on the same page as accelerator selection. The compensation is that you adapt the topology to the facility instead of asking a customer to rebuild its electrical system ^[11], which makes brownfield sites addressable without a construction budget. Huang's framing, that compute is revenue and demand is accelerating ^[3], is a statement about token output inside facilities purpose-built for 140 kW racks. Extending it to a telco closet is a different engineering problem, and so far only one person has published the number that makes it look like a market.