Optional context for teams evaluating GB300 NVL72 or the Rubin generation against the node-based design above.
Rack-scale collapses one tier into the rack. A GB300 NVL72 puts 72 GPUs into a single NVLink domain across nine switch trays, at 130 TB/s aggregate. Collectives that fit inside 72 GPUs never reach your leaf switches at all. The scale-out fabric still exists — 72 ConnectX-8 SuperNICs, four per compute tray, 800 Gb/s per GPU into either InfiniBand or Spectrum-X Ethernet — but the traffic it carries is different in character: less of it, in larger and burstier flows, crossing the spine.
Three design consequences follow.
The topology math starts from racks, not nodes. The rail-optimised arithmetic in Act 2 still applies, but the unit that attaches to rail n is a tray position within a rack rather than a node in a cabinet. Recompute leaf and spine counts from the number of racks and the per-rack NIC count; do not scale the node-based numbers.
The north-south attachment changes. The reference design pairs the rack with BlueField-3 DPUs for the north-south path — roughly 480 Gb/s per rack — and its own pair of management switches. Your frontend leaf sizing is now driven by rack count at that rate, not by 72 servers at 400 Gb/s, and the oversubscription arithmetic must be redone. It usually comes out better, because far more traffic stays inside the rack.
Power and cooling become the binding constraint on floor layout. At up to 142 kW per rack the question stops being "how many ports does the leaf have" and becomes "how many of these racks fit on this floor with this loop". Network design and facility design stop being sequential activities. Chapter 28's flow and temperature arithmetic is a prerequisite to drawing the row, not a consequence of it.
Looking one generation further. Vera Rubin NVL72 raises scale-out to 1.6 Tb/s per GPU on ConnectX-9 and moves the Ethernet fabric to Spectrum-6 at 102.4 Tb/s per switch, whose co-packaged optics change both the power budget and the failure modes of the physical layer — there is no pluggable module to swap, so optical faults are a fibre-and-connector conversation. Rack power rises far enough that 800 VDC distribution has been announced for the generation. NVLink 6 widens the scale-up domain again, pushing still more collective traffic inside the rack.
The structure in this chapter survives all of it. Four fabrics, sized from radix, with route policy as blast-radius control and an out-of-band path that does not depend on anything it is used to fix. What changes is the arithmetic — and the arithmetic is the part you were always going to redo anyway.