Everything above assumes discrete nodes attached to a scale-out fabric. Rack-scale systems change the shape of several gates. This section is optional; skip it unless you are commissioning GB300 NVL72 or planning for the generation after.
The rack is the unit. A GB300 NVL72 arrives as one machine: 18 compute trays, 9 NVLink switch trays, 8 power shelves, and a busbar, all on a single liquid loop. Several gates collapse or move as a result. The cooling gate becomes a hard prerequisite rather than a parallel workstream, because the rack has one flow requirement and there is no partial-population strategy that avoids it. The physical-layer gate splits in two: the NVLink domain inside the rack is factory-wired and verified as a unit, while the scale-out links — four ConnectX-8 SuperNICs per compute tray, 800 Gb/s per GPU — are the part that still needs a rail audit against your leaf map.
A new health domain: the NVLink fabric. Beyond per-node NHC, a rack-scale system has its own fabric to prove. The NVLink switch trays present a 72-GPU all-to-all domain, and its management plane has to be up and consistent before any job sees the rack. Expect to verify NVLink link state and error counters across all trays, confirm the fabric manager has claimed the domain and reports it healthy, and run an intra-rack bandwidth check as a distinct step from anything involving the scale-out fabric. A rack that passes scale-out link qualification but has a degraded NVLink lane will produce collective results that make no sense until you look inside.
# NVLink domain health on a rack-scale system
$ nvidia-smi nvlink --status | grep -c "Link .*: 100" # every link at rate?
$ nvidia-smi nvlink -e # error counters, expect zero
$ systemctl status nvidia-fabricmanager # domain claimed and healthy
$ nvbandwidth -t all_to_all_bidirectional # intra-rack, before scale-out
Out-of-band gets denser and more important. The reference design pairs the rack with its own out-of-band switches, because there are now tray BMCs, switch-tray controllers, power-shelf telemetry and leak sensors to reach, and because a rack-scale system that has lost its management path cannot be triaged by walking to it.
Looking ahead. The Vera Rubin generation keeps the same commissioning shape and moves every number. Scale-out goes to ConnectX-9 at 1.6 Tb/s per GPU; the Ethernet side moves to Spectrum-6 with co-packaged optics, which changes the physical-layer gate meaningfully — there is no pluggable transceiver to reseat, so optical faults are diagnosed and repaired differently, and the 24-hour soak reads external fibre and connector quality rather than module population. Rack power rises far enough that the power gate stops being a formality: 800 VDC distribution, announced for that generation, brings its own commissioning procedure and its own safety training. Plan the bring-up calendar for that, not for the one you ran last time.
The gates do not change. What changes is how expensive it is to discover, at gate five, that you should have finished gate zero.