This act is optional context. Skip it if you are working with discrete HGX nodes; read it if you are evaluating rack-scale systems or planning a build more than a year out. Treat every figure as a starting point to confirm against the current datasheet.

GB300 NVL72 is the rack-scale form of the same silicon. Seventy-two Blackwell Ultra GPUs and thirty-six Grace CPUs sit across 18 compute trays, with 9 NVLink switch trays -- two NVSwitch ASICs each -- closing a fully connected 72-GPU domain at 130 TB/s of aggregate NVLink bandwidth. The rack holds roughly 20 TB of HBM3e and 17 TB of LPDDR5X, and is budgeted at up to 142 kW from eight 33 kW power shelves. Scale-out is 72 ConnectX-8 SuperNICs, four per compute tray, at 800 Gb/s per GPU into either Quantum-X800 InfiniBand or Spectrum-X Ethernet; the north-south path runs on BlueField-3 DPUs, and a pair of small management switches carries out-of-band.

For the fabric engineer the important number is not the FLOPS. It is that a 72-GPU collective never touches your scale-out fabric. The ratio of intra-domain to inter-domain bandwidth widens by an order of magnitude compared with an 8-GPU node, which means fewer, larger, burstier flows crossing the spine -- a traffic profile that stresses adaptive routing more and stresses raw bisection less.

Vera Rubin moves every number again. The Rubin GPU carries 288 GB of HBM4 at up to 22 TB/s and 3.6 TB/s of NVLink 6 per GPU; the Vera CPU brings 88 custom Arm cores with 1.8 TB/s of NVLink-C2C to its GPUs. Scale-out moves to ConnectX-9 at 1.6 Tb/s per GPU, and the Ethernet side moves to Spectrum-6 -- 102.4 Tb/s per switch chip, built with co-packaged silicon photonics rather than pluggable transceivers, which NVIDIA quotes at roughly five times better network power efficiency. The rack-scale form ships as Vera Rubin NVL72 -- 72 Rubin GPU packages per rack. Earlier roadmap material called the same system NVL144, counting compute dies rather than packages; both names describe one rack of 72.

Two planning consequences are worth carrying now, well before any of this is on a purchase order. Co-packaged optics change what a physical-layer fault is -- there is no module to reseat, so the diagnostic path in Chapter 9 shifts toward fibre, connectors and the switch itself. And rack power rises far enough to change the room, which is why the 800 VDC distribution direction NVIDIA has announced for the generation belongs in facility conversations happening this year, not next.