Vera Rubin NVL72 is presented as one rack-scale system with 72 Rubin GPUs and 36 Vera CPUs. That architecture changes the planning unit from an accelerator card to a coordinated compute, networking and facilities block. To address “Start with the rack, not a single GPU”, build the bill of materials around complete rack or supported pod configurations, then map which elements are supplied by NVIDIA or the OEM and which remain the site operator’s responsibility. Test that result on NVIDIA Vera Rubin.

NVIDIA lists 20.7 TB of HBM4, 54 TB of LPDDR5X CPU memory and 260 TB/s of NVLink switch bandwidth for NVL72, while explicitly labeling the specification preliminary. Those numbers are useful for architecture comparison, not contractual acceptance criteria. To address “Keep preliminary specifications visibly preliminary”, record the publication date and revision of every specification used in the design, and require the purchase quote to identify the final supported configuration. Test that result on NVIDIA Vera Rubin.

Large aggregate memory does not mean every byte is equally reachable at the same latency or bandwidth. GPU HBM, CPU memory, KV-cache placement and offload paths serve different purposes in training and inference. To address “Model memory locality before capacity”, list model weights, KV cache, activation or optimizer state, CPU-side staging and dataset cache separately so memory placement decisions are tied to software behavior. Test that result on NVIDIA Vera Rubin.

NVLink 6 handles the tightly coupled scale-up domain inside the rack, while ConnectX-9 and the supported InfiniBand or Spectrum-X fabrics handle scale-out. Mixing those roles produces unrealistic network assumptions. To address “Distinguish scale-up from scale-out traffic”, create separate traffic budgets for intra-rack GPU communication, inter-rack collective traffic, storage access, management and client ingress instead of one aggregate bandwidth number. Test that result on NVIDIA Vera Rubin.

A fast accelerator fleet can wait on dataset reads, checkpoint writes, model distribution or KV-cache movement even when raw capacity is sufficient. Rubin-era infrastructure therefore needs explicit bandwidth and concurrency targets for the storage service. To address “Plan the storage path as part of utilization”, benchmark representative reads and writes at the intended node count, include metadata pressure and checkpoint bursts, and size the network path all the way from storage media to the compute rack. Test that result on NVIDIA Vera Rubin.

BlueField-4 is part of the Vera Rubin platform rather than a decorative management component. DPU placement can affect network isolation, infrastructure services and data movement architecture. To address “Use BlueField and data movement intentionally”, decide which functions belong on hosts, DPUs and external services before designing VLANs, security zones, storage paths or observability, then validate that design against the chosen software stack. Test that result on NVIDIA Vera Rubin.

NVIDIA describes ConnectX-9 with 1.6 Tb/s per-GPU bandwidth capability. Reaching useful application throughput still depends on switch fabric, cabling or optics, routing, collective patterns and oversubscription. To address “Treat 1.6 Tb/s networking as a topology question”, draw the physical and logical topology with port counts and failure domains, then calculate oversubscription at each stage rather than multiplying link-rate labels. Test that result on NVIDIA Vera Rubin.

A third-generation rack platform can only deliver its intended performance when the electrical and thermal systems support sustained operation. Mechanical, cooling and power changes are part of the Rubin platform story. To address “Design power and cooling together”, obtain final rack power and coolant or airflow requirements from the OEM, add facility headroom, and have qualified engineers validate distribution, redundancy and emergency behavior. Test that result on NVIDIA Vera Rubin.

AI workloads can move between steady inference, bursty reasoning, training and maintenance windows. A design that fits only the average load can fail during synchronized peaks or degraded-redundancy operation. To address “Preserve operating headroom”, define normal, peak and single-failure operating envelopes for compute, fabric, storage and facilities, then keep those thresholds visible in monitoring. Test that result on NVIDIA Vera Rubin.

Rubin introduces new CPUs, GPUs, networking generations and rack integration. Existing Blackwell workflows may transfer conceptually, but firmware, drivers, orchestration, monitoring and facility interfaces need a fresh validation cycle. To address “Plan migration from Blackwell as a system change”, create a migration matrix covering images, drivers, CUDA stack, NCCL, scheduler, network configuration, storage mount behavior, telemetry and rollback before moving production workloads. Test that result on NVIDIA Vera Rubin.