AGUASTO

NVIDIA / Rack-scale system

NVIDIA GB200 NVL72

NVIDIA reports GB200 NVL72 shipments began in 2024. Availability of a specific OEM build remains supplier-specific.

Original Aguasto conceptual illustration. Not manufacturer geometry, an operating Aguasto asset, or a measured benchmark.

Manufacturer specification reference

Architecture
Grace BlackwellNVIDIA GB200 NVL72 (external source) · NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (external source) · CUDA GPU Compute Capability (external source)
Category
Complete rack-scale referenceNVIDIA GB200 NVL72 (external source) · NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (external source) · CUDA GPU Compute Capability (external source)
Accelerator memory
13,400 GBPublished rounded 13.4 TB aggregate GPU HBM converted to decimal GB; CPU memory excluded.NVIDIA GB200 NVL72 (external source)
Memory type
HBM3eNVIDIA GB200 NVL72 (external source) · NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (external source) · CUDA GPU Compute Capability (external source)
Memory bandwidth
576 TB/sPublished aggregate GPU memory bandwidth across the rack, not bandwidth available to one GPU or NVLink bandwidth.NVIDIA GB200 NVL72 (external source)
Power
Not verified for this recordTotal operating rack draw not verified in the reviewed product source. No power estimate is derived from GPU count.NVIDIA GB200 NVL72 (external source) · NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (external source) · CUDA GPU Compute Capability (external source)
Precision formats
NVFP4, FP8, FP6, INT8, FP16, BF16, TF32, FP32, FP64NVIDIA GB200 NVL72 (external source)
Cooling / integration
Liquid-cooled rack architecture.NVIDIA GB200 NVL72 (external source) · NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (external source) · CUDA GPU Compute Capability (external source)

Configuration boundaries

  • A rounded decimal-TB rack total is preserved; it is not an exact allocatable memory guarantee.
  • Manufacturer throughput is a theoretical rack figure and does not predict delivered tokens per second.

Published arithmetic throughput

Manufacturer peaks, not Aguasto measurements or tokens per second
SpecificationPrecisionPublished peakBoundary & source
NVFP4 peak computeNVFP4720 PFLOPSWhole rack, manufacturer peak dense specificationManufacturer source ↗
NVFP4 peak computeNVFP41,440 PFLOPSWhole rack, manufacturer peak sparse specificationManufacturer source ↗

Deployment considerations

  • Treat rack HBM as distributed memory connected through a scale-up fabric, not one locally attached allocation.
  • Validate liquid-cooling interfaces, facility heat rejection, electrical redundancy and OEM acceptance tests.

Software ecosystem

CUDA · NVIDIA Mission Control. Confirm the exact model, runtime release, server integration and vendor support matrix before evaluating a deployment.

Continue the research

NVIDIA DynamoTensorRT-LLMKubernetes GPU orchestration — NVIDIA device pluginModel weights versus runtime memoryCompare compatible hardware categoriesWhy modular infrastructure?

Sources & verification

Reference revision 1 · Reviewed . Manufacturer claims and documentation are attributed; Aguasto has not run the cited products in production.

  1. NVIDIA GB200 NVL72 ↗

    72 Blackwell GPUs and 36 Grace CPUs; 13.4 TB aggregate HBM3e; 576 TB/s aggregate GPU memory bandwidth; Liquid cooling; Precision table; Mission Control. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  2. NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer ↗

    GB200 NVL72 first shipments in 2024; GB300 NVL72 shipments in 2025. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  3. CUDA GPU Compute Capability ↗

    CUDA product identification for H100, H200, B200, B300, GB200 and GB300. Publication: date not stated in the reviewed source. Verified: 2026-10-12.