← All chips

AI accelerator · Global

Nvidia GB200 NVL72

Rack-scale liquid-cooled system pairing 72 Blackwell GPUs with 36 Grace CPUs on NVLink 5 (1.4 exaflops FP4 sparse, 13.4TB HBM3e). Analyst and reporting consensus puts the price at roughly $3 million per rack.

GlobalShippingNvidiaBlackwell familyLaunched 2024-03TSMC 4NP

Pricing not publicly disclosed · researched 2026-08-27

Specifications

Sourced specs

SpecValueSource
Configuration72 Blackwell GPUs + 36 Grace CPUs (2,592 Arm Neoverse V2 cores), liquid-cooled rackvia nvidia.com2026-08-27
FP4 Tensor Core (rack)1,440 PFLOPS with sparsity / 720 PFLOPS dense (20 / 10 PFLOPS per GPU)via nvidia.com2026-08-27
FP8/FP6 Tensor Core (rack)720 PFLOPS with sparsity / 360 PFLOPS densevia nvidia.com2026-08-27
FP16/BF16 Tensor Core (rack)360 PFLOPS with sparsity / 180 PFLOPS densevia nvidia.com2026-08-27
TF32 (rack)180 PFLOPS with sparsityvia nvidia.com2026-08-27
FP32 / FP64 (rack)5,760 TFLOPS FP32 / 2,880 TFLOPS FP64via nvidia.com2026-08-27
Memory13.4TB HBM3e total, 576 TB/s aggregate (192GB / 8 TB/s per GPU); 17TB LPDDR5X CPU memoryvia nvidia.com2026-08-27
InterconnectNVLink 5: 1.8 TB/s per GPU, 130 TB/s total NVLink Switch bandwidth, 800 Gb/s Quantum-X800 / Spectrum-X800 networkingvia nvidia.com2026-08-27
Performance1.4 exaflops FP4 (sparse) per rack; up to 30x LLM inference vs 72x H100via nvidianews.nvidia.com2024-03-18

Pricing

Price history, newest first

DatePriceTypeSourceConfidence
2025-09~$3.1M per NVL72 rackMorgan Stanley: ~$3.1M per GB200 NVL72 rack vs ~$190k for an H100 rackAnalyst estimatevia Morgan Stanley via Wccftechmedium
2025~$3.2M per NVL72 rackSemiAnalysis: GB200 NVL72 servers between $3.2M and $3.5M (44k-49k per GPU)Analyst estimatevia Epoch AI (citing SemiAnalysis)medium
2024-07~$3M per NVL72 rackMid-2024 analyst consensus ~$3M (~$41.7k per GPU); widely cited Reuters-era 'over $3M' figureAnalyst estimatevia Epoch AI (compiling mid-2024 analyst estimates)low
2024-06~$2.6M per NVL72 rackHSBC estimated ~$2.6M per GB200 NVL72 (~36k per GPU)Analyst estimatevia Epoch AI (citing HSBC)low

vendor-official or multiple independent reports · single major report or several secondary · estimate or lone secondary source

Prices in USD, converted at each observation year's average exchange rate; the original figure and rate print under every converted price. How prices are sourced and graded ↗

Context

Notes

  • Nvidia does not publish a rack price; the $3M+ figure is analyst/reporting consensus.
  • Successor GB300 NVL72 shipping from December 2025; Morgan Stanley estimated the follow-on Vera Rubin VR200 NVL72 at ~$7.8M.

Where it sits

In the stack

All specs and prices link to their original sources, dated as observed. Figures are market observations, not offers; treat each as of its date.