Specifications
Sourced specs
| Spec | Value | Source |
|---|---|---|
| FP4 Tensor Core (rack) | 1,440 PFLOPS with sparsity / 1,080 PFLOPS dense (15 PFLOPS dense per GPU) | via nvidia.com ↗2026-08-27 |
| FP8/FP6 Tensor Core (rack) | 720 PFLOPS with sparsity / 360 PFLOPS dense | via nvidia.com ↗2026-08-27 |
| FP16/BF16 and TF32 (rack) | 360 PFLOPS FP16/BF16 / 180 PFLOPS TF32 (with sparsity) | via nvidia.com ↗2026-08-27 |
| FP32 / FP64 (rack) | 6 PFLOPS FP32 / 100 TFLOPS FP64 | via nvidia.com ↗2026-08-27 |
| Memory | 20TB HBM3e per rack at up to 576 TB/s (288GB / 8 TB/s per GPU) | via nvidia.com ↗2026-08-27 |
| Node compute (DGX B300) | 144 PFLOPS FP4 inference / 72 PFLOPS FP8 training per 8-GPU node | via docs.nvidia.com ↗2026-08-27 |
| Interconnect | NVLink 5: 1.8 TB/s per GPU, 130 TB/s rack; dual ConnectX-8 SuperNICs (800 Gb/s per GPU) | via nvidia.com ↗2026-08-27 |
| B300 TDP | Up to 1,400W max configurable (system configs commonly 1,100W) | via glennklockwood.com ↗2026 |
Pricing
Price history, newest first
| Date | Price | Type | Source | Confidence |
|---|---|---|---|---|
| 2026-04-30 | ~$965.5K from ¥7,000,000 at 2026 avg 7.25 CNY/USD per B300 (DGX-class) 8-GPU server, ChinaRestricted-supply B300 servers at ~7M yuan (~$1M) in China, nearly double the ~$550k US price | Street price | via Reuters ↗ | medium |
| 2025-12 | ~$3.7M per GB300 NVL72 rack~$1B Apple order equates to ~250 racks at $3.7-4M each from Dell and Supermicro (51-55k per GPU) | Street price | via TweakTown / TMTPost / DataCenterDynamics (Apple order) ↗ | medium |
| 2025-10-25 | ~$53K per B300 cardB300 cards reportedly around $53k; 8-GPU B300 server about $530k | Analyst estimate | via IFP analysis (reported figures) ↗ | low |
| 2025 | ~$3M per GB300 NVL72 rackHSBC: GB300 NVL72 at ~$3M, about 15% above its GB200 figure | Analyst estimate | via 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
- Shipping status as of Aug 2026: GB300 NVL72 racks 'Available Now' per Nvidia; first racks delivered to CoreWeave via Dell in December 2025; HGX/DGX B300 systems shipping through 2025-2026.
- China-market B30A derivative (single-die, four HBM stacks, ~12x H20 performance) reported at ~$22.5k per card (IFP, 2025-10); Reuters (2025-05-24) earlier reported a $6,500-8,000 expected price for a China Blackwell part. US approval was still pending into 2026.
- Next-generation Vera Rubin (VR200 NVL72) announced; Morgan Stanley BOM estimate ~$7.8M per rack.
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.