NVL72 ARCHITECTURE EXPLAINED
NVL72 packages 72 GPUs in a single rack with NVLink 5.0 fully connecting all GPUs. GB200 NVL72 delivers 1.4 exaFLOPS FP8 in a 120 kW rack. GB300 NVL72 increases to 2.1 exaFLOPS FP8 at 140 kW with next-gen NVSwitch and Vera CPU. Both function as a single logical GPU with 13.8 TB (GB200) or 16.5 TB (GB300) unified memory via NVLink.
SPEC COMPARISON
GB200 NVL72 uses 36 Grace CPU + 72 Blackwell GPU chips with 13.8 TB total HBM3e and 900 GB/s NVLink per GPU. GB300 NVL72 uses Vera CPUs and Blackwell Ultra GPUs with 16.5 TB HBM4, 1.8 TB/s NVLink, and 2.1x FP4 throughput. Both support liquid cooling at 40-50 kW per rack for a total of 120-140 kW rack power.
| Spec | GB200 NVL72 | GB300 NVL72 | Delta |
|---|---|---|---|
| GPUs | 72 Blackwell | 72 Blackwell Ultra | -- |
| Total HBM | 13.8 TB | 16.5 TB | +20% |
| FP8 TFLOPS | 1,440 | 2,100 | +46% |
| NVLink/GPU | 900 GB/s | 1.8 TB/s | 2x |
| Power | 120 kW | 140 kW | +17% |
| CPU | Grace | Vera | Next-gen |
PERFORMANCE COMPARISON
GB300 NVL72 delivers 46% higher FP8 throughput and 2.1x FP4 throughput versus GB200. Training throughput improves 30-50% across Llama 4 405B and GPT-scale models due to doubled NVLink bandwidth reducing pipeline bubble overhead. Inference throughput for long-context workloads improves 40-70% due to HBM4's 6.4 TB/s bandwidth per GPU versus HBM3e's 4.8 TB/s.
COST ANALYSIS
GB200 NVL72 pricing runs $2.5-3.5M per rack at list, $2.0-2.8M with volume discounts. GB300 NVL72 pricing starts at $3.5-4.5M with early 2026 allocations fully committed through Q3. Total cost of ownership including power and cooling adds $400,000-600,000/year per rack. GB300's 30-50% throughput improvement yields 15-25% better cost-per-million-tokens at scale.
WORKLOAD MAPPING
GB200 NVL72 is sufficient for Llama 4 405B training at 50-60% model FLOPS utilization. GB300 NVL72 enables frontier-scale models above 1T parameters with 65-75% utilization. Inference for 1M-token context workloads benefits disproportionately from GB300's doubled NVLink bandwidth. Training runs exceeding 30 days favor GB300 despite higher upfront cost.
DECISION FRAMEWORK
Buy GB200 NVL72 when shipping a product within 8-12 weeks, $3-5M budget, or training models under 500B parameters. Buy GB300 NVL72 when planning frontier research, total budget exceeds $6M, or model scaling beyond 500B is expected within 12 months. GB200 availability is immediate; GB300 requires 12-16 week lead time through Q3 2026.