Layer 2 · head to head
NVIDIA GB300 NVL72 vs NVIDIA GB200 NVL72
One generation, and 15-80 kW more per rack. The question is whether your facility can energise it.
| NVIDIA GB300 NVL72 | NVIDIA GB200 NVL72 | |
|---|---|---|
| Can you get it | Rentable by the hour from multiple clouds; purchasable at scale. | Rentable by the hour from multiple clouds; purchasable at scale. |
| Memory | Higher HBM3E capacity than GB200; per-rack figure not consistently published | 13.5 TB HBM3E across the rack |
| Bandwidth | Not published per rack | 576 TB/s aggregate |
| Compute | Not published in comparable FP4 terms | 1.44 exaFLOPS FP4 inference |
| Power | 135-200 kW per rack | ~120 kW nominal, 130-132 kW observed at full load |
| Interconnect | NVLink 5 | NVLink 5, 72 GPUs in one coherent domain |
| Software | CUDA | CUDA |
| Workload | both | both |
NVIDIA GB300 NVL72
For: The same buyers as GB200, one generation on, where throughput per megawatt is the binding constraint.
The catch: Pushes rack power to 135-200 kW. Very few facilities in the world can energise and cool that today, and the interconnection queue means new ones are a five-year decision.
Economics: NVIDIA claims up to 50x higher throughput per megawatt versus Hopper. Per-megawatt is the right frame in 2026 — see the Energy layer.
NVIDIA GB200 NVL72
For: Frontier training and the largest inference deployments, where a single 72-GPU coherent memory domain is the point.
The catch: Liquid cooling is a mandatory architectural requirement, not a preference. At 120-132 kW a rack it renders most enterprise halls structurally and electrically unable to host it. This is a datacentre decision, not a hardware decision.
Economics: NVIDIA publishes two cents per million tokens and a 15x ROI claim for this configuration. Treat vendor TCO as a ceiling, not a forecast.
Before either — can you power it?
135-200 kW per rack against ~120 kW nominal, 130-132 kW observed at full load. In 2026 that comparison usually matters more than the FLOPS one: the US interconnection queue exceeds 2,600 GW with waits approaching five years, and roughly 80% of projects withdraw before energising.
Want this whole layer as one table?
The accelerators on these pages as a single CSV — memory, bandwidth, compute, power, interconnect, software stack, availability, who each part is for and the catch — ready to sort, filter or load into a notebook.
Get the silicon table — free for research and evaluation
Four short lines and the download starts at once (CC-BY-NC-4.0, built from these pages the moment you ask). What you type is used only to answer the request and is never sold. Prefer Hugging Face? Request it there.
Using it inside a product, a client report or anything you sell? That needs a commercial licence.