NVIDIA · Blackwell · Layer 2
NVIDIA GB200 NVL72
Frontier training and the largest inference deployments, where a single 72-GPU coherent memory domain is the point.
rent Rentable by the hour from multiple clouds; purchasable at scale.
Specifications
- Memory
- 13.5 TB HBM3E across the rack
- Bandwidth
- 576 TB/s aggregate
- Compute
- 1.44 exaFLOPS FP4 inference
- Power
- ~120 kW nominal, 130-132 kW observed at full load
- Interconnect
- NVLink 5, 72 GPUs in one coherent domain
- Software
- CUDA
- Form factor
- rack
- Workload
- both
Verified 2026-09-09. Fields reading “not published” are exactly that — we do not estimate a figure a vendor withholds.
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.
The 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.
Cost per million tokens is the metric that decides most real purchases, and there is no neutral benchmark for it — every published figure comes from a company selling one side of the comparison. Read all of them, ours included, as claims rather than measurements.
The power question underneath this
A rack of frontier accelerators draws 120–200 kW against a 2026 average of about 27 kW, and the US grid interconnection queue exceeds 2,600 GW with waits approaching five years. Whether you can energise this part is now a harder question than whether you can buy it.
Layer 1 — Energy · The interconnection queue · Rack power density · Tokens per watt
Where you actually rent this
Choosing the part is the easy half. The same accelerator rents for wildly different money depending only on who you rent it from — the cheapest and dearest published on-demand price for an H100 sit 5.6× apart — and the provider, not the chip, is what decides whether you can leave later.
Compared against
- NVIDIA GB200 NVL72 vs Google TPU v7 (Ironwood) — Rack-scale coherent memory against a 9,216-chip inference pod. Different shapes of the same bet.Compare with Google TPU v7 (Ironwood)
- 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.Compare with NVIDIA GB300 NVL72
- AMD Helios (Instinct MI455X) vs NVIDIA GB200 NVL72 — 432 GB per GPU of ROCm against the CUDA rack everyone already runs. The memory argument at rack scale.Compare with AMD Helios (Instinct MI455X)
Related silicon
- NVIDIA B200 — 192 GB HBM3E, rent
- NVIDIA GB300 NVL72 — Higher HBM3E capacity than GB200; per-rack figure not consistently published, rent
- NVIDIA H100 — 80 GB HBM3, rent
- NVIDIA H200 — 141 GB HBM3E, rent
- NVIDIA Vera Rubin VR200 NVL144 — HBM4. NVIDIA quotes 20.7 TB per rack at NVL72 scale; the NVL144 CPX configuration is quoted at 100 TB of fast memory per rack., rent
- NVIDIA RTX PRO 6000 Blackwell — 96 GB GDDR7, buy
Sources
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.