GB300 NVL72

Rack scale, 72 liquid-cooled GPUs

Short answer. The GB300 NVL72 is a rack of 72 liquid-cooled GPUs with 20.7 TB of HBM3e (72 × 288 GB, 20 TB according to NVIDIA), assembled by Dell, HPE, Lenovo and Supermicro. It serves a complete frontier model on premises, for large enterprises that host their own sovereign AI cloud.

Hardware sheet · QDNA sovereign AI platform · Published on 30 July 2026 · Updated on 13 October 2026

Indicative price	on quotation: neither NVIDIA nor its integrators publish a price for the rack; analyst estimates reported in the press range from $3.7M to $6.5M and are not prices (price guide)
GPU	72 Blackwell Ultra and 36 Grace (2,592 Arm Neoverse V2 cores)
Memory	20.7 TB of HBM3e according to Lenovo and Supermicro (72 × 288 GB, i.e. 20,736 GB, which the sizing pages round to 20,700 GB); NVIDIA states 20 TB, Dell “up to 21 TB”; 37 TB of fast memory
Compute	1.08 ExaFLOPS dense FP4 (1.44 with sparsity)
NVLink	130 TB/s across a 72-GPU domain
Power draw	132 to 142 kW per rack (NVIDIA reference: 142 kW; HPE and Supermicro: 132 to 140 kW); liquid cooling, about 90% of the heat goes into the liquid
Form factor	full MGX rack (48U cabinet at Lenovo)

What the machine runs

A single 72-chip supercomputer serves frontier models and a private AI cloud. Recommended runtime: disaggregated vLLM with LiteLLM.

Positioning

Large enterprises host their own frontier model in a single rack, in a sovereign AI cloud.

Manufacturers

Rack integration from Dell (PowerEdge XE9712 in IRSS racks), HPE, Lenovo and Supermicro, with liquid cooling. NVIDIA also names contract manufacturers (Quanta, Foxconn, Wistron, Wiwynn, Inventec, Pegatron), whose ranges are not checked here.

GPUDirect Storage

On platforms equipped with ConnectX cards, GPUDirect Storage technology establishes a direct path between NVMe or NVMe over Fabric storage and GPU memory, bypassing the CPU buffer. Compatible storage arrays, such as NetApp, VAST, DDN or WEKA, feed training and large-scale RAG at full speed.

Support

Open stack, maintained by its communities. NVIDIA AI Enterprise option (NIM, NeMo, Triton) with a service-level agreement, licensed per GPU.

Which models fit on GB300 NVL72?

The machine offers 20,700 GB of memory. 7 of the 7 open models in the catalogue load on it, in the format shown.

Model	Billion parameters	Most precise format that fits	Sizing page
GLM 5.3	744	FP16	GLM 5.3 on GB300 NVL72
Kimi K3	2,800	FP16	Kimi K3 on GB300 NVL72
Kimi K2.7 Code	1,000	FP16	Kimi K2.7 Code on GB300 NVL72
DeepSeek V4	1,600	FP16	DeepSeek V4 on GB300 NVL72
Nemotron 3 Ultra	550	FP16	Nemotron 3 Ultra on GB300 NVL72
MiniMax M3	428	FP16	MiniMax M3 on GB300 NVL72
Qwen 3.8 27B	27	FP16	Qwen 3.8 27B on GB300 NVL72

See the full sizing matrix.

Frequently asked questions

How much does a configured gb300 nvl72 cost for an LLM?

Pricing depends on configuration and supplier. No public price is listed for this configuration; a scoping call helps compare supplier quotes.

Which LLM fits in a gb300 nvl72?

Depends on the quantization format and the input context. The table on the page indicates the recommended maximum model size.

Which runtimes support the gb300 nvl72?

vLLM, llama.cpp, Triton Inference Server, depending on the chosen framework. The model, format and GPU combinations measured by QDNA are published in the measurements section.