Rack-scale
GB300 NVL72
The GB300 NVL72 is the memory-dense rack: 288 GB of HBM3e per GPU and 20.7 TB across a 72-GPU coherent NVLink 5 domain, liquid-cooled.
Specification
- GPU memory
- 288 GB HBM3e/GPU · 20.7 TB per rack
- Intra-node
- 72-GPU coherent NVLink 5 domain
- Inter-node
- Quantum-X800 XDR between racks
- Node config
- Whole or half rack · liquid-cooled
- Best for
- Frontier MoE, reasoning-heavy inference
A 72-GPU coherent NVLink 5 domain at 1.8 TB/s per GPU inside the rack; Quantum-X800 InfiniBand XDR at 800G per link between racks, dedicated per tenant.
Why this machine
GB300 is what frontier MoE and reasoning-heavy inference ask for — a coherent domain large enough that expert routing does not become a network problem, with enough HBM per GPU that the KV cache does not force the model across more racks than the compute needs.
It is quoted per rack and delivered the same way as everything else on this site: our own liquid-cooled halls at 130 kW per rack, per-GPU power instrumentation, a 72-hour burn-in report you sign off, and hot spares on the same fabric with a 15-minute replacement SLA.
What it is for
Frontier MoE training
Expert routing that stays inside a 72-GPU coherent NVLink domain instead of crossing the network every step.
Reasoning-heavy inference
Long chains and large KV budgets on the densest HBM per GPU we rack.
Trillion-parameter serving
20.7 TB of HBM per rack, with Quantum-X800 XDR between racks and no per-tenant contention on it.
GB300 NVL72 questions
What does a GB300 NVL72 rack cost?
On request. Rack-scale systems are quoted per rack — that is a supply constraint, not a pricing one. A firm quote, a delivery date and the acceptance-test format come back within 48 hours.
How much HBM does a GB300 rack have?
288 GB of HBM3e per GPU, 20.7 TB across the 72-GPU rack — against 186 GB per GPU and 13.4 TB per rack on GB200.
When can I get one?
GB300 is on request. New capacity is commissioned in six-week blocks with the 72-hour burn-in included, so the quote carries a delivery date rather than an estimate.
Is the cooling a constraint on my side?
No. These are our own halls — direct-to-chip liquid cooling on 130 kW racks at a PUE of 1.12 annualized. We design, build and operate them rather than brokering someone else’s capacity.