72 Blackwell GPUs fused into one rack with 13.5TB shared memory and 30TB/s in-rack interconnect — a supernode in a single cabinet.
Fit an entire training supernode in one cabinet.
72 GPUs pre-integrated, ready to use.
13.5TB pool keeps huge models in one rack.
30TB/s bandwidth makes communication near-invisible.
Unified power and liquid cooling for long runs.
NVL72 interconnects within one rack at far higher bandwidth and lower latency than cross-node setups.
Both — ideal for single-rack residency of very large models.
Rack-scale needs advance booking; we schedule per project.
Yes, we tailor rack and network topology to your workload.