Nvidia: Vera Rubin NVL72 Boosts Inference Throughput per Cost Up to 30x with ~3% More HBM per GPU

The pitch is less about raw memory than about architectural leverage: Nvidia is arguing that co-design across the stack, not HBM volume, is now the binding constraint on inference economics as memory supply tightens. That reframes the competitive question away from who secures the most HBM toward who can extract more throughput from roughly the same memory footprint. The figures cited are vendor-selected benchmarks on specific workloads, so whether the gap holds across broader production inference remains the open question worth watching.