Nvidia says Vera Rubin NVL72 delivers up to 30x GB300 throughput per dollar

Nvidia said on November 5 that Vera Rubin NVL72 delivers up to 30 times the throughput per dollar of GB300 NVL72 on one agent workload, while high-bandwidth memory per GPU is up about 3%. The post names no ship date and no price.

At 125 tokens per second per user, Vera Rubin NVL72 delivers up to 30x higher throughput per dollar than GB300 NVL72, with HBM capacity per GPU up about 3%. Nvidia tied that figure to the SemiAnalysis AgentX DeepSeek V4 Pro workload. A separate line in the same post says GB300 NVL72 delivers up to 10x the throughput per dollar of an H200 platform at 100 tokens per second per user, with HBM capacity per GPU roughly doubling. The two comparisons are not the same jump.

The post treats memory pressure as the constraint. Nvidia said frontier models now have trillions of parameters, and that agent trajectories can process up to 3 million tokens in some cases. It said model weights and KV cache compete for HBM, the fastest and most expensive tier. CPU memory, local SSDs, and shared storage sit below it. The post says memory supply is tightening and prices are rising, and that the trend is expected to continue through the near term. It did not give a price index.

The claimed gain is software and placement, not a much larger memory chip. Nvidia said its NVFP4 format reduces the space weights take compared with FP8, which it called the lowest precision possible on Hopper. It said an NVL72 system can spread model weights across up to 72 GPUs, leaving more HBM for cache and concurrent users. It said Dynamo can move inactive context to CMX storage, which is shared across servers and costs less per bit than HBM. It did not state a CMX price or a capacity figure.

Nvidia also said it was first to adopt HBM3e and HBM4 in an accelerated computing stack. The November 5 post does not say Vera Rubin is shipping, and it does not name a customer or a delivery quarter. The 30x figure is an up-to result on one named workload at one response rate, not a general claim for every model.