Key figure: NVIDIA says HBM capacity per GPU rises about 3% in the Vera Rubin–GB300 comparison.
NVIDIA says HBM capacity per GPU rises about 3% in the Vera Rubin–GB300 comparison. Source: blogs.nvidia.com, 2026-11-05. Source: VERXERA card from blogs.nvidia.com

What matters

In its November 5, 2026 account of AI memory codesign, NVIDIA reports gains of up to 10x and 30x in throughput per dollar across two platform comparisons. These are company-reported results for a named workload at specified per-user token rates, not general performance guarantees. NVIDIA presents them alongside changes in high-bandwidth memory (HBM) capacity per GPU.

Verified facts

  • NVIDIA published its account of AI memory codesign on November 5, 2026.
  • On the SemiAnalysis AgentX DeepSeek V4 Pro workload, NVIDIA says GB300 NVL72 delivers up to 10x higher throughput per dollar than its H200 platform.
  • NVIDIA specifies that the GB300–H200 comparison uses 100 tokens per second per user, with HBM capacity per GPU roughly doubling.
  • For the same named workload, NVIDIA says Vera Rubin NVL72 delivers up to 30x higher throughput per dollar than GB300 NVL72 at 125 tokens per second per user.
  • NVIDIA says HBM capacity per GPU rises about 3% in the Vera Rubin–GB300 comparison.
  • NVIDIA describes full-stack codesign as designing GPUs, CPUs, networking, storage and software together to obtain more useful work from memory.

Analysis

The comparisons have different platform baselines and different per-user token rates. They therefore describe two distinct comparisons, not a single common-condition progression. Throughput per dollar and memory capacity also measure different things: the reported performance multiples should not be read as equivalent increases in memory. NVIDIA’s codesign description sets out its approach; these figures alone do not isolate each component’s contribution.

Editorial opinion

VERXERA opinion: NVIDIA’s November 5, 2026 account makes a specific case for looking beyond memory capacity alone. It reports up to 10x higher throughput per dollar for GB300 NVL72 versus H200 at 100 tokens per second per user, alongside roughly doubled HBM per GPU. For Vera Rubin NVL72 versus GB300, it reports up to 30x at 125 tokens per second per user with about 3% more HBM per GPU. Connecting those figures with NVIDIA’s description of coordinated hardware and software design explains the announcement’s significance: its argument concerns useful output relative to cost, not simply memory size. The strongest limitation is that these are company-reported, workload-specific maximum gains under different response-rate conditions. Deeper source reading is worthwhile for assessing those particular comparisons, rather than assuming universal gains.

This is VERXERA’s editorial interpretation, separated from the facts and analysis above. The reader keeps the final judgment.

Counterevidence and limits

This report rests on NVIDIA’s own announcement; the performance figures and publication date are the company’s statements.

Known uncertainties

How the reported maximum gains translate to workloads or per-user token rates beyond those named remains an open question.