NVIDIA published a technical post on 5 November 2026 describing full-stack memory co-design for Vera Rubin NVL72. The memory itself is a modest step up: HBM capacity is about 3% higher, at 288 GB of HBM4 per GPU with 19.2 TB/s of bandwidth.
The headline numbers need careful reading, because NVIDIA published two different "30x" claims in two different places — and they are not the same claim. The newsroom post says agentic throughput per dollar is up to 30x, at 125 tokens per second per user, on the SemiAnalysis AgentX workload running DeepSeek V4 Pro. The technical blog says up to 30x per megawatt, at 160 tokens per second per user. One is about cost, the other about power — merging them into a single "30x faster" line would be wrong.
Independent testing tells a smaller story. SemiAnalysis's own tests found up to 7x per megawatt, and the cost advantage falls to 1.4–3x at the 60–100 tokens per second that most providers actually serve. All of NVIDIA's figures are the vendor's own, vendor-run numbers.
Key facts
| Item | Details |
|---|---|
| Post | NVIDIA technical post, 5 November 2026 — full-stack memory co-design for Vera Rubin NVL72 |
| Memory | 288 GB HBM4 per GPU, 19.2 TB/s — HBM ~3% up |
| Claim 1 (newsroom) | Up to 30x agentic throughput PER DOLLAR at 125 tok/s/user (SemiAnalysis AgentX, DeepSeek V4 Pro workload) |
| Claim 2 (technical blog) | Up to 30x PER MEGAWATT at 160 tok/s/user |
| SemiAnalysis's own tests | Up to 7x per megawatt; cost advantage 1.4–3x at 60–100 tok/s most providers serve |
| Status of NVIDIA figures | Vendor's own, vendor-run — not independently verified |
What this means
The real story here is memory co-design, not a single speed number. Vera Rubin's HBM step-up is small (~3%), and the big gains NVIDIA claims come from the full stack — memory, interconnect and software working together on agentic workloads. The two "30x" figures describe different dimensions (cost vs power) at different speeds (125 vs 160 tokens per second per user), so any summary that quotes "30x" without saying which one is misleading.
SemiAnalysis's independent numbers are the useful anchor: up to 7x per megawatt, and a 1.4–3x cost advantage at the speeds providers actually serve. That is still a real improvement — just not the headline figure.
A note on verification
The memory figures and both 30x claims are as reported from NVIDIA's 5 November 2026 technical post and newsroom post; the two claims are kept separate here exactly as published. The SemiAnalysis test results are SemiAnalysis's own measurements, as reported.
FAQs
What did NVIDIA announce on 5 November 2026?
A technical post on full-stack memory co-design for Vera Rubin NVL72, with 288 GB HBM4 per GPU at 19.2 TB/s — HBM about 3% up.
What are the two "30x" claims?
The newsroom post claims up to 30x agentic throughput per dollar at 125 tokens per second per user; the technical blog claims up to 30x per megawatt at 160 tokens per second per user. They are different claims and must not be merged.
What did independent testing find?
SemiAnalysis's own tests found up to 7x per megawatt, with the cost advantage falling to 1.4–3x at the 60–100 tokens per second most providers serve.
Are NVIDIA's figures independently verified?
No. They are NVIDIA's own vendor-run numbers; the SemiAnalysis figures are the independent measurements in this story.
Source: NVIDIA technical post and newsroom post, 5 November 2026; SemiAnalysis AgentX tests. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!