HBM capacity per accelerator package: H100 80 GB; HGX B200 180 GB; B300 288 GB (HBM3E); Rubin Ultra ~1 TB HBM4E on a four-chiplet package (roadmap)
80 → 288 GB → ~1 TBobservedcontested
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | Blackwell capacity is SKU-specific and this row is the HGX B200 figure: NVIDIA's GB200 specification gives 372 GB per superchip (186 GB usable per Blackwell GPU), and 192 GB is the raw-stack boundary — three different boundaries for the same silicon (Chapter 7.2). Rubin Ultra's ~1 TB is package-level across four GPU compute chiplets, not memory on one GPU die; 2026 supply-chain reports suggest a revised dual-die layout — roadmap specs in flux. |
| Caveat | Announced figure (GTC 2025 roadmap: 1 TB HBM4e, four-die package) is no longer the supply chain's working assumption: TrendForce (2026-08-04) reports NVIDIA has been evaluating 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4 alternatives to the 12-Hi HBM4e baseline since early 3Q26, final spec unset; sell-side notes put the working SKU at 192–256 GB with the four-die package possibly deferred. NVIDIA has not issued a spec revision — treat as roadmap-in-flux, not a replacement spec. |
| As of | 2026-08 |
| Source | NVIDIA |
| Review | checking…review by 2026-10-14 · standard cadence |
| Recorded changes | last 2026-08-24 · 4 revisions tracked |
| Claim id | per-gpu-hbm-capacity-h100-80-b200-192-b300-288 |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.