GPU Dispatch · 2026-W32

NVIDIA confirms Vera Rubin NVL72 racks live at cloud partners; B200/B300 still the OEM HGX baseline

NVDA ecosystem — weekly dispatch (2026-W32)

### Material change this week

Official partner confirmation — NVIDIA's Vera Rubin blog states that Vera Rubin NVL72 production is ramping with racks running at partners CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius, across a supply chain spanning 350+ factory sites in 30 countries. This upgrades last week's press-report framing: the partner-live claim is now on NVIDIA's own blog.

Tokens/MW metric — CoreWeave's first DeepSeek-R1 benchmark on Vera Rubin NVL72, published via NVIDIA, reports ~10x tokens per second per megawatt versus Grace Blackwell NVL72. Treat as a partner/vendor workload result, not an independent audit.

Related NVIDIA posts on the same blog cluster: Google Cloud A5X instances on Vera Rubin NVL72; Nebius receiving first NVL72 systems in Finland; Spectrum-6 / Spectrum-X Ethernet Photonics called out for scale-out.

### What is shipping / generally available today

HGX B200 — Eight Blackwell SXM GPUs, 1.4 TB aggregate GPU memory, 5th-gen NVLink (1.8 TB/s GPU-to-GPU, 14.4 TB/s total), up to 108 PFLOPS FP4 sparse at the 8-GPU system level per NVIDIA's HGX table.

HGX B300 (Blackwell Ultra) — Shipping alongside B200 on NVIDIA's HGX page: eight Blackwell Ultra SXMs, 2.1 TB total memory, NVLink-5, higher FP4 system throughput (144 PFLOPS sparse).

H200 — 141 GB HBM3e at 4.8 TB/s, up to 700W configurable TDP (SXM). Still listed as available for Hopper-generation LLM/HPC memory-bound workloads.

GB200 / Grace Blackwell — Module/rack-scale NVL remains the current AI-factory form factor for fleets not yet on Vera Rubin NVL72 allocation.

### Forward visibility

Vera Rubin remains a rack/POD-scale platform (NVL72 + Vera CPU + NVLink 6 + Spectrum-6 / HBM4), not a drop-in SXM card swap for HGX B200 fleets. Early access is concentrated at named cloud partners; OEM channel breadth is still the open question for enterprise on-prem buyers.

### Supply chain

Blackwell and Rubin platforms depend on advanced packaging (TSMC) and HBM from Samsung, SK Hynix, and Micron. Automated news continues to stress HBM capacity constraints and long-term supply pacts — relevant as deployment bottlenecks, not investment signals. See the live supply-chain graph (F3).

---

Technical research for informational purposes only. Not financial advice. No investment recommendations. Specs sourced from vendor documentation; verify before engineering or procurement decisions.

The metric that matters for power-constrained AI factories is shifting to tokens per megawatt. For fleets without a Rubin allocation this quarter, B200/B300 and H200 remain the practical procurement baseline.

Helios/MI455X rivalry settles in; MI355X still the catalog CDNA4 SKU; Samsung posts HBM4E

Rivals — weekly dispatch (2026-W32)

### AMD Instinct / Helios (follow-through)

AMD's 2026-07-23 Helios launch remains the rivalry headline versus Vera Rubin NVL72:

• MI455X: 432 GB HBM4, 23.3 TB/s memory bandwidth, ~40 PF FP4 (AMD Helios materials)
• Helios: 72 GPUs + EPYC "Venice" + Pensando / UALoE; 31 TB HBM4 rack capacity; 260 TB/s scale-up bandwidth
• AMD-modeled rack claims vs Vera Rubin NVL72: up to 15% more AI compute, 50% more HBM, 50% more scale-out — vendor figures with footnotes

This week brings technical unpacking (package die stack, CDNA5 / ROCm ecosystem blogs) rather than a new product announcement. Catalog note: MI455X is still not in `gpu_products` — covered in brief only. Shipping catalog SKU remains MI355X (288 GB HBM3E, 8 TB/s, CDNA4, 1400W TBP).

Software — ROCm remains the adoption variable versus CUDA for open-ecosystem fleets.

### Samsung (HBM supply)

Samsung Semiconductor's HBM page now surfaces three generations relevant to AI racks:

• HBM3E — up to 36 GB (12H), up to 1,180 GB/s per stack (Blackwell / MI355X era)
• HBM4 — up to 36 GB, up to 3,300 GB/s (Vera Rubin / MI455X era)
• HBM4E — up to 64 GB (16H), up to 4 TB/s, up to 16 Gbps/pin (roadmap / next-gen stacks)

HBM allocation timelines remain the shared bottleneck across NVIDIA and AMD rack ramps.

### Huawei Ascend

Ascend 910B/910C serve China-domestic deployments under export-control constraints. No material English-language SKU revision surfaced this week.

---

Technical research for informational purposes only. Not financial advice. No investment recommendations. Specs sourced from vendor documentation; verify before engineering or procurement decisions.

Rivalry is rack-scale HBM4 (Helios vs Vera Rubin) plus HBM supply depth. Until MI455X is catalogued and channel-available, compare on ROCm readiness and allocation — not brochure FLOPS alone.