GPU Dispatch · 2026-W34

Microsoft begins Vera Rubin install; NVIDIA publishes Rubin GPU architecture deep-dive; B200/B300 still OEM HGX baseline

NVDA ecosystem — weekly dispatch (2026-W34)

### Material change this week

Microsoft Vera Rubin installation — On 2026-08-21, NVIDIA publicly marked Microsoft reaching a Vera Rubin deployment milestone (NVIDIA social / trade coverage). Treat this as confirmation that hyperscaler installation is underway for NVL72-class systems — still allocation-gated, not broad OEM-channel GA for every enterprise buyer.

Rubin architecture documentation — NVIDIA's technical blog *Inside NVIDIA Rubin GPU Architecture* publishes GPU-level specs useful for engineering comparisons:

• Up to 288 GB HBM4 with up to 22 TB/s peak bandwidth (12-Hi stacks)
• Up to 50 petaflops NVFP4 via third-generation Transformer Engine
• 336B transistors, 224 SMs, 896 Tensor Cores; dual compute dies linked by NV-HBI
• NVLink 6 scale-up (3,600 GB/s to NVLink Switch), NVLink-C2C 1,800 GB/s, PCIe Gen6 x16
• Vera Rubin NVL72 rack: cable-free MGX trays, liquid cooling, hot-swappable NVLink switch trays
• NVIDIA claims up to 10× agentic throughput per unit energy vs Blackwell (vendor/internal workload)

NVIDIA's 2026-05-31 newsroom release remains the primary official production statement (full production ramp; production shipments starting this fall).

### What remains the practical OEM / fleet baseline

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.

HGX B300 (Blackwell Ultra) — Shipping alongside B200: eight Blackwell Ultra SXMs, 2.1 TB total memory, NVLink-5, 144 PFLOPS FP4 sparse at system level.

H200 — 141 GB HBM3e at 4.8 TB/s, up to 700W configurable TDP (SXM).

GB200 / Grace Blackwell — Module/rack-scale NVL for fleets without a Vera Rubin allocation.

### Supply chain

Rubin depends on TSMC advanced packaging and HBM4 from Samsung, SK Hynix, and Micron. Ecosystem news continues to stress HBM capacity buildouts — operational constraint context only. 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.

Rubin moved from 'racks at partners' to named hyperscaler install milestones plus official architecture docs. For fleets without allocation, B200/B300 and H200 remain the planning baseline.

Helios/MI455X still in production awaiting late-Q3 shipments; MI355X remains catalog CDNA4 baseline

Rivals — weekly dispatch (2026-W34)

### Quiet on new AMD silicon this week

No new Instinct SKU or Helios datasheet revision material to our catalog. Carry forward:

• MI455X / Helios — AMD launch blog (2026-07-23): 432 GB HBM4, ~23.3 TB/s memory bandwidth, ~40 PF FP4, 72-GPU Helios rack with EPYC Venice + Pensando/UALoE. Status: in production; shipments expected end of Q3, broader ramp Q4 / into 2027 (AMD + trade coverage).
• Catalog note: MI455X is still not in `gpu_products`. Shipping baseline remains MI355X.

### AMD Instinct (current catalog)

MI355X — 288 GB HBM3E, 8 TB/s, CDNA4, 1400W TBP, OAM / UBB 2.0. Still the open-ecosystem HBM3E density play for fleets not on Helios allocation.

MI325X — 256 GB HBM3e refresh of the MI300 line.

Software — ROCm (including the Aug 4 CDNA5/Helios developer blog) remains the adoption variable versus CUDA.

### Samsung HBM

Samsung's HBM page lists HBM4E (up to 64 GB / 16H, up to 4 TB/s), HBM4, and HBM3E (up to 36 GB 12H, up to 1,180 GB/s per stack). HBM4 feeds Rubin/Helios; HBM4E is the next-stack roadmap.

### Huawei Ascend

Ascend 910B/910C remain the China-domestic path under export-control constraints.

---

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

Honest quiet week on AMD launches. Rivalry is still rack-scale HBM4 (Helios vs Vera Rubin) vs shipping catalog SKUs (MI355X vs B200). Watch late-Q3 Helios shipment confirmation.