GPU Dispatch · 2026-W31

Vera Rubin customer shipments reported; HGX B200/B300 remain the general-availability Blackwell baseline

NVDA ecosystem — weekly dispatch (2026-W31)

### Material change this week

Vera Rubin customer deliveries (reported) — Industry outlets (The Elec, Electronic Times) report that NVIDIA has begun shipping Vera Rubin AI platform / NVL72 racks to cloud customers, with named recipients including Google, Microsoft, Oracle, and CoreWeave. Treat these as press reports of early customer deployments, not a claim that Rubin is broadly available through OEM channels.

NVIDIA's own 2026-05-31 newsroom release remains the primary official source: Vera Rubin is ramping into full production with a broad MGX partner ecosystem, and production shipments are set to begin starting this fall. That official availability language has not been superseded by a newer NVIDIA product-page GA announcement as of this dispatch.

### 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) — Listed shipping alongside B200 on NVIDIA's HGX page: eight Blackwell Ultra SXMs, 2.1 TB total memory, NVLink-5 fabric, higher FP4 tensor throughput (144 PFLOPS sparse at system level).

H200 — 141 GB HBM3e at 4.8 TB/s, up to 700W configurable TDP (SXM). Still the pragmatic upgrade path from H100 for memory-bound LLM inference without a Blackwell/Rubin retool.

GB200 / Grace Blackwell — Module/rack-scale NVL configurations remain the AI-factory form factor for current Blackwell deployments that are not yet on Vera Rubin NVL72 allocation.

### Forward visibility

Vera Rubin is a rack/POD-scale platform (NVL72 + Vera CPU + Spectrum-X Ethernet Photonics / HBM4), not a drop-in SXM card swap for existing HGX B200 fleets. Plan cooling, fabric, and allocation lead times accordingly.

### Supply chain

Blackwell and Rubin platforms depend on advanced packaging (TSMC) and HBM from Samsung, SK Hynix, and Micron. Automated news this week continues to emphasize long-term HBM supply agreements between Korean memory makers and US AI platforms — relevant as capacity constraints, not as investment guidance. 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.

Until NVIDIA publishes a clearer GA/OEM channel statement, treat Rubin as early hyperscaler rollout. For fleets deploying in the next 1–2 quarters without a Rubin allocation, B200/B300 and H200 remain the practical SKU baseline.

AMD launches Instinct MI455X + Helios rack (HBM4); MI355X remains the shipping CDNA4 baseline

Rivals — weekly dispatch (2026-W31)

### Material change this week — AMD Helios / MI455X

AMD's 2026-07-23 Helios launch blog states:

• Instinct MI455X is the accelerator: 432 GB HBM4, 23.3 TB/s memory bandwidth, ~40 PF FP4 AI compute (per AMD Helios rackscale infographic).
• Helios co-designs 6th Gen EPYC "Venice" CPUs + MI455X + Pensando networking / UALoE, connecting 72 GPUs in one scale-up domain with 260 TB/s scale-up bandwidth.
• Rack-level claims vs NVIDIA Vera Rubin NVL72 (AMD-modeled): up to 15% more AI compute, 50% more HBM capacity, 50% more scale-out bandwidth — verify on your workload; these are vendor figures with footnotes.
• Graphic status on the launch page: IN PRODUCTION TODAY; broader partner/cloud adoption language (OpenAI, Meta, Microsoft, Oracle listed as at-scale AI partners on the blog).

Catalog note: MI455X is not yet in our `gpu_products` table — this week's dispatch covers it in the rivals brief only. Shipping baseline in our catalog remains MI355X.

### AMD Instinct (current catalog)

MI355X — 288 GB HBM3E, 8 TB/s, CDNA4, 1400W TBP, OAM / UBB 2.0 eight-GPU mesh with 2.3 TB aggregate HBM3E. Still the open-ecosystem HBM3E density play for fleets not yet on MI400/Helios allocation.

MI325X — 256 GB HBM3e refresh of the MI300 line.

Software — ROCm remains the adoption variable versus CUDA.

### Samsung (HBM supply)

Samsung Semiconductor's HBM page lists HBM3E (up to 36 GB 12H, up to 1,180 GB/s per stack) and HBM4 (up to 3,300 GB/s, 36 GB). HBM4 is the memory generation for both Vera Rubin and MI455X/Helios — allocation timelines gate rack ramps.

### Huawei Ascend

Ascend 910B/910C serve China-domestic deployments under export-control constraints. Public English specs remain limited.

---

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

Rivalry moved from single-GPU HBM3E density (MI355X vs B200) to rack-scale HBM4 platforms (Helios vs Vera Rubin). Until MI455X is in our catalog and generally available through your channel, compare on software stack readiness and allocation — not brochure FLOPS alone.