GPU Dispatch · 2026-W29

Vera Rubin in full production; HGX B200/B300 still the shipping Blackwell baseline

NVDA ecosystem — weekly dispatch (2026-W29)

### Material change this week

Vera Rubin full production — NVIDIA's 2026-05-31 newsroom release states Vera Rubin is ramping into full production with Taiwan ODMs and a broad MGX partner ecosystem (Dell, HPE, Lenovo, Supermicro, Foxconn, Quanta/QCT, Wistron, and others). Production shipments are set to begin starting this fall. Treat Rubin as a rack/POD-scale platform (NVL72 + Vera CPU + Spectrum-X Ethernet Photonics), not a drop-in SXM card replacement for existing HGX B200 fleets.

### What is shipping / generally available today

HGX B200 — Still listed as shipping: 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) — NVIDIA's HGX page now lists B300 as shipping alongside B200: eight Blackwell Ultra SXMs, 2.1 TB total memory, same NVLink-5 fabric numbers as B200, higher FP4 tensor throughput (144 PFLOPS sparse at system level).

H200 — First HBM3e Hopper GPU: 141 GB at 4.8 TB/s, up to 700W configurable TDP (SXM). Still the pragmatic path from H100 for memory-bound LLM inference.

GB200 / Grace Blackwell — Module/rack-scale NVL configurations remain the AI-factory form factor for current Blackwell deployments.

### Forward visibility (sourced)

NVIDIA's HGX product page now publishes preliminary HGX Rubin NVL8 / HGX Vera Rubin NVL8 tables (HBM4 aggregate memory, 6th-gen NVLink, NVFP4 system FLOPS). Specs are marked preliminary and subject to change — verify datasheets before any engineering commitment.

### Supply chain

Blackwell and Rubin platforms depend on advanced packaging (TSMC) and HBM from Samsung, SK Hynix, and Micron. See the live supply-chain graph (F3) for foundry/memory vendor links.

---

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

For fleets deploying in the next 1–2 quarters, B200/B300 and H200 remain the practical SKU baseline. Rubin's significance is rack-scale agentic throughput and HBM4/NVLink-6 — plan cooling, fabric, and allocation lead times, not a single-GPU swap.

MI355X still the open-ecosystem HBM3E density play; HBM4 race intensifies around Rubin

Rivals — weekly dispatch (2026-W29)

### AMD Instinct

MI355X — 288 GB HBM3E, 8 TB/s peak bandwidth, CDNA4, 1400W TBP, OAM. Platform materials describe eight OAMs on UBB 2.0 with 2.3 TB aggregate HBM3E and Infinity Fabric mesh. Firmware continues to receive platform maintenance releases (AMD docs through May 2026) — treat as operational maturity, not a new architecture drop.

MI325X — 256 GB HBM3e refresh of the MI300 line; still relevant for clusters deployed before MI350/MI355 ramps.

Software — ROCm compatibility for your model stack remains the gating adoption variable versus CUDA.

### Samsung (HBM supply)

Samsung Semiconductor's HBM page lists HBM3E (up to 36 GB 12H, up to 9.2 Gbps/pin, up to 1,180 GB/s per stack) and HBM4 (up to 3,300 GB/s, 36 GB, 12H). HBM4 is the memory generation NVIDIA associates with Vera Rubin platforms — allocation and qualification timelines matter more than peak brochure bandwidth for cluster ramps.

### Huawei Ascend

Ascend 910B/910C serve China-domestic training/inference where NVIDIA datacenter GPUs face export restrictions. Public English specs remain limited; verify regional availability before procurement planning.

### Ecosystem news context (this week)

Automated news ingest surfaces continued coverage of US export controls, China-domestic accelerator narratives, and TSMC advanced-packaging capacity — relevant as deployment constraints, not as investment signals.

---

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 still memory-capacity + software-stack + regional-availability. MI355X leads single-GPU HBM3E capacity vs a B200 GPU; NVIDIA leads on CUDA/HGX fleet depth. HBM4 qualification is the shared bottleneck for next-gen racks.