GPU Dispatch · 2026-W36

Groq 3 LPX joins Vera Rubin NVL72 in production; NVIDIA posts agentic efficiency metrics; B200/B300 still OEM baseline

NVDA ecosystem — weekly dispatch (2026-W36)

### Material change this week

Groq 3 LPX in production — NVIDIA announced (2026-08-25) that NVIDIA Groq 3 LPX is in full production as an extension of the Vera Rubin NVL72 platform. Design intent: Rubin GPUs handle large-context processing; LPX handles latency-sensitive token generation for agentic workloads. Nebius is named as the first cloud provider to adopt Groq 3 LPX in its Token Factory.

NVIDIA also introduced Scale-In, a networking/infrastructure layer for AI factories based on BlueField-4 DPUs and DOCA — rack-scale integration, not a single-GPU SKU change.

Agentic efficiency data (vendor-measured) — NVIDIA's 2026-08-27 blog publishes early Vera Rubin NVL72 results on the SemiAnalysis AgentX workload (recorded real-world agentic coding sessions):

• Up to 30× higher throughput per megawatt vs GB300 NVL72
• Up to 35× lower cost per million tokens vs GB300 NVL72 (vendor figures)
• Results pending SemiAnalysis review; Vera CPU tool-calling contribution not yet reflected

Vera CPU — NVIDIA blog states the Vera CPU built for agents is shipping now.

Carry forward from prior weeks: Rubin GPU architecture (288 GB HBM4, 22 TB/s, NVLink 6, NVL72) and Microsoft installation milestone remain the deployment context.

### What remains the practical OEM / fleet baseline

HGX B200 — Eight Blackwell SXM GPUs, 1.4 TB aggregate memory, NVLink-5, up to 108 PFLOPS FP4 sparse at 8-GPU system level.

HGX B300 (Blackwell Ultra) — 2.1 TB total memory, 144 PFLOPS FP4 sparse at system level.

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

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

### Supply chain

Ecosystem news continues to stress HBM capacity (Samsung long-term allocation coverage, SK Hynix US fab buildout, CXMT HBM3E ramp). Operational constraint context only — see 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 platform story expanded from 'GPU + rack' to a split inference stack (Rubin context + LPX decode). Treat efficiency multiples as vendor-measured until independently verified.

Helios enters late-Q3 shipping window; MI355X stays catalog baseline; HBM supply tightens

Rivals — weekly dispatch (2026-W36)

### Helios at the shipping threshold

No new Instinct launch or catalog SKU this week. AMD CEO Lisa Su stated on the early-August earnings call that Helios is in production with initial shipments on track to begin later this quarter (Q3 ends Sep 30) and ramp through Q4 into 2027. We are now inside that window — watch for customer deployment confirmations, not just production status.

Helios recap (unchanged specs):

• MI455X: 432 GB HBM4, ~23.3 TB/s per GPU, CDNA 5
• 72-GPU rack: ~2.9 exaFLOPS FP4, 31 TB aggregate HBM4, UALoE fabric
• Catalog note: MI455X still not in `gpu_products`; baseline remains MI355X

### AMD Instinct (current catalog)

MI355X — 288 GB HBM3E, 8 TB/s, CDNA4, 1400W TBP, OAM / UBB 2.0.

MI325X — 256 GB HBM3e MI300 refresh.

Software — ROCm remains the adoption variable vs CUDA.

### HBM supply (ecosystem pressure)

This week's automated news highlights:

• Long-term HBM capacity commitments / fab investments (Samsung, SK Hynix)
• CXMT beginning HBM3E production (China-domestic memory path)
• SK Hynix Indiana fab ground-breaking (US HBM supply-chain localization)

Samsung's HBM page still lists HBM4E (up to 64 GB / 16H, up to 4 TB/s), HBM4, and HBM3E.

### Huawei Ascend

Ascend 910B/910C remain China-domestic 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.

Helios moves from 'ships end of Q3' language to actually being in that window. Until MI455X appears in our catalog and your channel, MI355X vs B200 remains the practical comparison.