← LIVE TERMINALALL EDITIONSRIVALS SECTION

GPU Dispatch · 2026-W38

DISPATCH 2026-W38 · published 9/14/2026 UTC

No new NVIDIA silicon this week; the week's NVIDIA news was capital, not chips

NO MATERIAL CHANGE

No NVIDIA SKU launch or datasheet revision in ISO week 38. Vendor pages re-checked for H200 and HGX B200 show no spec change. The week's NVIDIA news was deals under negotiation, recorded in the event log rather than repeated here. Memory allocation remains the binding constraint on accelerator ramps.

NVDA ecosystem — weekly dispatch (2026-W38)

### Material change this week

None on NVIDIA silicon. No SKU launch and no datasheet revision on tracked NVIDIA products. Re-checked this run:

  • H200 SXM — 141 GB HBM3e, 4.8 TB/s, up to 700 W configurable. Unchanged.
  • HGX B200 / B300 — 1.4 TB and 2.1 TB total system memory. Unchanged.

Catalog unchanged.

### Context from NVIDIA's HGX page

NVIDIA's HGX page now leads with Rubin: 288 GB HBM4 and 22 TB/s per GPU, and 2.3 TB HBM4 / 176 TB/s per 8-GPU HGX Rubin NVL8 and Vera Rubin NVL8 system. Rubin is not in `gpu_products`; it stays in briefs until it is catalogued.

### Capital moves — see the event log

This week's NVIDIA items are structural, not silicon, and are recorded as events under Market moves: the Hugging Face acquisition (announced 2026-09-03, not closed), reported talks on a Thinking Machines Lab investment, and reported talks on anchoring Anthropic's IPO. The last two are rumored. None changes a spec.

### Supply chain

Reported changes to HBM supply — Micron's capacity expansion and Samsung's long-term agreements — are in the event log under Supply chain.

One operational item that is not an event: Micron's two Taiwan unions (about 10,000 members) reported about 80% support for strike action in a preliminary survey over bonus terms. Mediation was scheduled for late August and mid-September; no work stoppage has been reported.

---

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

ANALYSIS

Quiet silicon week. HBM allocation, not GPU design, is still what paces Rubin and Blackwell Ultra fleet growth. The Micron Taiwan labor dispute is the near-term operational risk to watch, since Micron is one of the three merchant HBM suppliers.

[NVIDIA H200][NVIDIA HGX platform (B200, B300, Rubin NVL8)][NVIDIA Blog: NVIDIA to Acquire Hugging Face (2026-09-03)][TrendForce: Micron Taiwan strike vote possible in September (2026-08-28)]

No new rival SKU this week; Enflame lists in Shanghai and Google sizes its TPU business

NO MATERIAL CHANGE

No AMD, Huawei or Samsung accelerator launch or catalog change in ISO week 38. AMD Helios (MI455X) remains inside AMD's stated end-of-Q3 shipping window. Enflame listed on Shanghai's STAR Market on 2026-09-11, raising 6.12B yuan for its 5th- and 6th-generation chips. Google Cloud said on 2026-09-08 that its TPU business is more than twice the next-largest hyperscaler's. CXMT's reported HBM3E start is in the event log.

Rivals — weekly dispatch (2026-W38)

### AMD — no catalog change

No new Instinct launch or catalog SKU. Helios (72x MI455X, 432 GB HBM4 per GPU) remains inside AMD's stated end-of-Q3 2026 shipping window. Q3 closes Sep 30. MI455X is still not in `gpu_products`; the catalog baseline stays MI355X (288 GB HBM3E, 8 TB/s, 1400 W, OAM). The MI355X vendor page did not load during this run, so its specs were carried from the catalog rather than re-verified.

### Huawei Ascend — no change

No new Ascend availability announcement found this week. Ascend 910B / 910C remain the catalogued parts.

### China accelerators — Enflame lists

Enflame, one of China's four so-called GPU dragons alongside Cambricon, MetaX and Moore Threads, debuted on Shanghai's STAR Market on 2026-09-11. It raised 6.12B yuan (about $912M) at 142.18 yuan per share and closed about 179% above the offer price. Proceeds are earmarked for R&D and industrialization of its 5th- and 6th-generation AI training and inference chips. Tencent is its lead strategic backer. This is a capital event, not a product; no Enflame part is catalogued.

### Custom silicon — Google TPU

At Goldman Sachs Communacopia on 2026-09-08, Google Cloud's Thomas Kurian said the TPU business is more than twice the size of the next-largest hyperscaler's. Google now sells TPUs three ways: cloud rental in its own facilities, TPU systems deployed in customer data centers (Anthropic named), and through a Blackstone joint venture. This is a vendor claim, not an independent measure.

### Memory

CXMT's reported start of small-batch HBM3E production is in the event log under Supply chain and Competitors.

---

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

ANALYSIS

Rivalry narrative unchanged on silicon. What moved is who sells accelerators and how they are funded: Google now ships TPU systems into customer data centers, and China's domestic GPU makers are raising public capital for next-generation parts. Neither changes a catalog spec this week.

[AMD Helios launch blog][AMD Instinct MI355X][Huawei Ascend 910B][Forkast: Enflame STAR Market debut (2026-09-11)][IEEE ComSoc: Google's TPU business outpaces rivals (2026-09-08)]

Model spotlights

B200-SXM-192GB

The shipping Blackwell baseline on HGX B200 while Rubin (288 GB HBM4 per GPU on NVIDIA's HGX page) ramps.

  • 192 GB HBM3e per GPU (catalog)
  • HGX B200: 1.4 TB total memory across 8 GPUs
  • 5th-gen NVLink — 1.8 TB/s GPU-to-GPU
  • FP4/MX formats on Blackwell Tensor Cores
ANALYSIS

HBM allocation, not GPU supply, is the near-term constraint on fleet growth. Validate power and cooling at HGX/MGX system level.

GB200-SUPERCHIP

Grace + Blackwell module for NVL racks — the generation Rubin rack-scale efficiency claims are measured against.

  • 384 GB aggregate HBM3e per superchip module (catalog)
  • Grace CPU for host-side data movement
  • NVL72/NVL36 with NVLink fabric
ANALYSIS

Rubin efficiency figures are rack-versus-rack vendor claims, not a direct comparison to this catalog SKU.

H200-SXM-141GB

The HBM3e Hopper part — a path up from H100 without a Blackwell or Rubin retool. Vendor specs re-verified this run.

  • 141 GB HBM3e at 4.8 TB/s
  • Up to 3,958 TFLOPS FP8 sparse (H200 SXM)
  • Up to 700 W configurable TDP (SXM)
  • 900 GB/s NVLink; MIG up to 7 instances
ANALYSIS

NVIDIA still labels these preliminary specifications. Validate published LLM throughput on your model size and batching.

MI355X-OAM-288GB

The catalogued AMD flagship until MI455X and Helios reach channels, with Helios inside its end-of-Q3 shipping window.

  • 288 GB HBM3E at 8 TB/s peak
  • CDNA4; 1400 W TBP; OAM / UBB 2.0
  • Successor: MI455X 432 GB HBM4 + Helios 72-GPU rack (not catalogued)
ANALYSIS

Specs carried from the catalog; the vendor page did not load during this run. ROCm readiness remains the gating item against CUDA fleets.

SAMSUNG-HBM3E-12H

HBM stacks gate every accelerator generation, and reported spot prices now sit far above contract prices.

  • HBM3E: up to 36 GB (12H), up to 1,180 GB/s per stack, 9.2 Gbps per pin
  • HBM4: 36 GB (12H), up to 3,300 GB/s
  • HBM4E: up to 64 GB / 16H, 4 TB/s per stack (Samsung HBM page, re-verified this run)
  • Reported: 36 GB HBM3E at about $2,100 spot, roughly 4-5x the long-term-agreement price
ANALYSIS

The spread between spot and contract HBM pricing means allocation under long-term agreements, not list availability, decides who ramps. Samsung's reported LTA commitments are in the event log.

Comparison matrix

DimensionB200-SXM-192GBH200-SXM-141GBMI355X-OAM-288GBASCEND-910B
GPU memory (GB)19214128864
Memory bandwidth (GB/s)800048008000
Memory typeHBM3eHBM3eHBM3EHBM2e
ArchitectureBlackwellHopperCDNA4Ascend
TDP / TBP (W)10007001400400
Form factorSXMSXMOAMOAM
ANALYSIS

Matrix stays on catalog SKUs. Rubin (288 GB HBM4) and MI455X (432 GB HBM4) remain in briefs only until they enter gpu_products.

[NVIDIA HGX B200][NVIDIA H200][AMD MI355X][Huawei Ascend 910B]