GPU Dispatch · 2026-W39

No NVIDIA silicon change in W39; the week's NVIDIA fact is an S-1 that shows it guaranteeing a customer's lease

NVDA ecosystem — weekly dispatch (2026-W39)

### Material change this week

None on NVIDIA silicon. No SKU launch and no datasheet revision on tracked NVIDIA products. Re-verified against vendor pages this run:

• H200 SXM — 141 GB, 4.8 TB/s, up to 700 W configurable, 3,958 TFLOPS FP8. Unchanged.
• H200 NVL — 141 GB, 4.8 TB/s, up to 600 W configurable, 3,341 TFLOPS FP8. Not catalogued separately.
• HGX B200 / B300 — 1.4 TB and 2.1 TB total system memory, 1.8 TB/s NVLink per GPU, 14.4 TB/s total. Unchanged.

Catalog unchanged.

### Rubin on the HGX page

NVIDIA's HGX page continues to lead with Rubin: 288 GB HBM4 and 22 TB/s per GPU, 2.3 TB HBM4 / 176 TB/s per 8-GPU HGX Rubin NVL8 and Vera Rubin NVL8, with 28.8 TB/s NVLink switch bandwidth on 6th-gen NVLink. Rubin is still not in `gpu_products`, so it stays in briefs until catalogued.

### Capital structure — see the event log

The week's NVIDIA item is a filing, not a part. Nscale Ltd filed an S-1 on 18 September 2026 to list on NYSE as NSCL. It records NVIDIA as a holder of more than 5% of Nscale's share capital via the Series B, a $300,000,000 SAFE that converted at Series C completion, and — the part worth reading twice — a guarantee of up to $860,340,000 of Nscale's obligations under an October 2025 net lease in Ward County, Texas, against warrants for up to 9,476,700 Series B preferred shares at $0.01. Contingent default warrants tied to that guarantee survive the offering.

That is a different exposure from selling parts to an arms-length buyer: NVIDIA is standing behind the site lease of a company that buys its GPUs. The filing also describes over 66,000 Vera Rubin NVL72 GPUs deploying with Microsoft from late 2027. Recorded as an event under Market moves and Supply chain.

### Supply chain

On the memory side, a supply-chain tracker places SK hynix in shipment of 16-high 48 GB HBM4 for Vera Rubin — 33% more memory per placement than the 12-high 36 GB stack, without a second placement's power, area or interposer cost. SK hynix has not confirmed it, so it is logged as rumoured. Separately, reports of SK hynix producing memory at Intel's Ohio campus drew a same-day statement from SK hynix that nothing has been finalised. Both are in the event log under Supply chain.

---

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

Second consecutive quiet silicon week. The interesting movement is in how compute capacity gets financed — a lease guarantee on a customer's data centre is infrastructure risk carried on the vendor's balance sheet, which is a structure to watch as neocloud build-outs scale, not a spec change.

Huawei pulls Ascend 960DT forward three quarters and launches an optical SuperPoD; AMD's Helios is in full production; Enflame lists in Shanghai

Rivals — weekly dispatch (2026-W39)

### Huawei: the scale-up domain, not the die

At HUAWEI CONNECT 2026 in Shanghai on 17 September 2026, Huawei launched the Atlas 960E SuperPoD, which it describes as the industry's first SuperPoD built on near-packaged optics:

• Scales to 4,096 NPUs under unified memory addressing
• 8 EFLOPS FP8, 16 EFLOPS FP4
• 5,500 Hi-ONE optical units, orthogonal architecture, full liquid cooling
• Claimed 99.8% system availability, >550 kW lower power than the conventional approach

In the same keynote, rotating chairman David Wang moved the Ascend 960DT to Q1 2027 — three quarters earlier than planned — with the 960PR in Q3 2027, Ascend 970 in 2028 and Ascend 980 in 2029, on a one-generation-a-year cadence. The roadmap dates come from the keynote rather than a Huawei product page, and no Ascend 960 part is available yet.

The competitive read is not per-chip specification, where Huawei remains constrained, but the scale-up domain. Answering rack-scale NVLink with optically interconnected pods measured in thousands of NPUs is an attempt to compete at the unit large training and inference jobs are actually bought in — and it is the axis where process-node disadvantage matters least, because it is bought back with interconnect and power.

The catalogued Ascend 910B (64 GB HBM2e, 400 W) is two generations behind that conversation; it stays in the matrix because it is what is in `gpu_products`.

### AMD: Helios in full production

AMD's Instinct MI455X carries 432 GB HBM4 on a 2 nm process, with peak MXFP8 and MXFP4 up to 4× the MI355X. The Helios rack pairs it with EPYC Venice and Pensando networking at 4,600 CPU cores and 18,000 GPU compute units per rack. AMD says Helios is in full production, with shipments later in the quarter ramping into Q4 2026 and H1 2027, and has put MI500 in 2027 and MI600 in 2028 on the roadmap.

The catalogued MI355X (288 GB HBM3E, 8 TB/s, 1400 W, CDNA4) remains the AMD flagship in `gpu_products`. AMD's vendor pages did not respond on two attempts this run, so MI355X catalog values are carried forward unverified this week; MI455X and Helios figures above are from launch coverage, not an AMD page.

### China: Enflame lists

Tencent-backed Enflame began trading on Shanghai's STAR Market on 11 September 2026 under ticker 688801, raising about 6.12 billion yuan (~$911M) and closing its debut sharply higher. Tencent holds about 20% and accounted for 83.8% of 2025 revenue — a concentration worth naming, since it makes the listing as much a captive-supply arrangement as an independent competitive entry. Proceeds are earmarked for fifth- and sixth-generation parts.

### Memory suppliers

Samsung's HBM page now lists HBM4 at 36 GB, up to 3,300 GB/s and 13.0 Gbps per pin across 2,048 I/O, and HBM4E at up to 64 GB and 4 TB/s in a 16-high stack at up to 16 Gbps. The catalogued HBM3E 12H part (36 GB, up to 1,180 GB/s, 9.2 Gbps) is unchanged.

---

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

Three rivals moved in one week and none of them moved on the axis NVIDIA is usually compared on. Huawei competed on interconnect, AMD on memory capacity per rack, Enflame on capital. The common thread is that the contested unit is now the rack or the pod, not the accelerator.