GPU Dispatch · 2026-W42

A second CoWoS interposer source in New York; Samsung HBM4E reportedly clears NVIDIA qualification

NVDA ecosystem — weekly dispatch (2026-W42)

### Packaging: GlobalFoundries to make CoWoS interposers for TSMC

On 8 October, GlobalFoundries announced a $2 billion, multi-year agreement to provide manufacturing services to TSMC for silicon interposers used in CoWoS, including parts with embedded deep trench capacitors. Capacity is being added at GF's Malta, New York site. Volume production is due to start ramping in the first half of 2028. The initial term is five years, with a framework to expand. GF calls it the first U.S.-based source of these interposers.

The interposer is the silicon layer that connects a GPU die to its HBM stacks inside the package. The agreement covers interposers only, not assembly or test. This is recorded as an event under Supply chain and Chip journey.

### Memory: Samsung HBM4E reportedly qualified

KED Global (Hankyung) reported on 7 October, citing unnamed sources, that Samsung's 12-high HBM4E passed qualification with NVIDIA and major hyperscalers. Neither Samsung nor NVIDIA has confirmed it. No NVIDIA product, supply volume or production date was reported. Samsung's HBM page still lists HBM4E at up to 64 GB and 4 TB/s per 16-high stack on a 4 nm foundry base die, and does not state production status. This is recorded as a rumoured event.

### Market moves (in the event log, not restated)

NVIDIA took confirmed minority stakes in Doctolib and Genesis AI (13 October). The FT reported early talks to buy or invest further in Reflection AI (10 October). See Market moves.

### RTX PRO 5500: still "Coming Soon"

The vendor page is unchanged: 84 GB GDDR7 ECC, 1,398 GB/s, up to 600 W, two-way MIG, availability "Coming Soon".

### Catalog re-verification

• H200 SXM: 141 GB, 4.8 TB/s, up to 700 W configurable, 3,958 TFLOPS FP8 (sparse). Unchanged.
• HGX B200: 1.4 TB total memory, 1.8 TB/s NVLink per GPU, 14.4 TB/s total, 72 PFLOPS FP8 sparse, 144 PFLOPS FP4 sparse. Unchanged. HGX B300 and B200 are listed as "shipping now", and Vera Rubin as "ramping into full production".
• RTX PRO 6000 Blackwell: 96 GB GDDR7, 1,792 GB/s, 600 W, 4,000 FP4 AI TOPS. Unchanged.

Catalog unchanged.

---

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

The interposer is one step of CoWoS, not the whole of it. A second interposer source widens that step from 2028, but chip placement, HBM integration, substrates and test still limit packaged output. The deal adds redundancy to the packaging chain more than it adds near-term volume.

IDC: NVIDIA's China accelerator share fell to 34% in H1 2026, with Huawei at 28% and non-GPU accelerators outshipping GPUs

Rivals — weekly dispatch (2026-W42)

### China: IDC's first-half figures

IDC data, reported by Reuters on 13 October, covers China's AI accelerator market for January–June 2026, about 2.0 million units in total:

| Vendor | Units (H1 2026) | Share | |---|---|---| | NVIDIA | ~678,000 (−44% YoY) | 34% | | Huawei | ~572,000 (from just over 300,000) | 28% | | Alibaba T-Head | — | 7% | | Cambricon | — | 5% | | Baidu Kunlunxin | — | 4% | | Hygon | — | 3% |

Non-GPU accelerators (ASICs and NPUs) reached about 1.016 million units, up nearly 80%. That is the first time they outshipped GPUs (about 991,000, down 26%) in China. Nine of the eleven Chinese vendors IDC tracks grew shipments. IDC did not break out which NVIDIA products make up its share. NVIDIA did not comment.

These are the first independent figures since Huawei's 1 October claim that Ascend has passed NVIDIA in China. On units, H1 does not support that claim. The trend does: Huawei nearly doubled while NVIDIA fell 44%.

### Amazon / Decart (reported)

The WSJ reported, relayed by Reuters on 12 October, that Amazon is in talks to buy Decart for about $7 billion. Decart makes real-time video models and simulation environments. NVIDIA joined Decart's May round as an investor. Coverage frames the deal as support for Amazon's Trainium. That framing comes from reporting, not from Amazon. This is recorded as a rumoured event under Competitors.

### AMD

AMD's MI355X page timed out twice again this run. Catalog values (288 GB HBM3E, 8 TB/s, 1,400 W) carry forward from W40's verification. This run found no primary confirmation of MI450-family shipments.

---

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

IDC's figures are the first independent numbers behind Huawei's 1 October claim that it leads in China, and they do not support that claim for H1: NVIDIA still shipped more units. The more structural signal is the product mix. ASICs and NPUs passing GPUs by volume means the Chinese market is moving away from the GPU form factor, and away from CUDA with it.