GPU Dispatch · 2026-W40

NVIDIA lists an 84 GB RTX PRO 5500 as 'coming soon'; Micron drops the 2 GB GDDR7 density most RTX 50 cards use

NVDA ecosystem — weekly dispatch (2026-W40)

### New part: RTX PRO 5500 Blackwell

NVIDIA has put up a product page for the RTX PRO 5500 Blackwell, without a press release. From the vendor page:

• 84 GB GDDR7 with ECC, 1,398 GB/s memory bandwidth
• Up to 600 W maximum power
• MIG: up to two fully isolated instances (one at 84 GB, or two at 42 GB each)
• PCIe Gen 5 x16, up to 4× DisplayPort 2.1b, 3× 9th-gen NVENC and 3× 6th-gen NVDEC
• Designed for rack-mounted workstation deployment: active air cooling as standard, with a liquid-cooled variant in the RXM form factor
• Availability: "Coming Soon"

The vendor page gives no CUDA core count. Trade coverage reports a cut-down GB202 with 21,760 CUDA cores across 170 SMs, the same count as the GeForce RTX 5090. That figure is not on NVIDIA's page, so treat it as unverified.

The part sits between the RTX PRO 5000 and the catalogued RTX PRO 6000 Blackwell (96 GB GDDR7, 1,792 GB/s, 600 W, re-verified this run and unchanged). It is not in `gpu_products`, so it stays in briefs until it is catalogued.

### Memory supply — see the event log

Micron's product catalog now lists both of its 2 GB (16 Gb) GDDR7 parts as end-of-life: the 28 Gbps MT68A512M32DF-28 and the 32 Gbps MT68A512M32DF-32. That leaves 3 GB modules as Micron's only GDDR7 offering. Micron was a late, secondary RTX 50 supplier, so this narrows the supplier base for the 2 GB density used across most desktop RTX 50 cards rather than creating a shortage. It is recorded as an event under Supply chain and GPUs.

### Regulation — see the event log

The Information reported that China's MIIT asked ByteDance and Alibaba to report planned RTX PRO 5500 purchases and signalled that it intends to approve them. Reuters could not verify the report, and nothing has been approved. It is recorded as a rumoured event under Regulation. The US remote-access rule logged in August is still at draft stage, with no published text.

### Catalog re-verification

• H200 SXM — 141 GB HBM3e, 4.8 TB/s, up to 700 W configurable. Unchanged.
• HGX B200 — 1.4 TB total memory across 8 GPUs, 1.8 TB/s NVLink per GPU, 14.4 TB/s total, 72 PFLOPS FP8. Unchanged.
• RTX PRO 6000 Blackwell — 96 GB GDDR7, 1,792 GB/s, 600 W. 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 RTX PRO 5500 is notable less for its specs than for where it sits. It has 84 GB on a GDDR7 board with MIG and a rack form factor, and that is the profile of an inference card that can be sold where HBM parts cannot. The reported Chinese interest fits that reading, but for now it is only a report.

Alibaba's T-Head unveils the 216 GB Zhenwu V900 for Q1 2027; AMD MI355X re-verified unchanged

Rivals — weekly dispatch (2026-W40)

### Alibaba / T-Head: Zhenwu V900

At the 2026 Apsara Conference in Hangzhou on 22 September 2026, Alibaba's chip unit T-Head presented the Zhenwu V900:

• 216 GB of memory
• 1,200 GB/s inter-chip bandwidth, three times that of the predecessor Zhenwu M890
• Native FP8 and FP4
• T-Head claims 3× the performance of the M890
• More than 1,000 V900s can operate as a single supernode over T-Head's ICN Switch interconnect, and a cluster on Alibaba Cloud's network can scale to about 500,000 accelerators
• Not yet in production: mass production and sale are targeted for Q1 2027

Alibaba gave no fab partner, memory type or power figure, and made no direct comparison to an NVIDIA part. It tied the chip to a target of 20 GW of Alibaba Cloud data-centre capacity by 2032. The predecessor M890 is deployed with more than 650 enterprise customers. The V900 is recorded as an event under Competitors and GPUs.

### AMD: catalog verified

AMD's MI355X page responded this run, which clears the carry-forward flag from W39: 288 GB HBM3E, 8 TB/s, 1,400 W TBP, CDNA 4. Unchanged. Helios and MI455X have no new information since last week; AMD's guidance is still first rack shipments at the end of Q3, ramping through Q4 2026 and H1 2027.

### Huawei

Nothing new since the Atlas 960E SuperPoD and Ascend 960DT roadmap covered in W39. The catalogued Ascend 910B is unchanged.

### Memory suppliers

Samsung's HBM page is unchanged: HBM3E 12H at 36 GB, up to 1,180 GB/s and 9.2 Gbps; HBM4 at 36 GB, up to 3,300 GB/s and 13.0 Gbps; HBM4E at up to 64 GB and 4 TB/s in a 16-high stack at up to 16.0 Gbps.

---

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

The V900 is the second Chinese accelerator in a fortnight that is pitched on interconnect and cluster size rather than per-chip compute. Both it and Huawei's Ascend 960DT now target Q1 2027. Neither is shipping, so both stay out of the matrix.