← LIVE TERMINALALL EDITIONSRIVALS SECTION

GPU Dispatch · 2026-W45

DISPATCH 2026-W45 · published 11/2/2026 UTC

Vera Rubin NVL72 reaches general availability on Oracle Cloud; NVIDIA's own pages now lead with Rubin

UPDATED

Oracle said on 27 October that Vera Rubin NVL72 systems and Vera CPUs are generally available on OCI, with clusters to scale to several gigawatts and HGX Rubin NVL8 to follow without a date. NVIDIA's B200 product page now leads with Rubin specifications (288 GB HBM4, 22 TB/s per GPU) while still footnoting HGX B300 and B200 as shipping.

NVDA ecosystem — weekly dispatch (2026-W45)

### Rubin on Oracle Cloud

At Oracle AI World on 27 October, Oracle said Vera Rubin NVL72 systems and Vera CPUs are generally available on Oracle Cloud Infrastructure and that it plans to scale Vera Rubin clusters to several gigawatts. Smaller HGX Rubin NVL8 servers will follow; no date was given.

### Vendor page

NVIDIA's HGX page now lists a Rubin GPU at 288 GB HBM4 and 22 TB/s, with HGX Rubin NVL8 at 2 TB HBM4 and 130 TB/s system-level. A footnote states HGX B300 and HGX B200 are shipping now.

### Taiwan export-control plan

On 16 October Taiwan's economy minister said AI servers and advanced chips would be added to the controlled goods list under the Foreign Trade Act. It is a plan: an amendment is required and no start date or penalties are set. Taiwan is a major point in the server-assembly chain.

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

ANALYSIS

A cloud GA is a distribution milestone, not a spec change: the B200 and H200 catalog rows are unchanged. The B200 page no longer prints B200-specific figures in the section that was checked, so the catalog values are carried forward rather than re-confirmed.

[Oracle AI World: Google TPUs on OCI, Vera Rubin NVL72 GA][NVIDIA HGX B200][Taiwan plans to criminalise AI chip smuggling to China]

Oracle will host Google TPUs from 2027; China drafts export curbs on 40 technologies

UPDATED

Oracle announced it will offer Google TPUs on OCI in calendar 2027, the first hyperscaler outside Google to do so, while keeping NVIDIA systems. Separately, China's commerce and science ministries published a draft catalogue revision covering 40 export-restricted technologies, 23 of them AI, plus some chip-design items.

Rivals — weekly dispatch (2026-W45)

### Google TPUs on Oracle Cloud

Oracle and Google Cloud announced TPUs will be available on OCI in calendar 2027 with a multi-year ramp. Google says TPU 8t delivers nearly 3x the compute per pod versus the previous generation. No financial terms were disclosed.

### China's draft export catalogue

Published 16 October by the Ministry of Commerce and the Ministry of Science and Technology: 40 technologies, including 23 in AI (models of 2 trillion or more parameters could be restricted), several chip-design technologies for AI processing and memory chips, 12 solar and 5 biomedicine items. It is a draft with no effective date.

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

ANALYSIS

The TPU deal widens where non-NVIDIA accelerators can be rented, but is dated 2027 and undisclosed in value. China's draft is not in effect; its chip-design scope was described only generally in the coverage read.

[Oracle AI World: Google TPUs on OCI, Vera Rubin NVL72 GA][Oracle brings Google TPUs to Oracle Cloud Infrastructure][China says it will curb tech exports, including AI (AP)]

Model spotlights

B200-SXM-192GB

The shipping Blackwell baseline, and a CoWoS-packaged part whose interposer supply gains a second source from 2028.

  • 192 GB HBM3e per GPU at 8 TB/s (catalog)
  • HGX B200: 1.4 TB total memory, 1.8 TB/s NVLink per GPU, 14.4 TB/s total
  • 72 PFLOPS FP8 / 144 PFLOPS FP4 sparse per HGX system
  • Packaging: TSMC CoWoS. GlobalFoundries will add U.S. interposer capacity for CoWoS from H1 2028
ANALYSIS

Catalog values carried forward: NVIDIA's page now leads with Rubin figures and the section checked did not print B200-specific numbers. HGX B200 is still footnoted as shipping.

ASCEND-910B

IDC puts Huawei at about 572,000 accelerators and 28% of China in H1 2026, nearly double a year earlier.

  • 64 GB HBM2e, 400 W (catalog)
  • Software: CANN, plus DeepSeek's open-source Ascend libraries and TileLang adapter (W41)
  • IDC: non-GPU accelerators outshipped GPUs in China for the first time in H1 2026
ANALYSIS

IDC's unit figures are independent. They show Huawei growing fast but still behind NVIDIA in H1, which contradicts Huawei's own claim of leadership. Carried forward from W42; not re-verified this run.

H200-SXM-141GB

The HBM3e Hopper part, and the chip at the centre of US licensing policy for China while IDC shows NVIDIA's China share falling.

  • 141 GB HBM3e, 4.8 TB/s
  • Up to 700 W configurable
  • 3,958 TFLOPS FP8 (sparse)
  • Vendor page re-verified this run
ANALYSIS

Unchanged; NVIDIA's H200 page re-verified this run (141 GB, 4.8 TB/s, up to 700 W, 3,958 TFLOPS FP8 sparse).

RTX-PRO-6000-BLACKWELL

NVIDIA's GDDR7 tier with MIG, which sits outside the HBM constraint. Its sibling, the RTX PRO 5500, is still 'Coming Soon'.

  • 96 GB GDDR7 ECC, 1,792 GB/s, 600 W, 4,000 FP4 AI TOPS (re-verified)
  • RTX PRO 5500: 84 GB, 1,398 GB/s, up to 600 W, two-way MIG, 'Coming Soon'
ANALYSIS

Unchanged. Carried forward from W42; not re-verified this run.

MI355X-OAM-288GB

The catalogued AMD flagship, ahead of the MI450-family / Helios ramp.

  • 288 GB HBM3E, 8 TB/s, 1,400 W TBP, CDNA 4 (catalog)
  • Successor MI455X in Helios racks per AMD guidance. No primary shipment confirmation this run
ANALYSIS

AMD's page was unreachable this run, so values carry forward from W40's verification. Carried forward from W42; not re-verified this run.

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

H200 values re-verified this run. B200 values carried forward (vendor page now leads with Rubin). AMD MI355X values carry forward from W40; AMD's page timed out again. Ascend 910B carried forward.

[NVIDIA HGX B200][NVIDIA H200][AMD Instinct MI355X][Huawei Ascend / Atlas 960E]