Revision History

Added configuration-specific HGX Rubin NVL8 specifications and comparisons with the NVL72 product tables

Update: added an official video.

Correction: Mission Control section rewritten to match NVIDIA's Mission Control 2.5 release notes, which give Vera Rubin NVL72 only early access bring-up support without Run:ai or the recovery engine

Auto-saved before update

Correction: Feynman TSMC A16 node was stated as an NVIDIA roadmap disclosure without a source (it is a press report; now attributed to TrendForce/Chosun Biz); Q1 FY27 earnings described as upcoming or by analyst consensus, replaced with NVIDIA's reported results

Correction: production shipments began August 2026 (not 'second half of 2026'); TrendForce April 2026 report misdescribed (it reported a dual-die Rubin Ultra plan); Kyber/NVL576 naming updated to NVIDIA's GTC 2026 scheme (Kyber NVL144, NVL576 = eight 72-GPU racks)

Internal links: nvidia_spectrum_6, nvidia_feynman, space_based_data_centers

Improve internal links (2026-10-13)

Correction: GB200 pairs two dual-die B200 GPUs with one Grace CPU

Correction: superchip pairs one Vera CPU with two Rubin GPUs (four dies), not two GPU dies; Vera is NVIDIA's first CPU with fully custom cores, not its second fully custom CPU; replaced an unsourced "first CPU with native FP8" claim and an unsourced x86 60-70% bandwidth comparison with NVIDIA-sourced figures. Update: BlueField-4 as the Sentry host in the Open Agent Safety Platform.

Verification fix (V9): NVIDIA GB300 NVL72 page showed up to 21 TB through 2025-11-12, 20 TB from 2025-11-13

Correction: attribute Rubin HBM4 and NVLink bandwidth to CES 2026 (22 TB/s, 1,580 TB/s, 3.6/260 TB/s) vs NVIDIA's September 2026 product page (19.2 TB/s, 1,400 TB/s, 3/216 TB/s); fix per-GPU FP8/FP16 figures (FP16/BF16 is 4 PF per GPU, not 8); drop unsourced 6.4 GT/s arithmetic; source the NVL144-to-NVL72 rename to CES 2026 instead of 'late December 2025'

Correction: GB300 NVL72 memory is 20 TB HBM3e (NVIDIA listed up to 21 TB in 2025; 20.7 TB is the Vera Rubin NVL72 HBM4 figure); GB200 NVL72 is 13.4 TB (13.5 TB in NVIDIA's 2024 spec); both racks list 576 TB/s; GTC Taipei keynote was June 1 Taipei time, with NVIDIA's releases dated May 31, 2026

Correction: fixes from an independent fact-check pass (see /corrections)

Add CUDA Toolkit 13.4 Rubin developer preview (compute capability 10.7 / SM_107) and CDMM default, plus the d-Matrix NVLink Fusion MGX collaboration

Correction: Groq 3 LPX SRAM is 500 MB per LPU per NVIDIA, not 512 MB per die; added the LPX blog as a source

Correction: 200 petaFLOPS figure appears only in the July post; drop unsupported cable-free attribution for the Pegatron tray

Add compute tray design subsection (PCB midplane, cable-free/hose-free/fanless, per-tray specs) with NVIDIA and ServeTheHome sourcing

Correction: NVIDIA's and Microsoft's Anthropic investment commitments were swapped (NVIDIA up to $10B, Microsoft up to $5B per NVIDIA's own blog); removed an unsupported 'radiation-tolerant' attribute and recited the FCC filing clause to SpaceNews

Added SpaceX Starmind AI1 satellite selection of Rubin GPUs and Vera CPUs, with Space-1 Vera Rubin Module context

Auto-saved before update

Added 6 contextual internal links

Auto-saved before update

Auto-saved before update

Auto-saved before update

Auto-saved before update

Auto-saved before update

No summary provided