A Complete Guide to NVIDIA’s Racks, and Then SemiAnalysis
Damnang
· Damnang’s Substack
· July 07, 2026 at 11:13
· ⏱ 21 min read
| Read on Substack ↗
Summary
The article argues that SemiAnalysis's claims of massive delays to NVIDIA's Kyber rack and Rubin Ultra generation are overblown and based on slippery language that can be walked back. The author provides a technical walkthrough of NVIDIA's chip, link, and rack naming conventions to decode the controversy, concluding that the midplane yield issue is unlikely to cause a 12-month delay and that NVIDIA's roadmap is intact, with the real question being the 2027 rack mix between Kyber and Oberon.
•NVIDIA's naming convention layers: chip (architecture, CPU, Ultra variant), link (NVLink scale-up vs. scale-out), and rack (Oberon, Kyber).
•Rubin uses two dies per package; Rubin Ultra adds more dies and HBM; the VR200 NVL72 (originally NVL144) is the current Rubin rack with 72 GPU packages.
•Kyber is the next rack debuting with Rubin Ultra in H2 2027, designed with vertical trays to fit 144 GPU packages and 600 kW power, using a large midplane instead of 20,000+ cables.
•The midplane is a 78-layer PCB (three 26-layer boards laminated) nearly 1 sq m, with trace spacing ≤25 microns, requiring tight 5% electrical tolerance.
•SemiAnalysis claimed Kyber NVL144 delayed to 2028, NVL576 delayed/small volume, NVL72x2 cancelled, and Rubin Ultra dies cut from four to two — all based on midplane manufacturability.
•NVIDIA responded only that its roadmap is intact, not addressing specific claims; the article notes each SemiAnalysis sentence is crafted to be walkable regardless of outcome.
•The author argues the midplane is already in production for VR200 NVL72 (smaller version) and that 78-layer lamination allows pre-testing of sub-boards, mitigating yield risk.
•The author will watch Taiwan monthly board supply chain revenue in H2 2026 for signs of Kyber ramp, as the right diagnostic for the real 2027 rack mix question.
The article defends NVIDIA's roadmap as intact, argues the SemiAnalysis delay claims are overblown, and notes NVIDIA has production-proven rack-scale scale-up with NVL72 since late 2024. The thesis su
The article defends NVIDIA's roadmap as intact, argues the SemiAnalysis delay claims are overblown, and notes NVIDIA has production-proven rack-scale scale-up with NVL72 since late 2024. The thesis supports that near-term revenue from Oberon-based Rubin racks will stay within NVIDIA.
Risk: If midplane yield turns out worse than expected or if hyperscaler pushback materializes, roadmap could still face delays not captured by this analysis.
The article discusses the midplane as a 78-layer PCB requiring advanced manufacturing capabilities that 'should start leaving traces at Taiwan's high-layer-count PCB makers and CCL suppliers from the
The article discusses the midplane as a 78-layer PCB requiring advanced manufacturing capabilities that 'should start leaving traces at Taiwan's high-layer-count PCB makers and CCL suppliers from the second half of this year.' TSMC is the primary advanced packaging and substrate supplier for NVIDIA's high-end racks, directly benefiting from Kyber/Oberon production.
Risk: Yield issues at the board level could limit upside; TSMC already has high concentration risk to NVIDIA.
The article mentions that Micron has been certified as a Vera Rubin HBM4 supplier alongside SK hynix and Samsung, with volume shipments announced. This confirms MU's participation in NVIDIA's next-gen
The article mentions that Micron has been certified as a Vera Rubin HBM4 supplier alongside SK hynix and Samsung, with volume shipments announced. This confirms MU's participation in NVIDIA's next-gen memory stack, which is a positive demand driver for HBM4.
Risk: HBM4 ramp challenges or pricing pressure from competition; the article itself notes SemiAnalysis previously incorrectly put Micron's share at 'essentially zero.'
This newsletter, published July 07, 2026,
features Damnang
discussing NVDA, TSM, MU.
3 trade ideas extracted by AI with direction and confidence scoring.