GS expert call (via X): Kyber NVL144 / Rubin Ultra racks pushed to 2028

Opening

Kyber NVL144 racks for Rubin Ultra GPUs pushed to 2028 — more than 12 months behind an original 2027 target — per a Goldman Sachs Nvidia 800VDC expert call relayed by @ParadisLabs. The same takeaways cut the 2030 US share of new datacenters expected on 800VDC architecture to 21% from 25%, citing strong AI inference growth that does not need training-class power density. Full 800VDC architectures are "not anticipated for at least 5 years"; near-term path is 800VDC for servers with traditional architecture for networking, storage, and lighting. Inference racks up to 150 kW are said to keep traditional architecture feasible. Primary GS note not opened this pass.

The numbers

Kyber NVL144 / Rubin Ultra — 2028 vs original 2027 target (>12 months slip).

800VDC share of new US datacenters in 2030 — 21% (was 25%).

Inference rack power cited as feasible on traditional architecture — up to 150 kW per rack.

Why it matters

A year-plus slip on the rack system meant to house Rubin Ultra is a delivery-timeline fact for the next NVIDIA scale-up node, separate from demand narratives. The 800VDC share cut is framed as inference mix, not a demand cut — but it does push the power-architecture upgrade path out.

Positioning

AI capex durability — Capex can stay elevated while specific rack/power SKUs slip; the call separates inference density from training density rather than calling an end to build.

Connects to

Astra / next-gen GPU framing — overnight Astra loop-transformer debate is about model shape; this print is about rack delivery timing for Rubin Ultra (discussion).