Overview
- At the OCP Global Summit, Nvidia said Samsung Foundry and Intel have joined the NVLink Fusion ecosystem for third‑party CPUs and accelerators.
- Samsung will provide design‑to‑manufacturing support so customer chips built on its processes can integrate Nvidia’s NVLink‑C2C interface and operate inside Nvidia GPU clusters.
- Nvidia confirmed Meta and Oracle will deploy its Spectrum‑X Ethernet platform for AI data centers, with Meta using Spectrum‑4 switches inside next‑generation Minipack3N hardware and Oracle targeting giga‑scale clusters.
- Nvidia outlined next‑gen systems based on its Rubin architecture, including NVL144 with 144 GPUs, a Rubin CPX inference part using GDDR7, and a goal to launch the Kyber supercomputer in 2027 with 576 GPUs.
- Samsung’s entry positions it to court custom‑AI chip clients as TSMC capacity remains tight, while OpenAI’s Broadcom‑developed chip is progressing with the manufacturing foundry yet to be disclosed.