The Blueprint For Brilliance: QCT's Modular Approach to AI Solutions

Author: Jack Lo, Products and Solutions Executive, QCT/Quanta Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

QCT (Quanta Cloud Technology), the cloud and data center subsidiary of Quanta Computer, presents its modular design philosophy for coping with the accelerating pace of AI chip releases -- from approximately 5 new chips per year before 2022 to 10+ per year since. The approach is split between proprietary ASIC/XPU designs (where hyperscaler architectures like Mont Blanc/Diablo promote some standardization) and general-purpose GPU servers where QCT applies modular design to enable rapid adoption across GPU and CPU generations. The modular design journey spans NVIDIA's MGX platform (from GR200 superchips in 2023 through NVL-72 rack-level products in 2025 to future Vera Rubin/Kyber architectures), PCIe-based accelerator servers using DC-MHS form factors, and QCT's own SGX modular design for 8-way GPU servers.

A key demonstration is the GB200-to-GB300 transition: by swapping GPU compute parts (including HBM and networking cards) on the same chassis, QCT reuses tooling, rack manifolds, and cable cartridges from the MGX 1.1 rack, minimizing generation-over-generation transition effort. QCT has implemented AI-assisted manufacturing, achieving 30% process efficiency improvement, 50% simulation turnaround time reduction, and up to 50% labor cost savings through automation -- all validated in NVIDIA Omniverse before production. For compute infrastructure, QCT is transitioning from MFLW and MDNO motherboard form factors to MSDNO (standardized DNO) as the primary modular architecture, expanding beyond x86 (Intel/AMD) to include ARM-based Neoverse V-series ecosystems in Gen6. The company also offers AI PaaS solutions with pre-configured, pre-validated software-defined infrastructure to lower adoption barriers.