MGX

MGX is NVIDIA's modular reference architecture for accelerated servers, a set of standardized chassis, tray, and interconnect designs. Server makers build on it instead of engineering a new machine from scratch for every CPU and GPU combination. It defines shared mechanical and electrical interfaces, so a vendor can swap in Grace or x86 CPUs, Blackwell or Hopper GPUs, and BlueField DPUs on the same base chassis.

The chassis come in 1U, 2U, and 4U sizes, in both air-cooled and liquid-cooled variants. NVIDIA states the shared design cuts development cost by up to three-quarters and brings development time down to about six months. Partners including ASRock Rack, ASUS, Gigabyte, Pegatron, QCT, and Supermicro ship MGX-based systems across more than 100 system configurations.

MGX explains why servers from different OEMs can share dimensions, power delivery, and cooling even with different GPUs installed. When comparing quotes across vendors, an MGX-based chassis is a useful anchor for judging whether a price difference comes from the components or from the mechanical design.

Sources

Last verified August 29, 2026.