Cost-efficient AI at scale is a software problem, given that all Nvidia competitors are lagging behind on software development. Not only that, but software will also become even more critical as heterogeneity in the data center becomes commonplace, Moreh CEO Gangwon Jo told EE Times in a recent interview.

South Korean startup Moreh was formed by a group of HPC experts in 2020 to build an AI software stack for AMD GPU clusters. AMD gaming GPUs were the cheapest way to build a supercomputer at the time, Jo said, but the AI software stack was missing. Moreh was founded to address this gap, with the overall aim of making AI training and inference at scale cost-effective by enabling alternatives to Nvidia hardware.