MatX

MatX designs hardware tailored for the world’s best AI models, focusing on high-throughput chips and optimized interconnects to maximize performance for large models. They target cost-efficient training and production inference for extremely large transformer-based models, offering scale-out capabilities across hundreds of thousands of chips and low-level hardware control for expert users. Their technology aims to enable faster access to state-of-the-art models and higher performance-per-dollar in data center AI workloads, with emphasis on 7B-class to 20B+-parameter dense and mixture-of-experts models and scalability to 10T-class models. Based on information presented, their operations appear global with emphasis on AI model training and inference hardware.