MatX develops high-throughput chips for large language models, targeting training, reinforcement learning, inference, and long-context workloads.