“Groq's LPU architecture is particularly well-suited for Mixture of Experts (MoE) models because it can be economically efficient with a small batch size, such as 10.”