“Cohere strategically constrains its model development to what can run efficiently on a maximum of two GPUs, believing the enterprise market cannot support larger models.”