“The multi-year development cycle for custom chips prevents even large AI labs from designing hardware for a specific future model, as model architectures evolve much faster.”