The industry overestimates how much value comes from access to in-silico foundation models and underestimates how much comes from target-specific benchmarking, model calibration, and experimental context. Better models help. Better framing, calibration, and selection matter more.