Nadella argues that firms need their own learning loops around models, including private evals, internal traces, and institutional memory they can keep when swapping a general model. It is worth reading as an AI control and governance pitch, but the post gives architecture language rather than proof that these loops are secure, portable, or operationally mature.