I'm always curious what threat model that is valid against TPMs is not also valid against CPUs and main memory.