On Windows Vista and above transactions are atomic by using the kernel's transaction manager to transact all of the filesystem and registry modifications into one big atomic action. Windows XP doesn't have a transaction manager, so the transaction is pseudo-managed by the update kernel code (which isn't really atomic, it's just compresses all of the failure conditions and catastrophic failure window into as small a window as possible). In Windows XP the update files are unzipped and written to disk fully before essentially deleting the old files and moving the new files on top.
For this reason there is a tiny chance that you botch the update in Windows XP and end up with a crazy mutant OS which would likely not boot (e.g. if ntdll.dll doesn't match ntoskrnl.sys or user32.dll doesn't match win32k.sys or you catch the OS between deleting kernel32.dll and putting the new version in %systemroot%). In Windows XP because everything is pre-unzipped you won't hit this due to checksum, time or disk-space reasons, and everything is cryptographically signed so bit-errors won't hurt you but there is a tiny window at the end of the update where a power-outage could potentially brick your system.
In Windows Vista and above, short of the transaction manager failing in some unforeseen catastrophic error there is a zero chance of corruption during Windows Update.
Also "atomic operations happen instantly" on what planet? atomic operations happen atomically. They can take years if they have to, the only restriction is that they either happen completely or don't happen at all, and that any parallel access sees either the previous state or the new state. Atomicity doesn't make any statements about operation length.