Well, 133
things per second, not necessarily instructions. That depends entirely on the architecture. Some might be able to pull off 1 cycle per instruction. Some will have all instructions having the same number of cycles (for example, all instructions take 5 cycles), perhaps for pipelining reasons. Some will have all instructions taking different amounts of time (instructions could be between 2 and 5 cycles, and each instruction could take a different amount of time depending on what happened (like branch versus no branch)). I may be wrong, but from some quick reading on Google, it looks like it's incredibly difficult with x86 to figure out how long instructions will take, particularly when you consider things like cache hits versus misses and that sort of thing.
Anyway, considering it's an Atom and can run AI War, he definitely meant 1.33 GHz (and he confirmed as much a couple posts ago).