I fully admit to having a weak spot for Serpent - it is self-bitslicing (see the submission package or the linux kernel tree), which in hindsight makes constant time software easier to write, and it was faster in hardware even when measured at the time, which is where we have ended up putting AES anyway (e.g. AES-NI etc).

BUT. On security margins, you could argue the Serpent designers were too conservative: https://eprint.iacr.org/2019/1492 It is also true that cryptanalytic attacks appear to fare slightly better against AES than Serpent. What does this mean? A brute force attack has the same number of operations as the claimed security level, say, 2^128 for 128-bit. An attack is something better than this: fewer operations. All of the attacks we know about achieve slightly less than this security level - which is nonetheless still impossible to do - but that comes at a cost: they need an infeasible amount of memory. In terms of numbers: 9000 TB to reduce 2^128 to 2^126 against full-round AES according to a quick check of wikipedia. For reference, the lightweight crypto competition considered 2^112 to be sufficient margin. 2^126 is still impossible.

In practice, the difference between Serpent and AES in terms of cryptanalytic security is meaningless. It is not an example of NIST picking a weaker algorithm deliberately, or I would argue, even unintentionally. It (AES) was faster when implemented in software for the 32-bit world that seemed to be the PC market at the time.