Chris Sawyer wrote 99% of the original RollerCoaster Tycoon in x86 assembly. RollerCoaster Tycoon (video game) - Wikipedia

Maybe. But the cost is abstraction. Assembly is as close to bare metal you can get, without going into complete binary mode.

The following examples are borrowed from https://courses.cs.washington.edu/courses/cse378/03wi/lectures/mips-asm-examples.html:

A simple program

i = N*N + 3*N

could in assembly be written as

lw     $t0, 4($gp)       # fetch N
mult   $t0, $t0, $t0     # N*N
lw     $t1, 4($gp)       # fetch N
ori    $t2, $zero, 3     # 3
mult   $t1, $t1, $t2     # 3*N
add    $t2, $t0, $t1     # N*N + 3*N
sw     $t2, 0($gp)       # i = ...

However, this isn’t an optimized solution.

lw     $t0, 4($gp)       # fetch N
add    $t1, $t0, $zero   # copy N to $t1
addi   $t1, $t1, 3       # N+3
mult   $t1, $t1, $t0     # N*(N+3)
sw     $t1, 0($gp)       # i = ...

See how complicated it gets real fast? Right here, we are only adding and multiplying numbers, nothing fancy at all. Your compiler would catch such an easy optimization, but you might have spend an hour to figure it out. That is what’s meant by abstraction. You sacrifice a bit of speed (sometimes it’s even negligible) in order to not have to write at such a low level.

Many web/network applications performance is dominated by IO concerns rather than raw number crunching. There are some unique cases where you can benefit from such optimisation, however they are usually quite rare and quite small in comparison to focusing on other areas like DB query optimisation, caching etc. Dropping down to assembly is a lot of pain and very expensive in terms of development/ maintenance for what is likely to be a tiny percentage of the possible benefits. At the very least it is worth testing and benchmarking to find out what is actually slow in your application design before working on optimising.