Is possible to a implement AES in a electromechanical machine in the 1900?

Can AES be implemented in a electromechanical machine in early 20th century?

Specifically does any machine from that era

• have the capability to implement byte-oriented S-Box, or to implement the optimized word-based S-Box that computes SubBytes and MixColumn in less steps?

• Have the capability to carry out ShiftRows in the memory (or storage) that's available to the machine?

• Have enough memory to load the entire program to perform key schedule expansion, forward cipher, and optionally inverse cipher?

And, what's the estimated throughput of such implementation?

yes, I believe that you could but you really wouldn't want to do so. I will say that one could create a SPN machine based just on mechanical stepping, but it'd be huge. Someone finally created Babbage's difference engine, so it's not a big step to assume you could do something similar. Of course, there's also that hexadecimal is not too inconvenient for this discussion.

The believe that the most difficult part of this machine would actually be the S-box because the mix columns and shift rows are fixed, so you'd need to just have a method to increment the S-BOX based result on the data presented to it. You'd need a crank, where you'd put in the data, key and then crank it until you've satisfied all of the rounds. This is the fundamental difference from the Enigma where the force of pushing the key stepped to the next value. In the AES-Enigma, you'd have to set the key, the data, and then crank a handle 2560 times (assuming the 10 rounds for AES 128, the 256 times you'd have to have to mechanically go through the S-box ).

The cool thing is that you'd see the output, much like do you for these: https://www.nayuki.io/page/aes-cipher-internals-in-excel

Answer = No.

Ref. early < 1950. I'm excluding all Lego, Meccano, Mega Bloks and cam based machines for obvious reasons.

This is the Lorenz rotor stream cipher machine (very Steampunk):-

Colossus and ENIAC were built to aid decrypting this keystream. Those computers were at the bleeding edge of cryptographic analysis, and millions of people's lives were in the balance.

12 metal wheels went round and around to create the key stream. How could a electromechanical device of the Colussus and ENIAC era manage to execute 14 rounds of 256 bit keys? And create a key schedule? And perform substitution/transposition to instil confusion/diffusion? The machines could do frequency analysis and some Monte Carlo simulations.

An Arduino board can do AES but has 48,000 transistors for it's core and then extra for the memory and GPIO. It generates little heat, consumes ~ 25mA and nothing moves. Mine are 100% reliable if I don't stroke the cat before hand. Colossus had a reliability of 50%.

Plus and most obviously, they didn't have the cryptographic knowledge to design a machine for a cipher that wouldn't be invented till almost the 21st century. Also consider that the famous Difference Engine was built in the 1980's ( not designed, but materialised). It calculates logarithmic and trigonometric coefficients. Today, not in the early 20th century. It's still can't encrypt a tweet.

It's nice and romantic to think that they could do that so long ago (and I guess that's what the anachronistic Steampunk genre targets), but my answer is no.