How is the Token/second calculated? I ask it a simple prompt and the model generated a 150 word (about 300 tokens?) answer in 17 seconds, then mentioning the speed of 408T/s.
Also, I guess this demo would feel real time if you could stream the outputs to the UI? Can this be done in your current setup?
Other comments give lots of examples of organizational complexity.
Here is an example of code complexity:
Changing an object from enum to String is trivial to do in a small codebase. Making this small change in a live system like Youtube becomes a 6 month long migration with multiple SWEs working on it, when you have to do it in a codebase with 10s of millions of lines of code.
A lot of complexity comes from code size, multiple teams may end up creating duplicating products/services if there is no standardization etc.