But at what point was game not saveable, at what point did Akobian
make a (big) mistake?
Qxf2 is a big mistake as after that it is forced mate in 7 moves. If So had played 24. Bxf6 there is no defense against all the mate threads and all black can do is prolong his suffering by throwing in some pieces.
[Event "US Chess Championships"]
[White "Wesley So (2773)"]
[Black "Varuzhan Akobian (2615)"]
[Opening "French Defense: Rubinstein Variation, Kasparov Attack"]
1. e4 e6 2. d4 d5 3. Nc3 dxe4 4. Nxe4 Nd7 5. Nf3 Ngf6 6. Nxf6+ Nxf6 7. c3 c5 8. Be3 cxd4 9. Bxd4 Be7 10. Bd3 O-O 11. Qc2 h6 12. O-O-O Qa5 13. Kb1 Rd8 14. Ne5 Bd7 15. Qe2 Bc6 16. Rhe1 Bd5 17. c4 Bxg2 18. Bc3 Qb6 19. Rg1 Bc6 20. Nxf7 Kxf7 21. Rxg7+ Kxg7 22. Qxe6 Qxf2?? 23. Qxe7+ Kg8 24. Bh7+!? (24. Bxf6 Qxf6 25. Qxf6 Bg2 26. Rg1 Rxd3 27. Rxg2+ Rg3 28. Rxg3+ Kh7 29. Qg7#)
However So did not play 24. Bxf6 but went for Bh7+ instead which is of course also winning. Most likely he saw that after Bh7+ lots of material would come off the board in a forced line and he would end up in an easily won endgame with queen plus 4 pawns vs. bishop plus 3 pawns
(no counterplay for black and no fortress either).
While mating quickly is more desirable for the beauty of chess, from a practical standpoint, going for a forced line which is easy to calculate and obviously won makes a lot of sense. 24. Bxf6 of course looks very good even without calculating concrete lines, but has more potential for error. Furthermore if So had played 24. Bxf6, he would have needed a bit more time for calculating consequences of that move and also Akobian might have fought on for a move or two. So in the end the second best move (Bh7+) might even have ended the game quicker, because it was obvious that white will win.
Back to Qxf2...
That's a curious choice as it leaves the bishop on e7 hanging and after that also the knight on f6 is hanging. I have no idea what Akobian was thinking here (probably desperation or time trouble?). He must (or should) have checked at least all the options of protecting the bishop on e7, like Qc5, Qc7, Re8 and Rd7. Both, Qc5 and Qc7 lose and it is easy to calculate this on the board. 22 .. Re8 seems to hold according to the computer. Black is still under a very dangerous looking attack and will have to find the correct moves to secure equality in the moves to come. 22 .. Rd7 will lose, but not as fast as 22 .. Qc5/c7. From a GM I would have expected to play one of Re8 or Rd7 at least.
Did Akobian go wrong before Qxf2?
Computers give more or less equality (small advantage for white) before that move, but from a human perspective I find 17 .. Bxg2 a strange choice as it opens the g file for white's rook. This will always be dangerous even if there is no immediate tactics.
Also I am not sure whether transferring the bishop from c8 to d5 via d7 and c6 was such a good idea. Wasted lots of tempi just to force white to play c4. Probably a better plan for black here was to go for a minority attack on the queen side with b5, b4...