DrMAL: <pencuse: 17. ... Nd5? is the move losing the game.> 17...Nd5 was actually the best move, here is eval.
Houdini_20_x64: 28/67 19:49 12,459,372,870
-0.47 17. ... Nd5 18.Rg3 Be6 19.Qd2 Kh8 20.Re4 Qd8
As I mentioned in Karpov vs Yusupov, 1989 Yusupov chose 13...e5 (instead of 13...b6) b6-Bb7 has inherent problem of weak c-pawn, and 9.Rc1 is dangerous this way. Idea of 13.e5 instead of 13...b6 was to simply work around this by opening c8-h3 diagonal for B on c8 instead.
Karpov got very powerful position for K-side attack through 14...exd4?! opening center, better was 14...e4 here is computer eval before and after pawn exchange.
Houdini_20_x64: 30/65 34:17 21,077,564,589
-0.18 14. ... e4 15.Nd2 Nf6 16.Qc2 Bg4 17.Rc1
Houdini_20_x64: 30/74 2:33:43 97,841,849,334
-0.42 15. ... Rd8 16.Re1 Qd6 17.h4 Nf8 18.Ne5
-0.47 15. ... Nf6 16.Re1 Qd6 17.Ne5 Nd5 18.Rg3
-0.47 15. ... Nb6 16.Re1 Qd6 17.Ne5 Nd5 18.Rg3
Magnus Carlsen was among those who got into trouble this way Topalov vs Carlsen, 2009 but Topalov did not take advantage of it well. It is positional error requiring more accurate defense. Yusupov did indeed play accurately, including bringing LSB over to defend K. This relates to what I wrote in Karpov vs Yusupov, 1989 near beginning of post about h6-Bh4 being added in. 6...h6 created some weakness on g6 and in position here with white B on b3 this adds to problem for black. Second mistake was 19...Bh7?! here is eval before and after.
Houdini_20_x64: 27/63 23:29 15,077,481,532
-0.47 19. ... Qe6 20.Ree3 Kh7 21.Qf3 Qf6 22.Bxd5
Houdini_20_x64: 29/74 42:43 27,291,409,472
+1.02 20.Qg4 Bg6 21.h4 h5 22.Qg5 Rae8 23.Ree3
+0.81 20.h4 Qe6 21.Ree3 Kh8 22.Qe2 Qf5 23.Ref3
As computer line hints 20...g5? was additional error, it was certainly the losing move, no need to verify with computer, black K is now open for execution. Game shows more theory, as well as providing another example of how QGD does not necessarily lead to dull, drawish positions.