I've heard that for major AAA releases, the hardware vendors will sometimes include hand-optimized shaders in the driver. When the game goes to load its shaders, the driver will look to see if they match one of the ones in its database, and instead load the hand optimized shader in place of the one provided by the game.
They may also go in and tweak some low level settings to work around bottlenecks introduced by the game. For instance if the game is trying to synchronize its state with the GPU every frame, and the vendor can prove that doing so isn't necessary, they may have the driver ignore the synchronization when it is running that specific game.
Why can't the game developers do this? Well in some cases they can, but sometimes the performance issues are too difficult for someone not intimately familiar with the workings of a specific GPU to diagnose. The game studio could either try and find an Nvidia GPU expert, or they could let an Nvidia employee with access to low level simulations of the card and knowledge of the system do it for them.