As discussed elsewhere, I think Tesla is using the same underlying GPU core, the GP102 or variant, but not the actual Drive PX2 modules themselves. Therefore NVIDIA can claim the same platform, but Musk referred to it as the Titan. Since Tesla is writing their own software stack completely that is portable between multiple processors, combined with the much lower number of sensors including the lack of LIDAR to process, Tesla can use a far cheaper hardware solution.
I could not find estimated pricing of the GP102 itself. Given that the retail price of a Titan X is $1,200, and unclear if Tesla needs the same size and speed of memory for AI processing, I suspect the price per chip is under $500. So the hardware cost over AP2 is somewhere around $1,500 to $2,000. Now, it may be that at the limited quantities that Tesla sells as compared to consumer electronics, it might actually be cheaper for Tesla to equip all cars with AP2 hardware than only those that order it. Especially before Model 3 volume hits its stride.