Technically, they aren't fabricating the chips themselves, but it should still be no more than low end double digits dollars per chip vs buying for triple digits from NVidia. IIRC they're fabricated at Samsung, but instead of having fabrication margin (NVidia doesn't own it's own foundry, either) on top of NVidia margin, there's just the fab margin (which is must closer to cost than what NVidia charges it's customers, obviously). The rest of the supply chain for the finished boards will be much the same (Tesla surely has the PCBs made somewhere in China, the components populated and soldered there too, etc), but swapping NVidia for Samsung is a no-brainer.
Tesla also could in theory take the design elsewhere (i.e. TSMC or GF) and get it fabricated there, though it would require probably a few million dollars worth of design rework for the different process and new masks and so on. Still an upside if they ever need it as it means that they can't be permanently cut off for any reason (AP1/MobilEye), or "blackmailed" (due to no other choice but to pay them) into higher prices (which I wouldn't put past NVidia to try, they have quite an abusive relationship with their GPU board partners ...), etc.