I think what should give Nvidia nightmares is that Tesla could set a precedent for the AI industry, and encourage others to develop their own custom chips. Their once-promising AI empire is collapsing.

Yeah, something tells me most car companies aren't going to employ entire chip design teams, stepping through minefields of regulations and patents for something that in 10-15 years will be akin to power steering. Not to mention that there is more than enough room for multiple companies to produce those kinds of chips, not only for cars - but boats, drones, planes, etc, nor the fact that the news of Tesla building it's own chip is more than a year old - the only difference is that might be AMD is involved. I think this comment from reddit sums everything pretty well:

The initial quote didn't mention AMD at all. This entire thing is basically CNBC hearing "Tesla is working with AMD and GF" while what actually was said that there is a trend of companies developing ASIC and contracting fabrication in many new fields like the automotive field with Tesla. Which is why they've issued a correction in like an hour saying just that. It's been known for a while that Tesla is developing their own hardware, they've done that throughout the development of their cars from over 90% being contracted out to less than 30% today. SpaceX did the same, Elon pushes for 100% or as close to that as you can get of vertical integration. This also means that in the long term they will develop their ASIC IP completely in house just like Apple, even if they'll license something from AMD it will be a stepping stone. Given that fact, licensing makes little sense now since you are already going through all the trouble. But the threat is enough for NVIDIA to drop the $8000 price tag of the Drive PX2 by a few notches. The biggest factor here is cost, Tesla wants every Tesla car to have self driving capabilities and the current hardware is very expensive. Tesla will have to drop the cost to well below $3000 for the entire autopilot including the sensors (which is why they use cheap cameras instead of LIDAR) for the model 3 and even then it's too expensive for its class. Tesla might very well be working with other vendors, but this entire story is conjured out of thin air and is a dictionary example of poor journalism. The timeline is also inconsistent, some people added the 2017-2018 AMD non console custom into the mix. This simply can't be Tesla, the new 2017 autopilot 2.0 has been gutted it's running on NVIDIA Drive PX2. For Tesla to come out with their custom ASIC even in 2018 it's way too short of a time frame especially since Autopilot 2.0 hardware isn't fully active yet. If Tesla will come out with an AMD based custom it's likely won't be before 2021-2022 which is the logical stepping stone for their next autopilot hardware revision as they replace those every ~5 years. People forget that the model S is a 5 year old car by now, Tesla isn't going to be putting millions and millions of dollars into hardware that will be on the road for 6-12 months.

The PX2 is too costly to put into low end cars with autopilot - given Tesla/Elon's insistence with every car having self-driving they need a way to drive costs down. Nvidia is subsidizing it's car software platform via hardware sales but Tesla wants it's own software. Having an $8000 chip in a $35,000 car just doesn't work. But for all the other companies who aren't building software teams or hardware teams with the likes of top engineers like Jim Keller - Nvidia's option is as competitive as Mobileye or anyone else in the industry.