Will the AI Chip Effort Create Shareholder Value?

[Rarity]

In @luvb2b's financials thread, @SBenson lamented Tesla's whole AI chip effort. I am creating a separate thread so as to keep @luvb2b's thread on topic.

In short, I think that the AI chip is the kind of project where Musk earns his keep as an entrepreneur. The project is not super expensive, and could fail, but has a high payoff if successful. In Tesla and SpaceX, Musk has implemented a few of these efforts, the staffing for which is at the cutting edge of their respective fields. Neither Tesla nor SpaceX had experience on these projects before undertaken.

(1) RF communications chip efforts at SpaceX for Starlink. The core staff comes from Broadcom.
(2) The advanced alloy effort at SpaceX (and Tesla) leading to the SX300 and SX500 superalloys for the Raptor engine. The core staff comes from Apple.
(3) The PICA-X ablative heat shield effort at SpaceX leading to the manufacture of advanced ablative heat shields. The core technology and original staffing comes from NASA.
(4) The AI chip effort at Tesla. The core staff comes from Apple.
(5) The AI software effort at Tesla. The core staff comes from OpenAI.

Oh, one other small issue - what is Tesla's expertise in chip making anyway?? How many chips did they make so far? Now ofcourse they are all ready for a cancer-cure-level kind of chip which surely no other traditional chip makers can make... See anything odd in this?

 

[EinSV]

@Rarity, great idea for a thread.

Another good example of Tesla succeeding in a space where it has little experience is Autopilot. By most accounts Autopilot was not on Tesla's agenda until 2013, and by late 2014 Tesla already had the most sophisticated commercially available systems in all of its production cars. A year later, after enough data was collected those systems were activated. In only two years after inception of the project Tesla had the most advanced system of any production car installed in its entire fleet dating back a year. Remarkable.

With the FSD chip Tesla assembled a world class team and if the specs pan out as advertised will have leapfrogged the competition, just as they did with Autopilot. According to Ark Investments, the chip would put Tesla three years ahead of the competition in terms of performance. The investment appears to be relatively modest -- and the return on investment could be staggeringly high if successful.

As you say "this is the kind of project where Musk earns his keep as an entrepreneur." He recognizes a need or opportunity, makes a decision to take advantage of it, builds a world-class team, and executes. Often very quickly even if not on his original overly ambitious timeline.

What is most amazing to me about the FSD chip is how little attention it is getting from investors. When Tesla trotted out the AP/FSD leadership team on the last earnings call the "analysts" didn't even bother asking about the chip. There is almost no coverage in the press (although it has gotten the attention of some forward looking investors like Ark). Seeing is believing I guess so perhaps this won't change until the chip is in cars running with FSD software.

 

[neroden]

A mass-produced Tensor Processing Unit (which is what this is) is a highly saleable product. As such, it could be highly profitable. I question whether Tesla will actually sell it to outsiders, though.

 

[Rarity]

A mass-produced Tensor Processing Unit (which is what this is) is a highly saleable product. As such, it could be highly profitable. I question whether Tesla will actually sell it to outsiders, though.

As far as I know, Tesla's TPU is tailored to Convolutional Neural Networks for use specifically in vision. My sense is that other efforts such as Nvidia's are not tailored so closely to CNNs and are more general solutions.

Edit: For example, IBM's chip is a somewhat more general solution.

IBM’s New Do-It-All Deep-Learning Chip

 

[ThereAre4Cars]

I consider a perfect example of Musk. It adds risk on top of a plateful of risk for long-term advantage. I hope it pays off but I don't like this particular choice personally because this kind of effort brings a lot of potential tooling, training and maintenance issues. Even Intel has been known to screw up the verification processes. Also, you are always competing against 3rd party solutions that could easily catch up in the future since neural net code is extremely generic and not specific to autopilot.

 

[ZachF]

A mass-produced Tensor Processing Unit (which is what this is) is a highly saleable product. As such, it could be highly profitable. I question whether Tesla will actually sell it to outsiders, though.

Would help this guy for sure: