Chip startups may create useful IP but poor equity returns

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:54:13

   All right. Dave, over to you, pal. Selfish question because I'm really, really curious. You were the first investor in Grok, and then NVIDIA bought it for $20 billion, another brilliant jump-off investment. But now you say you won't invest in chips. So just for the audience, Grok was an AI, a transformer accelerator dedicated chip, and now it's part of NVIDIA. So now you're not doing chips, but you are doing power.

2. 1:54:36

   Are you doing data center components, boxes? Like, why not chips, and why – how did you make that choice? You have to be able to go to markets where you look dumb. And when I basically found Grok or found Jonathan and then, you know, we started to work together, the company didn't exist when I wrote the first check. It was literally like a ground up startup, him, myself. and the third co-founder. It sounded really stupid.

3. 1:55:09

   And it sounded stupid to my partners. It sounded stupid to the industry. And they were all wrong. And I was right. Now it's like conventional wisdom that you should have a silicon startup. That's why it's completely not going to work. And it's not going to work for a couple of reasons. They are right about the fact that you're going to have a bunch of domain-specific silicon. They are right that you're going to have highly customized solutions that work in very specific operating environments. All of that is right. But you will not create enterprise value. And the reason you won't be able to create enterprise value is when you go and actually try to tape out a chip and then try to find production capacity, what you're going to find is all the memory's blocked, all the fab space is taken up, all the packaging capacity is absorbed, and now you're the nth person in line. And the

4. 1:55:58

   nth person in line is effectively DOA. So I think what you'll have is some really valuable IP. But again, that's going to go to zero because some other group will be able to direct an AI model against an AI model and say, design me the domain-specific architecture that I would use to optimize this for inference or for some other metric. So you're going to have all of these permutations. So silicon, the surface area of the number of silicon options will explode, But the ability to make equity returns will go to zero because that's now too obvious.

5. 1:56:33

   If you look down in the stack where people don't go because people think it's either too stupid or not worth it is in data centers because they think it's too blue collar and a lot of intellectuals will talk themselves out of lots of money. But that is where you should play. And then the much more speculative area is in applications, which is above. And, you know, TBD around how to make money in the application stack.

6. 1:56:57

   But those are the three areas. And you kind of have to be able to, you know, shuck and jive like ebb and flow. And it was obvious that the way to penetrate AI in 2015 was in the silicon layer to me. But to me, it's now totally saturated and no money will be made. And now I have to go down or go up. And so I chose to do both. But I could be wrong. But that's sort of how I think about it.