The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:46:36 · Jack Altman
[00:46:36] S1: Yeah. I mean, it has turned out that there was a period where you look back, and maybe we're still in that period, but there was certainly a period where you look back, and the right answer with investing was just put it all in the labs. Just buy the labs. And every round, people were like, it's expensive, but the the correct answer was just keep buying the labs. It has now turned out in the last eighteen months or whatever. The correct answer was just keep buying inference. You have modal, base 10, fireworks, foul together. It's just all worked. My partner, Eric Vitrea, had this line that I've been kind of quipping a lot lately because it's true, which is that it's all gonna work. It was like, is it this or that? And Eric's like, dude, it's all of it. And it doesn't mean that every company is gonna work, not every sector is gonna work, but in general, a crazy number of things are working. Inference has been a really great way to get an index bet on everything outside the labs. Jason just made the point, which is a big part of why these inference companies are doing so well, which is that the costs are just not sustainable. So you have that on one side, which is it's too expensive. And on the other side, you have this dynamic, which I think is we have now crossed the threshold in a lot of areas and more and more happening, where you get sort of like intelligence saturation, where it is now good enough to do the thing. To take a simple example, your tax return is filed correctly or it's not filed correctly. Once you have filed it correctly, throwing more intelligence at that problem doesn't do you any good. If your job is to hit a nail in with a hammer, your hammer's good enough making it a gold encrusted hammer that costs $30,000 Your whole job is just to get the nail in. And so we have more tasks like that. And as a result, you're gonna see more open source. As And a result, you're gonna see these inference companies. And there's obviously been, I can't remember who just mentioned it, but open source is doing tremendously well. By the way, this whole cost dynamic does not mean that open source just runs away with it. The labs are obviously also going to offer much cheaper versions of their own models. And I think people forget how cost advantaged they are in a lot of ways. Their access to compute is structurally very strong.
00:48:36 · Jack Altman, Jason Lemkin, Harry Stebbings
Their access to users is very strong. They have a lot of different ways to make money and therefore can subsidize certain costs in certain situations as it makes sense. So it doesn't mean that open source will dominate, but it does mean that open source is gonna be a big part of the market, and that means you're gonna have great inference companies. So I I'm long inference. [00:48:55] S2: I think open source has reached its maximum as a market share. Think it's gonna keep going down. [00:48:59] S1: But even if it does, even if, even if it goes down by 50% and total consumption goes up by 10 X. [00:49:05] S2: Yeah. It's still good investments. I'm not saying that there's not there's not an almost infinite amount of inference, but I think we've reached peak open weights. [00:49:11] S3: What what would... Jason, why? Sorry.
00:49:13 · Jason Lemkin
[00:49:13] S2: May I reasons. Jack hit one of them, and there's a second one. They're they're crystal clear. One is at the end of the day, Anthropic and OpenAI are just deciding what they want to price their non Max frontier models, and they have many ways to compete directly. There's no reason they can't be as cost competitive as they want to be. Anthropic's got its numbers if you 7,000 things that has 80% gross margins. Okay. We can make fun of that stuff, but they have the ability to be as competitive as they want to be. Right. It is. And Jack's right. There's for sure. There's a certain point where all that matters is resolving a task. But even that, I don't think is that is quite that simple, but they can price Sonnet Plus 5.5, which just came out. I just tested it. It's it's only like 20% cheaper, but that's 20% cheaper in one week. They can do 30. They can do 40 if they want. That's the reason number one. Reason number two is, boy, people really don't want Anthropic and OpenAI training on their own first party data. But they also, I mean, I just got back from Dreamforce and I got to tell you, I know there were a of blazers there and ill fitting suits and stuff. No one wants to run on open source models there. At least Chinese, China based, nobody. Not a single person is comfortable with it that I talk to. Founders aside, but no customers. So I think just those two trends mean we've hit the peak, right? To Jack's point, it's not that open... No, I mean, it's ridiculous to say it. No one at OpenAnthropic is stupid. They can twist the knobs and dials and do what they want and be as cost competitive and just... There are arguments for open... But I just... This anxiety at the C level is only going to increase. It's only going to increase, and it's only gonna increase as security becomes a bigger issue, it's only gonna increase as Astra five one was pulled back for security concerns. And no one had... Listen. I know it's fuddy duddies, but it is the real world in enterprise. I asked him. No one wants to use an open weight model on the floor that I talked to. Nobody. So I just think the market share has peaked. I'm not saying it's not material.
00:50:54 · Rory O'Driscoll, Jason Lemkin
[00:50:54] S0: Did you ask open weight, or did you say Chinese open weight? Because I think there's there's a bunch of things [00:50:59] S2: going on. I get your point. I'm simplifying that the vast majority of these models that we're consuming today are are China origin models. Right? [00:51:06] S0: Where some of those guys are also moving towards not being open weight themselves. But yes, I do think that's why the whole poolside Nvidia thing is interesting. It would be interesting to have a low cost US based alternative. And going back to Jeff, it's not open weight, but it is US and it's not an LLM. But for a certain class of use case, it is a direct competitor. Right? [00:51:27] S2: I'm not saying that they're destroyed to Jack's point. I just think they've peaked. They've peaked in market share. I think it will come down, and I do not... I don't even think it's gonna main modal or base 10 or others, but I think these market share charts that have seemed crazy this year, there's been a lot of sources, but but I think there's a lot of structural benefits that OpenAI and Anthropic have here, and they're they're gonna take take advantage of it in the coming months. Right? And OpenAI just hit 70,000,000,000, they said, enterprise. Like, it's it's it's a lot of traction there.
00:51:27 · Jason Lemkin, Rory O'Driscoll
[00:51:27] S2: I'm not saying that they're destroyed to Jack's point. I just think they've peaked. They've peaked in market share. I think it will come down, and I do not... I don't even think it's gonna main modal or base 10 or others, but I think these market share charts that have seemed crazy this year, there's been a lot of sources, but but I think there's a lot of structural benefits that OpenAI and Anthropic have here, and they're they're gonna take take advantage of it in the coming months. Right? And OpenAI just hit 70,000,000,000, they said, enterprise. Like, it's it's it's a lot of traction there. [00:51:51] S0: Pushing on that, and again, I don't know the answer is that provided you have opportunity at the frontier, you won't waste time with the second string stuff. To some extent, you're probably more likely to spend time on, as you say, doing the tax return, if you feel you're asymptotic on some of the frontier stuff, because the thing that you're allocating at the margin is your compute. And if you're open AI on Tropic, I'd love to know what the math is like on, do you take this next chunk of GPU and build a model for biology that can, know, cure cancer? Or do you take it and do, you know, Jack's tax return, which is a pretty finite, probably not a pretty finite task, actually, your benchmark partner, it's pretty infinite task. But within the context of AI, it's pretty finite, And super interesting set of trade offs there.
00:52:35 · Jack Altman, Rory O'Driscoll
[00:52:35] S1: Yeah, mean, I think a couple other reactions are one is that to the point of, you know, non American models, I do think that there's a lot of anxiety there. I also think we are starting to see and will continue to see a lot of enterprises post train their own models and draft off of open weights and use inference companies to make their own models and then run them themselves. And so I think... I I don't know how that will factor in, but I do think we will start to see more of that in The US. And I think that will have some amount of impact. The other thing is you can kind of... I often try to just kind of step back, blur my eyes, and what's the one thing here. And I think one of the things you could look at here in general, it's kind of all just gonna come down to like who's got the compute. If you look across all of the inference clouds, I think it's on the order of like a gigawatt or something like that. And I think open ion and tropic, high single digits each. And so you could say maybe one's a little bit more efficient than the other. Maybe one's got better token efficiency, one prices a little bit differently. But on some level, this will also just come down to all of the compute is firing all the time and who owns it, I think is gonna turn out to just be a dominantly important part of the equation. [00:53:46] S0: Yes. Compute share probably proxies to your token share, probably proxies with a little adjustment to your revenue share. [00:53:52] S1: At least within two x or something like that.