The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:26:10 · Dev Ittycheria, Jason Lemkin
[00:26:10] S1: Yeah. I actually tweeted out this morning. I actually thought that enterprises kind of can have their cake and eat it too. One, you get a US open source model that's near frontier intelligence. That's three to four times more cost effective, I e, cheaper. And that's music to the years of any enterprise because there's obviously gonna be some hesitation by using Chinese open source models, especially if you're in a regulated industry or industry that deals a lot of sensitive information. To me, this is a... Was a very noteworthy announcement. Again, I do wanna disclose the Sequoia's investor and re... Reflection, but I think it's great for the enterprise. [00:26:45] S2: That was a Dreamforce this year, right, and got to actually have the surprisingly large number of conversations with execs. Notwithstanding the incredible force of nature of Chinese based open weight models. When I nobody I talked to really wanted to use a Chinese source model, right or wrong, fair or not. Everybody that was, they felt like it was under duress for cost, but no one really wanted to use one. Again, right or wrong. So, you know, if they nail this right, it could be a torrent of demand.
00:26:45 · Jason Lemkin, Dev Ittycheria
[00:26:45] S2: That was a Dreamforce this year, right, and got to actually have the surprisingly large number of conversations with execs. Notwithstanding the incredible force of nature of Chinese based open weight models. When I nobody I talked to really wanted to use a Chinese source model, right or wrong, fair or not. Everybody that was, they felt like it was under duress for cost, but no one really wanted to use one. Again, right or wrong. So, you know, if they nail this right, it could be a torrent of demand. [00:27:10] S1: I agree. And I think this also reinforces a point that you don't need, you know, the frontier level intelligence for every workload. If it's close to the frontier, as Rory said, within six months of the latest models and addresses 90% of your use cases, like especially around reasoning use cases like coding and agentic workloads, I don't see why enterprises would not kind of want to talk to them immediately.
00:27:33 · Rory O'Driscoll
[00:27:33] S0: Ties back to the first topic at one level. So far the evidence has been that there hasn't been a massive push because people are lazy and the bills haven't been astronomical. And therefore they've just kind of used the frontier model and maybe downgraded to the Anthropic or OpenAI less expensive model levels. But I agree. I think the interesting thing now is because the numbers are getting so huge. I mean, if OpenTropicAreach making $70,000,000,000 in ARR, that's $140,000,000,000 that's more than Microsoft taxed for the longest time across both Windows and Office. That's a chunk of US corporate profits, right? And at that point in time, if you're a CFO, you probably come in to your CIO and say, I don't care that it's easier to choose you as Anthropic, dude. We need to shave $3,000,000 off this token bill. So start using JEV for 20% of your decisions, start using reflection or something for anything that's not frontier and figure out a way how to use Entropic for the hardship. I'd imagine, Dev, you guys will start to have those conversations and all your customers will. That's why, again, these growth rates are so interesting right now for Entropic. Can they still maintain that unparalleled growth rate in the context of this kind of chipping away? It's like, remember the old man in the sea, the guy's bringing back the fish, and then the the the other fish are known at bits of the body. You know, they're just taking away slugs of revenue from them.
00:28:50 · Harry Stebbings, Rory O'Driscoll, Dev Ittycheria
[00:28:50] S3: And then you've got JAV, and you've got a load of the other JAV competitors also, in a different way, attacking them. [00:28:56] S0: Yeah. And, the forward momentum has been such that all this is like, yeah. Thanks, Rory. That's super interesting. But did I mention it'll be 10 x last year and the year before? [00:29:04] S1: Other thing I think is important to understand, and Jason, I'm curious what you saw at Dreamforce, but I think there's a big gap between where the capabilities are and where actual usage is in the enterprise. I think there's a paucity of of real experience skill sets inside the enterprise. And so think we're gonna see Jevan's paradox kind of come into play where, basically, as these costs come down, people get more and more comfortable deploying it for more and more use cases across the enterprise. As they get more conversant and and skilled around these technologies, they'll start deploying it for more use cases. I see this even at MongoDB. And so I think you're going to see that frankly, this is a massive market and I think everyone can be successful. I mean, I won't comment on the growth rates of the frontier labs versus the open source model, but I don't see this like, I don't see this as being a zero sum game.
00:29:52 · Jason Lemkin
[00:29:52] S2: To answer this, the captain obvious answer to what I learned at Dreamforce, everyone I talked to, everyone was just overloaded with the amount of internal demand there is for tokens and whatever it is, it's more than they can imagine. Whether it's being used properly, whether it's being burned for nothing, whatever the cap is, demand is probably an order of magnitude more than they've figured out how to surface, right? That may lead to using sub optimal models. It may lead to a lot of waste in some ways, but everyone's got to manage it, which just that and the fear of using China based models. Again, right or wrong is just a huge opening if you can nail The only thing I will just add just for fun we can move on is man, just personally I'm getting more and more skeptical of published evals. Like of course you can make an eval look great. How fast was it? What were really the outputs? How well did it perform on a real workflow? How well did it perform on a mission critical workflow versus telling you an item was in stock when it wasn't, which is still kind of mission critical workflow. I found that every single eval needs like three asterisks and four daggers next to it today. So I am skeptical Beam is as great as they say it is, but I just asked while we're on if I get on open router and check before we end the podcast, but it's not there yet, I would try. But it's not that I don't think these are all great. I just think all the evals are so biased and focused on chess championships that I'm slightly skeptical.
00:31:07 · Harry Stebbings, Jason Lemkin
[00:31:07] S3: We have Nematron, we have Thinking Machines, we have Reflection. In twelve months' time, what percentage of token in open source flow will go through US versus Chinese? [00:31:20] S2: Well, look, just two things. As Dev made the point in the beginning, developers will change and and especially here. This is a lot easier than swapping out your database. Right? That that that could be done, but I... Having recently done it, it's some work, man. I gotta tell you. It's not... Even with an LLM, it's... I mean, listen, this is still work. You have to qualify your prompt. You have to redo your workflows. It's not nothing like the internet says, but I could see it being half in twelve months. Twelve months is a lot of time. And I think sovereignty plus U. S. Based is really important. I'm not saying it will happen because I don't know that again, I'm skeptical the parity is really there versus pretend there. But if if you're even x minus three months to Dev's point, it could be half in a year. A year is a long time. [00:31:58] S3: But that gives me a lot of confidence. I... I'm I'm an investor in fireworks, as is Sequoia. And I think the biggest concern I would have as if Fireworks or a base 10 or a Modal investor would do is that actually big enterprises are concerned about Chinese models, and that would prevent them from engaging with partners like Fireworks or Modal, and they then go to Frontier. And so the fact that you say 50% shows me that the the levels or quality you think is sufficient enough that it would be that rate, that makes me excited.