Frontier Labs Threatened by Kimi? Should the US Ban Chinese Open-Source Models & Stripe Buys PayPal

I totally agree with that, actually, Jason. I was curious to see what you'd say, is that there wasn't — you kind of led with the 'what do these new models mean?' I think Jason's take is exactly right. It's exactly what you'd expect. It turns out there's five main Chinese LLM companies and a bunch of followers. It turns out that aggressively funded companies with smart engineers are just going to keep cranking through and building new models. And yeah, they're not state-of-the-art compared to the frontier models, but they're, you know, six months behind depending on how you measure it. So yeah, actually no new news about that. But Jason's right, quite a lot of fun news about how parts of the US responded to that. We had the small-p political response. So that's one dimension. You know, the policy adviser for OpenAI, formerly from the Trump administration, making some comments on Twitter leading to a wonderful firestorm that we'll absolutely talk about. That's one thread. And then another thread is just talking about what these models start to reveal about the economics of a model company. I mean, Jason hinted at it. We lump all these models in together, right? But let's take two: some of the DeepSeek models you can run on your PC or your Mac or on a laptop, right? Conversely, Kimi K3 is, I think, a 2.8 trillion parameter model. It's a huge honking thing, and you know, you need myriads of GPUs just to run it. So they're not, quote-unquote, the same thing. That's much more comparable in size and therefore in terms of compute capacity to US frontier models. So there's a lot we can learn about. I think we'll kind of talk about the politics first and then maybe, oddly enough, talk about the inference implications and kind of as that goes into the opportunity for fireworks. So, lots of kind of downstream implications, but zooming out, nothing amazingly surprising in the news that after 3 years of competent execution along a pretty defined trend, we now have 3 years and 3 months of competent execution around a pretty defined trend.

It's Dean Ball, and he is currently, I think it's out of policy or communications director for OpenAI. He just started there two weeks ago. Before that he was at the Trump part of the administration, kind of on AI policy, and before that a bunch of Hoover Institute type stuff. And he, you know, he set off a firestorm with the tweet, and then he did a little bit of 'Oh, I can't really post because I'm now at OpenAI, everyone was mean to me because I posted a bunch of stuff,' which was frankly a little naive. Because the two comments about the tweet: one is you're in a senior role at OpenAI, and there was a hysterical tone to it, right? He used the word 'AI communism' and it was very kind of over-exaggerated. And then secondly, you know, when you start even hinting about — I mean, we saw this when Sarah Friar did it — when you start hinting about significant regulatory changes that will massively benefit you, you got to expect that everyone's going to say, 'Dude, of course you're going to say that. You're, you know, that's your side.' And if you make the expensive closed-source product that sells for, you know, 10, 20 bucks and the Chinese are shipping something for two bucks and you say, 'Well, totally independently, just speaking as a common citizen, I think they should ban this shit,' you got to expect that a whole bunch of people are going to say, 'Dude, you're not talking as a common citizen. You're talking as the provider of the company who will jack up our rates the minute the stuff gets banned.' So, you know, it was a little naive not to expect that level of blowback.

As they say in the meme, Jason, two weeks so far.

I didn't realize Bill had said that, though there's something very pleasing about that, which I'll mention in just a second, right? Because one of the fun things about this policy dispute, it brings out the hater in everybody, right? And you know, Dean Ball said what he said, and then two people who can be controversial came down strongly on the other side, and I support them both. The first was David Sacks, the former AI czar, who basically said, 'This is rubbish, stop.' And then the second one was Emil, whom you mentioned before, Emil Michael — I'm never sure of the pronunciation of his last name — who was a guy at the Defense Department who got totally sideways with Anthropic. I mean, what I like about that guy is that man knows how to hate, and one of his biggest hates for the last decade and a half has, of course, been Bill Gurley from his time at Uber. So I really find it... So if Bill and Emil are on the same side saying don't ban these models, then you got to know that there's got to be some truth in that. It's got to make you think, because that's an interesting lineup. But yeah, I think there — and I actually just saw literally as I came on, and look, this is the Trump administration, so things change every day. But Politico today basically saying some version of 'we're not going to ban these things on any significant basis,' which, as Jason points out, is very different than saying the White House decision support system will be run on Kimi, right? Even if it's Kimi hosted in California, I think we can take it for granted it won't be. Right? But conversely, if you're a startup doing inference on customer support queries for a very boring consumer product, there is no reason why you should pay marquee prices when something 10x cheaper is available, and it would be horribly bad policy to ban that. Right.

I don't think it's triggering to say that there are IP risks in this. I mean, at the risk of being kind of level-headed here, the data is very clear that, you know, technically important US companies, Boeing for example, suffer continual cyber attacks, many of which are attributable to sovereign state actors, including China. It's a thing. So we're not being, you know, we're not being sensationalist or alarmist. You know, it would be naive not to put it on the table. You're right on both counts. Second point, companies — I'm thinking about, can you... the problem proving a negative is, can you know, if you have an... remember these are open — I occasionally say open-source incorrectly — they are open-weight, which means you can see the weights and you can run it yourself, but you don't have technically the full definition of open-source in the context of an LLM, which means seeing the underlying training data, which you don't, right? But you have the open weights. The question is, if the model is being run in a trusted US inference company, Fireworks, some of those guys, right? And you could get into a long technical question of, look, what can it really do? Could it initiate tool use on — I mean, it probably speaks — could it initiate tool use on the customer side whereby the model sends a command back to the customer to exfiltrate their data? You can imagine being able to use these models fairly comfortably and being fairly secure that, fairly certain that you have blocked access and this can't happen. So I mean, I think you could, in logic, you could satisfy a technologist that the risk is not there. Whether you can satisfy a politician, whether you can satisfy someone who's just afraid of what they don't know, is, Jason, to your point, another question, right? I mean, and I think you're right, you have seen things like Huawei has effectively been prevented from selling to any cellular networks in Europe and the US because of this, as yet unprovable fear. So it's not crazy that there will be some level of, I think on the government side, some restrictions. I think a blanket ban would be massive overkill, to be very clear, right? But I think the interesting question it raises is this — well, two questions. First of all, it's also worth pointing out that while we're talking about banning Chinese open-weight models, the Chinese administration are talking about preventing those companies from selling those models to the US. Like, just like we don't let them buy Nvidia, they're not going to let us buy their open-source models. Which is kind of totally zany. We think they're trying to sell it to us and we don't want to buy it, and they think they shouldn't be selling it to us because it's so powerful. So that's kind of just weird in of itself. But I think the really interesting question here, and it gets to Thinking Machines, is: is the open-weight, low-cost LLM business a good business? And if it's a good business, why can't some red-blooded American company step up and give OpenAI and Anthropic a run for the money? And Jason, it's the point you made — if this is... I mean, where's Grok? Where's Gemini? Thinking Machines had an announcement last week. They announced a model. They didn't position it as, you know, state-of-the-art frontier, but they kind of, I think they made a comment on something that you can build upon. Inkling, I think it was called. So, one of the — and you know, at one point, Meta looked like they were going to go down this route, right? Is there a business? How can you make money? It's an interesting question. Can you make money as a maybe not completely open-weight but a low-cost US provider of these models and be competitive with those guys? Because you know, the open-weight models in China are getting 50, 70 billion valuations. Like, it's not Anthropic, but I wouldn't turn down a $50 billion outcome if someone could make a convincing case to me that a US company could do this. So I think that's one of the interesting questions here. Maybe it's because the dirty little secret is a lot of their advantage is distillation, which you can't legally do if you're US-based. So I do wonder, Jason, to your exact point, if there is a market for 80% cheaper intelligence, and that's roughly what we're looking at in terms of when you take into account the cost of inference, the difference between the bundled product that is a frontier model — you know, IP plus inference — and an open-source model where you dissociate the IP from the inference cost. If you're looking at an 80% cheaper opportunity and there's the mass demand for that, when is someone going to try and fill that demand in the US? I just, I don't know.

Yeah, that's because we say such nice things about them. Thank you for your check. But no, look, we're going to gloss by it and I don't have an answer, but it's a huge freaking question, right? There's this new category called LLM intelligence. Two companies existing as premium products, their combined market cap is $2 trillion. The combined revenue at this point is probably 100 billion, plus or minus. There are four or five other companies in the US that are capable and have proven their ability to build something roughly comparable. None of them are taking advantage of this. And there's five Chinese companies that have proven their ability to build something roughly comparable and they're, you know, cranking night and day to take advantage of it. Where are you, Google? Where are you, Reflection, as you say? Where are you, Thinking Machines? Where are you, Llama? I mean, the fact that there's four or five potential competitors — it's just fascinating.

Jason is answering the question, why is it a good time to sell? And you know, frankly, as you yourself, Jason, have said, and I say it too, the private market liquidity window opens so rarely that it's always a good idea to pay attention when it does. So I think it's easy to understand that side of it. I actually think the interesting side of the discussion is the other side. Why would someone want to buy, right? And I think when I saw that article, I was like, yeah, that makes sense. If you think about the last conversation we had, you know, what's in the zeitgeist right now, it's this whole idea of can I get escape model dependency, manage my costs, have a whole load of alternatives easily available to me, right? If I'm someone who makes my money as a hyperscaler hosting, especially someone that doesn't have their own in-house model they're pushing like Gemini. If I'm maybe Amazon in particular, who's made a business of saying I'm going to support all the models. If you had maybe even Microsoft now that the divorce is coming true from OpenAI, you know, maybe this would be a great product to own if I was a cloud hyperscaler. So I will admit I had that moment of — because you often see, and it's step out, you often see this in this kind of market. You have these businesses where intellectually over a 10-year period, you can say, 'Hey, margins are going to be tough in that business. It's going to be compressed, and maybe it won't at scale in the end be an amazingly valuable business because it won't be able to extract margin.' But on the other hand, when growth is so quick and people's urgency to adopt technology is so fast as it is right now, if you have a crucial piece of the plumbing at just the right time, when two or three people need that piece of plumbing, you can find yourself in a very interesting position in terms of M&A, right? Because it may well be that — and this is where the finance guys miss it — it may well be that the NPV of the company on a standalone basis is, you know, a couple of billion, not huge, but the value to it right now to a hyperscaler, if they could shift 10% market share in the enterprise to them over the next half a decade by saying, 'Dude, we are the cloud, we are the model-agnostic people, and we'll make it easy,' could be interesting. So I remember thinking, fuck, I wish I was in that — I mean, which is always how you know what a venture guy really thinks. It's like, damn, I bet you they could at a good offer right now. It's just interesting. So yeah, I'm with you. I think it's an interesting time to sell, interesting time to buy. I mean...

You know, it's always hard to... I mean, my first comment is whenever you get an offer, I always do the same thing. I say to the founders: one, the window's open. It doesn't open often. We should take it seriously. Two, I'm going to support you whatever you want to do. Three, if there's concerns that you have that you've been sitting on and not telling me, now would be a good time to share so we can make an informed decision, right? And then go away and think about it, and I have a whole process of how you talk to them about it, right? I don't think you should pressure people into selling. I think your job is to give them whatever experience you have to bring to bear, and then they'll make the decision. Because you know, in the end — and it's so funny because founders agonize about this when they think about control and they think, 'Oh my god, these people are going to make a sale' — even if as the VCs we have board control, 70% ownership, and a drag-along for the founders, Jason knows this: if the founders who are core to the business don't want to sell, it's not going to happen. So the first thing I tell the founders is, it's your decision, which I think is very empowering because it takes it away from 'people are going to make you' — you know, that's the beauty of being private, unlike being public where you don't have that degree of freedom. We could come back to that later. So yeah, I would say to them, take this seriously. Do you think you can be worth 3x this in three, four years? Do you think that's worth it? But yeah, I would definitely say take a day out of your life and think about this long and hard.

It was weird. I didn't think you were going that direction, but as often happens with you, and I listen to the whole thing, I'm like, I get it. I think what he's saying is this, Harry. Right. You know, when you face that sell decision, you don't not sell because you think you can make twice as much in a year, right? Because you just never know. And so in the end, even though I didn't think I'd agree with him, in the end I did. It's like what I think he's saying is, let's leave aside the 'what do you want to do with your life' questions from a return perspective. You know, don't think incrementally. If you have a chance to sell a company for six billion and make 600 million, and you think you can run it another three years and get, you know, 1.2 billion, that might — risk-adjusted — that mightn't be worth it if your current net worth is $40,000. And that's actually good financial advice. I mean, in other words, if you turn down a big-ass offer, you better be sure it can be way bigger. You better have high certainty and high bigness. I think that's a fair comment.

It wasn't what I thought — where I thought you were going, because I actually thought you were going to say, yeah, something that you also... Yeah. Don't sell also just because financially you feel you should. I mean, because the other thing you're not taking into account is that depends on the person. Some people just love running that company and frankly don't want to sell. It's their life's work. And I also think you have to respect — I mean, my point, you also have to respect that. I mean, again, this gets back to the, there's no one answer, the founder decides. I've known people who are like, 'This is my first hit, I'm going to take it,' who read exactly that logic. I am not...

I'm going to make 60 million bucks. I don't have one million bucks, and maybe I can make 120 in four years, but I'm taking the 60. And I've also known other people who, to a rounding error, have said this is what I want to do for the rest of my life. Why would I take that money? I've got—I've made three or four million in the secondary. I got my house. I'm done.

No, no, you do. No, you do first, because I'm just—

Yes. First of all, I agree. Yeah, inference is a huge—it goes back, ironically, to the prior comment on open-weight models. This kind of standalone inference is a big business, right? And you know, obviously, inference is both something that's done within the frontier model companies where they do their own inference, and people like Microsoft and Google provide the capex and provide the compute for that. But people like Fireworks and Base10 and Fire—they all make their money offering a variety of these open-weight models to third-party developers and enterprises that want to use open-source models to do AI, right? And as I said, the two trends go together. They're exploding because the open-source trend is exploding. So if you're Base10, if you're Fireworks, if you're Together, this is your moment, right? So yeah, this is—because this is how you access those. Because I can tell you one thing: going back to the discussion about open-weight models from China.

It's one thing to decide to use an open-weight model on Fireworks in the US. What you're not going to do is be using the API back to China, even if they'd let you. Right? So this is a one-to-one linkage between the open-source, the open-weight trend. Right? These are the companies that are benefiting massively from that trend, and it's not the only kind of route for inference. There are, you know, inference for US-based models, etc., etc. But the vast bulk of it is, oh my god, I'm sourcing Qwen, I'm sourcing Kimi, I'm hosting Qwen, Kimi. I want to use it as Cursor. I want to get someone to provide me some inference. These guys exist, and you know, they have lots of customer skew at the high end. I believe, you know, companies like Cursor are probably big customers of all these guys. At least they were until they were acquired by—and probably still are, right? So yeah, I mean, it's a great—candidly, I think, you know, it actually goes back to the point I made earlier. Some businesses where—and I think, you know, you can look at and say, oh my gosh, the cost—you know, you have margin compression in your future because you're buying your compute from the neoclouds, and you know, you're offering this product, and are you going to be scrunched? And margins, you know, were probably slow for a while, but now the beauty of it is demand is massive. So whatever compute you have today, whatever compute you've already signed up for—and these guys sign up for compute from the neoclouds in general—are starting to build their own. Whatever compute you own now, you can charge way more, which means what looked like a lowish gross margin business has now probably become a very attractive business. So not only are they probably going to grow 5x to a billion, but they're probably growing 5x to a billion with expanding gross margins.

Yeah. And just to be clear for people, what that means is—yeah, and that's exactly where I thought they'd be, and good on them, right? 30%. In other words, what they're saying—and this is going to be an—and this, and I agree with that sentence. It also means that the challenges I hinted at are there in the future, right? Because what they're saying is, if I'm buying data center compute and then effectively selling data center compute with hosted LLM, at some point I'm going to want to own my own data center assets to have more control of my destiny, which means vertically integrating downwards, which also means a ton more capex. So these are going to become way more capex-intensive businesses. There is a risk of commodification here, even though—even with massive complexity and massive capex.

And just to dimension that, because I agree, Jason. I mean, look, I'm an app investor. It hasn't—you know, you look back and you—what's going on here? You've got companies like, you know, Fireworks doing a billion dollars. There's very few apps companies doing that. And, you know, zooming out a million miles, my mental model is I divide the AI world up into three buckets. It's the making AI, the infrastructure layer, right? And you're right, the spend there is $800, $900 billion a year. Then there's the two foundation model companies themselves, and they're doing plus or minus $100 billion a year, right? And then taking those guys out, rounding up every other apps company, right? You struggle to make 40 or 50 bill. You struggle. You start with Cursor at four because I think coding is an app. You know, by the time you're chucking in Harvey, you're adding two, 300 million, right? It's amazing. I mean, just the difference in spend. And, you know, at some point, the people spending a trillion dollars a year are going to want some apps to pay for all this, right? But right now the volume has—it's been front-end loaded on the infrastructure side, and at some point the revenue has to match it. But right now, infra has been the place to be. Like, there's probably more money being spent on training data for the foundation models—you know, the Mercor, Surge, and that—than pretty much any app company outside of Cursor. In fact, probably the sum of all the apps companies outside of Cursor, right? You know, are probably less than the amount that Anthropic and OpenAI are spending on training data, which is just amazing.

I mean, yeah, maybe if you start throwing in on the other side the consumer products like Higgsfield, you get to roughly the same place, but it's astonishing the scale of the investment versus the scale of the apps at this point means that all the action's on the infra side for now.

Yeah. Or absolute state-of-the-art kind of small parts of the overall task. But agreed—little parts, right? Like, I mean, again, you're going to want to use the expensive tool for the expensive parts, right? And you're going to want to use the cheap tool for most of the parts, from the customized tool to your usage.

I think the answer, though, is really a derivative of the big question, which is—you know, because your statement, your companies are going to want their own model, is probably true, right? And the real question is not that. The real question is, will that be additive to the rough trajectory of the foundation models as it's established today? In other words, coming, you know, at or close to 100 billion combined revenue, growing nicely, or does it start to take away significantly? Because, you know, to answer your specific question, if the foundation models continue to grow—and we just saw an article in The Information that, you know, for all the training data companies, the vast majority of their revenue comes from the foundation models, to which your correct response is, no shit, of course it does, right? If that continues to grow and you have an additive market in enterprise of all these companies—you know, JP Morgan building the JP Morgan model on top—then, you know, it's net expansive, and net expansive is by definition good and reduces customer concentration. And I think that's what people like Mercor are forecasting, right? If, on the other hand—which is hard to contemplate today—if these enterprise models, if these open-weight models really impacted the growth rate of Anthropic and OpenAI, then obviously when your 80% customer slows down, it would have a significant impact on your growth rate, right? But, you know, if it's any consolation, Harry, if that happens, worrying about your Mercor valuation will be the least thing people are worried about, because you'll see an implosion of much bigger market cap entities, right? And that, frankly, is the billion-dollar question. You know, can these two foundation models maintain their growth trajectory, which is starting to become profitable, at least in the case of Anthropic, in the face of all this open-weight competition, in the face of this pushback on costs and basically kind of push for ROI? If they can maintain this trajectory for even another one or two years, then everything's fine and everyone's fine. And right now the data says they are. If you start to see slowdown on those two ARR growth rates, then, you know, all bets are off, because the pressure—because the amount of commitments they've made assuming that 10x growth rate continues will mean that even if it slips to a 2 or 3x growth rate, there's going to be a mad scramble.

Yes, Harry, it will impact. It might impact at 1% or 50%. What you're really saying—what the question you're really trying to ask is, does it produce a sustain—you know, does it reduce that growth rate to sub-100% within one or two years? Right? And the answer to that question is, I genuinely don't know. And if I did, I'd be trading that stock. Because let me be clear: if you know the answer to that question, that one question, you know the answer to the entire direction of the US stock market for the next two years, because all the hyperscaler RPO, all of it, is a function of the commitments they've gotten from the foundation model companies. And yeah, you can say if the open-weight models explode, there will be demand for inference. And yeah, you will have this kind of transition from, oh, I sold it to OpenAI, but I should have sold it to, I don't know, Cursor or Base10 or someone else, and the capex will get repurposed. But it will be a big-ass dislocation, and I just genuinely don't know. I mean, it's the million-dollar question.

And we don't have time for it, because I do think we should spend at least half the show on stuff other than AI model companies. But I think Jason's insight is correct about price. And if this was a software product with no gross cost of goods sold, that's what they'd do. I mean, Microsoft—and this is one of the big—I've seen a bunch of articles on this. Again, it's as you'd say, Jason, Captain Obvious, but it's worth emphasizing. In the great software wars of the last couple of decades, someone like Microsoft was able to just use price ruthlessly, because there was zero cost of goods sold, and they just bundled the browser in with the operating system, bundled all Office in together, didn't cost them anything, and it just wiped everyone else out. But as you're pointing out here, there are real—even at the margin, even after you've fully paid for your training costs, there are real physical costs to serve these models. And you got to cover your nut. You got to cover the marginal cost of the model, of the inference, which gets you to, I don't know, two, three bucks kind of blended average token. Then you got to recover the cost of the training, and you got to recover it pretty damn quick, because it only lasts, you know, 12, 24 months before it's obsolete. And then on top of that, you want to make extraordinary profits, because you're being valued at 20 times revenues. And if you're valued at 20 times revenues, you better be like Microsoft with 40% operating margins. So when you look at all that, you're right. You know, it's kind of back to the thing I said: you can squint at that and say, 'Ooh, there's lots of things that could go wrong here.' When you look at those fundamentals, the thing that's saving you right now, if you're growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut. Right? The minute that growth rate stops, Jason, if the only way you can keep that growth rate up is by lowering your kind of price per, then your gross margins start to deteriorate instead of continue to improve. That in itself would be a different ballgame. So you are right, price could solve it, but it would be a painful way to solve it.

Yeah. You remember the old days in software, Jason, when you'd have to say, 'I promise I'm only using this for personal use.' You remember that? And licensing, right? Well, if you're telling those nice Claude people that you're only using your personal subscription for personal use, they're going to find you, dude.

Well, yeah. It's just—Fable is very good.

Very good.

I mean, a vertical dentist company like—Benick making 500 million for his—oh, that is an AI. Sorry. Sorry. Sorry. That's an AI story.