Limitless: An AI Podcast - AI Bubble Update: Kimi K3 Sends China into Chaos

Episode Date: July 21, 2026

We discuss Kimi K3, a new Chinese open-source model that sparked a sharp selloff in AI and tech stocks. The real question: does it signals a real shift in the AI race between the U.S. and China? And who will the biggest winners be: model labs or the compute and infrastructure providers?------🌌 LIMITLESS HQ ⬇️EMAIL US:           [e-mail]:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------TIMESTAMPS0:00 Kimi K3 Shakes Markets0:49 Open Source Disruption2:54 China’s AI Message5:22 Demand Outstrips GPUs8:59 Open Source Backlash13:58 The Bubble Debate17:38 Infrastructure Wins Long Term22:36 Breakthroughs Still Loom24:48 Outro and Updates------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠Josh works with Anthropic as a contractor. All views expressed are his own and do not represent Anthropic, its leadership, or its affiliates. Nothing in this episode is investment advice.

China just released Kimi K3, the most powerful model to ever come out of the country. The question on everyone's mind is, is this going to create a bubble? Or is it already happening in front of us? Kimi K3 is the most impressive model, leapfrogging the frontier American models and some benchmarks, and doing so in a way that is significantly cheaper and open source. So the conversation over the entire weekend is not only is there a bubble, but how credible, how strong is the American mode as it comes to China? We always discuss the timelines and how America is normally ahead by, let's say, six to 12 months.

It seems like that timeline has kind of shrunken a little bit. And you see how these two countries are kind of juxtaposing the way they're going about it. There was two speeches that happened on Friday. One was from America blaming China on rigging votes. And the other was China very strongly presenting themselves as the up-and-coming AI leader. So there's a lot to unpack here, probably starting with the market, which was nothing but read over the last couple of days, because you have to assume there is a little bit more uncertainty than normal as it relates to just AI models in general and how it turns into China. I mean, look at this, look at the screen share. That is a wall of red. Yeah. So, so last week was basically China's retaliation to the American frontier models. And it

happened in a pretty big way. As you mentioned, Moonshot Labs, which was the creator of this model called Kimi K3, which is completely open source, is now the largest open source model ever. It comes in at around 2.8 trillion parameters, and it's super cheap to use when you compare it to Fable 5 and GBT 5.6. Now, the question on everyone's mind is, is it necessarily good enough? If you look at the NASDAQ market and how it responded last week, it shaved off hundreds of billions of dollars, like it was down completely in response to this model. The reason being is, it doesn't matter if it competes with Fable 5 or if it's as good as GPD 5.6. What matters is the cost.

When you look at Fable 5 specifically, you're paying around 50 different. 56 bucks per million tokens, input. And when you compare that to Kimmy K3, you're paying around 50 cents to a dollar. So if you're an enterprise that is burning tens to hundreds of millions of dollars, or if you're a software engineer that is spending all your tokens on an expensive $200 a month membership, you might want to consider an alternative model. And in this case, it'll be a Chinese model. Now, the craziest part about this is the valuation of Moonshot Labs in January of this year was $4.3 billion. Now, after they released Kimmy K3 just last week,

they announced that they're raising at a new valuation of $30 billion. That is still 30x cheaper than Open AI's $1 trillion valuation, and Anthropic is valued at the same amount. So the question in everyone's mind is, are the American labs overpriced or are the Chinese labs underpriced? We're going to have a series of Chinese IPOs coming from the AI. labs such as Deepseek and Moonshot over the next roughly six months or so. And the Chinese president Xi Jinping reinforced this in a very big way on Thursday, where he took the stage in Shanghai

in person at the World AI Conference the first time since 2018 to make his first public commentary on the AI landscape. And he had some pretty major takeaways for everyone. Yeah, like you mentioned, this is the first time since 2018 taking the stage. This is a fairly big deal. And his kind of thesis could be summarized in one line, which is, should not be a solo performance by a single country. It should be a symphony of international cooperation, followed by a shot at Washington, which said we should jointly oppose overstretching the national security concept in the field of AI or placing one country's security over that of others. And it was really interesting. I mean, I mentioned this in the intro briefly to see that kind of juxtaposed against

the other speech that was happening simultaneously on Friday night, but back in Washington, where President Trump, he went on to accuse Beijing of the largest compromise. of election data and history. So while one president is standing up in front of the AI conference for the first time offering the world basically free frontier intelligence through these open source weights, another one is complaining that we've been robbed and that we're not simultaneously while this kind of lead that we have is continuing to shrink. And there's this post, he says, I shared it with you just before recording. I loved it. It was from Travis Kalanick. And for those who don't know, Travis Kalanick, founder of Uber, he's the guy who as famously went into China and convinced them

somehow to allow him to let robo-taxies into this country is very familiar with how they work culturally. He said, in my experience, China rules different on these topics. If you're in China and you tell a business colleague, dude, they frauded me, they took my stuff. The natural response is, why did you let them? Get your shit together. And I think this is kind of beginning to show the two differences in cultures. We know that China doesn't have any regard for IP and therefore is able to distill the models, is therefore able to rip off all of our IP. And that's how they're able to accelerate so well. I think my question now, EJA, are they actually going to ever be able to flip the United States in terms of capabilities

if they continue to just distill and follow along? It feels like they can maybe get that lead closer and closer, but in order for them to actually break through and create something better, it will require some technical breakthrough that they haven't quite shown. So when we look at the market, does it feel overreactive that it's down so much, even though they don't actually have a frontier model and getting a true frontier model would imply a lot of breakthroughs that they've never been able to do. I mean, they can't even serve this model yet. They're overflowed with just Kimmy requests.

Yeah, and I think what you're referring to in that last sentence is Moonshot Labs announced over the weekend that they don't have enough GPUs to support everyone sign up and request. So they've had to pause new subscription. So it's in over-demand. And I actually have this really cool chart from the folks at Open Router who basically allow anyone and everyone to get access to these open models as well as American Frontier. models, and it shows that 58%, near 60% of users in America now are using open source Chinese models. So we're now flipped from using American frontier models in the majority of token

expenditure to Chinese models, which is very significant. But to answer your question more directly, Josh, it depends how you define what a breakthrough is. Now, a breakthrough in itself is Kimi K-3. It may not be Fable 5 worthy, but the fact that it's been able to catch up to within like, you know, striking distance of Fablify, but not quite as good, but at a fraction of the cost, could be considered a breakthrough. If you look at their white paper, you've seen that, like, there's a bunch of new research things that they've been able to figure out. I think they call it Kimmy Delta Retention, attenuation, or whatever hell that's called, which allowed them to efficiently scale two and a half times more efficiently than any other Frontier Lab has done.

Now, if you look at America, they haven't been able to do this. What they have been able to do is throw tens to hundreds of billions of dollars at the problem and hope that it would fix itself. So they've been able to make the frontier leap much more efficiently for a much cheaper cost. Now, if you ask me if the American labs stopped putting out new models, so let's say we never got anything later than Fable 5 from Anthropic going forwards, I think it would take Chinese labs significantly longer to catch up with them, simply because they have been distilling. You know, you talk about their disregard for IP.

That is primarily the way that they've been able to. able to catch up to America Frontier models. Now, my question that I throw back to you, maybe, I'm curious what your thought is, is distillation necessarily malpractice? I think everyone is kind of doing it. Like, I looked at Thinking Machine Labs last week, which is Miramorati X OpenAICTO's new lab, and they announced this new open source model. It's their flagship model got a lot of attention. Every American spokesperson was like, hey, American open source is so back. Look at this new cool model. If you look at the white paper, they basically distill Deepseek V4's architecture from China.

So we're kind of like all distilling and copying everyone's kind of trends and research to collectively make better models. And I don't necessarily know if it is a bad thing. The final thing I'll say is Fable 5 was restricted and banned for a while. It's only actually been fully live for like three weeks right now. There is no way Moonshot Labs was able to fully distill the model and create Kimmy K3 in that short time span. So they're doing something different, is what I'm saying. Yeah, well, Fable is like one model. I'm sure they have been doing this with others.

And I think when it comes to research, like algorithmic improvements similar to what Muramari's company was doing, it's a little bit different than blatant IP copying. Where like, sure, there are these technical techniques like the transformer that everybody uses, but then there is the actual IP built on top of it. And that is where, I mean, it is just illegal to do these things. It is illegal to steal other companies' information like this. And it's obvious that China doesn't have any regard. for this. So I would say it is different in the sense that this is just like blatant IP

sealing. And while it is sure a gray space and a lot of companies distill from a lot of places, this is just blatant illegal disregard for copyright. And I think that is like where you can comfortably draw a line if you are the United States. I worry that perhaps there is an overindexing coming just before we were recording each as you mentioned there's a possibility that open source models are going to start to have some pushback, perhaps from the government. There is some resistance happening towards that. And should that happen, I do worry what the world looks like then, because when you think about the costs of these things, like frontier intelligence models, you're paying, what, 10 to $50 per million tokens. And for these frontier open source models,

these Kimmy models, you're paying, what, a dollar, 50 cents, maybe $2 on that, yeah, 50 cents. And it's like that, if those start to get banned here in the United States, you're almost in a way forcing consumers to pay a significantly higher price for similar levels of intelligence. Like if you're paying a 50 times multiple for a model that is perhaps 10% better, that seems like it is problematic for just general users in the United States who don't want to pay perhaps $50 per million tokens, but they also want access to close to the frontier. There isn't really an option for them right now. And banning these open source models also doesn't seem like a good place to be either.

I think people are just tired of overpaying and they're starting to kind of like look at their wallets and budget spends and thinking, well, maybe if there is a sexier alternative, I'll go use it. I don't care if it's Chinese, it's open source after all. Now, it hasn't escaped to me that the irony of China being the open source proponents, the fact that in Xi Jinping's speech over the weekend, he basically said, like, hey, we need to be openly collaborating with each other. This can't be a solo kind of like country thing. And we also can't be imposing export restrictions. We've just got to just got to let people build, man. Compare that to like Trump's America or rather, you know, capitalist America who is supposedly meant to be proposing the same kinds of things, but they're actually like, you know, we might actually end up banning the Chinese open source models. Like, it's just ironic to me. But if you're looking at this objectively, the question that you should be asking is, are these models actually good? Like, what are they actually good for? If they're not Fable 5 worthy, like, what can I be using it for? And I think there's a really good example from David Sacks over the weekend here where he goes, Kimmy K3 just

fixed 15 critical security bugs that Codex and Fable refused to because of cyber guardrails. And over the weekend, I was using Kimi K3 quite a bit. And I will tell you, the number one feature has got nothing to do with how smart it is, or how long it is, or how cheap it is. It's the fact that I can ask it a question and it just answers. Or I can give it a task and it just does it. There is no relegation to a lower class model Kimi K2. It just kind of like does the thing. If I did the same thing for Faber 5, if I did the same thing for GPD 5.6, there is a very high probability that they would just say, sorry, I can't answer that. Or sorry, I'll give you a cheaper, crappier model to do your thing.

And it's quite freeing. It's kind of like the feature in itself. Like, I don't really care about the benchmarks. It's the fact that I have a near enough really good model that can just do the work. Is it slower? Yes. Does it use more tokens? Yeah.

Is it as cheap as I could have hoped? Maybe not. But I don't really care because it's unrestricted. I don't know if that makes sense. but like, yeah, that's my experience. Yeah, and this is the difficult part, too. It's because, like, that wasn't really always the intention either. It's like, if they wanted to release this model in the public,

the government would not allow it without these safeguards, and they tried and it got banned for almost a month. And that's part of the problem that I think a lot of these labs, and that's kind of what David Sachs is referencing now, is that if you have a frontier model, you are now burdened by the approval process of the U.S. government before it comes out, which will oftentimes include these restrictions that do set things back. And the reality is that, I mean,

Kimi K3 is basically close to frontier level cyber capability, and it's going to be freely downloaded soon. And if it's not Kimi K3, it will be Kimi K4 in a couple of months. And the reality is that while these safeguards are very important to have when they are truly at the frontier, the moment it becomes freely and openly available elsewhere, it loses all the advantage and then therefore puts everyone who is burdened by those safeguards at a disadvantage. And that's this really difficult thing is, if you're the AI labs, you want safe intelligence that is aligned with our best human interests. And you also want to offer people the best tools possible. And the U.S. government is saying, well, this is great, but they need to be

safe. We need to approve these things. And all the while, this is slowing things down. I mean, how long did it take for us to get GPT 5.6? It was sitting there ready for many, many weeks until they approved it. And during that time, we got Kimmy K3. And there's this really difficult disconnect that's getting hard to wrap my head around as it relates to the Chinese frontier labs, just kind of burning, scorching the earth. I mean, this was very much meta's plan when they first were trying to come up in the world of AI

where they were just going to scorch the earth, create the best open source models, distribute it to everybody, and lower the margins of the frontier labs. And the reality is that China's just kind of taking that and grabbing me by the horns and running with it. And I'd say doing a pretty good job so far. Yeah. Okay.

So something that we have discussed a lot on this show, Josh, is valuations, its IPOs, it's generating wealth in the market. Is there a bubble or not? You just get into it. Well, okay, yeah, there's definitely a bubble. The question is, should the bubble be popped? And this is a very visceral example that we're showing on the screen right now, right? So you have opening eyes valuation at the top, which is near a trillion dollars. And then you have DeepSeek, which is a fraction of that, at $71 billion. And then Moonshot this week who got a valuation upscale to $30 billion.

So it's sitting at 30x less the price of Open AI, despite being near frontier capability of their flagship model, GPD 5.6. Now, we know that there is a ton of frontier American IPOs that are happening by the end of this year. You've got OpenAI, you've got Anthropic, and you've got a few more chip companies as well. And then you've compared that to the Chinese IPO market, and you've got Deepseek that is aiming for a valuation that is considerably less than that. You've got Moonshot, Kimi K2, Kimi K3, sorry, doing the same. exact thing. And the question is, which one is overpriced or underpriced? Now, my guess, if I had to have a take here, would be that I think the Chinese frontier labs are going to bring the valuation of these open AIs and anthropics down in the medium term. And what I mean by that is they have

proven that you can get frontier capability for a fraction of the cost. But the premium that sits with American frontier labs sits in the research, sits in the talent, and sits in the fact. And that whoever is closest to building self-improving models will end up capturing quite a bit of future economic valuation that any of these AI models create. And right now, the best bets, if you had to place it on polymarket or wherever, would be with Anthropic and Open Air.

So it's very important. You can't just look at models right now. That's just a snapshot. You need to be thinking of Fable 6. You need to be thinking of GPT6. Can those models build themselves and build the next versions of themselves? If the answer is yes, then you could justify some of these valuations.

But it is a bit of a leap. It's a bit of a thought boy kind of take, I guess. I don't know. Do you think differently? Yeah, no, it could be true that both open AI is overvalued and moonshot is undervalued. And the reality is that they should exist somewhere in between. The big difference, the reason why we see this 28 times multiple on these different companies is because one has a tremendous amount of revenue and a predictable future amount

of revenue and the other does not. We just saw how Kimmy had a huge success. They were unable to satisfy the demand. They had to shut down the signups of new people using their product because they just don't have the GPUs. So there is just not enough compute available to them in order to earn a higher multiple at this moment. They're offering it for free.

There isn't predictable revenue drivers on it. It's just not possible. Open AI, on the other hand, is trading at a very high multiple on its revenue. But the reality is that, like you mentioned, it is somewhat close. You have to imagine getting closer and closer. closer to this self-recursive learning loop. And we have no idea what that's going to look like, but we can suspect and we can hypothesize that the totals demand for tokens is going to continue to be up only for basically forever or for at least a very long time. And should that be the

case, there are only a few companies that are able to satisfy that demand, let alone satisfy that the demand at a high quality. So it's really this race for, hey, MVIDIA. It's this race for GPUs. And it's funny because at the end of the day, the real winner here is Jensen. And I have a feel this is why he's been so supportive of open source in general because it just it requires a tremendous amount of GPUs it makes AI more accessible to more people and as other companies begin to make their own custom ASICs or their custom accelerators like Traneum and like the Google TPUs there is going to be this again increasing demand throughout open source communities of the GPUs and I don't know if I'm looking at the market right now I'm seeing things getting crushed I'm kind of I'm kind of bullish I'm

kind of like, wait a second, this is just a discount. Like none of the market sentiment, or I guess none of the market forces have really shifted in a meaningful way. The only thing that's changed meaningfully is the sentiment because people are a little bit concerned about where China stands. But the reality is to me is that while China has, you know, pretty great models that are very efficient and very open source, they're still so far away from the real frontier because the frontier is so much more than just intelligence per token. It's the infrastructure. It is the sales teams. It is the business development is the reputation. It's the American institutions trusting your product. And so much of that is already baked into open AI, into Anthropic, into these large AI labs,

that it doesn't seem to me like there is a problem. Like, okay, these models are great, but they can't serve them. They aren't going to pass the frontier. And Anthropic and Open AI, in the meantime, are just going to continue hurling so fast forward with very predictable revenues that are going to, I mean, expect it to go up only. And I think that makes me feel like this is a pretty actually decent buying opportunity and that maybe there is. like a little bubble pop, but that happened for 72 hours, now we're back on it way. It seems to be like the market has become very resilient. And even though it's short-sighted, there is this kind of weighing machine over a long period of time that will work itself out to continue to guide

upward, I think, is the idea. Yeah, I feel the same. When I look at valuations, when I see how volatile the stock market is, I do think it is based on a lot of people's emotions. But the fact of the matter is whether you end up with a frontier model that costs a lot of money or if you end up with an alternative model that costs a fraction of that, people are going to use the thing. People who want frontier intelligence and can't afford the Kimmy K3 intelligence deficit will end up using Fable 5 in paying for the extra cost because they get so much more, they make so much more money or they create so much more output on the back end. The people who are hobbyists, who are young founders and startups who can't afford the expensive models, will use Kimmy K3. It'll take

take longer, but they'll end up building amazing things. And Jevin's paradox will play out at the end of the day. People will want more tokens. They'll want to consume as much AI compute as they can for various different use cases. The winners at the end of the day are you who are listening to this. The winners at the end of the day are the businesses who want to use AI to accelerate their own purposes. And I think that if you're looking at where, you know, parking your money might make sense again, like not investment advice, it would be kind of like at the bottom of the stack, the people that are creating the infrastructure, the people that are helping serve these tokens, the people that are helping power these GPUs, the people that are creating the memory to remember all the context

that all these different models are going to create about you. The fact of the matter is, if you look at Trump, you look at his speeches and the policies that he's making, if you look at Xi Jinping, these guys are adversaries. He's got the opposite kind of take. The ultimate unifying thread is more consumption, more demand for more AI, more GPs, more infrastructure. And that is technically the sector of the companies that are going to end up winning. The most important Kimi announcement over the weekend is this one. Kimi K3 has received far more of them we expected, and our GPUs are feeling it. They've stopped signing members up because they've run out of GPUs.

Now, whether they're using Huawei's GPUs, which was announced at the same conference that Xi Jinping had a speech at or whether they're using Nvidia GPUs. The point is infrastructure is the moat, and that's where capital is headed. So whether memory stocks are down, whether these GPU stocks are, down, I think in the long term, I'm pretty bullish. I know I'm pretty bullish relatively on this show, but I think we both are. Please prove me wrong. I want to see the count of these is. Well, listen, it's been correct so far because, I mean, since the conception of the show, we have essentially been up only. And the market forces continue to point that way. You know,

it's like there is somewhat of a case we made that perhaps there being just two AI labs isn't going to work out great for everyone else. But the people who it works out for, like you mentioned is the people who are providing the infrastructure to build this out. Because even in the case, let's say Kimmy K4 is a true frontier model, they're still going to need a huge amount of compute and all of the other labs are going to need a tremendous amount of compute to keep up. And the thing with this competition is, like, so long as the demand is there, the amount of resources required to supply that demand is just going to continue to be up only. And regardless of these outcomes, regardless of the open source starting to ship away at closed source or closed source

maintaining its moat and just shooting ahead through this self-recursive learning, both of those realities require so much more compute, both to serve the larger models, and both to keep up with this increased demand. Clearly China can't do that. So I'd say safe from disaster, once again, we've survived our second deep seek moment, perhaps. It only took us 48 hours to kind of rebound and get back to it. But we'll see what comes next. I mean, there are rumors, Ejazz, and I don't want to speculate,

but I have heard that Ilya is beginning to wake up and has been receiving some pressure from investors to share what they've been working on. Iliates at Scareer. Yes, safe super intelligence. That guy co-founder, Open AI. He's been working quietly for how many years now. And I'm seeing on X all the time. It's like, hey, the investors at this company are kind of like getting a little restless. They know what's going on, but the world doesn't know and they want to get that bag secured. So I'm exceptionally excited for announcements like that, where perhaps there will be this huge sort of improvement in efficiency or algorithmic improvement or something new and novel that can really push things forward in a big way. So we are always a

single breakthrough away from this catastrophic event, either catastrophic up or down. We don't know. But for now, safe from explosion of the bubble for now. And again, like, I mean, China's just on their way and they're going to continue to keep firing with no regard for IP or any sort of American laws. And I think that's kind of where we could leave it. It's like, hey, Kimmy K3 is a great model. You can't use it right now because it's a little too good. But it's a great model nonetheless. And the American markets felt it.

And hopefully it keeps everyone on their toes moving forward very fast when it comes to iterating through these models. Yeah, that's pretty much it. I think if there was a single sentence that I would want to leave you folks with is we can't keep ignoring China. Like those days are gone. They are here to play in a very meaningful way. And they're leading the open source frontier,

which means that you can own your AI at home. The cost of hardware is going to get a lot cheaper, and you'll be able to have the privacy access and the cheap efficiency cost that comes with that. So keep an eye on the pulse, on the radar. If you can't do that, if you're not on X the entire time, hey, we've got this really fun podcast called Limitres. Subscribe to us.

We do four episodes a week. We post twice on a newsletter to hundreds of thousands of folks to cover anything and everything that is happening at the frontier of AI and technology. So if you want to keep up today, make sure you subscribe to us. Leave us a comment. If you disagree with anything that we've said today, or if you agree, we'd love to hear from our fans as well. We're always looking for where we might be wrong or different investment angles and stuff

like that. So please let us know in the comments. If you're listening to us on Spotify or Apple, again, please rate us, leave us a comment. It helps us out massively. And I think the final note, I will say, is Limitless has officially spun out. We are on our own. We're keeping them lights on at home ourselves. So if you are a company that has a product or service,

So if you know of someone that has a product or service, that you would like to share with our very palpable audience, please let us know. Reach out to us. Our email is linked in the description below or DM us on X. And I think that's it, Josh. Anything else? Man, that was pretty good. Dude, you got the adjos dialed now.

I think that was that covers just about everything and then some. Clearly, we've been doing this for a lot. This is what, episode 210 now. So yeah, you made this far. You know the deal. Thank you guys so much for watching. And we'll see you in the next one.