Loading summary
Transcript1 lines
- [00:00]
A
Welcome back to the rundown for another weekend deep Dive. Today we are talking about Kimik 3, a new Chinese AI model that sent shockwaves through Silicon Valley and Wall Street. Last week, a Chinese startup called Moonshot AI released a model that rivaled some of the best models from OpenAI and Anthropic. And within hours, tech stocks were selling off. So in today's episode we'll break down what Kimik3 actually is, why cheaper and open source AI models could completely change the economics of the AI, and who the biggest winners and losers could be of this new AI reality. We got a great one for you today. Let's dive in. Before we get into why markets are freaking out, let me tell you more about Kimike3. Kimi was released by a Beijing based AI startup called Moonshot AI. Moonshot was founded back in 2023 by 33 year old Yang Zhilin. And this guy has a pretty stacked resume. He got his PhD from Carnegie Mellon and he worked at both Meta and Google Brain. He eventually went China though to build his own AI lab and founded Moonshot AI. And for the last year and a half or so, Moonshot has mostly been living in Deep Seek Shadow. Deep Seek became the breakout Chinese AI startup after releasing their R1 model in early 2025, which also caused a market panic. What's interesting about that is that Moonshot also released their own model on the exact same day as deep seek R1. But nobody really cared at the time because everyone was obsessed with with Deepsea. Well, Fast forward to July 17, 2026 and moonshot finally got its moment. They released a model called Kimi K3 and this thing immediately caught the attention of both the tech industry and Wall Street. See, this model is almost as good as the leading edge models from Anthropic and OpenAI. According to independent benchmarks, it ranks third in the world behind only Anthropic's Fable 5 and OpenAI's GPT 5.6. And in some coding benchmarks, Kimi actually beats both those models. And just digging a little deeper. What makes Kimik 3 so impressive is that it's a massive model with 2.8 trillion parameters, making it the largest open weight AI model released so far. And we should talk about the open weight part here because that's pretty important too. An open weight model basically means that developers can download the underlying model and customize it however they want and run it on their own hardware instead of being forced to pay Moonshot for every time they use it. Moonshot is planning to Release the model weights of Kimik3 on July 27. So they might have already come out by the time you're listening to this episode. Now, what was so shocking about this model release is that a relatively unknown Chinese startup with about 300 employees and a $20 billion valuation built a model that was competitive with cutting edge models coming out of OpenAI and Anthropic, which have massive headcount and billions of dollars in funding. This is kind of like Kate Verde tying Spain at the World Cup. And the part that investors are paying attention to is the price. MoonShot charges about $15 per million output tokens for Kimmy K3. For some context here, OpenAI's Top Model cost about $30 per million output token, and Anthropic's Fable 5 cost about $50 per million output token. So Moonshot is able to offer something that gets pretty close to the frontier level performance while undercutting the price by roughly 50 to 70%. So now wall street is asking questions and panicking about the economics of AI. So let's talk about it. Okay, so why did this AI model from a relatively small Chinese startup cause tech stocks to sell off? See, the thing is, the entire AI trade and a big chunk of the stock market for that matter, is built around one major assumption. And that assumption is that the best AI models will stay scarce and expensive and that building models will require spending more and more money. You know, that's why OpenAI and Anthropic have raised tons of tons of money at near trillion dollar valuations. These AI labs have the best models in the world. And investors assume that because they have the best models, they can charge premium prices for years to come. Despite OpenAI anthropic not being profitable today, investors think that eventually they'll generate enormous profits. And there's like a ripple effect happening across the economy here. Because OpenAI, Anthropic are able to raise all this money. They've committed hundreds of billions of dollars towards things like chips and data centers and computing power, which has propped up the stock price of companies like Nvidia am, other chip stocks, not to mention all the companies benefiting from the AI data center build out, like power companies and electrical companies and construction companies. I mean, the AI spending boom is so big, it's literally showing up in US GDP numbers. So think of it like a chain. Premium models justify the premium prices, Premium prices justify raising a ton of money. All that money is then leading to the biggest CapEx spending in corporate history. And that CapEx is the revenue engine for a big chunk of the stock market today. So when Kimmy K3 bursts onto the scene, it's investors start questioning the assumption at the very top of that chain. Because if a 300 person startup in Beijing can build a model that's nearly as good as the best American models, well then what happens to the premium pricing? What if AI models just become a commodity at some point? And that's key, because when a product becomes a commodity, it's a race to the bottom when it comes to prices, and profits tend to go along with it. When you look back though, for the last three years, AI models have been the opposite of a commodity. Companies have paid millions of dollars to use models from OpenAI and Anthropic because there wasn't many alternatives that could do the same level of work. But now companies might be able to use a much cheaper open weight Chinese model for most of their workloads and only pay Anthropic or OpenAI for what they really need the absolute best models for. In fact, this is already starting to happen. Doordash said that they use the moonshot models for lower level work and they use Anthropic models for their hardest cutting edge tasks. Airbnb is also doing the same thing. Their customer service agents run mostly on Alibaba's Quinn model. CEO Brian Chesky says that that model is fast, cheap, and by the way, even Microsoft is reportedly testing whether Kimi K3 could eventually power some of their features inside Copilot. So American companies are starting to swap in Chinese models the same way your dad would buy the store branded cereal because it tastes basically the same as the name brand cereal, but cost half as much. And that's what Wall street is trying to price in right now. Could more and more American companies choose to ditch OpenAI and Anthropic and go with cheap Chinese AI models? This adds a level of uncertainty to the AI trade, which by the way, was already looking shaky these days. But here's the thing. The rise of cheap Chinese AI models isn't bad news for everybody. In fact, for some parts of the AI economy, cheaper intelligence might be the best thing that could possibly happen. So let's talk more about some of these potential winners and losers. Okay, so let's talk about some of the winners and losers, starting with the losers, because the most obvious losers here are are the frontier labs themselves, OpenAI and Anthropic. Their entire business model is selling premium intelligence at premium prices. So Kimi K3 is a direct attack on that pricing power and the timing Here couldn't be worse for both these companies because both of them have filed for an IPO at a potential trillion dollar plus valuation. It's probably a good thing that both these companies are still private, because if they were publicly traded, it would have been a bloodbath for both of them. One analyst estimated that the day after Kimmy dropped, about $314 billion got shaved off the valuation EST estimates for OpenAI and Anthropic. Now, to be fair, I don't think that Kimmy suddenly makes these companies irrelevant. I mean, their best models are still better overall and they've built massive ecosystems around products like ChatGPT, Claude and their coding tools. You know, Chat GPT is closing in on a billion users here, but the risk for them is their corporate business. If their corporate customers stop using their expensive models for everything, like the doordash example that I just gave, that could have a meaningful impact on their bottom line. And both these companies are leaning on the enterprise business to be the profit engine for the company down the line. If the launch of Kimmy K3 and other cheap Chinese models changes that picture, investors might start thinking twice before paying a trillion plus dollar valuation for these companies. And this repricing of OpenAI and Anthropic has spilled over into the rest of the AI trade. Companies like Nvidia, AMD and other AI names fell after the Kimmy K3 release. But I think this is where the panic starts to fall apart. Because cheaper AI might be bad for companies like OpenAI, Anthropic, but I think it could be great for companies providing the infrastructure that runs the AI. There's a concept called Jevons Paradox which basically says that when technology becomes cheaper and more efficient, people often end up using way more of it. We've seen this happen throughout history. As steam power got more efficient, we used more energy. As electricity got cheaper, we electrified everything. I mean, just look at the Internet itself. As Internet bandwidth got cheaper, instead of using less Internet, we're now streaming 4K Netflix while FaceTiming someone and scrolling TikTok at the same time. So the same thing could happen with AI. As AI becomes cheaper, every employee and small business might spin up multiple AI agents doing tasks for them. And every single one of these AI tasks still need hardware to run on. I'm talking memory and GPUs and data centers to sit in and electricity to power all of it. And that's where I think Kimi is very different from the Deep Seek Panic. When Deep Seek R1 came out in early 2025. The big fear was that AI models could get dramatically better while using far less computer. That's why Nvidia lost nearly $600 billion in market value in a single day after the Deep SEQ model came out. Investors thought that we would be in a world where we didn't need as much AI hardware as everyone expected. That's not what ultimately ended up happening. And that's why the Deep Sea panic was overblown. And this is where Kimi is so different. Right? Kimmy K3 is a massive model. It has 2.8 trillion parameters. And Bloomberg estimates that the full model requires 1.4 trillion terabytes of memory. So deploying something this large at scale requires high end AI chips and a ton of memory. And that's why Kimmy K3 could actually be good news for companies like Micron and sandisk and SK Hynix and Samsung and Nvidia and AMD and all the companies making chips and memory. Not to mention, this could also be a big boost for software companies that can make use of the cheap AI. But I want to zoom out for just a bit because there's an even bigger question hanging over all of this. And that question is, how did a Chinese startup build something so good despite years of US restrictions designed specifically to prevent this from happening? And does that mean that China has finally caught up to the US when it comes to AI innovation? Well, there are some people that are accusing the Chinese labs of flat out stealing. And the US government is starting to get involved. So how did a Chinese startup build a model this good despite years of US restrictions designed specifically to slow China down? Well, the answer to that question depends on who you ask. Some US officials and tech executives are basically saying that Chinese AI labs are stealing. There are two main allegations here. The first is something called distillation. You're going to hear that word a lot moving forward, so I'll try to do my best to explain it. Distillation is when you use someone else's AI model to train your own. The way it works is that you take a powerful teacher model like Anthropics Fable 5, and you ask it millions and millions of questions through their API. Then you take all those high quality answers and use them to train your own student model. And by doing this, the student model learns to imitate the teacher. So instead of spending years and billions of dollars figuring out how to make a model smart, you essentially just copy the answer straight from the teacher. Now, this is technically against the terms of services of AI companies like OpenAI and Anthropic. But there's nothing to stop these AI lives in China from doing it. So now the US government is getting involved. A top White House official publicly accused Moonshot of distilling Anthropic's Fable 5 model to build Kimmy K3. And Treasury Secretary Scott Bessen also jumped in, saying that sanctions and trade blacklists could be on the table for foreign companies found stealing American AI technology. Anthropic themselves have previously made similar allegations against Moonshot, AI Deepseek, and other Chinese labs. Moonshot, for their part, have not responded to these accusations. And look, distillation is only half of the story here. The other allegation is that Moonshot gained access to restricted Nvidia chips. Remember, the US Government has blocked Chinese companies from buying Nvidia's most powerful advanced Blackwell systems. The thinking there was that by blocking Chinese labs from getting access to the world's best AI chips, they would have a much harder time training cutting edge AI models. But according to the White House, Moonshot accessed Blackwell equipped servers in Thailand, and they may have used those servers to train their latest AI model. But you know, despite these accusations, there's some AI researchers that think that Kimmy's architectural improvements are real innovation and not just copying whatever anthropic did. Even OpenAI's president, Greg Brockman, admitted that Kimi is a strong model and that it was still too early to determine how much of its performance came from distillation. And honestly, I'm not sure if any of this matters. Maybe Moonshot did use distillation. Maybe they did get access to restricted Nvidia hardware in Thailand, and maybe they also developed some legitimate breakthrough on their own. I'm sure the real answer is a mix of all three. But to me, the bigger takeaway from the Kimmy story is that the US restrictions on China have not stopped Chinese companies from producing models that are competitive with America's best models. I mean, even Greg Brockman admitted himself that the gap between Chinese and US models is closing. The consensus used to be that China was six to nine months behind. Now they're probably closer to four months behind. And the key point that I want to make is that China is playing a very different game from the US when it comes to AI. The top AI labs here in the US are keeping their best model closed and they charge premium prices for access to those models. And I'm not just talking OpenAI and Anthropic. I'm also talking about Google and also Meta. Now, Meta tried to do the open weight thing with their Llama models, but now they're pivoting back to having a closed model. Chinese labs, on the other hand, are increasingly releasing capable models as open weight and pricing them aggressively to encourage developers all over the world to build on top of their models. What China wants to do is turn AI into a commodity. So Chinese tech gets a foothold across emerging markets, startups and companies all over the world. So that leaves the US government with a very difficult decision. Do they try to block Chinese models in the us? Do they tighten their chip restrictions and punish companies accused of distillation? Or should US tech companies compete directly with China by having their own cheap open weight models? It's a tough spot for the US government and tech companies to be in. They have to decide whether the right response is to block China, compete with China, or. Or maybe somehow do both. So what's my take here? Well, just like with the deep sea sell off, I think the market is overreacting to the Kimmy K3 thing. The big takeaway for me from this Kimmy thing is that AI models are probably going to become a commodity way faster than anyone expected. And yeah, that's probably bad news for someone invested in OpenAI and anthropic. But for the rest of the AI trade like chip stocks, memory stocks and the broader AI build out, I think this is great news. In fact, the biggest beneficiaries of cheap AI are might be companies that use the cheap intelligence to innovate. And you know, big picture here. I think the US needs to stop acting so shocked every time a Chinese company releases a good AI model. The US needs to start taking China's AI ambition seriously. China is clearly close enough to compete now and they are playing a very different game by releasing cheap open weight models that developers all over the world can build on. The US's answer so far to China has been regulation and chip export controls, but those clearly haven't worked. And simply banning Chinese models in the US could also backfire. You would cut off American companies from cheap AI, while companies all over the world could keep building on it. And by the way, this is crazy. Literally, as I was recording this episode, Nvidia CEO Jensen Huang is weighing in on this debate. He posted a letter signed by Nvidia, Meta, Microsoft and a bunch of other major tech companies arguing that America needs a strong open weight AI ecosystem. The argument that he makes in his letter is that the answer to China releasing cheap open models isn't to shut everything down with regulation. Instead, the US needs its own competitive open models, so developers around the world aren't forced to choose between an expensive closed American model and a cheap, open Chinese model. Now, look, obviously Jensen has some skin in the game here because the more AI models people use, the more Nvidia chips they need. But I think he's right about the broader point about getting behind open weight models, and most of the tech industry agrees. The two notable companies that didn't sign this letter, by the way, were OpenAI and Anthropic. And that shouldn't be a surprise, right? These are probably the two companies with the most to lose if powerful AI models become cheap and open. So I guess what I'm trying to say is that my biggest takeaway from Kimmy isn't that China just broke the AI trade, but what I think China is doing is forcing investors to rethink where the money in AI is actually going to be made. I think up until now everyone thought the winners were going to be anthropic in OpenAI, but now we're facing a reality where AI models might just be a commodity and the market now has to reprice that potential outcome. Well, all right, guys, that's it for today's weekend deep dive. Hope you guys enjoyed that one. Let me know in the comments on what you guys think about Kimi K3 and the rise of Chinese AI models. Do you think that cheaper open weight models are a threat to companies like OpenAI and Anthropic? Or do you think that this just makes the overall AI boom even bigger? Drop your thoughts on Spotify and YouTube and while you're at it, consider giving us a five star rating or thumbs up. You know, all that engagement engagement really does help us out and it helps other people find the show. Thank you guys so much for listening, watching and commenting. Shout out to Mike for all the work behind the scenes and we'll see you guys back here tomorrow.