The Chinese AI Model Rattling Silicon Valley and Washington

0:00AI is entering its most consequential phase where scale, safety and sovereignty will determine who leads and who lags. Join Bloomberg Tech in London on November 2nd and 3rd as global leaders across business, finance and policy examine the defining tradeoffs shaping the future of AI. Thank you to our presenting sponsor Salesforce and supporting sponsors IDA Ireland and Schneider Electric. Learn more at BloombergLive.com slash Tech London.

0:32Bloomberg Audio Studios, podcasts, radio, news. The battle for AI dominance between the U.S. and China has escalated sharply over the past few weeks after the surprise release of a new Chinese AI model took the tech world and Washington by storm. Kimi K3, which is a very powerful model. Kimi K3 model. This company's AI model, known as Kimi K3 model. For the last three days, I've said Kimi K3 many times. Kimi K3, developed by Chinese startup Moonshot AI. Kimi K3 is really making waves because it stacks up so well against offerings from OpenAI and Anthropic, which are among the most powerful AI models in existence.

1:19Mark Anderson is Bloomberg's Asia technology editor, based in Hong Kong. And all of this developed by a Chinese AI company that has access to far less compute and far less resources than deep-pocketed American competitors. Kimi K3 trails only OpenAI's GPT 5.6 and Anthropix Claude Fable 5. So it's the third most powerful model on the market today. Within hours of its release, Kimi K3 sent shockwaves through global markets, triggering memories of the deep-seek panic in early 2025. Remember, chip stocks fell into a bear market Friday. Because investors are now really reassessing whether or not all that CapEx by Samsung and SK Hynix, is that really justified given that China is doing it better and cheaper?

2:14Once markets settled, the questions began. about how Moonshot built Kimi in the first place. It raises the question about if this was done illegally and in a violation of those export controls. Did it use banned NVIDIA chips? Did it steal proprietary data from American rival Anthropic? This is what the director of the White House Office of Science and Technology, Michael Kratzios, he is saying that Moonshot allegedly developed a sophisticated internal platform to conduct large-scale distillation against U.S. models. And most importantly, what does yet another low-cost Chinese AI model mean for the race between the U.S.

2:58and China for global AI supremacy? I think the K3 moment tells us that the race is much closer than people previously thought. I think that China is about six months behind the U.S. That's not a lot of time. Which is not a lot of time, but it's also bringing a challenge to this perception that only American AI, only proprietary AI can be frontier, can be the most cutting edge AI. It's challenging this idea that America owns the most advanced AI models in a really exciting way.

3:34Welcome to the Big Take Asia from Bloomberg News. I'm Wan Ha. Every week, we take you inside some of the world's biggest and most powerful economies, and the markets, tycoons, and businesses that drive this ever-shifting region. Today on the show, what we know about Moonshot's Kimi model, what the company's stunning rise means for Silicon Valley's tech giants, and the high-stakes AI arms race between the U.S. and China.

4:06Kimi K3's launch didn't just rattle markets. Overnight, it became clear that China is intent on rewriting the rules of the global AI race. It's also upending long-held assumptions about how this technology will evolve. The prevailing view in the AI industry used to be that the companies with the biggest budgets, largest data centers, and access to the most advanced chips would lead. But with the release of DeepSeq's R1 model last year and Moonshot AI's latest offering, that narrative is being challenged. I want to get just a little geeky with you. Can you take us under the hood about this model and what's special about it?

4:48Sure. Geeky is good. Basically, it's got a really huge brain. So by brain, we mean parameters. It's got 2.8 trillion parameters. And what are parameters? I would describe parameters as the size of the model's brain. So it's essentially the amount of information that is stored inside an artificial intelligence model. The bigger the brain, the more information that is stored in it, the more information can be processed. You know, it's really as simple as bigger is better. And 2.8 trillion is certainly a very large number, which is essentially on par with the latest offerings from OpenAI and Anthropic.

5:26Kimi K3 isn't just notable for its brain size. It's also remarkable for the way it's being shared. While models like GPT and Claude largely operate behind closed doors, K3 was released to the public last month as an open-weight model, meaning anyone, a researcher, a startup, a government can access it. So basically, an open-weight model can be downloaded and customized. Users can look and see essentially what I would describe as the recipe of the model. It can see how the model works at a level that proprietary AI doesn't offer. Is that the same as open source? Open source goes one step further.

6:06Open source, I would describe it as a kitchen. Open source allows you to see all of the steps behind the recipe, the entire process by which the model was built. There are other open-weight AI models on the market, like those from DeepSeek. But right now, Kimi K3 is the biggest. And size isn't its only selling point. Kimi K3 comes with something called a massive context window. A feature, Mark says, is becoming increasingly important as AI models take on more complex tasks. The context window is essentially the temporary memory of an AI model. So it's the AI model's ability to retrieve information in a timely manner.

6:49I think we've all probably used Gemini or ChatGPT and tried to ask for things like, give me this recipe or what time should I put my toddler to bed? But when it comes to running a really complex task like build an entire app from scratch, it can take much, much longer. Most AI models have limited working memory. As a conversation or document gets longer, they can start to lose track of what came earlier. Imagine trying to talk to someone who can only remember the last 10 sentences you said. Kimmy K3, on the other hand, has a context window big enough to process nearly 800 ,000 words in a single prompt.

7:28That's like several full-length novels at once. This makes it extremely powerful for long legal documents, code bases, or research papers. And K3 isn't just powerful, it's also remarkably cheap to run. According to a benchmarking site that is widely regarded throughout the industry called Artificial Analysis, the cost of running Kimi K3 is far cheaper than running the latest offering from Anthropic or OpenAI. The most expensive model at the moment is Anthropix Claude Fable 5, which costs about$2.75 per task, according to artificial analysis. If you look at Kimi K3, I think it's about$1, something like$0.95.

8:18But it offers a very similar performance, and I think that is a key reason for its popularity. — K3's release has thrust its parent company, Moonshot AI, and its founder, Yang Chilin, into the spotlight. — When OpenAI shocked the world with ChatGPT, everyone was looking to Yang and Moonshot as China's best chance at matching that technology. And really, he was seen as the poster boy of Chinese AI up until DeepSeek shocked the world last January. And basically, ever since then, he's toiled in their shadow. But this release of Kimi K3 has brought him back to prominence. It's brought him back to the forefront of the conversation about Chinese AI.

9:03Now, the last time we saw this level of attention and certainly headlines about a Chinese AI, it was DeepSeek in 2025. Is this another DeepSeek moment, you know, DeepSeek 2.0 for China? For me, it's not. For me, the DeepSeek moment was really establishing that China was much, much closer than five years behind the U.S. in the global technology race. You know, before DeepSeek's R1 model was released last January, people really did believe that the U.S. had a years-long head start in AI and DeepSeek completely upended that perception, cut it down to just a few months. I think we are now in a place with Moonshot where people are just expecting more of these breakthroughs, and it's not as surprising as the DeepSeek moment was.

9:51But it is certainly exciting because it does show that Moonshot and other Chinese AI companies are able to use this distillation method to really improve their AI models and also offer them for much cheaper costs, which will spread AI throughout the world.

10:09After the break, the speculation over what may be powering Kimi K3's success and how Chinese AI advances are reshaping the strategies of global tech giants like Meta.

10:34China's AI ambitions were on full display this summer at a gathering at the World Expo Center in Shanghai. Thousands of executives, engineers, and investors packed the halls of the annual World Artificial Intelligence Conference in July. They had to jostle for space with humanoid robots and AI demos. And this year, the event drew the attention of China's top leader, President Xi Jinping. AI should be a trusted tool for humanity. In the meantime, we should jointly oppose overstretching the national security concept in the field of AI or placing one country's security over that of others. The very fact that she attended the World Artificial Intelligence Conference in Shanghai was very significant.

11:29Here's Mark Anderson, Bloomberg's Asia tech editor, again. So it shows how far up the national agenda AI has shot. And I think she wants to upend this global order and this global perception of AI as an American technology that is only available for people who can afford it. Wary of China's ambitions, Washington has steadily restricted the export of advanced AI chips to Chinese companies. Yet despite those constraints, Moonshot and DeepSeek have managed to build models that rival some of the world's best. How has Moonshot managed to achieve this level of performance, even though they've been cut off from the most advanced NVIDIA chips?

12:13How are they pulling this off? Well, Moonshot offered very little information about how it achieved the Kimi K3 breakthrough. But one thing that's worth noting is Anthropic, the maker of one of the most advanced AI models in the U.S., did say that it had identified these industrial-scale attempts by DeepSeek, Moonshot, and another Chinese AI company called Minimax. They basically said that these three companies are using a technique called distillation, which involves training a less capable model on the outputs of a stronger one. So essentially, they would set up thousands, if not tens of thousands, of accounts on Anthropics platforms and use those accounts to train their model.

12:57So really piggybacking off of the advances and the innovation that U.S. companies have paid for. Exactly. Training their models to be like a more advanced model that was developed using NVIDIA chips. The Trump administration says it's looking into whether Chinese AI companies, including Moonshot, gained access to advanced American-made chips in violation of U.S. export controls. Moonshot hasn't responded to those allegations. But China's foreign ministry said the country's advances in AI are the result of greater self-reliance and homegrown innovation.

13:40Mark says that even if Chinese companies are using techniques like distillation to develop new models, it doesn't mean they're leading the AI race. Replicating an existing model at scale is one thing. Building a frontier model from scratch is another. The common thinking is still that China's about six months behind the U.S. But, you know, with this technique of distillation, it's much harder to be at the forefront of the race without having anyone that you can piggyback off of. You know, if you have to be developing frontier models on your own, you need access to advanced chips. You need to be able to run your own training models, which it doesn't appear Kimi K3 has done.

14:21I mean, is that China's whole strategy? Is that, does it actually want to be the one who is making innovations, you know, from day one? Or would it rather much just follow and piggyback and do it better and cheaper? I think China is basically trying to buy time. It's trying to buy time to build this ecosystem where it can build its own chips. It can train its own models. Their strategy is to just pursue market share rather than to try and develop the most advanced AI using their very limited compute. I think the strategy here is let's get as many people in the world as possible to use our products.

14:56And this tension over open versus closed models is now front and center in Washington. Policymakers are debating whether tighter controls still make sense at a time when advanced AI is already widely accessible. It's a debate NVIDIA CEO Jensen Huang has been weighing in on. Here he is after meeting with U.S. lawmakers late last month. We need open weight for the American industry. We need open weights for security. We need open weights for safety. And it's really important for the vibrancy of the whole industry. And this week, Meta CEO Mark Zuckerberg published a lengthy manifesto supporting distillation.

15:37In it, he also encouraged the U.S. government not to restrict open source or open weight models. Zuckerberg and Meta also launched Muse Glimmer, an open-weight AI model built to run locally on laptops. I am proud to share that we are open-sourcing a new class of on-device models that we are calling Muse Glimmer. It's a 30 billion parameter dense model that runs on your laptop.

16:05Investors appear to be betting on that vision. Just last month, Chinese memory chip maker CXMT surged nearly 500 percent in its Shanghai trading debut. And after Alibaba unveiled its largest ever AI model, Quinn 3.8 Max, the company's shares jumped 7 percent, the most in nearly a month. That's a sign that investors believe that Beijing's push to localize its AI ecosystem is gaining real traction. It's trying to flood the world with brand recognition for Kimi K3, Kuen, and other advanced Chinese models so that when they do have this ecosystem intact in, who's to say, five to ten years, there will be no reliance on foreign companies like NVIDIA.

16:49They'll be able to do everything as they want. They'll be at the forefront of developing the most advanced models and they won't be piggybacking off the U.S. And Mark, looking ahead, what metrics should we be watching for to see whether this is actually a seismic shift in the AI race? I mean, I think a lot of people will point to the capabilities of these models and parameters. For me, it really is about uptake. It really is about market share. If you can convince American businesses you don't need to spend the money, all you need is Kimmy K3. It can do the coding. It can do prepare the presentations, whatever it is that, you know, the market is demanding.

17:28I think that's an incredibly important sign.

17:59Thank you.