# The Chinese AI Model Rattling Silicon Valley and Washington

An episode of [Big Take](https://invisiblenetwork.ai/podcast/1578096201), 12 Aug 2026, 18 min.

Audio: https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/825d4e29-b616-46f4-afd7-ae2b0013005c/b394d917-ac7d-46f8-8513-b4a4014f7a44/audio.mp3?utm_source=Podcast&in_playlist=8b1dd624-a026-43e9-8b57-ae2b00130066

The launch of Moonshot AI’s Kimi K3 model has taken the tech world — and Washington — by storm, raising urgent questions about how a Chinese startup can suddenly rival the world's most powerful AI systems. On today’s Big Take Asia podcast, host Oanh Ha speaks with Bloomberg’s Mark Anderson about the distillation method behind Moonshot’s breakthrough and what it means for the high-stakes AI race between the US and China. Read more: Moonshot Plans IPO in Six Months After China AI Breakthrough See omnystudio.com/listener for privacy information.

## Transcript

[0:02] Speaker 1: Bloomberg Audio Studios, podcasts, radio news.
[0:09] Speaker 2: The battle for AI dominance between the US and China
[0:12] Speaker 2: has escalated sharply over the past few weeks after the
[0:16] Speaker 2: surprise release of a new Chinese AI model took the
[0:19] Speaker 2: tech world and Washington by storm.
[0:22] Speaker 3: Kimmy K three, which is a very powerful model.
[0:25] Speaker 4: Kimmy K three months company's AI model known as Kimmy
[0:28] Speaker 4: K three.
[0:29] Speaker 1: For the last three days, I've said Kimmy K three
[0:32] Speaker 1: many times.
[0:33] Speaker 2: Kimmy K three, developed by Chinese startup Moonshot Ai.
[0:37] Speaker 5: Kimmy K three is really making waves because it stacks
[0:41] Speaker 5: up so well against offerings from Open Ai and Anthropic,
[0:45] Speaker 5: which are among the most powerful AI models in existence.
[0:49] Speaker 2: Mark Anderson is Bloomberg's Asia technology editor based in Hong Kong.
[0:54] Speaker 5: And all of this developed by a Chinese AI company
[0:57] Speaker 5: that has access to far less compute and far less
[1:01] Speaker 5: resources than depocketed American competitors. Kimmi K three trails only
[1:08] Speaker 5: Open Eyes GPT five point six and Anthropics Claude Fable five,
[1:13] Speaker 5: so it's the third most powerful model on the market today.
[1:21] Speaker 2: Within hours of its release, Kimmi K three sent shock
[1:24] Speaker 2: waves through global markets, triggering memories of the deep seak
[1:28] Speaker 2: panic in early twenty twenty five.
[1:30] Speaker 6: Remember, chip stocks fell into a bear market Friday.
[1:34] Speaker 3: Because investors are now really reassessing whether or not all
[1:37] Speaker 3: that capex by some song and sk heinis is that
[1:40] Speaker 3: really justified given that China is doing it better and cheaper.
[1:44] Speaker 2: Once markets settled, the questions began about how Moonshot built
[1:49] Speaker 2: Kimmi in the first place.
[1:51] Speaker 4: It raises the question about if this was done illegally
[1:55] Speaker 4: and in a violation of those export controls.
[1:59] Speaker 2: Did it use banned in video chips, did it steal
[2:02] Speaker 2: proprietary data from American rival andthropic?
[2:06] Speaker 4: This is what the director of the White House Office
[2:08] Speaker 4: of Science and Technology, Michael Kratzios, He is saying that
[2:11] Speaker 4: Moonshot allegedly developed a sophisticated internal platform to conduct large
[2:17] Speaker 4: scale distillation against US models.
[2:21] Speaker 2: And most importantly, what is yet another low cost Chinese
[2:25] Speaker 2: AI model mean for the race between the US and
[2:28] Speaker 2: China for global AI supremacy.
[2:30] Speaker 5: I think the K three moment tells us that the
[2:34] Speaker 5: race is much closer than people previously thought. I think
[2:38] Speaker 5: that China is about six months behind the US. That's
[2:41] Speaker 5: not a lot of time, which is not a lot
[2:43] Speaker 5: of time. But it's also bringing a challenge to this
[2:46] Speaker 5: perception that only American air, only proprietary AI can be frontier,
[2:51] Speaker 5: can be the most cutting edge AI. It's challenging this
[2:54] Speaker 5: idea that America owns the most advanced AI models in
[2:58] Speaker 5: a really exciting way.
[3:05] Speaker 2: Welcome to The Big Take Asia from Bloomberg News. I'm wanh.
[3:09] Speaker 2: Every week we take you inside some of the world's
[3:11] Speaker 2: biggest and most powerful economies and the markets, tycoons and
[3:15] Speaker 2: businesses that drive this ever shifting region. Today on the show,
[3:20] Speaker 2: what we know about Moonshot's Kimi model, What the company's
[3:24] Speaker 2: stunning rise means for Silicon Valley's tech giants, and the
[3:28] Speaker 2: high stakes AI arms race between the US and China.
[3:36] Speaker 2: Kimmy K three's launch didn't just rattle markets overnight. It
[3:41] Speaker 2: became clear that China is intent on rewriting the rules
[3:45] Speaker 2: of the global AI race. It's also upending long held
[3:48] Speaker 2: assumptions about how this technology will evolve. The prevailing view
[3:53] Speaker 2: in the AI industry used to be that the companies
[3:55] Speaker 2: with the biggest budgets, largest data centers, and access to
[3:59] Speaker 2: the most advanced chips would lead. But what the release
[4:03] Speaker 2: of Deep seeks are one model last year and Moonshot
[4:06] Speaker 2: AI's latest offering, that narrative is being challenged. I want
[4:11] Speaker 2: to get just a little geeky with you. Can you
[4:14] Speaker 2: take us under the hood about this model and what's
[4:17] Speaker 2: special about it?
[4:18] Speaker 5: Sure, geeky is good. Basically, it's got a really huge brain.
[4:23] Speaker 5: So by brain we mean parameters. It's got two point
[4:26] Speaker 5: eight trillion parameters.
[4:28] Speaker 2: And what are parameters?
[4:30] Speaker 5: I would describe parameters as the size of the model's brain.
[4:33] Speaker 5: So it's essentially the amount of information that is stored
[4:36] Speaker 5: inside an artificial intelligence model. The bigger the brain, the
[4:40] Speaker 5: more information that is stored in it, the more information
[4:43] Speaker 5: can be processed. You know, It's really as simple as
[4:46] Speaker 5: bigger is better, And two point eight trillion is certainly
[4:49] Speaker 5: a very large number, which is essentially on par with
[4:53] Speaker 5: the latest offerings from open ai and Anthropic.
[4:57] Speaker 2: Kimmik three isn't just notable for its brain side, it's
[5:00] Speaker 2: also remarkable for the way it's being shared. While models
[5:04] Speaker 2: like GPT and cloud largely operate behind closed doors, k
[5:08] Speaker 2: three was released to the public last month as an
[5:11] Speaker 2: open weight model, meaning anyone, a researcher, a startup, a
[5:15] Speaker 2: government can access it.
[5:17] Speaker 5: So basically, an open weight model can be downloaded and customized.
[5:21] Speaker 5: Users can look and see essentially what I would describe
[5:24] Speaker 5: as the recipe of the model. It can see how
[5:27] Speaker 5: the model works at a level that proprietary AI doesn't offer.
[5:32] Speaker 2: Is that the same as open source.
[5:34] Speaker 5: Open source goes one step further open source I would
[5:36] Speaker 5: describe it as a kitchen. Open source allows you to
[5:39] Speaker 5: see all of the steps behind the recipe, that the
[5:43] Speaker 5: entire process by which the model was built.
[5:46] Speaker 2: There are other open weight AI models on the market,
[5:49] Speaker 2: like those from deep Seek, but right now kimmyk three
[5:53] Speaker 2: is the biggest, and size isn't its only selling point.
[5:57] Speaker 2: Kimmik three comes with something called a massive context window,
[6:02] Speaker 2: a feature Mark says is becoming increasingly important as AI
[6:05] Speaker 2: models take on more complex tasks.
[6:10] Speaker 5: The context window is essentially the temporary memory of an
[6:13] Speaker 5: AI model, so it's the AI model's ability to retrieve
[6:16] Speaker 5: information in a timely manner. I think we've all probably
[6:20] Speaker 5: used Gemini or chat GPT and tried to ask for
[6:23] Speaker 5: things like give me this recipe or what time should
[6:26] Speaker 5: I put my toddler to bed? But when it comes
[6:29] Speaker 5: to running a really complex task like build an entire
[6:32] Speaker 5: app from scratch, it can take much much longer.
[6:36] Speaker 2: Most AI models have limited working memory. As a conversation
[6:41] Speaker 2: or document gets longer, they can start to lose track
[6:44] Speaker 2: of what came earlier. Imagine trying to talk to someone
[6:47] Speaker 2: who can only remember the last ten sentences you said.
[6:50] Speaker 2: Kimmy K three, on the other hand, has a context
[6:53] Speaker 2: window big enough to process nearly eight hundred thousand words
[6:57] Speaker 2: in a single prompt. That's like several full length novels
[7:00] Speaker 2: at once. This makes it extremely powerful for long legal documents,
[7:05] Speaker 2: code bases, or research papers. And K three isn't just powerful,
[7:10] Speaker 2: it's also remarkably cheap to run.
[7:13] Speaker 5: According to a benchmarking site that is widely regarded throughout
[7:17] Speaker 5: the industry called Artificial Analysis, the cost of running Kimmi
[7:22] Speaker 5: K three is far cheaper than running the latest offering
[7:27] Speaker 5: from Anthropic or OpenAI. The most expensive model at the
[7:31] Speaker 5: moment is Nthropics Claude Fable five, which costs about two
[7:36] Speaker 5: dollars and seventy five cents per task. According to Artificial Analysis,
[7:41] Speaker 5: if you look at Kimmy K three, I think it's
[7:45] Speaker 5: about a dollar something like ninety five cents, but it
[7:49] Speaker 5: offers a very similar performance, and I think that is
[7:53] Speaker 5: a key reason for its popularity.
[7:56] Speaker 2: K three's release has thrust its parent company, Moonshot Ai
[8:00] Speaker 2: and its founder Young Chilling into the spotlight.
[8:04] Speaker 5: When Open Ai shocked the world with chat Gipt, everyone
[8:07] Speaker 5: was looking to Yang and Moonshot as China's best chance
[8:12] Speaker 5: at matching that technology. And really he was seen as
[8:16] Speaker 5: the poster boy of Chinese AI up until Deep Seek
[8:19] Speaker 5: shocked the world last January, and basically ever since then
[8:23] Speaker 5: he's told in their shadow. But this release of Kimmy
[8:26] Speaker 5: K three has brought him back to prominence. It's brought
[8:29] Speaker 5: him back to the forefront of the conversation about Chinese AI.
[8:34] Speaker 2: Now, the last time we saw this level of attention
[8:38] Speaker 2: and certainly headlines about a Chinese AI it was Deep
[8:41] Speaker 2: Seek in twenty twenty five. Is this another Deep Seek moment?
[8:44] Speaker 2: You know, Deep Seek two point zero for China?
[8:47] Speaker 5: For me, it's not for me. The Deep Seek moment
[8:49] Speaker 5: was really establishing that China was much much closer than
[8:54] Speaker 5: five years behind the US in the global technology race.
[8:57] Speaker 5: You know, before Deep Seeks Our one model was released
[9:00] Speaker 5: last January, people really did believe that the US had
[9:03] Speaker 5: a year's long head start in AI and Deep Sea
[9:06] Speaker 5: completely upended that perception, cut it down to just a
[9:09] Speaker 5: few months. I think we are now in a place
[9:13] Speaker 5: with Moonshot where people are just expecting more of these breakthroughs,
[9:17] Speaker 5: and it's not as surprising as the Deep Seek moment was,
[9:21] Speaker 5: but it is certainly exciting because it does show that
[9:23] Speaker 5: Moonshot and other Chinese AI companies are able to use
[9:28] Speaker 5: this distillation method to really improve their AI models and
[9:32] Speaker 5: also offer them for much cheaper costs, which will spread
[9:35] Speaker 5: AI throughout the world.
[9:40] Speaker 2: After the break the speculation over what may be powering
[9:43] Speaker 2: Kimmy K three's success and how Chinese AI advances are
[9:47] Speaker 2: reshaping the strategies of global tech giants like Meta. China's
[10:05] Speaker 2: AI ambitions were on full display this summer at a
[10:08] Speaker 2: gathering at the World Expo Center in Shanghai. Thousands of executives, engineers,
[10:15] Speaker 2: and investors packed the halls of the annual World Artificial
[10:19] Speaker 2: Intelligence Conference in July. They had to jostle for space
[10:22] Speaker 2: with humanoid robots and AI demos, and this year the
[10:26] Speaker 2: event drew the attention of China's top leader, President Xi jinping.
[10:31] Speaker 6: AI should be a trusted tool for humanity.
[10:36] Speaker 2: In Vang Kung Fu fans.
[10:38] Speaker 6: In the meantime, we should jointly oppose overstretching the national
[10:43] Speaker 6: security concept in the feud of AI, or placing one
[10:48] Speaker 6: country's security over that of others.
[10:52] Speaker 5: The very fact that she attended the World Artificial Intelligence
[10:57] Speaker 5: Conference in Shanghai was very significant.
[10:59] Speaker 2: Here's Mark Anderson, Bloomberg's Asia Tech editor again.
[11:02] Speaker 5: So it shows how far up the national agenda AI
[11:05] Speaker 5: has shot. And I think she wants to upend this
[11:09] Speaker 5: global order in this global perception of AI as an
[11:12] Speaker 5: American technology that is only available for people who can
[11:17] Speaker 5: afford it.
[11:18] Speaker 2: Wary of China's ambitions, Washington has steadily restricted the export
[11:23] Speaker 2: of advanced AI chips to Chinese companies. Yet despite those constraints,
[11:28] Speaker 2: Moonshot and Deepseek have managed to build models that rival
[11:32] Speaker 2: some of the world's best How has Moonshot managed to
[11:37] Speaker 2: achieve this level of performance even though they've been cut
[11:40] Speaker 2: off from the most advanced nvidioships. How are they pulling
[11:43] Speaker 2: this off well?
[11:45] Speaker 5: Moonshot offered very little information about how it achieved the
[11:49] Speaker 5: Kimmy K three breakthrough, but one thing that's worth noting
[11:52] Speaker 5: is Anthropic, the maker of one of the most advanced
[11:56] Speaker 5: AI models in the US did say that it identified
[12:01] Speaker 5: these industrial scale attempts by deep Seek, Moonshot, at another
[12:06] Speaker 5: Chinese AI company called Minimax. They basically said that these
[12:10] Speaker 5: three companies are using a technique called distillation, which involves
[12:14] Speaker 5: training a less capable model on the outputs of a
[12:17] Speaker 5: stronger one. So essentially they would set up thousands, if
[12:20] Speaker 5: not tens of thousands of accounts on anthropics platforms and
[12:26] Speaker 5: use those accounts to train their.
[12:27] Speaker 2: Model, so really piggybacking off of the advances and the
[12:31] Speaker 2: innovation that US companies have paid.
[12:33] Speaker 5: For, exactly training their models to be like a more
[12:36] Speaker 5: advanced model that was developed using Invidia chips.
[12:42] Speaker 2: The Trump administration says it's looking into whether Chinese AI companies,
[12:47] Speaker 2: including Moonshot, gained access to advanced American made chips in
[12:52] Speaker 2: violation of US export controls. Moonshot hasn't responded to those allegations,
[12:58] Speaker 2: but China's Foreign ministries said the country advances in AI
[13:01] Speaker 2: are the result of greater self reliance and homegrown innovation.
[13:06] Speaker 1: Jung j.
[13:10] Speaker 2: Mark says that even if Chinese companies are using techniques
[13:14] Speaker 2: like distillation to develop new models, it doesn't mean they're
[13:17] Speaker 2: leading the AI race. Replicating an existing model at scale
[13:22] Speaker 2: is one thing, building a frontier model from scratch is another.
[13:27] Speaker 5: The common thinking is still that China is about six
[13:29] Speaker 5: months behind the US. But you know, with this technique
[13:33] Speaker 5: of distillation, it's much harder to be at the forefront
[13:36] Speaker 5: of the race without having anyone that you can piggyback
[13:39] Speaker 5: off of. You know, if you have to be developing
[13:41] Speaker 5: frontier models on your own, you need access to advanced chips,
[13:45] Speaker 5: You need to be able to run your own training models,
[13:48] Speaker 5: which it doesn't appear Kimmy K three has done.
[13:51] Speaker 2: I mean is is that China's whole strategy is that
[13:54] Speaker 2: does it actually want to be the one who is
[13:56] Speaker 2: making innovations, you know, from day one, or would it
[14:00] Speaker 2: rather much just follow and piggyback and do it better
[14:03] Speaker 2: and cheaper.
[14:04] Speaker 5: I think China is basically trying to buy time. It's
[14:06] Speaker 5: trying to buy time to build this ecosystem where it
[14:08] Speaker 5: can build its own chips, it can train its own models.
[14:11] Speaker 5: Their strategy is to just pursue market share rather than
[14:14] Speaker 5: to try and develop the most advanced AI using their
[14:18] Speaker 5: very limited compute. I think the strategy here is, let's
[14:21] Speaker 5: get as many people in the world as possible to
[14:24] Speaker 5: use our.
[14:25] Speaker 2: Products, and this tension over open versus closed models is
[14:30] Speaker 2: now front and center in Washington. Policy makers are debating
[14:35] Speaker 2: whether tighter controls still make sense at a time when
[14:38] Speaker 2: advanced AI is already widely accessible. It's a debate in
[14:42] Speaker 2: Vidia CEO Jensen Huang has been weighing in on here.
[14:46] Speaker 2: He is after meeting with US lawmakers late last month.
[14:49] Speaker 2: We need open weight for the American industry.
[14:52] Speaker 6: We need open weights for security, we need open ways
[14:54] Speaker 6: for safety.
[14:55] Speaker 1: And it's really important for the vibrancy of the whole industry.
[15:01] Speaker 2: And this week, Meta CEO Mark Zuckerberg published a lengthy
[15:04] Speaker 2: manifesto supporting distillation. In it, he also encouraged the US
[15:09] Speaker 2: government not to restrict open source or open weight models.
[15:14] Speaker 2: Zuckerberg and Meta also launched muse Glimmer, an open weight
[15:18] Speaker 2: AI model built to run locally on laptops.
[15:21] Speaker 1: I am proud to share that we are open sourcing
[15:24] Speaker 1: a new class of on device models that we are
[15:27] Speaker 1: calling muse Glimmer. It's a thirty billion parameter dense model
[15:30] Speaker 1: that runs on your laptop.
[15:32] Speaker 2: It's investors appear to be betting on that vision. Just
[15:38] Speaker 2: last month, Chinese memory chip maker CXMT surged nearly five
[15:42] Speaker 2: hundred percent in its Shanghai trading debut, and after Ali
[15:46] Speaker 2: Baba unveiled its largest ever AI model, Quinn three point
[15:49] Speaker 2: eight Max, the company's shares jump seven percent, the most
[15:54] Speaker 2: in nearly a month. That's a sign that investors believe
[15:57] Speaker 2: that Beijing's pushed to localize its AI ecosystem is gaining
[16:01] Speaker 2: real traction.
[16:03] Speaker 5: It's trying to flood the world with brand recognition for
[16:07] Speaker 5: Kimmy k three, Quen and other advanced Chinese models, so
[16:12] Speaker 5: that when they do have this ecosystem intact in who's
[16:15] Speaker 5: to say, five to ten years, there will be no
[16:17] Speaker 5: reliance on foreign companies like Nvideo. They'll be able to
[16:20] Speaker 5: do everything as they want. They'll be at the forefront
[16:23] Speaker 5: of developing the most advanced models, and they won't be
[16:27] Speaker 5: piggybacking off the US.
[16:29] Speaker 2: And mark Looking ahead, what metrics should we be watching
[16:32] Speaker 2: for to see whether this is actually a seismic shift
[16:36] Speaker 2: in the AI race.
[16:38] Speaker 5: I mean, I think a lot of people will point
[16:40] Speaker 5: to the capabilities of these models and parameters. For me,
[16:44] Speaker 5: it really is about uptake, It really is about market share.
[16:46] Speaker 5: If you can convince American businesses. You don't need to
[16:49] Speaker 5: spend the money. All you need is Kimmik three. It
[16:52] Speaker 5: can do the coding, it can do prepare the presentations,
[16:55] Speaker 5: whatever it is that the market is demanding. I think
[16:58] Speaker 5: that's an incredibly Hordan sign.
[17:06] Speaker 2: This is The Big Take Asia from Bloomberg News. I'm
[17:09] Speaker 2: wan ha. To get more from The Big Take and
[17:12] Speaker 2: unlimited access to all of Bloomberg dot Com, subscribe today
[17:15] Speaker 2: at Bloomberg dot com slash podcast Offer. If you liked
[17:18] Speaker 2: the episode, make sure to subscribe and review The Big
[17:21] Speaker 2: Take Asia wherever you listen to podcasts. It really helps
[17:24] Speaker 2: people find the show. Thanks for listening, See you next time.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.