Transcription
Tech stocks tumbled again this week over concerns about eye spending and semiconductor supply. But it was the surprise announcement from a Chinese startup drawing parallels to last year's deep seek moment. That's right. Joining us more is Mike Sheppard, Bloomberg's senior editor for technology and strategic industries. Excuse me. And Bloomberg's China correspondent Menlo. Thank you both so much for joining us.
I want to start with you. What is this announcement? What was the reaction to this moonshot AI? And you were over there with all these tech people. What what was the energy like?
Uh, the energy was really insane. It was so crowded today. And mind you, it was a very, very hot day. But it's draw drawn a huge amount of crowd at this world AI conference in Shanghai. I went to Kimmy's booth. I mean, moonshot AI's booth. That's the Chinese startup that was behind the new Kimmy K3 model. They just launched that model. It is the largest ever in China, 2.8 trillion parameters. And the company says that it outperforms all rivals Except for Anthropic Elitist Fable five, as well as OpenAI ChatGPT five for 5.6. So it's really, I think, peeking taken the world by surprise because of how fast China's AI model is closing the gap with the U.S., despite not being able to access some of these advanced chips from Nvidia. And that's why we saw a huge market reaction, the sell off that we saw across Asia, the cost is down more than 6% because investors are now really reassessing whether or not all that CapEx by Samsung and SK Hynix, 600 other billion dollars into new factories, is that really justified given that China is doing it better and cheaper?
Mike Shepherd I learned a new word this week. It's distillation. I heard it a lot at that Aspen security for it, but I guess it's kind of code for stealing or using models that are in existence, kind of replicating them, building on them. Uh, and there are a lot of accusations, I guess, that that's what's happening with a lot of these Chinese AI companies. Talk a bit about that and kind of give us your sense, put this into context, if you would, to where this new model stands, in contrast with the ones that we talk about so much here that are made in the US.
We have brought up an issue data that is increasingly used for support between Silicon Valley and China and the leading AI companies here. OpenAI and anthropic have increasingly beaten the drum. For policymakers to hear that, they think that some of the results that their models are generating are being extracted en masse in an unfair and improper fashion, a way that violates their terms of usage by Chinese startups, and then being used to develop a rival generation of chat bots. They haven't pointed the finger specifically at Kimi K3, but they have recently, and we've reported just a few weeks ago that, uh, anthropic itself was pointing the finger at Alibaba, accusing it of a large scale campaign of distillation. Now, to be clear, this is a technique that is commonly accepted when it comes to AI, when you have a, uh, larger model used to build a smaller model, a larger frontier model whose data are extracted and results are extracted to produce something that's a little smaller, more focused, and less advanced. But the claim from Silicon Valley is that, look, what's happening is that all of the physical investments that we're making and spending all this, all these billions of dollars on, uh, are being instead used, uh, by Chinese competitors to build this rival generation. Now, earlier today, we did hear from the Chinese deputy foreign minister rejecting those accusations. It was really the first and most prominent public statement from China's government with respect to that. But we can expect to see this come up between Presidents Donald Trump and XI Jinping, uh, if their planned meeting in September goes ahead.
Yeah. The timing of this, Mike, is really interesting. I mean, I'm assuming this is all related. This is why we saw this huge dip in chip stocks when this announcement came out. Is that just about this model, or does that kind of compound on some of the concerns about this CapEx investment, especially? It seems to me when you look at what China can bring to the table, if they have the technology side here in the U.S., we focus so much on the energy aspect, on building these data centers, on whether or not they can support them. It seems a lot easier to do coming out of a top down system like Beijing. Is that part of this concern as well?
It certainly is. And what you're seeing really is a return of that crisis of confidence and whether that AI trade will really sustain itself. And can all of the investments that these companies are making in what could be trillions of dollars in infrastructure for data centers and chips and other materials, will the will they keep going and can they justify it? And it's harder to justify if they're cheaper and just as good models out there being offered and also models that can be downloaded. What are the differences that you see with OpenAI and anthropic is, you know, their models are generally access to a website through an API. Um, you know, at an enterprise level with the Chinese approach. You've seen these startups use what's called open weight. That means you can just download it, modify it, adjust the parameters, and really tailor it to your needs. It's a lot harder to do that with some of the frontier models being developed by the US AI labs right now, and that is again something that people are also questioning the interoperability and functionality of, of what the US labs are producing. And then there's the cost amendment pointed to the fact that many of these, uh, Chinese companies are now offering their services at a much lower rate. And the companies here in the U.S. that are trying to buy these services are starting to look at the bottom line and thinking, hey, all this AI that we're telling our employees to use, it's starting to cost us. So that is, again, another factor in this. And it is, uh, causing this big downdraft across the markets here.
We compare these models. We compare the kind of technological advancements. It's also interesting to look at the contrast between the U.S. approach and China's approach. And that's something I talked about with Elizabeth Economy, who's a fellow at the Hoover Institution at Stanford University. A couple of days ago, we talked about the Chinese strategy compared to the U.S. strategy. Let's take a listen.
XI Jinping has a grand vision for how they're going to deploy AI in every sector of economy, the educational system and the logistics system, certainly in the military. I don't feel as though we have that vision of of taking all of our incredible, innovative work, you know, the most advanced models in the world. But how are we going to make them practical, practicable for our economy? Um, how are they going to be used? And again, in that transition, how do we make it not destabilizing?
Let me turn to you on this first, if I could. I'm curious sort of if you could expand on that a little bit. You heard from President XI yesterday at this conference. Talk a bit about kind of the way that the Chinese government is embracing advancements in AI and thinking about it, a kind of a larger strategic level.
Yeah, it's definitely an important priority for the Chinese government. In fact, they are grooming this entire AI sector as a next big export engine for China as the country transitions away from all the industries like property, which is really in a big slump for several years now. So they are pouring a lot of investments to, uh, provide the patient capital to groom all the startups in the sector. And it's all part of the five year plan that China had laid out earlier this year, right in March, at the two sessions, they had a I plus plan where they want to encourage adoption of AI across all sectors of the economy over the next decade or so. And just to give you a sense of how important I is, this world I Conference in Shanghai is in the ninth edition. And I remember in the early years not many journalists would even bother attending, but now it's a completely different story. Even presidency himself came to town to grade the event personally, and this really represents an upgrade to the status of this event, which in previous, uh, in the last two years, it was attended by Premier Li uh Li Chang, who is the number two leader in China and president. He gave this big speech this year. A lot of it was, of course, also a veiled message to the U.S. as well, because he talked about how when it comes to AI governance, the world needs to take a multilateral approach and the world needs to learn from one another. And AI development shouldn't be a solo development of one country and all of the Chinese state media. You look at the Communist Party mouthpiece People's Daily. They have been publishing editorials about how tech and I should not become a monopoly. It should really be treated as a global public good that is shared with the rest of the world. So again, presidency positioning himself as this leader of the Global South, uh, advocating for bridging the digital divide. So he announced a few key initiatives, including offering at least 5000 training opportunities in AI for people in the developed countries and allowing them some 30 countries to tap into this AI powered meteorological system to safeguard their homes, for example. So really. And also, yes, uh, a couple of days ago, there were some 29 countries who signed up to be part of this China led group called the world AI Cooperation Organization. And there's just yet another platform that could give China more sway in shaping global AI norms.
Mike, how does this impact American companies? I mean, obviously they're already down the road with a lot of these investments. Is there. They can't claw them back at this point, but they have to be looking at this and being very concerned. Is there a way for them to pivot? And then you also have the Trump administration starting to talk about having a new regulatory agency specifically for AI. Does that compound the difficulties American AI companies could now face?
Well, I'll take the second part first there, Christina, because it's the latest and there was a scoop from last night from our team here in DC. And the companies actually are would welcome some sort of a coherent approach to AI safety reviews. And at issue is, you know, they did not like the way that anthropic, uh, fable five and mythos five were essentially ordered off the market via export controls over, uh, concerns that the guardrails on them weren't adequate. And we saw the same sort of pressure put on open AI. But the companies had felt that it was, uh, overboard and also kind of ad hoc. And they approach and they would like to see a more coherent framework. And what's taking shape is, uh, you know, a Finra like model, one that would resemble what the broker dealer industry on Wall Street faces, that it's an industry funded and, and, uh, and not industry run, but industry certainly has a lot of input to how the rules of the road are set. And in this case, it would be the, uh, safety testing. And that would give them a little bit more of, uh, I guess, a clear path forward when it comes to dealing with the government and how far they can take their models, uh, as they develop them. Um, more broadly, the the question of how companies are reading this moment. Look, many of them are playing the long game. And all this week we heard from TSMC, they were reporting, look, we are seeing demand for AI chips in AR hardware going through 2030. We are it is going to be hard to meet. And that is a message that was echoed by SK Hynix during Ed Ludlow's interview with the CEO. Just, uh, last, uh, just for the little more than a week ago, as they were making their public market debut here in the U.S. and it's a message that's been reinforced by Nvidia Chief Executive Officer Jensen Wong. They just see the demand from hyperscalers here in the U.S. and even elsewhere as being off the charts and in video. Also quite a bit of a break in some of the other high, uh, uh, American companies caught a break, too, with the declassification of, uh, the United Arab Emirates, which is trying to establish itself as a real center of development for AI. And they see that as a new market that they can break into and try to export the American tech stack. But really, do you see this contrast the vision? So XI Jinping laying out something that is very, uh, multilateral. And President Donald Trump arguing, one that is very much America first, and it really doesn't have all the pieces together in the same way that you do see China putting on the table that.