How China Wins The AI War

0:00Support for today's show comes from Attio. Attio is the AI CRM that gets you from sign up to insights in seconds. You can connect your emails, calls, product data, and more. And Attio can instantly build your pipeline with enriched data and complete context. Whether you're running product-led growth or enterprise sales, Attio adapts to your business. Then, ask Attio to plan your next move. Spot deals at risk before they go sideways. Know what to say before every call and never miss a follow-up. Atio will search, update, and create across your data instantly so your customer context becomes more powerful than ever.

0:32Ask more from your CRM. Ask Atio. You can go to atio.com slash propg and you'll get 15 % off your first year. That's A-T-T-I-O dot com slash propg.

1:09See you next week. at adobe.com slash do that with acrobat. Support for the show comes from Amazon. There are the things you can plan for. A first birthday party, a movie marathon, a renter-friendly bathroom reno. And then there are the things you can never plan for. A surprise rainstorm, a Blu-ray player calling it quits, stick on tiles that looked way better on the package. For all things planned and unplanned, Amazon has you covered. You'll find low prices on everyday essentials and last-minute lifesavers. Shop Amazon and save on essentials. Save the everyday. Today's number, 15. That's the percentage of Americans who say they'd be willing to take a job where their direct supervisor is an AI program.

2:07We're not sure if that tells us how good American AI has gotten or how bad American bosses are. I'm sorry. Money market matters. If money is evil, then that building is hell. The show goes on! The folks in there are watching so soon! Welcome to Profity Markets. I'm Ed Elson. It is May 7th. Let's check in on yesterday's market vitals. The major indices extended their rally on reports that the US and Iran were reviewing a deal to end the war and gradually reopen the Strait of Hormuz. That news also sent Brent crude tumbling. Treasury yields also dropped. Meanwhile, AMD stock soared nearly 20 % after the chipmaker beat expectations and raised its guidance.

2:56OK, what else is happening? Chinese AI startup DeepSeek is targeting a$50 billion valuation in its first ever fundraising round. Leading the financing is China's biggest state-backed semiconductor investment vehicle known as the Big Fund. And the country's goal, build a full stack AI ecosystem that can rival the United States. DeepSeek first grabbed Wall Street's attention last year with R1, a powerful model that was built at a fraction of the cost of leading Silicon Valley LLMs. And last month, it released V4, a model that is now competitive with top US players on a number of benchmarks. So here to discuss DeepSeek and the state of the US versus China AI race, we are speaking with our friend Alice Han, director at GreenMantle and co-host of the China Decode podcast.

3:50Alice, thank you for joining us again. I kind of want to start with the valuation here,$50 billion, because at first glance, it seems quite low when you compare it to Anthropik, which is trading at a trillion. ChatGPT, OpenAI, sorry, also trading at a trillion on the secondary markets. These companies are ready to go public. And then I also know that this is an extremely powerful model. It's extremely popular. It's the leading model in China. So I guess that number$50 billion strikes me as small. I guess, would you agree? And then also, what else strikes you about this news? I completely agree, Ed.

4:28But this is a feature, not a bug, of Chinese valuations. Even if you look at Chinese tech companies historically, the valuations at the Ford P's, much lower than what you see in the MAG-7, what you see in the U.S. This is a feature of Chinese capital markets just not being as developed as they are in the U.S. And more importantly, in terms of the fundraising for DeepSeek itself, yes, there will be some private fundraising, but it's largely, I think, going to be state-led as opposed to what you're seeing in the U.S. with the anthropics and the open AIs of the world. So I think it's completely reasonable and within the bounds of what I expect that this valuation is orders of magnitude lower than what we're seeing in the States.

5:10More importantly, I think when you look at the capabilities, what has been interesting is that it has now announced, DeepSeq v4, that it can do a 1 million token context window. This is comparable to Gemini and what we're seeing out of Anthropic and OpenAI. But there's big, big questions moving forward about the hardware side of things, which we can get into. Will DeepSeq be able to access via Huawei and elsewhere the kind of chips it needs to continue to power its models. Well, that is exactly where I would like to go. And actually, I want to play a clip for you. This was an exchange recently between Jensen Huang and Dworkesh Patel, a podcaster who had Jensen Huang on his program.

5:56And they were talking about China and the fact that NVIDIA has been selling chips to China. And it got quite heated and it really pressed on this exact issue of, should we be selling chips to China if they're going to use those chips and use them to develop extremely powerful LLMs and AI models? Let's just play the clip and get your reaction. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation. Why? Is that? Because, I mean, currently you can have a model like DeepSeek that can run on any accelerator. Why would that stop being the case in the future? Well, suppose it doesn't.

6:31Suppose it's optimized for Huawei. Suppose it's optimized for their architecture. It would put others at a disadvantage. You described the situation. A company develops software, develops an AI model, and it runs best on the American tech stack. I saw that as good news. You set it up as a premise that it was bad news. I'm going to give you the bad news. That AI models around the world are developed and they run best on not American hardware. That is bad news for us. So I guess the striking thing here is, one, is how uncomfortable Jensen Huang appears to be when he gets this question about China.

7:07I mean, this is the first time I've seen him as defensive as he was in this exchange. But it does get to the heart of that question, like, should we be selling chips to China? And also, where are we in that stage? Because we know that the policy's gone off, on, off, on. We keep on changing our minds. if you could just dive into all of that for me. Yeah, well, firstly, I'm a big fan of both Jensen and Dorkesh, and I listened to that very attentively very recently. I would say just to break it down a little bit, what is important is that Huawei's Ascend 950PR, which just came out and it's been powering the DeepSeek V4, the big companies like ByteDance, Barber, Tencent are now rushing to order these chips.

7:53These are inference chips. And so just looking at the stats alone, they can perform better on inference by 2.87 times more than the existing NVIDIA chips that can be sold to China. These are the H20s, not the H200s. Although those have been approved, China hasn't allowed those imports as of yet. And secondly, they improved the multimodal generation efficiency by 60%. So just looking at what is available from NVIDIA compared to Huawei, apples and oranges, Huawei is doing better on the inference side of things. But if you compare that to the H200s or if you compare that to the Rubens, Huawei is still a great deal behind.

8:33But it seems like Beijing is trying to prioritize a domestic ecosystem, a domestic marketplace of inference chips, such that it will no longer have to rely on NVIDIA. So basically, exactly what Jensen Huang expressed as his big fear to Dukesh, which is that if you continue to push China with, for instance, this chip export controls, the chip equipment export controls that just recently have been announced against Huahong, which is a major Chinese semiconductor company, then you incentivize across the supply chain China to say, hey, we are not going to import your next generation, although they will be outdated of chips.

9:17We want to prioritize what is coming out of Huawei and SMIC. Now, what does that mean for Chinese AI companies? It means that, yes, to some extent, the performance will be behind what you see in America. But the way that I think about it is that it creates good enough alternatives. China is prioritizing inference, and inference is still an open game in terms of chip-making capabilities. Yes, NVIDIA is really the king when it comes to the training side of the chip infrastructure, but it's still unclear whether or not it can be the winner when it comes to inference chips. And here I think the Chinese companies like Huawei and SMIC could really give NVIDIA a run for its money in the long term.

10:00But right now, NVIDIA is the clear leader. A lot of U.S. companies like Broadcom, AMD are also clear winners. The last thing that I will end on is that currently there isn't enough chips being produced domestically by Chinese chip makers to meet the demand on compute and inference. And that is going to be in the short term the biggest bottleneck. China ordered, these Chinese companies rather, ordered 2 million H200s earlier this year because there was so much demand for its AI models. None of that has been approved, obviously, by Beijing. Right now, Huawei is saying that it can do 750 ,000 units of its Ascend chips.

10:42If you just look at the numbers alone, and if you even look at sort of the longer term statistics on how many chips China can produce and the compute output, China is, by some estimates, only going to be able to produce 2 % of what NVIDIA and TSMC can produce in 2027 by looking just at compute outputs. That is, I think, huge. But right now, the political priority trumps what is efficient. And the political priority from Beijing is to kickstart a domestic ecosystem that will be able to rival NVIDIA long-term. So just going back to some of those chips that you mentioned there. So you've got the NVIDIA H20, which was the chip that was essentially designed to be a dumber, slower chip than the highest end that is created by NVIDIA.

11:30That was designed for China. And you're telling us Huawei's now got a chip that is three times more powerful on the inference perspective, almost three times more powerful. So they're not interested in those H20 chips so much anymore, it seems, but still significantly behind the H200 and the Rubin. and those are NVIDIA's really fantastic chips, and they're still behind on that front. I guess the question then becomes like, how far behind are they? And also, how quickly have they improved recently? Because Huawei, I mean, it was a name in the AI world, but not that big of a name. And suddenly it seems to be gaining a lot of momentum and a lot of steam.

12:14As an observer, I'm reading about it a lot more. I mean, are they getting close to a point where they could actually rival NVIDIA's most advanced chips, which we do not allow to go to China? They're not yet close, but there are smart workarounds. So firstly, on the training front, obviously, there's a lot of news about Chinese AI companies distilling models coming out of Anthropic and OpenAI. That helps them. It's almost as if it's a student cheating on an exam, and it's just copying a smarter student, right? This is what is allowing them to really catch up On the inference front, and the inference is really when a student, having learned the information, takes the exam.

12:53So even the answers to the queries that you give to, say, an LLM or to an agent. Here, Chinese models are actually doing extremely well. And I think the long-term bet is that if a Huawei can produce a lot more in terms of chips, it can do these clusters. And you may have heard of this idea of this cluster architecture. So, for instance, if a Huawei chip is 10 times less compute efficient than an NVIDIA chip, maybe you cluster 10 of them to equal the equivalent of an NVIDIA. That, I think, will be the long-term bet. A lot of it rides on how much capacity Huawei and SMIC can do. Right now, and this connects to Iran, there is a real risk that helium, for instance, which is critical for chipmaking, is actually going to be a bottleneck and a headwind for Chinese chipmakers.

13:41There's another thing that could be a headwind is if the U.S. decides it's actually going to do more restrictions on chip equipment. China's even more reliant actually on the chip equipment providers, like AMD, for instance, and ASML. If there are more restrictions there, that could again set back China's domestic chip production. We'll be right back after the break. And by the way, we are heading out on tour at the end of the month. So if you want to get tickets to a show near you, head to ProfGMarketstour.com. It's going to be a lot of fun.

14:38that would be a waste of anyone's ad budget. So when you want to reach the right professionals, you can use LinkedIn ads. LinkedIn has grown to a network of over 1 billion professionals and 130 million decision makers, according to their data. That's where it stands apart from other ad buys. You can target your buyers by job title, industry, company role, seniority skills, company revenue, all so you can stop wasting budget on the wrong audience. That's why LinkedIn ads boasts one of the highest B2B return on ad spend of all online ad networks. Seriously, all of them. Spend$250 on your first campaign on LinkedIn ads and get a free$250 credit for the next one.

15:14Just go to linkedin.com slash Scott. That's linkedin.com slash Scott. Terms and conditions apply.

15:24Support for the show comes from Amazon. There are the things you can plan for. A first birthday party, a movie marathon, a renter-friendly bathroom reno. And then there are the things you can never plan for. A surprise rainstorm, a Blu-ray player calling it quits, stick-on tiles that looked way better on the package. For all things planned and unplanned, Amazon has you covered. You'll find low prices on everyday essentials and last-minute lifesavers. Shop Amazon and save on essentials. Save the everyday.

16:01Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the Unreal College Deal. Everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 Premium and a year of Xbox Game Pass Ultimate with a custom color Xbox wireless controller. Learn more at windows.com slash student offer. While supplies last, ends June 30th. terms at aka.ms slash college pc we're back with prof g markets when you think about all of the different i guess pieces of leverage or i guess all of the different races would be the the right way to describe it i mean you've got compute capability america's winning on that front it seems.

16:54You've got energy, you've got supply chain, you've got, I think, access to the Strait of Hormuz seems to be increasingly an important piece of this. And when you think about all of the different fronts on which the AI war is being waged, on which of them is China winning against America, and on which of them is America winning? So I think about it as a five-layer cake, and in this regard, I was happy that Jensen Huang also had a similar framework, you know, from upstream to downstream. So we start from the upstream, which is going to be, you know, basically the rare earths that are used, for instance, that power semiconductors and data centers.

17:41This is where China has, you know, huge dominance. We've seen that as of last year as well with the export restrictions. The second layer that we're going to look into is the energy layer, where China again has, it seems, on the electricity front a great deal of leverage. It has twice the amount of electricity output that the Americans have. That is going to be a bottleneck for American data center rollout as well moving forward. Number three is the data center infrastructure layer where the U.S. has done exceptionally well. China will only get to 60 gigawatts. It's targeting by 2030. The U.S.

18:17is basically already there as of now. And then the fourth layer is going to be the models themselves, where again, you know, based on the status quo right now, the U.S. is still leading, although that gap is at times closing because Chinese models are getting more efficient. They're distilling and improving at a rapid pace. And then it's the application layer where I actually think it's a bit of a mixed bag. China is making really interesting applications. This is why Meta wanted to buy and failed to buy Manus, its agentic AI that came out of mainland China and moved to Singapore. I think there will be really interesting applications coming out of China, especially on the agentic front.

19:01They could rival what you see in America. It's not just an LLM game and that China has proven that. So I see it as a five-layer cake right now. It's really mixed across the board. But China tends to dominate on some of these upstream as opposed to downstream. You mentioned Manus there, which is this company that Meta was trying to buy as a Chinese company, or I guess a Singapore-based company, but as Chinese roots, and then Beijing banned it, which seems to sort of lend itself to the argument that this is becoming quite a hostile race. It's fully a competition. If America tries to get access to stuff that's happening in China as it relates to AI, Beijing will ban it.

19:44They will stop it. Same thing is true over in America. And an analogy that I am increasingly hearing when it comes to the AI race is that it's sort of like getting your hands on a nuclear weapon. And this is more and more relevant the more we hear in the headlines. I mean, Anthropic coming out with Mythos, and we learned that there's this AI technology that could hack every single cybersecurity system in the world. It could literally bring the infrastructure, the digital infrastructure of a nation to its knees. It's seeming more and more that maybe it's a fair comparison, but I don't know, maybe it's a bridge too far.

20:24Do you think that that is a reasonable analogy when it comes to the AI race. Just to clarify, which analogy in particular you're referring to, Ed? The nuclear weapon. The nuclear weapons. You know, Kissinger, before he passed away a couple of years ago, wrote a book, as you probably know, with Eric Schmidt, in which he basically said that AI and autonomous weapons was going to be what nuclear weapons were during the Cold War when he was Secretary of State and National Security Advisor. And he was deeply worried, so much so that he, in the last few years of his life, went to Beijing to speak with Xi Jinping and high-level officials to set up the framework for strategic dialogue, which Biden did take up towards the end of his administration, a strategic dialogue between Beijing and Washington on AI arms controls.

21:13Now, there's an open-ended question as to whether or not when Trump meets with Xi likely next week in Beijing, if there is going to be a kickstart to that discussion, a new strategic dialogue will emerge. But I was just in Beijing last week, Ed, and what was very startling to me was the level of fear as to how effective U.S. AI technology has been in its application in the Iran conflict. There was a lot of concern as well about the use of CLAWD by the Department of War, by the defense community in America. And when Mythos came out, I've seen a consternation in China about what that means for Chinese national security.

21:59One thing that I will end on is that this is part of the reason why China has tried to make AI not just an economic issue, but also a national security issue. It also made quantum a national security issue, too. Quantum is also going to be key for decryption and encryption, to your point. If there is going to be a mythos that can hack into every system, quantum may be able to leapfrog that and create a new fortress. So these are all these ongoing concerns. But I think the view from Beijing right now is that there needs to be high-level strategic discussion. I won't be surprised if after this meeting between Trump and Xi, there is a revival from the Beijing side about ongoing strategic dialogue concerning AI weapons, because right now, things do look very scary from China's vantage point.

22:48Just going back to the Dworkesh Jensen Huang interview, I mean, it seemed like there was sort of this miscommunication in the dialogue where, on the one hand, the argument is being made to Jensen Huang that if AI is a nuke, if it is comparable to a nuke, you shouldn't be selling equipment that they are going to use to go and build that nuke and therefore nuke America. and then he's saying no they're already gonna they're already building it they they have the ability so do they do we want them to be using u.s made equipment or do we want them to be using china made equipment which i feel like didn't quite get to the heart of the disagreement and something i was thinking about this and i'd like to get your thoughts i mean running with the nuke analogy.

23:40China has nukes. They haven't dropped them on the US. And the reason they haven't dropped them on the US, I think, is because they don't want to. Mutually assured destruction. Mutually assured destruction. And so, to me, it seems as though this question of the AI race between America and China, increasingly, it's not actually about the technology. It's about diplomacy. And it's about the relationship between the two. And it's about getting us to a place where China doesn't feel that there is any reason to drop the AI equivalent of a nuclear bomb on America. And it seems to be that that's the trajectory where this is headed.

24:21This is a political discussion, not a technological one. I'd just be interested to get your thoughts on that and how this topic is kind of evolving. I love that question. As I was listening to that podcast, it was clear that neither of them were historians or politics majors as they approached the question. Because as you know, Ed, China built the nuclear with no American help. It was largely Soviet aid. And then they had engineers, physicists who came and built it by 64. They surprised the world by launching a successful nuclear test in China. That shows you historically that China can, in this AI age, create its own AI capabilities without American largely support or input.

25:08I think what they confuse is this question, can in the meantime Wall Street monetize China's AI development? Clearly, Jensen has an incentive to make that case, which benefits parts of America. But regardless of that argument, China will find a way to create a rival AI system. If that AI system is antagonistic or not, that question and its answer resides on your point, Ed, which I completely agree. Is there going to be continued strategic dialogue and diplomacy? Kissinger understood this and he was harping on about it even towards the end of his life. And I think it would be very, very sad and ultimately tragic for humanity if we don't have Washington and Beijing continue discussion about their capabilities and intentions as it pertains to AI.

26:00And more importantly, if there isn't track to diplomacy where these Chinese tech companies are also involved, because you also need experts in the room to help both sides assess the capabilities. My concern right now is those two sides, both on the expert side, as well as on the diplomatic side, are very far apart. And that, as Kissinger rightly predicted in his last book on AI is going to create massively tragic outcomes that we have yet to see far greater than we have yet to witness in our lifetime. But if we do ultimately come to a point where there is mutually assured destruction in AI, then we may be able to achieve equilibrium.

26:40But that rests on both sides having extremely strong and telegraphed AI capabilities. Alice Hahn, director at Green Mantle, co-host of the China Decode podcast. Alice, Trump and Xi are likely to meet next week. Maybe we'll get some answers when that happens. Thank you so much for joining us. Thanks so much, Ed.

27:04Here's a question for you. Is AI making us dumber? That is the question our research lead, Mia Silverio, asked in one of her recent substacks, and her findings were pretty conclusive. The answer is yes. According to Dr. Jared Horvath, a neuroscientist who recently spoke in front of Congress, every generation in history has been smarter than their parents' generation except for Gen Z. Young people underperform on, quote, every cognitive measure from IQ to literacy to basic memory skills. One look at the data will confirm that this is indeed true. Roughly half of American 12th graders are now scoring below a basic level of math, and roughly a third of them don't have basic reading skills.

27:55In fact, that is the worst rating that we've seen among that group in three decades. But it isn't just America. Over the past two decades, science, math, and reading scores among teenagers in high-income countries have all fallen nearly 5%. Meanwhile, literacy rates for adults have fallen roughly 3%. In sum, yes, we are literally getting dumber. Now, is this just because of AI? No. As Mia points out, there are other factors such as social media addiction and also the pandemic, which stunted learning for millions of children around the world. But we should also acknowledge the extent to which AI is still playing a role here and the fact that it actually is accelerating the trend.