WEBVTT - Magnificent Seven's Long Shadow Over Asia Tech 0:00:02.720 --> 0:00:07.480 You're listening to Asia Centric from Bloomberg Intelligence, the podcast 0:00:07.560 --> 0:00:10.640 that pulls back the curtain non global business so you 0:00:10.760 --> 0:00:15.160 can invest better across the Pacific rim. I'm Tom Corbett 0:00:15.160 --> 0:00:16.439 in Hong Kong, and. 0:00:16.360 --> 0:00:20.560 I'm John Lee. Artificial intelligence is generating a level of 0:00:20.600 --> 0:00:23.360 investor excitement not seen since the dot com era. 0:00:23.760 --> 0:00:27.479 Call it a feeding frenzy, call it a buying binge. 0:00:27.520 --> 0:00:31.160 The rush to board the AI bandwagon has powered America's 0:00:31.360 --> 0:00:36.320 Magnificent seven stocks to new heights, and giddy investors are 0:00:36.520 --> 0:00:37.880 rushing to get on board. 0:00:38.200 --> 0:00:42.319 Will artificial intelligence deliver on the high or investors at 0:00:42.440 --> 0:00:45.720 risk of being left empty handed? And what about Asian 0:00:45.800 --> 0:00:49.000 tech stocks? Can they catch up to their American counterparts. 0:00:49.280 --> 0:00:54.720 Let's bring in two Bloomberg Intelligence experts. Mandeep Seeing is 0:00:54.840 --> 0:00:58.600 Global head of Technology Research based in New York. 0:00:58.840 --> 0:01:01.800 And Robert Lee see Asia Tech Analysts based here in 0:01:01.840 --> 0:01:02.320 Hong Kong. 0:01:02.600 --> 0:01:05.280 Man Deep and Rob welcome to Asia Centric. 0:01:05.760 --> 0:01:08.640 Thanks for having us on and y'are excited to be 0:01:08.680 --> 0:01:09.920 on the podcast. 0:01:09.760 --> 0:01:13.040 Like wife Tom and John, Thanks for having us Mandy. 0:01:13.319 --> 0:01:15.680 Let's get straight into it. Are we in an AI 0:01:15.880 --> 0:01:16.760 driven tech bubble? 0:01:17.560 --> 0:01:20.840 Well? I think bubble is a term that gets used 0:01:21.080 --> 0:01:24.600 probably so frequently that in my lifetime I probably would 0:01:24.600 --> 0:01:27.240 have come across at least one hundred bubbles so far. 0:01:27.440 --> 0:01:31.440 So clearly, you know, there are times when things come 0:01:31.480 --> 0:01:35.280 across as fad and there isn't proper follow up in 0:01:35.360 --> 0:01:42.279 terms of monetizable opportunities. AI, especially generative AI, doesn't seem 0:01:42.319 --> 0:01:45.240 to be that type of fad. And the reason I 0:01:45.360 --> 0:01:48.840 say that is what we have seen so far is 0:01:48.960 --> 0:01:53.000 not only you know, in video really getting that pricing 0:01:53.080 --> 0:01:56.200 power in terms of the GPUs that are being used 0:01:56.200 --> 0:02:00.919 for AI training, but also follow up application, whether it's 0:02:01.440 --> 0:02:05.400 on the software side with chatbots that can streamline back 0:02:05.480 --> 0:02:09.680 end operations, contact center stuff, as well as other type 0:02:09.680 --> 0:02:15.400 of optimizations and copilot use for developers. So with Jenny 0:02:15.400 --> 0:02:19.280 and I, we are talking about workflows getting streamlined, and 0:02:19.320 --> 0:02:22.600 when we talk about workflows, they're across the board and 0:02:22.639 --> 0:02:24.320 we see a lot of potential over there. 0:02:25.320 --> 0:02:29.880 And you also mentioned these AI companies are monetizing artificial intelligence. 0:02:30.360 --> 0:02:33.680 Is that a key difference between say the tech bubble 0:02:33.760 --> 0:02:34.880 in nineteen ninety nine. 0:02:36.120 --> 0:02:39.680 I think so you are seeing a lot of upfront 0:02:39.760 --> 0:02:43.760 revenue for companies that are playing to this trend. And 0:02:44.120 --> 0:02:48.600 it is a secular tailwind in the sense that not 0:02:48.639 --> 0:02:52.160 only it's what Nvidia is doing right now, but think 0:02:52.240 --> 0:02:55.560 of the recurring revenue that these companies can generate through 0:02:55.600 --> 0:03:01.200 copilots or the subscription revenue streams that Microsoft is able 0:03:01.240 --> 0:03:04.600 to add on to its office bundle. I mean this 0:03:04.800 --> 0:03:08.799 is real. Companies are subscribing to this because they can 0:03:08.960 --> 0:03:12.919 see knowledge workers getting a productivity boost. So I would 0:03:12.919 --> 0:03:17.080 say whether it delivers on the promise that everyone feels 0:03:17.160 --> 0:03:19.800 this technology brings to the fore I mean we ourselves 0:03:19.880 --> 0:03:22.560 have done our forecast, a ten year forecast where we 0:03:22.600 --> 0:03:25.480 think it's going to drive one point three trillion in 0:03:25.480 --> 0:03:28.720 incremental revenue by twenty thirty two. So we think this 0:03:28.800 --> 0:03:30.760 is going to grow at a kegar of forty percent 0:03:30.800 --> 0:03:33.120 over the next ten years. But look, there will be 0:03:33.280 --> 0:03:36.600 ups and downs, as we have seen with any technology wave, 0:03:36.720 --> 0:03:39.920 and so far it seems to be delivering on the 0:03:39.960 --> 0:03:43.600 promise with revenue streams showing up in different companies that 0:03:43.680 --> 0:03:44.120 we follow. 0:03:45.360 --> 0:03:47.960 Robert Lee, I know that you've been through the dot 0:03:48.000 --> 0:03:51.560 com bubble as with the rest of us. What similarities 0:03:51.600 --> 0:03:54.120 and differences do you see between the AI craze now 0:03:54.160 --> 0:03:57.240 and the dot com bubble back then, Uh. 0:03:57.320 --> 0:03:59.600 Well, look at my reflection in the camera and a 0:03:59.640 --> 0:04:01.440 few more gray hairs, and I used to have I 0:04:01.480 --> 0:04:03.640 do remember the dot com bubble and I started my 0:04:03.840 --> 0:04:06.600 cell side career then. But I think the difference then 0:04:06.800 --> 0:04:10.280 was the sort of infrastructure was built out. The web 0:04:10.400 --> 0:04:13.839 was built out with little thought at that point as 0:04:13.840 --> 0:04:16.440 to how again, what's a business model, how is it 0:04:16.480 --> 0:04:21.000 going to be monetized? Whereas AI? You know, I guess 0:04:21.000 --> 0:04:23.159 to many people hadn't really thought or heard of AI 0:04:23.400 --> 0:04:26.680 until chat GPT came along about a year ago. But 0:04:26.800 --> 0:04:29.719 AI has been around for decades in a form of 0:04:29.839 --> 0:04:33.120 machine learning as it's referred to, and so there's been 0:04:33.160 --> 0:04:36.000 a lot of research at university level and at corporate 0:04:36.080 --> 0:04:39.359 level going into AI for many decades. What's changed and 0:04:39.400 --> 0:04:42.240 what's helped to accelerate it along the curve is obviously 0:04:42.279 --> 0:04:47.000 some fine tuning of the algorithms which were previously focused 0:04:47.000 --> 0:04:49.919 more on numerical data and now they can handle words 0:04:50.000 --> 0:04:53.680 and a language. And also the huge amounts of computing 0:04:53.800 --> 0:04:57.839 and storage power that is available today versus decades ago, 0:04:58.200 --> 0:05:02.160 and so that has enabled these sort of AI applications 0:05:02.279 --> 0:05:05.640 which have been there for many decades, but perhaps below 0:05:05.720 --> 0:05:09.239 the horizon, unknown and unseen so many people. It's allowed 0:05:09.240 --> 0:05:13.720 a proliferation of those applications and opening up substantial new 0:05:13.800 --> 0:05:16.880 markets and opportunities. In arguably what's going to be not 0:05:16.920 --> 0:05:20.280 only a multi year secular trend, but a multi decade 0:05:20.320 --> 0:05:23.000 secular trend. So it's a very very exciting place to 0:05:23.040 --> 0:05:26.200 be and very very different to what things look like 0:05:26.279 --> 0:05:27.600 in the late nineteen nineties. 0:05:28.680 --> 0:05:32.880 Manleep seeing does it surprise you how quickly AI has 0:05:33.200 --> 0:05:37.200 erupted under the investment landscape. 0:05:36.080 --> 0:05:40.600 Not really if you follow the trajectory of evolution, and 0:05:40.880 --> 0:05:43.440 as Rob alluded to, you know, machine learning has been 0:05:43.480 --> 0:05:46.920 around for a number of years. The whole large anguage 0:05:46.960 --> 0:05:51.880 model training is about ingesting large amounts of data, using 0:05:51.920 --> 0:05:56.760 the compute capacity and really putting guardrails around the technology 0:05:56.800 --> 0:06:00.400 to the point that it's not only productive, but uh, 0:06:00.680 --> 0:06:03.240 you know, you're not at the risk of it being misused. 0:06:03.279 --> 0:06:06.039 I think that is one of the biggest concerns that 0:06:06.320 --> 0:06:10.800 I see, is the potential for this to be misused 0:06:10.800 --> 0:06:15.360 by bad actors, and I think AI always had that promise. Clearly, 0:06:15.440 --> 0:06:18.520 if we have made strides in terms of you know, 0:06:18.839 --> 0:06:22.400 deploying large anguage models at scale and to be used 0:06:22.440 --> 0:06:26.359 in productivity apps. But the key is how do you 0:06:26.839 --> 0:06:30.240 ensure safety and you make sure that the guardrails are 0:06:30.279 --> 0:06:34.000 there for all kinds of use cases. And to me, 0:06:34.200 --> 0:06:37.920 that still remains to be proven, Man Dave. 0:06:38.279 --> 0:06:42.280 Nvidia is probably the most important stock right now and 0:06:42.440 --> 0:06:46.359 everyone eagerly awaits when it reports its quarterly earnings. Is 0:06:46.360 --> 0:06:50.360 the bullmarket totally relying on Nvida continuing to beat estimates 0:06:50.480 --> 0:06:52.320 or is there any other metrics we should look at. 0:06:53.080 --> 0:06:56.520 When you look at you know, team stock like Nvidia 0:06:56.680 --> 0:06:59.359 is it's supposter child for generative AI and for a 0:06:59.400 --> 0:07:02.560 good reason. And I think when you look at how 0:07:02.600 --> 0:07:07.480 they kind of became the name to resemble this wave 0:07:07.640 --> 0:07:11.600 of gen AI, it's because they had these chips that 0:07:11.640 --> 0:07:15.920 they had developed over the last thirty years around gaming 0:07:16.160 --> 0:07:20.680 and parallel processing, which was a different type of architecture 0:07:20.720 --> 0:07:24.320 when you compare it to the traditional CPU architecture that 0:07:24.480 --> 0:07:27.840 is used for pretty much all kinds of processing up 0:07:27.920 --> 0:07:32.800 until now. So this concept of accelerators and parallel processing 0:07:33.080 --> 0:07:36.600 is new when you think about the data centers, and 0:07:36.640 --> 0:07:39.160 the reason why it's taken off in such a big 0:07:39.200 --> 0:07:43.040 way is because when you're dealing with large aguage models. 0:07:43.120 --> 0:07:46.800 We're talking about a scale that we haven't used for 0:07:46.920 --> 0:07:50.960 processing before. Right now, you're seeing supply constraints. I mean, 0:07:51.000 --> 0:07:54.480 think of how fab capacity increases over time. It normally 0:07:54.520 --> 0:07:57.400 increases you know, mid single digit, because that's how the 0:07:57.480 --> 0:08:01.240 KAPEX investments used to be. How we are talking about 0:08:01.400 --> 0:08:05.920 insatiable demand for video chips and that to leading note chips. 0:08:05.920 --> 0:08:08.760 We're talking about three to five nanometer fabs that are 0:08:08.800 --> 0:08:12.640 needed for producing these accelerator chips and we don't have 0:08:12.680 --> 0:08:16.160 that capacity. Combine that with you know, COOS packaging that's 0:08:16.240 --> 0:08:19.960 needed for these chips. So clearly the supply constraints are 0:08:20.000 --> 0:08:23.040 what's driving the asps. But I won't pay too much 0:08:23.080 --> 0:08:26.640 attention to asps even though that's what's driving the high 0:08:26.760 --> 0:08:29.520 growth rates, the fifty percent plus kegres that we've seen 0:08:29.560 --> 0:08:32.719 with Nvidia. It all comes down to what is the 0:08:32.760 --> 0:08:35.840 addressable market for data centers over the next ten years, 0:08:35.840 --> 0:08:39.000 and based on our one point three trillion number, half 0:08:39.040 --> 0:08:42.679 of that is hardware needed for training and inferencing. So 0:08:42.720 --> 0:08:45.800 clearly there is a long runway when it comes to hardware. 0:08:47.320 --> 0:08:51.480 Man leeps saying how sustainable is Navidio's competitive advantage. 0:08:51.480 --> 0:08:54.760 Do you think now, look at the biggest buyers of 0:08:55.440 --> 0:09:00.280 Nvidia's chips, it's your hyperscaler companies, the companies that deploy 0:09:00.360 --> 0:09:04.480 them on cloud to be used for widespread consumption. Anytime 0:09:04.600 --> 0:09:09.520 you see such big concentration where your customers can become 0:09:09.520 --> 0:09:16.920 potential competitors, which is the case with Nvidia's buyers. Here, Amazon, Google, Microsoft, Meta, 0:09:17.400 --> 0:09:20.640 all of them have said they are developing their own 0:09:20.760 --> 0:09:24.600 chips because they think vertical integration would give them an 0:09:24.760 --> 0:09:27.920 edge over time, and we've seen that with the Apple model. 0:09:27.960 --> 0:09:30.800 I mean, think of why Apple's smartphone is so successful. 0:09:30.800 --> 0:09:33.240 It's a vertical integration from chips all the way to 0:09:33.480 --> 0:09:37.040 software and operating system. So clearly there is a risk. 0:09:37.400 --> 0:09:41.320 But right now, by the time these companies develop a 0:09:41.440 --> 0:09:45.240 version of their accelerator, Nvidia would be at least four 0:09:45.320 --> 0:09:48.400 or five generations ahead in terms of the performance. So 0:09:48.559 --> 0:09:51.760 that's why you know every three to four months, these companies, 0:09:51.920 --> 0:09:55.880 especially Nvidia NAMD, they will keep progressing to the point 0:09:55.880 --> 0:10:00.200 their processor, their accelerator performance will be ahead of what 0:10:00.240 --> 0:10:03.560 the hyperscillers are developing. And I think that kind of 0:10:03.640 --> 0:10:04.679 edge will remain. 0:10:04.440 --> 0:10:07.360 For a while Rob Lee, did you want to weigh in? 0:10:07.920 --> 0:10:11.160 I agree with that. I mean, the semiconductor is at 0:10:11.160 --> 0:10:14.920 the lynchpin of the global supply constraints at the moment 0:10:14.920 --> 0:10:18.400 within the AI sector, you know, particularly coming from TISMC, 0:10:18.960 --> 0:10:22.600 given the supply bottleneck that exists within that business, because 0:10:22.679 --> 0:10:27.679 TSMC is the only semi conductor foundry globally with true 0:10:27.760 --> 0:10:32.800 deep experience in leading edge technology derived on production equipment 0:10:32.800 --> 0:10:36.160 from ASML. So it's those two key companies which are 0:10:36.200 --> 0:10:39.640 at the center of this supply bottleneck which is benefiting 0:10:39.679 --> 0:10:43.800 in video at the moment and obviously causing this constraining factor. 0:10:44.360 --> 0:10:47.560 Is clearly benefiting in video, but also to the detriment 0:10:47.640 --> 0:10:51.920 of the Chinese given the export constraints, can't access leading 0:10:52.000 --> 0:10:55.560 edge technologies, and also given the supply constraints on the 0:10:55.600 --> 0:11:00.559 semiconductor founder of fabrication side, they also can't acts leading 0:11:00.640 --> 0:11:04.760 edge semiconductor equipment in order to fabricate their own chips. 0:11:05.240 --> 0:11:07.240 So I think I just at those two points in 0:11:07.280 --> 0:11:09.520 there before. 0:11:09.240 --> 0:11:11.280 We get onto like Asia Tech. But Mande, but I 0:11:11.280 --> 0:11:15.120 wanted to just broaden the discussion to like the stocks 0:11:15.160 --> 0:11:20.360 outside in VideA, How did you grade the Magnificent seven, Like, 0:11:20.480 --> 0:11:23.040 what scorecard would you give them to be an ABAC 0:11:23.600 --> 0:11:25.160 in terms of the different companies. 0:11:26.000 --> 0:11:30.240 Well so, my lens right now in terms of evaluating 0:11:30.320 --> 0:11:34.240 Magnificent seven is around which out of the magnificent seven 0:11:34.640 --> 0:11:39.040 have their own foundational models and based on that, Google, 0:11:39.440 --> 0:11:43.240 Meta and Microsoft because of their early Open AI partnership, 0:11:43.360 --> 0:11:47.040 stand out to me as companies that have their own 0:11:47.080 --> 0:11:51.720 foundational models natively deployed on their cloud and that gives 0:11:51.760 --> 0:11:54.400 them an edge because we are talking about generator AI. 0:11:54.920 --> 0:11:58.400 Amazon is the largest cloud player. Clearly they have a 0:11:58.400 --> 0:12:01.440 lot of potential to bene fit from the trend, but 0:12:01.520 --> 0:12:05.480 what they haven't figured out so far is their foundational 0:12:05.520 --> 0:12:09.320 model and what are they going to standardize on. Same 0:12:09.320 --> 0:12:12.160 thing with Apple. I mean, clearly Apple has the install 0:12:12.240 --> 0:12:15.520 based when it comes to you know, on device AI 0:12:15.920 --> 0:12:20.599 and inferencing, but we don't know much about their GENI strategy. 0:12:20.640 --> 0:12:24.280 And Nvidia clearly is the biggest beneficiary of the trend, 0:12:24.360 --> 0:12:28.120 so they get very high probably a plus so far. 0:12:28.320 --> 0:12:32.679 So that leaves Tesla. Tesla, I think is an interesting 0:12:32.760 --> 0:12:37.080 play a derivative of their early strategy around the FSD 0:12:37.200 --> 0:12:40.400 software which wasn't built on generative AI per se, but 0:12:40.720 --> 0:12:44.400 they have changed the latest version of their software to 0:12:44.640 --> 0:12:47.640 leverage generative AI, so clearly they have a headstart when 0:12:47.640 --> 0:12:50.920 it comes to generative AI on the automotive side. But 0:12:51.120 --> 0:12:54.600 I think if you use that framework, companies with their 0:12:54.640 --> 0:12:58.600 own foundational models will have a persistent advantage than the 0:12:58.640 --> 0:12:59.360 ones that don't. 0:13:00.800 --> 0:13:04.800 Our guests are man Deep Sing and Rob Lee, both 0:13:04.920 --> 0:13:09.200 tech analysts with Bloomberg Intelligence. Mandeep and Rob Let's talk 0:13:09.240 --> 0:13:13.600 about Taiwan Semiconductor. Some in the West refer to it 0:13:13.679 --> 0:13:17.480 as the most important company, the most important manufacturer that 0:13:17.480 --> 0:13:21.480 they've never heard of. Sixty to seventy percent of the 0:13:21.520 --> 0:13:25.800 global chip market is what it commands, talk about its 0:13:26.160 --> 0:13:30.960 positioning in the broader tech AI matrix and why should 0:13:30.960 --> 0:13:31.840 investors care? 0:13:32.800 --> 0:13:36.720 I mean, look, and Rob probably has better visibility on 0:13:36.720 --> 0:13:40.160 what TSMC is doing in the region. But from a 0:13:40.240 --> 0:13:44.560 Western perspective, there is this reliance on TSMC when it 0:13:44.600 --> 0:13:48.400 comes to all the fabulous chip makers, including Nvidia that 0:13:48.520 --> 0:13:51.120 we wave about so much. At the end of the day, 0:13:51.640 --> 0:13:55.480 they rely on TSMC to make those GPUs that there's 0:13:55.679 --> 0:13:59.080 so much demand for. So clearly that is a choke 0:13:59.160 --> 0:14:03.120 point when you think about deploying generative AI. But I 0:14:03.240 --> 0:14:06.280 have no doubt, and you know, we just published two 0:14:06.720 --> 0:14:10.320 deep dives supply chain and semicap equipment here at BI 0:14:10.400 --> 0:14:14.760 and part of the conclusion was that the advanced chief 0:14:14.800 --> 0:14:19.320 manufacturing capacity will diversify over the next ten years, and 0:14:19.600 --> 0:14:23.720 that's where you will see companies like TSMC as well 0:14:23.760 --> 0:14:28.440 as other big fab makers actually invest their CAPEX dollars 0:14:28.480 --> 0:14:32.160 in terms of expanding their fabs outside the East Asia region. 0:14:33.040 --> 0:14:36.960 Robbie, you have some interesting perspectives about TSMC, including how 0:14:37.000 --> 0:14:39.560 it's affected Taiwan's exports over the past decades. 0:14:40.240 --> 0:14:43.080 Yes, well, the I wish I had a very precise 0:14:43.200 --> 0:14:45.000 state for you here, but at some point in the 0:14:45.080 --> 0:14:50.240 nineteen sixties, I believe Taiwan's number one export was actually sugarcane, 0:14:50.560 --> 0:14:54.040 so it was largely an agricultural based economy. And although 0:14:54.040 --> 0:14:59.200 that's sixty plus years ago, through significant government support, through 0:14:59.400 --> 0:15:03.920 cooperation with the universities in pushing people or encouraging people 0:15:04.040 --> 0:15:07.880 to study STEM subjects and engineering, the transformation of the 0:15:07.880 --> 0:15:11.800 Taiwanese economies phenomenal used to be referred to as one 0:15:11.800 --> 0:15:14.960 of the Asian tigers, if you remember. So, going back 0:15:14.960 --> 0:15:16.840 to what Mandeep was saying as well, and just a 0:15:16.880 --> 0:15:21.080 little bit of historical context. Again, it's twenty years or 0:15:21.120 --> 0:15:23.440 so ago. At the beginning of my investing career, I 0:15:23.440 --> 0:15:25.960 think the view within the semiconductor industry was the real 0:15:26.040 --> 0:15:29.160 value is in the design companies like Armholdings which were 0:15:29.160 --> 0:15:32.160 still around then, and companies designing their chips. That's where 0:15:32.160 --> 0:15:35.520 the value was added and was captured. The assembly. The 0:15:35.560 --> 0:15:40.320 fabrication was important, but to a lesser extent, but as 0:15:40.360 --> 0:15:43.320 a result of something referred to as Moore's law, as 0:15:43.320 --> 0:15:46.600 the complexity of cease chips has increased to the drum 0:15:46.640 --> 0:15:48.960 beat which defines to grow for it within the industry, 0:15:49.240 --> 0:15:53.080 with transistor densities doubling every eighteenth to twenty four months. 0:15:54.080 --> 0:15:58.000 Sem conductors these days are fabricated at the atomic scale, 0:15:58.640 --> 0:16:03.120 so they're incredibly technically complex and it's incredibly difficult to 0:16:03.200 --> 0:16:07.800 fabricate at such a high level of miniaturization. And TSMC 0:16:07.960 --> 0:16:11.320 really is the only company on the planet which has 0:16:11.560 --> 0:16:15.000 really kept a commanding lead in fact enhanced and built 0:16:15.040 --> 0:16:18.400 its commanding lead over the competition over that period. So 0:16:18.440 --> 0:16:21.840 we've seen a transfer of value from the design, which 0:16:21.920 --> 0:16:25.840 is still important that increasingly to the fabrication or manufacturing side, 0:16:26.040 --> 0:16:29.120 and that is why TSMC on a global scale is 0:16:29.160 --> 0:16:33.080 so important. Obviously with an interesting overlay of geopolitics and 0:16:33.080 --> 0:16:35.840 everything else, which makes the story both more interesting and 0:16:35.880 --> 0:16:39.200 more complex to analyze. But TSMC is really the key 0:16:39.320 --> 0:16:42.680 lynchpin of the global technology sector at the moment, I 0:16:42.680 --> 0:16:44.200 would say, and. 0:16:44.200 --> 0:16:46.200 Robert, if you look at the share prices of Asian 0:16:46.280 --> 0:16:49.480 tech like TSMC definitely appears to have benefited from the 0:16:50.080 --> 0:16:54.080 AI trend, but some other stocks in your coverage seems 0:16:54.080 --> 0:16:57.560 to have really missed the AI hyph. I'm talking about 0:16:57.560 --> 0:17:02.080 the big Chinese tech companies Ali and tenset. Now, prior 0:17:02.120 --> 0:17:05.520 to recording this podcast, I did look at the Bloomberg 0:17:05.600 --> 0:17:09.359 terminal and if you roll back the clock three years ago, Ali, 0:17:09.440 --> 0:17:12.640 Barber and Tencent in terms of market cap were actually 0:17:12.640 --> 0:17:15.679 comparable to the Magnificent Seven. If you look at the 0:17:15.840 --> 0:17:19.080 average market cap of the Magnificent seven now, it's almost 0:17:19.119 --> 0:17:23.280 ten times the size of Ali Barber. What went wrong 0:17:23.600 --> 0:17:28.080 with China tech, especially as it relates to AI, right, how. 0:17:27.960 --> 0:17:28.680 Long have we got? 0:17:28.840 --> 0:17:28.879 Is? 0:17:29.960 --> 0:17:36.480 You don't answer for that? Well, clearly the China tech 0:17:36.520 --> 0:17:39.760 sectors being caught up with a lot of concerns on 0:17:39.800 --> 0:17:43.959 the regulatory front, a lot of domestic regulation with Chinese 0:17:44.000 --> 0:17:46.639 government want to, you know, to level up the playing 0:17:46.680 --> 0:17:49.200 field and with a lot of I think very well 0:17:49.280 --> 0:17:53.080 fought through regulation to protect the more vulnerable in society 0:17:53.640 --> 0:17:56.920 from the overarching power of these two Internet giants. That's 0:17:56.920 --> 0:18:00.560 one issue that's impacted their earnings outlook and clearly their 0:18:00.600 --> 0:18:04.280 share price. But on the AI side, I think whilst 0:18:04.400 --> 0:18:08.440 China has some very interesting technology, on the software side, 0:18:08.600 --> 0:18:11.560 there are two big differences with the US market. The 0:18:11.600 --> 0:18:15.159 first is it's a far more competitive space than in 0:18:15.200 --> 0:18:19.440 the US. So for example, when Chat GPT was announced 0:18:19.560 --> 0:18:22.960 roughly a year ago, you hadn't heard of any Chinese 0:18:23.080 --> 0:18:25.040 large language models. At that point there were probably were 0:18:25.040 --> 0:18:28.320 a few in development behind closed doors. As of October 0:18:28.400 --> 0:18:31.439 last year, there are two hundred and thirty eight. So 0:18:31.560 --> 0:18:34.920 this is an immensely competitive field in China, and these 0:18:35.000 --> 0:18:37.760 companies have come from a relative standing start to develop 0:18:38.000 --> 0:18:40.439 a large number of models, which also tells you that 0:18:40.480 --> 0:18:44.240 the relative technical barriers to entry are quite low, certainly 0:18:44.280 --> 0:18:47.920 within China. So again that has implications for the longer 0:18:48.000 --> 0:18:51.439 term growth potential and margin potential of these companies, and 0:18:51.480 --> 0:18:54.000 again I think that's another thing that's factored in. But 0:18:54.080 --> 0:18:56.480 I think the most important thing which is going to 0:18:56.520 --> 0:19:01.040 increasingly forestall their growth curve is the access to the 0:19:01.080 --> 0:19:04.000 core infrastructure in which these large language models are both 0:19:04.040 --> 0:19:07.679 trained and then the inferencing happens, and that is largely 0:19:07.920 --> 0:19:12.600 US derived technology. So whilst the chips may be fabricated 0:19:12.880 --> 0:19:16.840 in Taiwan, they are in Nvidia design chips or AMD 0:19:16.960 --> 0:19:20.840 design chips, and due to the export restrictions, it's increasingly 0:19:20.880 --> 0:19:24.160 difficult for Chinese companies to get hold of those, which 0:19:24.240 --> 0:19:27.760 makes them evra reliant on lower generation chips which are 0:19:27.880 --> 0:19:31.840 slower or more energy intensive, you know, their inferior in 0:19:31.880 --> 0:19:36.359 many respects, or it makes them more reliant on homegrown alternatives, 0:19:36.359 --> 0:19:39.600 which again are one to two generations behind, and they're 0:19:39.640 --> 0:19:42.280 also in short supply. So if you can't get the 0:19:42.280 --> 0:19:45.160 infrastructure to run these models off, then clearly that will 0:19:45.160 --> 0:19:48.199 impact your growth curve over the next few years. So 0:19:48.600 --> 0:19:50.679 these are some of the factors that have impacted the 0:19:50.680 --> 0:19:53.840 performance of these stocks. It's more complex than that, but 0:19:53.960 --> 0:19:56.720 maybe you can do another podcast and that another day. 0:19:58.240 --> 0:20:00.440 Robert, Have you even checked out any of them The 0:20:00.480 --> 0:20:06.480 mainland China apps relating to AI of course, yeah, and 0:20:06.520 --> 0:20:09.240 how do they compared to chapt This. 0:20:09.320 --> 0:20:14.119 Is purely anecdotal, But myself and a colleague this is 0:20:14.840 --> 0:20:17.960 into last year when ernie bot, which is Baidu's Chat 0:20:18.000 --> 0:20:22.119 GPT like janitor of a tool, was first launched. We 0:20:22.280 --> 0:20:24.960 compared that and you know, just to put a few 0:20:25.040 --> 0:20:29.320 questions running through GPT three and a half and ernie bot, 0:20:29.920 --> 0:20:33.399 and I think by dou is perfectly adept at answering 0:20:33.480 --> 0:20:38.320 any questions relating to China, to Chinese history, to Chinese culture. 0:20:39.680 --> 0:20:42.960 No major issues however, when it comes to answering questions 0:20:43.000 --> 0:20:45.840 on things that lie outside China. So for example, we 0:20:45.880 --> 0:20:48.320 asked it some questions on Bill Gates and just tell 0:20:48.400 --> 0:20:50.520 us a bit about him. You know, it gave two 0:20:50.560 --> 0:20:54.080 three lines, very simplistic lines without a lot of detail, 0:20:54.440 --> 0:20:58.639 comparing contrast to Chat GPT, giving significantly more information with 0:20:58.800 --> 0:21:02.480 a lot more opinion and a lot more depth and value. 0:21:02.520 --> 0:21:05.600 Add Now, why is that? There are two reasons? Again? 0:21:05.720 --> 0:21:08.720 I think one is the access to training data. Something 0:21:08.760 --> 0:21:11.720 like ninety percent of the world's worldwide Web content is 0:21:11.760 --> 0:21:15.000 in English and roughly ten percent is in Chinese, and 0:21:15.080 --> 0:21:17.840 I think the availability of training data is key to 0:21:17.880 --> 0:21:21.879 the effectiveness of these models. So the breadth and depth 0:21:21.920 --> 0:21:26.960 of training data available to Bido's early bot is insignificant 0:21:27.240 --> 0:21:30.960 compared to chat GPT, and that is reflected in their 0:21:30.960 --> 0:21:34.560 relative performance. And obviously, the Chinese data that anybot is 0:21:34.600 --> 0:21:37.560 trained on is largely China specific, which accounts for its 0:21:37.600 --> 0:21:42.240 superior results. The second issue, just briefly is obviously on 0:21:42.280 --> 0:21:45.840 the censorship rules. So obviously, if you asking anybot on 0:21:45.880 --> 0:21:49.600 anything that's politically sensitive, you're going to get a curtailed answer. 0:21:49.760 --> 0:21:51.280 Let's just leave that answer at that. 0:21:52.760 --> 0:21:56.359 Mandy, look into your crystal ball. Two years, five years, 0:21:56.359 --> 0:21:59.760 maybe ten years down the road. Where is this taking us? 0:21:59.800 --> 0:22:02.440 Where is AI going to lead us? Is it going 0:22:02.440 --> 0:22:04.720 to be to a good place or a not so 0:22:04.840 --> 0:22:05.439 good place? 0:22:06.119 --> 0:22:09.720 I think the answer there is simple. It definitely will 0:22:09.800 --> 0:22:12.600 lead us to a much better place in terms of 0:22:12.800 --> 0:22:18.480 the efficiencies and the productivity that we think generative AI promises. 0:22:19.200 --> 0:22:22.919 And the reason I feel fairly confident on that front 0:22:23.200 --> 0:22:25.760 is we have already seen the proof points when it 0:22:25.800 --> 0:22:30.840 comes to all the contextual knowledge that's embedded in these models. 0:22:30.960 --> 0:22:35.919 I mean, you're talking about intelligence systems that can understand 0:22:36.160 --> 0:22:40.119 human language and really never get tired, and you know 0:22:40.359 --> 0:22:43.000 it can respond in a very intelligent way. So when 0:22:43.040 --> 0:22:46.280 you apply that framework to different facets of our life, 0:22:46.800 --> 0:22:48.920 the world will be a much better place. The only 0:22:48.960 --> 0:22:52.560 thing I would put as a caveat is the guardrails 0:22:52.600 --> 0:22:56.560 need to be in place, because this can potentially be misused. 0:22:56.600 --> 0:22:59.639 All these foundational models are trained on a lot of 0:22:59.760 --> 0:23:03.320 bad things as well, which are embedded in the model. 0:23:03.480 --> 0:23:05.800 That's why the guardrails are very important. 0:23:06.960 --> 0:23:09.960 Rob Lee closing thoughts, Okay. 0:23:09.680 --> 0:23:11.800 Yeah, I don't disagree with any of that. I think 0:23:11.880 --> 0:23:14.680 if you look at who are the interested parties within 0:23:14.760 --> 0:23:19.000 the world of AI, there are the makers of the infrastructure, 0:23:19.280 --> 0:23:24.280 so clearly that's the in videos, TSMC's armholdings, etc. And 0:23:24.320 --> 0:23:26.480 they're clearly profiting from it at the moment as the 0:23:26.520 --> 0:23:29.040 infrastructure is built out. There are then as sort of 0:23:29.040 --> 0:23:31.639 mandeeps referred to the end users of AI, and I 0:23:31.680 --> 0:23:35.119 think there have been plenty of examples where the utilizing 0:23:35.160 --> 0:23:38.840 AI can drive significant cost and productivity savings and we're 0:23:38.840 --> 0:23:41.520 only on the cusp of that. So the potential benefits 0:23:41.840 --> 0:23:44.720 in the long term through the use of AI, you know, 0:23:44.800 --> 0:23:47.720 potentially immense. Then there are the software companies in the 0:23:47.720 --> 0:23:50.399 middle again, and I'm referring to the Chinese companies here. 0:23:50.760 --> 0:23:54.480 On one hand, they are having to spend significant amounts 0:23:54.520 --> 0:23:57.040 on building out the infrastructure, so it's costing them tens 0:23:57.040 --> 0:23:59.640 of billions of US dollars at a time when they're 0:23:59.640 --> 0:24:04.280