Transcript
Speaker: Welcome to HSBC Global Viewpoint, the podcast series that brings together business leaders and industry experts to explore the latest global insights, trends, and opportunities.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
Speaker: Thanks for listening.
Speaker: And now onto today's show.
Speaker: This is a podcast from HSBC Global Research, available on Apple Podcasts and Spotify.
Speaker: However you're listening, analyst notifications, disclosures and disclaimers must be viewed on the link attached to your media player.
Speaker: Welcome to Under the Banyan Tree, where we put Asian markets and economics in context.
Speaker: I'm Harold van der Linde, Head of Asian Equity Strategy here at HSBC.
Speaker: On today's podcast, we're going to bring you up to date with where we are in terms of the artificial intelligence AI rollout.
Speaker: But more importantly, we're going to focus on China's role within it.
Speaker: How do Chinese tech companies plan to make money from AI?
Speaker: And are they really in a position to rival US firms in this space?
Speaker: Joining me to help answer those questions are Frank Li, Head of Asia Technology Research, who joins us from Taiwan, and Shaolin Liu, who leads our China Internet Research Team from Singapore.
Speaker: Let's get the conversation started right here, under the banyan tree.
Speaker: Frank, Charlene, thanks for coming on to the podcast.
Speaker: Thanks, Harold.
Speaker: Thanks for having us.
Speaker: Frank, we want to get a good understanding of what is happening in AI and in particular, how the Chinese are stacking up versus the international place.
Speaker: Now, before we answer that question, I really want to take a step back.
Speaker: In the past, you spoke about that you see the development of the AI industry in three different stages.
Speaker: Can you remind me a little bit what those stages were and where we are?
Speaker: Sure, Harold.
Speaker: Right now, I would say, you know, everyone's investing in an AI, especially all the cloud providers.
Speaker: So what we kind of think right now is that this year is really the first stage.
Speaker: And the first stage is really about investing in the large language models and how you train them.
Speaker: And what's driving this training of the large-angle models is basically the usage of what they call GPUs.
Speaker: GPUs are graphic processing chips, right?
Speaker: So, for example, in a video game, you see something on your screen, a car that's driving or something like that.
Speaker: That's something that's processed by those particular kind of chips, right?
Speaker: Yeah, that's right, Harold.
Speaker: So actually, if you think about the GPU, it started off really being used heavily by gamers because the GPU chip allowed you to enhance and improve your gaming functionality.
Speaker: But then what they've discovered is that, you know, this GPU, because it's very high performance computing,
Speaker: has been used now for data centers, and it's also very functional to be used in AI servers, given the computing power that you need.
Speaker: And that explains why, for example, originally video gaming companies such as Nvidia is so prominent in making these chips, right?
Speaker: Yeah, I mean, I think right now, as far as stage one, NVIDIA in particular is almost a monopoly in terms of its dominance this year because it's really the sole provider of these GPUs and AI services this year.
Speaker: So the stage one is really, you're building out the chips and you're feeding the data into the service so that AKIs can use it.
Speaker: What is stage two and three then?
Speaker: Yeah, so I think next year is going to be interesting because you might probably start to see the beginnings of stage two, which we think is going to be more about an inference stage.
Speaker: So if you think about it, this year, everyone is building up their large line of models and needs to be trained.
Speaker: Next year, you're going to start putting in the inference capability because ultimately to use AI, you have to be able to infer from it.
Speaker: Right.
Speaker: And when it comes to inference, there's going to be potentially more possibilities.
Speaker: You know, the chip that's being potentially going to be used in inference is called an application specific IC, an ASIC chip.
Speaker: So an ASIC chip is something that is really specific to doing a particular action, right?
Speaker: So if I need AI to make me, for example, maps in a map making program or so, there's a particular chip that allows these processes to run in it.
Speaker: And that's what you're referring to, right?
Speaker: Yeah, I mean, typically ASIC chips are not as complex as the GPU and they have very specific functions.
Speaker: And that's also why if you look at the cloud providers, especially a lot of the US cloud providers, they are working on developing their own chip in-house as well.
Speaker: So while there is nothing they can do about the first stage, which is about the training, very dependent on NVIDIA, there is increasingly a push now to try and see if they can develop some of these ASIC chips in-house for their own use as you go into stage two.
Speaker: And the third stage, the last stage, what is that?
Speaker: Yeah, the third stage, I call this the blue sky scenario, right?
Speaker: And the reason why I call it blue sky scenario is that while stage one and stage two are very positive for the AI supply chain, it doesn't really help the overall hardware supply chain that much because the demand in terms of AI servers is
Speaker: in terms of number of units, while it takes up a lot of value and it's very expensive, the volume isn't that big.
Speaker: I mean, ultimately, you know, the smartphone PCs are much bigger markets in terms of units.
Speaker: And as we all know, smartphones and PCs have been challenging, but this is where the edge AI comes in.
Speaker: Edge AI, the idea is basically to provide AI functionality directly on your local device.
Speaker: So imagine if you will, that you can actually do AI functionality on smaller scale language models directly on your phone without the use of the cloud.
Speaker: So that ultimately could drive, I think, you know, a big replacement market again in smartphones and PCs.
Speaker: Okay, so I want to move to Charlene now, but just to quickly recap, basically, we now need chips that just can feed a lot of information and gather it and store it.
Speaker: We're going to a second stage whereby using that data, using specialty chips, you could say, to run all sorts of different applications of it.
Speaker: And later on, it's really, you're getting it onto a completely new iPhone or mobile phone.
Speaker: That's a later stage.
Speaker: That's broadly speaking, the kind of development that you're looking forward, right?
Speaker: Yes, that's right.
Speaker: Now, Charlene, you're looking at some of the Internet companies in China and clearly they've come out and have all sorts of monetization strategies.
Speaker: They want to make money out of this, right?
Speaker: Broadly speaking, what is the approach that these companies take in China?
Speaker: Sure.
Speaker: I think from our perspective, there probably comes down to three key monetization avenues for generative AI technology.
Speaker: The first one being AI computing service.
Speaker: And what it does is it supports and empowers AI applications from the large language models training to various AI embedded software services operations.
Speaker: One of such examples would be a photo editing app.
Speaker: And, you know, it has the intention to, for example, to speed up is image generation processing ability.
Speaker: And what they could do is they can then buy GPU usage from one of the cloud players, call it AliCloud.
Speaker: and pay for such services.
Speaker: Second would be model as a service.
Speaker: What it does is it enables developers and enterprises to deploy AI models.
Speaker: They can also design, tweak, and customize these models if they needed to do so.
Speaker: And obviously, internet platforms like Baidu, Alibaba, etc.
Speaker: Tencent would charge a fee for this kind of services.
Speaker: Last but not least, it's AI-generated content, AIGC applications.
Speaker: And this refers to how content generated by AI is being used.
Speaker: and help to enhance existing business to run more efficiently and obviously, you know, in turn to create more new revenue opportunities, so to speak.
Speaker: And I think one example is improving which online audience are being targeted and therefore attract more ad spending.
Speaker: Okay, so broadly speaking, either you make money out of it by saying, listen, I have the computing capabilities, you can rent that from me.
Speaker: Or companies just say, listen, we have all of this.
Speaker: Now I know exactly who and when to target to get them to buy a concert ticket or whatever that may be, right?
Speaker: So those are the models.
Speaker: Are these Chinese companies adopting all of these models or are they focusing on something else?
Speaker: And how do they differ when you look to, say, international companies?
Speaker: Sure.
Speaker: I mean, all these avenues that we laid out are ways that I think all platforms are monetizing to a certain extent.
Speaker: Obviously, some could be doing stronger in certain avenues than the others.
Speaker: But if we were to kind of stack them against the overseas players, I think one sort of obvious bottleneck for Chinese players is the gap in GPU capabilities and inventories.
Speaker: Overseas players have far bigger inventories than the Chinese counterparts.
Speaker: So if we were to kind of look at these three avenues from an AI computing power standpoint, then that would be a gap.
Speaker: And on model service, I think it's fair to say that both overseas and Chinese players offer a very similar range of services, but overseas players are one step ahead in terms of monetization.
Speaker: And in terms of AIGC applications, I would say that global peers are probably leading Chinese players in terms of addressable user base and also in terms of software as a service monetization and integration of large language models into existing products.
Speaker: And what about regulation, Charlene?
Speaker: What sort of differences do we see there between China and the AI names in the US and Europe?
Speaker: I think, you know, governments are all moving very rapidly, but at slightly different speeds.
Speaker: On the China front, China obviously had made the first announcement about the regulatory framework back in April 2023, and they started implementing by August 2023.
Speaker: And in September, they have already approved 12 large language models.
Speaker: Hmm.
Speaker: And on the Europe side, the EU AI Act has already been approved since May this year.
Speaker: And on US lawmakers' side, they're also seeking public comments around potential accountability measures for AI systems.
Speaker: Obviously, we see a few common threads here emerging around respective proposal made by Chinese and European regulators, and that's probably centered around accuracy and objectivity among the focus areas.
Speaker: Yeah, I can imagine that particular lawmakers are all over the place in terms of what kind of conclusions people can draw out of the information that AI is accessing yet.
Speaker: So broadly speaking, there is a regulatory changes.
Speaker: You say that some of the Chinese companies struggle to actually get the chips that they require now in order to store all that data and get AI running.
Speaker: And that comes back to Frank, right?
Speaker: Frank, how do the Chinese companies stack up in your hardware view on AI?
Speaker: Yeah, that's a good question, Harold.
Speaker: Especially now, as we look at the AI infrastructure and hardware development, it's still very much one-sided in the sense that most of the large Langmuid models today in China are all being driven by U.S. chips, right?
Speaker: U.S. companies providing that core chip that goes into it.
Speaker: Now, I think there's going to be more debate about what development's going to look like going forward now that they...
Speaker: I mean, what you're referring to is basically that America is trying to limit access of these chips to China, right?
Speaker: Because it wants to slow the development of the AI industry in China down and so that America as a nation, say, or Europe, stay ahead of the game.
Speaker: Yeah, so I think that's definitely a key issue.
Speaker: But aren't the Chinese making their own chips, new companies emerging that front?
Speaker: Well, the thing is, like the technology, when it comes to the process technology, the node technology that the Chinese are able to do and what they currently depend on, it's a big gap.
Speaker: So that's why I said earlier, like, you know, pretty much all the larger language models are being, all the chips that are going into it are all U.S. companies providing those chips.
Speaker: So basically what we're seeing is we're in the early stages of a build-out of AI.
Speaker: The model over the years will change away from just feeding AI with a lot of data, but towards making use of that data and finding all sorts of applications.
Speaker: The Chinese, like other markets, are trying to monetize on this in various ways, as Charlene has indicated.
Speaker: But the problem for them is that they're still very much dependent on supply of Western chips.
Speaker: And to a certain extent, that is...
Speaker: limited or maybe more difficult for them to obtain, right?
Speaker: And that slows down the development and the ability to monetize on this as well.
Speaker: Yes, I think that that's a good way of putting it, Harold.
Speaker: Okay, well, with that then, thanks, Charlene and Frank, and hope to discuss this matter at some point in time in the future again.
Speaker: Thank you again.
Speaker: Thanks, Harold.
Speaker: Well, we're going to have to wrap things up here, folks.
Speaker: Another very interesting discussion here under the banyan tree.
Speaker: And we thank you, as always, for joining us.
Speaker: Remember to check out our sister podcast, The Macro Brief and the ESG Brief, available wherever you get your podcast.
Speaker: We'll be back again, same time, next week.
Speaker: Thank you for joining us at HSBC Global Viewpoint.
Speaker: We hope you enjoyed the discussion.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
