Asia Centric by Bloomberg IntelligenceFinancial markets have been captivated by AI's opportunities since the launch of ChatGPT two years ago. Huge investments have flowed into established chip companies and the hyperscalers that make the infrastructure supporting AI – from Nvidia and TSMC to Alphabet, Microsoft and Amazon. But what about the frontier companies that could flourish as AI enters a new era? And why is Asia an ideal petri dish for this activity?
Esther Wong, founder of 3C AGI Partners, one of Asia's first AI-focused venture capital funds, and formerly a managing director at SenseTime, discusses the next phase of this technology and shares her outlook on the industry – data centers in space may not be as far-fetched as you think. She joins John Lee and Katia Dmitrieva on the Asia Centric podcast.
You're listening to Asia Centric from Bloomberg Intelligence, the podcast that explores the big ideas and trends moving money across the region. I'm Katidmitriva in Hong Kong, and.
I'm John Lee also in Hong Kong. Katya, The investment world has been captivated by artificial intelligence of the last two years.
Yeah, it was probably one of the biggest themes in twenty twenty four.
But if you look at where all the investment flows have gone, they've gone to the infrastructure backing AI. Now, I'm talking about the companies that design and manufacture the chips like Nvidia, Samson Electronics and of course TSMC, to the hyperscalis that own the data centers. But what about the real life, on the ground use cases of AI.
Yeah, that's a good question, especially at the start of twenty twenty five. You know, we've got to shift in geopolitics, some concerns maybe about AI's future. But our guests today can help us answer a lot of these questions. Joining us today is Esther Wang, founder of three c AGI Partners that's one of Asia's first AI focused venture capital funds. She's also one of the founding members of Sense Time that's one of the largest publicly traded AI companies that's listed on the Hong Kong Exchange and where she raised six billion dollars during her time there. Welcome Esther, thank you.
Thank you for having me.
Esther, you started one of Asia's first AI focused VC funds. Can you just give us a snapshot of you know, VC investing in Asia right now, because it seems like there's quite a few headwinds.
So yeah, this is a very good question because when we think about VC, what does VC do? So the traditional VC in the good old days, the strategy is more like we call it spread and prey. We invest in a lot of companies, we sprinkle, say like one hundred companies or one that two may come up to be the Nicks big thing. Right. That is the traditional mindset of how people do VC. In my case, it's a slightly different because AI is in such a critical and interesting juncture. I believe that we're slowly merging from AI one point zero, where AI is mostly a very on paper that is applicable to industrial users, only to now two years ago we enter the era of AAR two point zero that everybody, including yourself, I'm sure be using some sort of a on a daily basis, right that the chat gipt moments, so to speak, and drawing from experience in investment during the al one pointo era, which is almost years ago, I feel that the best way to do AI and the best way to capture this two point zero wave is to do incubation. And then when I think of incubation, I don't think of incubating a lot of small companies and sprinkle smaller tracks two one hundred names. Right, I think about identifying the right sector, a very specific sector or maybe very specific scientists I want to beat and I want to almost think about it as a co found the company with the founders that I like. So in this case I can drawn on all my previous experience and make sure that my venture the company is successful in the long run.
Can you tell us about some of those like some of the companies that you're looking at right now and why you feel that they're kind of undiscovered or you know why you believe in those companies?
Sure? So going back to AI, right, yes, is such a big few When you think about AI investment, anything and everything in the sun can be AI. So I would like to divide AI into four big categories. Starting from the bottom, we have the data, so basically what the ingredients of begging AI. So you have companies such as Scale AI that is you know, fourteen fifteen billion dollar companies already. And then a second layer is so called the infrastructure, the hardware layer. So that's why you have all the NVIDIAs of the world, and we believe that there are other alternatives to Nvidia that can emerging from this layer, which is quite interesting, so we're looking at that now. The third layer is the I would say, the operating system layer, so you have the finished product of those bottom two layer is things such as CHATGBT or mitro or co pilot. And then you have the last layer, which is the application layer. Application layer you have further divided into two B or T C. Some of the things that we're seeing these days, for instance, a time is driving or robotics. Right, these are the things that people use on a daily basis. So these are the four layers that we categorize AI in and then for my specific venture, I focus a lot on the lower two okay, the bottom two layer, and the reason is because again in my previous time, we actually built one of at the time Asia's largest AI data center. Okay, that's in Shanghai. So at the time nine hindsight's so small, but at the time it was really big. It was five x flop. So now the world is building hundred x of flop data centers. But just to give you a perspective, five x flop. At the time when we build a dojo Ilamas Dojo building the data center in the Tesla, it was about one point seven x of FLOP, so it was actually almost twice the size of dojo. So you can imagine we are quite advanced in that sense, right.
Just very international or like non tech savvy listeners. Rather, what is that sizing that you're referring to, that term in terms of size.
You can think, okay, so if you put in perspective, five x a flop is enough compute power that you can analyze every single video ever generated, since there is such a thing as video that was up to five years ago. That is the scale of the size. But of course, as you know, in the last few years, we have experience an explosion of data in terms of structural data and also in terms of unstructured data thanks to AI, so that's why now in hand sight five is not very such big of a number anymore. Just to give you some ideas, in China, when you think about AI data center, the air data center can consist of any clusters of GPU from let's say a few hundred to maybe a few thousand, right, or a large cluster is let's say ten thousand cluster together. GPUs can see a large cluster. But the world's largest cluster right now is Lama's Colossus, and they have one hundred thousand pieces together in single cluster. And then he's saying Elamus is saying. Google is saying the next cluster would be more than four hundred, five hundred thousand pieces of single cluster. So you can see that in this field, the extent of growth is truly astounding. But going back to the investment, I believe that there is a lot of ways to actually improve upon how AI is made. So the process of making AI itself, there's a lot of improvement, and that's why my fund actually focuses a lot on how to improve upon making of AI. I'll give you one example, right, So, going back to the data center in Shanghai, when we build it at the time that was almost four or five years ago. It already accounted for almost six percent of the entire Shanghai cities power consumption. That's a lot.
That is a lot.
Yeah, people are not aware that how exhaustive the building of AI itself is. Right, nothing about this number. Now extrapolly this to nowadays we fast forward five years today, building a AA data center is no longer a new concept. As you say, John, a lot of people is pouring money into this. So right now almost six percent of world's energy every year is on AI data center. That is a crazy number. Because Jenny I or AI two point zeroso this week literally started two three years ago.
It's gonna get worse, right absolutely.
And think about who are the user co U quote of this computer power? Who are the users of AA data center? Most of the AI data center right now is users for training training of models, right that the end product is models such as chachy b differences. So in the world there are like ten twenty companies that does that. So in China you have to buy do the by dance. And then in the US you have the Googles, you have the Oracles, you have the Opening of the world. But not that many companies that actually specialize in training AI, and yet AI is almost six percent of the world's power consumption. Nothing about the next phase of AI. The next phase of AI is not on training anymore. The next phase of AI is mostly uninferencing. And what I mean by that is inferencing is basically using of AI. So every time when you ask Chichipedia question, it gives you solid different answer, right because at the edge of the cloud, it is generating on a spot, inferencing an answer on the spot. And so imagine if all of us on the planet Earth, eight billion people is doing let's say, one hundred ten thousand infants every day, can you imagine the power consumption.
So are you saying that there's kind of companies you think or that you're looking at right now that could kind of reduce that power consumption.
Yeah, absolutely right, because the reason that it's so power consumptive, right, So, in the word of Jensen, the influencing power demand is that the demand for inferencing is literally one billion times higher than the demand for training. So obviously not sustainable, right, I mean, I've just read the will run out of power. So there are ways to reduce that, or there are ways to make the making of AI more sustainable. That's what my fund is all about. How do you make from a fundamental basis to make the making of AI more sustainable? And there are four ways to think about that. So number one AI two point zero is generated from this amazing paper that I'm sure you guys are heard of. Attentions all you need underlying the transformer architecture. So transformer architecture itself takes a lot of tokens encoding and decoding process. It takes a lot of data in and data out. So firstly we think about, are there any ways to refine the transformer algorithm itself so that it takes less data? Right? That's number one. Number two, if you look at any AI DC right now, a data center right now, the utilization rate for GPU is around maybe fifty percent, right, it's not even that high, right, And because they always have the build for peak capacity. So can we actually change that fifty percent to let's say fifty five, You know, you save a lot of trees that we're already. So that is another way to say, hardware wise, as we change from training to inferencing, is GPU the best solution to inference because think about it as a car. Training is like a truck. You need to carry a lot of stuff. You don't need to go super fast because the end consumption is scientists. They want to build an LLM. But inferance, you have to think about it like a Lamborghini. People do be very very fast. You don't have to carry a lot of stuff, but you need to be super fast. So are there alternative hardware that can do that? So we actually look into that. And also the third way to think about it is can the data center itself be improved? For instance, you know, fifty to eighty percent of the power. When we see this isn't center consumer a lot of power? What is it used for? It's actually used for cooling.
Yeah, that's right. We had kkrning office a couple of months ago and that was one of the things that they had mentioned, is that this is one of the biggest costs.
Absolutely, so can we make cooling more efficient? Let's say, for instance, can we have underwater data center because we now have three globally, can we have more? Or can we actually put the data center in space? I mean think about it, right, space is minus two undred and fifty degree celsius, so there's no need for cooling and there's actually unlimited amount of solar energy. The third way to think about is can we make the making of data center itself more fundamentally change the way we make data center? And then the fourth and find a way is how do you make the energy more sustainable or more renewable? And in that way we believe that there's going to be potential a lot of breakthrough, maybe in a nuclear because amongst all the alternative energy, nuclear actually has one of the highest EL ratio. So that's the four ways of how do you make the making of AI more sustainable so that it can enhance all the application to us? Because if we don't get that right, I mean, how can we think about the application layer? Is like if we cannot get the steam engine right, how can we think about, you know, having vacuum machines. So we need to make sure that the fundamental lay is right first.
Is there a lot of potential companies in Asia that are on this technological path and it sounds like a lot of money, sounds like a lot of capex.
So interestingly, a lot of the money that put in AI actually as reported by Bloomberg in Asia this year has been on the building of AI data center. It's the construction of data center, so buying of the land, building of the building, making sure all the infrastructures in place. So I think as the first half of this year is more than thirty billion even remember correctly, of course the world needs more AI data center. Yes, so in Asia a lot of the energy is based on that. But as an investment focused we don't focus on that because that takes a lot of money buying land itself. You know, it's it's.
It's it's a lot of money.
Yeah, So it's not it's not our game. I think, you know, bigger names like er I think Blessed actually just had the fun just to focus on that. So that's that's more like what they do. But what my fun out, what we want to focus is that what the frontier technology can actually falls into one of those four categories, and that in the SOULF doesn't take a lot of initial capital.
Okay.
So and then when you think about the talents in Asia, Asia versus US, and we have some of the best schools and best talents in Asia talent wise, I don't think that Asia is in any disadvantage compared to say, you know, the West.
In that regard, are there certain countries or regions or maybe even just cities that are doing it really well right now that are as you said, these sort of good examples of incubators.
So because my friend doesn't look at China, so I would say outside of China, when you think about AI infrastructure, you see that in Southeast Asia in particular, this has been a particularly active year. So Singapore has been a hub for a lot of these innovation and tech research, and there's a lot of scientists and startup teams out of Singapore. But when it comes to the application of these technologies that neighboring countries where they can actually build at the center, then you can apply some of those technologies too. So we're seeing applications in I think in particular Malaysia because I think in Juho there's a lot of though that't has been built, and also Indonesia. So these are the two countries with relatively chiap land and also relatively cheap natural resources, and so new technology can be a testing ground there. Right. So that's two that I can think of, and those are the investment by you know global investors. When I say global, I mean Microsoft just announced they're going to invest went from something billion dollars in Indonesia. Oracle is going to do the same. So you see that all the hyperscalers when they think about AI investment, they also invest in on a global basis, and soviacet Asia has been a very recent beneficiary since last year.
I will say Asia Centric is produced by Bloomberg Intelligence, where more than five hundred experienced analysts and strategists work around the clock to bring you timely, world class research. Our coverage spans two hundred market industries, currencies, commodities and industries, as well as over two thousand equities and credits. If you like what you hear, don't forget to subscribe and chairm. You know, Southeast Asia is quite diverse. There's a lot of different cultures also, and they don't all speak the same language. Does that pose a difficulty in investing in that region?
I think the difficulty is more like a two C level. You're absolutely right because when you look at historically, people cannot think about Serves Asia as one single country because they have very different political systems and they have very different consumer systems. So if you think about one big let's say company sharing economy, when they say the hsuth as Asia, it's extremely hard for them to get profitable prisons nicely because of the problem that you mentioned. Now, the problem is slightly alleviated when it comes to be like non consumer businesses because those are the ones that you can sort of get a relatively concentrated end user, So it can somewhat be alleviated. And I believe that even Service Asia is in this interesting great resett because AI is a great reset like nobody's with maybe China US right, No one is really late to the game because literally AA Tanos started two years ago. So it's a potential opportunity for countries around the world, of course, including Service Asia, to adopt and embrace this new air two pano by making sure the infrastructure is ready. And you're seeing that a lot of the money in you know, Malaysia, in Vietnam and in Indonesia actually across Service Asia, they're doing that also in Japan and Australia as well. I mean, the whole world's embracing that.
Well, you mentioned it a few times, so the China question, your fund does not invest in China. Why is that so?
My fund has a mandate of eight years. The focus is on long term investment and we need to have a very clear rules of engagement in terms of the area or the region that we invest in. Now, when you take a look at China in the last few years, you see one trend which is quite clear. That is that a lot of the US dollar fund has been retrieving and replacing. That you see R and B Fund dominating. And so when you look at the new formation of the R and B Fund, a lot of them expect by the local government because the government has a mandate to invest in these new technologies including AI and EV and whatnot. When you talk to the local government, these funds apart from financial return, which is also very important because they have a government nature, the mandate is not just on financial return. They have other matrix. For instance, they want to make sure that the relocation of a headquarter to the specific region, or they have some KPI for headcounts for specific kind of talents text requirement. So when you think about if the market is dominated by fund that is not purely financially returnd driven, then the dynamic of valuation would be quite different exit path is also going to be quite different. So that's why we focus on what we know best. So that's why my friend does not really put a lot of energy in China. But if there is an exceptional opportunity, well of course we still look at it. But the attention is not generally in China.
I guess if you're not in China, you can avoid a lot of the geopolitical risks, right. You know, there's this trade war that's basically been going on since twenty eighteen, twenty nineteen, sanctions on certain companies, certain products, tariffs. So how big do you think these geopolitical risks are kind of heading into twenty twenty five for like VC investors and particularly in AI and technology.
I mean thing about the big picture. Right, we grew up I would hope that we have very similar generation. So we grew up in the era of globalization. We grew up believing the world is getting flattered, the world's getting smaller, and that has been the world order for the last you know, twenty thirty years. And I think as an investor, whether you're a vcope or individual, right you sort of have to embrace the fact that perhaps for the next decade or so or even more. The war is going into a deglobalization trend. Whether you like it or not, it is. I mean, someone's telling me that twenty twenty four is the first year in modern democracy history that every single incumbent in a democratic country has been toppled this first time in modern history, and it's tend to be replaced by populists or nationalists, right, So that is actually a very interesting observation. I believe that the treind is here to stay. So as an investor, you have to be very connaisant by the fact that whatever you do, you have to be here to the local rules and regulations so that you don't really step on anyone's toest so to speak. That would apply for AI, that will apply for a lot of the high tech investment, and that will apply for just about any investment.
Yeah, do you get to avoid some of that at least by not investing in China or do you kind of face those risks as well?
To be fair, right, for the fund to not invest in China is more like a financial decision. But of course by focusing on one area, I mean, nowadays you don't see a lot of fund that focused both in China anglably. Usually when lp thinking about where to allocate the capitol, there's a mandate on whether you want to include certain countries or not, right yeah.
And some American university endowments pension funds cannot actually invest in China AI technology, right yeah.
Of course. So as a result, because of these supply and demand, naturally there'll be funds that separate themselves into different buckets. And you notice a lot of these China funds, they tend to the R and B more and more so R and B denominated.
I'd love to sort of take a step back and ask you another question. Now, a lot of our listeners in the public markets and AI has been a huge you know, there's been a lot of excitement. Some people call it the bubble, but there's been a lot of excitement. Now people are starting to get a little bit impatient. But when are we going to start seeing AI being commercialized? Like where are the killer apps? And what do you say about that?
This is such a great question because we know that it's just in any technology. Initially, there's always gonna be over excitement and there's a period of coming down and eventually you hit the equilibrium, right. We saw that with the last era of mobile Internet. So a lot of these new Internet companies that came out, some of them we no longer see. Doesn't mean they're not good companies such as you know, Nescape, Excite, you know, there's a lot of names belong to that category, and they all disappeared during the Great Bubble, but the impact of them linger and that's how eventually companies such as Google emerged as a result. So I think genuinely the same thing would apply to AI is that right now AI is trying to still find it's footing in terms of what would be the killer application. I believe that the reason that they have not emerged as of now is because number one, it's just too fast. The whole AI thing happened in less than two years, so people still trying to figure out how to use AI. In general, I would say amount eight billion population in the world, or maybe one point two billion population in the developed world. Even up to now, probably less than half actually uses AI on a regular basis. So it's just the timing is too show, so people need time to figure out. And secondly, the course of production of AI is still relatively high, say my daughter, she loves to use AI to make her own post cards and stuff. But if you want to subscribe to let's say make Journey or run Way or you know, Canva, it's like twenty years dollars a month. At the end of the day, these little costs just adds. A right, Kin Academy is another twenty years dolars a month, as you know, as all parents know. So all these things add up eventually. But can you imagine if the cause of marginal AI production i e. The couse of marginal intelligence is lower, then you can actually drive up significant adaptation. Just like if electricity is only available at let's say one hundred thousand dollars per ward, which was equivalent the case when it first came out, versus now it's down to you know, thirteen cents in Hong Kong, a lot of new application can come out. But we are not at that point yet. We're still at the point where electricity is worth one hundred thousand dollars per the equivalent of AI. So that's why for the first time here I was mentioning we focus a lot on how to make the marginal cost of intelligence lower so that AI can genuinely truly proliferate. And also when you think about the sectors that can most benefit from AI, I personally believe that there are two sectors that can make a huge impact from AI and the way to think bodies. What are the sectors that imagine if if you have access to all the world's knowledge, will make the most impact, Like what's the biggest delta?
Right?
And to me that too, one is education and one is healthcare, medical drug discovery. Right, those two would change the world because with AI, essentially my daughter and say your girl versus somebody from India, from Sudan, from London, from everywhere in or they can have the same access to knowledge and so you unlock intelligence, you put everybody on the same level, and that would be amazing, right, So that can also in turn unlock a lot of other potentials because human brain is one of the greatest things to be unlocked. So that is one. And number two is why do I believe healthcare Because I mean, this year's Nobel Price winner goes to the chemistry goes to protein folding, where there are two groups that they use different approaches, but we believe that with enough these AI tri and eraror the feedback loop. As long as becomes faster, personalized drug can be a reality. Beating cancer can a reality. I am on the cam that I believe that within a lifetime we can actually find a cure at this current rate, because only another ten twenties to try a drug. You can just use a few.
Minutes, Yeah, you know, the healthcare space. Though it also highlights a lot of ethical issues with using AI. You know, one of the last stories that I worked on before moving to Hong Kong was about AI that was kind of this one hospital North Dakota was actually building an AI room in their hospital, and there were kind of a lot of issues with that, like in terms of monitoring, in terms of like how the information is stored and shared. So I wonder how you navigate that as an investor, Like when you're looking at a company, you know, for example, facial recognition technology and how that could be used in negative ways, how do you navigate that as an investor?
If you look at the history of all technology innovation, there's always two sized to his stories. So every technology innovation can be used for good and can be use for bad. That does not mean that we should just stop pursuing it. But it just means that we need have very clear rules of engagement. Right. So I believe that for the example that you use, for instance, a patient's data, who should have access to the patient's data the patients who is producing the data himself or herself should have very clear understanding who is using it. Right, cannot be just used without his or her knowledge, or even with the knowledge of using it, can I monetize from it? Even so, I believe that people need to take agency of their own data in the age of AI. I think this is such an interesting and very important point, and that actually leads to a bigger point. I think AI can potentially have a lot of benefit of course for education, healthcare, it cannot also have a lot of potential risks as well. So it's just one of those things that the government is not working fast enough in a way to produce the rules to regulate AI. And I can understand why because again it happened in the last two years. I mean, sometimes it takes a government two years to convene a meeting. And this is not just in Asia or US, It's just a global issue. It just happens too fast. So I think that this is actually A very important thing is it needs more regulation. And now the regulation is being set by a few hyperscalars right rather in the yeast of the West. It's the same thing sept by a few hyperscalers. And I believe that it needs a lot more involvement by third parties, angios, people that actually uses it, civilians and also the government. So we need that dialogue, and that dialogue cannot be just within the country. It should be a global dialogue.
Esther, before we let you go, I wanted to ask, now, before you became this you know, venture capital Titan in Hong Kong, you were actually working for an investment bank. You were in institutional equity sales, you were colleagues. Yes, we just found out and to be afraid of I'm getting very insecure and depressed because I'm still in finance and now you're this Titan investing. But look, it must have been such a big jump to leave that cushy job to join sense Time. What went through your mind and why did you make that decision?
So I was in banking for twenty years before I started the startup, which was sense Time. I was in San Sian for seven years. I remember I was in my last role. I was actually almost forty years old and I was actually pregnant at time.
Wow.
So that propelled me to have a sense of urgency because I realized that the world is changing rapidly in front of my eyes. Either I can be the changer or I can be the change gee. And since I know that I'm going to become a parent, I want to be part of the change. I want to be the changer and not the change ee. So I always love reading papers. By the way, this is just my curiosity. I always loved reading research papers, and at the time I encounter one of them, it was the first time in the human history that a computer can do things better than human and that is in terms of a recognition of the face right. And so I look at the offer of the of the paper. I was expecting m I t or Stanford professors, and to my very pleasant surprise, it was a professor all from Chinese University of Hong Kong. So that's how I got in touch with Professor Tang and his team. At the time it was fifty one hundred people, and I felt so compelled that form the next generation. I want to be the changer so I took that leap of faith. I took a literally eighty percent pay cut.
Well okay, and I thought that, And you're pregnant at the same time as well.
Yes, I was pregnant, so I was flying literally three times a week when I was nursing. But I just felt like, in life, right, you're gonna look back and say, what if? Right, the what if I are the one that kills you? So what if I don't join? Would I regret more so if the answers yes, And yeah, just just take a leap. And that's what I did.
Okay, great story.
Yeah, that's a perfect, perfect ending. Thank you for joining us.
Okay, thank you, thanks for having me.
Thank you.
You've been listening to Asia Centric from Bloomberg Intelligence. My name is Katya Dmitrieva. You can find me at Katya d m I on x or Katyamitrieva on LinkedIn.
And this is John Lee and you can find me on LinkedIn by searching John Lee at Bloomberg.
Clara Chan produced this podcast, and you can find us on Apple Podcasts, Spotify, or wherever you like to listen. See you next time.
