UBS AI Podcast - CEO Series - Ep. 4 (Dr. Kai-Fu Lee, CEO of 01.AI)
The UBS AI podcast featuring Dr. Kai-Fu Lee highlights significant advancements in generative AI, suggesting that its capabilities surpass previous AI iterations. This technological leap is poised to increase productivity, especially for firms ready to leverage these advancements. Per the full note source, Dr. Lee predicts that slow adopters of AI technology risk being outpaced and ultimately eliminated from the market. Given the rapid evolution in AI and its implications on various sectors, traders should be attentive to how these developments could influence currency movements, particularly in technology-linked economies.
What the desk is arguing
The desk posits that the emergent capabilities of generative AI may lead to disruptions across sectors, impacting currency valuations tied to technology. Dr. Kai-Fu Lee emphasizes that the current wave of AI is 'infinitely more' powerful than past iterations, hinting at a fundamental shift in productivity. The implications for firms and jobs could lead to currency volatility as markets anticipate which economies will adapt successfully.
The potential for generative AI to enhance productivity across industries presents a unique opportunity, particularly for digitally savvy companies. The risk of companies failing to adapt is a critical concern, as noted by Dr. Lee, who stated that this could lead to job losses and business failures. As generative AI becomes more integral to business operations, currencies linked to dominant tech economies might see significant movements.
Where it sits in our coverage
Our consensus target for the EUR/USD is 1.075, with a range between 1.04 and 1.12. Several firms have projected their targets as follows: - jpmorgan: 1.10 (Mar26) - bofa: 1.04 (Mar26)
This view maintains alignment with the broader forecast trends, where jpmorgan is forecasting a stronger dollar as firms embrace generative AI technologies, placing us firmly within the upper range of the spread.
How other firms see it
Our analysis is aligned with firms such as jpmorgan, who see the potential for a stronger dollar as AI-driven productivity rises, while bofa expresses a more cautious view, anticipating limited movements. The ongoing developments in generative AI will likely exert influence on major currency pairs, particularly USD/EUR, closely linked to economic health and technological advancement.
What the calendar says
With no upcoming economic events in the next 30 days, the focus remains on the broader implications of AI advancements as they reshape various industries and economy-wide productivity metrics.
01Generative AI represents a significant shift in productivity capabilities.
02Slow adopters risk obsolescence, affecting job markets and industry dynamics.
03Currency movements could reflect changes in economic landscapes shaped by AI advancements.
04Traders should monitor technology-linked currency pairs for volatility.
Market implications
Traders should watch the EUR/USD as AI developments could influence economic outlooks and currency strength. A shift towards stronger productivity could push the EUR/USD toward the upper range of our forecast. Attention should also be given to reports on tech sector earnings as a signal of AI adoption rates.
Risks to this view
A slower-than-expected adoption of AI technology by firms could invalidate the bullish outlook. Additionally, unforeseen regulatory challenges or economic downturns may inhibit the expected productivity boost, prompting a reevaluation of currency forecasts.
ubs
Hello and very nice to see you, Kai-Fu. Thank you so much for making time. Very much appreciated.
Thank you for inviting me, Ulrike. So Kai-Fu, as someone who has been at the forefront of AI for more than four decades as an executive in some of the largest U.S. tech giants, Apple, Microsoft, and then Google, as an author of two bestselling books on AI, as an investor into Chinese technology companies, and now as an AI entrepreneur, I'm curious, how do you see this current wave of generative AI compared to prior AI cycles? Thank you.
I very much agree with your comment that it's completely underhyped right now. And power of generative AI is infinitely more than any AI I have ever seen. The differences are that we now have a technology that is approaching general intelligence, and that means it can reason about any domain.
It can converse with us. It can start to take over many of our tasks using agent technology. So its IQ is higher than most of us for most tasks, and soon it will do most tasks better than most of us.
And that is a tremendous boost in productivity, a tremendous opportunity for companies that have the digital transformation and assets ready to leverage this technology, tremendous for investment and entrepreneurship, and the horrible news for those who are too slow to adapt because their companies will be eliminated and their jobs may be replaced. Too slow to adapt. I think that is certainly something that does not apply to you.
It seems like with founding Zerodot AI, you are actually somewhat closing a circle for you personally. But I'm curious, what motivated you to start an AI company? Right.
So at Sinovation, which is a VC in China, we've already invested in 13 AI unicorns, although most of them are AI 1.0 or pre-generative AI. And even then, they were creating a lot of value and their value at a billion dollars or higher. But when generative AI came about, and when I saw the way with which chat GPT took over the world by storm, then I saw that this was only the beginning, that it was going from chatbots to reasoning engines, to acting agents, to innovators, and then to organization replacements.
And this is happening roughly one per year. Of course, back when I started Zerodot AI, I didn't see steps two, three, four, or five, but I saw how powerful, how much more powerful the engine in chat GPT was compared to any other AI technology I had seen. Older AI was fragile.
It was limited to one domain. It was expensive to implement. Chat GPT, as an example, was general, and it could be easily tuned to any domain.
And we could see the roadmap arriving. It was further reinforced by scaling law, which meant if you throw more computing power and more data, it gets smarter automatically. And now, as you said earlier in your talk, AI is teaching AI, so that arguably is accelerated.
So when I saw all of these things and the prospect of some of the things that hadn't happened, I thought that if I had spent my whole life in AI already, and that if I just sit this one out and act passively as an investor, I might make some money, but I would not have the impact to build the greatest technologies and products to change people's lives and truly feel the fulfillment of my life's dream. So this wasn't something I could invest in. It had to be something I would build.
And how would you describe, and fascinating to fulfill your personal dream, so what would you describe as your mode, your vector of differentiation from some of the larger foundational models that we now see proliferating? Right. The excitement and the challenge coexist in generative AI, right?
The excitement is through all the change. And the challenge is adopting one's company to these changes. So when we started out, we wanted to build our own foundation models.
And we did very good models among the top ranked in the world. But pretty soon it became clear that only a small number of companies would have the resources to continue to build these giant models. And that much more sensible from a business standpoint, that is solving real problems and making real money, is going to be in building of platforms and applications.
So it's sort of like compared with building the engine for a car or an airplane is perhaps the most important thing, but the engine doesn't fly, the airplane flies. And building of an operating systems kernel may be the most exciting and challenging, but an operating system kernel is unusable. It takes a Windows or an Office to be usable.
And the world needs but one engine or a small number of engines, one operating system kernel, or perhaps a small number of them. And the world needs but a small number of large language models. So we felt that in order to have our investors realize their returns, the time has come for us to spend more energy on tuning, refining these models, adopting them for industries, understanding real problems and solving them.
And of course, it wasn't just from that reason alone that drove us to this direction. It was also in talking to a lot of companies that we realized the value of AI will also change. It started being, it's free, ChatGPT, then pay me a subscription, $10 a month, then pay by the API, by token use, and then it's paid by the labor force with which you're contributing.
That is, AI is now able to do tasks that would take white-collar professionals one or two hours to do. It could do in seconds. And if a reasoning engine takes 10 minutes or 20 minutes, it might do something that would take a researcher maybe weeks.
So that is real value. So now I think AI is at a point when people think about it as AI worker. So we've gone from a free product to subscription to API, now to paying for AI worker.
And we're now looking forward to AI outcome. Because once you have a bunch of AI workers who can be orchestrated with a corporate level goal and execute the tasks relatively autonomously, then they are functioning as departments, divisions, groups in a company. And therefore, the money they save and the money they make should be partially or on a percentage basis compensated for the people who provide the AI with which organizations are being replaced or improved with AI.
So charging per outcome is the next big thing on the horizon. So as we go from charging per worker, AI worker, to charging per corporate AI outcome, that's where the money is. So in order to build a generative AI company, one has to look at that as the biggest opportunity.
So those are some of the reasons we decided to pivot the company towards building solutions for businesses that can go from AI workers to AI outcomes. So in our CIO AI white paper, we differentiate between three layers of the value chain, the enabling layer, which are many chips, AI data centers, the intelligence layer, and then the application layer. And it seems you started off with the intelligence layer, but now have pivoted to really solving problems, concrete enterprise problems on the application layer in conjunction with using what you built on the intelligence layer.
Is that the right way to frame what you've built? Yes, it is. I would say there is a platform layer between them as well.
So the intelligence layer is the model perhaps built by OpenAI or Anthropic or DeepSeek. And then the models are not directly useful for corporate. That's something many people didn't fully realize.
They figured, hey, I can use chat GPT. That means if I plug it into my company, it just works. But it's not quite like that.
There is additional need to train on additional data for your industry or your company. That might be done usually with fun tuning, reinforcement learning, or occasionally continuous pre-training. So these skill sets do require our model expertise, but they are not going to be doable by a traditional company.
So that's kind of one layer of work on top of the engine layer, but not quite at the application layer that is tweaking the models to make them work well for the company, industry and applications you need. And then I think there's another layer, or maybe these are the same layer. They're the thing that make models work.
Another layer that is about building, making it easy to build corporate applications. Even if you have a model that's tuned for your industry and application, how will you make an intelligent agent that can screen resumes, make offers, negotiate with candidates or an agent that can schedule trips and deal with emergencies and change them and alert all the right people? These are functions that are highly valuable and require reasoning skills.
And I think we're going to be at a stage where these apps can be built in a day, or maybe in many cases an hour. And that requires what I would call a set of tools. There are a number of open source tools already, LandGraph, NAN, and Crew AI are some examples.
But I think corporates will need something that is easy to learn, easy to build, and has great support so that these will function like a GitHub, like a Visual Studio. So that is the platform layer, that is when your company buys AI, it's no longer buying an API or a model from another company. It's buying a set of tools with which you can build your AI agents, which not only answers questions, writes documents, but also acts for you.
It will make offers to candidates, screen resumes, contact people, email people, pay accounts payable, buy tickets, reserve hotels. So that's when AI becomes actionable and becomes as powerful as a human worker. So I think my company now is adding to the tools.
So they are industry knowledgeable, and corporate knowledgeable, and then having the tools to build agents that build on corporate databases, requirements, and so on. And of course, we will build applications, but we think the final outcome is that there will be standard applications, and then also industries will be able to build their own applications. So enhancing the model and the platform application building layer, and then the applications themselves, that's the business that we're currently in.
Makes sense. And when we think about, maybe when we pivot now, sort of similar question, but pivot away from your perspective as an entrepreneur to your perspective as an investor, when we think about these different buckets of investing opportunities, layers, including the tooling layer that you just mentioned, where do you see as an investor, the biggest opportunities over the next decade? I think there are still a lot of opportunities.
In terms of investing in smaller companies, entrepreneurs, I think now is a great time to invest in companies that are building agentic applications. And I typically say agentic applications, because just to build another writer or a chatbot or a PowerPoint generator, those are useful, but they are not providing the greatest value. It's hard to extract.
Think about it. Think about what applications might I have a hope one day of extracting an outcome-based revenue business model, or what, at least today, we can think about extracting value because it provides AI worker. And an AI worker or an AI outcome clearly requires agentic capabilities.
That is the capabilities of not just answering a question or writing a document, but being able to take on a relatively complex task and use reasoning capability to do equally well or better than humans. We are currently seeing, at least in Silicon Valley, a very rapid emergence of companies that are building agentic solutions in many, many areas, legal, customer service, banking. These will usually start in industries that are already digital or companies that already have done the digital transformation.
So building apps to solve problems for these companies will be the lowest hanging fruit because it's relatively easy to build something. And also, if one can align the AI capabilities to improving a corporate level KPI, such as more revenue, more throughput, lower cost, et cetera, then it becomes a clear equation that companies are going to pay for that because they're seeing good business outcomes. We're seeing companies rapidly, more rapidly than in the mobile era, going into enterprise level monthly ARRs of tens of millions of dollars at a speed that's never been seen before because people are willing to pay the money on a monthly or per use or outcome basis that can lead to positive business outcomes.
So I think agentic applications, targeting industries where it's easy to show results that have already digital virtual businesses are the lowest hanging fruit. But also we think traditional business in every one of them will undergo the agentic generative AI revolution. And I think betting on companies building products, applications, and tools is a great time.
And just to clarify, the reason we're seeing this burst in agentic AI companies is because the underlying reasoning engines are becoming powerful enough so that one can go from a human hard-coded workflow AI to an AI generated and therefore robust agentic reasoning AI. A workflow AI is when a human writes the rules, ask the person where they want to go, ask them how much they want to pay for the hotel, ask them if they want to fly in the morning or evening, much as a travel website, but now you have AI working with a human created workflow. A reasoning engine is when you can just describe the outcome you want.
I need to go to Shanghai for these two meetings, take care of it. What you would say to a capable assistant and then the agentic AI can take care of it for you. So that moving from workflow through the agentic AI to real agentic AI that is reasoning and somewhat close to autonomous is the turning point at which we can see AI is now basically printing money for these startups that do that.
And then lastly, I think another way to make money is really a long, short play. Look at companies that are embracing and adopting AI in a very progressive way and companies that have chosen to stay out. And you long the first one, short the second one, and that's another short way to make money.
So there are many, many others, but these are some examples. That's an interesting last comment. I mean, if you look at the markets and particular stock price performance since the launch of ChetCPT, it's quite surprising to see the divergence of performance and the dividing line, as you said, certainly seems to be AI with NVIDIA being up 10x and call center companies in Europe and also in the U.S. being down 60-70% because those have already been placed partially by AI over the last two years.
So certainly a lens that we think is very important to pay attention to. Agreed. And maybe pivoting from investment domains now to regional differences, you've been very prescient with your prediction that the U.S. and China would be at the forefront of AI innovation in your book, AI Superpowers.
I'm curious now when you look at the landscape, what do you see as some of the most misunderstood portions of this East-West discourse on AI? Oh, there are so many. First one is that some people believe it's static.
U.S. is ahead, so it will always be ahead. For China's caught up, it will get even better. But I think the core to understand is that every country has strength and weaknesses.
The core strength of American technology companies is that they are very good with taking risks on a breakthrough technology and taking a big bet on the visionary view. What's wrong about Chinese companies is the engineering capability. And when it becomes proven there's a business value, there's huge tenacity and perseverance and hard work to do whatever it takes to grab its fair share.
What has then transpired was that we saw the Americans invented deep learning back 10 years ago or so. But China really built the companies that made a lot of money from deep learning, in particular with computer vision. But that was old AI.
With generative AI, U.S. leaped way ahead again with chat GPT and open AI. And now with DeepSea, China's basically catching up by saying, okay, you guys were betting big before something was proven, and this time it worked, you're way ahead, let's go figure out how to make this thing work fast, work efficiently, and how to make money with it. So DeepSea really is a case in point that shows this whole game is dynamic.
Whenever some new technology comes to fore, most likely U.S. would be ahead for a little while. And China will figure it out because it's quite capable in research and even stronger in engineering, and even stronger in building applications. So I think we're now beyond the first inning of this generative AI, which U.S. led with the technology.
But we're now in inning two where China basically through DeepSea said, hey, we figured it out. We built a research product model that's almost as good as the top American models, and it's a lot faster and cheaper. And also that's partly a result of not being able to afford or get the most advanced chips that Chinese companies were forced to become efficient.
So we're now in inning two. There are many more innings to be played out, but I think we'll continue to see. Whenever there's a breakthrough new approach or technology, U.S. will be ahead.
But if it's about engineering efficacy and or proliferating building applications and super apps, China would tend to be better. So that's kind of the first point I want to make. Second thing is that I think a lot of American policymakers and in fact, CEOs and technologists felt that, OK, what's important is the top end GPUs.
If China doesn't get a top end GPUs, it won't be able to produce top end models. And that was proven to be wrong because the Chinese companies were basically forced to use fewer GPUs or inferior GPUs to try to build equally good products, which succeeded through DeepSeek. So I think that's another maybe a surprise to some people.
But I think the first stage of the sanctions really didn't produce the outcomes that that were predicted originally. And I think going forward, there are so I think a lot of that is sort of explaining how China caught up. But I think there are also some areas where I worry about the Chinese ecosystem.
One is that the Chinese VC ecosystem isn't as healthy as it used to be 10 years ago. There's less money, fewer capable VCs, harder to raise money. In Q1 of 2025, U.S. raised at least some U.S. generated AI companies raised at least 40 times more money than Chinese generated AI companies.
So the East raised money, the continued virtuous cycle of VCs who run great companies, make money, entrepreneurs who become VCs, that's currently being rebuilt in China. And that, I think, is a deficit for China and an advantage for the U.S. The last point I'll make is rather surprising to a lot of people, is that China seems to now represent open source and U.S. represents closed source, at least as far as language models.
And I think that is an interesting situation because U.S. invented open source. Most open source successes were American stories. But for some reason, the best American models from XAI, OpenAI, Google are all closed source and also anthropic, are all closed source.
Meta doesn't have a competitive model, so I leave them out. The best Chinese models, DeepSeek and Quen, are both completely open source. They're not open source and inferior model, charge money for the top model.
They open source every model. And that has created a very rapidly growing ecosystem, not just in China, but throughout the world. It's important to note that the speed of adoption of GenAI in the rest of the world, that is not counting in North America, is so much faster.
It's like 20 years faster than the adoption of the Internet. So globally, GenAI is being used. And in many countries, DeepSeek and Chinese models are getting a lot of success because it's free, it's cheaper.
You can take the model, you can adopt it. Unlike the OpenAI and other models, which are can't even be downloaded or touched or modified easily and kept on one's premises. And also one has to submit the data up to the cloud, which may potentially be undesirable for some companies.
And also smaller countries may be poorer and can't afford to use the American models or subscribe to GDPT. And DeepSeek offers a good solution. And also entrepreneurs and universities globally really love DeepSeek.
And also they like Twint from China. So I think very interestingly, open source, at least as far as models go, is becoming an advantage of China, which was not something people would predict. And to the second point that you're making, one often says necessity is the mother of invention, and certainly Chinese large language models have proven that case.
If we think about the limited hardware resources, and yet if you look at some of the leaderboards, the Hugging Face leaderboard, which is probably the most widely used one, it's really two nations that dominate the top 10. It's the US and China. And again, China, despite having access to the number of GPUs that the US companies have.
I know we're running close to the end of our session, but I would like to ask you one last question, which is how do you use AI personally in your private life, but also professionally? And is there one piece of advice that you could give us how to make sure that we are using AI in the best possible way? Sure.
As a personal user, I don't really write anything important by myself only anymore. I always use the top engines to either elaborate or edit or improve what I write. And also, even including PowerPoints, well, my company actually has a PowerPoint-like product that can generate PowerPoints that are almost good enough for me to use, which is something I didn't anticipate would be possible because I have very high standards.
So AI is really taking over content production. And also, whenever I have photos, my daughter's wedding photos, I use editing tools that are AI. I create funny videos.
So almost everything I do, I do with AI for content. And another interesting aspect is that I haven't programmed for a long time, right? Even though I was trained as a software engineer and AI researcher because of my managerial positions, I haven't programmed for about 25 years.
But now I can program again because now AI can be my partner. I'm getting rusty with a lot of the programming languages and details and debugging. And now AI can do a lot of that.
And I can use a lot of tools like Cursor, GitHub to be able to do things. Not that I program professionally, but it helps me understand the technology better. And also, I think it's a marvel to see how good a job AI can do for these things.
And also, my company, ZeroOne.AI, we built a platform for building these applications. So now I don't have to actually... Because if you use Cursor or GitHub, you're still writing a traditional piece of code.
But now I want agentic AI. And now my company's platform allows agentic AI to be programmed with very little requirement of writing code. Drag and drop and graphical interfaces makes it very easy to build something that I talked about, whether it's for travel planning or reimbursement or resume screening.
So I can really be fully active as a software engineer, which I didn't think was possible. Now that, finally, you said to give one advice, I would say that I would guess most people haven't gone as deeply as I have, and it's not necessary necessarily to do that. But one advice I would give is you really should learn how to ask good questions.
People call it prompt engineering, but really generative AI models are this huge network of knowledge in order to tease out the answer that you want. And it's in there, and it's probably a brilliant answer. You need to know how to ask the right question.
Things like you need to ask AI to role play. You need to provide specific details. You need to give it guidance on how to give the answer.
You need to load in any factual data that if you're afraid of hallucination. So how to ask a question is much different than in real life. A good question might be a thousand words, and actually, it might not be humanly generated.
You might want to use a generative AI to generate a prompt that will get you the answer. But I think this is all worth doing because it doesn't require any programming or understanding of expertise on AI or software engineering. But just go to your favorite search engine and ask about how to ask a good question, how to do prompt engineering, and you'll be surprised how much you can tease out and get out of these models and learn from it.
I think this couldn't be a better closing note to say the human mode may be for some time how to ask the right questions. And with all the abilities that you're now using to enhance your own capabilities, I also wonder whether 01.ai may indeed be just one of many more AI companies that you'll start in the future. So thank you very much, Kaifu, for your time.
Ladies and gentlemen, Dr. Kaifu Lee. To our listeners and clients, thank you for your time and engagement.
If you have enjoyed this episode, you may subscribe to UBS Market Moves, available on Apple Podcasts and Spotify for additional content. Until next time, where again, we decode the future of AI, layer by layer.