🌪️Kai-Fu Lee on U.S. AGI Hegemony, OpenAI's Monopolistic Ambitions, and China's Path to Innovation
The former head of Google China and CEO of Chinese AI startup 01.AI said whoever first develops AGI will become a commercial monopoly, and Sam Altman might become the greatest monopolist in history.
The race to AGI isn’t just about creating the smartest machine — it’s about who gets to rewrite the rules of the game.
At the forefront, OpenAI is vying to be the ultimate monopolist, wielding vast resources and an unrelenting drive for dominance. Google, while rich in research and talent, faces an uphill battle to defend its turf. Meta, ever the disruptor, bets on open source to carve out its niche. NVIDIA reaps the rewards as chips fuel this arms race.
From a broader, global perspective, while the U.S. continues to lead in the invention of cutting-edge technologies, China is playing the long game with cost-effective models and rapid deployment.
These insights come from Kai-Fu Lee, a computer scientist with a career at Google, Microsoft, and Apple, and now the CEO of Chinese AI startup 01.AI. In a recent interview with Qianwang, a Tencent News-owned tech media, Lee cautioned that whoever achieves AGI first won’t just lead in technology — they’ll hold the keys to an unprecedented monopoly.
Sam Altman might become the greatest monopolist in history…His strategy, ambition, and well-thought-out plans are admirable.
Whoever first develops AGI that can dominate its competitors will not only achieve a technological milestone but also inevitably become a commercial monopoly.
China’s latecomer advantage lies in its Product-Market Fit experts—product managers and engineers who create super applications that ignite waves of innovation.
A healthy ecosystem distributes value: applications generate the most, platforms earn moderately, and chips earn the least. AI hasn’t achieved this yet.
Lowering inference costs by tens of times will unleash an explosion of applications, driving innovation and user adoption.
Below is an excerpt from the translation of Lee’s interview. Kudos to Benita Zhang张小珺, an exceptional Chinese journalist and podcaster from Tencent News. You can find the original Chinese story here, and follow the author’s future coverage here.
AGI Monopolist
Qianwang: During an interview with foreign media, you mentioned that the biggest difference between Chinese LLM companies and Silicon Valley giants is the ability to create cost-effective models and inference engines. Why are Chinese teams better at this?
Kai-Fu Lee: I can tell you that we’ve competed with Google and OpenAI for talent — fresh Ph.D. graduates — and we’ve lost more often than we’ve won.
These Ph.D. students, while in school, typically work with just a few GPUs for their dissertations. Then Google or OpenAI comes along and says, “Join us, and we’ll give you access to 5,000 GPUs, or even 10,000 GPUs, to work on your dream projects for three to six months.” For these students, the hardest part during their Ph.D. was the lack of GPU resources at school. So naturally, they are drawn to the company that can provide abundant GPU resources, alongside data and real-world application scenarios.
On our side, we say, “Come work with us; we have quite a few GPUs, maybe a few thousand.” But do we have 5,000 or 10,000 GPUs? No. So why would they choose us? At Google or OpenAI, one person might have access to 5,000 GPUs.
Their strategy is to attract brilliant Ph.D. graduates by offering vast resources and giving them the freedom to pursue bold, imaginative projects. Every few months, they ask: “Hey, has anyone made something interesting? Can we scale it down and integrate it into a product?” This is a viable approach, especially if your goal is AGI. They recruit the smartest people, give them massive resources, and burn through GPUs without hesitation, creating tons of fascinating results.
But scaling these results down is very difficult. It’s like designing the most luxurious, beautiful, and grand kitchen and then trying to fit it into a small apartment. Or creating the fastest, largest engine and attempting to fit it into a tiny electric car. It just doesn’t work.
So, the design needs to account for this from the outset. Our goal isn’t to build the world’s most expensive AGI but rather to create top-tier models that feature cost-effective inference. Only with such models can our applications be widely adopted. We make this our focus and are upfront about it with recruits. Some say, “If you can’t guarantee me 5,000 GPUs, why should I join you?” But others say, “I want to do grounded research and practical innovation; you’re a better fit for me.” While we often lose out to OpenAI or Google, those who join us align with our vision and DNA, working together to achieve our goals.
In one of my books, I explicitly stated that the U.S. often leads China by a significant margin in the invention of cutting-edge technologies. However, when it comes to implementation, China can catch up and even surpass the U.S. I used the example of the mobile internet: invented in the U.S., but China’s apps outperformed their American counterparts. Similarly, deep learning and convolutional neural networks were invented in the U.S., but China’s applications, execution, and resulting unicorns are on par with the U.S.
The same applies today. Who invented GPT-4? The U.S. Who invented o1? The U.S. But every time they invent one, we can catch up with it. On one hand, China has many optimization methodologies; on the other, our speed of implementation is not only faster but also often more cost-effective.
Everyone wants to pursue AGI. But aiming to be the first to achieve AGI and completely dominate others? That’s a dream we neither have nor should have.
Qianwang: If someone were to be the first to create an AGI that surpasses others, what could that lead to? Would it be a scientific dream or a commercial one?
Kai-Fu Lee: It’s both. If you train a super-brain, it wouldn’t just have today’s reasoning and thinking capabilities. In the future, it could develop creativity, independent thinking, and even self-awareness. The next steps might involve understanding multimodal systems and world models, eventually evolving into embodied intelligence that interacts with the physical world — essentially turning science fiction into reality.
Even without needing to reach such extremes, let’s frame AGI as purely software-based. If it possesses thinking, reasoning, innovation, independent learning abilities, and creativity far superior to humans, AGI could invent countless new things.
One of those inventions could be: “Design a business model to crush other LLM companies for me”; “Develop a PR strategy to make everyone believe we are the most trustworthy company”; or, in the hands of malicious actors, “Create a cyberattack strategy to paralyze my competitors.” For those with harmful intentions, AGI could enable numerous other malicious actions.
Whoever first develops an AGI that can dominate its competitors will not only achieve a technological milestone but will also inevitably become a commercial monopoly. Moreover, such a breakthrough would likely fuel ambitions to become the ultimate monopolist.
Qianwang: So, we don’t yet know what super applications AGI might bring, but it will definitely create a monopolist?
Kai-Fu Lee: If only one company manages to create AGI, then it will inevitably become a monopolist. It could use this tool to secure its monopoly. In the past, it was hard to imagine something like Windows dominating everything: making all Macs disappear or monopolizing the mobile operating system. Today, such scenarios seem far-fetched. But AGI is essentially a brain. Once it’s created, it could give rise to super applications, or even invent entirely new things for you.
There’s also a belief within the AGI community that by the time we reach something like GPT-6, it might have the capabilities we just discussed. If your AGI can independently think and invent, you could instruct it to help maintain your monopoly, maximize commercial profits, give you an edge, and undercut your competitors. That’s a very tempting prospect.
Throughout history, every deeply ambitious individual has likely dreamed of such absolute dominance, but whether that’s good for the world is another matter.
China’s Path to Innovation
Qianwang: If we (China) don’t pursue AGI and choose another path, how might that scenario play out?
Kai-Fu Lee: The first monopolistic AGI creator’s dream would be to develop the most powerful brain, monopolize all the money, and leave nothing for others. This monopoly would be far more extensive than those of past giants like Microsoft or Google.
But is there an alternative? One option is to build an ecosystem as a moat. A monopolist might not prioritize reducing inference costs or encouraging the emergence of diverse applications. But as users, we would prefer more applications, wouldn’t we?
In the short term — over the next one to two years — inference costs will drop, likely by a factor of 10 annually. Through concerted efforts, it might even decrease by 20 to 50 times. This would ignite a wave of applications: suddenly, you could have search engines, social media, entertainment, e-commerce, and more. Applications would sprout up like mushrooms after a rain.
Now, what happens in an ecosystem dominated purely by a U.S.-based AGI monopolist? If the goal is an AGI monopoly, the approach is simple: throw the most money into buying the most GPUs, burn through them until you achieve AGI, and once achieved, you win the game. But this creates side effects: NVIDIA, for example, would make an extraordinary amount of money because all the resources are poured into GPUs.
A healthy ecosystem, by contrast, should distribute value differently: chipmakers earning the least, platforms earning a moderate amount, and applications earning the most. The platform itself might earn more than any single application, but the combined value of all applications should surpass the platform’s earnings.
This was true for PCs, the internet, mobile internet, and cloud computing. Is this true for AI? Absolutely not.
Today’s AI ecosystem looks like an inverted pyramid: GPUs (chipmakers) rake in $75 billion, cloud providers $10 billion, and applications like ChatGPT only $5 billion. If this imbalance persists, AI-first applications won’t emerge rapidly. Users won’t benefit, application developers won’t achieve Product-Market Fit (PMF) quickly, and it will be harder to generate revenue or secure funding. This breaks the positive feedback loop needed for a healthy ecosystem.
In a proper ecosystem, value flows like this: applications generate the most value > cloud platforms > chips/GPUs. Users benefit first, pay for apps, which then improve and demand more from platforms, driving platform innovation, which in turn spurs chip advancements, creating a virtuous cycle.
Why aren’t we seeing this with AI applications today? The inference costs are simply too high. If we can lower inference costs for top-tier models, we can reach a tipping point for an explosion of PMF-aligned applications. Once these applications flourish, they will build their moats quickly through brand recognition and user data.
Why are WeChat, Douyin (TikTok), and Facebook so hard to beat today? Because they developed applications, user bases, and moats early, even when they were relatively weak.
If AGI is inevitable—and let’s assume it might emerge in seven years—by the time OpenAI or another entity achieves total dominance, we should already have robust social applications, search engines, agents, and hardware in place. If they come to crush us, at least we’ll have the ability to resist.
Qianwang: Why do you say AGI is still at least seven years away?
Kai-Fu Lee: Recently, there was an article titled Situational Awareness: The Decade Ahead by a former OpenAI employee who boldly claimed AGI would arrive in three years. However, some of their assumptions are debatable. I find Epoch AI’s analysis more reliable. In their August article, Can AI Scaling Continue Through 2030, they examined four critical factors:
How quickly can data continue to increase?
How fast can the world scale up its production of GPUs and HBM (high-bandwidth memory)—essentially memory capacity?
How much can computational throughput grow? As neural networks scale, the bottleneck may shift from GPU availability to data transmission. For example, transferring data between one and two GPUs is manageable, but scaling from one million GPUs to two million GPUs creates massive transmission challenges.
How much electricity is available globally?
Their calculations suggest AGI is more likely to arrive around 2030. They argue that the leap in progress from GPT-2 to GPT-4, which occurred between 2019 and 2023 (a span of four years), represents a similar level of advancement needed to go from GPT-4 to AGI. This leap, roughly equivalent to advancing from kindergarten-level intelligence to a high school student, mirrors the next leap required: from a high school student to a genius-level AGI.
If the first leap (kindergarten to high school) took four years, the second leap (high school to genius) would require around seven years. This reflects diminishing returns — progress continues but at a slower pace and at a much higher cost.
We can expect that while the Scaling Law (the principle of improvement with increasing computational resources) may still hold, it will be slower and prohibitively expensive compared to previous advancements.
How much would achieving AGI cost? According to the article, the figure is astronomically high — so high that it might be beyond the reach of any single company.
Qianwang: So, our second path is to rely on ecosystems to resist AGI dominance, is that correct?
Kai-Fu Lee: Exactly. If every application company has the potential to develop its own AGI, competition will make monopolization more difficult.
Qianwang: Does that mean application companies should first establish their ecosystems and generate commercial profits before pursuing AGI, rather than chasing AGI from the start?
Kai-Fu Lee: Yes. Without building ecosystems or achieving commercialization, solely chasing AGI could leave you without resources to continue before you ever realize your dream.
Qianwang: But I have a question: if our primary goal isn’t AGI and we don’t push the capabilities of our models as aggressively as in the AGI main track, can we still achieve an explosion in applications?
Kai-Fu Lee: Five months ago, OpenAI launched its GPT-4o model, priced at $10 per million tokens. Today, our Yi-Lightning model has surpassed the May version of GPT-4o. In just five months, our API pricing has dropped by tens of times while still leaving room for profitability. This kind of accessible pricing has the potential to spark more applications.
Qianwang: Based on what you’ve shared, you’ve outlined a technological philosophy: if the U.S. is likely to achieve AGI dominance, how can China come together to counter it? Besides lowering inference costs, what else should we do?
Kai-Fu Lee: First, we should fully utilize models that are only five months behind the U.S., allowing our applications to outpace theirs and take the lead through latecomer advantage.
I am optimistic that Chinese apps will eventually surpass American ones. While it hasn’t happened yet, it will in the future.
This has already occurred twice — first with mobile internet and then with AI 1.0. What makes China especially strong today is its pool of people skilled at achieving Product-Market Fit (PMF). Our latecomer advantage lies in these PMF experts. They are product managers and engineers who, when paired with model specialists, will ultimately create super applications that ignite a wave of innovation.
Never Underestimate OpenAI
Qianwang: Industry peers say you are very familiar with global matters, having held senior management positions at major companies like Google, Microsoft, and Apple, and maintaining close interactions with these organizations. They asked me to inquire about overseas developments:
What advancements could emerge in the next 1-2 years that haven’t yet gained consensus?
When might we see significant changes on the application side?
Kai-Fu Lee: I just returned from Silicon Valley, where I met with many people. Here are some of the major insights:
OpenAI still has many valuable technologies it hasn’t unveiled. We must not underestimate it. Its GPT-5 training hasn’t gone smoothly, but to secure funding, it released o1. OpenAI still holds many cards and isn’t in a rush to play them. Each time it unveils a new card, global tech companies, including those in China, observe closely, speculate, and develop competing solutions. Even if they can’t match it entirely, they can achieve 80-90% parity. Because of this, OpenAI doesn’t want to exhaust its cards prematurely. It plans to save them for when AGI seems within reach and can confidently deploy them.
This is an incredibly powerful company in terms of technical strength. I feel the gap between OpenAI and Google isn’t narrowing — on the contrary, it’s widening. A company with such extensive technical reserves can strategically decide when to release its technologies, such as during a funding round or to demonstrate strength. However, the challenges GPT-5 has faced during training suggest that the prediction of achieving AGI in three years might be overly optimistic. GPT-5 was originally supposed to be released by now, but at this rate, even if it does launch, it won’t be for another six months at the earliest. In the meantime, o1 has stepped in to complete this funding round.
o1 itself hasn’t brought significant improvements in reasoning or understanding, but it has implemented a dual system for fast and slow thinking, which will complement each other in the future. For a technologist, 1.5 years ago, all the focus was on pretraining because it was extremely difficult — something that very few companies had done. Now, many companies have learned how to do it. Afterward, they discovered that another hard problem was excelling in post-training, particularly reinforcement learning. Today, very few companies are good at reinforcement learning. Now, OpenAI has introduced a third challenge: scaling laws in inference.
The most impressive thing about o1 is its scaling law for inference — longer inference times lead to better reasoning outcomes, something previous versions of ChatGPT didn’t have. This has expanded the scope of thinking for many people.
Qianwang: A new goldmine has opened.
Kai-Fu Lee: This goldmine is highly beneficial and stimulating for the entire world. I predict that advancements in the inference stage will go far beyond the technology in o1 today. Within our internal discussions, we speculated about how o1 was built, and our team came up with three different solutions. Multiply this by 20-30 other top-tier companies, and even larger ones than ours. Hundreds of solutions will emerge, including things that have never been done or thought of before.
Returning to the startup ecosystem, a “standard Silicon Valley consensus” now is the article published by Sequoia Capital (Sequoia US). Have you read it?
Qianwang: The one released after o1? Generative AI’s Act o1: The Agentic Reasoning Era Begins.
Kai-Fu Lee: Yes, that’s the one. It uses o1 as the title, but you need to read to the end, where the investment logic is laid out very clearly. Entrepreneurs in China can take a look. This aligns with what I’ve gathered from my conversations with Silicon Valley entrepreneurs and investors.
The sentiment in the U.S. is that the pretraining of foundation models will be concentrated in the hands of a few companies. Investing in new efforts to build models may seem exciting, but the returns may not be the best. From an infrastructure (Infra) perspective, they believe that some of the good companies have already emerged and that it’s not the most attractive investment area. The focus is shifting to apps and “Service as a Software.” This Silicon Valley consensus could influence investment approaches in the U.S. and might also be followed by Chinese VCs.
I don’t completely agree with everything stated in the article. In my view:
Models: If you have an advantage in inference, you can develop unique strategies.
Infrastructure: U.S. VCs don’t fully understand where the value lies in Infra, so there’s still opportunity.
Apps: While I agree that investing in apps is important, we shouldn’t be overly optimistic.
In the past, people often spoke of PMF (Product-Market Fit) for apps. In the LLM era, I add two letters: TC-PMF (Technological Cost × Product-Market Fit). Beyond PMF, we must account for TC—how advanced the technology needs to be (e.g., multimodal capabilities or video processing), inference optimization to reduce costs and improve performance, and predictions on who can develop the technology, when, and at what cost. Furthermore, when will the cost become low enough, and how can we maintain sensitivity to frontline markets? Integrating these elements is a highly complex task.
Whoever first identifies TC-PMF will create the “Super App” of the LLM era, much like TikTok in the mobile internet era.
A good app entrepreneur needs to be market-sensitive, understand how to build a product, identify its demands on models and technology, know how to fine-tune the model, and determine when it can be fine-tuned effectively. Additionally, they must assess whether the inference cost will be low enough when the tuning is complete. The difficulty has doubled compared to before.
As for 01.AI, I am relatively confident because the model is in our own hands, making it slightly simpler to align these points.
Qianwang: You are quite familiar with the moves of major players. Could you comment on the actions of overseas giants like Google, Microsoft, NVIDIA, Tesla, and domestic giants like ByteDance, which is going all out?
Kai-Fu Lee: NVIDIA is undoubtedly the biggest winner right now. However, it may face challenges as more GPUs transition from training to inference—can it maintain its edge? That’s the question. It might, but I don’t know for sure.
Meta is the biggest disruptor. When it can’t win at something, it goes open-source, using this strategy to carve out a position. I really admire how they allocate a large team to focus on ads, earn money from advertising, and then use open-source as a wedge. Their technology doesn’t match OpenAI’s, but by going open-source, they’ve positioned themselves well, attracting widespread adoption of their models. Later, they can explore opportunities to push further. While I don’t think Mark Zuckerberg deeply understands AI, his tactic — if you can’t beat them, open-source it —has worked twice. The first time was TensorFlow versus PyTorch, and this time seems equally successful. Meta’s position is solid.
Microsoft is in the best position. On one hand, it is making substantial profits from AI; on the other, it holds a significant stake in OpenAI, which gives it both offensive and defensive capabilities. However, its challenge lies in its own models, which haven’t performed well. While OpenAI and Microsoft are partners now, this honeymoon phase may not last. OpenAI likely has a Plan B, and if Microsoft doesn’t develop one, it could face trouble in the future. For now, Microsoft is in an excellent short- to mid-term position. In the long term, if it fails to build its own models and falls out with OpenAI, it could face serious challenges. At present, Microsoft and NVIDIA are the biggest beneficiaries.
OpenAI is an exceptionally powerful, monopolistic company. Sam Altman (OpenAI CEO) might become the greatest monopolist in history. I’m not saying this pejoratively — just stating the facts. While he hasn’t achieved monopoly status yet, his strategy, ambition, and well-thought-out plans are admirable. From an industry perspective, however, it’s also concerning.
xAI, led by Elon Musk, is incredibly execution-focused. Musk runs his company much like a Chinese company. I know some of the key members of his team — they’re extremely capable and work tirelessly for him. His current achievements replicate some of the early technologies of OpenAI and Google at an impressive speed, which is rare in the U.S. Whether he can integrate Tesla’s embodiment technologies and autonomous driving with xAI remains to be seen. Regardless, Elon Musk is a dark horse we shouldn’t underestimate.
Google’s situation is more bittersweet. Theoretically, it should be the strongest player — its papers on LLMs are the most impressive, and its reinforcement learning research through DeepMind is unparalleled. Yet combining these hasn’t yielded significant results.
Meanwhile, Google faces a multi-front challenge in search.
On one side, LLMs are pulling some users away, as they now turn to ChatGPT for queries instead of search engines. More significantly, a growing number of users are bypassing Google entirely for shopping-related searches and going directly to Amazon. In China, it’s long been the case that people shop on Taobao or Pinduoduo and use Baidu for other searches. Now the same trend is emerging in the U.S., with users going directly to Amazon for commercial searches.
On top of this, Google is caught in a dilemma about whether to integrate LLMs into search. There are three approaches:
Replace search entirely: This would dismantle Google’s entire ad business, effectively shutting the company down.
Keep it separate: This creates two entry points, which is counterintuitive — why split a single process into two?
Merge the two: Display a Gemini Overview alongside traditional search results. But this approach leads to reduced ad revenue and a suboptimal user experience. If I ask a question and just get an overview, it doesn’t solve my problem. Users want a clear answer, not an overview. This turns Google’s offering into a mishmash — search results with lots of links, ads, and an overview, but no straightforward answer.
This is Google’s current strategy. But there are other strange things, like Gemini not answering election-related questions. Why? I’m not sure. There have also been issues like AI recommending eating glue or rocks — problems that persist without clear reasons.
In the short term, I’m not optimistic about Google. However, Google has strong technological foundations. Whether it can rebound remains uncertain.
Qianwang: Is Perplexity a viable product model? Could it disrupt or replace Google?
Kai-Fu Lee: It’s a very good product. It’s not that Google can’t create something like Perplexity — Google could easily build one in no time. However, due to the considerations we just discussed, it can’t afford to do so. Perplexity can afford to make less money or even lose money to attract a large user base. Google, on the other hand, generates $0.016 (1.6 cents) in revenue for every search.
Currently, Perplexity Pro charges $20 for its users, but the cost of providing searches, especially using GPT-4o, is very high and cannot cover those costs. Perplexity doesn’t mind though; it’s willing to lose money on free users and even subsidize paid users if it means acquiring more users and leveraging that growth for investment. Google, however, cannot retaliate in the same way. This is Perplexity’s greatest advantage — it’s a well-executed product, though not extraordinarily groundbreaking.
What Perplexity does exceptionally well is addressing trust issues — or at least giving users the illusion that trust issues have been resolved. It uses a citation model, which makes users feel confident in the answers. When they see a response with numerous citations they can click on, they feel assured of its reliability, thinking it has no hallucinations. However, this is not entirely true — we’ve evaluated Perplexity’s hallucination rate, and it’s still quite high. But users see citations and feel more at ease. This is an interesting and effective user experience trick worth learning from.
In terms of use cases, if you’re conducting research or seeking insights, Perplexity’s user interface (UI) is spot on. It integrates various visuals, videos, clickable articles, extensions, and citations, like a digital library. Perplexity positions itself with a company vision to be a “Swiss Army knife” — a tool that can handle everything from opening bottles to cutting paper, doing a dozen different things in one package.
That said, I tend to believe what Larry Page (one of Google’s founders) once said. Early in Google’s history, I attended a meeting where Larry Page stated: “Our search results — a bunch of links popping up after you type in a query — isn’t the right way to do it. The correct search model should be asking a question and receiving a single, correct answer.”
With today’s technology, we are closer to achieving this “one question, one answer” model. However, creating an extremely complex research tool might not cater to the needs of everyone. While it’s an excellent tool for analysts, researchers, professors, students, and journalists, it’s not necessarily what the average user wants.
So, while I acknowledge and appreciate Perplexity, the idea of it replacing Google remains highly challenging.
Qianwang: Perplexity seems to be gradually evolving into a content product rather than a search portal.
Kai-Fu Lee: However, it has also faced lawsuits from content producers; The New York Times recently sued them. This is a significant issue in the U.S., and The New York Times is likely to pursue legal action against other companies as well.
Qianwang: Back to OpenAI, will Ilya’s (former Chief Scientist of OpenAI) departure have a significant impact on the company?
Kai-Fu Lee: The departure of members from the management team not only hasn’t negatively affected OpenAI but might even allow Sam Altman to execute decisions faster. Those who left didn’t doubt OpenAI’s ability to become a great company or question its product or technological strength. Instead, they prioritized creating a safe and responsible AGI over building a commercially successful LLM product.
Sam Altman is focused on the latter, while they were focused on the former. You can imagine that if a CEO’s goals don’t align with his direct employees, he’ll face numerous obstacles in execution. With their departure, he can fill the team with people he trusts and whose philosophies or ambitions align with his, thereby improving execution. However, the safety of OpenAI’s products will undoubtedly decline.
AI Apps to Dominate 2025
Qianwang: Are there any promising applications or potential super apps emerging now?
Kai-Fu Lee: When a super app first launches, it doesn’t start as a super app; it evolves into one over time.
In the era of LLMs, the concept of a super app may differ, and we should maintain this mindset. Every mobile internet application will be reimagined and rewritten. If you invest in the disruptors, you win. Disruption might involve capturing the business of current leaders, but that’s the lowest level. We hope to see the creation of new models that integrate more user needs, generate more revenue per user, attract more users, and potentially evolve into super apps. The first stage is to redo existing applications with some extensions. Perplexity, for instance, is a reimagined version of Google. The next stage involves unprecedented applications.
Qianwang: Like TikTok during the mobile internet era?
Kai-Fu Lee: TikTok, Didi, Meituan — what would be their equivalents in AI-first applications? Beyond this, there’s potential for some truly “super” Super App opportunities.
There’s a plausible theory, though not guaranteed to happen: from a user interface perspective, there should be an intelligent assistant interacting with you. This assistant wouldn’t just answer your questions but also get tasks done for you.
If this day arrives, the intelligent assistant could become the greatest Super App of all time. It could connect and satisfy all your needs — for instance, if you want to buy something, it decides which e-commerce platform to use; if you want to travel, it books tickets for you.
It won’t start out being that powerful, but over time, users may come to realize that their assistant understands them better than they understand themselves. Whether for work or personal matters, people would gradually entrust and rely on it entirely.
This represents a shift from graphical user interfaces (GUIs) to conversational user interfaces (CUIs), and eventually to a delegation-based user interface. You would delegate tasks to your assistant.
If your assistant is smarter, more capable, more knowledgeable, and understands you better than you do yourself, why wouldn’t you let it handle everything for you?
Qianwang: Could today’s ChatGPT become this kind of Super App?
Kai-Fu Lee: They can certainly aspire to become such a Super App, which is entirely reasonable. If this day comes, all the dominant apps today will face challenges. For example, if you’re an e-commerce platform and users no longer treat you as the entry point, instead accessing you through an assistant, you become just a passive bidding warehouse. Your value will plummet.
Of course, during the rise of such assistants, e-commerce platforms may resist cooperation. If the assistant can’t collaborate with these platforms, what happens then? This scenario somewhat mirrors how Toutiao evolved from being a news aggregator to a central content platform (like TikTok). In the same way that Toutiao progressed, today’s chatbots could progress similarly.
In the future, the App Store might no longer exist, nor might e-commerce platforms or standalone apps. Why would we need them if the assistant handles everything for us?
A larger opportunity could arise if hardware changes. Smartphones, for instance, could remain the primary carrier of these assistants. The advantage is portability and the ability to store personal data. The downside, however, is that today’s phones cannot achieve always on and always listening.
When I delegate tasks to my assistant, I don’t want to pull out my phone and select an app. I should simply state a command, and it should respond — whether through a screen, smart glasses, or earphones. This would motivate more frequent interaction and delegation to the assistant.
The goal would be to reduce a 40-50 second process to just 1 second.
Qianwang: When do you think we’ll see a global explosion of applications?
Kai-Fu Lee: For 2C (consumer-facing) applications, likely within the first half of next year.
Qianwang: Is this based on a feeling, or are there signs?
Kai-Fu Lee: It can be calculated based on inference costs. Right now, inference costs have dropped to a level where it’s feasible. In the U.S., domestically in China, and elsewhere, costs have come down—not as low as ours, but still relatively low. This should enable the creation of some products with Product-Market Fit (PMF).
Another indicator is the shift in venture capital sentiment. Many VCs are starting to say they’re open to investing in applications. For entrepreneurs, securing funding is the first priority—without it, how can applications explode?
Qianwang: Would you still invest in such projects as an investor?
Kai-Fu Lee: Yes, I would.
Qianwang: Finally, if we are to face a potentially dominant U.S. AGI hegemony, what would you like to say to your peers in China, especially given the intense domestic competition?
Kai-Fu Lee: Currently, the “Six Little Tigers” (leading Chinese AI companies) are pursuing different paths and aren’t directly competing with each other—some focus on domestic 2B (business-facing) applications, others on overseas 2C (consumer-facing) applications, some on entertainment, and others on chatbots. Each company hopes to encourage the others and develop successfully.
It’s already very clear that following OpenAI’s path is a significant challenge. There’s a high probability that we won’t be the first to “burn through” and achieve AGI with a super-LLM. However, there’s more than one way to win. Each company can forge its unique path, learning from America’s strong innovations and explorations while ensuring our business models and strengths remain distinctive.
The arrival of the app era is great news for China. China’s understanding of app methodologies, its insights, and the sheer number of people capable of executing these strategies far exceed those in the U.S. Simultaneously, through multidimensional efforts in “models + AI infrastructure + applications,” we aim to rebalance the current generative AI “triangle” ecosystem. By redirecting semiconductor-derived profits back to the application layer, we can help the industry return to a healthy, virtuous cycle.
In the long term, this shift is inevitable, but it will take time. From this perspective, Chinese teams possess a vast market, diverse application scenarios, and exceptional execution capabilities. These factors combined offer us a unique opportunity to lead in the AI-first era.
Our time has come.