The AI veteran discusses the industry’s evolution, highlighting opportunities in cost-efficient models and the race to turn innovation into impact.

Held in Beijing from November 28–29, the 36Kr WISE2024 summit drew a stellar crowd, securing a strong response from China’s business community. Now in its 12th year, the event has consistently chronicled the evolution and resilience of Chinese enterprises amid shifting economic landscapes.

2024 has been a year of tempered growth, marked by ambiguity and a pragmatic search for new economic drivers. Industries face a growing need to adapt as emerging shifts challenge entrenched norms. This year’s WISE conference embraced the theme “Hard But Right Thing,” urging participants to confront questions about what constitutes the “right thing” in today’s complex environment—a topic of particular resonance in the rapidly advancing artificial intelligence industry.

In the span of a year, seismic changes have defined the AI 2.0 era of large language models . Once-reliable scaling law paradigms are now under scrutiny, and the debate over whether game-changing applications will first arise in consumer or enterprise markets remains unresolved. What was celebrated as progress yesterday may now be reevaluated as a misstep.

Kai-Fu Lee, a leading voice in the field, offered insights at the summit. With an extensive background as a global executive at Apple, Microsoft, and Google, as well as a track record of backing over a dozen Chinese AI companies, Lee brings a unique perspective. Now at the helm of 01.AI, an AI unicorn he founded, Lee’s multifaceted experience positioned him to dissect the challenges and opportunities of this pivotal moment in AI development.

During the conference, Lee shared a conversation with Feng Dagang, CEO of 36Kr.

The following interview has been edited and consolidated for brevity and clarity.

36Kr: Speaking of the “right thing,” what do you think were the right things you’ve done over the past two years? What logic drives your decisions?

Kai-Fu Lee (KL): I think the most important thing I’ve done is starting 01.AI. While I can’t yet declare it a comprehensive success, it represents a way of making the right choice.

I deeply believe in Steve Jobs’ philosophy: life cannot be planned endlessly for the future. But if you look back, you’ll find that everything you’ve done can connect into a line. If each of your decisions follows your heart, you will eventually find they were correct and paved the way for the future.

At the age of 20, I chose to dedicate my life to AI. Since then, I’ve built various AI-driven products, founded Microsoft Research Asia upon returning to China, and later created Sinovation Ventures. I’ve seen the brilliance of AI 1.0 and the challenges it faced. I believe every action I take now is something I must do.

When facing the future, one piece of advice is: don’t wait until everything is certain to make a decision. If you do, by the time you start, others will have already seen the same thing.

How can you make decisions when things are uncertain? History offers us great inspiration. We can see how PC and mobile internet applications emerged, the opportunities and challenges in the consumer and business sectors, the paths AI companies have taken, what sparked initial waves of innovation, the challenges they faced, and whether they ultimately became great companies.

These lessons and learnings are invaluable. I advise everyone that making the “right” decision doesn’t guarantee ultimate correctness. Instead, it ensures you’ve done your homework, know what you want to pursue, and are ready to chase it with unwavering determination.

36Kr: When making decisions, do you feel conflicted? Many entrepreneurs hesitate because their ventures’ outcomes are uncertain. Yet, there’s immense excitement and joy in seeing past decisions connect to today’s must-do opportunities.

KL: That’s a great way to put it. When you need to determine if something is a must-do, ask yourself: “Which would you regret more—failing after trying or not trying at all?”

Most of my major life decisions have been based on this principle, including this time. My career hasn’t been without failures, but what I’m deeply proud of is bringing together exceptional talent at every turning point. Some of these individuals have gone on to build great companies, such as Kuaishou, or excelled as venture capitalists and executives at major firms, like Jiang Fan.

Success isn’t always about creating a company with trillion-dollar valuations. Sometimes, history offers you the opportunity to nurture talent and improve the overall environment. While other achievements also matter, I believe in continuously trying, understanding my heart, moving on without regret, and spotting the next opportunity.

36Kr: You’ve recently visited Southeast Asia and the Middle East, among other regions. What are your impressions? Do people in other countries view AI differently from us?

KL: Silicon Valley remains a key benchmark. It has firmly established itself as a global leader. Here are a few observations:

1. OpenAI will continue its R&D, but the pace of model iteration will slow compared to before. While OpenAI has profoundly impacted the world, its significance lies not only in its advancements but in revealing a second scaling law in the inference stage, beyond the pre-training stage.
2. The AI 2.0 ecosystem in the US has rapidly developed, largely because ChatGPT originated there and swept the globe. Integration into various applications happened swiftly, educating users and establishing awareness that chatbots aren’t just for conversations—they can accomplish meaningful tasks.

In the AI 1.0 era, China actively embraced new technologies and led the world in applications. However, in the AI 2.0 era, while ChatGPT has dominated, it hasn’t been open to China. Although several Chinese companies are making great strides in large model development, China has yet to experience its own ChatGPT moment. Many users remain unfamiliar with these technologies and haven’t formed usage habits—this must change in the coming year.

Globally, the differences align with expectations. The US leads in technical understanding, followed by China, while Southeast Asia and the Middle East face significant challenges in market comprehension.

Unlike operating systems, which are purely technical, large models like ChatGPT involve language, culture, laws, values, and religion. In non-Western countries, these models sometimes fail to meet expectations, as biases in US-dominated training data can result in perceived inaccuracies or discrimination. For instance, questions about Palestinian rights might yield responses seen as biased, reflecting the skewed data sources.

Thus, I foresee a trend of “one country, one model,” where each region develops models reflecting its local values, laws, and cultural characteristics. Unlike the universal applicability of operating systems, large models may struggle to achieve global uniformity.

36Kr: That’s enlightening. Previously, the idea of each country having its own large model seemed to stem from trade disputes. But on a fundamental level, it seems nearly impossible to create a single model that accommodates all cultures, given the significant differences between nations.

KL: Exactly. China and the US have different market leaders. People often think China’s market size, technological strength, and entrepreneurial talent give it an edge. However, I believe that every wealthy country, like Saudi Arabia, and every nation with a large user base, such as India and Indonesia, will eventually have its own large model.

These models might be developed by companies from other countries or fine-tuned locally using open-source models. Take ChatGPT as an example—especially considering how US companies operate. It prioritizes serving the US, then Europe, with other regions left for “later.” This “later” creates opportunities, allowing entrepreneurs in these countries, or those outside the US, to seize this window of time and develop their own large models.

36Kr: Last year, scaling law was deemed a breakthrough and it indeed delivered remarkable results. However, this is no longer the consensus this year—it seems to have reached its limit. Do we have a new consensus now? What is it?

KL: I believe the scaling law still holds, but now you need to spend ten times the money to achieve incremental gains. The improvements are there, but they’re not as impactful as before.

The scaling law hasn’t stopped, but investing heavily into it is no longer a good business proposition. For example, training GPT-4 cost USD 100 million, GPT-5 might cost USD 1 billion, GPT-6 could reach USD 10 billion, and GPT-7 might cost USD 100 billion. These numbers are staggering, and I don’t think most companies will pursue this path. OpenAI might, but this isn’t a suitable direction for China. Spending vast sums on uncertain outcomes isn’t aligned with the operational habits of Chinese tech giants, let alone feasible for startups.

Another key point is that continuing to push model parameters and building ever-larger models will primarily drive up Nvidia’s stock prices. US tech giants are still aggressively buying GPUs, but this isn’t beneficial for the ecosystem as a whole.

The real question we need to ask is: how can every entrepreneur in this room build AI-native applications? The biggest bottleneck today is the high cost. Just over a year ago, GPT-4 cost USD 75 per million tokens. Today, 01.AI’s Yi-Lightning model costs RMB 0.99 (USD 0.14) per million tokens. Yi-Lightning not only outperforms GPT-4 as of its 2023 launch but also surpasses GPT-4o from May of this year—500 times cheaper than GPT-4’s initial price and 30 times lower than GPT-4o’s.

It’s clear that industry costs are declining. If you’re starting an AI venture and feel that models are still too expensive or inadequate, I can confidently tell you that just by looking at changes from last year to this year, you can predict the future. Within 18 months, costs dropped by 500 times, and model capabilities improved significantly. If you think it’s expensive now, there’s a 99% chance it won’t be next year—and in another year, not only will it be affordable, but it will support your desired applications.

So our thinking must change. Last year, myself included, we all focused on the scaling law—getting more GPUs and building larger models. Now, it’s clear the application era has arrived. The focus should be on making excellent models fast, cheap, and accessible to everyone, ushering in an era of inclusivity and entrepreneurial prosperity in the AI 2.0 era.

36Kr: The figure of 500 times cheaper is incredibly persuasive—almost free.

KL: Exactly. And we will continue developing better models that will be even cheaper next year than they are now. Imagine when PCs or mobile phones first emerged—they thrilled us. Today, if a PC improves and becomes 30% cheaper than last year, we’d be delighted. The same goes for phones. But here, we’re talking about a 500-fold difference. Imagine the economic impact a car, a house, or a phone could achieve with this level of cost reduction.

36Kr: We can also see that 01.AI’s strategy has remained consistent. You’ve stuck to the “three-in-one” strategy. Do you believe this approach represents the “right thing” in the AI industry?

KL: Yes, if we were to separate the tasks—one team for applications, one for models, and another for AI infrastructure—the results wouldn’t be as efficient.

If you agree with what we discussed earlier—our goal is to significantly lower costs to enable flourishing applications—then the “three-in-one” strategy is essential.

Of course, for a startup, doing one thing well is already difficult—doing three things at once seems daunting. How do we achieve this? Through vertical integration. By optimizing infrastructure and models simultaneously before they solidify, we achieve faster, cheaper, and more efficient outcomes.

For example, to make large models affordable, the first question is: why are they so expensive? The main culprit is GPU costs. Can we use fewer GPUs? We addressed this by leveraging memory as a substitute. We implemented a memory cache system to record calculations that might be reused, loading these cached results directly when needed. This approach maximizes the reuse of prior computations. At the same time, during model design, we consider architecture, the number of GPUs, memory size, and the cache system to create an inference-friendly model.

Most model companies focus on “training-friendly” models or the largest models under the scaling law. In contrast, we design for ultra-fast inference during training.

Many of humanity’s greatest, world-changing products came through vertical integration. The iPhone was created this way—a single team optimizing everything. The Mac was vertically integrated. Tesla and SpaceX are also examples. I believe that in an industry that hasn’t yet standardized or solidified, vertical integration is the best way to break through.

That doesn’t mean 01.AI aims to be the world’s largest infrastructure company, model company, or app company. Instead, we focus on optimizing vertical integration for the models we want to build, concentrating on key products such as the Yi-Lightning model, consumer-facing apps, and upcoming enterprise solutions.

36Kr: Where should startups allocate their resources, given the current trend of “burning money” in the AI industry? Where do you think money should be allocated to bring the greatest value?

KL: The term “burning money” carries a somewhat negative connotation. What we advocate is smart growth, even if the application doesn’t generate revenue immediately.

For example, AI-powered productivity tools—such as those for content creation or video editing—already have users willing to pay. This is a proven route. In such cases, there’s less need to worry about inference costs being too high. If user payments exceed operational costs, it’s sustainable and offers hope for a quick return on investment (ROI).

Another strategy is to accumulate users first without charging them initially. While some may call this “burning money,” it can lead to massive user bases, which unlock significant value later. However, this route carries high risks—it requires significant funding and comes with a high failure rate.

For startups pursuing this path, I have two key recommendations:

1. Understand the risks: A high failure rate is inherent in this strategy.
2. Don’t spend recklessly: Validate your product-market fit (PMF) first. Look for healthy user retention and reasonable customer acquisition costs before scaling.

I don’t think China’s chatbot products have achieved reasonable PMF this year. Big corporations can afford to “burn money,” but startups cannot. Competing head-on with giants is a surefire way to fail. Consumer-facing apps must launch at the right time. For example, last year, GPT-4 cost USD 75 per million tokens. If you built an AI-powered search engine back then, each query might have cost USD 1—unsustainable.

Today, using the Yi-Lightning model reduces those costs to one five-hundredth of GPT-4’s original price. This makes achieving PMF viable. Timing is critical: starting too early might kill your startup, starting too late misses the opportunity. Savvy entrepreneurs should anticipate cost declines and align their application development accordingly.

Unlike big companies that hoard resources, 01.AI leverages its model-infrastructure-application triad to build better products efficiently. This vertical integration allows startups to adapt flexibly to market changes, iterating quickly to improve outcomes.

Startups don’t have the “innovator’s dilemma” burden that large companies do, which often makes them disruptive forces. This is why ChatGPT emerged from OpenAI, not Microsoft or Google.

36Kr: Does this mean that 01.AI won’t engage in burning money for consumer-facing super applications? You’ve also launched some solutions—do you think B2B is the best path for AI adoption in China today?

KL: We’ll pursue the user growth strategies I described earlier, but we won’t burn money recklessly. If PMF hasn’t been achieved or user acquisition costs are too high, we won’t spend heavily. But once PMF is validated, we’ll act decisively while keeping costs under control, especially given today’s challenging funding environment.

We don’t reject the idea of creating a consumer-facing application that’s non-monetized to serve a high daily active userbase, but we’ll approach this with great caution.

I also advise entrepreneurs to apply similar principles—don’t assume that burning money to reach one million daily active users will attract investors. Today’s venture capitalists are much smarter than they were during the early days of mobile internet when everyone was still exploring.

Returning to B2B, there are opportunities, but there are also risks. If every project requires a request for proposal and follows the project-based approach of AI 1.0, it may lead to unsustainable business models where companies lose money on each deal.

At 01.AI, we aim to spend cautiously and operate frugally in both B2B and B2C segments. For consumer applications, we avoid reckless spending. For enterprise projects, we focus on creating value—developing standardized products, fostering long-term client relationships, or securing lighthouse clients to attract more customers. Such projects are worth pursuing.

In the B2B space, we’re committed to providing complete solutions. If a company only offers models for sale, it’s unlikely to find buyers today. The dream of selling models for big money faded over a year and a half ago—it’s no longer viable.

Enterprise clients have clear needs: cost reduction and efficiency improvement to solve real business problems. These challenges might include customer service, user acquisition in e-commerce, or product promotion. The best applications create immediate value—generating revenue or cutting costs for clients—making them eager to pay. This is our B2B principle.

36Kr: Regarding OpenAI’s approach of spending hundreds of millions—or even billions—is there an alternative path for Chinese companies? What might that look like?

KL: I believe there isn’t just one path. Today, companies are pursuing diverse strategies—some focus domestically, some internationally, some target consumer markets, others focus on B2B, healthcare, and beyond.

In the early days of mobile internet, companies proudly identified themselves as mobile internet firms. Today, that’s no longer the case. ByteDance and Meituan, for example, don’t define themselves as mobile internet companies anymore—they have found their unique directions.

The same applies to large models. China’s four “little AI dragons” of the past started as computer vision companies but eventually found their own business paths. The ultimate test for any company is whether it can build a sustainable business model from large models—one that withstands scrutiny in the secondary market, generates revenue, achieves growth, and becomes profitable. Only then can it claim to have found its true identity.

36Kr: Is it difficult for companies that originate in China to enter Western markets? If so, can they succeed in other regions?

KL: Entering the US market is extremely difficult—nearly impossible. However, if you’re developing an excellent application where your proprietary model is the underlying technology, that could still work.

Beyond Western markets, some countries do offer opportunities. Southeast Asia, the Middle East, and Latin America, for example, are less influenced by US restrictions on Chinese models. However, building a giant company in these markets is still very challenging.

Language is another significant hurdle. While many Chinese teams are proficient in English, entering markets that require Arabic, Spanish, Portuguese, or other languages presents substantial challenges. Entrepreneurs must carefully assess whether such directions are feasible and recognize the potential ceiling these markets impose.

36Kr: One last question: what’s your outlook for the AI industry in 2025? Will it improve or face more challenges? Will many companies fail or thrive? What advice do you have for entrepreneurs?

KL: The landscape will definitely differ from last year’s. In 2022, we witnessed the “war of a hundred models.” This year, only a few remain. Yet, Chinese entrepreneurs demonstrate extraordinary resilience, adaptability, and openness. I’ve seen several companies successfully pivot from pre-training models to other paths, and they are doing okay.

The transformations are ongoing and will continue in the future. Over the next year, I’m most optimistic about applications. From Sinovation Ventures’ perspective, we initially invested in large models, then infrastructure, and now our focus has shifted to applications. As inference costs decline, we’re entering an era of flourishing applications. Entrepreneurs who leverage powerful, fast, and affordable models to build innovative applications are making the right choice.

#AIInnovation #ScalingLaw #TechEntrepreneurship #AIApplications #FutureOfAI