The current excitement over AI models is not the first wave of enthusiasm you’ve observed. How has this been different? The pace of this wave of AI development has clearly exceeded my expectations.

AI has been developing for nearly 40 years, if we start from the 1980s, and I have lived through more than one developmental cycle.

In the 1980s, Japan launched its ambitious Fifth Generation Computer Systems initiative. People thought it would bring sweeping changes, but it ultimately fell short of expectations.

Over the past decade, China has had its own wave of enthusiasm around the “four AI tigers”: SenseTime, Megvii, CloudWalk and Yitu, with their focus on computer vision technology. That, too, did not deliver what many had expected.

The current wave is different. Its progress has exceeded expectations, and its impact extends far beyond AI itself. AI is no longer just an industry; it is a variable that will affect almost everything, as we are already seeing in how it assists mathematics research.

To understand this wave, we need to look at how the roles of data, models and computing power have changed over time. We often discuss all three together without making it clear that their importance shifts in different phases of AI development.

The AlexNet model in the 2012 ImageNet competition showed the importance of data. The introduction of the Transformer architecture in 2017 showed the importance of model design. And Nvidia’s rise over the past few years reflects the significant role that computing power has played more recently.

My view is that data will become more important again over the next three to five years, perhaps longer. Internet text is depleting fast for AI training, while vast amounts of scientific data remain under-explored.

I also think we need to reconsider what we mean by “artificial intelligence”. Much of the discussion today treats it as artificial human intelligence. I think that is the wrong way to frame it. I think when we talk about AI, we’re actually talking about machine intelligence.

Machine intelligence and human intelligence have overlapping capabilities, but they need not develop along the same path. If we constantly measure machines against human intelligence, we risk starting from a mistaken premise. Changing that perspective could change how we think about many questions in AI.

You predicted that 90 per cent of AI applications and technologies will disappear within the next decade. Which AI tools are truly useful, and which ones are just there to address “pseudo-demands”?

I made that prediction against a very specific backdrop. At that time, everyone was just building extensions on top of ChatGPT and other GPTs. Even OpenAI wants to turn ChatGPT into a platform for developers to build applications on top of it.

If you look at it today, a chat tool itself is a very small application. If you view AI so narrowly that you only derive applications from a chat platform, you are only capturing a very small fraction of what AI can do for the world. If we calculate by token volume today, the vast majority are already being used elsewhere; they are not on those initial ChatGPT-based applications.

I believe AI will be able to do lots of things for us. As its capabilities expand, applications built around a chatbot platform to perform a few limited functions may struggle to be sustained.

In AI applications, at least two areas deserve particular attention. The first is coding. I prefer to describe it as “coding” rather than “writing software”.

Coding used to be seen as a specialised skill belonging to software engineers. Now, anyone with an idea could turn that idea into something tangible with the help of AI coding tools. This explains why I believe AI coding opens up a wide range of use scenarios and will create value far beyond the traditional software industry.

The second area is scientific research. Computers were widely used in research from their earliest days, and AI could now further change research practices.

In the past, people might ask how much training someone needed to become a scientist, or whether someone who had not attended a leading institution such as MIT was qualified to conduct research.

AI could lower these barriers and unlock potential that previously went unrecognised. That is what I consider a seismic change: people who cannot code may be able to build software, and people without traditional research training may be able to contribute to scientific discoveries.

Journalists also have to confront these changes. I do not believe good journalists will be easily replaced by AI, but repetitive work that lacks independent thinking or creativity could certainly be replaced. I use AI tools intensively, but I still create all my own presentation decks.

The AI race between the United States and China has intensified, with a tendency towards decoupling at both the model and computing levels. How do you picture the future of China’s AI industry against this backdrop?

I am very optimistic about China’s AI development over the long term, while recognising that there are challenges in the short term.

Unlike many other technologies, AI will affect all of society and is difficult to confine to one place. It is inherently global. I agree with China’s view that AI should be a global public good. It is already too late to “decouple” China and the US in AI models and compute.

For one thing, China’s open models are approaching the capabilities of leading US proprietary models in certain areas. Disruptive breakthroughs may still emerge, and it is difficult at this stage to say who has an absolute lead. For another, China’s domestically developed computing chips have already reached a certain level of capability.

In an open market, Chinese computing hardware would struggle to compete with US products on price-performance and technological sophistication. But the US has disrupted that open market.

Under the current conditions, China’s computing hardware has to be sufficient for itself, and it is basically sufficient. US restrictions may therefore not be a catastrophic problem; instead, the restrictions helped accelerate the development of China’s chip capabilities.

China’s leading AI companies have been sticking to open-source or open-weight models, while closed-source remains the dominant format in the US. Why has there been such a divergence, and which approach will win?

We cannot simply apply the experience of traditional software – open-source versus closed-source – when it comes to AI.

Traditional open-source software offers an alternative in a market where commercial software has been relatively mature. AI is different: open and closed approaches have emerged side by side while the industry itself is still taking shape.

There is also a worldwide debate over whether some models described as “open-source” actually qualify, since they release only their weights. Applying the traditional software definition of “open-source” may not accurately describe what a model “opens”. I prefer the term “open-resource” to describe model openness.

Training each model requires substantial resources, including computing power and electricity. Some countries, even if they have the money, may not have access to sufficient electricity, computing capacity or talent.

Making a model open shares the benefits of those resources, allowing people who cannot afford to make the same investment themselves to use the resulting capabilities.

Open and closed models are not in a zero-sum contest, nor must one eliminate the other. Competition between them can help both advance.

A company that operates efficiently and has a viable business model can also grow while making its models openly available. Valuation considerations and business conditions may differ between China and the US, so openness may affect their companies’ valuations differently. DeepSeek, for example, has maintained a high valuation despite making its models openly available.

AI tokenises data. Are token prices low enough to maintain market demand while covering the computing cost? Will they drop further as technology develops? How will the token industry appear in the US and China given the gap in resource costs?

A key step in AI advancement is whether token prices can fall further. This is closely tied to the importance of data represented as tokens.

As I have said before, we should think about data in terms of tokens, and share and use data through tokens.

The cost of tokens reflects several factors: computing power, electricity, models and crucially, the value of the underlying data. It should not be understood simply as how much electricity or computing capacity has been consumed to generate the tokens.

Over time, electricity and computing expenses should account for a smaller share of token costs, while the value attributed to data should account for a larger share. I believe tokens will become increasingly affordable while still commanding a reasonable price that reflects the value of the data behind them.

China and the US may gradually diverge in their approach to tokens and their underlying business ecosystems.

If we look only at electricity and computing costs, their understanding of token value is likely to converge. Although these costs vary between locations, industrial production and circulation tend to keep them within a certain comparable range over time.

The real difference lies in the value of data. One important contribution of this wave of AI development is that it has shown us how significant data is as a factor of production.

Tokens should be cheap, but they are not yet cheap enough. Model training is generally expensive, and those costs must, in principle, eventually be spread across each token.

Token prices in China are relatively low for several reasons. The average cost of computing resources used to train models is lower. China’s electricity supply and renewable energy development also provide wiggle room, particularly when it comes to supporting concentrated model training. Compared with the US, China has less reason to worry about electricity constraints.

I am reluctant to call China’s electricity abundance an “advantage”. Having the wiggle room is one thing, whether it could be put to good use is another.

When it comes to AI and computing, semiconductors are always a focus. As the US retains tight control over chip exports to China, especially the most advanced units, home-made hardware has achieved breakthroughs in recent years. How do you evaluate the actual progress of the domestic chip sector?

We should not view China’s domestically developed chips narrowly as a response to the US refusing to sell advanced chips to China. Instead, we should also ask how chips made in China can serve users around the world.

I believe Nvidia deserves tribute. Without Nvidia, we might all still be exploring in the dark. It initially developed chips for gaming rather than AI. But through its sustained investment in computing power, it was well positioned when the AI industry emerged and helped explore a path for the entire industry.

Today, once we know the chips are serving AI, the difficulty actually drops a lot. AI chips can be categorised into two types of demand: training and inference.

We know that chips need to support model training and inference, and we can design and optimise them around clearly defined requirements. Their design requirements differ, allowing developers to make more targeted trade-offs between performance, efficiency and application needs.

New entrants no longer have to begin by figuring out what their chips could be used for. Nvidia undertook that exploration for the industry, lowering the barriers for those that followed.

Chinese chip developers know which requirements they are designing for and understand how training and inference work. That allows them to optimise chips effectively for tasks such as inference. I am fully confident in China’s ability to develop inference chips.

It is difficult to say what chips will look like in 10 or 20 years, because we do not know what AI will look like then. We must all continue to innovate and face the next set of challenges together.

I do not believe the concept of “overtaking on a bend” actually exists. Today, the US and China can jointly solve the challenges we are facing as humanity.

You founded Alibaba Cloud in 2009, when most people considered cloud computing an unattainable concept. What prompted you to stick to this decision?

Looking back at the industrial revolutions, it was the steam engine that set change in motion, and electricity that amplified its impact. The information technology revolution followed a similar pattern. If computing had remained confined to mainframes, it would have been like the industrial revolution stopping at the steam engine: its impact would never have become so widespread.

Cloud computing is a revolution in how computing power is delivered, much like the transition from steam power to electricity.

Once you truly believe in that direction, you have the motivation to persevere. I once thought that if I didn’t finish the work myself, others could have carried it forward. I was fortunate enough to see the cloud computing sector thrive.

The mobile internet’s explosive growth gave us tremendous support. Without it, cloud computing might have faltered. Then AI emerged, making the cloud’s importance even clearer. Nvidia’s massive chip sales to cloud providers are one example. I have always believed that cloud computing is an enduring business, not something ephemeral.

Building cloud computing, like developing AI, requires wholehearted conviction. If there is even the slightest weakness in that conviction – something that may not yet amount to doubt – difficulties can overwhelm you. It is like a crack in an aircraft’s surface that gradually widens under pressure.

An employee once told me that he worked hard, frequently stayed late and was willing to be “the last to leave” the company. My response was that being “the last to leave” is completely different from being “the last man standing”. The former ultimately leaves; the latter remains through the hardest times.

Another colleague who later left to start an AI company described the feeling perfectly. He said he felt like a group of us were holding hands trying to hug a massive tree, but we had no idea how thick the tree actually was, or when our hands would finally meet. The uncertainty was beyond imagination.

US hyperscalers have long maintained higher margins than their Chinese peers. Does this mean China’s cloud industry is smaller? How can China’s cloud computing market become bigger?

China’s cloud computing market is no smaller than that of the US, and it has technological capabilities, but the industry as a whole remains underdeveloped.

The US has Amazon, Microsoft and Google, as well as Oracle, which is increasingly focused on its cloud business. China has relatively fewer companies with comparable scale and cloud capabilities, with Alibaba Cloud being the most prominent.

I would like to see two more companies similar to Alibaba Cloud emerge, giving China at least three significant cloud providers operating at scale. That would help the industry flourish, strengthen economies of scale and create fuller competition. Alibaba Cloud has played a leading role, but one company cannot drive innovation across the entire market alone.

Competition does not have to mean suppliers continually undercutting one another on price. There are two kinds of competition. Destructive competition can ruin a market. Healthy competition encourages companies to improve their operations and efficiency, giving customers better prices and services while allowing providers to earn reasonable profits. Low prices don’t necessarily come at the expense of suppliers’ profitability.

Fuller competition can also spur creativity across companies. When others continually push you to improve, efficiency, service quality and profit margins can rise together.

Furthermore, in Chinese Premier [Li Qiang]’s government work report this year, the term “public cloud” was heavily emphasised. This was the first time it was used in such a report. In China, we used to constantly debate “public cloud” versus “private cloud”, which is not really helpful to industry development. It is a matter of public service versus proprietary networks. Computing must become a public service. No public service can exist if costs remain high, so it must scale to become cost-effective.

You have served as the director of Zhejiang Lab since 2023. The research institute focuses on AI and computing, including space computing. What drove you to pivot towards space computing?

The rationale is very simple: the problems facing Earth – climate change, extreme weather, geological disasters like landslides – can no longer be solved by only looking at them from the ground. You have to go to space to see the whole picture. Going into space is not just about exploring space itself, it is also about solving problems on Earth.

Short-term weather forecasting for aviation, for example, requires observations made by satellites in space to see the full picture.

When observations are made and data is collected in space, we also need to consider where the computing should take place. If a satellite collects data, sends it back to Earth for processing and then delivers the results to those who need them, that can delay time-sensitive work. It is like having to bring surgical instruments from a distant location when an operation is already needed.

Putting computing power in space would allow data generated there to be processed promptly. AI has made this need more urgent: there is already an abundance of data in space, but what once was missing was the ability to analyse and use it in the right place. Now with AI, we are able to process all this data, hence the urgency to build space computing facilities.

There is a substantial gap between satellite observations and the use of the resulting data. AI could help process and use more of it.

Cameras and other observation payloads on some remote-sensing satellites sit idle roughly 90 per cent of the time because they cannot continuously collect and transmit everything. Meanwhile, roughly 90 per cent of data collected in space is lost because bandwidth is insufficient to transmit it back to Earth.

Cloud computing mainly addresses computing needs on Earth. Space computing has its own requirements and development logic. It should not be understood simply as moving terrestrial data centres into orbit or finding another location to address electricity constraints on Earth.

We have to be very clear about how our approach differs from Elon Musk’s. Space computing is not simply moving ground-based data centres up to space. If Musk says he wants to put a data centre in space, today he is the only one who can do it because he is the only one with that kind of launch capacity.

We are not doing that. We are solving the fundamental issue of processing space-native data directly in space.

The first questions are how data generated in space should be processed and how it can be turned into timely, usable results. I have also discussed this in Geneva. “Putting computing power in space” and “sending data centres into space” may sound similar, but the approach is different.

Sending data centres into space may seem like a daunting task. But space computing involves much more than that, because processing space data requires an entire value chain, with opportunities for a wide range of participants.