Jacky Cao Reiterates Alibaba's AI Strategy at VivaTech

Jack Ma, Chairman of Alibaba Group. Image processed by AI.

At this year's VivaTech conference, Alibaba Chairman Joe Tsai systematically outlined Alibaba's long-term AI vision during a "fireside chat," marking his second public reflection on Alibaba since the Yale summit in late May.

From a macro perspective, we are fully committed to AI—the logic is straightforward.

Cai Hongxin stated that global GDP exceeds $100 trillion, and at least half of it will come from contributions by human intelligence and human productivity. “That $50 trillion is the total market for AI—far larger than any company’s IT budget and bigger than the entire software market.”

Everyone is talking about going all-in on AI, and Cainiao is no exception—he has summarized Alibaba’s strategic layout as encompassing everything except the energy layer, including chips, cloud infrastructure, models, and applications.

We primarily focus on four layers, but we do not engage with the underlying energy layer, as China’s energy efficiency is high and costs are low.

In Cai Jingxin’s view, the nearly full-stack approach stems from the uncertainty of the future, as no one can currently define where the ultimate value will be anchored—whether in chips, cloud infrastructure, or the model layer. “We chose to participate comprehensively, so that no matter where the value ultimately settles, we’ll be there.”

More aggressive than Alibaba are the U.S. cloud giants, which have nearly fully invested in infrastructure, with combined capital expenditures of $800 billion by 2027—a move criticized by short sellers as a “bubble.” Cai Jingxin not only rejects the bubble theory but also emphasizes that Chinese companies must increase their investment in infrastructure.

"The investment numbers are indeed impressive," said Charlie Cheung. "We need to look back at the total market size of $50 trillion—that’s what gives us reason to stay optimistic."

When discussing open source, Cai Jingxin first mentioned the Trump administration’s recent halt of Anthropic’s most powerful model, stating that this is precisely the consequence of putting all your eggs in one basket. In his view, the models from Google, OpenAI, and Anthropic have all become closed-source, and now it is Chinese companies that are pursuing the open-source path.

You truly cannot base trust on the assumption that a third-party government will never act against your interests.

Below is the condensed version of the Charlie Cheung interview:

Question: Alibaba has undergone significant changes over the years, such as its achievements in open-source large models. However, many people still think of you merely as a B2B and B2C platform. Could you share the overall development journey of the group?

Cai Jingxin: When Alibaba started in 1999, it was indeed a B2B platform. The idea at the time was simple: to bring China’s small manufacturers and trading companies online so they could wholesale their products to the world. Later, we entered the B2C space with Taobao, which is now China’s largest consumer e-commerce platform.

Question: How many consumers is this service designed for?

Cai Chongxin: 820 million Chinese consumers, and this platform helps European companies and brands sell approximately €30 billion worth of goods to Chinese consumers each year. But the story doesn’t end there—we’ve also made significant investments in AI and cloud technology.

We began investing in cloud technology 17 years ago, but it was out of necessity. At the time, our e-commerce business generated massive amounts of data daily, and relying on third-party databases and storage solutions meant we’d end up giving away all our profits to technology vendors. So we decided to build our own proprietary technology to manage this data—and that’s how our cloud business began.

From a macro perspective, we are fully committed to AI, and the logic is straightforward.

If you ask me how big the AI market is, I’d say it’s far larger than any company’s IT budget and much bigger than the software market. That’s because AI is essentially producing human intelligence and productivity. Today, global GDP exceeds $100 trillion, and at least half of that—$50 trillion—is tied to human productivity and human intelligence. That’s the total addressable market for AI. So we must commit fully.

Question: Do you really believe AI can boost productivity? Many people have invested a lot of money but haven’t seen results yet.

Caesar Zhang: Many CEOs will tell you that engineers are consuming a large number of tokens, and costs are rising. But I want to say that we are on the brink of a true explosion in productivity.

Take our company as an example: some engineers are super users of AI—they don’t just use programming tools to perform their core duties, but also explore a wide range of new applications with them. Give engineers a toy, and they’ll discover even more uses, often without realizing the company is footing the bill for these expenditures. This is the current reality.

But deep down, I’m convinced that this is more of a belief—that artificially created intelligent units can add value to human intelligence. It’s almost a faith; I don’t want to convince you that this will definitely happen, but we ourselves hold this conviction firmly.

Question: Returning to Alibaba’s strategy, which layer of AI are you investing in the most—infrastructure, models, or cloud services?

Cai Hongxin: We primarily focus on four layers, but we do not engage with the underlying energy layer, as China’s energy efficiency is high and costs are low.

We truly entered the field starting from the chip level, which is the first layer; the second layer is the infrastructure layer, corresponding to our cloud business; the third layer is the model layer, such as Qwen, which is now one of the most popular open-source models globally; and the fourth layer is the application layer, where we have a complete digital life ecosystem—including e-commerce, food delivery, local services, travel, maps, and more—where AI capabilities can be directly integrated to serve users.

The benefit of doing this is that we don't bet on a single赛道.

Today, people see pure model companies valued highly, as if all the value resides in the model layer. But over the next five to ten years, it’s unclear where the true value will ultimately settle—whether in chips, cloud infrastructure, models, or applications. We’ve chosen to participate across all layers, ensuring we’re present no matter where the value ends up.

Question: When it comes to AI infrastructure, given the massive investments being made, do you think there’s a bubble? Do we really need this much computing power, especially since some models are more efficient and require fewer resources?

Charlie Cheung: I don’t think it’s a bubble. The scale of investment is astonishing. Just looking at the major U.S. hyperscale cloud providers, the combined capital expenditures of the top four or five companies next year will exceed $800 billion, and could surpass $1 trillion the year after. Such massive levels of investment naturally raise concerns about potential overcapacity.

But when we look back at the total market size of $50 trillion, that’s reason to stay optimistic.

Moreover, in China, our investment in AI infrastructure and supply chains is still insufficient; theoretically, all Chinese companies should increase their investments. Of course, we won't reach the investment levels of U.S. hyperscale vendors, but our investment magnitude is already very substantial.

Question: Why hasn't it reached the level of the United States?

Cai Jingxin: Sometimes, we are constrained by capital, depending on how much free cash flow we can generate. Fortunately, Alibaba is one of the few companies with a core e-commerce business that generates approximately $25 billion in free cash flow annually, enabling us to invest in AI. So, we are in a relatively strong position.

Question: Does e-commerce still account for 80% to 85% of Alibaba's total revenue?

Cai Jingxin: Yes, e-commerce platform revenue still accounts for over 80%, generating stable cash flow that enables us to invest in the future, primarily in AI and cloud computing.

Question: Qwen is an open-source model. Which customers are you primarily targeting, and how do you assist them?

Cai Chongxin: Over the past few weeks, I’ve spoken with numerous corporate executives, CEOs, and scientists in Europe, and the word that has come up most frequently is “sovereignty.”

But what is sovereignty?

Ask ten Europeans, and you might get twelve different answers. For me, the core comes down to two things.

First, technological independence. Everyone is concerned about the risk of a "one-click shutdown," fearing excessive reliance on a country's technology, which could be turned off at any moment. We saw a real-life example of this just in the past few days.

Second is data privacy. People want to use AI technologies but wish to retain full ownership of their data and use it within their own environment, building a firewall to protect it.

I believe open source is precisely what solves these two issues. It is essentially free software that you can download to your own data center, or even onto a laptop. At that point, it has no connection to the original manufacturer—we can’t even figure out how to charge for it. This achieves independence.

More importantly, by using open-source models, you can further train, fine-tune, and post-train them with your own data, keeping the entire process and all data fully confidential behind your firewall. This is crucial for European companies.

But I want to emphasize that open source is not a cure-all or the only path, but it is a practical way to achieve a certain degree of sovereignty.

Interestingly, today the open-source movement is actually being driven by Chinese companies, while major U.S. players have closed off their models. They want you to access their systems via API, yet you have no idea where your data goes. When you chat with a bot, your most private questions and thoughts are fed into their data pools to further train the models—yet the flow of your data remains completely opaque to you.

Question: To be honest, European sovereignty is now a major concern—we’ve just realized how overly dependent we are on U.S. technology. I acknowledge the benefits of open source, but I still worry about the risk to Europe of being cut off from access to models in the future.

Cai Hongxin: You're right—this concern cannot be completely eliminated. Simply put, you can't base your trust on the assumption that a third-party government will never act against your interests. But the issue is that now all your eggs are in one basket.

Why not choose the second basket and keep your eggs separated? Even if Europe may eventually develop its own basket in the long term, at least right now you have two baskets.

A: That's true. How do you collaborate with German companies, and what do you help them with?

Cai Hongxin: These German manufacturing companies are very interesting. In the Chinese market, they are all customers of Alibaba Cloud. We collaborate with them in the manufacturing sector, covering areas such as design, testing, and quality control.

I believe this will become a very important field in the future. Currently, most AI applications are either consumer products like ChatGPT or tools like Copilot designed for programmers and knowledge workers. However, in the future, these manufacturing companies will be extremely valuable, as they have accumulated unique, high-quality data through their production processes—data that can be used to train proprietary models and optimize manufacturing workflows.

We collaborate with companies such as BMW, Siemens, and Bosch. Last week, I attended the Bosch ConnectedWorld conference, where they are using AI to develop advanced driver assistance and autonomous driving technologies, requiring substantial computing power.

A lot of interesting things are happening in manufacturing.

Question: Can I understand it this way: U.S. export controls on advanced chips have instead created opportunities for you?

Cai Hongxin: You can understand it that way. There are two paths here:

First, they can directly adopt our open-source model and deploy it on their own infrastructure, such as data centers. However, our infrastructure was developed in close integration with the model, offering high efficiency and enabling customers to train models effectively. If they use our open-source model, they can still purchase computing power from us—there is a symbiotic relationship between the model and the infrastructure. This is one path forward.

Another option is that a number of reasoning platform companies have recently emerged, offering users a choice of multiple models. You don’t necessarily have to use Qwen—as long as there is an agreement between the model provider and the platform, and the weights are made available in a private environment, customers can access these models through the platforms.

Question: Let me ask a more philosophical question. What are your thoughts on the balance between AI, large language models, and humanity—or even the future of human nature? What state will humanity be in over the next decade?

Charlie Cheung: Today, I was chatting with my colleagues at the Paris office. We just moved into our new office on the upper floors of a beautiful building. I looked out the window and saw a café where people were sitting outside, enjoying coffee and soaking in the lovely weather.

I pointed out the window to my colleague and said, "This is the future of AI."

You might think they’re sipping coffee and having fun, as if they aren’t working—but in reality, they’ve deployed agents to work on their behalf. While you sleep, agents continue working for you. Imagine this boost in productivity: you have someone working for you 24/7, every day of the week.

Question: This is similar to the idea held by some people in Silicon Valley—that many people won’t need to work because agents and robots will do it for them.

Cai Hongxin: I believe this will undoubtedly free up people’s time to enjoy life, spend time with family, and engage in more entertainment. This is also one of the reasons I place such strong emphasis on live entertainment. When people spend less time in the office, where will they go? They won’t just stay at home—they’ll want to attend concerts, watch football matches, or catch basketball games.

Question: Chinese people are known for their diligence. Chinese engineers, even with the advent of agents and AI, still work long hours.

Cai Hongxin: There will always be people who work harder than others, but I believe most people internally wish to enjoy life a bit more and spend more time with their families.

This article is from the WeChat public account "Tencent Technology," authored by Su Yang.