At the 2026 Apsara Conference held in Hangzhou on September 22, Alibaba Group CEO Eddie Wu delivered a keynote speech outlining his vision for the era of "machine intelligence." He stated that machines are becoming the primary drivers of thought, intelligence is transforming into a scalable commodity, and the total volume of machine thinking will eventually exceed that of humans by more than 1,000 times. The last comparable shift occurred during the Industrial Revolution, when power was converted into a scalable commodity. Humanity successively invented the steam engine, the internal combustion engine, and electricity, and built modern industry on this foundation.
Wu argued that the defining product of the machine intelligence era has not yet emerged. Today's AI coding resembles the electric light bulb of 1882—it replaces existing workflows but is not yet sufficient to create a new era. Helping humans write code or generate reports is only the initial stage of machine intelligence. When the supply of machine intelligence becomes infinitely abundant, its significance to human society will extend far beyond simply substituting for existing mental labor. This transformation will be more profound than the Industrial Revolution.
Like the electrical age, which later gave rise to air conditioning, washing machines, refrigerators, and countless other inventions, the machine intelligence era will also spawn new products and breakthroughs. However, regardless of how many new inventions follow, the first necessary step is to build enough power stations and lay a sufficiently extensive power grid. Around 1900, many household appliances were already entering homes, but the entire world's annual electricity output at that time would be exhausted in just two hours by today's standards. The machine intelligence era requires similarly robust infrastructure.
Wu emphasized that the three foundational pillars of the machine intelligence era are AI models, AI chips, and AI cloud—the prerequisites for scalable machine thinking. Around these pillars, Alibaba announced its latest progress and goals. First, in AI models, the Qwen team has made progress in recursive self-improvement (RSI), with future models planned to scale to 5-10 trillion parameters, aiming to handle more complex and longer-horizon tasks and advance toward artificial superintelligence (ASI). Second, in AI chips, T-Head has unveiled its most powerful domestic AI chip, the Zhenwu V900, delivering three times the computing power of the M890. A single cluster can now scale to 500,000 cards, and T-Head's annual AI chip shipments are expected to rise significantly. Third, in AI cloud, customer demand for AI capabilities is extremely strong, with medium- and long-term demand far exceeding current supply. Alibaba will work with all partners to invest fully in AI infrastructure, targeting global data center capacity operated by Alibaba Cloud to exceed 20 gigawatts by 2032.
Below is the full text of the speech: Welcome to the 2026 Apsara Conference. Another year has passed, and over the past 12 months, AI technology has advanced faster than anyone anticipated, with development cycles accelerating. A year ago, we judged that AGI was just the starting point, and that AI would continue evolving toward ASI, capable of self-iteration. Since then, everyone has grown accustomed to moving from vibe coding to vibe working. AI has achieved major breakthroughs in long-horizon tasks, and the technical path toward autonomous evolution has become increasingly clear. As AI unlocks greater capabilities and penetrates more scenarios, we are witnessing a profound shift: machines are becoming the main actors in thinking, and intelligence is becoming a scalable commodity.
The last time we saw such a change was when the Industrial Revolution turned power into a scalable commodity. Humans invented the steam engine, the internal combustion engine, and electricity, and built modern industries around them. This time, it is thinking that is being industrialized. In the past, the amount of thinking a complex problem could receive was limited by the total human intellect and how value was distributed. Today, machines have broken those limits. As ASI gradually becomes a reality and AI infrastructure expands, thinking will become a commodity supplied at scale, just like power. In the future, we will enter the era of machine intelligence. What I want to discuss today is our thinking about this era and Alibaba's strategic choices.
First, I want to clarify a concept. People are used to calling AI "artificial intelligence." The word "artificial" in English implies something man-made, like synthetic diamonds or artificial leather. When discussing AI, we tend to understand it through human capabilities, hoping to create AI that thinks like humans. In the past, we always asked: "Does it resemble a human?" "Can it think like one?" The Turing test once became the standard for measuring AI. Let us revisit the Industrial Revolution. Early steam engines and internal combustion engines replicated what humans and horses could do, such as pumping water from mines, weaving cloth, and hauling loads. That is where the term "horsepower" comes from. But soon, the power of machines far exceeded mere physical labor. No number of people or horses could make a train speed along, let alone send airplanes and rockets into the sky. Machine power was not simply "artificial physical labor"; it created an entirely different species of capability. The depth, breadth, and density of machine power enabled humanity to accomplish things we never even imagined possible. Likewise, today's machine intelligence is not a replacement for human intelligence but an entirely different kind of entity. AI helping people write code or generate reports is only the early stage. When the supply of machine intelligence becomes infinitely abundant, its impact on human society will go far beyond replacing mental labor. This change will be even more profound than the Industrial Revolution.
I have two judgments about the future machine intelligence era. My first judgment: in the future, the total volume of thinking supplied by machines will be more than 1,000 times greater than the total thinking volume of all humans. This has already happened in the physical world. From the invention of the steam engine, which opened a new era, to today, machines now bear 99.9% of all physical work. In the future, human thinking will continue to grow, but machine thinking will expand at a far faster pace and is expected to assume 99.9% of all cognitive work. We can reason through both supply and demand. On the supply side, top-tier human thinking is extremely scarce. The number of world-renowned scientists and experts is very limited. Training a top scientist or expert requires decades of study and practice, plus talent and luck. As a result, many fields of research have long suffered from insufficient intellectual investment. For example, progeria in newborns affects only dozens of new cases worldwide each year, with only a few hundred recorded patients globally. Such rare diseases struggle to attract adequate research resources. Under the commercial logic of the existing pharmaceutical industry, few companies are willing to spend a decade or more and invest enormous sums to develop a new drug for just a few hundred patients. But when AI reaches the level of expertise needed for professional research, as long as computing power is plentiful and cheap, every niche field can deploy millions of agents to continuously study, simulate, and verify, exploring different approaches. Research directions that previously suffered from insufficient investment will, for the first time, gain access to massive intellectual supply. Top-tier thinking will shift from being a luxury to a commodity that can be supplied at scale. Imagine the change when every niche field has millions of AI scientists and experts working around the clock to innovate and solve problems. On the demand side, in the industrial age, how many goods humans could consume depended on population size. Coffee needs someone to drink it, cars need drivers, phones need users. The physical goods one person can consume in a lifetime are finite. But as AI capability grows stronger, this relationship will fundamentally change. Imagine a future scenario: we ask AI to build a spaceship to Mars. AI can break down such an ultra-complex, long-horizon task into tens of millions of subtasks, with millions of agents working continuously without rest until completion. Humans only need to propose an idea and a goal, and they can mobilize enormous intellectual resources. Every super-individual can have a ten-thousand-fold intellectual leverage. The consumption of thinking will no longer be proportional to population size but will scale with the depth of machine thinking and the scale of supply. The total volume of machine thinking will vastly exceed that of humans. Today, machine thinking accounts for less than 3% of human thinking. If future machine thinking is 1,000 times human thinking, a simple projection suggests machine thinking has at least tens of thousands of times of growth ahead.
My second judgment: the representative product of the machine intelligence era has not yet appeared. Electricity's earliest use was to replace kerosene lamps and candles for lighting. In 1882, Edison built the Pearl Street Station, lighting roughly 400 electric lamps in the surrounding neighborhood. Interestingly, in the early days, Edison gave away electricity when selling light bulbs, much like how agents are bundled with tokens today. Many world-changing things came later. Air conditioning appeared in 1902, followed by washing machines and refrigerators entering households. The first computer did not arrive until 1946, more than sixty years after those 400 lights on Pearl Street. Of course, the pace of the AI era will be much faster than that. Today's AI coding may be like the early light bulb of the machine intelligence era—it replaces existing work in human society, such as traditional programming and report writing. AI is now entering the broader office market, but merely replacing existing jobs cannot create a new era. Today, it is difficult to predict what new products and inventions will emerge once machine intelligence explodes, just as standing under the electric lights of 1882, people could hardly imagine the full picture of the electrical age. But no matter how many new inventions come later, the first step is always to build enough power stations and lay a sufficiently wide grid. Around 1900, many appliances were already entering homes, but the entire world's annual electricity output at that time would only last two hours by today's standards. The machine intelligence era likewise requires powerful infrastructure construction. The three pillars of the machine intelligence era are AI models, AI chips, and AI cloud. They are the prerequisites for the scalable supply of machine thinking. Only with sufficiently complete and advanced AI infrastructure can a steady stream of AI products and inventions emerge at the application layer. Judging from the history of electricity, the future demand for AI infrastructure relative to today's scale is almost limitless. Alibaba will firmly invest in infrastructure represented by AI models, AI chips, and AI cloud. This is our long-term strategic choice.
The first pillar is AI models. Last year, we talked about AI moving from autonomous action to self-iteration. This year, we see the technical path to ASI becoming clearer. From coding and office work to scientific research, models are taking on more and more real-world tasks. These tasks and their feedback make models stronger. AI researchers are gradually mapping out a concrete path to superintelligence—recursive self-improvement, or RSI. Models discover their own weaknesses from real feedback, design their own experiments, build their own data, evaluate results, and loop to drive their own evolution. Currently, our Qwen team is exploring RSI and has made some progress. The Qwen team is advancing research on model architecture and data optimization, planning to train new models with 5-10 trillion parameters, with the goal of handling more complex, longer-horizon tasks and moving toward ASI. If the direction of LLM base models is to build highly intelligent, self-evolving brains, multimodal capability is another vital direction for machine intelligence development. We believe a sufficiently smart brain alone is not enough; AI also needs perception and interaction capabilities. Humans communicate through sound, images, expressions, and actions, and express intent and emotion through these channels. Models need stronger multimodal capability to understand and express this information like humans, aligning with human culture, aesthetics, and values. In the future, people should be able to converse with AI as naturally as with another person, without facing a complex interface. Only then can AI serve humanity better. Therefore, integrated understanding-and-generation multimodal models are another key research direction for us.
The second pillar is AI chips. If tokens are the electricity of the AI era, chips are the generators. In the future machine intelligence era, demand for tokens will have almost no upper limit, so chips must continuously improve performance and expand supply. Today, T-Head is making comprehensive arrangements for data center chips: the Zhenwu series covers GPU chips, the Yitian series covers CPU chips, the Panmai series covers smart network interface cards, and there is also the ICN interconnect chip. Core chips for building ultra-large-scale AI clusters are now fully covered. At the same time, we are jointly optimizing chips, servers, supernodes, networks, models, and inference systems to continuously improve the token production capacity and efficiency of the entire AI cluster. Alibaba's self-developed M890 AI supernode has successfully supported efficient inference for large models with over 2 trillion parameters, making us one of the very few companies globally with such capability. Starting this quarter, Alibaba Cloud has begun large-scale deployment of our AI supernodes. Today, we will release the next-generation AI chip, Zhenwu V900. This is currently the most powerful AI chip in China in terms of computing performance, delivering performance three times that of the M890. A single cluster built on it can scale to 500,000 cards, supporting training and inference for frontier models. Given the maturity of T-Head's chip product line and widespread customer adoption, we expect T-Head's annual AI chip shipments to rise significantly. In the future, intelligence should be everywhere. AI will further integrate into desktop and mobile endpoints. On the desktop side, our open-source Qwen 27B model is the most popular model among developers globally. We will continue to optimize such models to support local deployment by developers and enterprises. On the mobile side, today we officially release Qwen Intelligence, providing AI solutions for partners so that mobile devices can handle various complex tasks.
The third pillar is AI cloud. Chips produce tokens, and the cloud delivers tokens wherever they are needed. Models require continuous training and inference, agents need to execute long-running tasks, and AI-developed software will multiply. All of these will become "residents" of the cloud, continuously consuming computing power. We need to build an AI cloud designed for the sustained operation of models and agents, based on these new workloads. Today, we barely notice the power grid itself. Anywhere you are, plug into a socket and electricity flows. Without the grid and sockets, life as we know it would be unbearable for even a day. Using machine intelligence in the future should be just as convenient. Any device connected to the cloud should have instant access to machine intelligence anywhere and anytime. To achieve this, we need to build a global, highly efficient super-network that jointly optimizes GPUs, CPUs, networking, storage, databases, and various tools. The more complete this network becomes, the more users and agents will join, creating scale and network effects. Customer demand for AI is currently extremely strong, and we are fully deploying AI computing capacity to meet it, which is why Alibaba Cloud's revenue continues to accelerate. However, we see that medium- and long-term industry demand far exceeds our supply capacity. The current global shortage in AI data center supply chains limits the speed of our computing capacity growth. To this end, Alibaba will work with all partners to invest fully in AI infrastructure and build an AI cloud for the machine intelligence era. Our goal is to have Alibaba Cloud's global data center capacity exceed 20 gigawatts by 2032 to meet rapidly growing AI demand.
Finally, when machines become the main drivers of thinking, what changes will unfold? We can look back at history. The Industrial Revolution shifted more and more physical labor onto machines. After being freed from heavy manual labor, humans developed sports such as football and basketball. The modern Olympics, the World Cup, and the NBA are all products of the post-Industrial Revolution era. History shows this pattern repeating. The rise of printing triggered an explosion in cultural creation. After the camera appeared, painting did not disappear—painters stopped pursuing mere likeness, which actually gave birth to Impressionism and abstract art. The phonograph did not put singers out of work; instead, it turned music into the most widespread popular art. When machines take on more of what humans must do, people gain time to do what they truly want to do. In the future, machine intelligence will open up even greater space for human curiosity and creativity. Today, it is hard to describe exactly what the future will look like. Just as a farmer working in a field 300 years ago could hardly imagine that people would one day walk into a place called a "gym" and pay to perform physical labor. Last year, we said that the more powerful AI becomes, the more powerful humans become. Today, I still hold that view. This is also the meaning of our full commitment to the machine intelligence era: leave the heavy tasks to machines, and leave time, creativity, and the appreciation of beautiful things to humans. All of this is just beginning now. I wish everyone a productive and enjoyable Apsara Conference. Thank you.
Comments