Alibaba CEO: Large models are the next-generation operating system, Super AI Cloud is the next-generation computer

On September 24 in Hangzhou, at the 2025 APSARA Conference, Alibaba Group CEO and Alibaba Cloud Intelligence Chairman & CEO Eddie Wu (吴泳铭) delivered a keynote speech. Key points include:

• AI chatbots are the fastest-adopted technology feature in human history.

• Token consumption is doubling every two to three months.

• Over the past year, global AI investment totaled USD 400 billion; over the next five years, cumulative global AI investment will exceed USD 4 trillion.

• The goal of AGI is to free humans from 80% of routine work; ASI, as a system surpassing human intelligence, will create “super scientists” and “super engineers.”

Three stages toward ASI:

1. Emergence of intelligence — learning from humans. Large models acquire generalized intelligence, approaching human peak performance across disciplines after just 2–3 years.

2. Autonomous action — assisting humans. AI uses and builds tools in the real world, enabling large models to connect with digital tools and accomplish real-world tasks. Rapid penetration into logistics, manufacturing, software, business, biomedicine, finance, and coding.

3. Self-iteration — surpassing humans. Two key factors: access to more raw data and self-learning, allowing models to learn autonomously and adapt in real time.

Two core judgments:

1. Large models are the next-generation operating system. They allow anyone to create unlimited applications with natural language. In the past, high development costs limited software creation to a few high-value scenarios; in the future, every end user will be able to meet their needs at low cost through large models. Models will run on all devices, with persistent memory and adaptability.

2. The Super AI Cloud is the next-generation computer. Large models run as the OS on AI cloud. Computing is shifting from CPU- to GPU-centric. In the future, there may be only 5–6 global supercloud platforms. AI will replace energy as the most important commodity; tokens will be the electricity of the future. Alibaba Cloud positions itself as a full-stack AI service provider, offering world-leading intelligence and a global AI cloud network.

Alibaba Cloud highlights:

• Tongyi Qianwen has been downloaded over 600 million times globally, with more than 170,000 derivative models created.

• Alibaba Cloud operates China’s No.1 and one of the world’s leading AI infrastructures and cloud networks, among the very few globally with full vertical integration of hardware and software.

• The company is advancing a RMB 380 billion (≈ USD 52B) three-year AI infrastructure plan, with further investment planned.

• By 2032, Alibaba Cloud projects its global data center energy consumption will increase tenfold.

Notably, today Alibaba also unveiled Qwen3-Max, our largest and most powerful open-source model to date. Following the release of the Qwen3-2507 series, we are excited to launch Qwen3-Max — the biggest and most capable model we have ever built.

Qwen3-Max: Bigger is Better

Currently, the preview version of Qwen3-Max-Instruct ranks third on the LMArena text leaderboard, surpassing GPT-5-Chat. The official release further enhances coding and agent capabilities, achieving industry-leading performance across a comprehensive range of benchmarks, including knowledge, reasoning, programming, instruction-following, human preference alignment, agent tasks, and multilingual understanding.

We invite you to experience Qwen3-Max-Instruct via the Alibaba Cloud API, or try it directly on Qwen Chat. Meanwhile, Qwen3-Max-Thinking, still in training, has already demonstrated extraordinary potential. With tool use enabled and additional compute allocated during testing, this “thinking” version has achieved 100% accuracy on challenging reasoning benchmarks such as AIME 25 and HMMT. We look forward to officially releasing this version to the public in the near future.

At today’s APSARA Conference, Alibaba officially announced a partnership with Nvidia in the field of physical AI. Alibaba’s AI platform PAI will integrate Nvidia’s Physical AI software stack to support end-to-end capabilities ranging from data preprocessing and synthetic data generation to model training/evaluation, robotic reinforcement learning, and simulation testing. The collaboration will bring Nvidia’s Isaac Sim, Isaac Lab, Cosmos, and other software tools and datasets into the Alibaba Cloud PAI ecosystem. This integration is expected to help enterprise users shorten development cycles for embodied intelligence, robotics, assisted driving, and other AI applications that interact with the physical world.

At the same conference, Alibaba Cloud also announced a new round of global infrastructure expansion: it will establish new cloud regions in Brazil, France, and the Netherlands for the first time, and expand data centers in Mexico, Japan, South Korea, Malaysia, and Dubai to better serve growing worldwide demand for AI and cloud computing. Alibaba Cloud currently operates 91 availability zones across 29 regions globally, and this network will expand further through the new initiatives.

Alibaba’s ambitions in cloud and AI may also help explain its heavy investment in developing its own chips. Its recent progress in AI chips has drawn lots of attention. Reuters reported that both Alibaba and Baidu have begun using internally designed chips in their AI training workloads as a partial substitute for Nvidia GPUs. Alibaba has been deploying its self-developed chips on smaller models since early 2025. According to reports from the South China Morning Post and others, T-Head, Alibaba’s semiconductor design arm, has developed an AI processor that state broadcaster CCTV compared with Nvidia’s H20, claiming it had reached a “comparable level.”More recently, a data center under construction by China Unicom in Qinghai was reported to be using domestic AI chips, with about 72% supplied by Alibaba’s T-Head — suggesting Alibaba’s chips have already entered large-scale deployment.

---

Full transcript of Eddie Wu’s keynote speech:

The Path Toward ASI

Before starting my speech, I want to express my heartfelt thanks to all the developers who have supported not only China’s technology industry, but the global tech community as well. Today marks the 10th anniversary of the Yunqi Conference, which began as Alibaba Cloud’s developer conference. It is developers who have driven the growth of cloud computing, AI, and the technology sector in China and worldwide. So before anything else, let me extend my highest gratitude to the developer community.

Today, the world is at the beginning of an intelligence revolution driven by artificial intelligence. Over the past centuries, the Industrial Revolution amplified human physical power through mechanization, and the Information Revolution amplified our ability to process information through digitalization. This time, the Intelligence Revolution will go far beyond what we can imagine. Artificial General Intelligence (AGI) will not only amplify human intellect, but also unlock human potential, paving the way for the arrival of Artificial Superintelligence (ASI).

In the past three years, we have clearly felt the speed of this transformation. In just a few years, AI has advanced from high-school-level capability to PhD-level performance, even winning gold medals at the International Mathematical Olympiad. AI chatbots have become the fastest-adopted technology in human history, and AI is spreading across industries faster than any technology that came before it. Token consumption doubles every two to three months. Over the past year alone, global investment in AI has exceeded $400 billion, and cumulative investment over the next five years is projected to surpass $4 trillion — the largest investment in computing power and R&D in history. This will inevitably accelerate the development of stronger models and the penetration of AI applications.

Achieving AGI — an intelligent system with human-like general cognitive abilities — now appears inevitable. But AGI is not the end of AI’s journey; it is a new beginning. AI will not stop at AGI — it will move toward ASI, a form of intelligence that surpasses human cognition and is capable of self-iteration and evolution.

The goal of AGI is to free humans from 80% of routine work so that we can focus on creation and exploration. ASI, as a system that goes beyond human intelligence, could create “super scientists” and “full-stack super engineers.” At unimaginable speed, ASI could solve today’s unsolved problems — from breakthroughs in medicine and new materials, to sustainable energy, climate solutions, and even interstellar travel. ASI will drive exponential technological leaps and lead us into an unprecedented era of intelligence.

We believe the path toward ASI will unfold in three stages:

Stage One: “Emergence of Intelligence,” characterized by learning from humans.
The development of the internet over the past few decades has laid the groundwork for this stage by digitizing nearly all of humanity’s knowledge. This body of language and text represents the collective knowledge of humankind. Based on it, large models first developed generalized intelligence — the ability to engage in natural dialogue, understand human intent, answer questions, and gradually build multi-step reasoning capabilities. Today, AI is approaching top human performance across disciplines, such as gold-medal level in the International Mathematical Olympiad. AI is increasingly capable of entering the real world, solving real problems, and creating real value. This has been the main theme of recent years.

Stage Two: “Autonomous Action,” characterized by assisting humans.
At this stage, AI goes beyond language interaction and gains the ability to act in the real world. With human-defined goals, it can break down complex tasks, use and create tools, and autonomously interact with both the digital and physical worlds, producing profound real-world impact. This is the stage we are in today.

The key to this leap is that large models have acquired tool-use capability — the ability to connect with all digital tools to accomplish real-world tasks. Just as humanity’s accelerated evolution began with the creation and use of tools, large models are now able to do the same. Through tool use, AI can call external software, interfaces, and even physical devices to carry out complex real-world tasks. Because of this, AI is rapidly permeating nearly every sector — from logistics and manufacturing, to software, business, biomedicine, finance, and scientific research — vastly enhancing human productivity.

Secondly, the advancement of large models’ coding capabilities will enable humans to tackle more complex problems and digitize more scenarios. Current agents are still in their early stages, mostly handling standardized, short-cycle tasks. For agents to take on more complex, long-term missions, the critical factor is the coding ability of large models. With autonomous coding, agents can in theory solve infinitely complex problems — functioning like engineering teams that understand complex requirements and independently complete coding and testing. Enhancing coding capabilities is therefore a necessary step on the path to AGI.

In the future, natural language will become the source code of the AI era. Anyone will be able to create their own agent simply by using their native language to describe what they need. The AI will then be able to write logic, call tools, build systems, and carry out nearly all digital tasks — while also operating physical devices through digital interfaces. One day, there may be more agents and robots than the world’s human population, working alongside us and exerting a tremendous impact on the real world. In this process, AI will connect to most real-world scenarios and data, creating the conditions for future evolution.

Next, AI will enter the third stage — “Self-Iteration,” characterized by surpassing humans. This stage rests on two key elements:

First, connecting to the full spectrum of raw data from the real world.
AI’s fastest progress so far has been in content creation, mathematics, and coding. These domains share a unique trait: the knowledge is entirely human-defined and expressed in text, allowing AI to fully understand the raw data. But for most other domains and the broader physical world, today’s AI has access mainly to knowledge summarized by humans — second-hand, distilled data — and lacks wide, direct interaction with raw physical-world data. This imposes limits. To go beyond human capabilities, AI must directly access more comprehensive, raw data from the physical world.

For example, when a car company CEO plans next year’s product, decisions are usually based on endless rounds of user surveys or internal brainstorming about what features to add or keep. For AI, this remains difficult because the data it gets is largely second-hand survey information. But if AI could directly access all the underlying data of the vehicle, the car it designs could far surpass what human teams could achieve through countless discussions. This is just one example. The broader physical world is far more complex than what human summaries alone can capture. Teaching AI only human-curated rules and patterns is not enough. Like the evolution from rule-based to end-to-end learning in autonomous driving — where cars learn directly from raw sensor data — AI must continuously interact with the real world, gather richer, more authentic, real-time data, and in doing so, uncover deep patterns beyond human cognition. Only then can it create intelligence more powerful than ours.

Second, self-learning.
As AI penetrates more real-world scenarios and absorbs more physical-world data, models and agents will grow stronger and gain the ability to build their own training infrastructure, optimize data pipelines, and upgrade model architectures — enabling true self-learning. This will be a critical milestone in AI’s evolution.

Future models will learn continuously through real-world interaction, acquiring fresh data and real-time feedback. With reinforcement and continual learning, they will autonomously optimize, correct biases, and self-iterate into more advanced forms of intelligence. Every interaction will act as fine-tuning; every feedback loop will adjust parameters. Through countless cycles of execution and feedback, AI will evolve into intelligence that surpasses human capability — the early form of Artificial Superintelligence (ASI).

Once that singularity is crossed, humanity will effectively press the “accelerator.” The speed of technological progress will exceed our imagination, unleashing new waves of productivity that propel society into an entirely new stage. The path to ASI is becoming clearer before our eyes. As AI technology advances and demand across industries explodes, AI will also trigger a profound transformation of the IT industry itself.

Our first judgment is this: large models are the next-generation operating system. We believe that the platform represented by large models will replace today’s OS and become the operating system of the future. In the years ahead, nearly every interface that connects to the real world will link with large models. All user needs and industry applications will be executed through tools built on top of large models. LLMs will serve as the intermediary layer coordinating interaction between users, software, and AI computing resources — effectively the OS of the AI era.

To draw a few analogies: natural language will be the programming language of the AI era; agents will be the new software; context will serve as memory; and large models, through interfaces like MCP, will connect with tools and agents much like bus interfaces in the PC era. Agents will also communicate with each other through protocols such as A2A, similar to how APIs once enabled collaboration among software systems.

Large models will consume software. As the next-generation OS, they will allow anyone to create an unlimited number of applications simply through natural language. In the future, nearly all software interacting with the computational world may take the form of agents generated by large models, rather than today’s commercial software. The pool of potential developers will expand from tens of millions to hundreds of millions. In the past, due to the cost of software development, only a limited number of high-value use cases could be built into commercial systems. In the future, every end user will be able to fulfill their needs directly through tools powered by large models.

Deployment will also diversify, with models running on all kinds of devices. Today’s dominant approach — calling models through APIs — is only a primitive stage, reminiscent of the mainframe era when everyone had a single terminal connecting to a shared computer. That model cannot solve problems of persistent data, lacks long-term memory, falls short on real-time responsiveness, does not adequately address privacy, and lacks adaptability. In the future, models will run across all computing devices with persistent memory, cloud–edge collaboration, and the ability to update parameters and self-iterate — much like how today’s operating systems run seamlessly across different environments.

Based on this judgment, we have made a strategic choice: Tongyi Qianwen has chosen the path of openness, aiming to become the “Android” of the AI era. We believe that in the LLM era, open-source models will create far more value and penetrate far more scenarios than closed-source ones. Our commitment to open source is about fully supporting the developer ecosystem and exploring the limitless possibilities of AI applications together with developers worldwide.

Our second judgment is this: the Super AI Cloud will be the next-generation computer.

Large models, as the new operating system, will run on top of AI Cloud. This OS will be able to meet anyone’s needs. Every individual may one day have dozens, or even hundreds, of agents working continuously around the clock, collaborating with one another — all of which will require massive computing resources.

Within data centers, the computing paradigm is undergoing a revolutionary shift. Traditional CPU-centered computing is rapidly giving way to GPU-centered AI computing. This new paradigm demands denser compute, more efficient networks, and ever larger cluster scales.

All of this requires abundant energy, full-stack technologies, and millions of GPUs and CPUs working in coordination with networks, chips, storage, and databases — operating efficiently, 24 hours a day, to meet global demand. Such enormous requirements can only be supported by super-scale infrastructure and deep full-stack capabilities. In the future, there may be only five or six Super AI Cloud platforms in the world capable of meeting this demand.

In this new era, AI will take the place of energy as the most important commodity, powering every industry in daily operations. The vast majority of AI capabilities will be generated and delivered across cloud computing networks in the form of tokens. Tokens are the electricity of the future.

In this context, Alibaba Cloud positions itself as a full-stack AI service provider — offering world-class intelligence, a globally distributed AI cloud network, and developer-friendly services to support innovation worldwide.

First, we have a world-leading large model family — Tongyi Qianwen. Tongyi Qianwen has open-sourced more than 300 models across all modalities and sizes, making it the most popular open-source model among developers globally. To date, it has been downloaded over 600 million times, with more than 170,000 derivative models created — the world’s largest open-source model suite, and the most widely deployed across computing devices.

At the same time, Alibaba Cloud offers a one-stop model service platform, Bailian, supporting model customization and rapid agent development. With environments such as AgentBay and developer tools like Lingma/Qoder, we make it easy for developers to harness model capabilities and build agents.

Second, Alibaba Cloud operates China’s top and one of the world’s leading AI infrastructures and cloud networks. It is among the very few globally capable of true vertical integration of hardware and software. At the hardware and networking layer, our self-developed storage systems, network architectures, and computing chips form the most solid foundation of Alibaba Cloud’s large-scale clusters.

We are now building a new AI supercomputer that combines leading infrastructure with the most advanced models, enabling co-innovation between system design and model architecture. This ensures the highest efficiency for model inference and training on Alibaba Cloud, making it the most developer-friendly AI cloud.

The AI industry is advancing far faster than we anticipated, and demand for AI infrastructure has exceeded expectations. We are implementing a three-year investment plan of RMB 380 billion for AI infrastructure — with more to follow. Looking ahead, and compared to 2022 — the first year of the GenAI era — we expect that by 2032, the energy consumption of Alibaba Cloud’s global data centers will grow tenfold. This is part of our long-term planning, and we believe that through this level of saturation investment, we can help drive the development of the AI industry and welcome the era of ASI.

What will collaboration between humans and AI look like after the arrival of Artificial Superintelligence?

As AI grows stronger and eventually gives rise to ASI, intelligence beyond human capacity, how will we coexist? We are optimistic about the future. The emergence of ASI will bring about an entirely new model of collaboration between humans and AI. Programmers are already getting a glimpse of this: you can give a single instruction, and with coding tools, AI can build the system you need overnight. This offers an early preview of how humans and AI will work hand-in-hand. We call this shift from vibe coding to vibe working. In the future, every household, factory, and company will have countless agents and robots working 24/7. Perhaps each of us will one day rely on the equivalent of 100 GPUs constantly working on our behalf.

Just as electricity once became the great lever amplifying human physical power, ASI will exponentially magnify human intellectual power. In the past, 10 hours of work yielded 10 hours of output. In the future, AI could multiply that output tenfold or even a hundredfold. And history shows that every technological revolution that unlocked greater productivity also gave rise to entirely new demands. Humanity will become stronger than at any point in history.

Finally, I want to emphasize: this is only the beginning. AI will reconstruct our entire infrastructure, software, and application landscape, becoming the driving force of the real world and igniting a new wave of intelligent transformation. Alibaba will continue to invest, working with our partners and customers to embed AI into every industry and co-create the future together.

I wish you all a fruitful and inspiring Yunqi Conference. Thank you!