How did Alibaba become the Android of the AI era?

Author | Chai Xuchen

Editor | Wang Xiaojun

After telling the thrilling story of super artificial intelligence "ASI" and presenting an ALL IN posture with an investment of 380 billion yuan over three years, Alibaba has once again taken center stage in the global AI arena.

At the 2025 Yunqi Conference, Alibaba CEO Eric Wu discussed the prospects of ASI, believing that "large models are the next generation of operating systems," which will replace the current OS, and AI models will accelerate their penetration into all devices; he asserted that super AI clouds will become the next generation of computers, predicting that there may only be 5-6 super cloud computing platforms in the future.

Based on this foreseeable vision, Alibaba made a strategic choice: to increase infrastructure investment and open-source Tongyi Qianwen, with the ambition to become "the Android of the AI era." This means that Alibaba must maintain a rapid iterative pace in model development and computing power to secure a winning edge in such a competitive landscape.

Achieving this is not easy. In a post-conference interview, Alibaba Cloud CTO Zhou Jingren candidly told Wall Street Insight that competition in the AI market is extremely fierce, with models related to Open AI, Google, and Claude accelerating rapidly, and the Tongyi model is constantly chasing global leaders.

Zhou Jingren pointed out that as the industry enters an acceleration phase, players are competing not just on the capabilities of a few models, but more importantly, on rapid iterative innovation.

However, Alibaba Cloud remains confident. In Zhou Jingren's view, the competition between models today is actually a competition of systems; the competition in cloud computing is also a competition of models, and these two are inseparable. Alibaba Cloud is one of the few companies globally capable of full-stack self-research and joint innovation in both large models and cloud computing.

Alibaba Cloud Vice President Zhang Qi told Wall Street Insight that Alibaba Cloud possesses full-stack AI capabilities, including underlying computing power, cloud layout, and a family of large models. "From these three aspects, the only companies globally that can simultaneously have a layout in all three layers are Alibaba and Google."

Regarding ASI, Eric Wu revealed that Alibaba Cloud is fully committed to building a brand new AI supercomputer that can achieve collaborative innovation in infrastructure design and model architecture, ensuring the highest efficiency when calling and training large models on Alibaba Cloud.

It can be said that Alibaba Cloud is now sharpening its knives, ready to engage in fierce competition with its rivals in this vast new blue ocean, and is preparing to become the new ballast for the entire Alibaba in the future.

The following is a transcript of the conversation with Alibaba Cloud CTO Zhou Jingren and Vice President Zhang Qi (edited):

Q: Why has Alibaba been able to maintain a high pace of model releases this year? Will there be a unified model like Open AI's GPT-5 that ends this rich model landscape?

Zhou Jingren: The progress of AI models globally is accelerating. The Tongyi model has been chasing global leaders. In this industry, one can see that models related to Open AI, Google, and Claude are all accelerating AI has made many breakthroughs in the past, but today it has entered an acceleration phase. Today, the competition is not just about the capabilities of a few models, but more importantly, about rapid iterative innovation. Unconsciously, everyone is accelerating the efficiency of model iteration today. The second aspect is the inevitable trend of evolution from single-modal models to multi-modal models, which is closely related to human intelligence.

Q: The AI market competition is very fierce. After Alibaba Cloud achieved the number one market share, does it have unique strategies and ideas for its approach and competitive advantages in the AI cloud space?

Zhou Jingren: Alibaba Cloud is one of the few companies globally that can achieve full-stack self-research and joint innovation in both large models and cloud computing. This is our unique advantage.

The development of Alibaba's AI models, Agents, and innovations in AI infrastructure are interconnected and mutually reinforcing. The competition between models today is actually a competition of systems; the competition in cloud computing today is also a competition of models. These two are inseparable.

Q: In the current situation where R&D personnel have limited energy, how can we achieve innovations at a more fundamental level, like deepseek?

Zhou Jingren: The innovation of the entire model cluster is not dispersed; it is all interconnected. Sometimes, it is necessary to achieve optimal performance in a specific task scenario within a single modality to enhance the overall capability of a model. The development of all models must be part of the overall optimization of the evolution of large models.

Since the beginning of this year, there has been rapid progress, and several generations of models have already developed. Each generation of models has seen significant improvements in capability, and we are also actively working on the development of the next generation of models. For instance, Tongyi Qianwen Next has made a lot of innovations in its architecture. Once we launch it, the entire community will adapt and experiment around the new architecture.

The development of models is a gradual process, not a logic of holding back major moves. All overseas manufacturers will gradually develop, and today there is a need to accelerate the speed of model iteration and innovation.

Zhang Qi: The Tongyi Qianwen 3-MAX released this morning ranks third in all global model rankings.

Today, we have over 300 models in Qianwen because we have both the best models and the strongest cloud. From the perspective of AI cloud, Alibaba Cloud is the only Chinese player among the top four globally. Currently, we are also the only cloud vendor in China with international influence and strong international competitiveness.

International vendors like Salesforce and SPA choose Alibaba Cloud as their main partner in China or Asia, and we are strategically positioned to exert full-stack efforts.

Q: This year, some vendors have integrated Agent capabilities into their models. What will the relationship between models and Agent capabilities be in the future?

Zhou Jingren: Actually, there is no clear boundary here. Our model services will also possess Agent capabilities. The model's ability to perform search functions inherently makes it an Agent Today's discussion on the development of intelligent agents is industry-oriented, requiring a deep understanding of the knowledge systems of each current industry. Some core capabilities of the agents provided by BaiLian will gradually be integrated into Tongyi Qianwen and WanXiang, meaning that the underlying models will become increasingly powerful. However, the use of tools at the business layer and the tuning at the business layer still require intelligent agents to implement and solve these issues.

Q: How to stimulate the creativity of AI scientists?

Zhou Jingren: The entire Tongyi Laboratory is relatively open and encourages everyone to pursue new innovations. The development directions of the entire industry today are not like a few years ago when overseas companies dominated; there is an increasing consensus across the industry. However, consensus does not mean that it can be realized or completed.

What we need today is to effectively plan a series of related tasks according to priority, and through the joint optimization of our systems and algorithms, we can push the work forward.

Today, we are advancing in a parallel manner. Ultimately, we hope to achieve excellence in specific areas; otherwise, the significant enhancement of model capabilities will also encounter bottlenecks. The underlying logic behind this has not changed significantly among various companies; it's just that others are not as open as we are.

We have an important understanding from the developer's perspective: why provide so many models and parameters? We know that developers have diverse needs, allowing everyone to choose the best model for their own scenarios and integrate it into their systems. We genuinely hope to jointly promote the development of the AI industry with developers and enterprises.

Q: In the face of competitors calling tokens on a scale of billions, how does Alibaba Cloud maintain foresight in a competitive environment?

Zhang Qi: In the AI cloud market, including underlying infrastructure and token calls, Alibaba Cloud is number one in China, equivalent to the sum of the second to fourth places. According to a Sullivan survey of the top 500 companies in China, over 70% have adopted generative AI, with a penetration rate of 53% for Alibaba Cloud and Tongyi Qianwen.

Q: Alibaba Cloud is full-stack, but what is the core focus?

Zhang Qi: Regarding full-stack AI, looking at underlying computing power, cloud layout, and the family of large models, globally, only Alibaba and Google have simultaneous layouts in these three areas.

Zhou Jingren: The so-called full-stack self-developed system of Alibaba Cloud is not something we just started today. We proposed "Model as a Service" long before others began providing model services. Since then, our direction for cloud development has been integrated with AI models. We were already discussing the important integration of AI and cloud before the explosion of ChatGPT.

Q: Now that there are no major divergences in the major technical routes, what is the key factor for Alibaba to maintain its lead in models?

Zhou Jingren: I do not believe that the speed of innovation has slowed down at all; global investment in this area is also accelerating, confirming that the upper limit of AI models has not yet been reached. We are still continuously accelerating and innovating As for achieving ASI, there are many challenges that need to be addressed. Currently, the model is complex overall, including processing capabilities and deep thinking abilities. We have made progress in areas such as mathematics, coding, and other scenarios. To truly enable the rapid integration of various tools, the methods of model training, the patterns of model innovation, and the structure of the model may all undergo a series of changes. Ultimately, the model must be able to learn autonomously, utilize feedback through interaction with the world, collect feedback, and effectively use that feedback for further evolution and upgrading of the model.

In this process, we need to gradually cultivate a continuous learning and self-improvement process for the model from the current generations of model development. There are architectural challenges, system challenges, and algorithmic challenges involved.

Q: With various companies following up on MaaS, does Alibaba have any differentiation?

Zhou Jingren: "Model services" is not a simple concept. Achieving extreme elasticity, extreme performance, and extreme high throughput in model services is very difficult. We often say that model services are the elastic computing of the AI era.

Before discussing the performance of MaaS, we must first talk about accuracy. For the same model, customers sometimes wonder why the results differ when using different services on different platforms. The underlying issue is accuracy alignment. We have very strict processes in these areas. The Tongyi model and a series of services ensure native support for the best accuracy.

Today, the services of the Tongyi model, along with other model services on Alibaba Cloud, achieve very high-quality results. Of course, there are diverse requirements for model service throughput, latency, and cost. Different enterprises have varied demands for model services. Some companies pursue extreme efficiency, which may refer to model effectiveness or latency. Others seek extreme cost-effectiveness and may have flexible trade-offs regarding the timing of model services.

Currently, only a few companies globally can achieve this.

Q: What are the boundaries of the memory aspect related to multimodal capabilities in Agent?

Zhou Jingren: "Memory" here is a broad term. We first hope that the model's responses can be familiar with the context, which is simple. However, human memory is not just about what happened yesterday; it may also involve events from last year or even ten years ago, managing vast amounts of information. AI needs to support memory across various modalities, not just text, but also today's videos, past audio, etc.

Additionally, we need to layer memory, similar to human memory, from specific records of past events to abstraction, forming experiences, habits, and even becoming part of one's personality. This involves short-term memory and multiple memories, which is an intelligent process and not purely rule-based. There is still much research to be done in this area. We believe that creating an intelligent agent, particularly for enterprise applications, still requires relevant memory.

Q: What is the priority of evaluation dimensions for the Tongyi large model internally? Zhou Jingren: First of all, our evaluation system for all technical work is more about capability, including the capability of the models. On the path to ASI, the most important thing is to have breakthroughs in technology and capability, and to be able to make further technical breakthroughs and progress towards the ASI direction mentioned by Eddie Wu