The competition behind the scenes is a systems engineering endeavor.
Author | Chai Xuchen
Editor | Wang Xiaojuan
After unveiling the compelling narrative of Artificial Super Intelligence (ASI) and demonstrating a bold, three-year RMB 380 billion commitment, Alibaba has once again taken center stage in the global AI arena.
At the 2025 Apsara Conference, Alibaba’s Chief Executive Officer Wu Yongming discussed the prospects of ASI. He remarked that ‘large models are the operating systems of the next generation,’ replacing the current OS paradigm, with AI models set to rapidly permeate all devices. Furthermore, he asserted that super AI clouds will become the computers of the next generation, predicting that there may only be 5-6 global super cloud computing platforms in the future.
Based on this foreseeable vision, Alibaba made a strategic decision: to double down on infrastructure investment and open-source its Qwen model, with the aim of becoming the ‘Android of the AI era.’ This implies that Alibaba must maintain a rapid pace of iteration across model development and computational capabilities to secure a competitive edge in such an intensely contested field.
Achieving this is no easy task. In a post-event interview, Zhou Jingren, CTO of Alibaba Cloud, candidly told Wall Street News that AI market competition is exceptionally fierce, with models like OpenAI, Google, and Claude accelerating their advancements, while Alibaba’s Qwen continues to closely compete with global leaders.
Zhou Jingren pointed out that as the industry enters an acceleration phase, players are now competing not just on the capabilities of individual models but, more importantly, on rapid innovation and iteration.
Nevertheless, Alibaba Cloud remains confident. In Zhou’s view, today’s competition between models is essentially a competition of systems; similarly, the competition among cloud providers is also tied to the performance of models. These two aspects are inseparable. Alibaba Cloud is one of the few companies globally that can achieve full-stack self-research and integrated innovation in both large models and cloud computing.
Zhang Qi, Vice President of Alibaba Cloud, told Wall Street Insight that Alibaba Cloud possesses full-stack AI capabilities, including underlying computing power, cloud infrastructure, and a large model family. “From these three aspects, the only companies globally that have layouts across all three layers are Alibaba and Google.”
In response to ASI, Wu Yongming revealed that Alibaba Cloud is building an entirely new AI supercomputer. This machine can drive collaborative innovation in both fundamental architecture design and model architecture, ensuring the highest efficiency when invoking and training large models on Alibaba Cloud.
It can be said that at this moment, Alibaba Cloud is sharpening its tools, ready to fiercely compete with rivals in this vast new blue ocean, while preparing to become a new cornerstone for the entire Alibaba Group in the future.
Below is the edited transcript of the conversation with Zhou Jingren, CTO of Alibaba Cloud, and Zhang Qi, Vice President:
Question: Why has Alibaba been able to maintain a high pace of model releases this year? Will there be a unified model like OpenAI’s GPT-5 that could potentially bring an end to such a diverse range of models?
Zhou Jingren: The progress of AI models globally is accelerating. The Tongyi model has been keeping pace with global leaders. Observers may notice that models from OpenAI, Google, and Claude-related projects are all speeding up their advancements.
AI has seen numerous breakthroughs in the past, but today it has entered an acceleration phase. What is being compared now is not just the capability of a few individual models; more importantly, it is the rapid iteration and innovation. Unconsciously, everyone is speeding up the efficiency of model iteration. Secondly, the evolution from single-modal models to multi-modal ones is an inevitable trend, closely related to human intelligence.
Question: The competition in the AI market is extremely fierce. After achieving the top market share, does Alibaba Cloud have any unique strategies or ideas regarding its approach and competitive advantages in AI cloud services?
Zhou Jingren: Alibaba Cloud is one of the few global companies capable of fully self-developing and jointly innovating 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 integrated and mutually reinforcing. Today’s competition among models is essentially a competition of systems; similarly, competition in cloud services is also a competition of models. These two are inseparable.
Question: Given the limited resources of R&D personnel today, how can we achieve innovation at a more fundamental paradigm level like DeepSeek?
Zhou Jingren: The innovation in the entire model cluster is not dispersed; everything is interconnected behind the scenes. Sometimes, it is necessary to optimize a specific task within a single modality to achieve the best possible results, which in turn enhances the overall capability of the model. All advancements in these models are necessarily part of the joint optimization of the evolution of the entire Tongyi large model system.
From the beginning of this year until now, there has been rapid progress, with several generations of models developed. Each generation of models has seen significant performance improvements, and we are also actively working on the development of the next-generation model. For example, the Tongyi Qianwen Next has introduced substantial innovations in its architecture. Once launched, the entire community has been adapting to the new architecture and conducting related experiments.
The development of models is an incremental process, not one focused on sudden breakthroughs. All overseas manufacturers will gradually develop, and today the focus is on accelerating the iteration and innovation speed of models.
Zhang Qi: This morning, the Tongyi Qianwen 3-MAX was ranked third in all global model leaderboards.
Today, we have over 300 models under the Qianwen series because we possess both the best models and the strongest cloud services. In terms of AI cloud, Alibaba Cloud is the only Chinese company among the top four globally. Currently, we are also the only cloud provider in China with international influence and strong global competitiveness.
International companies such as Salesforce and SPA primarily choose Alibaba Cloud as their main partner in China or across Asia. We are strategically positioned to exert full-stack efforts across the board.
Question: This year, some vendors have integrated Agent capabilities into their models. What will be the relationship between models and Agent capabilities in the future?
Zhou Jingren: There isn’t a clear-cut boundary solution for this. Our model services will inherently possess Agent capabilities. Once a model has search functionality, it essentially becomes an Agent.
The development of intelligent agents discussed today is industry-specific. This requires deep knowledge of each industry’s system. Some core Agent capabilities provided by Bailian will gradually be integrated into Tongyi Qianwen and Wanxiang. This means that the underlying models will become increasingly powerful. However, for business-layer tool usage and fine-tuning, specialized business-layer Agents are still needed for implementation and problem-solving.
Question: How to stimulate the creativity of AI scientists?
Zhou Jingren: The overall environment within the Tongyi Lab is relatively open, encouraging everyone to pursue new innovations. Today’s development trends in the industry, unlike a few years ago when overseas companies dominated, have reached an increasing consensus across the sector. However, having a consensus does not necessarily mean it can be realized or completed.
What we need today is to more effectively carry out a series of tasks according to priority and make relevant plans. Through the joint optimization of our systems and algorithms, we can drive these efforts forward.
Today, our approach is one of simultaneous progress on multiple fronts. Ultimately, we aim to push individual aspects to their limits; otherwise, even significant improvements in large model capabilities will hit bottlenecks. Behind this logic, there isn’t much variation in the general direction taken by various companies—except that they are not as open as we are.
One key understanding we have is from the developer’s perspective. Why do we provide so many models and parameters? We understand that developers have diverse needs and want them to choose the best model for their specific scenarios to integrate into their own systems. In the truest sense, we hope to collaborate with developers and enterprises to jointly advance the development of the AI industry.
Question: Facing competitors in markets where Token calls reach hundreds of millions, how does Alibaba Cloud maintain its forward-looking position in such a fiercely competitive environment?
Zhang Qi: In the AI cloud market—including underlying infrastructure and Token calls—Alibaba Cloud is number one in China, equaling the combined total of the second to fourth places. According to Frost & Sullivan research of China’s top 500 companies, over 70% have already adopted generative AI, and the penetration rate of Alibaba Cloud and Tongyi Qwen is 53%.
Question: Alibaba Cloud covers the entire stack, but what is the core focus?
Zhang Qi: Regarding full-stack AI, including underlying computing power, cloud infrastructure, and the family of large models, only Alibaba and Google globally have layouts across all three layers simultaneously.
Zhou Jingren: Alibaba Cloud’s so-called fully self-developed stack system did not just start today. Long before others were providing model services, we proposed Model as a Service. From that point onward, our cloud development direction has been integrated with AI models. Even before ChatGPT emerged, we were discussing the crucial integration of AI and cloud technologies.
Question: There are no major disagreements on the overarching technical roadmap at present. What is the most critical factor for Alibaba to maintain its leadership in models?
Zhou Jingren: I don’t think the pace of innovation has slowed down at all. Global investments in this area are accelerating, and they continue to demonstrate that the upper limits of AI models have not yet been reached. We are continuously accelerating and innovating.
As for achieving ASI (Artificial Superintelligence), there are numerous challenges that need to be addressed. Currently, the overall complexity of models, including processing capabilities and deep-thinking abilities, shows progress in some areas such as mathematics, coding, and other scenarios. Truly enabling rapid integration of various tools, methods of model training, patterns of model innovation, and model architectures may undergo a series of significant changes. Ultimately, models must achieve autonomous learning, self-improvement through feedback, interaction with the world to gather feedback, and utilize that feedback for further evolution and upgrading.
In our current trajectory of generational model development, we aim to gradually establish a process where models continuously learn and self-improve. This journey presents architectural, systemic, and algorithmic challenges.
Question: MaaS (Model as a Service) is being pursued by many companies. Does Alibaba have any differentiation?
Zhou Jingren: ‘Model service’ is not just a simple concept. Achieving ultimate elasticity, top-tier performance, and extremely high throughput in model services today is highly challenging. We often refer to model service as the elastic computing of the AI era.
Before discussing the performance of MaaS, we must first address precision. For the same model, customers sometimes find varying results when using different platforms or services. This discrepancy primarily stems from precision alignment. In these aspects, we have very rigorous processes in place. The series of services provided by the Tongyi model ensures native support for optimal precision.
Today, the services offered by the Tongyi model, along with other model services on Alibaba Cloud, deliver exceptionally high-quality outcomes. Of course, enterprises also have diverse requirements for model services regarding throughput, latency, cost, and more. Some companies prioritize extreme efficiency—sometimes referring to model effectiveness, other times to latency. Others seek the best cost-performance ratio and may flexibly trade off certain aspects of model service timing.
Today, only a few companies globally can achieve this.
Question: What are the boundary points for Agent’s memory-related multimodal functionalities currently?
Zhou Jingren: ‘Memory’ here is a broad term. First, we hope that the model’s responses can be well-informed about the context, which is relatively simple. However, human memory not only includes what happened yesterday but also events from last year or even ten years ago, involving the management of ultra-large-scale information. AI needs to support memory across various modalities—not just text, but also today’s videos, past audio, and so on.
Moreover, we need to stratify memory, much like human memory. From specific records of what happened in the past, abstraction is required—abstracting experiences, habits, or even elements that have become part of one’s personality today. This involves short-term memory, multi-layered memory, and it represents an intelligent process rather than pure rule-based operations. There is still much research to be done in this area. We believe that for creating an intelligent agent or applying it within enterprises, relevant memory capabilities remain necessary.
Question: What are the internal priority dimensions for evaluating large models like Tongyi?
Zhou Jingren: First, our direct evaluation system for all technical work focuses more on capability, including the model’s ability. On the path toward Artificial Super Intelligence (ASI), the most critical step is to achieve breakthroughs in technology and capability, making further technological progress towards the ASI direction as Wu Yongming mentioned.