Alibaba has made a strategic shift in its AI development, shifting from pursuing an open source ecosystem to commercial monetization. Key members of the Qwen team, including Lin Junyang and Hu Binyuan, have resigned due to strategic differences. Former Alibaba Cloud CTO Zhou Jingren has taken over, and CEO Wu Yongming has established the "Alibaba Token Hub" and formed an AI strategy committee to align model development with cloud business revenue targets. The company has clearly shifted its focus to prioritize MaaS and commercialization. The transformation logic is a very correct choice at present. Although Qwen's open source has gained global developer reputation, the model itself is not profitable. The main source of Alibaba Cloud's AI revenue is still selling GPU computing power, with MaaS accounting for a small proportion and thin profits. Having top-notch model capabilities and a large amount of GPU computing power earns meager profits, which is unsustainable and a failure in business. After the reform, Alibaba's strategy in AI has become highly consistent with ByteDance's strategy in AI. Doubao was closed source from the beginning, and the Volcano Engine was built around AI monetization as its logic. The number of Token calls is the highest in the world, and the monetization of Tokens is at the forefront. Although the model capability is not the first, the monetization rate is not worse than that of QWEN. The cost of transformation is that the soul of Qwen's open source ecosystem, Lin Junyang, has left. His departure may shake community confidence and trigger a chain of talent loss. More importantly, MiniMax、 Competitors such as Zhipu have surpassed Qwen in code generation, and the model capability itself is under pressure. At this point, if the product strength is not strong enough, customers will only turn to competitors. At the same time, the ByteDance volcano engine is growing rapidly, and has taken the lead in the distribution of the token consumption driven cloud sales model. The gross margin is much higher than Alibaba Cloud's GPU sales. So looking at Alibaba's future AI direction from the current perspective, Alibaba is likely to choose the route of "closed source flagship model+deep binding of cloud services", similar to the models of Microsoft Azure+OpenAI and Anthropic+AWS. Our own model combination can generate more gross profit, rather than wasting resources on selling GPU computing power. Prioritize the implementation of AI applications in e-commerce scenarios. Open source will not be completely abandoned, it is entirely possible to emulate Google, with top-level models closed source and their own applications, and small models open source. This wave of Agent explosion has fully connected the monetization chain of AI, and the token consumption brought by Agent AI far exceeds that of traditional chat. Alibaba's enterprise genes are not in the C-end users, but in the adoption of B-end users. If it can occupy a place on the enterprise level Agent platform, the MaaS ceiling will be greatly raised, and the profit margin will also be significantly improved. Let me give you an example that is easier for everyone to understand. The computing power is in the gold mine. Originally, Alibaba contracted the tools and mines to others, and also opened sourced the mining methods. They own the mine to earn rent and tool fees. But the price of gold has risen, and Alibaba has decided to take action on its own after calculating the accounts. In short, if the direction is correct, success or failure depends on whether the model's ability can rebuild its advantages. The current situation is somewhat similar to Kimi's situation at that time, and we must work hard on research and development. Kimi's transformation at that time was to stop advertising and shift the model from closed source to open source. But for big companies, flagship closed source is actually the right path, because they have enough computing power to complete monetization and monetize tokens. And whether the collaboration between Alibaba's e-commerce and cloud can truly run the commercial closed loop, and then sell this solution to the B-end At this point, I have to admire ByteDance. From the beginning, I had already thought it through and decided not to blindly follow the trend of open source models, hoard cards, own applications, and tokens for personal use. After the solution was successfully implemented, I would sell it to the B-side through the Volcano Engine to complete the commercial loop. The transformation of AI and organizational structure adjustment were just learning ByteDance's strategy that was set from the beginning.