TMTPOST -- Li Feifei, Vice President of Alibaba Cloud Intelligence Group and head of the database product division, noted that as DeepSeek gains traction, AI competition is shifting back to traditional system engineering, particularly distributed systems. Li emphasized that system evolution inevitably follows this trend. He also predicted that large-scale AI pretraining will become centralized and dominated by a few major players. "Globally, only a handful of companies will remain in this space," he said. Given that no AI company has yet achieved a self-sustaining business model, Li described large-scale pretraining as a money-burning endeavor—one that will only become profitable if it crosses the threshold into Artificial General Intelligence (AGI). However, he acknowledged that DeepSeek’s emergence has significantly advanced open-source AI models, enabling complex logical reasoning while reducing inference costs. To address these challenges, Alibaba Cloud is optimizing heterogeneous resource pools that combine GPUs and CPUs. The key goal for database development is to maximize GPU efficiency by reserving them for critical computing and caching tasks, while offloading secondary processes to a three-tier system of CPU, memory, and storage pools—ultimately making online inference more cost-effective.