AI需求侧核心逻辑正式向多模态大模型延展-国产算力认知强化！Tokens消耗

Core Insights - The recent release of multimodal models, particularly Sora2, is considered a "revolutionary" milestone in the industry, enhancing user engagement and willingness to pay for AI-generated content [1][2] Group 1: International Developments - OpenAI launched the Sora2/Pro App on October 1, supporting up to 15 seconds of text-to-video generation, achieving the top position in the US App Store within three days [1] - The developer conference on October 7 announced that ChatGPT can now directly access third-party applications, marking a shift from a single dialogue tool to an AI application and social platform [1] - xAI introduced the "Imagine" visual generation module on October 6, enhancing its capabilities in creating high-quality images and videos from text [1] - Anthropic released the Claude Sonnet 4.5 programming model on September 30, emphasizing its ability to build "production-ready" AI agents [1] Group 2: Domestic Developments - Kuaishou's Ling2.5Turbo topped the global video generation model rankings on October 2, showcasing its international leadership in video generation and content quality [2] - ByteDance partnered with UCLA on October 2 to launch Self-Forcing++ video generation technology, significantly improving visual stability [2] - Tencent released and open-sourced the mixed Yuan Image 3.0 on September 28, quickly rising to the top of the Hugging Face leaderboard [2] Group 3: Domestic Computing Power Investment Logic - The rise of domestic computing power is driven by demand from AI applications, marking a shift from supply-side to demand-side dynamics [3] - DeepSeek's release of DeepSeek-V3.2-Exp on September 30 demonstrated lower inference costs and compatibility with domestic chip ecosystems [3] - Alibaba's open-source Qwen3-VL series multimodal model, released on October 4, achieved zero-day adaptation with domestic chips, accelerating the local hardware ecosystem [3] Group 4: Investment Recommendations - Recommendations for cloud computing power include companies like Cambrian, Haiguang Information, and Chipone [4] - For edge computing power, companies such as Amlogic and Rockchip are recommended [4]