这些变化可以概括为多个方面。其一，基础设施巨变：从"训练竞赛"到"推理革命"，能源与芯片格局面 临重塑。 周鸿祎判断，AI产业的发展动力将发生根本性转移。过去行业拼的是"谁家模型更博学"，而2026年，企 业将不再频繁训练大模型，而是通过"推理应用"直接"雇佣"AI解决实际问题。无论是影视创作还是城市 调度，高频、长流程的推理任务将推动推理算力需求呈现"短期百倍级"增长，其规模与增速将全面超越 训练算力。 这一转变将引发连锁反应。首先，AI芯片市场"英伟达一家独大"的单极格局将被打破，形成"英伟达主 导训练，多家厂商分食推理"的双轨产业格局。专用推理芯片（ASIC）将凭借成本优势在细分场景快速 渗透。其次，制约发展的核心瓶颈将从算力芯片转向稳定充足的电力供给，全球科技竞争将升级为"能 源大战"。他特别指出，中国依托"东数西算"国家工程与绿色电力能力，已在这场新竞争中占据先发优 势。 第二，模型智能升维：从"博学工具"到"深思伙伴"，中国开源生态崛起为全球核心。 在模型演进层面，周鸿祎预言AI将完成从"静态工具"到"持续进化系统"的蜕变。2026年，传统的"预训 练+微调"模式将让位于 "通用基座+行业专精+推 ...

Core Viewpoint - The development of an open-source ecosystem in artificial intelligence (AI) is crucial for enhancing China's AI industry, with open-source large models playing a significant role as AI becomes a global public good and enters a stage of inclusivity [1][3]. Group 1: Open Source Ecosystem Development - Beijing has the foundational capabilities to build an "open-source capital," hosting over half of the country's open-source startups across key areas such as foundational platforms, large models, and robotics [3]. - The city possesses the largest information software industry in the country, with a robust industrial chain that supports open-source development [3]. - Continuous improvement of top-level design encourages enterprises to actively participate in international open-source projects and supports the establishment of open-source organizations in Beijing [3]. Group 2: Recommendations for AI Open Source Ecosystem - It is recommended to strategically position in frontier areas such as intelligent agents and embodied intelligence to accelerate technological breakthroughs and ecosystem development [3]. - Creating a favorable open-source culture is essential, encouraging institutions to practice open-source principles in the AI field [3]. - Integration of inclusive computing resources and the establishment of a joint operation mechanism are necessary for the ecosystem [3]. - Strengthening communication and collaboration within existing AI open-source communities is vital to create a neutral and authoritative platform [3]. Group 3: AI Open Source Community - The AI open-source community has become a key component for the practical application of large models, exemplified by the Jingzhi community initiated by the China Academy of Information and Communications Technology [5]. - As of now, the Jingzhi community has over 2,800 open-source models covering areas such as computer vision, natural language processing, speech processing, and multimodal applications [5]. - The community has aggregated more than 380 datasets, including text, images, and audio-visual data, and provides over 500 petaflops of computing power from resources like NVIDIA, Huawei Ascend, and Baidu Kunlun [5].