2026年全国两会，全国政协委员、360集团创始人周鸿祎重点关注四个方向：一是优化推理算力布局，夯实人工智能产业发展底座；二是"智能体技术 普惠+懂AI懂业务的人才培育"双轮驱动，加速推进"人工智能+"行动落地；三是推广安全智能体的广泛应用，筑牢新兴领域国家安全屏障；四是协同 完善数据流通安全合规体系，推动数据、网络、AI一体化安全能力提升。 周鸿祎表示，过去两年，产业焦点集中于大模型的预训练，对高端训练芯片的需求极为迫切。当前，随着基础模型能力普遍越过及格线，行业正迈入"人 工智能+"的应用时代，全国算力的需求结构发生了根本性变化。 "大语言模型聊天的算力和智能体实际干活时所需的算力，是无法相提并论的。"周鸿祎解释，当一个智能体真正开始为企业干活——撰写一部短剧，分析 一份财报，或是自动完成一笔复杂交易，其消耗的推理算力将是简单对话的几百倍甚至上千倍。他表示，一旦进入大模型应用阶段，推理算力的需求将呈 指数级增长。 "过去重视训练算力是合理的，但现在基座模型已过及格线，行业应用更应聚焦推理算力。"他解释，智能体执行任务时需反复分解步骤、试错搜索， Token消耗可达聊天场景的数百倍。周鸿祎表示，推理芯片对互 ...

2026.03. 03 本文字数：734，阅读时长大约2分钟 封图 | 受访者供图 2026年全国两会即将召开，全国政协委员、360集团创始人周鸿祎拟围绕推理算力、智能体技术与 人才发展、智能安全三方面提交提案。 周鸿祎认为，我国经历"百模大战"后，有了很多"国际一流"的开源模型，国家主导的训练算力稳步提 升，推理算力需求在"百亿智能体时代"呈指数级增长，专用推理芯片是我国芯片产业实现差异化突围 的重要方向。 目前，我国算力中心面向推理任务的专用集群存在缺口，区域间供需适配有待优化，专用推理芯片技 术也亟需突破。周鸿祎建议优化推理算力布局：国家出台推理算力布局指导政策，依据各地场景密 度、算力缺口、能源保障能力，建立"全国统筹 + 区域细化" 的推理算力布局体系，在重点产业集聚 区域，建设低时延、高密度的推理算力集群。 另外，强化一体化调度，推动跨层级、跨区域的算力资源动态调配，提升推理算力利用效率。鼓励专 用推理芯片的国产化发展，重点突破高精度、低时延、多模态的芯片技术，实现产业链自主可控，支 持智能体技术的深度应用。 采访中，周鸿祎表示，国产大模型未达到及格线时，重视训练算力是合理的，但如今极度消耗推理算 ...

行 业 及 产 业 行 业 研 究 / 行 业 点 评 相关研究 《春节海内外大模型更新全梳理！壁仞科 技深度发布！——计算机行业周报 20260216-20260220》 2026/02/23 《模型会吞噬软件吗？——计算机行业周 报 20260202-20260206》 2026/02/07 证券分析师 黄忠煌 A0230519110001 huangzh@swsresearch.com 洪依真 A0230519060003 hongyz@swsresearch.com 刘洋 A0230513050006 liuyang2@swsresearch.com 研究支持 崔航 A0230524080005 cuihang@swsresearch.com 曹峥 A0230525040002 caozheng@swsresearch.com 陈晴华 A0230525100001 chenqh@swsresearch.com 罗宇琦 A0230124070004 luoyq@swsresearch.com 联系人 王开元 A0230125030001 wangky@swsresearch.com 2026 年 02 ...

行 业 及 产 业 行 业 研 究 / 行 业 点 评 相关研究 《春节海内外大模型更新全梳理！壁仞科 2026 年 02 月 28 日 从国产算力变化到 LPU！DS 新模 型前瞻！ 看好 ——计算机行业周报 20260223-20260227 技深度发布！——计算机行业周报 20260216-20260220》 2026/02/23 《模型会吞噬软件吗？——计算机行业周 报 20260202-20260206》 2026/02/07 证券分析师 黄忠煌 A0230519110001 huangzh@swsresearch.com 洪依真 A0230519060003 hongyz@swsresearch.com 刘洋 A0230513050006 liuyang2@swsresearch.com 研究支持 崔航 A0230524080005 cuihang@swsresearch.com 曹峥 A0230525040002 caozheng@swsresearch.com 陈晴华 A0230525100001 chenqh@swsresearch.com 罗宇琦 A0230124070004 luoyq@ ...

在"企业和个人如何快速使用AI"方面，周鸿祎表示，现在面临的问题是都在用AI助手，或者把AI当搜索用，个人如何打造专属私人的智能体？OpenClaw 的启发是要简单化。 "智能体只有做得更加专业，能够直接给企业带来价值，企业才会愿意付费使用。"周鸿祎强调。 【导读】全国政协委员、三六零创始人周鸿祎：将关注AI赋能安全等方向 中国基金报记者 卢鸰 全国政协委员、三六零创始人周鸿祎2月26日下午在接受媒体集体采访时表示，今年全国两会期间，将关注AI赋能安全、AI在中国如何落地、企业和个人 如何快速使用AI等方向。 "以Anthropic为例，通过AI编程、AI查找漏洞，可以解决很多原来安全上不能解决的问题，所以，我建议关注AI智能体。"周鸿祎称。 据其介绍，三六零已经做了几十种、上万个AI安全智能体，这些智能体能够挖掘软件漏洞，抵御其他国家的黑客智能体。 对于"AI在中国如何落地"，周鸿祎表示，一定要把算力分成训练算力和推理算力，训练算力在规模上可能还有一定的空间，而推理算力的发展空间是无限 的。 "所以，希望各地在发展算力方面能够偏向推理算力。从国家产业政策来看，在芯片政策上不能都追英伟达的高端训练芯片，推理芯 ...

建议重点关注—#端侧推理核心+G端本地部署业务核心的【云天励飞】 建议重点关注一#端侧推理核心+G端本地部署业务核心的【云天励飞】 1OpenClaw不是普通的AI工具，它更像是一个能一站式搭建业务的智能机器人。 1OpenClaw不是普通的AI工具，它更像是一个能一站式搭建业务的智能机器人。 OpenClaw爆火，AI闭环更进一步，推理算力需求持续提升 OpenClaw爆火，AI闭环更进一步，推理算力需求持续提升 传统AI工具如ChatGPT，大多是单一功能，而且不同工具之间没有记忆联动。 传统AI工具如ChatGPT，大多是单一功能，而且不同工具之间没有记忆联动 比如用这个工具写内容，用那个工具做SEO，彼此之间不通气，只能完成碎片化 比如用这个工具写内容，用那个工具做SEO，彼此之间不通气，只能完成碎片化 OpenClaw爆火，AI闭环更进一步，推理算力需求持续提升 OpenClaw爆火，AI闭环更进一步，推理算力需求持续提升 建议重点关注—#端侧推理核心+G端本地部署业务核心的【云天励飞】 建议重点关注一#端侧推理核心+G端本地部署业务核心的【云天励飞】 1OpenClaw不是普通的AI工具，它更像是 ...

Core Insights - The AI industry is transitioning from a "training-driven" phase to a "reasoning-driven" phase, with reasoning computing power becoming the core element for the commercialization of AI [1][2] - Sunrise, a domestic AI chip company, has launched its new generation reasoning GPU chip, the Qihang S3, aiming for a target of "one cent per million tokens" [1][5] - The next decade will see reasoning infrastructure as the foundational base for China's AI era, emphasizing the need for cost-effective and scalable reasoning capabilities [1][9] Group 1: Reasoning Computing Power - Reasoning computing power is essential for the practical application of AI, with predictions indicating that by 2026, reasoning computing will account for 66% of AI computing, surpassing training computing for the first time [2][4] - The shift towards reasoning-driven AI is crucial for enhancing the efficiency of AI services in the real economy [2][3] Group 2: Sunrise's Innovations - Sunrise is the first company in China to focus on reasoning GPUs, having developed its first chip, Qihang S1, in 2018, and has since released the Qihang S2 and Qihang S3, which are optimized for large model reasoning scenarios [3][5] - The Qihang S3 chip aims to achieve over ten times improvement in reasoning cost-effectiveness, with current costs at approximately 0.57 yuan per million tokens, better than the market average [5][6] Group 3: Industry Challenges and Solutions - The industry faces challenges such as low resource utilization, insufficient adaptation efficiency, and complex operations, with over 40% GPU idle rates under traditional architectures [6][8] - Sunrise is collaborating with partners to create a reasoning system-level solution that optimizes both hardware and software to address these challenges and improve computing efficiency [6][8] Group 4: Market Potential and Future Trends - The demand for reasoning tokens is expected to grow exponentially, with a significant market opportunity for specialized reasoning GPUs [6][9] - The reduction of reasoning costs is projected to lead to a massive increase in AI applications, with estimates suggesting that a 50% cost reduction could trigger widespread adoption [8][9]

Core Insights - The founder of 360 Group, Zhou Hongyi, predicts that by 2026, the world will enter the "hundred billion intelligent agent" era, and China is well-positioned to seize this strategic opportunity [1][4] - 360 Group is set to launch a "short drama intelligent agent" that allows users to generate large-scale animated films from scripts, significantly lowering the barriers to content creation [1][2] Group 1: AI Evolution and Market Dynamics - Zhou Hongyi believes that 2024 will be a year focused on large models, while 2025 will be a transition period. Large models, primarily in the form of "chatbots," struggle to address complex business problems directly [1] - The "five-force model" proposed by Zhou includes "electricity—computing power—intelligence + human power—productivity," emphasizing that converting general computing power into specialized intelligence is crucial for practical applications [1] - The industry often confuses "training computing power" with "inference computing power," with the latter expected to see exponential growth in demand as intelligent agents are applied to complex tasks like short drama production and education [2] Group 2: Transformation of Internet and Business Models - The rise of intelligent agents will fundamentally change how humans interact with software and the internet, leading to a bifurcation into two types of internet: one for human use and another for intelligent agents [3] - Traditional e-commerce models will shift from "humans finding goods" to an agent-based model where intelligent agents handle the entire transaction process, resulting in increased transactions occurring between agents rather than between humans and screens [3] - New trust and settlement systems will emerge in the intelligent agent economy, necessitating advancements in identity verification, transaction security, and automated settlement, which will leverage technologies like blockchain and smart contracts [3] Group 3: China's Strategic Position - China possesses robust electrical infrastructure, a complete industrial system, and excellent open-source model ecosystems, positioning it to capitalize on the opportunities presented by the hundred billion intelligent agent era [4] - There is a call for companies to foster an "AI-native" culture, transforming individuals who embrace AI into "super individuals," while also emphasizing the importance of maintaining safety standards to mitigate risks associated with collective intelligence [4]

Core Viewpoint - OpenAI and Cerebras are collaborating to deploy a 750 MW wafer-scale system, which will become the world's largest high-speed AI inference platform by 2028, with a project value exceeding $10 billion [1]. Group 1: Collaboration and Market Demand - The partnership between OpenAI and Cerebras signifies a strong market demand for inference computing power and highlights the increasing importance of inference speed among tech giants [1]. - Cerebras, founded in 2015, aims to create the fastest AI inference and training platform, with its CS-2 and CS-3 systems already applied in various fields such as medical research and cryptography [4]. Group 2: Technological Advancements - Cerebras' unique system integrates massive computing power, memory, and bandwidth into a single giant chip, eliminating traditional hardware bottlenecks that limit inference speed [4]. - The response speed of large language models based on Cerebras technology can be up to 15 times faster than those based on GPU systems for code and voice chat tasks [4]. Group 3: Industry Trends - The tech industry's history shows that speed has played a crucial role in technology adoption, with significant advancements in processing frequency and internet connectivity driving the growth of personal computing and modern internet [5]. - Low-latency inference solutions provide faster response times and more natural interactions, enhancing productivity in the AI-driven market [5]. Group 4: Competitive Landscape - In December 2025, AI chip startup Groq announced a non-exclusive licensing agreement with NVIDIA, valued at $20 billion, marking NVIDIA's largest transaction to date [5]. - NVIDIA plans to integrate Groq's low-latency processors into its AI factory architecture to support a broader range of AI inference and real-time workloads [6].

Group 1 - The core theme of the China Wealth Management 50 Forum 2025 Annual Meeting is "Towards a Financial Powerhouse in the 14th Five-Year Plan" [1][4] - Alibaba Cloud's strategic direction focuses on "full-stack AI cloud" and "globalization," emphasizing a complete system construction from underlying chips and infrastructure to model applications [1][4] Group 2 - The competition between China and the US in various model fields is characterized by mutual strengths and weaknesses, with China showing a clear leading advantage in niche areas such as autonomous driving and embodied intelligence [3][6] - Future demand for reasoning computing power is expected to surpass training computing power, indicating a "reverse" trend [3][6] - The relationship between cloud and AI is described as a mutually reinforcing "flywheel," where financial institutions need to build a dual-wheel system of "large flywheel driving intent understanding and small flywheel executing" to achieve deep collaboration and integrate AI into professional workflows [3][6]