即刻App年轻人的同好社区
下载
置顶旧认知(2023-2024)
投资AI 应用 = 投资大模型公司 = OpenAI、Anthropic、智谱、月之暗面…
新认知(2025-2026)
投资AI 应用 = 投资完整的 Agent 系统 = 模型 + 工具调用 + 长程记忆 + 商业化场景 + 用户入口
投资AI 应用 = 投资大模型公司 = OpenAI、Anthropic、智谱、月之暗面…
新认知(2025-2026)
投资AI 应用 = 投资完整的 Agent 系统 = 模型 + 工具调用 + 长程记忆 + 商业化场景 + 用户入口
10 01
Hermes Agent GUI 客户端来了，现在 Agent 的主流是 GUI 。
它不只是 GUI 化——是把整个 Hermes 的使用体验重新组织了一遍。体验提升大很多。
它不只是 GUI 化——是把整个 Hermes 的使用体验重新组织了一遍。体验提升大很多。
Hermes Desktop — The Agent That Grows With You
1 02
Indigo: 每次黄老板的分享 AI Factory 都是重头戏！我们已经从追求 GPU 发展到了追求“G 兆瓦”，NVIDIA DSX™ 平台为 AI 工厂提供了全套方案，这是智能时代的「重工业」⚡️它不再局限于单一硬件，而是涵盖了将电力、土地、并网、冷却、光模块、GPU、HBM、先进封装、晶圆代工、网络以及云容量等压缩进同一个可交付系统的能力。NeoCloud 就是新时代的“硅基”地产开发商，财富密码都在图4️⃣
1 11
黄仁勋在台北举行的 GTC 2026 大会上的主题演讲，围绕“智能体 AI（Agentic AI）时代的全面到来”展开，发布了全新一代面向智能体计算的系统架构、处理器（CPU）、软件开发工具包以及个人电脑与机器人参考平台。本次演讲向行业宣告，现代计算的底层逻辑已由英伟达与台湾半导体供应链共同重构。无论是云端的数据中心、案头的个人电脑，还是工厂里的具身机器人，其底座都正在转变为以“智能体驱动、算力即营收、异构系统级设计”为核心的新工业体系。
1. 核心趋势：从生成式 AI 迈向智能体 AI (Agentic AI)
- 实用 AI 的普及与生产力激增：黄仁勋指出，AI 已经度过了早期的技术验证阶段，以智能体（Agents）为核心的实用 AI 正式落地。以软件工程领域为例，2026年 GitHub 上的代码提交量（Commits）较此前近乎三倍增长，AI 并没有减少就业，反而通过释放高达数万亿美元的生产力价值，促使企业雇佣更多工程师。
- 新型计算范式（Computing Pattern）：未来的计算机不再依赖传统的“点击与键入”应用流，而是由用户表达意图，再由大语言模型（LLM）作为“大脑”，结合外围控制框架（Harness）、工作内存、长期内存及各类工具共同执行任务。
2. 重磅硬件：Vera Rubin 架构与全栈 AI 工厂
- Vera Rubin 平台全面投产：作为接替 Grace Blackwell 的下一代超大规模计算平台，Vera Rubin 并非单一芯片，而是专为智能体时代打造的分布式、异构的多机架 Pod 级超级计算机。该系统现已进入全面量产阶段。
- 工业级 AI 工厂蓝图（NVIDIA DSX）：针对动辄数百亿美元、高达吉瓦（Gigawatt）级能耗的现代化 AI 数据中心，英伟达推出 DSX 基础设施蓝图。利用 Omniverse 数字孪生在开工前进行全仿真验证，引入 45°C 高温液冷、动态功率分配以及动态电网调节技术，实现“每瓦特代币（Tokens per Watt）”收益率的最大化。
- 供应链的质跃：得益于制程与组装工艺的革新，此前组装一台 Grace Blackwell 机架需要 2 小时，而现在组装 Vera Rubin 机架仅需 5 分钟。
3. 首款面向智能体的处理器：NVIDIA Vera CPU
- 颠覆传统 CPU 设计理念：以往的 CPU 是为人类（以秒为单位）设计的，而智能体（以纳秒为单位）在调用工具、检索数据库时对延迟极度敏感。
- Vera CPU 核心技术指标：
极致的单线程性能：拥有超高 IPC（每时钟周期指令数），单周期内可解码并执行 10 条指令。
高吞吐无缝互联：基于 88 核单片架构，摒弃了传统芯粒（Chiplet）带来的延迟损耗，核心间通信速度提升 50%。
低内存延迟：全球首款支持纠错功能（ECC）的 LPDDR5X 数据中心 CPU，内存延迟比传统 x86 架构降低 40%。
极高的能效比：将更多功耗预算留给直接产生营收的 GPU 计算。
- 实际性能表现：在处理主流的企业级负载时，Vera CPU 的数据库 SQL 查询速度提升至 3 倍，纽约证券交易所的实时流处理速度更是达到了 6 倍。
4. 软件与开源生态：企业级智能体工具包与 Neotron 3 Ultra
- 英伟达企业级智能体工具包（NEMO Agent Toolkit）：包含模型、框架、工具与运行环境四大模块，协助企业安全、合规地构建私有智能体系统。全新开
- 大模型 Neotron 3 Ultra：英伟达发布了新一代开源基础模型，该模型基于状态空间模型（SSM）与混合专家模型（MoE）的混合架构。其运行速度提高了 5 倍，算力成本降低了 30%。英伟达不仅开源模型，还一并公开了全部训练脚本与数据集。
- 现场展示了与 Cadence 联合打造的芯片设计验证智能体（Super Agent），将原本需要数周的 RTL 芯片代码验证与Debug周期，大幅缩短至数小时，效率提升超 40 倍。
5. 个人电脑（PC）的重构：RTX Spark 系列
- 重塑 40 年未变的 PC 架构：英伟达与微软、联发科（MediaTek）深度协作，推出全新的 RTX Spark 移动计算芯片（N1X）。该芯片基于台积电 3nm 工艺，集成了黑格尔（Blackwell）架构 RTX GPU、20 核 Grace CPU 以及 128GB 统一内存，提供 1 Petaflop 的本地 AI 算力，100% 兼容 Windows 与 CUDA 生态。
- 全线智能体硬件家族：除了 RTX Spark 笔记本外，还推出了可全天候本地运行智能体的桌面台式机，以及拥有 768GB 内存、具备 20 Petaflops 算力的桌面级工作站（DGX Station for Windows），用于本地训练和部署万亿参数模型。黄仁勋预测，十年后的 PC 将不再是单纯的软件启动器，而是成为类似家用超级计算机和全天候生活辅助智能体的形态（类比智能手机对传统手机的颠覆）。
6. 具身智能与物理 AI (Physical AI)
- Cosmos 3 物理 AI 基础模型：针对具身智能面临的“第一人称视角数据匮乏”痛点，英伟达推出 Cosmos 3 开源模型。采用扩散与自回归混合变换器（Mixture of Transformers）架构，能完美理解物理世界的视觉、声音与动作，生成符合物理定律的合成视频，以此作为机器人策略训练的闭环模拟器。
- 智能驾驶 AlphaMo 2：推出首款具备高级推理与自解释能力的自动驾驶大模型，现已接入全球近 80% 的汽车制造供应链及 97% 的出行服务网络。
- Isaac GR00T 人型机器人参考平台：为了降低高等教育与学术界的研发门槛，英伟达推出了高度集成的 GR00T 人形机器人参考硬件（单手 25 自由度，全身 31 自由度）。配合 Jetsen Thor 计算机、Isaac Lab 仿真环境和物理 AI 模组，研发人员可以在数小时内完成开发环境搭建并部署动作策略。
1. 核心趋势：从生成式 AI 迈向智能体 AI (Agentic AI)
- 实用 AI 的普及与生产力激增：黄仁勋指出，AI 已经度过了早期的技术验证阶段，以智能体（Agents）为核心的实用 AI 正式落地。以软件工程领域为例，2026年 GitHub 上的代码提交量（Commits）较此前近乎三倍增长，AI 并没有减少就业，反而通过释放高达数万亿美元的生产力价值，促使企业雇佣更多工程师。
- 新型计算范式（Computing Pattern）：未来的计算机不再依赖传统的“点击与键入”应用流，而是由用户表达意图，再由大语言模型（LLM）作为“大脑”，结合外围控制框架（Harness）、工作内存、长期内存及各类工具共同执行任务。
2. 重磅硬件：Vera Rubin 架构与全栈 AI 工厂
- Vera Rubin 平台全面投产：作为接替 Grace Blackwell 的下一代超大规模计算平台，Vera Rubin 并非单一芯片，而是专为智能体时代打造的分布式、异构的多机架 Pod 级超级计算机。该系统现已进入全面量产阶段。
- 工业级 AI 工厂蓝图（NVIDIA DSX）：针对动辄数百亿美元、高达吉瓦（Gigawatt）级能耗的现代化 AI 数据中心，英伟达推出 DSX 基础设施蓝图。利用 Omniverse 数字孪生在开工前进行全仿真验证，引入 45°C 高温液冷、动态功率分配以及动态电网调节技术，实现“每瓦特代币（Tokens per Watt）”收益率的最大化。
- 供应链的质跃：得益于制程与组装工艺的革新，此前组装一台 Grace Blackwell 机架需要 2 小时，而现在组装 Vera Rubin 机架仅需 5 分钟。
3. 首款面向智能体的处理器：NVIDIA Vera CPU
- 颠覆传统 CPU 设计理念：以往的 CPU 是为人类（以秒为单位）设计的，而智能体（以纳秒为单位）在调用工具、检索数据库时对延迟极度敏感。
- Vera CPU 核心技术指标：
极致的单线程性能：拥有超高 IPC（每时钟周期指令数），单周期内可解码并执行 10 条指令。
高吞吐无缝互联：基于 88 核单片架构，摒弃了传统芯粒（Chiplet）带来的延迟损耗，核心间通信速度提升 50%。
低内存延迟：全球首款支持纠错功能（ECC）的 LPDDR5X 数据中心 CPU，内存延迟比传统 x86 架构降低 40%。
极高的能效比：将更多功耗预算留给直接产生营收的 GPU 计算。
- 实际性能表现：在处理主流的企业级负载时，Vera CPU 的数据库 SQL 查询速度提升至 3 倍，纽约证券交易所的实时流处理速度更是达到了 6 倍。
4. 软件与开源生态：企业级智能体工具包与 Neotron 3 Ultra
- 英伟达企业级智能体工具包（NEMO Agent Toolkit）：包含模型、框架、工具与运行环境四大模块，协助企业安全、合规地构建私有智能体系统。全新开
- 大模型 Neotron 3 Ultra：英伟达发布了新一代开源基础模型，该模型基于状态空间模型（SSM）与混合专家模型（MoE）的混合架构。其运行速度提高了 5 倍，算力成本降低了 30%。英伟达不仅开源模型，还一并公开了全部训练脚本与数据集。
- 现场展示了与 Cadence 联合打造的芯片设计验证智能体（Super Agent），将原本需要数周的 RTL 芯片代码验证与Debug周期，大幅缩短至数小时，效率提升超 40 倍。
5. 个人电脑（PC）的重构：RTX Spark 系列
- 重塑 40 年未变的 PC 架构：英伟达与微软、联发科（MediaTek）深度协作，推出全新的 RTX Spark 移动计算芯片（N1X）。该芯片基于台积电 3nm 工艺，集成了黑格尔（Blackwell）架构 RTX GPU、20 核 Grace CPU 以及 128GB 统一内存，提供 1 Petaflop 的本地 AI 算力，100% 兼容 Windows 与 CUDA 生态。
- 全线智能体硬件家族：除了 RTX Spark 笔记本外，还推出了可全天候本地运行智能体的桌面台式机，以及拥有 768GB 内存、具备 20 Petaflops 算力的桌面级工作站（DGX Station for Windows），用于本地训练和部署万亿参数模型。黄仁勋预测，十年后的 PC 将不再是单纯的软件启动器，而是成为类似家用超级计算机和全天候生活辅助智能体的形态（类比智能手机对传统手机的颠覆）。
6. 具身智能与物理 AI (Physical AI)
- Cosmos 3 物理 AI 基础模型：针对具身智能面临的“第一人称视角数据匮乏”痛点，英伟达推出 Cosmos 3 开源模型。采用扩散与自回归混合变换器（Mixture of Transformers）架构，能完美理解物理世界的视觉、声音与动作，生成符合物理定律的合成视频，以此作为机器人策略训练的闭环模拟器。
- 智能驾驶 AlphaMo 2：推出首款具备高级推理与自解释能力的自动驾驶大模型，现已接入全球近 80% 的汽车制造供应链及 97% 的出行服务网络。
- Isaac GR00T 人型机器人参考平台：为了降低高等教育与学术界的研发门槛，英伟达推出了高度集成的 GR00T 人形机器人参考硬件（单手 25 自由度，全身 31 自由度）。配合 Jetsen Thor 计算机、Isaac Lab 仿真环境和物理 AI 模组，研发人员可以在数小时内完成开发环境搭建并部署动作策略。
4 02
FT: Tencent moves closer to launching AI agent for China’s most-used app
Maker of WeChat has fallen behind domestic rivals in artificial intelligence models
Tencent is moving closer to launching an embedded AI agent for WeChat, the ubiquitous app used by China’s 1.4bn people for everything from messaging and social media to ride-hailing and payments.
The company is testing a prototype of the agent, which can help users complete tasks within the app, and aims to start the compliance process required before a public launch as soon as this month, according to two people with knowledge of the plan.
After that, Tencent plans to test the agent on a small group of outside users before initiating a phased rollout. A public launch date has not been set because the amount of time required for the compliance process is uncertain.
Tencent shares rose 6 per cent in Hong Kong on Tuesday after the FT report was published.
Users will be able to access the chat box for the AI agent by swiping right on the main WeChat screen, according to a person who has seen an early demonstration.
They can then enter instructions for the agent to automatically tap into WeChat’s millions of mini-apps — the bedrock of the app’s broad functionality — and complete tasks such as finding a café and ordering a drink based on certain flavour and price requirements.
The successful launch of an AI agent on China’s most-used app could help Tencent leapfrog its domestic rivals, which have built stronger models and already implemented agents across their platforms.
Alibaba has integrated its ecommerce, travel and mapping services into its Qwen AI app, while ByteDance has added agentic functions such as online shopping to its Doubao app.
While Doubao and Qwen’s monthly active user numbers are far behind WeChat’s 1.4bn, they are growing rapidly, putting pressure on Tencent to release its AI agent sooner, according to the people familiar with the company’s plan. Tencent already has a chatbot with search functions, Yuanbao, embedded in WeChat.
The company had made the rollout its highest strategic priority, said one of the people familiar with the plan. Tencent’s executives are careful about getting every aspect of the AI agent right, which could lead to prolonged testing and many rounds of changes before the eventual launch.
Although the prototype was able to complete tasks smoothly, challenges remained in finding enough computing power to support a mass rollout, the person said. China’s internet giants had relied on Nvidia chips to power their AI data centres before they were banned for export to China.
“Tencent was too conservative to stock up on Nvidia’s chips before the US export ban, while domestic semiconductor supply remains tight,” said the person.
Early internal estimates suggest a full rollout of an AI agent on WeChat will be very costly, and it is unclear whether Tencent can generate enough revenue in the near term to cover the expense.
Recommended
Artificial intelligence
China’s big tech groups miss out on AI stock market frenzy
WeChat, known as Weixin in Chinese, has evolved since its launch as a messaging app in 2011 to become an indispensable platform embedded in nearly every aspect of daily life in China.
Through mini-apps that businesses and government entities set up, users can order delivery from restaurants, pay utility bills and schedule doctor appointments, meaning users rarely need to leave the app for everyday transactions.
In turn, Tencent has been able to gather vast amounts of behavioural data from WeChat’s billion-plus users.
“Beyond foundation models, it has become increasingly evident that agentic AI represents a breakthrough use case,” said Tencent’s president Martin Lau in an earnings call with analysts last month. “Our platform inherently has many benefits of hosting AI agents.”
Tencent declined to comment.
Maker of WeChat has fallen behind domestic rivals in artificial intelligence models
Tencent is moving closer to launching an embedded AI agent for WeChat, the ubiquitous app used by China’s 1.4bn people for everything from messaging and social media to ride-hailing and payments.
The company is testing a prototype of the agent, which can help users complete tasks within the app, and aims to start the compliance process required before a public launch as soon as this month, according to two people with knowledge of the plan.
After that, Tencent plans to test the agent on a small group of outside users before initiating a phased rollout. A public launch date has not been set because the amount of time required for the compliance process is uncertain.
Tencent shares rose 6 per cent in Hong Kong on Tuesday after the FT report was published.
Users will be able to access the chat box for the AI agent by swiping right on the main WeChat screen, according to a person who has seen an early demonstration.
They can then enter instructions for the agent to automatically tap into WeChat’s millions of mini-apps — the bedrock of the app’s broad functionality — and complete tasks such as finding a café and ordering a drink based on certain flavour and price requirements.
The successful launch of an AI agent on China’s most-used app could help Tencent leapfrog its domestic rivals, which have built stronger models and already implemented agents across their platforms.
Alibaba has integrated its ecommerce, travel and mapping services into its Qwen AI app, while ByteDance has added agentic functions such as online shopping to its Doubao app.
While Doubao and Qwen’s monthly active user numbers are far behind WeChat’s 1.4bn, they are growing rapidly, putting pressure on Tencent to release its AI agent sooner, according to the people familiar with the company’s plan. Tencent already has a chatbot with search functions, Yuanbao, embedded in WeChat.
The company had made the rollout its highest strategic priority, said one of the people familiar with the plan. Tencent’s executives are careful about getting every aspect of the AI agent right, which could lead to prolonged testing and many rounds of changes before the eventual launch.
Although the prototype was able to complete tasks smoothly, challenges remained in finding enough computing power to support a mass rollout, the person said. China’s internet giants had relied on Nvidia chips to power their AI data centres before they were banned for export to China.
“Tencent was too conservative to stock up on Nvidia’s chips before the US export ban, while domestic semiconductor supply remains tight,” said the person.
Early internal estimates suggest a full rollout of an AI agent on WeChat will be very costly, and it is unclear whether Tencent can generate enough revenue in the near term to cover the expense.
Recommended
Artificial intelligence
China’s big tech groups miss out on AI stock market frenzy
WeChat, known as Weixin in Chinese, has evolved since its launch as a messaging app in 2011 to become an indispensable platform embedded in nearly every aspect of daily life in China.
Through mini-apps that businesses and government entities set up, users can order delivery from restaurants, pay utility bills and schedule doctor appointments, meaning users rarely need to leave the app for everyday transactions.
In turn, Tencent has been able to gather vast amounts of behavioural data from WeChat’s billion-plus users.
“Beyond foundation models, it has become increasingly evident that agentic AI represents a breakthrough use case,” said Tencent’s president Martin Lau in an earnings call with analysts last month. “Our platform inherently has many benefits of hosting AI agents.”
Tencent declined to comment.
2 00
