宇树科技物理AI具身智能

10月22日，2026 世界物理 AI 大会暨宇树合作伙伴生态大会在杭州举行。宇树科技董事长、总经理兼 CTO 王兴兴发表主旨演讲，抛出一个重磅判断：物理 AI 目前已看到"被击穿"的苗头。

王兴兴指出，对顶尖大模型公司而言，某种意义上内部已基本实现模型自进化——模型自己写程序、采集数据，人只需适当引导参与。但公众感受不深，是因为从技术突破到扩散至每个人使用，需要 5-10 年时间。

他认为，具身智能迈过 ChatGPT 时刻的最核心挑战在于"AI 模型的输入输出和真实机器人的对齐"——即模型在虚拟世界中的推理能力如何精准映射到物理世界的传感器和执行器上。

科大讯飞1024开发者节国产AI

10月22日，2026 科大讯飞全球 1024 开发者节北京科博展在北京国家会议中心二期开幕。本届科博展以"以全国产 AI 赋能千行百业"为主题，集中展示讯飞星火 X2.6、AIPC、AI 眼镜等最新产品。

应用场景覆盖多个行业：

• 教育：AI 黑板、智能批改系统
• 医疗：中医 AI 辅助分析平台
• 汽车：智能座舱交互方案
• 企业：智能招采、智能决策平台、阡陌 AI 原生客服平台

讯飞还展示了机器人生态和开发者社区的最新进展，强调从底层芯片到上层应用的全栈国产化能力。

OpenAIGoogle Cloud企业AI

OpenAI 宣布与 Google Cloud 达成合作，将其前沿模型及 Codex 智能体工作流引入 Google Cloud 的 Model Garden。企业客户可在未来几周内通过 Google Cloud Marketplace 购买相关模型，并使用现有谷歌云支出承诺额度支付。

双方还计划将 OpenAI 模型接入 Google Antigravity 和 Gemini Agent，为企业提供更多模型选择。OpenAI 表示，此次合作将结合谷歌云的访问控制、数据治理工具及高性能基础设施，支持企业构建 AI 智能体工作流。

值得注意的是，微软仍是 OpenAI 的主要云合作伙伴，此次合作标志着 OpenAI 正在推进多云战略。

NVIDIADSX PODAI工厂

在 GTC 柏林大会上，英伟达发布了 DSX POD 参考设计——一套面向完整 Vera Rubin AI 工厂的模块化集装箱方案。计算、网络、线缆、机架级供电与液冷管路，全部由合作伙伴在工厂内预装集成，完成功能验证后整体发货。

核心数据：

• 场地就绪后部署周期从约 6 个月缩短至 14 天
• 配套 DSX Max LPS 方案，同等 Token 输出下最高节电 30%
• 采用闭环液冷 + 干式冷却，降低耗水
• 生态伙伴包括戴尔、HPE、联想、施耐德电气、富士康、广达等
• 合作伙伴计划 2027 年初在全球交付首批模块化 AI 工厂

黄仁勋表示："AI 工厂只有在开始产出 Token 时才创造价值，而不是在建造完工时。集装箱本身就是 AI 工厂。"

微软AI安全网络安全

微软在 2026 年 9 月威胁情报报告中披露，黑客组织 Storm-0940 使用 AI 框架 Artex 对全球数百家组织发起自动化攻击。这标志着 AI 工具被恶意行为者系统化利用的趋势正在加速。

Artex 框架能够自动化完成攻击链中的多个环节，包括目标侦察、漏洞扫描和载荷投递。微软警告企业应加强对 AI 驱动攻击的防御能力，特别是在身份验证和访问控制层面。

Unitree RoboticsPhysical AIEmbodied Intelligence

On October 22, the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference was held in Hangzhou. Unitree Robotics Chairman, General Manager and CTO Wang Xingxing delivered a keynote, offering a heavyweight judgment: physical AI is already showing signs of being "broken through".

Wang Xingxing noted that for top large-model companies, model self-evolution has basically been achieved internally in a sense—models write their own programs and collect data, with humans only needing to guide and participate appropriately. But the public does not feel it deeply because it takes 5-10 years from a technical breakthrough to diffusion to everyone's use.

He believes the core challenge for embodied intelligence to pass the ChatGPT moment is "aligning the inputs and outputs of AI models with real robots"—that is, how the model's reasoning ability in the virtual world maps precisely onto sensors and actuators in the physical world.

iFlytek1024 Developer FestivalDomestic AI

On October 22, the 2026 iFlytek Global 1024 Developer Festival Beijing Science and Technology Expo opened at the second phase of the China National Convention Center in Beijing. Under the theme "Empowering all industries with fully domestic AI", the expo showcases the latest products including iFlytek Spark X2.6, AIPC and AI glasses.

Application scenarios cover multiple industries:

• Education: AI blackboards and intelligent grading systems
• Healthcare: a TCM AI-assisted analysis platform
• Automotive: smart cockpit interaction solutions
• Enterprise: intelligent procurement, intelligent decision-making platforms and the Qianmo AI-native customer service platform

iFlytek also showed the latest progress in its robotics ecosystem and developer community, emphasizing full-stack domestic capabilities from underlying chips to upper-layer applications.

OpenAIGoogle CloudEnterprise AI

OpenAI announced a partnership with Google Cloud to bring its frontier models and Codex agent workflows to Google Cloud's Model Garden. Enterprise customers can purchase the models through Google Cloud Marketplace within the next few weeks and pay using existing Google Cloud spending commitments.

The two sides also plan to integrate OpenAI models into Google Antigravity and Gemini Agent, giving enterprises more model choices. OpenAI said the partnership will combine Google Cloud's access controls, data governance tools and high-performance infrastructure to help enterprises build AI agent workflows.

Notably, Microsoft remains OpenAI's primary cloud partner; this partnership signals that OpenAI is advancing a multi-cloud strategy.

NVIDIADSX PODAI Factory

At the GTC Berlin conference, NVIDIA released the DSX POD reference design—a modular containerized solution for a complete Vera Rubin AI factory. Compute, networking, cabling, rack-level power and liquid-cooling piping are all pre-installed and integrated by partners in the factory, then shipped as a whole after functional verification.

Core figures:

• Deployment cycle cut from about 6 months to 14 days after site readiness
• Paired with the DSX Max LPS solution, saving up to 30% power for the same token output
• Uses closed-loop liquid cooling plus dry cooling to reduce water consumption
• Ecosystem partners include Dell, HPE, Lenovo, Schneider Electric, Foxconn and Quanta
• Partners plan to deliver the first modular AI factories worldwide in early 2027

Jensen Huang said: "An AI factory only creates value when it starts producing tokens, not when construction is finished. The container itself is the AI factory."

MicrosoftAI SecurityCybersecurity

In its September 2026 threat intelligence report, Microsoft disclosed that the hacking group Storm-0940 used the AI framework Artex to launch automated attacks against hundreds of organizations worldwide. This marks an accelerating trend of AI tools being systematically exploited by malicious actors.

The Artex framework can automate multiple stages of the attack chain, including target reconnaissance, vulnerability scanning and payload delivery. Microsoft warned enterprises to strengthen defenses against AI-driven attacks, especially in authentication and access control.