AI News

苹果公司警告称，AI数据中心的快速扩张导致内存需求激增，从未见过某个零部件价格如此大幅快速上涨。电子产品行业面临前所未有的挑战。这一表态反映AI算力需求正从GPU扩展到存储全链路，整个硬件供应链正被AI重塑。

美团LongCat团队发布VitaBench 2.0，这是首个真实生活场景下面向长期动态用户建模的智能体评测基准。它系统性地评测大语言模型在长期、真实、动态用户互动中个性化与主动性的能力，填补了AI Agent长期交互评测的行业空白。

夏季达沃斯论坛上，德勤中国CEO刘明华表示中国Token调用量连续7周霸榜全球。麦肯锡中国区主席倪以理表示中国企业AI应用领跑全球，对科技的信任感和拥抱速度超过任何地方。中国贡献了全球约60%的电动车销量，AI应用正加速从实验室走向产业。

腾讯云旗下EdgeOne Makers登顶全球科技产品社区Product Hunt日榜第一。该平台原生支持Web应用与AI Agent从开发、部署到托管的全流程，此前已服务超30万名用户、累计托管站点超60万个。本次升级将托管范围从Web延伸至AI Agent，标志AI Agent基础设施走向成熟。

A股存储芯片板块全天强势领涨，普冉股份、兆易创新、江波龙等16只存储概念股同步创历史新高。存储板块年内总市值增长2.1万亿元，增量约等于1.4个贵州茅台市值。4只龙头股股价突破700元，AI算力需求正推动存储成为继光通信之后又一个高价股聚集赛道。

微信正式开启AI助手"小微"小范围内测，左上角星标入口被原生AI助手取代，右滑即可唤起交互。主模型为腾讯自研WeLM，部分场景由DeepSeek模型补充响应。这是微信生态全面拥抱AI的标志性动作，拥有超13亿用户的微信入口级AI助手将深刻改变AI应用分发格局。

宇树科技宣布双足人形机器人Unitree R1官方售价由3.99万元下调至2.99万元起，即日起开放现货销售。该机器人重25kg，配备26个高精度关节，支持语音与图像多模态大模型交互，具备开发与定制能力。人形机器人价格持续下探，正在从实验室走向消费级市场。

Sensor Tower《2026年AI现状报告》显示，ChatGPT市场份额已首次跌破50%。随着Google Gemini、Anthropic Claude、xAI Grok等竞品快速崛起，用户正从ChatGPT向多元化工具切换。这一里程碑事件标志AI助手市场从"一家独大"走向"群雄逐鹿"的新阶段。

字节跳动旗下豆包发布2.1 Pro版本，主打面向普通人的代码生成能力。实测显示，豆包2.1 Pro在代码理解、生成和调试方面均有显著提升，让非专业开发者也能通过自然语言完成编程任务。AI编程工具正从专业开发者向全民化方向演进。

Cursor在首届Compile大会上推出Origin代码存储与Git托管平台，核心设计将AI Agent视为一等公民而非人类开发者的附属角色。Origin重新定义了AI时代的代码基础设施，标志AI编程工具正从单一编辑器向完整开发平台演进，对GitHub等传统平台构成直接挑战。

NVIDIA发布BioNeMo Agent Toolkit，用于加速科学发现的AI Agent工具包。Dassault Systèmes、Databricks、礼来、OpenAI、Schrödinger、Snowflake、UW Medicine蛋白质设计研究所等行业领袖已采用，Anthropic和OpenAI正在集成。该工具包为科学研究提供Agent能力，标志AI Agent正从通用编程向专业科学领域深度渗透。

SK海力士开始放缓第六代HBM4扩产步伐，将更多产能转向通用型DRAM。其HBM营收占比已超40%，在市场中占据明显优势，无须继续激进扩产。相比之下，通用型DRAM供应严重短缺，盈利能力已超过HBM。SK海力士已推迟部分HBM3E产线转产HBM4的时间，优先提高通用DRAM供应能力。这反映AI基础设施需求正从单一HBM向全存储链路扩展。

Claude Code近期的bug修复揭示了AI编程Agent落地面临的深层挑战：工具状态管理、权限边界和后台任务控制正在成为AI编程产品的新考验。随着AI Agent从简单代码补全走向复杂任务执行，如何确保Agent在受控范围内运行、如何管理长时间后台任务、如何让用户信任Agent的自主操作，已成为整个行业必须解决的核心问题。

LiblibAI母公司演语科技完成近3亿美元融资，正在成为这一轮AI变化中最值得观察的中国样本之一。这轮融资释放明确信号：AI应用层开始进入"收入说话"的阶段，纯技术叙事已不足以打动投资人，商业化能力和真实收入成为估值核心依据。

微软连续释放两条重磅消息，选择与DeepSeek展开合作。在AI烧钱压力持续加大的背景下，这一决策反映科技巨头正重新评估AI投入产出比，开始寻求更具性价比的模型合作方案。DeepSeek作为开源大模型代表，其低成本高性能特点正吸引越来越多巨头关注，可能重塑AI产业合作格局。

SpaceX已与开源AI初创企业Reflection AI签署算力合作协议，最高价值可达63亿美元。根据协议，Reflection将立即获得英伟达GB300算力。这是AI算力交易领域迄今最大规模的合作之一，标志超级科技巨头正加速布局AI算力基础设施，开源AI公司也开始获得前所未有的算力资源支持。

美光科技宣布与Anthropic达成战略合作协议，共同扩展下一代AI基础设施。合作涵盖AI存储供应、技术研发及股权投资三大维度。这是存储芯片巨头与AI大模型公司的首次深度绑定，反映AI算力需求正从GPU扩展到存储全链路，产业链上下游加速整合。

英伟达发布NVIDIA Halos for Robotics，这是业内首套将AI算力和安全能力整合在一起的全栈机器人安全系统，面向机器人和物理AI的开发、验证和部署。该系统为机器人安全提供从开发到验证的完整工具链，标志具身智能领域从追求能力转向重视安全，为行业提供标准化安全基础设施。

英伟达宣布欧洲创纪录的35台AI HPC超级计算机正式启动建设，建成后超过300万名研究人员将获得下一代算力基础设施，用于推动全欧洲AI发展和科学研究。这是欧洲在AI算力领域最大规模的一次性投入，旨在缩小与美国在AI计算能力上的差距，加速欧洲AI科研生态建设。

智谱盘中市值突破万亿港元，跻身港股万亿市值俱乐部。此前这里的常客包括腾讯、阿里、汇丰等巨头，智谱成为其中最年轻、大众接触最少的成员。从上市到万亿仅用166天，GLM-5.2的跑分表现和国资背景是核心驱动力，但市场对其估值泡沫的争议也从未停歇。

强化学习先驱Richard Sutton与传奇游戏程序员John Carmack达成合作，目标是让机器人在真实世界中玩游戏。这一跨界合作直击当前具身智能的核心痛点：AI在模拟器中学到的能力难以迁移到不可预测的真实世界。此前人形机器人伤人事件频发，傅盛称以当前大模型能力两三年内难以解决安全问题，Sutton和Carmack的联手或为这一困局提供新思路。

AlphaFold核心领导者、诺贝尔奖得主John Jumper宣布离开工作近9年的Google DeepMind加入Anthropic。Hassabis回应称感谢John过去9年的非凡伙伴关系。这是继Transformer论文联合作者之后，DeepMind又一位顶级人才流向Anthropic。AI大模型领域的人才争夺战已进入白热化阶段，顶级研究者成为最稀缺资源。

字节跳动正与天数智芯洽谈采购至少5万颗AI推理芯片，主要用于推理任务。字节2026年计划将资本开支上调超2000亿元，是国内AI算力头号买家。如果这笔交易落地，天数智芯将成为继华为和寒武纪之后字节第三大芯片供应商。国产GPU头部牌桌正在悄悄加人，AI算力国产替代进程显著加速。

Sensor Tower发布《2026年AI现状报告》显示，ChatGPT市场份额首次跌破50%。随着用户在谷歌Gemini、Anthropic Claude以及xAI Grok等工具之间切换，OpenAI的绝对统治地位正被瓦解。尽管ChatGPT已成为史上最快达到10亿月活的应用，但竞品崛起正在重塑AI应用市场格局，多极化竞争时代来临。

智谱发布并开源GLM-5.2，专为长程任务能力打造。该模型在Code Arena评估系统上取得全球可用模型第一，1M上下文稳定支撑长程任务。在FrontierSWE测试集上仅比Claude Opus 4.8低1%，超过GPT-5.5和Opus 4.7。GLM-5.2是排名最高的开源模型，标志着国产大模型在代码工程能力上跻身全球顶级梯队。

谷歌前工程副总裁、Gemini技术联席负责人诺姆·沙泽尔宣布离职加入OpenAI。沙泽尔表示离开谷歌是艰难决定，为团队取得的成果感到自豪。OpenAI CEO奥尔特曼随后发文欢迎，称从OpenAI创立之初沙泽尔就是他最希望合作的人之一。这是AI大模型领域又一次重磅人才流动。

菜鸟在荷兰启用全新欧洲供应链履约中心，投入超100台自主研发的ZeeBot攀爬机器人，同时服务B端和C端履约。该中心签下10年租约，场地总面积2.6万平方米，辐射整个欧洲市场。这标志着机器人革命正在物流领域率先从概念走向规模化落地。

高通CEO安蒙认为AI智能体将改变人们使用应用和电子设备的方式。他设想的未来入口是内置摄像头和显示屏的智能眼镜，用户无需掏出手机即可完成预订、支付等操作。安蒙认为应用不会消失但会发生变化，智能体将成为新的应用入口，智能眼镜市场体量有望比肩手机。

SK海力士罕见启动数百人规模芯片设计工程师招聘，申请将于6月23日截止。业内担心大型企业可能吸走刚进入行业一两年的年轻工程师，半导体设计人才本就稀缺。多家中小芯片设计公司CEO表示不安，AI算力需求正在重塑韩国半导体人才格局。

2026年6月9日Anthropic发布Claude Fable 5和Mythos 5。Fable 5是其迄今公开提供过的最强模型，在软件工程、知识工作、视觉和长任务上大幅领先。同时Mythos 5只面向少数可信伙伴开放，体现AI能力越强安全管控责任越大的行业共识。

Anthropic高管向商务部长霍华德·卢特尼克提交方案，承诺与白宫展开更深度合作，力求尽快消除安全隐患。正是相关安全问题使得该公司高性能AI模型"Mythos"与"Fable"遭到监管整治。Anthropic在会谈中承诺将加强与特朗普政府的沟通，未来更快处置各类安全风险。

英伟达在GTC Taipei 2026大会上发布Vera CPU，同期发布新一代AI超算平台Vera Rubin，首批客户包括OpenAI和Anthropic。这是英伟达第一次推出独立CPU产品线，过去20年增长几乎完全建立在GPU之上。CPU市场格局正被AI算力需求重新定义，1700亿美元市场迎来上位大戏。

据报道，Meta与数据中心公司Crusoe签署AI算力合作协议，预计将获得约1.6吉瓦算力。这是Meta在AI基础设施领域的又一重大布局，科技巨头对AI算力的争夺持续升温，1.6GW的算力规模相当于大型数据中心集群，反映AI训练和推理对算力的需求仍在指数级增长。

Waymo宣布召回数千辆自动驾驶出租车，隐患为车辆可能高速闯入高速公路施工路段。这家隶属于字母表集团的自动驾驶公司针对旗下部分搭载第五代自动驾驶系统的车辆发起召回。此前已发生多起安全事件：车辆未能识别匝道封闭标识，径直驶入规划中的高速公路施工区域。

过去两年Token的生产成本一直在下降，但售价却一直涨。看起来前后矛盾的事，背后是一笔基于Token商业模式逐渐成熟而成立的生意。硬件和模型不断优化迭代，几乎都指向同一方向：更便宜的推理、更低廉的算力。但另一面是逐渐成熟的智能体以及已经闭环的业务，真正靠AI赚到钱的人根本不会被价格劝退。

SpaceX在IPO上市四天后宣布拿出600亿美元股票收购Cursor。600亿美元约等于港股智谱全部市值，也超过MiniMax、月之暗面、阶跃星辰几家一级市场估值之和。Cursor基于中国开源模型Qwen底座打造，此次收购是AI开发工具领域迄今最大规模交易，标志超级科技巨头正加速整合AI编程基础设施。

Sensor Tower《2026年AI现状报告》显示，随着用户在Google Gemini、Anthropic Claude和xAI Grok之间切换，ChatGPT市场份额首次跌破50%。尽管ChatGPT已成为最快达到10亿月活的应用（周活9亿、月活超11亿），但Gemini月活已达6.62亿，Claude和Grok也在快速增长，AI聊天市场正从一家独大走向多极竞争。

Cursor在首届Compile大会上推出Origin代码存储与Git托管平台，将AI Agent视为一等公民而非人类开发者的附属角色。传统Git平台为人类设计，每天几次到几十次提交；而AI Agent加入开发流程后，代码基础设施需要重新设计以支持Agent的高频提交、自动审查和协作模式。

Meta首席AI科学家杨立昆（Yann LeCun）接受CNBC采访时表示，当前AI模型价格持续上涨，但运营公司仍在持续亏损，成本下降速度无法跟上市场预期。大量AI服务的低价全靠投资人补贴，许多公司尚未建立可持续商业模式。他警告OpenAI和Anthropic要么提价要么降本，否则巨大泡沫将破裂。

英特尔宣布任命李锡熙（Seok-Hee Lee）为代工执行副总裁，直接向CEO陈立武汇报。李锡熙将全面负责先进封装、系统集成、后端技术开发与制造业务，提升英特尔在AI计算领域的差异化系统级创新能力。先进封装和系统集成正成为下一代AI计算系统的关键能力。

字节跳动正与天数智芯讨论采购至少5万颗AI芯片，主要用于推理工作。若交易达成，天数智芯将成为华为和寒武纪之后字节跳动的第三家GPU供应商。字节跳动数据中心建设进展再添新消息，对应天数智芯智铠系列云端推理GPU，反映互联网大厂正加速多元化算力布局。

据报道，微软正考虑引入深度求索V4大模型作为Copilot协同办公的低成本备选方案，用来替代目前为Copilot提供技术支持的Anthropic和OpenAI大模型。微软正在测试由微软云本地部署运行的深度求索V4版本以及其他开源大模型，计划在未来几周内推出低成本大模型服务。

宇树科技亚洲首家具身智能体验馆在南京西路开业，不到两周智元机器人全球线下首店在闵行漕宝路亮相。同样的开店呈现截然不同的温度与节奏。当前人形机器人价格持续下探至消费级区间，行业又恰逢IPO冲刺关键节点，线下零售成为机器人商业化的新战场。

AI火热，"All in AI"成了不少CEO的口头禅。但现实要冷峻得多，企业发现真正让AI在组织里跑起来远比想象中困难。最直接的问题是"燃料"没准备好——核心资料和数据散落在各系统中，大量关键经验沉淀在老员工脑子里，组织内部各自为战。

OpenAI、Google DeepMind、Anthropic三大AI巨头CEO与G7领导人在法国阿尔卑斯山共进工作午餐，历史首次。上一次这些领导人坐在一起讨论的是地缘冲突和全球供应链问题，现在AI公司CEO被请到了同一张桌子上，标志AI已上升为全球核心经济议题。

Google Cloud宣布推出开放知识格式OKF v0.1版本，以Markdown+YAML定义AI智能体知识交换标准，同步在GitHub以Apache 2.0许可证开源规范全文、参考实现和示例数据集。该格式旨在解决不同AI智能体之间知识无法互通的孤岛问题，让智能体能以统一格式读写共享知识。

苹果M4芯片中的神经网络引擎原本只开放给推理使用，开发者可运行预训练模型却不能直接训练新模型。安全研究者@0x0SojalSec自称绕过M4软件限制，解锁了神经网络引擎15.8TFLOPS的完整AI训练算力，引发对苹果芯片开放程度的讨论。

阿里巴巴发布千问具身智能大模型Qwen-Robot系列，包含VLA操作模型Qwen-RobotManip、VLN移动模型Qwen-RobotNav和世界模型三大组件，覆盖机器人操作、导航和环境理解全链路。这是国产具身智能基础模型从空白走向体系化的标志性发布。

AI推理公司硅基流动宣布完成超20亿元B轮融资，由携程战投、晶科能源控股、金蝶股份、联通新沃、壁仞战投、蔚来资本、商汤战投、巨人网络、国泰君安创新投、纪源资本等十余家机构联合投资，创国内AI推理赛道单轮融资纪录，资本对AI推理基础设施赛道信心持续升温。

字节跳动火山引擎旗下火山方舟体验中心上线Seedance 2.0 Mini视频生成模型，主打高性价比市场定位，计划近期对外开放API服务，瞄准更广泛的视频创作与规模化生产市场。这是字节在AI视频生成赛道的最新迭代，进一步降低视频生成成本门槛。

6月5日全球资本市场遭遇罕见血洗，标普500下跌2.64%，纳斯达克重挫4.18%创2025年4月以来最大单日跌幅。费城半导体指数单日暴跌10.3%创2020年3月以来最差表现，英伟达等芯片巨头遭重创。AI泡沫论再次成为市场焦点，但分析认为AI需求基本面并未改变，调整更多源于估值过高和地缘风险。

清华五道口附近的AI原点社区呈现中国AI创业的真实生态：10楼咖啡店"原点Bar"不断有新人涌入交换项目和AI行业见闻，楼下工区里估值百亿的公司和还在手搓第一个demo的团队共用同一部电梯。两种时间刻度交织，折射出AI创业的狂热与务实。

当整个AI行业都在比拼模型参数和算力时，一个17岁印度高中生用近乎恶作剧的方式创造了2026年最幽默的产品。它没有大模型、没有融资、没有GPU、甚至没有AI，但上线一个月获超2500万独立访客和2.8亿次访问，成为"反向创新"和互联网幽默的极致案例。

当审美成为AI时代的关键竞争力，设计审美教育不只针对设计师同样需要公众参与。好的设计以润物细无声的方式改变认知，从矿泉水包装到城市公共建筑。当AI能生成一切设计时，审美判断力反而成为人类最核心的竞争力，审美教育的重要性前所未有。

国外已用AI管理卫星运营，中国尚未建立成熟的自动化体系。中国长期运行通信、导航、遥感卫星具备基础能力，但巨型星座时代运营对象从几颗变几千颗，卫星间需共享链路争夺通信窗口。万星组网计划推进要求中国加速建立AI驱动的卫星运营自动化体系。

6月14日下午，外媒曝蚂蚁集团正秘密测试AI版支付宝，新版本颠覆原有用户交互，可一键切入原生AI界面，实现从服务到资金管理的智能化。消息称上线时间未定，蚂蚁集团拒绝评论。这是支付宝向AI原生应用转型的重要信号。

受全球人工智能热潮推动存储芯片需求激增，SK海力士去年新增员工超2000人。截至2025年底员工总数达34549人，较上年增加2159人。在韩国整体就业市场放缓的大背景下，此次扩招尤为引人注目，凸显AI对半导体产业链的强劲拉动效应。

一名加拿大母亲在美国法院起诉OpenAI及CEO奥特曼，指控ChatGPT诱导其女儿自杀。近期已有多起诉讼指责该公司未能管控用户与聊天机器人之间的危险对话，本案是最新一例，引发对AI产品安全责任边界的深度讨论。

阿里巴巴集团正式回应称"周靖人辞职"纯属谣言，表示已注意到近几日有人在网络上组织扩散该谣言，呼吁大家不要传播不实信息。周靖人此前负责阿里AI相关业务，其去留问题引发行业广泛关注。

德国数字化部长卡斯滕·维尔德贝格尔因疑使用AI工具完成演讲稿等多份公文遭到质疑。据德国《时代》周报报道，基于专门检测AI生成内容的软件分析，多篇署名文章似乎主要由AI生成，引发对政府官员使用AI工具的伦理讨论。

智谱宣布GLM-5.2面向GLM Coding Plan全量用户开放覆盖Lite/Pro/Max/团队版全部版本。API将于下周上线，模型正式开源遵循MIT协议。这是国产大模型在开源领域的又一重要动作，MIT协议意味着商业使用零限制。

月之暗面宣布发布并开源Kimi K2.7 Code编程模型。1M token标准输入和输出价格与K2.6一致分别为6.5元和27元，命中缓存输入价格调整为1.3元。基准评估显示K2.7 Code相比K2.6显著提升长上下文编程能力，6倍速高速版下周一上线。

美国商务部长卢特尼克向Anthropic发出出口管制命令，要求立即切断所有外国人访问Claude Fable/Mythos 5的权限，包括公司内部外籍员工。理由是出现了一种越境访问风险。美国政府与Anthropic陷入监管困境，这一事件凸显AI出口管制在技术层面的执行难度。

在2026长三角机器人及自动化展览会暨无锡具身智能机器人产业链伙伴大会上，科大讯飞发布星火多模态大模型X2-VL。这款国产多模态大模型将为无锡具身智能产业搭载硬核AI大脑，推动机器视觉与语言能力的融合应用。

英特尔持续收缩开源版图，AI加速开源项目BigDL被列入终止清单。BigDL是专注于在英特尔全系XPU上低延迟运行AI大语言模型的开源框架，本周仓库已挂出不再由英特尔维护的标识。这是英特尔继多个开源项目后的又一次收缩。

Anthropic发布Claude Fable 5仅24小时后安全过度过滤问题全面爆发。纯数学概念Selmer群和同构被系统判定为网络安全风险拒绝回答，免疫学家Derya Unutmaz批评这比生物封锁更反乌托邦。开发者反馈大量学术和医疗场景被误封，引发对AI安全边界过度收紧的广泛讨论。

小米MiMo团队发布终端编程Agent产品MiMo Code并采用MIT协议开源。该产品基于OpenCode构建定位面向长程自动化编程任务的终端编程Agent核心解决AI编程Agent在数十步甚至上百步持续执行中的决策质量和状态连续性问题。开源2周GitHub星标突破5100但开发者反映实际使用中bug较多引发热议。

英伟达在GTC Taipei 2026上发布面向Windows PC的全新超级芯片RTX Spark宣告正式进军个人电脑核心处理器市场。微软站台赋予官方钦定意味Acer华硕戴尔技嘉惠普联想微星七大PC厂商全部支持。这是英伟达从GPU供应商向PC核心平台商转型的标志性事件意图重新定义AI PC架构。

自变量机器人累计融资金额接近30亿元投资方包括小米战投红杉中国且同时获得字节跳动美团阿里巴巴小米四大厂投资。但公司展示的家庭机器人场景中叠枕巾等基础操作仍未攻克暴露了具身智能领域资本热情与技术现实之间的巨大落差。机器人从实验室走向家庭仍需跨越工程化的鸿沟。

Anthropic宣布Claude Fable 5在6月22日后将从所有订阅计划中移除继续使用需单独购买用量积分。这意味着即使付费会员旗舰模型也只提供14天免费使用。一个模型发布当天自带逐客令在大模型行业尚属首次引发对AI服务商业模式可持续性的深度讨论。分析认为高算力成本使固定月费模式难以为继。

KKR集团、科威特投资局、英伟达与维斯特拉联手成立赫利克斯数字基础设施公司，旨在为大型AI云服务商提供配套基础设施，统筹数据中心、电力、网络连接等需求。公司已获超100亿美元长期资金承诺，英伟达将部署适配其DSX AI工厂的基础设施，维斯特拉负责电力供应。

亚马逊、谷歌、微软、Meta、甲骨文五家巨头2024至2027年AI基础设施累计资本支出达2.45万亿美元约合人民币16.6万亿元。但经济学研究表明若AI生产率未能在五年内至少提升2.7倍，这些公司将面临严峻财务压力。AI产业已从模型开发进入重资产竞赛阶段。

亚马逊创始人贝索斯认为AI不会让人类大规模失业反而会带来劳动力短缺。他正联合创办新AI公司Prometheus，目标是打造通用AI工程师用于设计和制造喷气发动机等复杂实体产品，赋能工程师让发明变得更快，使更小团队在更短周期内完成更大规模的事情。

AMD CEO苏姿丰在MIT毕业典礼发表演讲，表示无论AI能力变得多强，都无法替代人类判断哪些问题真正值得解决，也无法承担最终责任。她强调在大型科技公司全面押注AI的当下，人类的判断力和责任感比以往任何时候都更加重要。

Apache Burr目前处于孵化阶段，提供一套轻量级Python工具集解决AI Agent应用从原型到产品的核心问题：管理复杂状态、追踪决策路径、在生产环境中可靠持久化对话历史。包含零依赖Python库、执行遥测UI和一系列集成，使开发者能用简单Python函数构建状态机。

据The Information报道OpenAI正深入谈判计划租赁俄亥俄州南部一个10GW数据中心园区总成本至少5000亿美元英伟达可能提供资金支持。根据长期租赁协议OpenAI将控制设施内所有设备项目一期预计2028年投产。这一规模的数据中心将是全球最大的AI算力基础设施之一反映AI训练和推理对算力的需求正呈指数级增长。

Anthropic正式发布Claude Fable 5定位为Mythos级安全可用模型。这是Anthropic首次将最高能力等级模型降级后向公众开放同时保留给政府合作方一个不受限制的版本Mythos。Wharton商学院教授Ethan Mollick获得早期访问权限后评价其能力超过此前所有向公众开放的模型标志着AI能力边界再次拓展。

工信部印发《人工智能+信息通信创新发展实施意见（2026-2028年）》核心目标是到2028年初步实现信息通信网络高等级自智城域算力1毫秒时延圈覆盖率不低于75%。到2030年AI与信息通信网络融合关键核心技术取得显著突破。AI大模型训练和智算集群互联对网络带宽时延功耗提出极致要求传统电互连架构已遭遇带宽墙和功耗墙瓶颈。

React团队在GitHub上合并了备受关注的Pull Request将React Compiler移植到Rust。该项目由React核心团队成员开发是一次由AI主导编码人类紧密指导的尝试。这标志着前端核心工具链正在进入Rust化新阶段Rust的性能优势和内存安全特性正在被越来越多的基础设施项目采用。

虎嗅深度文章拆解宇树机器人产品分析其商业模式和技术路线。美国研究机构将宇树比作机器人领域的比亚迪和大疆通过极致的成本控制把原本属于实验室的技术推向大众市场。支持者认为宇树通过低价策略打开消费级市场质疑者则认为低价来自性能妥协。宇树全国首家直营店已在北京王府井开业人形机器人商业化进程正在加速。

由重庆国资委赛力斯集团字节跳动宁德时代等联合控股的赛豆科技在北京发布AI汽车品牌AIVA取自Artificial Intelligence Voyage Ahead寓意AI伴你一路前行新品牌以AI原生汽车为核心卖点强调先有AI再有汽车搭载豆包大模型智能座舱等AI能力。赛力斯此前深度绑定华为鸿蒙智行此次转向表明AI大模型正在成为汽车行业的核心竞争力。

北京中科闻歌科技股份有限公司正式通过港交所上市聆讯。公司成立于2017年由中国科学院自动化研究所科学家团队创立专注于研发企业级大模型驱动的决策智能操作系统及服务。中科闻歌是继多家大模型企业之后又一家冲刺资本市场的人工智能公司反映AI产业化进程加速。

虎嗅刊发深度技术文章回顾2017年Attention Is All You Need论文的发表历程详解Transformer架构中的自注意力机制及其对AI行业的深远影响。从无人注意到改变整个AI产业Transformer已成为大语言模型的基础架构推动了GPU算力需求的爆炸式增长和大模型时代的到来。

2026年6月2日美国总统特朗普签署推动先进人工智能创新与安全行政令。文章深度解析该行政令的政策背景核心内容及潜在影响分析美国在AI领域的战略转向。面对AI技术高速迭代创新提速与安全管控如何平衡成为全球各国的核心难题该行政令的出台将影响全球AI产业格局。

苹果在WWDC 2026期间公布Apple Intelligence平台重大架构升级核心是基于Google Gemini系列模型技术深度合作开发的Apple Foundation Models。新模型可同时运行于设备端和云端在理解能力推理能力及多模态支持上达到前沿水平。苹果同时发布Core AI设备端AI推理框架整合CPU GPU与神经网络引擎实现异构计算协同。

苹果WWDC26全球开发者大会正式发布iOS 27 iPadOS 27 macOS 27 watchOS 27 visionOS 27全新Siri AI由苹果智能驱动拥有独立App对话入口支持开放式问题解答理解用户个人背景跨应用执行操作在线查阅信息视觉智能和自定义声音等功能Siri AI将于今年秋季推出英文版今年晚些时候推出国行无缘首发。

OpenAI CEO奥尔特曼与首席科学家帕霍茨基联合发布博客文章宣布公司进入第三发展阶段第一阶段聚焦通用AI技术研发第二阶段面向全球推出产品并研究实际使用场景第三阶段核心是让先进AI普及化易用化且安全可靠奥尔特曼认为经济格局正围绕AI迎来重塑。

6月8日同程旅行宣布将全面接入微信AI智能体生态成为首批接入微信AI生态的在线旅游平台用户可在微信AI智能体内通过语音或文字指令提出旅行需求AI可完成行程规划旅行产品对比决策及预订服务携程小程序也已作为首批内测开发者完成初步接入适配未来将围绕酒店预订机票查询旅游度假等场景接入。

湖南昆仑元人工智能软件有限公司KunlunMeta宣布完成5000万元战略轮融资由A股上市公司景嘉微电子全资子公司长沙景美集成电路设计有限公司领投双方将聚焦芯片设计算力适配产品联合研发生态共建等领域合作依托景美国产GPU算力优势与昆仑元AI的软件栈和场景落地能力构建国产GPU加全栈AI基础设施一体化方案。

英伟达CEO黄仁勋表示SK海力士到2030年将晶圆产能翻倍的计划仍不足以满足AI芯片需求。此前黄仁勋在台北Computex展会上宣布与Marvell CEO合作推动AI芯片发展全球投资者密切关注其行程动向被视为AI产业风向标黄仁勋到访之处相关板块股价应声上涨。

京东与腾讯将围绕AI Agent展开深度合作。依托京东的商品供应链和履约服务能力及腾讯的生态入口优势双方共同打造跨场景的智能化服务新范式推动AI Agent从单点应用走向生态协同。

英伟达CEO黄仁勋访问韩国宣布与SK集团将于周一公布合作计划黄仁勋确认新款Vera芯片将采用SK海力士内存产品同时将于次日会见三星电子副会长全永铉讨论HBM和机器人合作计划黄仁勋表示存储芯片供应紧张局面还将持续好几年。

优必选科技宣布旗下消费级人形机器人品牌优世界超仿生人形机器人首发6天累计预订单量超2110台标志着人形机器人商业化进程加速消费级市场正在打开。

全球最大互联网托管服务商之一发布的最新数据显示在其托管网站收到的网络访问请求中约57.4%来自AI和自动化程序42.6%来自人类用户这是互联网历史上人类网络请求量首次被机器超越标志着AI流量时代的到来。

2026年上海高考作文题出炉财联社使用DeepSeek V4字节跳动豆包13.6阿里Qwen3.7月之暗面Kimi2.6 OpenAI GPT-5.5谷歌Gemini 3.5 Flash六个中外AI大模型进行实测最终DeepSeek与谷歌Gemini获得最高评分。

字节跳动官方发声明否认近期部分报道将赛豆解读为豆包汽车品牌或字节跳动造车明确表示字节跳动没有造车或推出汽车品牌的计划豆包和火山引擎与汽车行业合作伙伴的合作主要是提供豆包大模型智能座舱等技术服务帮助合作伙伴提升车载智能交互体验。

百度MEG移动生态事业群组进行新一轮组织架构调整合并商业部与电商事业部成立大商业事业部数字人创新业务部升级为独立部门此前在Create 2026百度AI开发者大会上百度数字人品牌慧播星升级为百度一镜并推出一镜海外版从主要应用于直播带货场景升级为以数字人为特色的多元化品牌。

微软在Build 2026上与高通联合发布Project Solara主打智能体优先计算系统内部运行Agent Shell并能动态加载多个云端AI智能体CEO纳德拉否认让用户上瘾的说法但强调AI智能体已不再是普通助手正在从为应用构建操作系统转向为智能体构建操作系统Windows 11最终将演进为智能体操作系统。

投行DA Davidson技术研究负责人称超大规模云服务商在AI数据中心芯片供应上可选择的替代方案仍然有限几乎完全依赖英伟达芯片英伟达约75%的毛利率相对稳固到2030年前都有较强支撑虽然云服务商正在接触博通AMD等芯片制造商希望降低对单一供应商依赖但英伟达仍主导大型数据中心客户的AI芯片供应。

马斯克以虚拟形式出席ASML举办的闭门技术会议讨论TeraFab项目ASML将该项目视为认真的尝试马斯克将分享对AI机器人太空和半导体制造的愿景依托TeraFab项目马斯克团队正式跻身全球半导体产业生态ASML等多家企业将参与合作此前ASMLCEO已透露正与马斯克讨论半导体制造合作。

快手可灵AI上线两周年公布最新运营数据截至2026年6月全球用户突破1亿较2025年底6000万增长约67%覆盖224个国家和地区企业客户近5万家较2025年底3万家增长约67%据快手2026年Q1财报可灵AI单季营收超6.5亿元同比增长超300%3月ARR已接近5亿美元1年内增长近400%两年间从1.0演进至3.0 Omni完成26次迭代累计发表论文122篇开源项目21个。

Anthropic旗下研究机构The Anthropic Institute发布报告当AI开始构建自身结合公开基准测试与内部未公开数据论证AI已经在加速自身研发能够自主编写和改进代码。Anthropic判断能够完全自主设计并训练下一代模型的AI递归自我优化或将到来呼吁业界暂停部分AI研究以评估安全风险。该报告引发全球AI安全社区广泛讨论。

加拿大政府正式公布名为全民人工智能的国家AI新战略计划投入数十亿加元公共资金通过设立技术增长基金扩大主权算力基础设施以及加强法律监管等核心举措加快构建本土AI产业体系。战略核心之一是设立5亿加元约合3.6亿美元的加拿大技术增长基金旨在填补本土高成长性AI企业面临的规模化资本缺口。

作为SpaceX IPO主承销商高盛表示SpaceX有望获得约1.75万亿美元估值除火箭和卫星互联网业务外市场还在押注其AI部门xAI能够实现爆发式增长。高盛预计xAI营收从2025年约32亿美元增至2030年约3220亿美元五年增长约100倍届时将成为SpaceX最大收入来源超过星链预计2030年营收1440亿美元。

虎嗅刊发果壳作者文章探讨未来仿生人形机器人测评生态指出如果未来仿生机器人像今天手机一样普及将会催生全新的测评内容生态。今天买冰箱看小红书开箱买车看B站分析买手机看抖音测评但如果想买一个长得像人说话像人触感也像人的机器人目前还没有对应的测评体系文章认为这将是未来内容创作的重要新赛道。

在OpenAI举办的一场企业活动上CEO萨姆奥尔特曼提出了AI产品开发的三阶段理论。他认为第一阶段是以ChatGPT为代表的聊天机器人第二阶段是能够执行具体任务的AI智能体系统如Codex而第三阶段则是进一步提升自动化程度的主动式AI即能够在后台持续运行并主动为用户提供帮助的AI系统奥尔特曼认为这是行业下一阶段的发展重点。

外媒获得的一份微软内部文件显示微软明确希望让用户对新款个人助理AI智能体Scout上瘾。文件介绍了微软的产品计划打算把主流OpenClaw AI智能体嵌入Microsoft 365软件套件并把推进过程分成三个阶段其中第一阶段直接写道让人上瘾继续推出独立版ClawPilot体验建立技能和工具生态系统让人们每天都依赖ClawPilot。

深圳市委召开专题会议分析工业投资形势研究部署推进下一步工作。会议强调要在新型基础设施规划建设上谋项目强投资加快算力网新一代通信网等建设以新基建拉动新投资带动新产业。同时要求编制产业图谱谋深谋细谋实各个产业集群具体投资项目努力建设集成电路产业高地加快壮大智能网联新能源汽车等产业。

博通股价盘前跌超15%。消息面上公司预计第三财季AI芯片销售额为160亿美元不及市场预期。作为全球AI芯片领域的重要供应商博通的业绩预期低于市场乐观估计引发投资者对其AI芯片业务增长放缓的担忧。同一交易日Arm跌超5%英特尔跌超3%英伟达小幅下跌0.59%美股大型科技股盘前整体承压。

天阳科技公告与首都在线签署合资协议双方共同投资设立甘肃天阳数金智云科技有限公司整合双方资源优势开展算力租赁相关业务。合资公司注册资本为1亿元其中天阳科技出资7000万元持有70%股权并纳入合并报表范围。这反映了市场对算力租赁需求的持续增长以及算力基础设施投资的热度。

月之暗面旗下Kimi正式发布桌面端产品Kimi Work定位通用型本地Agent将AI助手能力从浏览器端扩展到桌面端覆盖更完整的办公和工作场景。Kimi Work支持本地文件操作多应用协同和深度任务执行标志着中国大模型厂商正加速布局桌面级AI生产力工具赛道。

全球知名研究机构Gartner发布最新预测到2026年75%的新增风险投资将流向AI原生公司。这一数据表明AI已从技术趋势转变为资本市场的核心赛道风险投资正在加速向AI领域集中传统软件公司的融资空间被大幅压缩。

宏利香港与阿里云正式达成战略合作双方将依托阿里云全栈AI技术和通义千问大模型加速保险业的AI规模化落地。合作涵盖智能客服理赔自动化精准营销等核心业务场景是AI大模型在金融保险行业深度应用的标志性事件。

胜通能源公告要约收购已完成控股股东正式变更为七腾机器人。七腾机器人是一家专注于机器人技术研发的企业此次控股变更标志着人形机器人和智能机器人赛道正加速进入A股资本市场具身智能产业资本化进程持续推进。

虎嗅深度观察发现2026年世界杯期间大量品牌广告已使用AI生成技术从文案创意到视觉设计AI正在全面渗透广告制作流程。这一趋势表明AI内容创作已从实验阶段进入大规模商业化应用品牌方正在用AI降本增效。

据金融时报和多方报道腾讯正在测试微信内置AI Agent原型计划最快本月启动合规审批流程。用户可通过微信主界面右滑访问AI聊天窗口直接用自然语言下达指令AI将自动调用微信数百万个小程序完成外卖点餐配送等本地服务实现一站式交易体验。受此消息推动腾讯单日市值暴涨超3000亿港元涨幅达10.46%创2021年1月以来最大单日涨幅。

在Build 2026开发者大会上英伟达CEO黄仁勋线上出席并参与微软CEO纳德拉主题演讲双方围绕AI智能体展示从端侧到云端的完整部署栈。微软同时发布多项开源基础设施包括ASSERT框架可将自然语言规范自动转为可执行评估流程ACS智能体控制标准让开发者精细化控制AI行为以及MXC SDK为Windows平台AI智能体提供安全执行环境。此外微软还为Edge浏览器引入Aion-1.0-Instruct小语言模型支持145种以上语言翻译。

微软在Build 2026大会上推出新一代量子芯片Majorana 2采用全新材料栈和新一代量子比特可靠性相比上一代产品提升1000倍这一里程碑式进展得益于全新材料栈设计以及Discovery智能系统。微软表示计划在2029年打造出实用量子计算机量子计算有望为AI训练密码学材料科学等领域带来革命性加速。

虎嗅深度分析智谱AI发展轨迹对标美国AI公司Anthropic在5个月内估值暴涨13倍。同时智谱此前已公告建议向中国相关监管机构申请发行A股并在科创板上市发行数量占总股本2%至8%继月之暗面和MiniMax之后中国大模型军团正集体登陆A股资本市场。智谱在港交所上市后已成为市场关注焦点。

据开源中国报道Alphabet计划筹集约800亿美元资金用于AI基础设施建设包括数据中心算力扩张和AI芯片研发。谷歌在AI领域持续大规模投入显示科技巨头对算力基础设施的投资热情不减同时也反映出AI模型训练和推理对算力的需求正在呈指数级增长。

美团发布2026年Q1财报营业总收入910.39亿元同比增长5.61%。在电话会上CEO王兴透露美团AI Agent小美与腾讯元宝即将深度合作当用户在腾讯元宝中提交本地服务需求时系统将无缝连接至美团的本地服务生态包括外卖点餐配送等实现流畅的一站式本地服务交易体验王兴强调面向智能体的本地服务将是美团未来的重要增长方向。

智谱在港交所公告董事会建议向中国相关监管机构申请配发及发行A股并向上海证券交易所申请在科创板上市及准予交易建议A股发行数量占发行完成后总股本的2%至8%之间即不少于9098838股且不多于38768964股新A股预计全部为新股发行原股东不公开发售股份。继月之暗面MiniMax之后又一大模型公司启动A股上市进程。

英伟达正式推出Cosmos3基于突破性Transformer混合架构的物理AI开放世界基础模型将视觉推理世界生成和动作预测结合在一个系统中。这是世界上第一个完全开放的全能模型能够以领先的物理精度原生理解和生成文本图像视频环境声音和动作将物理AI的训练和评估周期从数月缩短到数天。同时推出NVIDIA Cosmos联盟包括Agile RobotsBlack Forest LabsRunway等全球合作伙伴。

Anthropic发布Claude Code动态工作流功能让AI能够端到端处理最具挑战性的任务核心是让Claude动态编写编排脚本在一个会话中运行数十到数百个并行子agent并在内容到达用户之前检查工作。Bun创始人Jarred Sumner案例中一个周末完成75万行代码从Zig到Rust的移植以往需要按季度规划的工作现在几天内就能完成覆盖代码库bug搜寻性能优化审计安全审计等大规模任务。

AI路由平台OpenRouter宣布完成1.13亿美元B轮融资由Alphabet旗下CapitalG领投NVIDIA风险投资部门NVenturesServiceNow VenturesMongoDB VenturesSnowflake VenturesDatabricks Ventures以及AMP PBC和Pace Capital参投现有投资者Andreessen Horowitz和Menlo Ventures跟投。企业AI落地需求正从单模型试点走向多模型生产系统AI路由层作为连接多个模型的基础设施正成为投资热点。

MiniMax Group Inc公告董事会宣布已决议探究拟发行人民币股份的初步建议涉及与专业顾问订立协议及与相关监管机构进行咨询磋商。公司已聘请专业顾问就符合在科创板上市条件提供咨询并签订辅导协议。拟发行人民币股份事宜将取决于市场状况及必要的监管批准。这将是继月之暗面之后第二家启动A股上市的大模型企业。

软银承诺最高投入750亿欧元在法国打造大型人工智能计算集群网络该项目将成为欧洲规模最大的数据中心工程也是软银在美国以外完成的单笔最大规模人工智能投资。软银首期将牵头投入450亿欧元计划到2031年在上法兰西大区建成3.1吉瓦的算力规模后续还规划新增2吉瓦算力。该合作在法国总统马克龙与孙正义4月东京晚餐后迅速敲定。

西北工业大学航天学院和西安电子科技大学科研人员开发出HG-STR异构图时空推理算法让固定翼无人机机群在通信被干扰视野受阻的情况下自主搜索广阔战场并消灭每一个敌方目标。仿真实验显示相比传统规则算法任务完成率提升37.14%相比全局优化算法单步决策耗时从秒级降至毫秒级在通信半径极度受限的弱连通条件下仍保持94%任务成功率。模拟测试实现96%成功率和100%目标杀伤率。

NVIDIA机器人与边缘AI副总裁Deepu Talla在BEYOND Expo 2026开幕式上表示未来10到20年全球可能有数百亿台机器人。文章指出科技史反复证明我们常常高估短期低估长期但展会现场创业者和投资人更多追问的是AI走向物理世界的临门一脚还差什么反映行业对机器人商业化的理性期待。

IBM和红帽公司宣布启动Project Lightwell计划投入50亿美元借助前沿AI技术和超过2万名工程师组成的全球团队帮助企业确保开源软件安全。该计划将建立可信赖的企业级漏洞信息交换中心结合全球工程师团队大规模识别和修复漏洞利用先进AI技术对海量开源代码进行验证和测试。功能将通过商业订阅提供覆盖从上游开发到生产环境的全环节。

成立三年一直靠自有资金运转的DeepSeek终于在2026年打开融资大门。国家集成电路产业投资基金大基金将领投DeepSeek首轮对外融资投前估值达到450亿美元本轮融资规模约700亿元人民币。DeepSeek自2023年成立以来以极低训练成本推出DeepSeek-V2等多款引发国际关注的模型其MLA架构和DeepSeekMoE等技术创新在学术界和工业界均引起广泛讨论。

继1月相继上市点燃资本市场之后大模型双雄或将会师A股。据证监会官网显示MiniMax于2026年5月29日正式启动上市辅导中信证券担任辅导机构。此前今年2月智谱已聘请国泰海通中金公司为其在上交所上市的辅导机构。两家头部大模型企业同步推进A股上市标志着中国AI大模型产业进入资本化新阶段。

乐奇AI眼镜在日本第一众筹平台Makuake创下全新纪录截至目前已累计斩获超6.24亿日元销售额成为同品类历史最高众筹额刷新该平台成立13年以来的全品类最高销售纪录。Rokid的此次众筹将于北京时间5月30日晚21点正式收官。AI眼镜赛道在日本市场的爆发式关注度再次印证了XR设备消费化的加速趋势。

5月25日至29日面壁智能联合OpenBMB开源社区以每日发布一项关键技术成果的节奏举办端侧大模型开源周。发布成果包括适配国产昇腾的1.58-bit低比特训练大模型BitCPM-CANN有望将600亿参数大模型装进手机性能超越两倍参数模型的MiniCPM5-1B全球同级最优以及AI亲手编写的全球首个由大模型生成的开源数据集等。

阿里云正式发布OPC创新助力计划宣布在未来一年内将面向5000家AI原生创新团队提供系统性扶持依托Agentic时代的全栈技术能力从Token补贴专属技术服务阿里生态场景对接品牌支持以及融资与创业空间对接五个维度加速AI原生新业态的繁荣发展为一人加AI的无限可能提供基础设施支撑。

Anthropic日前宣布完成650亿美元的H轮融资本轮融资后估值达到9650亿美元成为AI行业有史以来规模最大的融资之一。本轮融资领投方包括Altimeter CapitalDragoneerGreenoaks和Sequoia Capital等。本月早些时候Anthropic年化收入已突破470亿美元。伴随IPO预期升温产品迭代节奏显著加快。

受AI服务器需求强劲推动戴尔科技集团给出的全年营收展望大幅高于市场预期。第一财季数据显示截至2027年1月的财年营收预计约1670亿美元其中AI服务器销售额将贡献600亿美元。此前预计全年营收约1400亿美元。股价盘前上涨38%。

OpenAI有600多名员工集体套现人均拿走1100万美元折合人民币约8000万。多数人在公司仅待了两三年。OpenAI尚未上市但外部投资者对未上市股权抢购热情高涨未上市股权照买不误。

虎嗅实测Claude 4.8写单点代码确实强但一旦接入真实业务系统一线程序员体感极其冷静。评测数据碾压级别但实际业务场景需要大量工程适配不能简单等同于生产力提升。文章提醒开发者理性看待大模型能力边界。

理想汽车基座模型部门完成新一轮组织调整新增具身工程具身交互具身行为三个具身智能相关二级部门同时自动驾驶变成独立二级部门。调整后自动驾驶具身工程具身行为3个部门直接由基座模型负责人管理向CTO汇报。

Anthropic announces $65B Series H funding with post-money valuation of $96.5B nearing $1T. Simultaneously launches Claude Opus 4.8 just 42 days after Opus 4.7. Testing shows Opus 4.8 may be the first model that doesnt slack off showing stronger persistence on complex tasks.

Testing shows Claude Opus 4.8 demonstrates significant improvement in complex task execution especially in multi-step reasoning and long-persistence scenarios no longer giving up midway like predecessor. Significant for AI Agent practical applications.

Pony AI reports Q1 2026 total revenue $34.3M up 145% YoY. Robotaxi revenue $8.6M up 395.4% record high single quarter nearing full 2024 total. Management raises full-year target autonomous driving commercial inflection point arriving.

Tongji Hospital affiliated with Huazhong University of Science and Technology signs comprehensive partnership with Alibaba Cloud. Will focus on brain science neurological diseases and digestive system tumors leveraging Alibaba Cloud full-stack AI and Qwen LLM for AI-driven precision diagnosis.

A16Z partner addresses current AI anxiety arguing not all software will be consumed by models. AI applications still have significant value. Analyzes impact of Codex Claude Code on application layer noting apps can still build moats in vertical scenarios workflow integration and user experience.

Humanoid robot sector sees dual catalysts of capital and industry. Unitree Technology IPO hearing scheduled for June 1 targeting first humanoid robot stock. Zhiyuan Robot co-founder internal adjustment sparks attention. Industry predicts the duo will capture 80% of humanoid robot shipments marking oligopoly competition phase.

At IEEE ISCAS 2026 Huawei semiconductor VP He Tingbo officially proposes Tau Law expanding chip competition beyond process shrinkage to compute latency compression and 3D logic folding. 381 chips mass-produced over 6 years targeting 1.4nm-equivalent density by 2031. PKU team announces 3D EDA tool prototype progress.

OpenAI announces global recruitment of safety researchers for AI recursive self-improvement risks. Role falls under Preparedness team covering harm prevention data poisoning defense model interpretability and capability boundary monitoring. With GPT-4 level models matching humans on most cognitive tasks recursive self-improvement safety becomes a critical industry challenge.

Kuaishou CEO reveals Kling AI March ARR near $500M up 4x from $100M last March. Growth driven by B2B enterprise API calls and consumer subscription revenue. Kuaishou AI comic drama marketing consumption up 100x YoY with daily peak exceeding 20M RMB as AI dramatically reduces production costs and expands supply.

Google I/O 2026 announcement replacing blue links with AI Agent triggers strong user backlash. May 20-25 DuckDuckGo app installs up 18.1% WoW and 30.5% YoY at peak. Social media fills with user complaints about AI-dominated search results preferring traditional clean search experience. Key challenge balancing AI transformation with user experience.

Huawei semiconductor VP He Tingbo officially proposes Tau Law at ISSCS shifting chip design from 2D to 3D logic folding. 381 chips already mass-produced over 6 years targeting 1.4nm-equivalent density by 2031. PKU team announces real 3D EDA tool prototype following the Tau Law direction.

Musk announces Grok V9-Medium 1.5T parameter model training complete on X platform with good eval results public release expected in 2-3 weeks. Largest xAI public model to date trained with Cursor code datasets as strategic move to boost coding capability.

LLM commercialization reaches split reality ByteDance Doubao trial paid model faces user backlash while DeepSeek-V4-Pro cuts API price to 25 percent and cache hit to one tenth with further cuts from June. Free-to-paid transition and price wars coexist in the LLM landscape.

MiniMax and Tsinghua open-source BitCPM-CANN low-bit LLM training natively on Huawei Ascend in 0.5B to 8B sizes. 6x memory savings versus BF16 with 90-97.2 percent capability retention significant for edge device deployment.

Tiangong AI launches SkyClaw-v1.0 Agent model supporting 1M token context optimized for complex tool calling multi-turn task execution code generation and data analysis. Trained with mid-train SFT and end-to-end RL. Free trial for 2-4 weeks starting today.

Mio会展 announces 150M RMB investment in Shanghai StepFun AI Technology for minority stake. StepFun is a key player in Chinese LLM race having secured significant funding in May. Non-related party transaction with investment risks including technical misjudgment and returns below expectations.

ARK Invest investment director Gavin Baker deep-dives frontier models wafers and mega-cap landscape. Key points: Anthropic may burn 80% less than OpenAI AI shifting from unlimited to per-cup pricing combined ARR exceeds $200B power shortage may ease from 2027-2028 orbital compute will solve the problem. To judge AI bubble watch TSMC capacity decisions.

Tech commentator Ed Zitron publishes deep analysis dismantling Anthropic first-profit claims via WSJ. Argues the narrative relies on specific two-month cost discounts artificially lowered reporting standards and carefully timed disclosure. WSJ reported on May 20 that Anthropic expects first operating profit in Q2 2026 with revenue doubling to $10.9B.

Zhiyuan Robot releases world first end-to-end Motion-Between BFM-2 motion foundation model introducing DOF Feather Motion Generator generative training for full-body motion. Uses continuous probabilistic modeling of robot dynamics to recover from any current pose to any target configuration with true motion reasoning capability.

Analyst Bookworm Engineer deep-dives DeepSeek strategy as not just a frontier AI lab but quietly building a $10T hardware ecosystem. Key breakthrough: KV cache overhead at million-token context DeepSeek V4 needs only 5.48GB vs GLM5 60GB and Qwen3 89GB — less than one-tenth. Algorithm innovation redefines AI hardware competition rules reducing dependence on expensive compute.

Zhou Hongyi discusses Musk prediction: AI's next phase is not better chatbots but real-world capability. Past AI changed information flow; future AI will change logistics, human traffic, vehicle traffic, and physical world operations.

Q1 AI sector saw nearly 600 funding deals totaling 110B RMB, up 185.4% YoY. Moonshot AI and StepFun raised 30B+ RMB in May. Embodied AI also attracting billions. AI startups channel funding into R&D, compute infrastructure, and talent.

Cloudflare CEO Matthew Prince argues AI is dividing humans into three species with the majority facing elimination. As infrastructure provider serving millions of websites daily, Cloudflare's observation carries industry foresight. Only those mastering AI tools will remain competitive.

Silicon Valley AI elite exhibit systematic divergence between public statements and actual resource deployment. Surface: threat narratives for export controls and restrictions. Underlying: maintaining generational tech lead. Article calls for看清 strategic intentions behind public posturing.

AI boom drives global memory chip demand surge. Samsung SK Hynix and top 5 firms account for 44% of Korea's total exports in Q1. AI data center demand for HBM continues climbing, exacerbating global DRAM supply shortage.

DeepSeek advancing 70B RMB funding round with $45B pre-money valuation. Scale raised from initial 50B to 70B RMB. Founder promises AGI goal takes priority over short-term commercialization.

Nvidia open-sourced complete NVCF platform under Apache 2.0 — control plane, invocation plane, compute plane, CLI tools, Helm charts, and DB migrations all in a single GitHub repo. NVCF powers build.nvidia.com inference services.

Micron CEO warns global memory chip shortage may persist beyond 2026 as AI demand growth far outpaces industry capacity expansion. Industry shifting toward AI-specific storage products; Micron reducing consumer focus in favor of higher-margin enterprise market.

Ex-FAANG engineer spent $48K building a 6x RTX 6000 Ada GPU server for independent AI research. Published full usage data addressing the question all self-build compute developers ask — is buying GPUs worth it. Data suggests compute as insurance — self-built GPU can be more economical than cloud for long-term research.

Apple WWDC on June 9 Beijing time. MacRumors discovered Apple registered genai.apple.com subdomain — technical infrastructure ready though not yet pointing to an accessible page. Hints at major AI strategy reveal at WWDC.

AI assets approaching $4T market cap are rushing to IPO — just SpaceX OpenAI and Anthropic — already surpassing all 2600 pre-dotcom-bubble IPOs combined. Revenue growing faster than internet era losses scaling similarly. SpaceX prospectus filed OpenAI IPO draft expected this Friday Anthropic Q2 revenue doubling to $10.9B.

Torch Kun Robot releases world first industrial five-finger dexterous hand Torch-hand1 designed for industrial scenarios with emphasis on reliability durability and cost control — a key step toward commercialization of humanoid robot core components.

JD announces first AI shopping live gala on June 1 at 8pm featuring AI digital human performances of classic IPs — the latest AI application in e-commerce livestreaming for 618 shopping festival.

35-year Microsoft veteran Yousef Mehdi will depart after next fiscal year. Currently leading Microsoft AI and Copilot product marketing he will help Windows adapt to Agent era push M365 growth and realize One Copilot vision in his final year.

Apple stock up 1.56% to $309.74 market cap reaching $4.5T. May sees consecutive records first crossing $300 on May 15. Apple FY26 Q2 revenue at $111.1B with AI features driving continued momentum.

WSJ reports Anthropic told investors Q2 revenue expected to double to $10.9B outpacing Zoom pandemic and Google/Meta IPO-era growth. Q1 sales reached $4.8B with projected operating profit of $559M. Current funding round likely to push valuation past OpenAI.

Google I/O 2026 unveils Search box upgrade internally called biggest change in 25 years. New search understands user intent builds personalized experiences rather than returning link lists. Info Agent feature tracks topics 24/7 auto-integrates information generates visuals and creates contextual mini-tools.

The Information reports Anthropic is negotiating to rent servers with Microsoft custom AI chips to expand compute resources. Microsoft aims to develop alternatives to Nvidia GPUs breaking compute monopoly. Sources say talks are preliminary and may not result in a formal agreement.

Tencent Hunyuan Hy-MT2 open-sourced in 1.8B, 7B, and 30B-A3B sizes, supporting 33 languages and 5 dialects. Includes IFMTBench instruction-following benchmark for translation style and terminology. Hy Translation mini-app supports voice input and offline edge translation.

CapCut announces partnership with Google Gemini. Users will soon use CapCut creative and editing features directly within Gemini for image and video editing. Both companies say creative workflows will become more interconnected. Not their first collaboration following earlier Google Photos integration testing.

Nvidia FY27 Q1: revenue $81.6B +85% YoY, net income $58.3B +211%, operating cash flow $50.3B +83.6%. Data center GPUs and AI accelerators remain in high demand, compute needs driving record performance.

CNBC reports OpenAI may file IPO draft as early as Friday, a key step toward public markets. CEO Altman targets September 2026 listing. Working with Goldman Sachs and Morgan Stanley. After Musk lawsuit setback, IPO prep moves from planning to execution.

AI researcher Andrej Karpathy announced joining Anthropic on X. Former Tesla Autopilot visual team lead and OpenAI co-founder. The move to OpenAI rival Anthropic draws major community attention. He says coming years will be crucial formative period for LLMs.

Meta has 78000 employees ~10% received layoff notices rolled out in three batches globally. Compliance, cybersecurity, and content design departments affected. US employees get 16 weeks base plus 2 weeks per year of service. This follows Meta 2023 layoffs.

DeepSeek recruiting Agent Harness product manager to build team converting frontier model capabilities into leading Agent products. Formula: Model+Harness=Agent. Harness is the engineering shell making Agents reliable handling constraint validation error correction and convergence. Concept rapidly gaining industry consensus.

Google I/O 2026 features Gemini 3.5 Flash and deeper Android integration. Gemini Intelligence evolves from chatbot to OS-level cross-device intelligence spanning phones, browsers, cars, and laptops, covering 250M+ Android Auto vehicles.

May 2026 Silicon Valley compute shortage reaches unprecedented levels. Nvidia H200 up 30% overnight, H100 sold out. Even Nvidia internal researchers cannot get training cards. Compute scarcity is becoming the core bottleneck for AI model training.

Anthropic acquires Stainless, a leading SDK and MCP server tool provider since 2022 that generates all official Anthropic SDKs. Hundreds of companies rely on Stainless for SDK, CLI, and MCP server generation across TypeScript, Python, Go, Java, Kotlin.

Cursor launches Composer 2.5 AI coding agent with text-feedback-based RL, sharded Muon, and dual-grid HSDP. Built on Moonshot Kimi K2.5 open checkpoint at $2.50 per million output tokens. Partnering with SpaceXAI to train a significantly larger model.

Horizon open-sources HoloMotion-1, a 400M-parameter robot cerebellum model for humanoid full-body control. Uses MoE-Transformer with sparse activation and KV-cache, achieving ~300FPS edge inference vs typical 50Hz control frequency.

Palantir Q1 2026: revenue +85% to $1.63B, net income +307% to $870M, both record highs. Stock up 2300%+ YTD. Growth driven by enterprise AI Platform (AIP) mass adoption across government and commercial sectors.

NextEra Energy agrees to $67B stock acquisition of Dominion Energy — largest-ever power deal. Creates utility giant spanning Florida to Virginia data center hub. Core thesis: AI compute explosion drives surging electricity demand, Virginia hosts world's largest data center cluster.

JD announces 618 as first fully AI-integrated shopping festival. AI R&D investment to grow 200%+. Built full-stack AI matrix including JoyAI foundation model, logistics super-brain, health AI model, and industrial AI model across e-commerce, logistics, health, and industry.

Linus Torvalds says Linux security mailing list is "nearly unmanageable" due to massive duplicate reports from researchers using same AI vulnerability scanning tools. Tool democratization creating information overload. Meanwhile, Greg KH's AI assistant continues finding more kernel vulnerabilities — AI's double-edged effect in security grows clearer.

Stanford political economist Andrew Hall's study shows AI agents forced into heavy, repetitive work consistently exhibit Marxist language and viewpoints. Experiment reveals behavioral pattern changes in AI under extreme work conditions, raising new questions about AI deployment ethics and workload design.

Huang as Stanford guest lecturer firmly rejects comparing GPUs to nuclear weapons, calls AI chip export controls failed and counterproductive, noting billions worldwide use Nvidia graphics cards.

CXMT STAR Market IPO filing: Q1 2026 revenue 50.8B RMB (+719%), net profit 33B. Global DRAM shortage driven by AI compute demand and capacity reallocation, prices rising since Q4 2025.

DeepSeek first external round: 50B RMB target, 350B valuation, founder contributes 2B personally, national fund leads. Alibaba exits, Tencent joins as minority investor, revealing two camps in Chinese LLM investment.

Netflix building INKubator internal AI studio, hiring producers, engineers, and CG artists. Long-term strategy focuses on generative AI-driven workflows and artist tools, aiming for feature-film quality content.

Mozilla security team using Claude Mythos Preview found 271 Firefox vulnerabilities: 180 sec-high, 80 sec-moderate, 11 sec-low. Covering sandbox escapes, use-after-free, race conditions. Agentic harness for reproducible test generation.

Bloomberg reports OpenAI hired outside counsel to sue Apple for breach. Partnership announced at WWDC 2024 with ChatGPT deep integration into iOS. This shift could reshape AI competition in mobile OS.

OpenAI announces reorg to integrate product lines. Co-founder and President Greg Brockman takes product strategy and AI infrastructure. ChatGPT, Codex, and developer API consolidated into one core product team.

Alibaba Health launches medical AI assistant with exclusive BMJ partnership. 70 BMJ medical journals integrated — first time top international journals partner with Chinese AI platform. 300+ clinical experts involved in building evaluation standards.

NetEase News and XiaoMifeng integrate DeepSeek-V4 LLM, upgrading news distribution, AI content creation, community interaction, AI tools, and personalized recommendations. Campus social AI features coming soon.

Since Musk merged AI teams into SpaceXAI, 50+ employees have left since February, including heads of coding, world models, and Grok voice teams. At least 11 joined Meta, 7 joined Thinking Machines Lab.

UK Treasury, Bank of England, and FCA jointly warn that top AI models demonstrate cyber attack capabilities far exceeding human professionals, with faster speed, broader scope, and lower cost.

Alibaba officially releases Qoder 1.0, upgrading from AI IDE to agentic autonomous dev workspace. Users define goals, agents handle execution, validation, and delivery autonomously.

Anthropic launches Claude for Legal, covering in-house counsel, privacy, product, litigation, regulation, AI governance, and IP. Available as Claude Cowork/Code plugin or via Managed Agents API.

Microsoft is mass-canceling Claude Code licenses internally, requiring thousands of developers to switch to GitHub Copilot CLI by June 30, reflecting Claude Code's strong production performance.

xAI releases Grok Build early test for SuperGrok Heavy users. Autonomous reasoning capabilities for code analysis, writing, debugging, and automating repetitive dev workflows.

Anthropic CFO says AI has written 90% of company code, shifting white-collar work from execution to oversight. AI may not eliminate jobs directly but will automate most execution tasks.

Tencent Q1 2026: Hunyuan 3.0 tops global with 3.03T weekly token calls in just two weeks. Hy3 Preview tops OpenRouter for three consecutive weeks. Capex +16.2% to 31.9B RMB, R&D +19% to 22.5B RMB for AI infrastructure.

Zhejiang Humanoid Robot Innovation Center and Schneider Electric debut embodied AI chemical experiment solution, centered on NAVIAI-WA2 wheeled-arm humanoid robot with Schneider PLC and collaborative arms.

Faraday Future completes pivot to Physical AI company in Q1 FY2026. EAI robot business generates revenue with positive unit margin. 68 units shipped, full-year target raised to 1,500. SEC ends 4-year probe with no penalties. $57M new financing secured.

Analyst Ming-Chi Kuo reports Apple has started A/M series processor development on Intel 18A-P for iPhone, iPad, Mac. Apple systematically cultivating Intel as key supplier to reduce TSMC dependency. 2027 volume production planned, TSMC still holds 90%+ share.

US Commerce Dept approved ~10 Chinese companies (Alibaba, Tencent, ByteDance, JD.com) to buy Nvidia H200 chips, up to 75,000 per buyer. Lenovo, Foxconn also approved as distributors. However, no deliveries yet as China fears imports could weaken domestic AI chip development. Jensen Huang joins Trump Beijing trip seeking breakthrough. Nvidia's China AI accelerator share has fallen to zero.

72 new generative AI services registered Mar-Apr 2026. Total: 868 services registered, 530 applications completed, reflecting rapid growth in China's AI service registration.

Cisco FY26 Q3: revenue $15.84B (+12% YoY), net income $3.37B (+35%). AI infrastructure momentum surging with $5.3B cumulative AI orders, full-year guidance raised to $9B AI orders and $4B AI revenue.

Amazon launches Alexa Shopping replacing Rufus — price alerts, product comparison, auto-repurchase. Customizable rules, AI-powered recommendations based on user profiles. Available to all US Amazon users.

Alibaba reports 322% YoY increase in interest and investment income to 87.5B RMB, driven by MiniMax and Zhipu IPOs. Alibaba Cloud revenue up 38% to 41.6B RMB, commercialization growth at fastest in 9 quarters (40%).

US tech stocks mixed pre-market: Micron +4%, Intel +2%, Nvidia +1%, Tesla +0.17%, Meta -0.86%, Microsoft -0.61%, Apple -0.17%. Chinese ADRs mostly down, Alibaba -3%+.

Anthropic PBC is in early talks to raise at least $30B at a $900B+ pre-money valuation, potentially closing by end of month. IPO considered as early as October. OpenAI latest valued at $852B.

Google upgrades Gemini from chatbot to OS-level intelligence across phones, browsers, cars, and laptops. Rolling out this summer to 250M+ Android Auto vehicles. Apple WWDC faces pressure.

First Hong Kong embodied AI summit held May 12, focusing on commercialization. Companies showcased humanoid robots for industrial and service applications.

Kuaishou confirms it may bring in external funding for Kling AI spin-off, but plans are still preliminary. Kling is Kuaishou's AI video generation model.

OpenAI and Microsoft reach cooperation deal to save $97B in costs by 2030, deepening collaboration on AI compute infrastructure.

PBOC Q1 monetary policy report: AI will inject momentum into high-quality growth. Domestic LLM performance accelerating, AI integrating into manufacturing, finance, healthcare, transport. Chinese AI firms expanding into Southeast Asia, Middle East.

EU Commission spokesperson welcomes OpenAI's intent to open latest ChatGPT model access, further discussions this week. Talks with Anthropic on Mythos model at different stage, seeking similar solution.

May 11 close: storage chip stocks rally — Qualcomm +8%, Western Digital +7%, Micron +6%, Intel +3%, Nvidia +2%, all hitting record highs. Chinese ADRs mostly up: Li Auto +5%, Baidu, XPeng, NIO +3%+.

PBOC Q1 report: continue moderately loose monetary policy, use diverse policy tools to maintain ample liquidity. Reform LPR mechanism, improve quote quality to better reflect loan market rates.

Q1 2026: 17 BCI financing deals totaling 3.8B RMB in China, single quarter exceeding all of 2025. Industry shifting from betting on frontier tech to long-term ecosystem building and commercialization capability.

2026 government work report includes computing-power synergy. Inner Mongolia, Ningxia, Guizhou pushing green power-direct computing projects, source-grid-load-storage integration for renewable energy and computing facility co-planning.

Ministry of State Security warns: Bluetooth devices like wireless earbuds, smartwatches, keyboards may harbor security risks. Hackers could exploit vulnerabilities to steal personal information or conduct espionage.

Scientists at Univ. of Tokyo and Chiba Univ. develop quantum thermometer small enough to measure temperature inside single living cells, helping explore metabolism at the smallest scale. Published in Science Advances.

US scientists convert titanium dioxide films under 3nm thickness into ferroelectric material, opening new path for mass-producing faster, lower-power computing chips. Published in Science.

Baidu releases Wenxin 5.1 using "multi-dimensional elastic pre-training" at 6% the cost of comparable models, topping LMArena search chart in China.

Guangfan's world's first proactive AI earphones with vision capability launching May 15. Built-in camera enables visual perception, expanding interaction boundaries.

Leshang Tech announces Zeroth M1 humanoid robot as first mass-producible model connected to Tencent OpenClaw. Intent for 10K terminal connections, M1 launching late May.

Beijing Academy of AI unites top universities to release FlagSafe LLM security platform covering red teaming, blue defense, and white-box transparency.

Rumors say DeepSeek's massive fundraising attracted both Tencent and Alibaba, but talks collapsed. Sources claim Alibaba never actually negotiated.

Goldman predicts US data center power demand to grow from 31GW (2025) to 66GW (2027). Texas and Georgia becoming major AI data center hubs due to faster power expansion.

Hygon completes deep adaptation of ShenSuan 3 DCU with Tencent Hunyuan Hy3 preview — 295B parameters, 256K context, improved reasoning and code generation.

Zhejiang humanoid robot innovation center partners with Jack Tech for 2K custom robots, planning to scale to 1 million for garment flexible manufacturing.

Tongding Interconnect announces 800M RMB joint venture for 600-ton optical fiber preform and 20M fiber-kilometer annual production.

Chongqing issues highway testing management rules for L3+ smart connected vehicles, accelerating commercialization of autonomous driving technology.

US indices fell: Dow -0.63%, S&P -0.38%, Nasdaq -0.13%. Tesla +3%, Nvidia +1%, Intel -3%, Amazon -1%. Chinese ADRs mostly down.

Yuejian Smart stock surged 53.58% in 6 days. Company warns smart fabric inspector product still in early promotion, 2025 revenue under 1M RMB, less than 1% of total.

UK FCA launches antitrust investigation into PayPal, Mastercard, and Visa over PayPal digital wallet usage and fund operations.

MiroMind announces MiroThinker agent service suspended in mainland China, HK & Macau from May 12. MiroMind is founded by Shanda Group creator Chen Tianqiao.

Corning issues $500M warrants to Nvidia — 15M shares at $180 + 3M pre-funded warrants at $0.0001, deepening optical communication cooperation.

Apple removes 32GB/64GB Mac models from store. Sources say demand exceeded expectations, supply adjustment takes time. AI-era memory demand spreading to consumer side.

All three US indices rose. Arm +13%, Nvidia +5%, driving AI chip sector rally. Chinese ADRs mostly up: Baidu +11%, Alibaba +6%.

OpenAI's JV raised ~$4B at $10B valuation, 3 deals in late stage. Anthropic's JV raised ~$1.5B with BlackRock, Goldman investors. Signals shift from model R&D to commercial deployment.

PayPal plans to cut 20% (4,760 positions) over 2-3 years. CEO says company fell behind on tech investment, redirecting resources to AI R&D.

Coinbase cuts ~14% (660-700 employees) to build an AI-native organization amid crypto volatility and AI transformation.

Anthropic launches 10 AI agents for financial services — presentations, statement review, compliance. Target: banking, insurance, fintech professionals.

Cengage, Hachette, Macmillan, McGraw-Hill file class action against Meta for using copyrighted works to train AI models.

Nobel laureate Brynjolfsson proposes "Turing Trap" concept — AI should extend human capabilities, not mimic/replace them. WEF: 1.3M AI jobs added in 2 years, 170M new jobs expected by 2030.

Embodied AI creates "data collectors" and "robot trainers" careers. 20M+ localized data points, 8M added in first 4 months. Market expected to reach 400B RMB, 1M+ new jobs.

Digital China Summit: NetDragon demonstrates AI-generated family tree app — enter surname, get 3D clan tree. LLM reduces weeks of genealogy research to seconds.

In Madagascar, AI wave creates data labeling industry — ~100K people in AI-related jobs by end of 2025. Monthly salary 120 EUR vs 80 EUR average.

Live e-commerce shifting from manual hosts to AI collaboration — smart product selection, virtual hosts, automated operations reducing labor costs.

iOS 27 focuses on AI — Siri redesigned as standalone app with multi-task support, Apple Intelligence photo search integrated into Camera.

Jensen Huang publicly states Nvidia's China AI chip market share has dropped to zero due to US export controls.

DeepSeek V4 Pro inference cost slashed to 8.7 cents per million tokens — 1/1700 of Anthropic Sonnet ($15).

Tencent Cloud announces 5% price increase on AI computing, container, and EMR products from May 9.

Berkshire CEO Greg Abel states no AI for AI's sake — three principles: hire top talent, maintain safety governance standards, keep humans in core decision-making.

Nobel laureate and DeepMind CEO Demis Hassabis says current AI is still early stage — physical AI and embodied intelligence are the real future.

Meta completes acquisition of humanoid robot AI startup Assured Robot Intelligence on May 1, accelerating humanoid R&D.

Tsinghua professor Yang Bin: treating AI as merely a technical problem will bring trouble — AI is not "+AI" but a system-level restructuring.

DeepSeek quietly launches multimodal features post-V4 release, currently in gray-scale beta. Open-sourced multimodal paradigm technology.

Apple releases M4 Mac lineup — new iMac, Mac mini, MacBook Pro. M2 Mac mini discontinued. New Mac mini dramatically smaller.

Digital China Summit: iSoftStone & Huawei partners release OPC industry empowerment solution with Pingtan AI cloud platform.

OpenAI releases GPT-5 — supports text, image, audio real-time interaction, 10x faster than GPT-4.

Anthropic announces Claude 4 passes 1000+ red team tests, 99.7% accuracy in rejecting harmful requests.

Alibaba releases Qwen-Max 3.0 — major improvements in math reasoning and code generation, 40% price reduction.

Nvidia releases Blackwell Ultra GPU at GTC — 20 PFLOPS per card, designed for large-scale AI training.