国内新闻
中国半数token消耗靠AI短剧？豆包高管回应
近期，一项“短剧Token消耗占比55%”的数据被多家媒体研究机构转发，并形容为“国内一半算力用来造爽剧”，引起热议。该数据原始出处为中文在线在6月3日举行的北京线下路演活动及线上电话会议，“在国内所有Token消耗份额中，AI短剧与视频生成高居首位，占比高达55%；其次是电商与营销，占比24%；而此前因Vibe Coding流行而被寄予厚望的软件开发领域，在Claude等模型编程能力持续推进下也仅占15%。”
对此，豆包公关负责人刘星公开表示该数据“有些不靠谱”，称字节总算力中仅约10%用于视频生成模型的训练和推理。DataEye研究院负责人刘尊指出，AI短剧消耗的多是通用模型、中低端卡的推理算力，与科研用的高端训练算力是两条相对分割的供给线，不能简单理解成“短剧用了Token，科研就没得用”。
DeepSeek开源推理框架DeepInfer，披露单集群日吞吐155万亿Token
10月20日，DeepSeek宣布开源推理框架DeepInfer，这是支撑其大模型高吞吐、低延迟线上推理服务的核心技术。同步发布的技术报告披露了关键运营数据：基于该框架，DeepSeek V4.1 Flash已实现大规模推理服务，单一集群服务吞吐达155万亿tokens/天，整体收入为每卡每小时38美元（按谷价）。此前10月19日，DeepSeek还公布了API网页搜索与抓取费用：每1000次成功请求5元。
DeepSeek发布Flash新版本
10月16日，DeepSeek V4.1 Flash更新至V4.1-Flash-1017版本，官方网页端、APP与API同步变更。新模型在Terminal-Bench 4.0上从31.2提升至42.3，DeepSWE v1.1从74.2提升至77.0，ProgramBench从20.3提升至29.1，同时缩短了任务运行时间与输出长度。新版本重点提升了游戏开发、3D建模等从零构建完整项目的能力，并针对DeepSeek Harness、腾讯CodeBuddy等平台框架进行了优化。
Today AI累计完成超2亿美元融资
10月22日，个人AI助理公司Today AI完成超2亿美元融资，投资方包括IDG资本、腾讯、五源资本、蚂蚁集团、红杉中国、传音控股、同歌创投、上海交大菡源资产。最近一个月，个人AI助理在海外成为继AI Coding之后又一个备受追捧的应用方向。据悉，Today AI是由前Teambition创始人齐俊元推出的个人级AI Agent操作系统。
MiniMax港股破发
10月22日，MiniMax盘中跌5.6%至162.2港元，跌穿165港元的发行价，为今年1月上市以来首次破发。该股3月曾创下1330港元高点，以最高位计算市值已蒸发近88%。
MiniMax发布新的Flash模型
10月18日，MiniMax正式更新M3.1 Flash。官方称该模型在Agentic任务、编程、Computer Use及网络安全能力上实现代际提升，推理速度提升近3倍，使用成本降低70%。在DeepSWE v1.1上取得73.7%，OSWorld V2和Agents‘Last Exam两项电脑操作基准均位列国产模型第一，CyberGym网络安全基准取得82%。定价为输入每百万Token 1元、输出每百万Token 4元，缓存读取0.1元。
豆包手机助手限制金融类App自动操作
10月21日，豆包手机助手官网更新拒绝操作应用清单，主动对金融类应用限制相关操作能力，范围包括银行、信托、支付、互联网金融、信贷等领域。此前公示的拒绝操作名单已包括工商银行、招商银行、重庆农商行等多家金融机构。已明确声明拒绝的应用，Agent将不会对其进行操作。
DeepSeek考虑入局具身智能，梁文锋正密集与相关创业者交流
知情人士透露，目前梁文锋正在考虑DeepSeek是否要进一步进入具身智能领域，方案可能包括亲自组建团队下场研发，也可能通过投资相关企业切入。此前，梁文锋已投资宇树科技，另据彭博社报道，DeepSeek计划领投腾讯前AI科学家胡瀚创立的机器人公司Winverse AI。不过，该公司不能算严格意义上的具身智能公司，更偏底层infra，做模型加速相关。据另一位知情人士表述，DeepSeek考虑具身智能并不令人惊讶，智谱与Kimi等AI公司均有具身智能相关业务储备。具身智能，已经成为全球AI公司争夺的新战场。（凤凰网科技）
消息称DeepSeek拟在最新一轮融资中筹得约1000亿元
有市场消息称，国内大模型公司DeepSeek拟在最新一轮融资中筹得约1000亿元人民币(约150亿美元)，腾讯和宁德时代为本轮出资最多的投资方，交易预计在未来数日内正式完成。截至目前，DeepSeek及腾讯等相关投资方均未公开回应。消息人士透露，本轮募资目标最初设定为500亿元，此后经多轮上调至约1000亿元，目前尚未官宣，最终金额仍未完全确定。这笔资金将继续用于模型训练、算力建设和人才招募，以支撑DeepSeek参与全球竞争。也有部分消息人士判断，本轮融资将为公司筹备中的科创板IPO铺平道路。
宇树科技发布人形机器人“H2-X功夫传奇”，最长续航约8小时
10月21日，宇树科技宣布发布人形机器人“H2-X功夫传奇”，最长续航时间约8小时。宇树科技创始人王兴兴同期预测，物理AI将于明年迎来拐点。此外，宇树科技最畅销的人形机器人累计生产下线已达1-2万台。
长鑫4F²DRAM架构引发技术路线分歧，美光明确不跟
长鑫存储本月推出4F²DRAM架构，据称单位面积晶体管数量可比主流6F²架构更多。当地时间10月21日，美光科技总裁兼首席技术与产品官Scott DeBoer回应称，美光将沿6F²路径并开发3D DRAM，不做4F²，理由是复杂度与成本上升，并表示美光研究了4F²20年，认为该方向“没有可行微缩路径”。受此消息及黄仁勋称对合作持开放态度影响，长鑫科技10月23日涨超7%。
Kimi K3.1 Pro发布
10月24日凌晨，月之暗面推出Kimi K3.1 Pro。Kimi称，这是其迄今能力最强的3万亿级参数模型，支持100万上下文，在代码、3D创作、金融研究和其他长程Agent能力上的提升尤为显著。Kimi K3.1 Pro是Kimi K3的后训练版本。通过增强后训练算力和技术优化，Kimi K3.1 Pro模型在所有领域都获得了大规模RL（强化学习）带来的全面提升。
寒武纪前三季度营收净利润双双翻倍
10月23日晚，芯片龙头寒武纪披露2026年三季报。前三季度，公司实现营业收入95.26亿元，同比增长106.76%；归母净利润37.15亿元，同比增长131.49%；扣非归母净利润35.23亿元，同比增长148.30%。
第三季度，公司实现营业收入35.31亿元，同比增长104.47%；归母净利润14.04亿元，同比增长147.74%；扣非归母净利润13.58亿元，同比增长168.13%。当季经营现金流净额为-8367.79万元。前三季度，公司非经常性损益净额为1.91亿元，其中第三季度为4602.54万元。
国家数据集团成立
10月20日，国家数据集团有限公司在河北雄安新区揭牌成立，成为第100家国务院国资委监管央企。集团由国务院、雄安新区管委会、中国电信、中国联通、中国移动、国家电网、南方电网共同出资组建，以数据资源开发利用为主责主业，定位为数据要素服务国家队、数据基础设施互联互通引领者。
金蝶灵基全线升级，付费用户已达3706家
2026全球创见者大会上，金蝶企业智能体操作系统灵基（Lingee）发布了新一轮核心能力更新，包括灵基Manage、企业上下文和ERP连接器，帮助每家企业加速成为AI原生组织。截至目前，灵基的付费用户已达3706家。并联合大模型厂商、高校发布国内首个面向企业管理场景的Agent评测基准《Lingee Bench》。
同时，大会展示了金蝶与阶跃星辰在企业智能场景上的共创成果：阶跃模型已落地于金蝶灵基企业数据分析和AI速记两大场景。其中，语言模型在自然语言查询经营数据、根因分析等任务中准确度测评居前；语音模型融入了金蝶研产供销人财税专业语料，在企业管理场景语音识别测评中居于前列。
国际资讯
美国政府下令：禁采DeepSeek
10月19日，美国国土安全部（DHS）公开一份针对DeepSeek的正式禁令，要求美国联邦民事行政机构停止采购DeepSeek相关产品和服务，并从现有信息技术系统中移除相关产品。此次禁令不仅针对DeepSeek官方产品，还要求清除采用DeepSeek技术实现方案开发的第三方产品和服务。涉及的企业包括杭州深度求索人工智能基础技术研究有限公司（DeepSeek）及其关联实体。
根据美国《联邦公报》公开的文件，这项命令由美国国土安全部长马克韦恩·马林（Markwayne Mullin）于2026年7月24日签发，10月19日对外公开，10月20日正式刊登。文件编号为2026-21447，案卷编号为DHS-2026-0628，正式名称为《关于从联邦民事行政机构信息技术系统及采购活动中排除和移除DeepSeek产品的命令通知》。
Meta删除允许AI抓取新闻数据的条款
10月18日，Meta宣布已正式删除其最新商业服务条款中关于“使用企业域名内容训练AI模型”的争议性表述。该条款原允许Meta抓取拥有其企业账户的网站内容用于产品开发与AI训练，引发法律界质疑。Meta回应称，此次更新仅为整合既有广告条款，并未扩大网络爬取范围；现行条款已明确限定数据收集仅服务于广告及商业产品，且重申尊重robots.txt协议。此举系在听取多方反馈后作出的澄清与修正。
OpenAI模型将登陆谷歌云，包括GPT‑6 Astra等前沿模型、Codex代码Agent
10月20日，OpenAI宣布与谷歌云合作，将OpenAI前沿模型和CodexAgent工作流引入Google Cloud的Model Garden。Google Cloud客户可通过Model Garden购买OpenAI模型，并在Google Cloud Marketplace完成交易，已承诺的GCP消费额度可计入采购费用。OpenAI模型也将很快在Google Antigravity和Gemini Agent中推出，计划在未来几周内正式上线。微软仍是OpenAI的主要云合作伙伴。
Anthropic推出个人AI助手Claude Duet
10月20日，Anthropic推出个人AI代理应用Claude Duet研究预览版，面向美国Pro和Max订阅用户开放等候名单。Duet基于Claude Opus 5.5，以独立应用形式登陆iOS、Android及claude.ai。它会在每天早晨推送重要事项简报，持续监控已连接应用，仅在需要用户注意时发送消息，并可在用户离开时继续完成任务。Duet内置安全分类器和系统提示，拒绝分享信用卡号、政府身份证明等最高敏感层级信息。
特斯拉第三季度净利润大跌43%
10月21日，特斯拉发布财报，第三季度净利润同比下降43%，营收280.5亿美元同比微降0.2%，营业利润从16.2亿美元降至7.18亿美元，降幅56%。特斯拉表示，营业利润下降主要受AI及其他研发项目相关运营费用增加、股权激励费用上升影响。当季全球交付486532辆，同比下降约2%。
英伟达GTC柏林发布DSX POD，数据中心部署时间从6个月缩至14天
北京时间10月21日，英伟达CEO黄仁勋在GTC柏林演讲中发布DSX POD AI工厂参考设计。该产品在模块化集装箱内建造，交付前已完成集成与验证，将于2027年初全球交付。黄仁勋称，相比传统部署方式，从开始部署到产生Token的时间可从6个月缩至最少14天。同时宣布与供应链伙伴扩大产能、升级Isaac机器人平台、推出面向AIAgent的OpenShell参考架构。
AI板块近期遭遇明显抛售，Anthropic会推迟“超级IPO”吗？
AI独角兽Anthropic正站在上市的历史关口，但市场逆风正在加剧不确定性。
据路透社报道，Claude聊天机器人的开发商Anthropic计划于11月中旬完成IPO，目标估值高达2万亿美元，有望超越SpaceX今年早些时候1.77万亿美元的上市规模，成为史上最大科技IPO之一。然而，美国证券交易委员会（SEC）要求公司在启动路演前至少15天公开提交招股说明书，这意味着Anthropic留给自己的窗口期正在迅速收窄。
与此同时，AI板块近期遭遇明显抛售——CoreWeave股价两周内暴跌约30%，美光过去一个月下跌17%——市场对AI支出可持续性的疑虑持续升温。多位市场人士表示，Anthropic可能面临推迟上市的压力，但也有分析师指出，该公司对资本的迫切需求或将推动其坚持原定计划。预测市场目前显示，Anthropic年内上市的概率约为80%。
Meta旗下Muse下载量突破1000万
10月21日，Sensor Tower数据显示，Meta旗下AI助手Muse自9月推出以来累计下载量达1090万次，其中1020万次来自美国，成为上市后40天内美国下载量达到1000万次的移动应用之一。此前仅11款应用曾达到这一里程碑。
Anthropic推出个人AgentClaude Duet预览版
10月22日前后，Anthropic推出个人AgentClaude Duet预览版。此前Meta Muse、xAI GrokBot、OpenAI Dots等已进入个人Agent赛道，国内阿里千问Go、腾讯元宝助手、字节“小豆”等也在推进中。
苹果智能中国测试版将于10月底上线
据报道，库克称苹果智能中国测试版将于10月底上线，接入千问，不含Siri AI。
融资
新石器完成4亿美元E轮融资，超募一倍
10月20日，无人配送企业新石器（Neolix）完成4亿美元E轮融资，由博裕投资与某国家级基金联合领投，阿里AEF NextGen香港基金、淡马锡Vertex Growth、广汽资本、正大机器人等参投，老股东磊石资本、软银亚洲、亦庄国投、高榕等跟投。创始人余恩源透露，原计划仅融2亿美元，实际超募一倍。国家人工智能产业投资基金由此首次将无人配送企业纳入投资组合。
东山精密寻求港股上市，募资至多107亿港元
10月20日，据港交所公告，苏州东山精密制造拟在香港发行约1.138亿股，最高发行价每股93.80港元，预计募资不超过107亿港元（14亿美元），10月28日挂牌。基石投资者包括Aurora SF、永联投资、摩根资产管理（亚太）和Alibaba Investment。
FORT Robotics获约2000万美元新增融资
实体AI安全平台FORT Robotics获现有投资者约2000万美元新增承诺，Tiger Global领投约1000万美元，Mark Cuban Companies参与。公司正通过与Newbury Street II合并寻求纳斯达克上市，代码预计为“FROB”。
亿铸科技完成超10亿元新一轮融资
存算一体AI芯片公司亿铸科技宣布完成超10亿元新一轮融资，基于3D DRAM的全数字通用存算一体AI芯片已完成流片。本轮融资将用于研发投入、产品迭代及产业化推进。
伊塔动力完成超亿元种子及种子+轮融资
专注于具身智能操作平台与动力系统研发的初创企业“伊塔动力”完成超亿元种子和种子+轮融资。
趋势
TrendForce：2026年服务器CPU出货增幅预估31%，Agentic AI推高需求
10月22日，TrendForce集邦咨询指出，随着AI服务厂商积极推出AI Agent相关应用，服务器CPU需求自2025年下半年起稳步成长，预估2026年出货年增幅达31%。2027年由于AI推理需求持续成长、Agentic AI服务渗透率提升，出货量有望再年增逾40%，x86阵营主导的市场格局可能出现变化。
中信证券：AI科技企业CDS利差走阔，尚未演变为行业性信用风险
中信证券研报称，AI资本开支规模持续增加，相关主体融资成本日渐高企，推动投资者对AI科技企业信用风险的关注和对冲需求同步上升，AI科技企业CDS市场呈现交易量与利差双升态势。但目前尚未演变为行业性信用风险的全面抬升，主要AI相关企业CDS利差呈现梯度分化。
商务部等9部门：加速脑机接口、具身智能等前沿科技成果向市场转化
10月19日，商务部等9部门发布关于加快推进科技兴商的意见，提出加快新型显示、智能终端等领域关键技术研发，加速脑机接口、具身智能等前沿科技成果向市场转化，支持智能仓储、数智供应链等领域商贸企业申报专精特新企业。
国家统计局：“人工智能+”行动有力带动新兴领域投资和出口
10月19日，国家统计局新闻发言人王冠华表示，前三季度“人工智能+”行动有力带动了新兴领域投资和出口。电子电路制造业、集成电路制造业、信息服务业投资分别增长58.7%、14.2%和20.2%；电子元件、电线电缆等产品出口分别增长70%和25.3%。
工信部：600亿元国家AI基金投出一批项目，人工智能企业超6600家
10月19日，工信部科技司司长魏巍在国新办发布会上介绍，600亿元国家人工智能产业投资基金已投出一批优质项目。2025年中国AI核心产业规模超1.2万亿元，规模以上制造业企业AI技术应用普及率超30%，截至今年6月底AI企业数量超6600家。已披露的三笔直投项目均落在头部大模型公司：参与DeepSeek首轮融资出资约9.8亿元，投后估值3500亿元；参与月之暗面F轮融资，持股约1.14%。
商务部等9部门发文推进科技兴商
10月19日，商务部等9部门发布《关于加快推进科技兴商的意见》，提出18条具体举措，涵盖科技赋能商品消费提质、科技推动服务消费扩容、科技畅通商贸流通链条、科技塑造未来商业场景四方面。意见提出深入实施“人工智能+制造”专项行动，丰富智能手机、电脑、家居、穿戴设备、机器人等智能商品供给；推动脑机接口、具身智能等智能终端攻关迭代；开展“人工智能+文化和旅游”应用试点等。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
National news
Half of China's token consumes Ai shorts? Doubao executive response
Recently, the data on a short-time token consumption ratio of 55 per cent were relayed by media research institutions and described as “half of the country's calculus is used to make a cool play”, giving rise to much debate. The original source of this data was the Chinese-language online Beijing offline show and online teleconference held on 3 June. “Ai short play and video production accounted for up to 55 per cent of all Token consumption in the country, followed by electricians and marketing, 24 per cent, and only 15 per cent of the software development that had previously been given high hopes because of the Vibe Coding epidemic, and only 15 per cent of model programming capacity in Claude”.
In response, Liu Xing, the head of Doubao public relations, publicly stated that the data were “some of the uneconomical” and stated that only about 10 per cent of byte's total power was used for training and reasoning in video generation models. The head of the DataEye Institute, Liu Zhang, pointed out that the Ai short play consumes much of the reasoning of a generic model, medium- and low-end card, and that the high-level training calculus for scientific research is a two-part supply line that cannot be simply understood as “token for short plays and no use for scientific research”.
DeepSeek open source reasoning framework DeepInfer, disclosure of 155 trillion Tokens per day in single-clusters
On 20 October, DeepSeek announced the Open Source Infer framework, which is the core technology that underpins its LLM high-input, low-delayed on-line reasoning services. The technical report publish synchronized disclosed key operating data: based on this framework, DeepSeek V4.1 Flash has achieved large-scale reasoning services, with a single cluster service of 155 trillion tokens/day, with an overall income of $38 per card per hour (at valley prices). On 19 October, DeepSeek also announced the API search and capture costs: $5 per 1,000 successful requests.
DeepSeekpublishFlash new version
On 16 October, DeepSeek V4.1 Flash was updated to V4.1-Flash-1017, with the official web page, APP and API synchronized. The new model was upgraded from 31.2 to 42.3 on Terminal-Bench 4.0, DeepSWE v1.1 from 74.2 to 77.0, and ProgrammeBench from 20.3 to 29.1, while reducing the operation time and output length of the mission. The new version focuses on enhancing the capacity of game development and 3D modelling to build complete projects from zero and optimizes platform frameworks such as DeepSeek Harness and TencentCodeBuddy.
Today AI cumulatively completed over $200 million in financing
On 22 October, Today AI, a personal assistant company, completed over $200 million in financing, with investments from IDG capital, Tencent, five-source capital, ant groups, Redwood China, audio-transmitting shares, cosmopolitan creation and Shanghai for large source assets. In the last month, PAI assistants have become a more popular application after AI Coding. Today AI is known to be a personal-level AI Agent operating system introduced by Zhi Joon, a former founder of Teambition.
Mini Max's stock has collapsed.
On 22 October, Mini Max dropped between HK$ 5.6 per cent and HK$ 162.2, breaking the distribution price of HK$ 165 for the first time since it was launched in January this year. The unit reached a high point of HK$ 1330 in March, evaporating nearly 88 per cent of its market value at the highest level.
Mini Maxpublish New Flash Model
On 18 October, Mini Max officially updated M3.1 Flash. Officially, the model has increased intergenerationally in terms of its mission, programming, Computer Use and network security capabilities, increasing almost threefold in reasoning and reducing the cost of use by 70 per cent. 73.7 per cent on DeepSWE v1.1, OSWorld V2 and Ages' Last Exam were ranked first in the national production model and 82 per cent in CyberGym network security. The price is one dollar per million Token, four dollars per million Token and one dollar per million cache.
Doubao Cellular Assistant restrains App from financial self-operation
On 21 October, the network of Doubao mobile phone assistants updated the list of denials of application, proactively restricting operational capabilities for financial applications, including in the areas of banking, trust, payment, Internet finance and credit. The previously published list of refusals has included a number of financial institutions, such as the Business Bank, the Business Bank and the Chongqing Chamber of Commerce. Applications that have been expressly declared rejected will not be operated by Agent.
DeepSeek is thinking of being smart, and Liang Wenbing is in intensive communication with relevant entrepreneurs
According to the source, Leung Wenfeng is currently considering whether DeepSeek is going further into the area of smartness, which may involve the formation of a team in person to develop it, or through investment-related enterprises. Prior to this, Liang Wensai had invested in Wootree Technology and, according to Bloomberg, DeepSeek was planning to take over Winverse AI, a robotic company founded by Tencent ex-Ai scientists. However, the company cannot qualify as a strictly smart company, more than a bottom-infra, for model acceleration. According to another source, it is not surprising that DeepSeek considers being smart and that Zhipu and other AI companies, such as Kimi, have smart-related business reserves. Smart as a new battleground for global AI. (phoenix technology)
News says DeepSeek is going to raise about $100 billion in the latest round of financing.
According to market information, the domestic LLM company DeepSeek intends to raise approximately RMB 100 billion (about $15 billion) in the latest round of financing, and the Tencent and Ninde investors who contributed most to the current round are expected to formally complete the deal in the coming days. To date, none of the relevant investors, DeepSeek and Tencent, have responded publicly. According to the source, the target for this round was initially set at $50 billion and has since been revised upwards by more than a hundred billion yuan. This funding will continue to be used for model training, numeracy and talent recruitment to support DeepSeek in global competition. Some sources have also concluded that this round of financing will pave the way for the creation of IPOs in the company ' s pipeline.
publish man-form robot "H2-X Kung Fu Legend" for a maximum of eight hours
On 21 October, publish man-made robot “H2-X Kung Fu Legend” was announced by Woo-ki Technology for a maximum duration of approximately eight hours. Wang Hing Hing Xing, founder of Wow tree technology, predicts that Physical AI will be at a turning point next year. In addition, the most well-selling human robotics in Uyuki technology have accumulated down to 12,000.
The long-duration 4F2DRAM structure has led to a technical divide.
A 4F2DRAM structure was launched this month by long-storage storage, with an alleged larger number of area transistors per unit than the mainstream 6F2 structure. On 21 October, local time, Scott DeBoer, President and Chief Technology and Products Officer of American Light Technology, responded that Mizuoka would follow the 6F2 path and develop 3D DRAM, instead of 4F2, on the grounds of increased complexity and cost, and indicated that it had studied 4F220 and considered that the direction “no viable micro-micropath”. As a result of this information and the call by Huang In-hoon for openness to cooperation, Changsung technology rose by 7 per cent on 23 October.
Kimi K3.1 Propublish
In the early hours of 24 October, Kimi K3.1 Pro was launched on the dark side of the moon. Kimi claims that this is its most capable 3 trillion-billion-degree parameter model to date, which supports 1 million context, and that this is a particularly significant increase in code, 3D creation, financial research and other long-range Agent capabilities. Kimi K3.1 Pro is a post-training version of Kimi K3. The Kimi K3.1 Pro model, through enhanced training in arithmetic and technological optimization, has been fully upgraded by large-scale RL (enhanced learning) in all areas.
The net profits from the first three quarters of Cold War doubled.
On the evening of 23 October, the Quarter 2026 was revealed by the leader of the chip. During the first three quarters, companies achieved a business income of $9.526 billion, an increase of 106.76 per cent over the same period; a net profit of $3.715 billion to the mother, an increase of 13.14 per cent over the same period; and a net profit of $3.523 billion to the non-mother, an increase of 148.30 per cent over the same period.
In the third quarter, companies achieved a business income of 3,531 million yuan, an increase of 104.47 per cent over the same period; a net profit of 1,404 million yuan for the mother, an increase of 147.74 per cent over the same period; and a net profit of 1,358 million yuan for the non-mother, an increase of 168.13 per cent over the same period. Net operating cash flows for the current season amounted to $83.677 million. In the first three quarters, net non-recurring corporate gains and losses amounted to $191 million, of which $4.625.4 million were recorded in the third quarter.
Establishment of national data groups
On 20 October, the National Data Group Ltd. was inaugurated as the 100th State Council National Investment Commission to oversee the firm in Hebei Yuan District. The Consortium is co-financed by the State Council, the Yuan Xin District Control Commission, China Telecommunications, China Connectivity, China Movement, the National Electricity Network and the Southern Power Network. It is based on the development and utilization of data resources as the main source of responsibility. It is located as the national team serving data elements and as the lead agency for data infrastructure connectivity.
paid users have reached 3706.
At the Global Conference of Founders in 2026, Lingee publish, a new round of core competency upgrades at the AI agent operating system of the Fluttershy Company, including the Ling Ki Manage, the context and the ERP connectors, helped each enterprise to accelerate the process of becoming an AI native organization. To date, there are 3706 users of paid in Lingiki. The first Agent benchmark, Lingee Bench, for a business-management scenario in the country, is organized by a LLM manufacturer and university publish.
At the same time, the Congress showcased the co-results of the Golden Flutters and the Step Star on the enterprise smart scene: the step-by-step model has landed in two scenarios, namely, the Golden Fluttershy Source Enterprise Data Analysis and the AI Stenography. Among these, the language model is pre-enumerated for accuracy in natural language reference operations, analysis of root causes, etc.; the voice model is integrated into the gold butterflies researcher's specialty in taxation, and is at the forefront of voice recognition of business management scenarios.
International information
United States Government Order: Ban DeepSeek
On 19 October, the United States Department of Homeland Security (DHS) released an official ban on DeepSeek, requiring the United States federal civil administration to stop purchasing DeepSeek-related products and services and remove them from existing information technology systems. The ban not only targets the official products of DeepSeek, but also requires the removal of third-party products and services developed through the DeepSeek technology. Among the enterprises involved are Hangzhou ' s Deep Search for Artificial Intelligence Basic Technology Research Ltd. (DeepSeek) and its associated entities.
According to documents published in the United States Federal Gazette, this order was issued on 24 July 2026 by the United States Secretary of Homeland Security, Markwayne Mullin, and was made public on 19 October and officially published on 20 October. Document No. 2026-21447, Case No. DHS-2026-0628, officially known as the Notice of Order on the Exclusion and Removal of DeepSeek Products from Information Technology Systems and Procurement Activities of the Federal Civil Administration.
Meta removes the clause that allows AI to access news data.
On 18 October, Meta announced that it had formally deleted the controversial expression in its latest commercial service provision, “the AI Model for Training in the Use of Enterprise Field Names”. This provision, which had allowed Meta to access the content of its business accounts for product development and AI training, raised legal questions. Meta responded that the update only consolidated the existing advertising provisions and did not expand the reach of the Internet; the current provisions clearly limited data collection to advertising and commercial products and reaffirmed respect for the robots.txt agreement. This was clarified and amended following multiple feedbacks.
Open AI models will land Google Clouds, including GPT-6 Astra, Cordex code Agent
On 20 October, Open AI announced that it would work with Google Clouds to introduce the Open AI front model and CodexAgentworkflow to Model Garden in Google Cloud. Google Cloud clients can purchase the Open AI model through Model Garden and complete the transaction at Google Cloud Marketplace, where the promised GCP consumption of Z quota can be charged to procurement costs. The Open AI model will soon be launched in Google Antigravity and Gemini Agent and is scheduled to be officially launched in the coming weeks. Microsoft remains the main cloud partner for Open AI.
Anthropic launched personal AI assistant Claude Duet
On 20 October, Anthropic launched the personal AI proxy application Claude Duet research preview, which opened the waiting list for both Pro and Max subscribers in the United States. Duet, based on Claude Opus 5.5, landed in IOS, Android and Claude.ai in independent applications. It sends a briefing on important matters every morning, monitors connected applications on an ongoing basis, sends messages only when they require the attention of the user and continues to complete the mission when the user leaves. Duet has built in a security taxonomy and a system alert, and refuses to share the most sensitive levels of information, such as credit card numbers and government identification.
Tesla's net profit dropped by 43% in the third quarter.
On 21 October, Tesla publish, the third quarter, net profits declined by 43 per cent in relation to each other, $28.5 billion in revenue by 0.2 per cent over the same period and operating profits by 56 per cent from $1.62 billion to $718 million. According to Tesla, the decline in business profits was mainly affected by increased operating costs associated with AI and other R & D projects, as well as increased costs of equity incentives. Global deliveries of 486532 were made during the season, a decrease of approximately 2 per cent over the same period.
GTC Berlin publishDSX POD, data centre deployment reduced from 6 months to 14 days
Beijing time, October 21, publishDSX POD AI plant reference design at the GTC Berlin speech by CEO Wong In-hoon, Weida. The product was built in modular containers and integrated and validated prior to delivery, to be delivered globally in early 2027. Wong In-hoon stated that the time from the beginning of deployment to the generation of Token could be reduced from six months to a minimum of 14 days compared to the traditional deployment. It also announced the expansion of capacity with supply chain partners, the upgrading of the Isaac robot platform and the launch of the OpenShell reference structure for AIAgent.
Anthropic will postpone the SuperIPO because of the recent apparent sale of the AI plate?
The AI unicorn Anthropic is at a historical juncture in the market, but the reverse wind in the market is increasing uncertainty.
According to Reuters, Claude chat robot developer Anthropic plans to complete IPO by mid-November, with a target value of up to $2 trillion, which is expected to go beyond SpaceX’s $177 trillion listing earlier this year, making it one of the largest technology IPOs in history. However, the requirement of the United States Securities and Exchange Commission (SEC) to publicly submit an equity statement at least 15 days prior to the launch of the road show means that the window period that Anthropic left for itself is shrinking rapidly.
At the same time, the AI plate has recently experienced a marked sale — a sharp drop of about 30 per cent in CoreWeave's share price within two weeks, and a 17 per cent fall in American light over the past month — and the market's doubts about the sustainability of AI's expenditure continue to rise. Several marketers indicated that Anthropic might be under pressure to delay listing, but analysts also pointed out that the company ' s urgent need for capital might push it to adhere to the original plan. The forecast market currently shows a probability of listing about 80 per cent in Anthropic during the year.
Muse downloads under Meta exceeded 10 million.
On 21 October, Sensor Tower data showed that Muse, an assistant under Meta ' s flag, had accumulated 10.9 million downloads since its launch in September, of which 10.2 million were from the United States, which was one of the mobile applications with 10 million downloads within 40 days of being listed. That milestone had previously been reached with only 11 applications.
Anthropic presents a personal Agent Claude Duet preview Version
Around 22 October, Anthropic launched the personal Agent Claude Duet preview. Formerly, Meta Muse, XAI GrokBot, Open AI Dots, among others, have entered the personal Agent track, and the country's Ali, Go, Tencent, and byte “beans” are advancing.
Apple Smart China Test will be online by the end of October.
Cook reportedly stated that Apple Smart China test version would be on line by the end of October, with access to Chizu, excluding Siri AI.
Financing
New Stones completes the $400 million E round, doubles the raise.
On 20 October, Neolix, an undistributed enterprise, completed a $400 million E round of financing, which was co-financed with a national fund, with investments from the AEF NextGen Hong Kong Fund, Vertex Growth, Guangxi Capital, and Real Robots, as well as investments from old shareholders in stone capital, soft and silver Asia, Japan, and high-tech investments. As revealed by the founders, the original plan was only $200 million, with actual over-mobilization doubling. For the first time, the National Fund for Investment in the Artificial Intelligence Industry has included undistributed enterprises in its portfolio.
Toshiyama sought out Hong Kong shares to raise up to HK$10.7 billion
On 20 October, as announced by HKBO, Sioux Tungshan ' s sophisticated manufacture is proposed to issue approximately $113.8 million shares in Hong Kong at a maximum cost of HK$93.80 per share, and is expected to raise no more than HK$10.7 billion (US$1.4 billion) and to be registered on 28 October. Capitalist investors include Aurora SF, EMU, Morgan Asset Management (Asia-Pacific) and Alibaba Investment.
Fort Robotics received approximately $20 million in new financing
Fort Robotics, the entity ' s security platform, received an additional commitment of approximately $20 million from existing investors, and Tiger Global contributed about $10 million, with the participation of Mark Cuban Companies. The company is seeking to market NASDAQ through a merger with Newbury Street II, which is expected to be “FROB”.
Billionaire cast technology completes over $1 billion in new financing.
The depository AI chip company Cast Technology announced the completion of over $1 billion in new funding, and the 3D DRAM-based full-digit universal storage of AI chip was completed. This round of financing will be used for R & D inputs, product overlay and industrialization.
Ita powered over billions of yuan in seed and seed + round finance
A start-up enterprise focused on the development of smart-operative platforms and power systems, The Ita Dynamics, completed a multi-billion dollar seed and seed + wheel financing.
Trends
TrendForce: Estimated 31 per cent increase in the delivery of CPU from server in 2026, Agentic AI boosting demand
On 22 October, the TrendForce Group Advisory noted that, with the active introduction of AI Agent-related applications by AI service providers, the demand for servers for CPU had grown steadily since the second half of 2025, with an estimated 31 per cent increase in the year 2026. As a result of the growing demand for AI reasoning and the increased penetration of Agenic AI services in 2027, the volume of deliveries is expected to increase by more than 40 per cent a year, with a possible change in market patterns led by the x86 camp.
CSI: AI Technology Enterprise CDS spreads widely and has not yet evolved into an industry credit risk
According to CSI, the scale of capital expenditure in AI continues to increase, and the related subject-matter financing costs are increasing, prompting investors to increase their interest in and hedge against the credit risk of AI-technology enterprises, and the A-technology-based CDS market shows a double upswing in turnover and spreads. However, it has not yet evolved into a full-scale rise in credit risk for the industry, with gradients for the major AI-related enterprises, the CDS spreads.
9 sectors, such as the Ministry of Commerce: Accelerating brain interfaces, smart cutting edges, etc.
On 19 October, publish, Department of Commerce and others, commented on the need to accelerate the advancement of science and technology, proposing to accelerate the development of key technologies in the areas of new displays, smart terminals, and to accelerate the transformation of cutting-edge technological achievements, such as brain interfaces and smarts, to the market and to support the declaration of niche new business enterprises in the areas of smart warehousing, digital supply chains, etc.
National Statistical Office: Operation Artificial Intelligence+ is a powerful engine of investment and exports in emerging areas
On 19 October, Wang Guanghua, Press Spokesman for the National Statistical Office, stated that the “Advisible Intelligence Plus” initiative in the first three quarters had given a strong impetus to investment and exports in new and emerging areas. Electronics, integrated circuit manufacturing and investment in information services grew by 58.7 per cent, 14.2 per cent and 20.2 per cent, respectively; exports of electronics components, cables, etc. grew by 70 per cent and 25.3 per cent, respectively.
Ministry of Industry and Communications: The 60 billion National AI Fund has invested in a number of projects with over 6,600 artificial intelligence enterprises
On 19 October, the Director of the Science and Technology Department of the Ministry of Industry and Communications, Wei Qiang, presented at the publish meeting of the New Zone Office of the United Nations High Commissioner for Human Rights, and the 60 billion yuan National Fund for Investment in Artificial Intelligence Industries has launched a number of high-quality projects. In 2025, China ' s core AI industry was over 1.2 trillion yuan, with over 30 per cent access to AI technology for manufacturing and above, and over 6,600 by the end of June of this year. All three of the projects disclosed fell on the head of LLM: participation in DeepSeek ' s first round of financing was approximately 980 million yuan, valued at 350 billion yuan; participation in the month ' s dark side of F-wheel financing was about 1.14 per cent.
9 departments such as Commerce have sent letters to promote technology.
On 19 October, publish, the Ministry of Commerce and others, issued an Opinion on Accelerating the Development of Science and Technology, proposing 18 specific initiatives covering the promotion of the consumption of science and technology-enabled goods, the expansion of the consumption of services by science and technology, the smooth flow of technology and technology, and the shaping of future business landscapes. It proposes to further implement the AIS+ Manufacturing initiative, which will enrich the supply of smart commodities such as smart phones, computers, homes, wear equipment, robots, etc.; promote brain interfaces, smart end-uses such as smarts; and pilot applications such as AIS+Culture and Tourism.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.