晚安咖啡GoodNightCoffee

Podcast · 生活休闲

晚安咖啡GoodNightCoffee

每日更新全球商业与科技热点,中英双语轻松掌握

主播 NullName 出品 nullName 生活休闲 · 休闲 8,992 订阅 1380 集 16小时前更新

在小宇宙收听

简介

"晚安咖啡" 一杯陪你从清晨到深夜的神奇咖啡。 早晨,来杯清醒提神的商业浓缩, 读懂全球科技脉搏, 创业风云、商业变革, 让你的早安问候都带着点儿专业范。 午间,来杯香醇的娱乐拿铁, 影视剧追踪,明星动态, 让你的话题库永远新鲜, 办公室八卦,从此不用怕。 夜晚,来杯温暖特调, 治愈故事,生活感悟, 让一天的疲惫, 在耳机里慢慢融化。 每一杯咖啡的温度, 都能让你的碎片时间闪闪发光。 洗漱、通勤、午休、健身, 每个平凡时刻, 都能遇见不平凡的分享。 这是一杯, 让你在生活中保持清醒, 但不会让你失眠的咖啡。 你的随身资讯管家, 每天都为你精心冲泡, 一杯又一杯, 不见不散。

《晚安咖啡》科技夜话(英文版):英伟达财报超预期,Anthropic IPO火爆

《晚安咖啡》科技夜话(英文版):英伟达财报超预期,Anthropic IPO火爆

晚安咖啡GoodNightCoffee

Hi everyone~~ Good evening! It’s Wednesday, November 18, 2026. Welcome to Goodnight Coffee—I’m your host, An’an! Tonight I’ll walk you through some interesting business and tech stories. We’ll keep it light and fun, and wrap up today on an upbeat note. Ready? Let’s get started! @@@ According to CNBC, NVIDIA shares rose 3.72% in Wednesday premarket trading after the company once again reported better-than-expected results. Its third-quarter revenue was $115.84 billion, more than $6 billion above analysts’ $109.10 billion estimate; fourth-quarter revenue guidance was $130 billion, above LSEG’s $125.1 billion consensus. Boosted by the results, new cloud providers CoreWeave, Nebius, and Iren rose 6.1%, 8.2%, and 7.2%, respectively. Shortly before the earnings release, CEO Jensen Huang, who resigned last week, made his first public statement, saying he was willing to return to the company after several prominent tech figures voiced support. But new board chair Harvey Jones—NVIDIA’s first-ever board chair—said on the earnings call that the board does not plan to have Huang serve as CEO again. Interim CEO Colette Kress told CNBC that hyperscalers are urgently seeking more available compute, that claims of a shift in customer types are not true, and that all types of customers are still working with NVIDIA, whose customer base is very broad. The US Commerce Department announced regulatory adjustments on Monday to restrict Chinese companies’ remote access to compute in overseas data centers, aiming to stop China from obtaining NVIDIA’s most advanced chips. NVIDIA said in its earnings commentary that its outlook assumes no data center compute revenue from China. @@@ According to a report from Jiemian News published with authorization from 36Kr, Gerstner, CEO of Altimeter Capital, said the firm applied to subscribe for $5 billion worth of shares in Anthropic’s IPO—its largest allocation request in 20 years. The IPO is 10 to 20 times oversubscribed, and the order book is more than five times the size of the previous largest IPO order. However, the request has not yet been filled, and the oversubscription multiple is based on its own market judgment. Anthropic plans to price the deal local time on November 18 and list on Nasdaq the next day under the ticker ANT, with an offering range of $70 to $78. The IPO could raise up to $64.4 billion, plus a private placement of about $26 billion, for a total of up to $90.4 billion. At the top end, the fully diluted valuation would be about $1.4 trillion. Investors including Sequoia, Iconiq, Lightspeed, Accel, and Altimeter plan to subscribe billions of dollars. Private placement participants include Google and NVIDIA at $10 billion each, AMD at $2.5 billion, Broadcom, Micron, and Microsoft at $1 billion each, and a TSMC affiliate at $500 million. After completion, Amazon and Google will hold 14.7% and 13.7% of outstanding shares, respectively. @@@ According to Readhub, Microsoft has launched six AI models spanning coding, cybersecurity, image generation, and voice. Its coding model MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at only 40% of the price; against Claude Opus 5.5, it reaches 96% performance at just 5% of the cost. Cybersecurity model MAI Cyber 2 Flash outperformed GPT-6 Astra in relevant security tests, with per-task cost less than 2% of the latter’s. Image model MAI Image 2.7 ranked first in image editing. Mustafa Suleyman, Microsoft’s AI chief, said the company is accelerating work on its own general AI models to reduce dependence on external models, and plans to open model weights to enterprises so customers can customize models with their own data. @@@ According to CNBC, JPMorgan is testing with Hyundai Card how AI agents can handle financial transactions without direct human intervention and assessing whether they could eventually negotiate and make payments with each other. The two companies have completed an initial proof of concept simulating an FX transaction. In the test, Hyundai Card’s AI agent requested an FX transaction to maintain euro liquidity in its Kinexys blockchain deposit account. JPMorgan said this was the first test of an autonomous bank agent versus a client agent on a private blockchain for corporate treasury management. Wee Kee Toh, global head of Kinexys Insights at JPMorgan, said the next step will be to test autonomous agents going beyond executing transactions and interacting with one another in the ecosystem—the biggest part of expanding the first phase. The two companies plan to launch a more complex phase involving multiple AI agents in the first half of 2027. After the initial trial, JPMorgan has not set a commercialization timetable. Toh said the biggest obstacle is not the technology itself, but governance, internal controls, and accountability and risk without direct human oversight. AvePoint’s 2026 State of AI report shows that unauthorized autonomous agent behavior is the third most common AI agent-related security vulnerability, affecting 34.1% of companies that reported such incidents. CEO Tianyi Jiang warned that agent interactions can amplify errors; in the worst case, a compromised or hallucinating agent gives bad data to another, potentially triggering a high-speed cascade of erroneous fund transfers before humans notice. @@@ According to TechWeb, JD.com announced unaudited results for the three months ended September 30, 2026: revenue was RMB309.6 billion, up 3.5% year over year; operating profit was RMB5.2 billion, compared with a loss of RMB1.1 billion a year earlier, swinging to profit; net income attributable to shareholders was RMB6.5 billion, compared with RMB5.3 billion a year earlier; non-GAAP net income attributable to shareholders was RMB8.3 billion, up 44%. By segment, JD Retail revenue was RMB257.966 billion, up 2.9%, with operating profit of RMB15.1 billion and a 5.8% margin; JD Logistics revenue was RMB63.953 billion, up 16.1%; new business revenue was RMB7.831 billion, down 49.8%, with an operating loss of RMB9.642 billion, narrowing from RMB15.736 billion a year earlier, as JD Food Delivery and Jingxi drove a clear reduction in losses. Marketing expenses were RMB15.5 billion, down 26.2%; operating margin was 1.7%, compared with negative 0.4% a year earlier. JD completed its voluntary public takeover offer for CECONOMY at €1.335 billion, acquiring 59.8% of share capital and voting rights; CECONOMY has been consolidated since November 12. As of September 30, cash and short-term investments totaled RMB214.7 billion; third-quarter free cash flow was negative RMB23.861 billion. @@@ According to TechWeb, after new US export control rules took effect, a $480 million loan for ByteDance to buy NVIDIA GPUs in Malaysia was halted. At least five banks evaluating the deal have internally discussed withdrawing or not participating, and one banker said the loan is “effectively dead.” The financing was borrowed by Singapore cloud provider Aolani Cloud and led by DBS Group; the funds were to buy NVIDIA GPUs in Malaysia and then lease the compute to ByteDance for remote use. On November 16, the US Bureau of Industry and Security designated Malaysia, Singapore, Thailand, Indonesia, and Brazil as high-risk transshipment countries; exports, reexports, or in-country transfers of advanced AI chips to these countries require US licenses. The new rules also regulate remote access: if an operating entity knows or should know that the end user is in China, and the aggregate processing performance exceeds 12 million—roughly equivalent to 700 to 800 H100s—the relevant chip supply must undergo license review. Banks are not worried about ByteDance’s ability to repay, but that once they lend, they could become liable for violating US export controls. Aolani said it is coordinating with regulators and financing partners to ensure compliance. As of publication, ByteDance had not publicly commented. @@@ According to Huxiu 24 Hours, Zong Fuli has made another personnel adjustment, with 173 managers removed from their posts. @@@ Okay, that’s it for today’s show! Thank you for spending time with me and chatting about so many light and fun little stories. I hope you’re in a good mood today too. If you enjoyed the show, you’re very welcome to subscribe and share it with your friends—it helps us a lot. See you next time!

9 分钟 23 次播放 16小时前

《晚安咖啡》科技夜话:英伟达财报超预期,Anthropic IPO火爆

《晚安咖啡》科技夜话:英伟达财报超预期,Anthropic IPO火爆

晚安咖啡GoodNightCoffee

嗨~~大家晚上好呀!今天是2026年11月18日,星期三,欢迎来到晚安咖啡,我是主播安安!今晚我会陪你一起看看有趣的商业科技动态,轻松聊、开心听,给今天画个元气满满的句号。准备好啦,我们开始吧! @@@ 据CNBC报道,英伟达股价周三盘前交易中上涨3.72%,此前公司再次公布超预期财报。其第三季度营收为1158.4亿美元,比分析师预期的1091.0亿美元高出逾60亿美元;第四季度营收指引为1300亿美元,高于LSEG统计的1251亿美元共识。受业绩带动,新型云服务商CoreWeave、Nebius和Iren股价分别上涨6.1%、8.2%和7.2%。在财报发布前不久,上周辞职的CEO黄仁勋首次公开表态,称在多位知名科技人士声援后,他愿意回归公司;但新任、也是英伟达首位董事会主席Harvey Jones在财报电话会上表示,董事会不打算让他重新担任CEO。临时CEO Colette Kress对CNBC表示,超大规模云服务商正急切寻求更多可用算力,所谓客户类型发生变化并不属实,各类客户都仍在与英伟达合作,公司客户群非常广泛。美国商务部周一宣布调整法规,限制中国企业远程访问海外数据中心算力,以阻止中国获取英伟达最先进芯片;英伟达则在财报评论中称,其展望不假设来自中国的任何数据中心计算收入。 @@@ 据界面新闻经36氪授权发布的消息,Altimeter Capital首席执行官格斯纳称,公司申请认购Anthropic IPO价值50亿美元的股票,为其20年最大配售申请;此次IPO超额认购10至20倍,订单规模是此前最大IPO订单五倍以上,但申请尚未完成购买,超额倍数基于其市场判断。Anthropic计划当地时间11月18日定价,次日以代码ANT在纳斯达克挂牌,发行区间70至78美元,IPO最多募资644亿美元,另安排约260亿美元私募配售,合计最高904亿美元,按上限定价完全稀释估值约1.4万亿美元。红杉、Iconiq、光速、Accel和Altimeter等投资方计划认购数十亿美元。私募配售方包括谷歌、英伟达各100亿美元,AMD25亿美元,博通、美光、微软各10亿美元,台积电关联5亿美元;完成后亚马逊、谷歌分别持有流通股14.7%和13.7%。 @@@ 据Readhub报道,微软推出六款AI模型,覆盖编程、网络安全、图像生成及语音等领域。其中,编程模型MAI Code 2的性能可达GPT-6.1 Sol的98%,价格仅为其40%;相较Claude Opus 5.5,其性能达到96%,成本仅为5%。网络安全模型MAI Cyber 2 Flash在相关安全测试中的表现超过GPT-6 Astra,单项任务成本不到后者的2%。图像模型MAI Image 2.7位列图像编辑排名第一。微软AI负责人苏莱曼称,公司正加速研发自主通用AI模型,减少对外部模型的依赖,并计划向企业开放模型权重,支持客户利用自有数据定制模型。 @@@ 据CNBC报道,摩根大通正与韩国现代信用卡测试AI代理如何在无直接人工干预下处理金融交易,并评估其最终能否相互谈判和付款。两家公司已完成一项模拟外汇交易的初步概念验证。测试中,现代信用卡的AI代理请求进行外汇交易,以维持其在Kinexys区块链存款账户中的欧元流动性;摩根大通称,这是首个基于私有区块链、面向企业财资管理的自主银行代理对客户代理测试。摩根大通Kinexys Insights全球主管Wee Kee Toh表示,下一步将测试自主代理不再局限于执行交易、而是在生态系统中彼此互动,这是扩大第一阶段范围的最大部分;两家公司计划2027年上半年启动涉及多个AI代理的更复杂阶段。初步试验完成后,摩根大通尚未设定商业化时间表。Toh称,技术本身并非最大障碍,更难的是治理、内部控制以及无直接人工监督下的责任和风险。AvePoint《2026年AI现状》报告显示,未经授权的自主代理行为是第三常见的AI代理相关安全漏洞,影响34.1%报告此类事件的公司。其首席执行官Tianyi Jiang警告,代理互动可能放大错误,最坏情况下,一个被入侵或产生幻觉的代理向另一个提供错误数据,可能在人类察觉前引发高速连锁的错误资金转移。 @@@ 据TechWeb报道,京东集团发布截至2026年9月30日止三个月未经审计业绩:收入3096亿元,同比增长3.5%;经营利润52亿元,上年同期亏损11亿元,实现扭亏;归母净利润65亿元,上年同期53亿元;非GAAP归母净利润83亿元,同比增长44%。分业务看,京东零售收入2579.66亿元,同比增长2.9%,经营利润151亿元,利润率5.8%;京东物流收入639.53亿元,同比增长16.1%;新业务收入78.31亿元,同比下降49.8%,经营亏损96.42亿元,较上年同期157.36亿元收窄,京东外卖和京喜带动明显减亏。营销开支155亿元,同比下降26.2%;经营利润率1.7%,上年同期为负0.4%。京东以13.35亿欧元完成对CECONOMY的自愿公开收购要约,取得59.8%股本及投票权,后者自11月12日起并表。截至9月30日,现金及短期投资合计2147亿元;第三季度自由现金流为负238.61亿元。 @@@ 据TechWeb消息,美国出口管制新规生效后,字节跳动一笔用于在马来西亚采购英伟达GPU的4.8亿美元贷款被叫停。至少五家评估该交易的银行已内部讨论退出或不参与,有银行人士称贷款“实际上已经死亡”。融资由新加坡云服务商Aolani Cloud借款,星展集团牵头安排,资金用于在马来西亚采购英伟达GPU,再以算力租赁供字节跳动远程调用。11月16日,美国商务部工业与安全局将马来西亚、新加坡、泰国、印尼和巴西列为转运高风险国家,向这些国家出口、再出口或境内转移先进AI芯片须申请美国许可;新规还把远程访问纳入管制:若经营实体知道或应知最终用户位于中国,且算力总处理性能超过1200万、约相当于700至800块H100,相关芯片供应须接受许可审查。银行担忧的并非字节还款能力,而是一旦放贷,自身可能成为违反美国出口管制的责任主体。Aolani称正与监管机构和融资伙伴协调确保合规。截至发稿,字节跳动未公开置评。 @@@ 据虎嗅24小时报道,宗馥莉再次进行人事调整,此次共有173名管理人员被免职。 @@@ 好啦,今天的节目到这里就结束啦!谢谢你们陪我聊了这么多轻松有趣的小事,希望今天也有好心情。如果您喜欢我们的节目,也非常欢迎订阅和转发给你的小伙伴,这对我们非常有帮助。那我们下期再见吧!

9 分钟 44 次播放 16小时前

《晚安咖啡》科技夜话:Anthropic拟万亿美元IPO,苹果2027推智能眼镜,阿里发Qwen4-Max

《晚安咖啡》科技夜话:Anthropic拟万亿美元IPO,苹果2027推智能眼镜,阿里发Qwen4-Max

晚安咖啡GoodNightCoffee

嗨~~大家晚上好呀!今天是2026年11月16日,星期一。欢迎来到晚安咖啡,我是主播安安~今天咱们一起聊聊最新的商业科技动态,干货满满,别走开哦! @@@ 据The Verge报道,围绕英伟达联合创始人、前首席执行官黄仁勋的事态正快速演变。报道提到,有传闻称黄仁勋可能正在寻找重返CEO职位的路径。与此同时,Megaspeed发表声明称,Alice Huang自2025年8月以来未在该公司工作。报道称,这份声明引发的疑问多于它能回答的问题。The Verge还表示,其Stepback新闻通讯最新一期已梳理相关情况,而自当天上午以来,事件又出现多项新进展。 @@@ cnBeta援引彭博社报道,Anthropic预计11月19日或20日登陆纳斯达克,寻求1.4万亿美元估值。公司计划通过IPO和一项独立私募配售合计融资最多904亿美元,规模将超过SpaceX IPO的750亿美元基础融资,成为有史以来最大上市交易;知情人士称,此次IPO已获近七倍超额认购。Anthropic拟发行8.25亿股,发行价每股70至78美元。截至10月31日,公司年化营收810亿美元,较9月底增加100亿美元;现金及等价物超过1300亿美元。高盛预计其2030年才能实现年度自由现金流为正;到2045年支出承诺超9000亿美元,其中约710亿美元在2027年9月底前到期。上周五,CEO阿莫代伊与摩根大通CEO戴蒙在纽约IPO活动上同台推介;知情人士称,Anthropic不准中国大陆和香港机构投资者参与IPO认购。此前黄仁勋声誉危机致英伟达股价重挫,Anthropic被寄望接替英伟达成为AI新标杆。 @@@ 据The Verge报道,彭博社记者马克·古尔曼称,苹果计划在2027年6月的全球开发者大会(WWDC)上发布首款智能眼镜,并在当年年底前开始出货。该产品不配备增强现实功能,核心卖点是新版Siri人工智能和iPhone应用生态。原型代号N50,目前运行精简版visionOS,采用基于旧款智能手表芯片的新处理器。眼镜配有类似MagSafe的扁平矩形磁吸充电器,大小约相当于一粒Tic Tac,用户戴着眼镜也能充电。苹果还强调外观、隐私和音质,员工称其音质达到行业领先水平,支持通话、音乐播放和FaceTime音频。眼镜具备Wi-Fi,离开iPhone后,部分非第三方应用功能仍可继续使用,例如通过Wi-Fi继续听音频,但它不会成为完全独立的设备。新版Siri将增加对应用的细粒度控制,苹果先让开发者在iPhone上适配,以期眼镜推出后能直接使用相关功能。 @@@ 据The Verge报道,美国联合发射联盟(ULA)的“火神”(Vulcan)火箭在一次发射中发生异常,未能将40颗亚马逊“Leo”卫星送入轨道。ULA在社交平台X发表声明称,根据发射主管收到的数据,这枚从41号航天发射复合体升空的火箭在飞行10分钟后出现异常,导致亚马逊任务未能入轨;目前尚不清楚故障原因,公司承诺在调查过程中保持透明。这据报是ULA成立20年来首次任务失败。此次发射原计划是亚马逊向ULA购买的38次发射中的第一次。“Leo”是亚马逊对标SpaceX“星链”的卫星项目,但此前进展艰难,迄今仅有少量所需卫星入轨。亚马逊表示,任何发射都有风险,首次任务尤其如此,ULA一直是其可信赖的合作伙伴,相信ULA会从此次任务中吸取教训,并落实“火神”的长期计划。 @@@ 据TechWeb援引快科技消息,美国加州北区联邦地区法院法官诺埃尔·怀斯11月14日全盘驳回三星、美光、SK海力士的撤诉动议,认定原告已合理指控三家违反《谢尔曼法》第一条、达成非法协议,DRAM集体反垄断诉讼进入实质审理。诉讼由美国五州17名消费者及小企业于2026年6月提起,指控三家协同限制常规DRAM供应、推高价格。诉状称,截至2025年第三季度,三家合计占全球DRAM营收逾91%;自2022年底同步削减传统消费级DRAM产能,2023年将大量晶圆产能转向HBM,18个月内传统DRAM价格累涨约697%。法院列举的同步行为包括三星与SK海力士拒绝长期供货协议、三家推迟或叫停新晶圆厂、同季度减产DDR4、美光2025年12月关闭30年历史的Crucial品牌、2026年1月上线高度相似客户审核机制,并提及其1998至2002年曾涉价格操纵案。三家企业称决策为适配AI市场变化,且类似争议2018年已被驳回;法官认为本案事实和证据有本质区别。三家须在2026年12月7日前提交完整答辩文件,首次管理会议定于2027年1月12日。 @@@ 据新浪科技报道,11月16日,据千问消息,阿里巴巴发布全新一代基座大模型Qwen4-Max,总参数2.4T,具备原生视觉能力,支持百万上下文。该模型基于新的架构和训练优化,在智能体、编程等关键能力上持续进化,并在金融、法律等专业垂直领域实现从“信息检索”到“专业交付”。上线期间,千问APP会员可限免体验,使用Qwen4-Max不消耗会员额度。相关报道显示,其预览版已于11月11日首发千问AI平台Token Plan,同步接入千问办公、Qoder、千问APP和QwenStudio,预览窗口持续至11月25日24时,主要用于能力验证和收集反馈。另据澎湃新闻报道,Qwen4-Max已上线千问AI平台提供API服务,缓存命中每百万Tokens低至0.5元,完成单任务平均费用较上代模型下降六成;用户也可通过阿里Token Plan调用,Qwen4系列中小尺寸模型将陆续开源。 @@@ 据新浪科技报道,腾讯文化发文披露,腾讯在28岁生日当天将一年一度总办LinkTime做成特别播客,总办就AI时代机会、挑战与选择发声。讲话提出,所有核心业务团队都要面向AI时代再创业:AI native业务、模型和产品要逆流而上,核心业务一半人守住、一半人打新战役,组织扁平化、以小团队作战,打破山头和中间传话阻碍。腾讯称要做以人为本的AI,坚持造福人类和社会、由人控制、赋能创造力;AI仍处早期,算力基建需自建且价格飞涨,投入会兼顾产出比。其AI金字塔以模型为核心,已进入国内第一梯队;小微定位私人AI助理,目标是“token自由”;元宝未来定位决策和资讯参谋;WorkBuddy在8个月内成为中国提供最多token的单一AI产品。总办还要求心态归零、加快反馈决策和行动、提升人才密度,并提到游戏、金融、安全等团队已与模型Co-Design带来Hy能力提升。 @@@ 据新浪科技报道,华为终端11月16日上午宣布,鸿蒙终端设备数已超过1亿台。此前,华为常务董事、终端BG董事长余承东通过社交媒体宣布,全新子品牌OSTA星芒将于11月20日在“鸿蒙智行享界V8及星芒品牌发布会”上正式亮相。消息称,星芒将共享鸿蒙操作系统、麒麟芯片技术,以及华为覆盖线上线下的渠道与售后体系。 @@@ 新浪科技消息,11月16日,因算力资源供给紧张,DeepSeek面向普通用户的App及网页版对话服务推出限额措施:存续超过7天且无活跃交互的长对话将无法接续;单次输入文字上限约4万字;每日对话额度用尽会提示“因服务器繁忙,使用额度已达上限”。分析认为,此举意在回收闲置会话占用的KV缓存,优先保障活跃用户推理。10月以来其服务多次承压,官方状态页记录8起故障,发生在10月2日、8日、12日(两次)、15日、20日、27日和11月3日,多为响应变慢等性能降级;10月15日故障持续超4小时,C端服务及V4 Pro、V4.1 Flash API同步受限,事后对API用户赠金补偿。8月至10月,V4.1 Flash API可用性99.46%,V4 Pro为99.74%。DeepSeek近期开放全家桶应用API,API侧实行峰谷分时定价和独立并发限额。OpenRouter周榜显示,10月19日至25日,V4.1-Flash-1017以33.6万亿Token周调用量居首,环比增801%。 @@@ 据新浪科技报道,11月12日超聚变创业板IPO过会,从5月22日受理到过会仅用174天,80亿元募资额使其成为创业板年内最大IPO项目。公司营收从2023年250.92亿元增至2025年582.46亿元,2026年上半年为372.86亿元,同比增20.78%;AI服务器收入占比从25.03%升至53.29%。但主营业务毛利率从14.27%降至2025年8.58%,净利率1.77%;2026年上半年毛利率回升至15.49%,净利润21.08亿元,已超2025年全年两倍。2025年研发费用率2.02%,低于中科曙光、紫光股份和浪潮信息。2026年上半年经营现金流净额-156.85亿元,应收账款153.03亿元,较2025年末增超170%,存货356.62亿元,增68%。截至2026年7月末在手订单301.13亿元,其中AI服务器占73.70%。 @@@ 好啦,今天的节目到这里就结束啦!谢谢你的陪伴,愿你今晚有个好心情~如果您喜欢我们的节目,也非常欢迎订阅和转发给你的小伙伴,这对我们非常有帮助。那我们下期再见吧!

12 分钟 99+ 次播放 2天前

《晚安咖啡》科技夜话(英文版):Anthropic拟万亿美元IPO,苹果2027推智能眼镜,阿里发Qwen4-Max

《晚安咖啡》科技夜话(英文版):Anthropic拟万亿美元IPO,苹果2027推智能眼镜,阿里发Qwen4-Max

晚安咖啡GoodNightCoffee

Hi there~~ Good evening, everyone! Today is Monday, November 16, 2026. Welcome to Goodnight Coffee. I'm your host, An'an. Today we're going to talk about the latest business and tech developments—packed with useful insights, so stay tuned! @@@ The Verge reports that events surrounding Nvidia co-founder and former CEO Jensen Huang are evolving rapidly. The report notes rumors that Jensen Huang may be looking for a path back to the CEO role. Meanwhile, Megaspeed issued a statement saying that Alice Huang has not worked at the company since August 2025. The report says the statement raises more questions than it answers. The Verge also says the latest issue of its Stepback newsletter has reviewed the situation, and since that morning, several new developments have emerged. @@@ cnBeta, citing Bloomberg, reports that Anthropic is expected to list on Nasdaq on November 19 or 20, seeking a $1.4 trillion valuation. The company plans to raise up to $90.4 billion through an IPO and a separate private placement, which would exceed SpaceX's $75 billion base IPO raise and become the largest listing ever; people familiar with the matter say the IPO is nearly seven times oversubscribed. Anthropic plans to issue 825 million shares at $70 to $78 each. As of October 31, the company's annualized revenue was $81 billion, up $10 billion from the end of September; cash and equivalents exceed $130 billion. Goldman Sachs expects it to achieve positive annual free cash flow only in 2030; spending commitments by 2045 exceed $900 billion, about $71 billion of which comes due before the end of September 2027. Last Friday, CEO Amodei and JPMorgan CEO Dimon appeared together at an IPO event in New York to pitch the deal; people familiar with the matter say Anthropic will not allow institutional investors from mainland China and Hong Kong to participate in IPO subscriptions. Earlier, Jensen Huang's reputation crisis caused Nvidia shares to slump, and Anthropic is expected to succeed Nvidia as the new AI benchmark. @@@ The Verge reports that Bloomberg's Mark Gurman says Apple plans to unveil its first smart glasses at its Worldwide Developers Conference in June 2027 and begin shipping them by the end of that year. The product will not have augmented reality features; its core selling points are the new Siri AI and the iPhone app ecosystem. The prototype is codenamed N50, currently runs a stripped-down version of visionOS, and uses a new processor based on an older smartwatch chip. The glasses come with a flat rectangular MagSafe-like magnetic charger about the size of a Tic Tac, and users can charge them while wearing them. Apple also emphasizes design, privacy and audio quality. Employees say its audio quality is industry-leading, and it supports calls, music playback and FaceTime audio. The glasses have Wi-Fi; after leaving the iPhone, some non-third-party app functions can still be used, such as continuing to listen to audio via Wi-Fi, but it will not be a fully independent device. The new Siri will add granular app controls; Apple is first having developers adapt on the iPhone so that related functions can be used directly once the glasses launch. @@@ The Verge reports that United Launch Alliance's Vulcan rocket experienced an anomaly during a launch and failed to send 40 Amazon Leo satellites into orbit. ULA said in a statement on X that, according to data received by the launch director, the rocket, which lifted off from Space Launch Complex 41, experienced an anomaly 10 minutes after liftoff, causing the Amazon mission to fail to reach orbit; the cause is not yet known, and the company promised transparency during the investigation. This is reportedly ULA's first mission failure in its 20-year history. The launch was originally planned as the first of 38 launches Amazon purchased from ULA. Leo is Amazon's satellite project aimed at competing with SpaceX's Starlink, but it has struggled, with only a small number of the needed satellites in orbit so far. Amazon says any launch carries risk, especially a first mission; ULA has always been a trusted partner, and it believes ULA will learn from this mission and carry out Vulcan's long-term plan. @@@ TechWeb, citing Kuaikeji, reports that on November 14, Judge Noelle Wise of the U.S. District Court for the Northern District of California denied in full motions to dismiss by Samsung, Micron and SK Hynix, finding that the plaintiffs have plausibly alleged that the three violated Section 1 of the Sherman Act and reached an illegal agreement, moving the DRAM class-action antitrust lawsuit into substantive proceedings. The lawsuit was filed in June 2026 by 17 consumers and small businesses in five U.S. states, alleging that the three coordinated to restrict conventional DRAM supply and push up prices. The complaint says that as of the third quarter of 2025, the three together accounted for more than 91% of global DRAM revenue; since late 2022, they synchronized cuts to conventional consumer DRAM capacity, and in 2023 shifted substantial wafer capacity to HBM, with conventional DRAM prices cumulatively rising about 697% in 18 months. The court listed synchronized conduct including Samsung and SK Hynix refusing long-term supply agreements, the three delaying or halting new fabs, cutting DDR4 production in the same quarter, Micron closing its 30-year-old Crucial brand in December 2025, launching highly similar customer audit mechanisms in January 2026, and mentioned that they were involved in a price-fixing case from 1998 to 2002. The three companies say their decisions were adapting to changes in the AI market and that similar disputes were dismissed in 2018; the judge found the facts and evidence in this case fundamentally different. The three must submit complete answer documents by December 7, 2026, and the first management conference is set for January 12, 2027. @@@ Sina Tech reports that on November 16, according to Qwen, Alibaba released its new-generation foundation model Qwen4-Max, with 2.4 trillion total parameters, native vision capabilities and support for a million-token context. Based on a new architecture and training optimizations, the model continues to evolve in key capabilities such as agents and coding, and in professional verticals such as finance and law it moves from "information retrieval" to "professional delivery." During the launch period, Qwen app members can try it free for a limited time, and using Qwen4-Max does not consume membership quota. Related reports show that its preview version first launched on November 11 on the Qwen AI platform Token Plan and was simultaneously integrated into Qwen Office, Qoder, the Qwen app and QwenStudio; the preview window runs until 24:00 on November 25, mainly for capability validation and collecting feedback. The Paper also reports that Qwen4-Max is now available via API on the Qwen AI platform, with cache hits as low as RMB0.5 per million tokens, and the average cost per completed task down 60% from the previous generation; users can also call it through Alibaba Token Plan, and the small- and medium-sized models in the Qwen4 series will be open-sourced successively. @@@ Sina Tech reports that an article from Tencent Culture disclosed that on its 28th birthday, Tencent turned its annual Executive Office LinkTime into a special podcast, with the Executive Office speaking out on the opportunities, challenges and choices in the AI era. The remarks proposed that all core business teams must start a second venture for the AI era: AI-native businesses, models and products must go upstream against the current; half of the people in core businesses should defend the core, while the other half should fight new battles; the organization should be flattened and operate in small teams, breaking down fiefdoms and barriers of middle-layer relayed messaging. Tencent said it wants to build human-centered AI, adhering to benefiting humanity and society, being controlled by humans and empowering creativity; AI is still in its early stages, compute infrastructure must be self-built and prices are soaring, and investment will balance input and output efficiency. Its AI pyramid is centered on models and has entered China's first tier; Xiaowei is positioned as a personal AI assistant with the goal of "token freedom"; Yuanbao will be positioned as a decision-making and information adviser; WorkBuddy became the single AI product in China providing the most tokens within eight months. The Executive Office also required a zero-based mindset, faster feedback, decision-making and action, and higher talent density, and mentioned that teams in gaming, finance, security and other areas have already co-designed with models, bringing improvements in Hy capabilities. @@@ Sina Tech reports that Huawei Terminal announced on the morning of November 16 that HarmonyOS terminal devices have exceeded 100 million units. Previously, Huawei's Executive Director and Chairman of the Terminal BG Yu Chengdong announced via social media that the new sub-brand OSTA Xingmang will officially debut on November 20 at the "HarmonyOS Intelligent Mobility Stelato V8 and Xingmang Brand Launch Event." Reports say Xingmang will share HarmonyOS, Kirin chip technology, and Huawei's online and offline channel and after-sales systems. @@@ Sina Tech reports that on November 16, due to tight supply of compute resources, DeepSeek introduced usage limits for its consumer app and web chat services: long conversations that have lasted more than seven days without active interaction can no longer be continued; single input text is capped at about 40,000 characters; when the daily chat quota is exhausted, users will see the prompt "Server busy, usage limit reached." Analysts believe the move is intended to reclaim KV cache occupied by idle sessions and prioritize inference for active users. Since October, its services have come under pressure multiple times; the official status page has recorded eight incidents, occurring on October 2, October 8, October 12 (twice), October 15, October 20, October 27 and November 3, mostly performance degradation such as slower responses; the October 15 incident lasted more than four hours, with consumer services and the V4 Pro and V4.1 Flash APIs simultaneously restricted, and API users were later compensated with credits. From August to October, V4.1 Flash API availability was 99.46%, and V4 Pro was 99.74%. DeepSeek recently opened APIs for its full suite of applications, with time-of-use peak/off-peak pricing and separate concurrency limits on the API side. OpenRouter's weekly ranking shows that from October 19 to 25, V4.1-Flash-1017 ranked first with 33.6 trillion tokens in weekly calls, up 801% week-on-week. @@@ Sina Tech reports that on November 12, xFusion's ChiNext IPO was approved at the listing review meeting. It took only 174 days from acceptance on May 22 to approval, and its RMB8 billion fundraising makes it the largest ChiNext IPO project this year. The company's revenue rose from RMB25.092 billion in 2023 to RMB58.246 billion in 2025, and was RMB37.286 billion in the first half of 2026, up 20.78% year-on-year; AI server revenue share rose from 25.03% to 53.29%. However, its main business gross margin fell from 14.27% to 8.58% in 2025, with a net margin of 1.77%; in the first half of 2026, gross margin rebounded to 15.49%, and net profit was RMB2.108 billion, more than twice the full-year 2025 figure. Its 2025 R&D expense ratio was 2.02%, lower than Sugon, Unisplendour and Inspur. In the first half of 2026, net operating cash flow was -RMB15.685 billion; accounts receivable were RMB15.303 billion, up more than 170% from the end of 2025; inventory was RMB35.662 billion, up 68%. As of the end of July 2026, its backlog was RMB30.113 billion, of which AI servers accounted for 73.70%. @@@ All right, that's all for today's show! Thank you for spending time with us. I hope you're in a great mood tonight. If you like our show, you're very welcome to subscribe and share it with your friends—it really helps us. See you next time!

13 分钟 34 次播放 2天前

评价

还没有人评价。听过的话,留个分数和一句短评吧。

播放列表

自动播放下一个

播放列表还是空的

去找些喜欢的单集添加进来吧