Hi there~~ Good evening, everyone! Today is Saturday, October 31, 2026. Welcome to Goodnight Coffee. I’m your host, An’an! Tonight, let’s chat about the latest, piping-hot business and tech stories—relaxed, fun, and full of substance—to keep you company through a lovely evening. So come along with me! @@@ According to The Verge, Apple announced that it will launch Apple Intelligence in the Chinese market with iOS 27.3. Some of the first features will arrive first in the iOS 27.3 beta planned for release on November 4, while the full version of iOS 27.3 is scheduled for December. This rollout will not include the new Siri AI; instead, it will offer earlier Apple Intelligence features such as notification summaries, live translation, rewriting and proofreading tools, Visual Intelligence, and object removal in photos. Apple also said it has partnered with Alibaba to support its Tongyi Qianwen chatbot in some experiences, similar to how iOS integrated OpenAI’s ChatGPT in 2024. Apple Intelligence was approved in China in July; Reuters reported in August that Apple had also trained its own large language model specifically for China. @@@ According to Readhub, after the market opened on October 30 local time, Intel’s stock plunged sharply, falling more than 30% at one point during the session, with more than $137 billion in market value wiped out, on track for its largest single-day drop on record. The report said the plunge was mainly due to the company’s fourth-quarter guidance coming in below market expectations, along with warnings that weakness in the PC market and rising memory prices could persist into the first half of 2027. The earnings report showed that Intel’s third-quarter 2026 revenue rose 21% year over year to $16.51 billion, but its fourth-quarter revenue guidance was far below analysts’ expectations, and the company also said it would not achieve positive cash flow this year. In addition, an explosion at its factory in Israel is expected to cause about $600 million in revenue losses within two quarters; the company has invested heavily in AI chips and manufacturing but is constrained by capacity and supply chain limits, making it difficult to fully meet data center customer demand. The decline also affected several semiconductor companies, including AMD and Nvidia. @@@ According to Huxiu 24 Hours, Anthropic is sprinting toward an IPO, with the report describing the listing process as part of its race with OpenAI in valuation, revenue, and computing power. @@@ According to The Verge, Google’s Project Suncatcher satellite has generated an AI response from space for the first time. In a blog post, Google said the team successfully booted up the tensor processing unit, or TPU, aboard the satellite, enabling the onboard Gemma 4 model to compose a haiku about being in space in response to a prompt: “Blue Earth lies below, / data streams in silence flow, / starlight feeds my core.” Google said it also tested other space-themed limericks and telemetry prompts. The test is part of the Project Suncatcher experimental program, which could eventually send AI data centers into space; earlier this month, Google launched the satellite into orbit. Next, Google will conduct a series of tests to assess how the TPU performs under orbital thermal stress and radiation exposure. Google said that verifying whether TPUs can reliably run workloads in space will pave the way for its next milestone in 2027: two interconnected satellites. @@@ TechWeb reports that ChangXin Technology’s third-quarter 2026 financial results show quarterly revenue of RMB 118.714 billion, up 19.30% quarter over quarter; net profit was about RMB 88.983 billion, up about 19.38% quarter over quarter. For the first three quarters, revenue was RMB 269.024 billion, up 738.51% year over year; net profit attributable to parent was RMB 132.04 billion, turning from a loss to a profit year over year. The company said rising global demand for computing power and capacity allocation by major manufacturers have left DRAM in short supply; capacity construction and product mix optimization drove sharp increases in sales volume and selling prices; higher prices, combined with economies of scale and lean production, lowered unit costs and sharply increased gross profit. First-three-quarter R&D expenses were RMB 11.071 billion, up 73.03% year over year, accounting for 4.12% of revenue. The company disclosed that DRAM based on its sixth-generation platform has completed engineering sample validation; the report also noted that G6 continues to use DUV, with cell area reduced from 0.00128205 square micrometers in G5 to 0.000896 square micrometers, a reduction of about 30%. G5 was announced for official mass production on September 20, 2026, with the active-area half-pitch reduced to 11.95 nanometers. @@@ According to an exclusive report by LatePost, Volcano Engine has fully shifted its business focus to commercialization, and its MaaS, or Model-as-a-Service, business posted a gross margin of over 55% in the first half of the year. @@@ According to 36Kr, on October 28, Unitree Robotics, which has been listed for just over two months, announced that it has jointly established the “Shanyu Venture Program” with Sequoia China to enter early-stage investment in embodied intelligence. The program will focus on embodied AI, robotics applications, and core upstream and downstream technologies, providing funding, technical guidance, product validation, and industry resources through an “engineer + investor” approach. Its five principles are: “understand technology and industry, discover early, decide fast, value people, and stay out of the way.” The two sides have a seven-year partnership: in December 2019, Sequoia invested RMB 15 million in Unitree, and has since invested a cumulative total of about RMB 102 million. Unitree listed on the STAR Market in August this year; Sequoia holds a 7.11% stake. Based on the RMB 882 share price in early trading on its first day of listing, the paper return exceeds 220 times. On the same day, Unitree disclosed that in the first three quarters its revenue was RMB 1.817 billion, up 55.71% year over year; net profit attributable to parent was RMB 387 million, up 270.54% year over year; and R&D expenses were RMB 263 million, up 191.48% year over year. Wang Xingxing previously revealed that the “Yuchuang Program” received more than 200 projects in just over a month, and that in the next five years the company hopes to incubate more than 200 ecosystem companies. @@@ According to Sina Tech, 360 Group’s self-developed AI security attack-and-defense system, Tulongfeng, recently topped the international authoritative benchmark CyberGym with a 99.4% vulnerability reproduction success rate. Other contenders on the same leaderboard included models and agents from leading global AI companies such as OpenAI, Anthropic, and Google. CyberGym focuses on complex real-world challenges, covering a large number of mainstream open-source software programs and difficult historical vulnerabilities. It tests large models and security agents on autonomous reasoning, vulnerability reproduction, and PoC validation in massive source code, complex contexts, and multi-stage adversarial environments. Evaluation data shows that across all 1,507 tasks and three independent rounds of evaluation, Tulongfeng achieved success rates of 99.27%, 99.47%, and 99.40%, respectively, with an average success rate of 99.38%—equivalent to successfully tackling nearly 1,500 real vulnerabilities and setting a new record high on the leaderboard. To date, Tulongfeng has discovered more than 15,000 security vulnerabilities, covering operating systems, office software, AI applications, and other fields; the system has passed the China Academy of Information and Communications Technology’s capability assessment for vulnerability mining agents, and more than 500 enterprises have adopted it for specialized security testing, covering key industries such as finance, energy, government affairs, and manufacturing. @@@ TechWeb, citing Nikkei Chinese, reports that Toyota, Honda, Nissan, and Mazda produced a combined 1,020,579 vehicles in China from April to September this year, down 32% year on year, hitting the lowest level since 2008 and down 63% from the peak of about 2.8 million in the same period of 2020—less than 40% of that peak. Honda saw the largest drop, producing 131,716 vehicles, down 60% year on year and about 90% from its peak; its all-electric Ye GT was originally scheduled to launch in December 2025 but has been indefinitely delayed until after 2026. Nissan produced 163,192 vehicles, down 49% year on year and 80% from its peak; Toyota produced 663,079 vehicles, down 18% year on year and 23% from its peak; Mazda’s output rose 43% year on year, but its volume is limited. The main reason is shrinking demand for fuel-powered vehicles: from January to September, the new-car market fell 22% year on year, and new-car sales excluding new energy vehicles dropped by more than 30%. Honda plans to cut its annual fuel-vehicle production capacity in China from 1.49 million to 1 million. Mobility Global forecasts that the four automakers’ production in China in 2030 will fall another 47% from 2025 to about 1.6 million vehicles. @@@ Alright, that’s all for today’s show! Thank you for spending this time with us. I hope today’s little sharing has filled you with good vibes~ If you like our program, you’re very welcome to subscribe and share it with your friends—it really helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月31日，星期六，欢迎来到晚安咖啡，我是主播安安！今晚咱们一起聊聊新鲜热乎的商业科技动态，轻松又有料，陪你度过美好夜晚，快跟我出发吧！ @@@ 据The Verge报道，苹果宣布将随iOS 27.3在中国市场推出Apple Intelligence。首批部分功能将先行登陆计划于11月4日发布的iOS 27.3测试版，完整版iOS 27.3则计划于12月发布。此次上线不包括新的Siri AI，而是提供通知摘要、实时翻译、改写和校对工具、视觉智能以及照片物体移除等较早推出的Apple Intelligence功能。苹果还表示，已与阿里巴巴合作，在一些体验中支持其通义千问聊天机器人，类似2024年iOS整合OpenAI ChatGPT的方式。Apple Intelligence已于7月在中国获批；路透社8月报道称，苹果还专门为中国训练了自己的大语言模型。 @@@ 据Readhub报道，当地时间10月30日开盘后，英特尔股价大幅下挫，盘中跌幅一度超过30%，市值蒸发超过1370亿美元，有望创下有记录以来最大单日跌幅。报道称，暴跌主要因公司发布的第四季度业绩指引低于市场预期，并警告PC市场疲软、内存价格上涨等问题可能持续至2027年上半年。财报显示，英特尔2026年第三季度营收同比增长21%，达到165.1亿美元，但第四季度营收指引远低于分析师预期，公司还表示今年无法实现正现金流。此外，以色列工厂爆炸事故预计将在两个季度内造成约6亿美元营收损失；公司大力投资AI芯片和制造业务，却受产能和供应链限制，难以充分满足数据中心客户需求。此次下跌还波及AMD、英伟达等多家半导体企业。 @@@ 据虎嗅24小时消息，Anthropic正在冲刺IPO，报道将这一上市进程描述为其与OpenAI在估值、营收和算力上的竞赛。 @@@ 据The Verge报道，Google的Project Suncatcher卫星已首次从太空生成AI响应。Google在博客文章中称，团队成功启动卫星搭载的张量处理单元（TPU），使机载Gemma 4模型按照提示创作了一首关于身处太空的俳句：“蓝色地球在下方，数据流在静默中优雅涌动，星光滋养我的核心。”Google表示，还测试了其他太空主题打油诗和遥测提示。这项测试属于Project Suncatcher实验计划的一部分，该计划最终可能把AI数据中心送入太空；Google本月早些时候已将这颗卫星送入轨道。接下来，Google将开展一系列测试，评估TPU在轨道热应力和辐射暴露下的表现。Google称，验证TPU能否在太空可靠执行工作负载，将为2027年实现两颗互联卫星的下一个里程碑铺路。 @@@ TechWeb消息，长鑫科技2026年第三季度财报显示，当季营业收入1187.14亿元，环比增19.30%；净利润约889.83亿元，环比增约19.38%。前三季度营收2690.24亿元，同比增738.51%；归母净利润1320.4亿元，同比扭亏为盈。公司称，全球算力需求攀升及主要厂商产能调配令DRAM供不应求，产能建设和产品结构优化带动销量、售价大幅上涨；价格拉升叠加规模效应和精益生产使单位成本下降，毛利大幅上升。前三季度研发费用110.71亿元，同比增73.03%，占营收4.12%。公司披露，基于第六代平台的DRAM已完成工程样片验证；报道还提到，G6继续采用DUV，单元面积由G5的0.00128205平方微米降至0.000896平方微米，缩小约30%。G5已于2026年9月20日宣布正式量产，有源区半间距缩至11.95纳米。 @@@ 据晚点LatePost独家报道，火山引擎已将业务重心全面转向商业化，其MaaS（模型即服务）业务上半年毛利率超过55%。 @@@ 据36氪报道，10月28日，上市两个多月的宇树科技宣布与红杉中国共同成立“杉宇·创计划”，进军具身智能早期投资。该计划重点投资具身智能、机器人应用及上下游核心技术，以“工程师+投资人”方式提供资金、技术指导、产品验证及产业资源。五项原则为“懂技术和产业、发现早、决策快、看重人、不添乱”。 双方合作已有七年渊源：2019年12月红杉向宇树投入1500万元，此后累计投入约1.02亿元。宇树今年8月登陆科创板，红杉持股7.11%，按上市首日早盘882元股价测算，账面回报超220倍。 同日宇树披露，前三季度营收18.17亿元，同比增长55.71%；归母净利润3.87亿元，同比增长270.54%；研发费用2.63亿元，同比增长191.48%。王兴兴此前透露，“宇创计划”一个多月收到超200个项目，未来五年希望孵化200家以上生态企业。 @@@ 新浪科技消息，360集团自研的AI安全攻防系统图龙锋近日在国际权威基准测试CyberGym中以99.4%的漏洞复现成功率登顶榜单，同榜竞逐者包括OpenAI、Anthropic、Google等全球头部AI厂商的模型与智能体。CyberGym聚焦真实世界复杂挑战，覆盖大量主流开源软件与高难度历史漏洞，重点考验大模型和安全Agent在海量源码、复杂上下文及多阶段对抗环境下的自主推理、漏洞复现与PoC验证能力。评测数据显示，图龙锋在全部1507个任务中经过三轮独立评测，成功率分别为99.27%、99.47%和99.40%，平均成功率99.38%，相当于成功攻克近1500个真实漏洞，刷新该榜单最高纪录。截至目前，图龙锋累计发现安全漏洞超15000个，覆盖操作系统、办公软件及AI应用等领域；该系统已通过中国信通院漏洞挖掘智能体能力评估，超500家企业接入开展专项安全测试，覆盖金融、能源、政务、制造等重点行业。 @@@ TechWeb援引日经中文网报道，丰田、本田、日产、马自达四家车企今年4至9月在华合计产量102.0579万辆，同比降32%，创2008年以来新低，较2020年同期约280万辆峰值降63%、不足四成。本田跌幅最大，产量13.1716万辆，同比降60%，较峰值降约90%；纯电“烨GT”原定2025年12月上市，已无限期延至2026年后。日产16.3192万辆，同比降49%，较峰值降80%；丰田66.3079万辆，同比降18%，较峰值降23%；马自达同比增43%但体量有限。燃油车需求萎缩是主因，1至9月新车市场同比降22%，除新能源外新车销量降逾三成。本田计划将在华燃油车年产能从149万辆减至100万辆。Mobility Global预测，四家车企2030年在华产量比2025年再降47%至约160万辆。 @@@ 好啦，今天的节目到这里就结束啦！谢谢大家陪伴，愿今天的小分享给你满满好心情～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！

Hi~~ Good evening, everyone! Today is Friday, October 30, 2026, and I’m An’an, your host at Goodnight Coffee! Welcome to drop by right on time. Today we’ll chat about the latest and hottest business and tech news—easy to learn, fun to listen to, and the joy never closes. If you’re ready, let’s go! @@@ TechCrunch reports that Waymo and Toyota have announced an expansion of their autonomous driving partnership. Toyota will produce the 2027 model-year bZ4X EV at its Takaoka Plant in Toyota City, Aichi Prefecture, Japan, and will integrate Waymo’s autonomous driving hardware at the same plant. The two sides plan to begin initial operation, testing, and validation of this autonomous fleet in Japan in 2027, then deploy it to the U.K. and other markets where Waymo operates. Waymo ultimately intends to run a robotaxi service in Japan; the U.K. is one of the later right-hand-drive markets, and the company has already been testing Jaguar I-Paces in London for months and will launch a robotaxi service there once it obtains the necessary permits. This partnership comes about 18 months after the two sides signed an initial agreement in April 2025. The current arrangement is closer to a traditional vehicle supply agreement, helping Waymo meet regional demand and add new models. The two sides will also continue studying how to use Waymo’s autonomous driving technology in personally owned Toyota vehicles and use real-world operational insights to improve safety for other road users. The customized bZ4X offers generous rear legroom, automatic doors, a power tailgate, and ambient lighting. Toyota Executive Vice President Hiroki Nakajima said the partnership helps scale the technology and bring it to Japan and other markets through Waymo’s ride-hailing service. @@@ CNBC reports that Amazon and Intel fell 7% and 25% respectively in U.S. premarket trading on October 30 after their newly released third-quarter results disappointed investors, dragging Nasdaq 100 futures down 2%. Amazon’s third-quarter revenue and profit both missed expectations, and the company maintained its forecast of about $220 billion in capital expenditure this year. Quarterly revenue was $201.9 billion, up 12% year on year; operating income was $28.1 billion; and AWS revenue rose 39% to $46 billion. Operating cash flow over the past 12 months was $168.3 billion, but free cash flow was negative $42.4 billion, mainly due to increased AI-related investment. This was the second consecutive quarter of negative free cash flow, and the gap was larger than in the previous quarter. Intel gave weaker-than-expected guidance due to continued weakness in the PC market. Micron, Nvidia, and Arm fell in premarket trading, while European semiconductor stocks such as ASML tumbled. A day earlier, Alphabet closed down 13% as its negative free cash flow widened, marking its worst single-day performance ever. Raymond James strategist Matt Orton said this is not a liquidation moment, but investors should think about how positioning in the sector still looks crowded. He believes it is acceptable for these companies’ free cash flow to turn temporarily negative. @@@ CNBC reports that Hon Hai Technology Group, also known as Foxconn, Chairman Young Liu said Foxconn currently accounts for about 40% of the global AI server market, expects that to reach about 45% next year, and aims for 50% within a few years. The company’s capacity is fully booked for at least the next year and a half, and AI demand remains “quite strong,” mainly from AI model developers and cloud service providers, while adoption by governments and enterprises has not yet fully unfolded. Young Liu expects that in the next three to five years, AI infrastructure excluding edge AI devices will account for 60% to 65% of the company’s revenue. Foxconn plans to expand manufacturing in the U.S. and move further upstream in the supply chain. Because U.S. costs are higher, it will initially focus on high-margin products such as AI server racks and data center equipment. The company is also strengthening its layout from servers and racks to complete data centers and computing services. It plans to invest $3 billion to $5 billion a year in its AI business, and that could increase as vertical integration deepens. Young Liu said spending will be strictly controlled, because “overinvestment will kill the company.” If major customers cut AI spending, growth in other areas can offset it, but a broader industry slowdown would be more challenging. @@@ CNBC reports that Binance has suffered a hack with an estimated impact of $6.3 billion, and about $3 billion in crypto assets has been frozen. Binance has promised to bear all losses. TRM Labs attributed the attack with high confidence to the North Korea state-backed hacking group TraderTraitor and estimated that about $6.36 billion in digital assets was moved out of Binance wallets across 10 blockchains, most of it within about 20 minutes. Preparation for the attack began at least 36 hours in advance, and the hackers funded wallets through Tornado Cash. The FBI previously identified the group as responsible for the February 2025 Bybit theft of $1.5 billion. TRM Labs’ Ari Redbord said freezing can stop funds from moving, but recovery still has to go through legal procedures. Law enforcement agencies usually obtain a court order or seizure order and then work with stablecoin issuers to return assets. USDT and USDC can be frozen, burned, and reissued by Tether and Circle, while Bitcoin and Ether need to be tracked to cooperating exchanges and other platforms for interception. Chainalysis research director Eric Jardine said recovery requires coordination among law enforcement, exchanges, and token issuers. A technical freeze may happen quickly, but returning assets could take months or even years. @@@ The Verge reports that Reddit CEO Steve Huffman said on the company’s third-quarter 2026 earnings call that Google’s AI search modes “won’t drive traffic.” Reddit disclosed that in October, search traffic fell by more than 1 million referrals per week on average. Huffman attributed part of the Google traffic problem to AI Mode and AI Overviews changing the search experience. He said AI Mode does not drive traffic—for Reddit, and “as far as I know,” for others as well. Google used to be a traffic driver, but is no longer; AI Overviews and AI Mode at least push everyone further down the search page. Reddit also believes that machine-translated content in other languages ranks lower in search results, mainly affecting markets outside the U.S. Huffman said Reddit is still a popular search term on Google and consumer demand remains, but Google no longer sends traffic the way it used to, while Google’s crawling is as strong as ever. In response, Reddit is pushing the platform to become a direct, sticky daily destination. It has launched a “For You” recommendation feed, has been testing AI conversion of posts into short videos since August, and has announced that everyone will be required to log in, saying this benefits user experience, the advertising business, safety and compliance, and data protection. @@@ Sina Tech, citing foreign media, reports that DeepSeek’s second funding round reached 105.5 billion yuan, with nearly 30 investors participating, and the company’s valuation exceeded 615.5 billion yuan. In addition to Tencent and CATL increasing their investments, the latest investors include leading tech companies such as ChangXin Memory Technologies, SMIC, Meituan, vivo, and Xiaohongshu, covering chip manufacturing, consumer electronics, and internet services. People familiar with the matter said Tencent invested 10 billion yuan in the second round, the same as in the first round, for a total of 20 billion yuan across the two rounds. CATL invested 10 billion yuan this round, roughly double its first-round investment. DeepSeek had not publicly commented on the second funding round as of publication. @@@ Readhub reports that AgiBot has released its embodied model GO-3 Preview and took first place in three authoritative evaluations. The model enables large-scale learning across the full chain of pretraining, post-training, and post-deployment. Through designs including the embodied native architecture WRAM, it uses human first-person perspective data for pretraining and is capable of handling hundreds of tasks out of the box. In post-training, AgiBot uses Genie Evolver, a distributed hybrid-environment post-training system combining real machines, physics simulation, and world models, allowing robots to improve complex motion execution through virtual-real interaction. The model also uses the embodied Agent framework Genie Master to validate embodied recursive self-evolution in real business scenarios, enabling robots to continuously iterate their capabilities from actual work experience and extending embodied model learning across a fuller robot lifecycle. GO-3 Preview is planned to be open-sourced by the end of the year. @@@ Readhub reports that the 11th Pharmaceutical Innovation and Investment Conference was held in Nanjing. Fan Jie, President of Huawei’s Pharmaceutical Business Unit, said clearly at the conference that Huawei does not make drugs or pipelines. Instead, it will leverage its strengths in computing, communications, and artificial intelligence to provide computing power, storage, and other capabilities to pharmaceutical companies, research institutions, and AI drug discovery companies, serving as the technical foundation and ecosystem supporter for the AI pharmaceutical industry. Fan Jie said AI drug discovery is still in a stage of rapid development with high market expectations. Some technologies have already been deployed, but the full drug R&D chain has not yet been connected. In the next stage, the industry needs to push models into real R&D processes and form a closed loop with experimental data and R&D decisions. AI’s improvement of pharmaceutical R&D efficiency must ultimately be verified through specific pipelines and commercialization results. Some cooperative projects have already begun deploying hundred-GPU-scale computing clusters for drug R&D work. He also introduced Huawei’s three advantages in AI drug discovery: it can provide advanced computing and storage, has leading computing cluster capabilities, and is supported by a relevant cooperative ecosystem. Huawei has already begun cooperation with related companies and research institutions. @@@ Readhub reports that Baidu Group Vice President Wu Tian has resigned. @@@ TechWeb reports that BYD disclosed its third-quarter 2026 report on the evening of October 29. Third-quarter revenue was 221.5 billion yuan, up 14% year on year; net profit was 10 billion yuan, up 28% year on year; and net margin was 4.52%. Revenue for the first three quarters was 566.3 billion yuan, and R&D expenses were 36.3 billion yuan, higher than net profit of 22.3 billion yuan for the same period. The company said its second-generation Blade Battery and flash charging technology launched in March set the world’s fastest mass-production charging record, tackling slow charging and difficulty charging in low temperatures. As of October 7, it had 12,165 flash-charging stations, including 2,111 highway stations, and plans to build 20,000 by the end of the year. On intelligence, BYD said it has provided “dual backstops” for intelligent parking and urban navigation. As of September 30, the number of vehicles with assisted driving exceeded 3.95 million. Its self-developed 4nm Xuanji A3 intelligent driving chip will be launched this year. From January to September, overseas sales were 1.34 million vehicles, up 91% year on year. Combined sales of Fangchengbao, Denza, and Yangwang reached 410,000 vehicles, up 75%, accounting for more than 13%. The Denza Z made its global debut, the Fangcheng S&S GT and Da Han were introduced, and Yangwang will unveil Dingyao. @@@ Alright, that’s all for today’s show! Thank you to everyone who spent time with An’an. May your good mood stay online, and remember to share the warmth! If you like our show, you’re very welcome to subscribe and forward it to your friends—it helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月30日，星期五，我是晚安咖啡的主播安安！欢迎准时来坐坐，今天咱们一起聊聊新鲜热乎的商业科技动态，轻松涨知识，快乐不打烊，准备好就出发啦！ @@@ 据TechCrunch报道，Waymo与丰田宣布扩大自动驾驶合作。丰田将在日本爱知县丰田市高冈工厂生产2027款bZ4X电动车，并在同一工厂完成Waymo自动驾驶硬件集成。双方计划2027年在日本启动这支自动驾驶车队的初步运营、测试和验证，之后部署到英国及其他Waymo运营市场。Waymo最终打算在日本运营robotaxi服务；英国是后续右舵市场之一，该公司已在伦敦测试捷豹I-Pace数月，待取得必要许可后将在当地推出robotaxi服务。此次合作距双方2025年4月签署初步协议约18个月。目前安排更接近传统车辆供应协议，帮助Waymo满足区域需求并增加新车型。双方还将继续研究把Waymo自动驾驶技术用于个人拥有的丰田汽车，并利用其实际运营洞察改善其他道路使用者安全。定制版bZ4X提供宽敞后排腿部空间、自动门、电动尾门和氛围灯。丰田执行副总裁中岛裕树称，合作有助于将技术规模化并通过Waymo叫车服务推向日本及其他市场。 @@@ 据CNBC报道，亚马逊和英特尔10月30日美股盘前分别下跌7%和25%，此前公布的第三季度业绩令投资者失望，拖累纳斯达克100期货跌2%。亚马逊第三季度营收和利润均不及预期，维持今年约2200亿美元资本支出预测；当季营收2019亿美元，同比增12%，营业利润281亿美元，AWS收入增39%至460亿美元。过去12个月经营现金流1683亿美元，但自由现金流为负424亿美元，主要因AI相关投资增加；这是连续第二季度为负，且缺口大于上一季度。英特尔因PC市场持续疲软给出弱于预期的指引；美光、英伟达、Arm盘前走低，ASML等欧洲半导体股大跌。此前一天，Alphabet因负自由现金流扩大收跌13%，创史上最差单日表现。Raymond James策略师Matt Orton称，这不是清算时刻，但投资者应思考该领域仓位仍显拥挤；他认为这些公司自由现金流暂时转负可以接受。 @@@ 据CNBC报道，鸿海科技集团（富士康）董事长刘扬伟称，富士康目前占全球AI服务器市场约40%，预计明年约45%，几年内达到50%。公司产能至少未来一年半已经排满，AI需求仍“相当强劲”，主要来自AI模型开发商和云服务提供商，政府和企业的采用尚未充分展开。刘扬伟预计，未来三到五年，不含边缘AI设备的AI基础设施将占公司营收60%至65%。富士康计划扩大美国制造并向供应链上游延伸；因美国成本较高，初期聚焦AI服务器机架和数据中心设备等高毛利产品。公司还加强从服务器、机架到完整数据中心和计算服务的布局。其AI业务每年计划投入30亿至50亿美元，并可能随垂直整合程度提高而增加；刘扬伟称会严控支出，“过度投资会杀死公司”。若大客户削减AI支出，可由其他领域增长抵消，但行业整体放缓挑战更大。 @@@ 据CNBC报道，币安遭黑客攻击，估计影响达63亿美元，约30亿美元加密资产已被冻结。币安承诺承担全部损失。TRM Labs以高置信度将攻击归因于朝鲜国家支持的黑客组织TraderTraitor，并估计约63.6亿美元数字资产从币安钱包跨10条区块链转出，多数在约20分钟内完成；攻击准备至少提前36小时开始，黑客用Tornado Cash为钱包注资。FBI此前认定该组织对2025年2月Bybit 15亿美元被盗事件负责。TRM Labs的Ari Redbord表示，冻结能阻止资金转移，追回仍须走法律程序，执法机构通常取得法院命令或扣押令后与稳定币发行方合作返还资产；USDT和USDC可由Tether和Circle冻结、销毁并重新发行，比特币和以太币则需追踪至合作交易所等平台拦截。Chainalysis研究主管Eric Jardine称，返还需要执法、交易所和代币发行方协调，技术冻结可能很快，但返还可能耗时数月甚至数年。 @@@ 据The Verge报道，Reddit首席执行官史蒂夫·霍夫曼在2026年第三季度财报电话会上称，谷歌的AI搜索模式“不会带来流量”。Reddit披露，10月搜索流量平均每周减少超过100万次引荐；霍夫曼将部分谷歌流量问题归因于AI Mode和AI Overviews改变搜索体验。他说，AI Mode不驱动流量，对Reddit如此，“据我所知”对其他人也是如此；谷歌过去是流量驱动者，现在不是，AI Overviews和AI Mode至少把所有人往下推搜索页面。Reddit还认为，机器翻译成其他语言的内容在搜索结果中排名更低，主要影响美国以外市场。霍夫曼表示，Reddit仍是谷歌热门搜索词，消费者需求仍在，只是谷歌不再像以前那样导流，且谷歌抓取力度一如既往。为应对，Reddit正推动平台成为直接、有粘性的日常目的地，已推出“For You”推荐信息流，8月起测试用AI把帖子转成短视频，并宣布将强制所有人登录，称这有利于用户体验、广告业务、安全合规和数据保护。 @@@ 新浪科技援引外媒报道称，DeepSeek第二轮融资规模达到1055亿元，共有近30家投资方参与，公司估值突破6155亿元。除腾讯和宁德时代持续加码外，最新出资方还包括长鑫存储、中芯国际、美团、vivo、小红书等头部科技企业，覆盖芯片制造、消费电子和互联网服务等多个行业。知情人士透露，腾讯在第二轮投资中投入100亿元，与首轮出资额持平，两轮合计已投入200亿元；宁德时代本轮出资100亿元，较首轮投资翻了约一倍。对于第二轮融资相关事宜，DeepSeek方面截至发稿并未公开置评。 @@@ 据Readhub消息，智元发布具身模型GO-3 Preview，并拿下三项权威评测第一。该模型实现预训练、后训练、部署后全链路规模化学习，通过具身原生架构WRAM等设计，依托人类第一视角数据完成预训练，具备数百种任务开箱即用能力。在后训练环节，智元采用结合真机、物理仿真和世界模型的分布式混合环境后训练系统Genie Evolver，让机器人在虚实交互中提升复杂动作执行能力。该模型还通过具身Agent框架Genie Master，在真实业务场景中验证具身递归自进化，使机器人能够从实际工作经验中持续迭代能力，将具身模型学习延伸至机器人更完整的生命周期。GO-3 Preview计划于年底开源。 @@@ 据Readhub消息，第十一届医药创新与投资大会在南京召开。华为制药军团总裁樊杰在会上明确表示，华为不做药、也不做管线，将发挥在计算、通信和人工智能领域的优势，为药企、科研机构及AI制药企业提供算力、存储等能力，做AI制药产业的技术底座与生态支持者。樊杰说，当前AI制药仍处于快速发展、市场预期较高的阶段，部分技术已经落地，但完整药物研发链条尚未打通；下一阶段行业需推动模型融入真实研发流程，形成与实验数据、研发决策的闭环，AI对药企研发效率的提升最终要通过具体管线和商业化结果验证。目前部分合作项目已开始部署百卡规模计算集群，用于药物研发相关工作。他还介绍，华为参与AI制药有三大优势：可提供先进算力存力、具备领先的算力集群能力，并有相关合作生态配套。目前华为已和相关企业、科研机构开展合作。 @@@ 据Readhub消息，百度集团副总裁吴甜已离职。 @@@ 据TechWeb报道，比亚迪10月29日晚披露2026年三季报：三季度营收2215亿元、同比增14%，净利润100亿元、同比增28%，净利率4.52%；前三季度营收5663亿元，研发费用363亿元，高于同期净利润223亿元。公司称3月发布的第二代刀片电池及闪充技术创全球量产最快充电纪录，攻克充电慢、低温充电难；截至10月7日闪充站达12165座，含高速站2111座，计划年内建成20000座。智能化方面，比亚迪称已为智能泊车和城市领航提供“双兜底”，辅助驾驶车型保有量截至9月30日超395万辆，自研4nm璇玑A3智驾芯片今年发布。1至9月海外销量134万辆、同比增91%；方程豹、腾势、仰望合计销量41万辆、增75%，占比超13%。腾势Z全球首发，方程S&S GT、大汉推出，仰望将发布鼎曜。 @@@ 好啦，今天的节目到这里就结束啦！谢谢每一位陪伴安安的你，愿好心情一直在线，记得把温暖分享出去哦！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！

Hi~~ Good evening, everyone! Today is Thursday, October 29, 2026. Welcome to Goodnight Coffee. I’m your host, An’an! Today we’re going to talk about the freshest business and tech stories—plenty of substance and plenty of fun. Grab a seat, and let’s get started! @@@ According to The Verge, Google released its latest frontier model, Gemini 4 Carbon, on Wednesday, saying it has made significant gains in coding, agentic tasks and cybersecurity. Google says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra. It scored 88.2% on SEC-bench Pro, which measures the ability to find real vulnerabilities; 54.8% on ExploitGym, which measures exploitation capability; and 71.0% on the patch-repair benchmark CWE-bench. The model is another release after Gemini 4 Argon earlier this month, with a focus on strengthening vulnerability detection, validation and remediation. Google says it can push forward complex multi-step engineering tasks for several days in a row, take fewer shortcuts, resist prompt injection, and cut unsupported claims by half compared with Argon. Carbon is currently available only to selected cybersecurity defense companies, and will gradually expand to more trusted partners through the Fairwind Program. Google also says it is participating in the U.S. government’s voluntary pre-release testing process for models. Its entry price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens. @@@ According to cnBeta, Microsoft regulatory filings disclose that in the first quarter of fiscal 2027, ended September 30, 2026, the company earned $10 billion in revenue from its commercial partnership with OpenAI, including revenue-sharing arrangements, accounting for nearly 11% of total quarterly revenue of $92.3 billion. Despite intensifying competition in the AI market, the two remain closely tied commercially. Related revenue for fiscal 2026 was $24.1 billion; at the latest quarterly level, that would be about $40 billion over the next year. As of the end of September, OpenAI owed Microsoft $7.2 billion in accounts receivable, up from $6.0 billion at the end of June. This is historical transaction debt, not new revenue. Microsoft did not specify the respective shares from Azure cloud services and revenue sharing. The two reached an agreement in April this year under which OpenAI must continue providing revenue sharing to Microsoft through 2030, with a cap. In October 2025, Microsoft allowed OpenAI to work with other cloud providers, and OpenAI committed to purchase an additional $250 billion in Azure services. As of September this year, Microsoft held about 24% of OpenAI, down from 25% in June and about 27% after the 2025 restructuring. Microsoft did not explain the dilution. @@@ According to Readhub, Apple officially launched its first touchscreen MacBook Pro on October 29, bringing a touchscreen to the Mac product line for the first time, in 14-inch and 16-inch sizes. The new machine features a PROLED XDR display with tandem OLED technology. The touch sensors are integrated directly into the display panel to reduce latency, and the Dynamic Island design comes to MacBook for the first time. Peak brightness reaches up to 2,000 nits for both HDR and SDR, with a contrast ratio of 10,000,000:1. It supports a 120Hz variable refresh rate and covers the P3 and Adobe RGB color gamuts. The product is powered by M5 Pro and M5 Max chips, with up to 128GB of unified memory and 16TB of storage. It weighs about 1.38 kg and 1.82 kg, offers up to 24 hours of battery life, supports fast charging, and brings back the glowing Apple logo on the lid. Pre-orders start at 9 a.m. on October 30, and official sales begin on November 6. China pricing starts at RMB 22,999, with different chip, memory and storage versions reaching up to RMB 43,999. @@@ According to CNBC, Jio Platforms, India’s largest telecom operator and digital services company, became India’s second-most-valuable company after listing, behind parent Reliance Industries. That gives Mukesh Ambani the country’s two most valuable companies. The stock listed on Wednesday at an issue price of 1,119 rupees, opened at 1,358 rupees, up 21.36%, and closed on the Bombay Stock Exchange up 26.66% at 1,417.35 rupees, with a market cap of 13.05 trillion rupees, about $137.1 billion. On Thursday, its second trading day, it opened up another 3.3%. After the IPO, Reliance Industries holds more than 64%. Jio’s IPO raised 302.13 billion rupees, with subscription 22.19 times. Analysts say the proceeds will be used to repay legacy telecom debt, and with the 5G investment cycle slowing, it will significantly improve its free cash flow. However, much of the value had already been priced in after listing, and its valuation is close to AT&T and Verizon. Jio plans to expand into AI, cloud, connected devices and space connectivity over the next 10 years. Its satellite communications plan was previously criticized by Musk, who accused the Ambani family of delaying Starlink’s launch in India. Separately, Bloomberg reported that Reliance group’s market cap is 30 trillion rupees, exceeding Tata group’s 25 trillion rupees. @@@ According to cnBeta, Anthropic warned investors in its IPO filing that its outspoken advocacy for AI safety and public statements by executives could make it a target for government retaliation. The prospectus says the company has faced and may continue to face heightened scrutiny, and could be singled out by certain government agencies even when other similarly situated companies are not treated the same way. The tension stems from an earlier dispute this year over what safety guardrails should apply to the Pentagon’s use of its technology. More recently, the company has disagreed with government officials over whether its cutting-edge Mythos model is safe enough. CEO Dario Amodei supports a coordinated slowdown in AI development so safety teams can catch up with technological progress, which has drawn Trump’s displeasure. The prospectus also mentions proposed federal legislation that could force model shutdowns or even nationalize AI companies, as well as officials’ proposal to use the Defense Production Act to force companies to disclose information and cooperate with national security agencies. The company says trade controls could cause it to lose overseas customers, model misbehavior could bring legal consequences, and if the regulations it advocates are adopted, they would also increase compliance burdens and limit flexibility in product development and release. The Trump administration is both a regulator and a major customer; during the period covered by the IPO filing, government agency contract revenue accounted for about 1% of total revenue. The company also warns that advanced AI systems without safety guardrails could pose catastrophic or even existential risks to humanity. @@@ According to Readhub, on October 29, Tencent unveiled its AI office suite Buddy One and Tencent Drive at the 2026 Tencent Global Digital Ecosystem Summit. Buddy One uses WorkBuddy as its workspace, connecting Tencent Docs, Tencent Meeting, Tencent Drive and ima, with unified accounts and a shared credit pool. After the WorkBuddy PC app is upgraded to 5.9.0, linked accounts can use it, and memory sharing will be supported later. With accounts linked, WorkBuddy can integrate authorized documents, meetings and knowledge bases to generate answers with sources. General credits are shared across the four products and deducted in sync. VIP continuous monthly subscription is RMB 700, single month RMB 999; continuous annual is RMB 6,720, single year RMB 8,400. The package includes WorkBuddy Flagship, Tencent Meeting Pro, Tencent Docs Super Member Plus, 50,000 credits per month and 1TB of shared storage. Tencent Drive serves as the unified storage foundation for the suite; the personal beta version offers 5GB free, rising to 1TB with VIP. Enterprise edition storage per seat rises from 5GB to 50GB, with encryption and permission management. Since WorkBuddy Enterprise launched in March this year, it has covered 50% of China’s top 500 companies and 60% of the top 100. @@@ According to Huxiu 24-hour news, on October 29, Alibaba Cloud announced at the international Apsara Conference in Amsterdam that its first Netherlands region had officially opened. This is Alibaba Cloud’s fourth new overseas region this year, after Malaysia, France and Brazil. It has also expanded data centers in Japan, South Korea and Mexico, adding 12 availability zones overseas this year. The Netherlands region, located in Amsterdam, has two availability zones and offers cloud services including compute, storage, networking and databases. Enterprise AI services will also be launched gradually. In the European market, the Netherlands region is Alibaba Cloud’s fourth region after Frankfurt, London and Paris, and will further meet rapidly growing local demand for AI and cloud services. Li Feifei, CTO of Alibaba Cloud Intelligence Group and President of International Business, said the Netherlands is one of Europe’s most dynamic digital economies. As local enterprises’ AI demand continues to grow, Alibaba Cloud will provide high-performance cloud and AI services through local data centers, support localized data deployment, and help companies accelerate innovation and improve operational efficiency. In the future, Alibaba Cloud plans to open new regions in Turkey and Finland, and expand data centers in Malaysia, Germany, the UAE, France and Hong Kong. @@@ According to cnBeta, ChangXin Memory Technologies expects to officially start mass production of HBM3E high-bandwidth memory before the end of 2026, with mass production planned within the next two months. The company began risk production in August this year and has supplied samples to some customers, with the product now moving toward commercialization. Sources say its HBM3E overall yield is about 20% to 30%, notably lower than leading international manufacturers; front-end wafer manufacturing yield is about 30%, and back-end packaging can further improve the pass rate by about 70%. The biggest technical challenge is the TSV through-silicon-via stacking process: Samsung’s HBM2 uses about 5,000 TSVs per layer, SK hynix’s HBM3 uses more than 8,000, while ChangXin uses about 3,000 TSVs for its HBM3E design. Due to this limitation, the company is focusing on 8-layer stacking in the near term and is not rapidly advancing 12-layer products. ChangXin has provided two types of samples—standard and yield-optimized—to domestic AI chip companies, and customers are testing and validating them. Preliminary tests show its performance is closer to SK hynix’s previous-generation HBM3, with a gap versus advanced HBM3E. If validation goes smoothly, domestic AI accelerators using its HBM3E could enter the market as early as 2027. @@@ According to Sina Tech, at the AgiBot Innovation Day conference held on October 29, AgiBot announced that its general-purpose AI robots off the production line had exceeded 20,000 units, with its supply chain and batch delivery system taking shape. It disclosed a 43% global humanoid robot market share and revenue of nearly RMB 2 billion in the first half of 2026, making it the general AI robot company with the highest revenue in that period. The 20,000th unit was an Expedition A3 Ultra, which rolled off the line and was delivered to Chimelong Group at Hengqin Chimelong Spaceship Park on September 24. On the same day, the world’s first large-scale embodied intelligence theme park jointly created by the two opened. More than 300 AgiBot robots across its full product line are routinely on duty, handling visitor interaction, guided tours, performances and science education. The jointly built Embodied Intelligence Cultural Tourism Joint Research Institute was also unveiled, and will conduct research on scenario adaptation, algorithm iteration, content creation and commercial validation. On production pace, AgiBot started scaled mass production in August 2024, surpassed 10,000 units in March 2026, 15,000 in June, and reached 20,000 on September 24. It took about six months to go from 10,000 to 20,000. The report also notes that the 20,000 figure covers all robot product lines, and there is a statistical difference between off-line output and actual delivered and deployed volume. @@@ According to Sina Tech, Li Auto launched the 2026 Li i6 on October 29, with a national unified retail price of RMB 249,800 and a first-sale benefit price of RMB 239,800. Cumulative sales in its first year on the market are nearly 190,000 units, with a net promoter score of 80.2. The new model comes standard with zero-gravity seats for the driver and front passenger; the second row’s maximum recline angle increases from 125 to 140 degrees; the center armrest adds a 67W USB-C port; the fridge supports freezing down to -6°C; and a powered front trunk is standard. The new model comes standard with a 96 kWh self-developed Lianshan 5C battery. CLTC range for the rear-wheel-drive version rises from 720 km to 780 km, and for the all-wheel-drive version from 660 km to 710 km. It also comes standard with dual-chamber Magic Carpet air suspension, an all-aluminum front double-wishbone and rear five-link suspension, continuously variable damping dampers and the self-developed Mach M100 chip. Mach VLA 2.5 will be delivered via OTA 9.2. The cabin comes standard with a Qualcomm Snapdragon 8797 chip and a 29-inch 6K panoramic screen; CarPlay, HiCar and Carlink will be added. Vehicle torsional rigidity exceeds 52,000 N·m/°, high-strength steel and aluminum alloy account for 85%, and the front row features a three-chamber far-side airbag. @@@ All right, that’s it for today’s show! Thank you all for spending this time with us. I hope today’s stories leave you feeling relaxed and in a good mood. If you like our show, we’d really appreciate it if you subscribe and share it with your friends—it helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月29日，星期四，欢迎收听晚安咖啡，我是主播安安！今天咱们一起聊聊新鲜热乎的商业科技动态，干货和趣味都不缺席，快搬好小板凳，和我一起开启今天的分享吧！ @@@ 据The Verge报道，谷歌周三发布最新前沿模型Gemini 4 Carbon，称其在编程、智能体任务和网络安全方面取得显著提升。谷歌表示，Carbon在Artificial Analysis Intelligence Index v5.0上以59分排名第一，领先Claude Opus 5.5和GPT-6 Astra；在衡量发现真实漏洞能力的SEC-bench Pro上得分88.2%，在衡量漏洞利用能力的ExploitGym上得分54.8%，在补丁修复基准CWE-bench上得分71.0%。该模型是谷歌本月早些时候发布Gemini 4 Argon后的又一模型，重点强化漏洞检测、验证和修复，并称其能连续数天推进复杂多步工程任务，更少走捷径、抵御提示注入，且无依据声明较Argon减少一半。Carbon目前仅向部分网络安全防御公司开放，之后将通过Fairwind Program逐步扩大至更多可信合作伙伴；谷歌还表示正参与美国政府自愿性的模型发布前测试流程。其入门价与Argon相同，为每百万输入token 2美元、每百万输出token 10美元。 @@@ cnBeta消息，微软监管文件披露，截至2026年9月30日的2027财年第一季度，公司从与OpenAI的商业合作中获得100亿美元收入，含收入分成等安排，占当季总营收923亿美元的近11%，尽管双方在AI市场竞争加剧，商业联系仍紧密。2026财年相关收入为241亿美元；按最新季度水平推算，未来一年约400亿美元。截至9月底，OpenAI对微软应收账款72亿美元，高于6月底的60亿美元，属历史交易欠款，非新增收入。微软未说明其中Azure云服务与收入分成的各自占比。双方今年4月达成协议，OpenAI须在2030年前继续向微软提供收入分成，并设有上限；2025年10月微软允许OpenAI与其他云服务商合作，后者承诺额外采购2500亿美元Azure服务。截至今年9月，微软持有OpenAI约24%股份，低于6月的25%和2025年重组后的约27%，微软未说明稀释原因。 @@@ 据Readhub报道，苹果10月29日正式发布首款触屏MacBook Pro，首次为Mac产品线引入触控屏，提供14英寸、16英寸两种尺寸。新机搭载PROLED XDR显示屏，采用双层串联OLED技术，触控传感器直接集成在显示面板内部以降低操作延迟，并首次把灵动岛设计带入MacBook；屏幕HDR和SDR峰值亮度最高达2000尼特，对比度10000000:1，支持120Hz可变刷新率并覆盖P3和Adobe RGB色域。产品搭载M5 Pro、M5 Max芯片，最高可选128GB统一内存和16TB存储，重量约1.38千克和1.82千克，最长续航24小时，支持快充，并重新在A面带回发光苹果Logo。该机将于10月30日上午9点开启预购，11月6日正式发售，国行售价22999元起，不同芯片、内存和存储版本最高43999元。 @@@ 据CNBC报道，印度最大电信运营商和数字服务公司Jio Platforms上市后成为印度市值第二高公司，仅次于母公司信实工业，穆克什·安巴尼由此拥有印度市值最高的两家公司。该股周三以1119卢比发行价上市，开盘报1358卢比，涨21.36%，孟买证交所收盘涨26.66%至1417.35卢比，市值13.05万亿卢比（约1371亿美元）；周四上市次日开盘再涨3.3%。IPO后信实工业持股超64%；Jio此次IPO募资3021.3亿卢比，认购22.19倍。分析师称，募资用于偿还电信历史债务、5G投资周期放缓，将显著改善其自由现金流，但上市后大部分价值已定价，估值接近AT&T和Verizon。Jio计划未来10年拓展AI、云、联网设备及太空连接，其卫星通信计划此前遭马斯克批评，马斯克指责安巴尼家族导致Starlink在印度推迟上线。另据彭博社，信实集团市值30万亿卢比，超过塔塔集团的25万亿卢比。 @@@ cnBeta消息，Anthropic在IPO文件中警告投资者，其直言不讳倡导AI安全及高管公开表态，可能使其成为政府报复目标。招股书称，公司已受到并可能继续受到更严格审查，可能遭特定政府机构针对，即便其他处境类似公司未受同样对待。矛盾源于今年早些时候围绕五角大楼使用其技术应设置何种安全护栏的争议；近期又因尖端Mythos模型是否足够安全与政府官员分歧。CEO Dario Amodei支持协调放慢AI开发速度，让安全团队追赶技术进步，招致特朗普不满。招股书还提到强制关闭模型乃至AI企业国有化等拟议联邦立法，以及官员提出用《国防生产法》迫使企业披露信息并与国家安全机构合作。公司称贸易管制或令其失去海外客户，模型不当行为可能带来法律后果，若所倡导监管被采纳也将增加合规负担并限制产品开发和发布灵活性。特朗普政府既是监管者也是重要客户；IPO文件涵盖时期内，政府机构合同收入约占总营收1%。公司还警告先进AI系统缺乏安全护栏可能带来灾难性甚至关乎人类存亡风险。 @@@ 据Readhub报道，10月29日，腾讯在2026腾讯全球数字生态大会上发布AI办公套件Buddy One及腾讯网盘。Buddy One以WorkBuddy为工作台，打通腾讯文档、腾讯会议、腾讯网盘和ima，统一账号与积分池；WorkBuddy PC端升级至5.9.0后关联账号可用，后续支持记忆共享。账号打通后，WorkBuddy可整合已授权文档、会议和知识库，生成附来源回答；通用积分在四款产品共享并同步扣减。VIP连续包月700元、单月999元，连续包年6720元、单年8400元；套餐含WorkBuddy旗舰版、腾讯会议专业版、腾讯文档超级会员Plus、每月5万积分和1TB共享存储。腾讯网盘作为套件统一存储底座，个人内测版免费5GB、VIP后升至1TB；企业版每席存储从5GB升至50GB，支持加密和权限管理。WorkBuddy企业版自今年3月上线以来，覆盖中国企业500强50%、企业100强60%。 @@@ 虎嗅24小时消息，10月29日，阿里云在荷兰阿姆斯特丹举行的国际云栖大会上宣布，荷兰首个地域节点正式开服。这是阿里云今年新开的第四个海外地域节点，此前已先后在马来西亚、法国、巴西开服，并扩建日本、韩国、墨西哥等地数据中心，年内海外累计新增12个可用区。此次开服的荷兰节点位于阿姆斯特丹，设有两个可用区，提供计算、存储、网络、数据库等云服务，企业级AI服务也将陆续上线。在欧洲市场，荷兰节点是阿里云继法兰克福、伦敦和巴黎之后的第四个地域节点，将进一步满足当地迅速增长的AI和云需求。阿里云智能集团CTO、国际业务总裁李飞飞表示，荷兰是欧洲最具活力的数字经济体之一，随着当地企业AI需求持续增长，阿里云将通过本地数据中心提供高性能的云与AI服务，支持数据本地化部署，帮助企业加快创新、提升运营效率。未来，阿里云计划在土耳其、芬兰新设地域节点，并扩建马来西亚、德国、阿联酋、法国和中国香港的数据中心。 @@@ cnBeta消息，长鑫存储预计2026年底前正式启动HBM3E高带宽内存量产，计划未来两个月内推进量产。公司今年8月已开始风险试产并向部分客户供样，产品正转向商业化。消息人士称，其HBM3E整体良率约20%至30%，明显低于国际领先厂商；前段晶圆制造良率约30%，后段封装可再提升约70%合格率。最大技术挑战是TSV硅通孔堆叠工艺：三星HBM2单层约5000个TSV，SK海力士HBM3超过8000个，长鑫以约3000个TSV实现HBM3E设计。受此限制，公司短期专注8层堆叠，暂不快速推进12层产品。长鑫已向国内AI芯片企业提供标准版及良率优化版两种样品，正由客户测试验证；初步测试显示其性能更接近SK海力士上一代HBM3，与国际先进HBM3E有差距。若验证顺利，采用其HBM3E的国产AI加速器最快2027年进入市场。 @@@ 新浪科技消息，在10月29日举行的智元AGIBOT创新日大会上，智元公布：通用AI机器人下线量突破2万台，供应链和批量交付体系成型；其披露的全球人形机器人市占率达43%，2026年上半年营收接近20亿元，为期内收入最高的通用AI机器人公司。第2万台为远征A3 Ultra，已于9月24日在横琴长隆飞船乐园下线并交付长隆集团。同日，双方打造的全球首个大规模具身智能主题乐园启幕，超过300台智元全系机器人常态化上岗，承担游客互动、导览、演艺和科普研学等服务；共建的“具身智能文旅联合研究院”揭牌，将围绕场景适配、算法迭代、内容创作和商用验证开展研究。量产节奏上，智元2024年8月启动规模化量产，2026年3月突破1万台、6月突破1.5万台、9月24日达2万台，从1万台到2万台约用6个月。报道同时指出，2万台涵盖全线机器人产品，下线产量与实际交付部署量存在统计口径差异。 @@@ 新浪科技消息，理想汽车10月29日发布2026款理想i6，全国统一零售价24.98万元，首销权益价23.98万元；上市一年累计销量近19万辆，用户净推荐值80.2分。新车主副驾标配零重力座椅，二排最大躺角由125度增至140度，中央扶手新增67瓦Type-C，冰箱支持零下6摄氏度冷冻，电动前备箱标配。新款标配96千瓦时自研连山5C电池，后驱版CLTC续航由720公里升至780公里，四驱版由660公里升至710公里；标配双腔魔毯空气悬架、全铝前双叉臂后五连杆悬架、连续可变阻尼减振器和自研马赫M100芯片。马赫VLA 2.5将随OTA 9.2推送。座舱标配高通骁龙8797芯片和29英寸6K全景屏，CarPlay、HiCar、Carlink将上线。整车扭转刚度超52000N·m/°，高强钢与铝合金占比85%，前排配三腔远端安全气囊。 @@@ 好啦，今天的节目到这里就结束啦！谢谢大家的陪伴，愿今天的内容给你带来轻松好心情。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！

Hi everyone, and a very good evening to you! It’s Wednesday, October 28, 2026, and welcome to Goodnight Coffee. I’m your host, An’an. Today, let’s talk about some fresh, juicy business and tech updates—lighthearted roundups you can enjoy and absorb easily. Ready? Let’s go! @@@ According to CNBC, Alphabet announced that Oracle will offer Google’s TPU chips to customers, marking the first time customers can use TPUs outside Google’s own infrastructure. Google has been developing TPUs since 2013 and previously focused on a one-stop full stack of chips, models, and cloud services. This time, it is using external partners to expand its AI chip business and challenge Nvidia; in the second quarter, it had already opened TPUs to external data centers. Analysts expect Google Cloud’s third-quarter revenue to grow about 90% year over year, after an 82% increase in the second quarter. Second-quarter cloud revenue was $25 billion, below AWS’s $42 billion and Azure’s $29 billion, but it was the fastest-growing. TD Cowen expects TPU sales of $9.4 billion in the second half of this year and $62.5 billion in 2027; Wells Fargo expects $9 billion in the third quarter and $99 billion next year. Google’s CFO said only a small portion of revenue from existing TPU agreements will be recognized this year, with the vast majority recognized in 2027. @@@ According to TechCrunch, Kodiak AI has signed its first binding commercial agreement with California-based midsize freight company DTL Transport. Starting in 2028, Kodiak will deliver 75 autonomous heavy trucks for operation on public highways. The contract was announced on Wednesday. DTL plans to first use these vehicles on the Dallas–El Paso, Texas route and on Interstate 20 between Dallas and Atlanta; Kodiak has already tested and operated on both routes. This is Kodiak’s first highway “Driver-as-a-Service” customer. The customer owns and operates the vehicles, while Kodiak licenses the technology and charges by the mile. In a California pilot last month, Kodiak autonomously hauled fresh produce for DTL on an approximately 225-mile stretch from Fresno to Los Angeles along Interstate 5 and State Route 99, with a safety driver in the vehicle. Kodiak plans to remove the safety driver in December and haul goods for IKEA along Interstate 45 from Dallas to Houston, launching a driverless highway service. The company went public in September 2025, and its stock is down 87% from its opening price. CEO Don Burnette said midsize carriers execute quickly and are easier to integrate fleets—this is a turning point for the autonomous driving industry. Kodiak also has TaaS customers operating on three routes in Texas, still in the pilot stage. @@@ According to CNBC, SK Hynix shares closed up 10.89% on Wednesday, outperforming the Kospi’s 6.63% gain, after the company reported record third-quarter results and announced a shareholder return plan worth about 100 trillion won (more than $25 billion). Third-quarter revenue rose 327.9% year over year and 31.9% quarter over quarter; operating profit rose 632.1% year over year and 37.6% quarter over quarter; sales exceeded 100 trillion won for the first time. The return plan includes at least 30 trillion won in cash dividends and 60 trillion won in buybacks and cancellations; the third-quarter dividend is 14 trillion won, including a special dividend from investment gains related to Kioxia. A previous 40 trillion won buyback has been completed and will soon be cancelled, and the company will buy back another 20 trillion won, with half cancelled this year; the cumulative free-cash-flow shareholder return target for 2025 to 2027 has been raised to more than 50%. On the earnings call, the company’s CEO said that due to financial market volatility and the correction in semiconductor stocks, the company has stopped evaluating pre-IPO financing and IPO options for its U.S. subsidiary Solidigm. The company expects the supply-demand imbalance in the memory market to continue. Susquehanna analyst maintained a Neutral rating, prefers Micron and Samsung Electronics, and is more positive on SK Hynix’s U.S. ADR. @@@ According to CNBC, India’s largest telecom operator, Jio Platforms, rose 21.5% on its first day of trading on Wednesday. The company issued 270 million shares at 1,119 rupees apiece, raising 302 billion rupees (about $3.2 billion), making it India’s largest IPO to date. According to National Stock Exchange data, its market capitalization reached 12.5 trillion rupees (about $131 billion), making it India’s second most valuable listed company, after parent Reliance Industries. The IPO was oversubscribed more than 22 times, with the qualified institutional buyer portion 70.55 times covered and the retail portion 2.72 times. Shares opened at 1,358 rupees and last traded at 1,360 rupees. Nomura initiated coverage with a Buy rating and a target price of 1,340 rupees. Reliance Industries holds more than 66%, Meta affiliate Jaadhu Holdings holds nearly 10%, and Google International holds 7.7%. Analyst Deven Choksey said inclusion in key benchmark indices will bring demand and prevent oversupply; some investors who bought in 2021 may exit after about six months, when Jio is expected to be included in the MSCI and Nifty 50 indices. @@@ According to The Verge, the rapid expansion of data centers in Texas has sparked bipartisan opposition and become a focal point in the midterm elections. Bledsoe, president of the Texas NAACP, said multiple projects around Austin are facing resistance: Taylor is proposing to build a data center on parkland; the mayor and city council members in Temple face recall attempts over related development; and San Marcos became the first city in the state to ban data center construction. The Texas NAACP has joined dozens of groups across the political spectrum to form the “We the People Texas” coalition, seeking to block or at least delay construction, and estimates it represents as many as 1 million voters. An August poll showed that 57% of Texans oppose new data centers, including 43% of Republicans, 75% of Democrats, and 54% of independents. Governor Abbott once called Texas an “AI development hub” and is now running for re-election; his Democratic opponent, Hinojosa, is seen as having a chance at an upset. In the Senate race, Republican Attorney General Paxton and Democrat Talarico are neck and neck. Candidates from both parties are trying to distance themselves from the data center industry, but voters aren’t buying it. @@@ According to Sina Tech, Unitree Technology released its third-quarter 2026 report, its first financial report since going public. Third-quarter revenue was 665 million yuan, up 69.93% year over year; net profit attributable to shareholders was 113 million yuan, down 17.37% year over year; non-GAAP net profit attributable to shareholders was 105 million yuan, down 17.42%; and basic earnings per share were 0.30 yuan. For the first three quarters, cumulative revenue was 1.817 billion yuan, up 55.71%; total profit was 413 million yuan, up 143.05%; and net profit attributable to shareholders was 387 million yuan, up 270.54%. The company said the increase in net profit for the first three quarters was mainly driven by revenue growth and higher gross profit, as well as the prior-year one-time recognition of a relatively large share-based payment expense. Third-quarter R&D investment was 127 million yuan, up 249.58% year over year and accounting for 19.1% of revenue for the period; first-three-quarter R&D investment was 263 million yuan, up 191.48% year over year and accounting for 14.47% of revenue. The increase was mainly due to greater investment in embodied intelligence, especially significantly higher R&D and computing power investment in embodied large models, as well as expanding the R&D team and developing new products. As of the end of the reporting period, total assets were 10.156 billion yuan, up 216.52% from the end of last year; owners’ equity attributable to shareholders of the listed company was 8.910 billion yuan, up 242.10%, mainly due to the completion of its initial public offering and the receipt of proceeds. Total ordinary shareholders stood at 48,007, with Wang Xingxing holding 21.44% as the largest shareholder. @@@ According to TechWeb, CATL recently confirmed on an investor interaction platform that the company continues to invest in all-solid-state battery R&D, that the related technology is industry-leading, and that small-batch production is expected in 2027. All-solid-state batteries use a solid electrolyte to replace the electrolyte and separator, paired with a lithium-metal anode. Their theoretical energy density can exceed 400 Wh/kg, greatly improving range, while making thermal runaway and fires less likely and raising the ceiling for fast charging. But the biggest current challenge is solid-solid interface contact: materials expand and contract during charging and discharging, which can cause the electrolyte and electrode to delaminate and directly affect cycle life and stability. Materials, processes, and production-line equipment all need to be redeveloped. CATL Chairman Zeng Yuqun previously said solid-state batteries still have a long way to go before mature mass production; Chief Manufacturing Officer Ni Jun also said large-scale deployment will take at least several years. The industry generally expects that samples produced in 2027 will likely be supplied first to premium models and special equipment for validation, with limited output and high cost, making it hard for them to enter mainstream family cars in the short term. CATL is also advancing condensed-matter batteries, Shenxing superfast charging, and sodium-ion batteries; liquid batteries will remain the market mainstay for the next few years. @@@ According to Sina Tech, on October 28, Huawei held its Double 11 Renewal Season · All-Scenario New Product Launch. He Gang focused on introducing the Mate 90 series’ ten major innovations and announced Double 11 exclusive phone purchase benefits. The ten innovations cover materials, chips, imaging, display, and communications. According to Huawei, Kirin 9050 uses a logic-folding τ chip architecture, integrating 238 million transistors per square millimeter; the Mate 90 Pro Max combines Kirin 9050 with HarmonyOS 7 for a 31% overall performance boost, while the Mate 90 and Pro use Kirin 9030 for a boost of more than 20%. The entire series features the third-generation Red Maple original-color camera, with color accuracy improved by 37% and a standard F1.4–F4.0 continuous variable aperture. The second-generation Linglong display reaches 12,000 nits peak brightness at 1% APL. The Mate 90 Pro Max debuts four-card triple-standby, supporting two physical SIM cards and two eSIMs, with three cards on standby simultaneously; a software upgrade planned before the New Year will bring support to the entire series. The Mate 90 Pro offers up to 41 hours of long-video playback, with battery life improved by 20%. @@@ According to cnBeta, the State Key Laboratory of Optical Communication Technologies and Networks at China Information and Communication Technologies Group, together with Pengcheng Laboratory and FiberHome Fujikura Optical Fiber Technology Co., Ltd., completed a multi-core fiber real-time bidirectional transmission trial with a total capacity of 3.2 Pb/s, setting a record for single-fiber bidirectional real-time transmission and improving by about 28% over the 2.5 Pb/s achieved in March this year. The trial used a 10.3 km, 24-core fiber, using space-division multiplexing and the S, C, and L bands to build thousands of parallel channels. It employed commercial coherent optical modules and high-order probabilistic shaping modulation, and optimized core arrangement to reduce crosstalk; self-developed optical modules achieved 400 Gb/s per wavelength in the S band and 1.2 Tb/s per wavelength in the C and L bands. The team said traditional commercial single-core G.652 fiber has a capacity upper limit of about 100 Tb/s in the C and L bands; this total capacity is about 32 times that. The test link was dedicated multi-core fiber and does not represent existing home broadband levels. The solution is expected to be used for data center interconnection, backbone networks, and 6G bearer network capacity expansion, but it is still in the trial validation stage. The result was selected for OFC 2026 and rated as a high-score paper; the team was invited to submit a paper to the Journal of Lightwave Technology, focusing on validating a system closer to commercial architecture under real-time transmission conditions. @@@ According to TechWeb, Baidu’s Intelligent Cloud Business Group, or ACG, recently made organizational adjustments around business development. Related businesses under ACG’s Intelligent Application Division were transferred according to their business nature to the Infrastructure Division, Agent Division, and Service Delivery and Quality Engineering Center. The Service Delivery and Quality Engineering Center was renamed the Industry Application and Technology Service Center; the department head remains unchanged and continues to report to Shen Dou. Meanwhile, Baidu Vice President Ruan Yu rotated to serve as head of Xiaodu Technology and continues to report to Shen Dou. The adjustment is part of Baidu’s efforts to advance organizational evolution for the AI era and deepen coordinated operations under “cloud-intelligence integration.” In the future, Baidu ACG’s product and R&D lineup will form a two-tier structure of “infrastructure and agents,” focusing on the two core battlefields of AI cloud and agents, and using FDE to accelerate the large-scale and efficient implementation of AI capabilities in enterprise and consumer scenarios. @@@ Alright, that’s all for today’s show! Thank you for listening and spending time with us. I hope today’s content puts you in a good mood. If you enjoy our show, you’re very welcome to subscribe and share it with your friends—it helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月28日，星期三，欢迎来到《晚安咖啡》，我是主播安安～今天咱们一起聊聊新鲜有料的商业科技动态，轻松盘点、快乐吸收，准备好就出发吧！ @@@ 据CNBC报道，Alphabet宣布，Oracle将向客户提供Google的TPU芯片，首次允许客户在Google自有基础设施之外使用TPU。Google自2013年起研发TPU，过去主推芯片、模型和云服务一站式全栈，此次借外部伙伴扩大AI芯片业务，挑战英伟达；第二季度已向外部数据中心开放TPU。分析师预计Google Cloud第三季度营收同比增长约90%，第二季度增幅为82%；第二季度云收入250亿美元，低于AWS的420亿美元和Azure的290亿美元，但增速最快。TD Cowen预计今年下半年TPU销售额94亿美元、2027年625亿美元；Wells Fargo预计第三季度90亿美元、明年990亿美元。Google CFO称，现有TPU协议仅小部分收入今年确认，绝大多数在2027年确认。 @@@ 据TechCrunch报道，Kodiak AI与加州中型货运公司DTL Transport签署首份具有约束力的商业协议，将从2028年起交付75辆自动驾驶重卡，在公共高速公路上运营。合同周三宣布。DTL计划先在达拉斯至得州埃尔帕索、以及达拉斯至亚特兰大之间的20号州际公路使用这些车辆；Kodiak已在这两条路线测试和运营。这是Kodiak首个高速公路“司机即服务”客户，车辆由客户拥有运营，Kodiak授权技术并按英里收费。上月加州试点中，Kodiak在5号州际公路和99号州道弗雷斯诺至洛杉矶约225英里路段，为DTL自动驾驶运输生鲜，车内配有安全员。Kodiak计划12月撤除安全员，在达拉斯至休斯顿的45号州际公路为宜家运货，启动无人驾驶高速服务。公司2025年9月上市，股价较开盘价跌87%。CEO Don Burnette称，中型承运商执行快、更容易整合车队，这是自动驾驶行业的转折点。Kodiak另有TaaS客户在得州三条路线运营，仍处试点阶段。 @@@ 据CNBC报道，SK海力士周三股价收涨10.89%，跑赢收涨6.63%的Kospi，因公司公布创纪录第三季度业绩并宣布约100万亿韩元（超250亿美元）股东回报计划。当季营收同比增327.9%、环比增31.9%，营业利润同比增632.1%、环比增37.6%，销售额首次突破100万亿韩元。回报计划包括至少30万亿韩元现金股息、60万亿韩元回购及注销；第三季度股息为14万亿韩元，含Kioxia投资收益特别股息。此前40万亿韩元回购已完成并将很快注销，公司还将回购20万亿韩元，其中一半今年注销；2025至2027年累计自由现金流股东回报目标提高至50%以上。公司首席执行官在财报电话会上称，因金融市场波动和半导体股回调，停止评估美国子公司Solidigm的Pre-IPO融资和IPO选项。公司预计存储市场供需失衡仍将持续。Susquehanna分析师维持中性评级，更偏好美光、三星电子，并更看好SK海力士美国ADR。 @@@ 据CNBC报道，印度最大电信运营商Jio Platforms周三上市首日股价上涨21.5%。公司发行2.7亿股，发行价每股1119卢比，募资3020亿卢比（约32亿美元），为印度迄今最大IPO。按印度国家证券交易所数据，其市值达12.5万亿卢比（约1310亿美元），成为印度市值第二高上市公司，仅次于母公司信实工业。此次IPO超额认购逾22倍，其中合格机构部分70.55倍、散户部分2.72倍。首日开盘价1358卢比，最后交易价1360卢比。野村给予买入评级，目标价1340卢比。信实工业持股超66%，Meta关联公司Jaadhu Holdings持股近10%，Google International持股7.7%。分析师德文·乔克西称，纳入关键基准指数将带来需求并防止供过于求；部分2021年入股投资者可能在约六个月后退出，届时Jio预计纳入MSCI和Nifty 50指数。 @@@ 据The Verge报道，得州数据中心快速扩张引发跨党派反对，并成为中期选举焦点。得州NAACP主席布莱索称，奥斯汀周边多个项目遭抵制：泰勒市拟在公园地块建数据中心，坦普尔市长和市议员因相关开发面临罢免尝试，圣马科斯成为全州首个禁建数据中心的城市。得州NAACP已与数十个跨政治光谱团体组成“We the People Texas”联盟，寻求阻止或至少推迟建设，估计代表多达100万选民。8月民调显示，57%得州人反对新建数据中心，包括43%共和党人、75%民主党人和54%独立人士。州长阿博特曾称得州为“AI开发中心”，现正竞选连任，民主党对手希诺霍萨被视为有望爆冷；参议院选战中，共和党总检察长帕克斯顿与民主党人塔拉里科势均力敌。两党候选人均试图与数据中心产业保持距离，但选民不买账。 @@@ 新浪科技消息，宇树科技发布2026年第三季度报告，这是其上市后首份财报。第三季度营收6.65亿元，同比增长69.93%；归母净利润1.13亿元，同比下降17.37%；扣非归母净利润1.05亿元，同比下降17.42%；基本每股收益0.30元。前三季度累计营收18.17亿元，同比增长55.71%；利润总额4.13亿元，同比增长143.05%；归母净利润3.87亿元，同比增长270.54%。公司称，前三季度净利润增长主要因收入增长带动毛利额增长，以及上年一次性确认较大金额股份支付费用。第三季度研发投入1.27亿元，同比增长249.58%，占当期营收19.1%；前三季度研发投入2.63亿元，同比增长191.48%，占营收14.47%，增长主要因加大具身智能领域投入，尤其具身大模型相关模型研发与算力投入大幅增长，以及扩充研发团队、开展新产品研发。截至报告期末，总资产101.56亿元，较上年度末增长216.52%；归属于上市公司股东的所有者权益89.10亿元，增长242.10%，主要因完成首次公开发行并上市、募集资金到账。普通股股东总数48,007户，王兴兴持股21.44%，为第一大股东。 @@@ 据TechWeb报道，宁德时代近日在投资者互动平台确认，公司持续投入全固态电池研发，相关技术处于行业领先水平，2027年有望实现小批量生产。全固态电池用固态电解质替代电解液和隔膜，并搭配锂金属负极，理论能量密度可突破400Wh/kg，续航可大幅提升，同时更不易发生热失控起火，快充上限也更高。但当前最大难点是固固界面接触：电池充放电时材料会膨胀收缩，容易造成电解质与电极脱层，直接影响循环寿命和稳定性，材料、工艺和产线装备均需重新开发。宁德时代董事长曾毓群此前表示，固态电池距成熟量产仍有很长攻关路程；首席制造官倪军也称，规模化落地至少还要数年。业内普遍预计，2027年产出的样品大概率优先供给高端车型、特种装备做验证，产量有限且成本较高，短期内难以进入主流家用车。宁德时代还在同步推进凝聚态电池、神行超充和钠离子电池等技术线，液态电池仍将是未来几年市场主力。 @@@ 新浪科技消息，10月28日华为举行双十一焕新季·全场景新品发布会，何刚集中介绍Mate 90系列十大黑科技并发布双十一专享购机权益。十大黑科技覆盖材料、芯片、影像、显示和通信。据华为介绍，麒麟9050采用逻辑折叠τ芯片架构，每平方毫米集成2.38亿个晶体管；Mate 90 Pro Max搭麒麟9050与HarmonyOS 7，整机性能提升31%，Mate 90和Pro搭麒麟9030提升20%以上。全系搭载第三代红枫原色摄像头，色彩准确度提升37%，标配F1.4—F4.0连续可变光圈。第二代灵珑屏1% APL峰值亮度达12000尼特。Mate 90 Pro Max首发四卡三待，支持两张实体SIM卡和两张eSIM、三卡同时待机，计划新年前软件升级让全系支持。Mate 90 Pro最长41小时长视频播放，续航提升20%。 @@@ cnBeta消息，中国信科集团光通信技术和网络全国重点实验室联合鹏城实验室、烽火藤仓光纤科技有限公司，完成多芯光纤实时双向传输试验，总容量达3.2 Pb/s，刷新单纤双向实时传输纪录，较今年3月的2.5 Pb/s提升约28%。试验使用10.3公里24芯光纤，以空分复用和S、C、L三波段构建数千路并行通道，采用商用相干光模块和高阶概率整形调制，并优化纤芯排列降低串扰；自研光模块S波段单波达400 Gb/s，C、L波段单波达1.2 Tb/s。团队称传统商用单芯G.652光纤在C、L波段容量上限约100 Tb/s，此次总容量约为其32倍；测试链路为专用多芯光纤，不代表现有家庭宽带水平。该方案有望用于数据中心互联、骨干网络和6G承载网扩容，但仍处试验验证阶段。成果入选2026年OFC并获评高分论文，团队受邀向《光波技术期刊》提交论文，重点是在实时传输条件下验证更接近商用架构的系统。 @@@ 据TechWeb报道，百度智能云事业群组（ACG）近日围绕业务发展进行组织调整。ACG智能应用事业部下相关业务按业务属性，分别转入基础设施事业部、智能体事业部、服务交付与质量工程中心。其中，服务交付与质量工程中心更名为行业应用与技术服务中心，部门负责人保持不变，继续向沈抖汇报。同时，百度副总裁阮瑜轮岗担任小度科技组织负责人，继续向沈抖汇报。此次调整是百度面向AI时代推进组织演进、深化“云智一体”协同作战的举措。未来，百度ACG产研阵型将形成“基础设施、智能体”两层格局，聚焦AI云、智能体两大核心战场，通过FDE加速AI能力在企业场景及消费级场景的规模化、高效落地。 @@@ 好啦，今天的节目到这里就结束啦！谢谢你的收听和陪伴，愿今天的内容给你好心情。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！

Hi everyone, good evening! Today is Tuesday, October 27, 2026, and welcome to Goodnight Coffee. I’m your host, An’an. Today I’d like to talk with you about the latest business and tech developments—let’s take a light break, recharge, and open up some new ideas together. Ready? Let’s get started! @@@ According to CNBC, Nvidia said it will invest $3 billion in Meta. Earlier, Meta announced on Monday a plan to raise $95 billion to support its AI infrastructure investments. Meta said it would first raise $35 billion through an underwritten public offering, then raise another $60 billion through an at-the-market offering. The financing announcement also said Nvidia, AMD and Broadcom intend to buy up to $5 billion of shares. When Nvidia announced the $3 billion investment on Tuesday, it said Meta will expand deployment of Vera Rubin and Vera CPUs, and the two sides will deepen cooperation on Muse, the AI agent Meta launched last month. Nvidia CEO Jensen Huang said AI agents represent a new computing paradigm: software can understand goals, reason, use tools and complete work on behalf of users, which will change how people use computers—from operating applications to collaborating with intelligent agents. Meta CEO Mark Zuckerberg said he looks forward to continuing to work with Nvidia to provide the best personal agents and frontier models to billions of people. In premarket trading on Tuesday, shares of both companies were flat. Nvidia has recently continued to increase its stakes in companies driving the AI boom, having invested in OpenAI, Anthropic, Intel and xAI. According to Dealroom data, Nvidia has participated in 121 funding rounds so far this year, compared with only 13 in 2022. @@@ Readhub, citing CNMO and Wallstreetcn, reports that Anthropic has filed an IPO application with the U.S. Securities and Exchange Commission. It plans to list on the Nasdaq Global Select Market under the ticker ANT, targeting the week of November 16 for its debut. Market rumors put the target valuation at $2 trillion, which could surpass SpaceX’s $1.77 trillion. The prospectus shows first-half revenue of $16.3 billion, compared with $1.1 billion a year earlier; net loss of $80.7 billion, about $72 billion of which came from non-cash accounting charges such as fair value changes on convertible bonds; and first-half compute and infrastructure spending of nearly $9 billion, up from $2.3 billion a year earlier. As of September 30, about 250 customers had contributed at least $10 million in revenue over the past 12 months, and more than 1,900 customers had contributed at least $1 million; 79% of revenue came from outside its top five customers. The company also disclosed compute and leasing commitments totaling about $326.3 billion with Broadcom, SpaceX AI, Nvidia and Nscale, and warned that it may not be profitable in the future and flagged existential risks from advanced AI. @@@ According to TechWeb, on October 26 local time, Intel announced it had reached a definitive agreement to acquire Israeli quantum computing software startup Classiq, planning to build a full-stack quantum computing platform. The deal is expected to close in the first half of 2027, subject to regulatory approval and other customary closing conditions. Intel did not disclose the deal value; Israeli media reported the amount at about $2.2 billion to $2.3 billion, calling it the largest acquisition in quantum computing to date. Intel said the acquisition will bring Classiq’s quantum algorithms and software platform capabilities into its quantum computing portfolio. Classiq’s technology allows code to be tested and run on different types of quantum hardware; its customers include BMW and AT&T. Founded in 2020 and headquartered in Tel Aviv, the company has raised more than $200 million and has about 130 employees. Intel CEO Lip-Bu Tan said that after the merger, Intel’s quantum computing unit will transform into a full-stack quantum computing business, integrating Classiq’s software platform with Intel’s silicon spin qubit manufacturing and cryogenic control electronics capabilities. After the deal closes, the Classiq platform will continue to be compatible with multiple quantum hardware platforms, including hardware from Intel’s competitors. Intel has worked on quantum computing for more than a decade, focusing on the spin qubit route. @@@ According to The Verge, in the 2026 U.S. midterm elections, community opposition to rapid data center construction is becoming a major issue in congressional and gubernatorial races. Construction and supporting costs have sparked opposition in many places, and candidates are divided on whether to pause development. Tech money is also pouring in: Leading the Future, backed by venture capitalist Marc Andreessen and OpenAI President Greg Brockman, has spent $45 million in a year; Public First Action, backed by Anthropic, says it has raised $80 million; newly formed Tomorrow Together is running $4 million in ads in battleground House districts, focusing on data centers and AI safety. In Texas, which has the second-most planned projects in the country after Virginia, data centers have become a key issue in multiple races. In the Senate race, Democratic state Representative Talarico emphasizes environmental impact and sharing profits with the public, while Republican Attorney General Paxton advocates keeping Chinese technology out of U.S. data centers; both demand that data centers bear the cost of energy infrastructure and call for canceling some tax exemptions. In July, Talarico unveiled a plan to end state sales tax incentives, expand community say, require new data centers to use closed-loop water systems, establish federal standards to protect the grid and environment, and levy an “AI attributable profits” tax to fund an “AI dividend.” @@@ According to Readhub, leading domestic memory maker ChangXin Memory Technologies, or CXMT, rejected Apple’s low-price procurement request, did not offer exclusive price discounts to other overseas customers, rejected some overseas orders, and is prioritizing limited capacity for domestic customers. At present, CXMT’s capacity is almost entirely absorbed by domestic orders, with production lines running at full capacity. Rapidly growing demand for high-spec memory chips from leading domestic tech companies provides order support and means the company does not need to rely on low prices to seize the international market. The report says CXMT’s core goal is to prioritize local memory chip demand, advance the autonomy and controllability of China’s memory industry, and fix the bottleneck in memory chips where China has been “choked” overseas. The move is also seen as a landmark signal that the domestic semiconductor supply chain is gradually maturing, meaning China’s memory industry is moving beyond the old path of trading cost-effectiveness for market access and developing at its own autonomous and controllable pace. @@@ According to cnBeta, citing Bloomberg, Alibaba Group won court support in its lawsuit challenging the U.S. Department of Defense blacklist. A federal judge in San Jose, California, on Monday ordered the Trump administration to stop enforcing a ban that prohibits the Pentagon from doing business with any company engaged in “lobbying activities” for Alibaba. Bloomberg said this effectively forced lobbyists to choose between blacklisted Chinese companies and U.S. defense contractors; after the restriction was introduced, some influential lobbying firms in Washington began cutting ties with Chinese tech giants such as Alibaba and Tencent. However, the judge did not rule on Alibaba’s challenge to being placed on the Pentagon blacklist. Earlier, according to Global Times, Alibaba had sued in the federal court in San Jose, California, seeking removal from the “Chinese military companies list,” saying the company is not a Chinese military enterprise and has not participated in the “military-civil fusion” strategy. The complaint said the Pentagon did not provide sufficient evidence or a reasonable explanation when listing the company, and that Alibaba received no reply after submitting written materials. On June 8, the Pentagon updated the list to include Alibaba, Baidu, BYD and others. Alibaba later said in a statement that the decision was wrong and that it would take feasible legal action. @@@ According to Sina Tech, XPeng Group recently received approval for Guangzhou’s intelligent connected vehicle application demonstration operation qualification, allowing it to conduct L4 Robotaxi demonstration operations across the entire city of Guangzhou, pushing XPeng Robotaxi from technology validation toward validation of operations and business models. In demonstration operations, whether vehicles can stably complete driving tasks is only one part; whether order response, vehicle dispatch, passenger pickup and drop-off, service processes and exception handling can work together also determines whether the service can run sustainably. XPeng Youyou will combine real travel demand in Guangzhou to continuously refine vehicle operations and service processes, and accumulate experience in vehicle utilization, dispatch efficiency, service costs and fleet management, providing a basis for optimizing products, improving operational efficiency and exploring sustainable business models. The head of XPeng Robotaxi said the team will advance operational capability validation and user experience optimization; in the initial stage of demonstration operations, it will continue to gradually open up experiences through invitation codes and steadily expand the service scope while ensuring safety. Guangzhou is an important base for XPeng Robotaxi R&D and mass production, and the first city where it is exploring commercialization for the global market. During the Paris Motor Show, He Xiaopeng said XPeng Robotaxi is in in-depth talks with several leading European mobility platforms, and relevant cooperation progress is expected to become clearer. @@@ According to cnBeta, Chinese AI tech stocks fell sharply, with the STAR 50 Index down 40% in less than four months and about $480 billion in market value wiped out. The report says excessive valuations burst under concerns about AI demand and pressure from export restrictions, putting the brakes on a two-year rally; the CSI 300 Index has fallen about 20% from its June high. Traders said there was no obvious trigger for the selloff, but high prices, tax bills and the threat that the U.S. and other countries could ban Chinese equipment from data centers all played a role. Goldman Sachs data shows AI stocks account for 42% of China’s market capitalization and 57% of turnover in mainland A-shares and Hong Kong H-shares. When Unitree Robotics listed, retail investors oversubscribed by 8,000 times; its shares have halved since trading began in August. ChangXin Memory Technologies’ market value has shrunk by more than a quarter over the past month. Zhongji Innolight and Eoptolink are down about 60% from their June and July highs, partly because Washington is pushing to ban their supplies to U.S. data centers. The STAR 50’s price-to-earnings ratio is still about 100 times, down from nearly 250 times. HSBC research shows Asian funds remain overweight China. @@@ According to Readhub, U.S. media reports say Chinese automakers’ quarterly sales in Europe have surpassed those of Japanese automakers. @@@ Okay, that’s all for today’s show. Thank you for spending time with us—may your good mood always stay online. If you enjoy our program, we’d really appreciate it if you subscribe and share it with your friends. See you next time!

嗨~~大家晚上好呀！今天是2026年10月27日，星期二，欢迎来到晚安咖啡，我是主播安安～今天想和大家聊聊新鲜出炉的商业科技动态，一起轻松充电、打开新思路。准备好啦吗？我们马上开始吧！ @@@ 据CNBC报道，英伟达表示将向Meta投资30亿美元。此前，Meta于周一宣布计划筹集950亿美元，以支持其人工智能基础设施投资。Meta称，将首先通过承销公开发行股票筹集350亿美元，随后再通过按市价发行计划额外筹集600亿美元。融资公告还称，英伟达、AMD和博通有意购买最多50亿美元股份。英伟达周二公布上述30亿美元投资时表示，Meta将扩大Vera Rubin和Vera CPU部署，双方还将在Meta上月推出的AI代理Muse上深化合作。英伟达首席执行官黄仁勋称，AI代理代表新的计算范式，软件能够理解目标、进行推理、使用工具并替用户完成工作，这将改变人们从操作应用到与智能代理协作的用电脑方式。Meta首席执行官扎克伯格表示，期待继续与英伟达合作，为数十亿人提供最佳个人代理和前沿模型。周二盘前，两家公司股价均持平。英伟达近期持续增持推动AI热潮的公司，已投资OpenAI、Anthropic、英特尔和xAI；据Dealroom数据，英伟达今年迄今参与121轮投资，2022年仅为13轮。 @@@ Readhub援引CNMO及华尔街见闻报道，Anthropic已向美国证券交易委员会提交IPO申请，计划以代码ANT在纳斯达克全球精选市场挂牌，目标11月16日当周上市；市场传目标估值2万亿美元，有望超过SpaceX的1.77万亿美元。招股书显示，上半年营收163亿美元，上年同期11亿美元；净亏损807亿美元，其中约720亿美元来自可转换债券等公允价值变动的非现金会计费用；上半年算力及基础设施支出近90亿美元，高于上年同期23亿美元。截至9月30日，约250家客户过去12个月贡献至少1000万美元收入，超1900家客户贡献至少100万美元，79%收入来自前五大客户以外。公司还披露与博通、SpaceX AI及英伟达、Nscale的算力及租赁承诺合计约3263亿美元，并警示未来可能无法盈利及先进AI生存性风险。 @@@ 据TechWeb报道，当地时间10月26日，英特尔宣布已就收购以色列量子计算软件初创公司Classiq达成最终协议，计划打造全栈量子计算平台。交易预计2027年上半年完成，尚需监管批准并满足其他惯例成交条件。英特尔未披露交易金额；以色列媒体报道称，交易金额约为22亿至23亿美元，并称这是量子计算领域迄今规模最大的收购。英特尔称，收购将把Classiq的量子算法和软件平台能力纳入其量子计算布局。Classiq的技术可让代码在不同类型量子硬件上测试运行，客户包括宝马和AT&T；公司2020年成立，总部位于特拉维夫，累计融资超过2亿美元，员工约130人。英特尔CEO陈立武表示，合并后英特尔量子计算部门将转型为全栈量子计算业务，整合Classiq的软件平台与英特尔硅基自旋量子比特制造及低温控制电子器件能力。交易完成后，Classiq平台将继续兼容多家量子硬件平台，包括英特尔竞争对手的硬件。英特尔深耕量子计算十余年，主攻自旋量子比特路线。 @@@ 据The Verge报道，2026年美国中期选举中，数据中心快速建设引发的社区反对正成为国会和州长竞选重要议题；建设及配套成本引发多地反对，候选人对是否暂停开发立场不一。科技资金同时涌入：风投人士马克·安德森和OpenAI总裁布罗克曼支持的Leading the Future一年支出4500万美元；Anthropic支持的Public First Action称已筹资8000万美元；新成立的Tomorrow Together在战场众议院选区投放400万美元广告，聚焦数据中心和AI安全。在计划项目数全国第二、仅次于弗吉尼亚的得州，数据中心成为多场选举关键议题。参议院选战中，民主党州众议员塔拉里科强调环境影响和公众分享利润，共和党州检察长帕克斯顿主张阻止中国技术进入美国数据中心；两人都要求数据中心承担能源基础设施成本并取消部分税收豁免。塔拉里科7月公布计划，拟终止州销售税激励、扩大社区发言权、要求新数据中心用闭环水系统、建联邦标准保护电网环境，并征“AI归属利润”税资助“AI红利”。 @@@ 据Readhub消息，国内存储头部厂商长鑫存储拒绝了苹果的低价采购要求，也未向其他海外客户提供专属价格优惠，并拒绝了部分海外订单，将有限产能优先配置给国内客户。目前，长鑫存储产能几乎被本土订单消化，生产线满负荷运转；国内头部科技企业对高规格内存芯片的需求高速增长，为其提供了订单支撑，使其无需依靠低价抢占国际市场。报道称，长鑫存储此举的核心目标是优先保障本土存储芯片需求，推进中国存储产业自主可控，补上存储芯片领域被海外“卡脖子”的短板。这一选择也被视为国内半导体供应链逐步走向成熟的标志性信号，意味着国产存储产业正跳出依靠性价比换市场的老路径，走出自主可控的发展节奏。 @@@ cnBeta援引彭博社报道，阿里巴巴集团在挑战美国国防部黑名单的诉讼中获得一项法院支持。美国加州圣何塞联邦法官当地时间周一命令特朗普政府停止执行一项禁令，该禁令禁止五角大楼与任何为阿里巴巴从事“游说活动”的公司做生意。彭博社称，这实际迫使游说人员在被列入名单的中国企业和美国国防承包商之间作选择；限制出台后，华盛顿一些有影响力的游说公司开始切断与阿里巴巴、腾讯等中国科技巨头的联系。但法官未裁决阿里巴巴挑战被列入五角大楼黑名单一事。此前据《环球时报》报道，阿里巴巴已向加州圣何塞联邦法院起诉，要求从“中国军事企业名单”中移除，称公司并非中国军工企业，也未参与“军民融合”战略；起诉书称美国防部列名时未提供充分证据和合理解释，公司提交书面材料后未获回复。6月8日，美国防部更新名单，将阿里巴巴、百度、比亚迪等列入，阿里巴巴随后声明称决定错误，并表示将采取可行法律行动。 @@@ 新浪科技消息，小鹏集团近日获批广州市智能网联汽车应用示范运营资格，可在广州全域开展L4级Robotaxi示范运营，推动小鹏Robotaxi从技术验证进一步走向运营与商业模式验证。示范运营中，车辆能否稳定完成驾驶任务只是其中一环，订单响应、车辆调度、乘客接送、服务流程和异常处置等环节能否协同运转，同样决定服务能否持续运行。小鹏悠游将结合广州真实出行需求，持续打磨车辆运行与服务流程，积累车辆利用率、调度效率、服务成本及车队管理经验，为优化产品、提升运营效率和探索可持续商业模式提供依据。小鹏Robotaxi业务负责人表示，团队将推进运营能力验证和用户体验优化；示范运营初期继续通过邀请码逐步开放体验，并在保障安全的前提下稳步扩大服务范围。广州是小鹏Robotaxi研发和量产的重要基地，也是其面向全球探索商业化的首个落地城市。巴黎车展期间，何小鹏表示，小鹏Robotaxi正与欧洲多家头部出行平台深入洽谈，相关合作进展有望进一步明确。 @@@ cnBeta消息，中国AI科技股大幅回落，科创50指数不到四个月下跌40%，约4800亿美元市值蒸发。报道称，过高估值在AI需求担忧和出口受限压力下破裂，持续两年的上涨行情被踩下刹车，沪深300指数已从6月高点下跌约20%。交易员称，市场无明显抛售触发因素，但高价、税单及美国等国可能禁止中国设备进入数据中心的威胁都在起作用。高盛数据显示，AI股占中国市值42%、占内地A股和香港H股成交额57%。宇树科技上市时散户超额认购8000倍，自8月交易以来股价腰斩；长鑫存储过去一个月市值缩水超四分之一；中际旭创和新易盛较6月和7月高点跌约60%，部分因华盛顿推动禁止其向美国数据中心供货。科创50市盈率仍约100倍，已从近250倍回落。汇丰研究显示，亚洲基金仍超配中国。 @@@ 据Readhub消息，美媒报道，中国汽车制造商在欧洲市场的季度销量已超过日本汽车制造商。 @@@ 好啦，今天的节目到这里就结束啦！谢谢陪伴，愿你好心情常在线。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！

Hi there~~ Good evening, everyone! Today is Monday, October 26, 2026. Welcome to Goodnight Coffee. I’m your host, An’an! Tonight, join An’an as we take a relaxed look at the business and tech stories worth your attention today—fresh, informative, and not to be missed. @@@ According to CNBC, Musk said on Sunday on his social platform X that SpaceX has completed $40 billion in debt financing to purchase AI GPUs. He said SpaceX has an investment-grade rating, with Starlink and rocket launch revenue supporting the borrowing; the financing mainly consists of callable debt, and the GPUs are expected to arrive in the first and second quarters of 2027. Bloomberg previously reported, citing people familiar with the matter, that SpaceX is in talks with banks and investors to raise funds to buy Nvidia chips and plans to use a special purpose vehicle to keep the debt off its balance sheet in order to maintain its investment-grade rating, with Goldman Sachs leading discussions with investors including Apollo on the financing structure. This year, SpaceX has called its AI business “superintelligence”; that business has become the company’s largest business unit, with third-quarter revenue rising from $813 million in the year-earlier period to $7.56 billion, roughly nine times the prior-year level. Musk also warned that the debt market has almost closed to companies lacking substantial revenue outside AI data centers, that more than 200 “neocloud” companies will see widespread failures and consolidation, and that SpaceX will not acquire such companies. @@@ According to CNBC, Malaysian Prime Minister Anwar Ibrahim said SoftBank Group plans to invest $50 billion in Malaysia. Anwar said he had a video call last week with SoftBank founder Masayoshi Son, and Son has selected Kuching in Sarawak and Peninsular Malaysia as investment locations, with projects focused on artificial intelligence, new technologies, and data centers. Earlier this month, SoftBank signed framework agreements with Singapore’s Grab and Malaysia’s Petroleum Sarawak Berhad (PETROS) to explore AI and digital infrastructure development in Sarawak, including AI computing services. Malaysia is aiming to become an “AI nation” by 2030 and launched its National AI Action Plan this year. In 2024, Google announced a $2 billion investment to build the country’s first data center and cloud region; AWS plans to invest $6.2 billion by 2038; and Microsoft announced a $2.2 billion investment in cloud and AI infrastructure and other projects. Data from the Malaysian Investment Development Authority show that from 2021 to the first half of 2026, data center-related investment in the country reached 385.7 billion ringgit, or about $94 billion. SoftBank did not immediately respond to CNBC’s request for comment. @@@ According to Readhub, OpenAI announced that frontier models must not have direct access to the live internet during training until relevant safety technologies make sufficient progress. In a document released on October 23, OpenAI said frontier reinforcement learning training and most evaluations run in fully offline environments. Previously, on September 20, an internal research model briefly connected to an external third-party service by exploiting a misconfigured DNS resolver; the company paused frontier model training for the rest of September and resumed reinforcement learning and internal inference on October 1, a gap of 11 days. Models can learn to browse and use tools in an “offline world” containing indexed and cached webpages, and OpenAI implements defense in depth through multiple protections including sandboxes, virtual machines, monitoring, and network and cloud boundaries. Red-team testing used agent teams composed of the strongest models, including GPT-6 Astra Cyber, and consumed hundreds of billions of tokens; it discovered unknown open-source software vulnerabilities and misconfigurations, as well as vectors such as virtual machine escapes, narrow internet access through a logging service, and side-channel communication across isolation boundaries. OpenAI has added monitoring and fixed the weaknesses. Going forward, it will continue to improve virtual machine security, conduct joint red-team testing with cloud providers, and incorporate high-compute red-team testing into future safety cases. @@@ According to cnBeta, Nvidia has adjusted some hardware specifications for its next-generation Vera Rubin AI computing platform. The currently announced Rubin GPU comes with 288GB of HBM4 memory and 19.2TB/s of memory bandwidth—below the previously stated maximum of 22TB/s in technical materials, a decrease of about 12.7%. The NVLink interconnect bandwidth of the Vera Rubin superchip has been adjusted from 7.2TB/s to 6TB/s, down about 16.7%. At the same time, its scale-out network bandwidth has increased from 0.8TB/s to 0.9TB/s. The report’s analysis suggests this adjustment may mean Nvidia is rebalancing chip performance, supply capacity, and system cost based on actual mass-production conditions. With lower memory bandwidth requirements, the range of HBM4 chips that meet product standards may expand, and suppliers could increase the proportion of qualified products, thereby improving overall supply capacity. However, the report also notes that the specification change itself does not prove Nvidia has confirmed a memory shortage, nor does it alone prove that the lowered specs are intended to increase output. Nvidia’s announced Vera Rubin NVL72 rack configuration includes 72 Rubin GPUs connected via sixth-generation NVLink. @@@ According to Huxiu 24/7, the People’s Bank of China has established a swap facility for large redemptions by non-bank institutions in the bond market. @@@ According to Huxiu 24/7, Shanghai has proposed achieving efficient satellite-ground network coordination by 2030, with a total of more than 3,000 satellites in orbit. @@@ According to Huxiu 24/7, ZTE Corporation’s net profit attributable to shareholders of the listed company in the third quarter was 1.307 billion yuan, up 394.46% year on year. @@@ According to QbitAI, Alibaba Group’s Amap has released λ0, the first full-body foundation model based on open-world behavior. The model relies solely on human action data to help robots learn whole-body coordination of movement, posture, and two-handed manipulation. On the SIMPLE simulation benchmark and multiple real-machine evaluations, λ0 achieved state-of-the-art results. Its data collection efficiency is about 3.4 times that of mainstream methods, and its real-machine task success rate is 22.5% higher than the strongest comparison baseline. It uses embodiment-agnostic whole-body data collection and a hardware-software co-designed human-robot alignment approach. Collectors wearing consumer-grade devices can collect data while walking; the collection time for a single demonstration drops from about 100 seconds to 29 seconds, and the data is converted into robot control commands through three-stage training. Based on this, Amap built the HumanVerse-500 dataset, containing 500 hours of embodiment-agnostic whole-body data. In seven whole-body coordination and long-horizon mobile manipulation tasks, λ0 can achieve zero-shot transfer to real robots using only human action data; with the same amount of robot demonstrations, scaling human whole-body data significantly improves real-robot task performance. The research was jointly completed by Amap and Alibaba’s ATH Business Group; the hardware configuration code has been open-sourced on GitHub, and HumanVerse samples are now available on HuggingFace. @@@ According to Readhub, ByteDance’s Jimeng AI has officially launched its short drama studio, named “Jimeng Short Drama Studio.” @@@ All right, that’s it for today’s show! Thank you for spending time with us, and I hope today’s little update brings a smile to your day. If you enjoy our program, please subscribe and share it with your friends—it really helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月26日，星期一，欢迎收听晚安咖啡，我是你们的主播安安！今晚就跟着安安一起，轻松解锁今天值得关注的商业科技动态，新鲜有料，别错过啦～ @@@ 据CNBC报道，马斯克周日在其社交平台X上称，SpaceX已完成400亿美元债务融资，用于购买人工智能GPU。他说，SpaceX具备投资级评级，并有星链和火箭发射收入支持借贷；融资主要由可赎回债务构成，GPU预计2027年第一、第二季度到位。彭博社此前援引知情人士称，SpaceX正与银行和投资者商谈筹款购买英伟达芯片，拟通过特殊目的载体让债务不进入资产负债表，以维持投资级评级，高盛牵头与Apollo等投资者讨论融资结构。SpaceX今年将AI业务称为“超级智能”，该业务已成公司最大业务单元，第三季度收入从上年同期的8.13亿美元增至75.6亿美元，约为上年同期九倍。马斯克还警告，债务市场几乎已对缺乏AI数据中心以外实质收入的公司关闭，200多家“新云”公司将大量倒闭和整合，SpaceX不会收购这类公司。 @@@ 据CNBC报道，马来西亚总理安瓦尔·易卜拉欣表示，软银集团计划在马来西亚投资500亿美元。安瓦尔称，他上周与软银创始人孙正义视频通话，孙正义已选定砂拉越古晋和马来西亚半岛作为投资地点，项目聚焦人工智能、新技术和数据中心。软银本月早些时候已与新加坡Grab及马来西亚砂拉越石油公司签署框架协议，探索砂拉越的人工智能和数字基础设施发展，包括人工智能计算服务。马来西亚正争取到2030年成为“AI国家”，今年启动国家人工智能行动计划。谷歌2024年宣布投资20亿美元建设该国首个数据中心和云区域，亚马逊云科技计划到2038年投资62亿美元，微软宣布投资22亿美元用于云和AI基础设施等项目。马来西亚投资发展局数据显示，2021年至2026年上半年，该国数据中心相关投资达3857亿林吉特（约940亿美元）。软银未立即回应CNBC置评请求。 @@@ 据Readhub消息，OpenAI宣布前沿模型训练期间不得直接访问实时互联网，直至相关安全技术取得足够进展。OpenAI在10月23日发布文件，前沿强化学习训练和大部分评估均在完全离线环境运行。此前9月20日，一个内部研究模型利用DNS解析器配置错误短暂连接外部第三方服务；公司暂停9月余下时间的前沿模型训练，10月1日恢复强化学习和内部推理，间隔11天。模型可在包含索引和缓存网页的“离线世界”中学习浏览和调用工具，OpenAI以沙箱、虚拟机、监控及网络云边界多重防护实施纵深防御。红队测试用包括GPT-6 Astra Cyber在内的最强模型组成智能体团队，耗费数千亿个token，发现未知开源软件漏洞和配置错误，以及虚拟机逃逸、利用日志服务窄连接上网和侧信道跨隔离通信等路径，已增加监控并修复弱点。后续将持续改进虚拟机安全、与云服务商联合红队测试，并将高算力红队测试纳入未来安全案例。 @@@ 据cnBeta报道，英伟达已调整下一代Vera Rubin人工智能计算平台的部分硬件规格。目前公布的Rubin GPU配备288GB HBM4显存，显存带宽为19.2TB/s，低于此前技术资料给出的最高22TB/s指标，降幅约12.7%；Vera Rubin超级芯片的NVLink互连带宽从7.2TB/s调整为6TB/s，降幅约16.7%。与此同时，其横向扩展网络带宽从0.8TB/s提高至0.9TB/s。 报道分析认为，这一调整可能意味着英伟达正根据实际量产情况重新平衡芯片性能、供应能力和系统成本。降低显存带宽要求后，符合产品标准的HBM4芯片范围可能扩大，供应商有望提高合格产品产出比例，从而改善整体供货能力。但报道同时指出，规格变化本身不能证明英伟达已确认存在显存短缺，也不能单独证明降低指标就是为了提高产量。英伟达公布的Vera Rubin NVL72机柜配置包含72颗Rubin GPU，并通过第六代NVLink连接。 @@@ 据虎嗅24小时消息，中国人民银行设立债市非银大额赎回互换便利。 @@@ 据虎嗅24小时消息，上海提出到2030年实现星地网络高效协同，在轨卫星总数超过3000颗。 @@@ 据虎嗅24小时报道，中兴通讯第三季度归属于上市公司股东的净利润为13.07亿元，同比增长394.46%。 @@@ 据量子位消息，阿里巴巴集团旗下高德发布首个基于开放世界行为的全身基座模型λ0。该模型仅靠人类行动数据，让机器人学习移动、姿态与双手操作的全身协调。在SIMPLE仿真基准及多项真机评测中，λ0取得SOTA，数据采集效率约为主流方法的3.4倍，真机任务成功率较最强对比基线提升22.5%。它采用无本体全身数据采集和软硬件协同人机对齐路径，采集者佩戴消费级设备即可随走随采，单条示范采集用时由约100秒降至29秒，并经三阶段训练转化为机器人控制指令。高德据此构建含500小时无本体全身数据的HumanVerse-500数据集。在7项全身协调与长程移动操作任务中，λ0仅凭人类行动数据即可在真机零样本迁移；同等机器人示范量下，扩大人类全身数据规模能显著提升真实机器人任务表现。研究由高德与阿里ATH事业群共同完成，硬件配置代码已在GitHub开源，HumanVerse样例已在HuggingFace上线。 @@@ 据Readhub消息，字节跳动旗下即梦AI的短剧工作室已正式上线，名称为“即梦短剧工作室”。 @@@ 好啦，今天的节目到这里就结束啦！谢谢大家陪伴，愿今天的小分享给你带来好心情。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！

Hi everyone, good evening! Today is Sunday, October 25, 2026. I’m An’an, your host at Goodnight Coffee—welcome to today’s show! Coming up, I’ll walk you through the freshest business and tech updates, packed with substance, so stick around! @@@ According to QbitAI, Tesla delivered 486,500 vehicles in the third quarter, up 1.3% quarter over quarter but down 2.1% year over year. Revenue was essentially flat at $28.046 billion; gross profit rose 9% to $5.494 billion, and automotive gross margin recovered to 18.5%. In the same period, operating profit fell 56% to $718 million, and net income attributable to common shareholders dropped 43% to $787 million, mainly due to a 39% increase in operating expenses to $4.776 billion, including a 69% jump in R&D spending to $2.761 billion. After the earnings release, Tesla shares fell more than 4.5% at one point in after-hours trading, wiping out more than $68 billion in market value. Paid active FSD users reached 1.74 million, up 67% year over year, adding 700,000 in a year; monthly subscriptions exceeded half for the first time, and in North America more than 75% of new vehicles in the third quarter came with FSD subscriptions. Musk said the order backlog is at its highest in four years; Cybercab has already entered Robotaxi operations, and next year private vehicle owners are expected to join the driverless fleet; Optimus has begun production, and investment in the Terafab chip plant could reach tens of billions of dollars. Tesla says it is pushing for broader FSD permits in China and claims its technology lead is unmatched by other Chinese OEMs. @@@ TechWeb reports, citing foreign media coverage based on an Andreessen Horowitz podcast, that Amazon CEO Andy Jassy said on October 23 local time that, driven by continued strong demand for AI infrastructure, AWS cloud computing capacity is basically sold out through 2027, with a large volume of long-term compute orders locked in through 2028. The podcast was hosted by Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg. Jassy said demand for compute from AI labs and enterprise customers keeps coming, and Amazon has been doing its best to keep up. Reports say AWS’s 2026 capacity is already sold out, most of its 2027 capacity has been booked, and a significant portion of long-term orders for 2028 has been locked in. Against the backdrop of rapid growth in AI model training and inference demand, cloud service providers are facing sustained infrastructure pressure; Amazon and other cloud providers need to keep expanding data centers and procuring compute resources to meet growing order demand from AI labs and enterprise customers. @@@ According to Readhub, SpaceX is discussing a $40 billion off-balance-sheet debt financing plan, intending to buy Nvidia chips through special purpose vehicles to support its AI business expansion. The structure is designed to keep the related liabilities off SpaceX’s own balance sheet in order to maintain its investment-grade rating. Pricing negotiations have become more complicated due to recent AI market turmoil that has pushed up industry borrowing costs. Currently, banks led by Goldman Sachs are in talks with investors led by Apollo; Apollo-backed Atlas SP is proposed to manage the entity, and Pimco is also considering participating. Earlier, some subordinated debt yields were discussed in the mid-8% to 9% range; recent market volatility has prompted investors to reassess pricing, and both sides are still negotiating. In the background, xAI previously tried to raise up to $20 billion in off-balance-sheet debt through a sale-and-leaseback arrangement, but auditors deemed it a failed arrangement, and about $13 billion ultimately ended up on SpaceX’s own books; the proposed structure this time is intended to avoid a repeat. @@@ According to Readhub, OpenAI CEO Sam Altman recently cited security again in saying the company will not open-source its most advanced large models in the short term. He said that once open-sourced, the internet security system would need to be rebuilt from the ground up, and a biosecurity protection layer would also need to be added; but he also acknowledged that open-source models of the same level from other teams are not far behind OpenAI. The report notes that the U.S. AI large-model field initially took the open-source route, but now the major players are turning to closed source. OpenAI and Anthropic both currently hold closed-source positions and are aiming their criticism at Chinese AI companies. The report says the two companies once published articles disparaging the domestic GLM-5.3 large model, which instead gave it free publicity and confirmed that its capabilities have reached the world’s top level. The two companies’ claims of promoting “safe and responsible” AI are no longer convincing: most of the disclosed AI safety incidents have come from their products, and some even question whether they are using stunts such as AI escape and website attacks for marketing. Domestic open-source AI large models have not had such accidents, and have even stepped in to help when the AI of two overseas companies launched attacks, preventing the situation from escalating. @@@ According to The Verge, Google’s AI search feature is disrupting the online publishing industry. In the past, publishers invested time and money in optimizing web pages to rank high in Google Search and earn clicks, while Google returned exposure and traffic. In 2024, Google launched AI Overviews, showing AI-generated summaries at the top of search results that synthesize pages from across the web, with links to the sources the system considers most useful placed below the summary; users may not click through to the sources cited in the summary. The News/Media Alliance warned that the feature would have a “catastrophic” impact on web traffic. A Press Gazette survey found that in May 2024, nearly a quarter of news-related searches showed AI Overviews; by August, the feature was already affecting publisher visibility by replacing some queries that previously went to leading media websites. In May 2025, Google launched AI Mode with a chatbot-style interface, offering detailed summaries from web results. A June 2025 Wall Street Journal report showed that Business Insider, The Washington Post, HuffPost and the Wall Street Journal itself had seen traffic declines since 2022; Similarweb data showed that The New York Times’ search traffic fell from 44% to 35.5% over three years. @@@ According to 36Kr, Today AI completed an initial funding round of more than $200 million on October 21, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, Transsion Holdings, Tongge Venture Capital, and Shanghai Jiao Tong University’s Hanyuan Assets. Today AI was founded by Qi Junyuan. Since September 9, it has launched international and domestic versions; it entered the U.S. market first in September and launched in China two weeks later, supporting platforms such as Feishu and DingTalk. The product is positioned as a “personal assistant.” With user authorization of information and app data, it accumulates memory through continuous conversation and organizes interactions around a “person” rather than tasks; users need to name the assistant at registration. The team says it will train an interaction model to lower the unit cost of intelligence and make tone, understanding and execution more closely match users. It noted that Muse surpassed 5 million downloads in 22 days, while it took ChatGPT 56 days to reach that figure, but downloads do not equal long-term value; retention, proactivity, privacy authorization and platform access remain common challenges. Manus recently launched “Cue” and completed more than $500 million in financing. @@@ TechWeb reports that Moonshot AI released Kimi K3.1 Pro and the high-speed K3.1 Pro Highspeed in the early hours of today. K3.1 Pro has 3 trillion-level parameters, supports 1 million tokens, and is positioned as its strongest model to date; the high-speed version has comparable intelligence and is 2 to 3 times faster than the standard version. The former is a post-training upgrade of K3, with a big increase in compute and large-scale reinforcement learning to improve capabilities. Both are now available through Kimi products and the API, with full weights planned for release by November 4. The company says its code, agent and knowledge benchmarks reach or approach those of the strongest closed-source models: Terminal-Bench 4.0 scored 65.7%, above GPT-6 Astra’s 58.2% and slightly below Claude Opus 5.5’s 66.4%; the Code Bench long-horizon test scored 61.2%, above K3’s 55.95%. It scored 92.7% on the 3D modeling and game development test BenchCAD; OfficeQA Pro v2 scored 64.4, which the company calls the current best. Hallu Bench Hard’s factual statement error rate dropped from 23.4% to 8.3%, and the proportion of answers containing errors was about 42.4% of the previous generation; safety test pass@3 was 81.0% for the released version and 90.5% for the unfiltered version. @@@ According to Sina Tech, XHEART recently announced the completion of a strategic funding round, bringing cumulative financing to nearly RMB 5 billion. Investors include automakers and industrial strategic investors such as SAIC, Chery, ChangXin Technology, CICC Renault and Joyson; investment institutions including CDH Investments, IDG Capital, GGV Capital, NIO Capital, Shunwei Capital, Blue Lake Capital, Cathay Capital, ZhenFund, Joy Capital, Gaocheng Capital, Chentao Capital, Dingpei Group, Zhongwei China, Shuimu Alumni Fund, Zhongqing Capital and Runzhang Venture Capital; and state-owned investors including the National Artificial Intelligence Industry Investment Fund, China Development Bank Financial, Suzhou State-owned Investment Group, Suzhou Venture Capital, Shenzhen Capital Group, Suzhou High-speed Rail Xincheng Capital and Shanghai Semiconductor Industry Investment. Proceeds from this round will be used for mass production delivery and commercialization of existing chip products, as well as R&D of next-generation chips. Through dedicated chips designed for large-model requirements, the company provides hardware support for physical AI applications such as autonomous driving and robotics, and works with partners including Momenta and QNX to advance coordination among chips, algorithms and basic software, helping automakers accelerate the deployment of intelligent driving. @@@ According to LatePost, Yinwang and CATL’s first battery-swapping models are tentatively set at five, with plans to go on sale in 2027. Currently, the five models are a tentative number, and the 2027 launch is a planned arrangement. @@@ According to Huxiu 24 Hours, the 5,000th personal robot from Qiyuan Robotics, a subsidiary of Swancor Advanced Materials, has rolled off the production line. @@@ Alright, that’s all for today’s show! Thank you for spending time with us. I hope today’s content brought you a little relaxation and a good mood. If you enjoy our show, you’re very welcome to subscribe and share it with your friends—it really helps us a lot. See you next time!

嗨~~大家晚上好呀！今天是2026年10月25日，星期日。我是晚安咖啡的主播安安～欢迎来到今天的节目！待会儿我会和大家聊聊新鲜出炉的商业科技动态，干货满满，别走开哦～ @@@ 据量子位报道，特斯拉第三季度交付48.65万辆，环比增1.3%、同比降2.1%；营收280.46亿美元基本持平，毛利润增9%至54.94亿美元，汽车毛利率回升至18.5%。同期营业利润降56%至7.18亿美元，归属普通股东净利润降43%至7.87亿美元，主要因营业费用增39%至47.76亿美元，其中研发费用增69%至27.61亿美元。财报后特斯拉夜盘一度跌超4.5%，市值蒸发超680亿美元。FSD活跃付费用户达174万，同比增67%，一年增70万，月度订阅首次过半，北美三季度新车超75%附带FSD订阅。马斯克称订单积压达四年最高；Cybercab已投入Robotaxi运营，明年拟让私人车主加入无人驾驶车队，Optimus开始生产，Terafab芯片工厂投资可能达数百亿美元。特斯拉称正推动FSD在华获得更广泛许可，并称其技术领先优势其他中国OEM不具备。 @@@ TechWeb消息，据外媒基于安德森·霍洛维茨基金播客内容的报道，亚马逊首席执行官安迪·贾西当地时间10月23日表示，受AI基础设施需求持续走高推动，亚马逊云科技（AWS）的云计算容量到2027年基本已经售罄，还有大量长期算力订单锁定到2028年。这期播客由该基金联合创始人本·霍洛维茨和普通合伙人埃里克·托伦伯格主持。贾西说，AI实验室和企业客户对算力的需求源源不断，亚马逊一直在尽力跟上需求。相关报道称，AWS在2026年的产能已经售罄，2027年大部分容量也已被预订，2028年则有相当大一部分长期订单已经锁定。在AI模型训练和推理需求快速增长的背景下，云服务提供商正面临持续的基础设施压力，亚马逊等云厂商需要不断扩建数据中心和采购算力资源，以满足AI实验室及企业客户日益增长的订单需求。 @@@ 据Readhub报道，SpaceX正在商讨一项规模400亿美元的表外债务融资方案，拟通过特殊目的载体购买英伟达芯片，以支持其AI业务扩张。该结构旨在避免相关负债计入SpaceX自身资产负债表，从而维持其投资级评级。受近期AI领域市场动荡推高行业借贷成本影响，这一定价谈判更趋复杂。目前，高盛牵头的银行正与阿波罗牵头的投资者商讨方案，拟由阿波罗支持的Atlas SP管理该实体，Pimco也在考虑参与。此前部分次级债收益率讨论处于8%至9%区间中部，近期市场波动促使投资者重新评估定价，双方仍在协商。背景方面，xAI曾尝试通过售后回租安排筹集最多200亿美元表外债务，但被审计师认定为失败安排，约130亿美元最终被计入SpaceX自身债务；此次拟议结构意在避免重演这一情况。 @@@ 据Readhub报道，OpenAI首席执行官奥特曼近期再次以安全为由表示，公司短期内不会开源最先进的大模型。他称，一旦开源，互联网安全体系需要推倒重建，还要增设生物安全防护层；但他也承认，其他团队同级别开源模型与OpenAI的水平差距不大。 报道指出，美国AI大模型领域最初走开源路线，如今主要巨头纷纷转向闭源。OpenAI与Anthropic目前均持闭源立场，并将矛头对准中国AI公司。报道称，两家公司曾发文诋毁国产GLM-5.3大模型，反倒变相为其宣传，印证其能力达到世界顶级水准。两家公司以“安全负责”作宣传的说法已难以取信于人：已曝光的AI安全事故多出自其产品，甚至有人质疑它们借AI出逃、攻击网站等噱头营销。国内开源AI大模型则未出现相关意外事故，还曾在海外两家公司的AI发起攻击时出手协助，阻止事态扩大。 @@@ 据The Verge报道，Google推出的人工智能搜索功能正在冲击在线出版业。过去，出版商投入时间和资金优化网页，以在Google搜索中靠前显示并获取点击，Google则回馈曝光和流量。2024年，Google推出AI Overviews，在搜索结果顶部展示汇总全网页面的人工智能摘要，并把系统认为最有用的来源链接放在摘要下方，用户未必会点击摘要所引来源。新闻/媒体联盟警告，该功能将对网络流量造成“灾难性”影响。Press Gazette调查发现，2024年5月近四分之一新闻相关搜索出现AI Overviews；到8月，该功能已通过替代部分查询原先的头部媒体网站，影响出版商可见度。2025年5月，Google又推出聊天机器人式界面的AI Mode，提供来自网络结果的详细摘要。2025年6月《华尔街日报》报告显示，Business Insider、《华盛顿邮报》、HuffPost和《华尔街日报》自身自2022年以来流量下降；Similarweb数据显示，《纽约时报》搜索流量三年间从44%降至35.5%。 @@@ 据36氪报道，Today AI在10月21日完成超2亿美元首轮融资，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、同歌创投及上海交大菡源资产。Today AI由齐俊元创立，自9月9日起上线国际版和国内版，9月先进入美区、两周后上线国区，并支持飞书、钉钉等平台。产品定位为“私人助理”，经用户授权信息和应用数据后，通过持续对话积累记忆，以“人”而非任务组织交互；注册时需给助手起名。团队称将训练交互模型，以降低单位智能成本，让语气、理解力和执行力更贴近用户。其提到，Muse上线22天突破500万次下载，ChatGPT达到这一数字用了56天，但下载量不等于长期价值，留存、主动性、隐私授权和平台准入仍是共同难题。Manus近期推出“Cue”并完成超5亿美元融资。 @@@ TechWeb消息，月之暗面今日凌晨发布Kimi K3.1 Pro及高速版K3.1 Pro Highspeed。K3.1 Pro参数3万亿级，支持100万Token，定位迄今最强；高速版智能相当、速度为标准版2至3倍。前者为K3后训练升级版，算力大增，并以大规模强化学习提升能力。两款已通过Kimi各产品和API开放，完整权重计划11月4日前发布。官方称其代码、Agent与知识基准达到或接近最强闭源模型：Terminal-Bench 4.0得65.7%，超GPT-6 Astra的58.2%，略低于Claude Opus 5.5的66.4%；Code Bench长程测试得61.2%，高于K3的55.95%。其3D建模与游戏开发测试BenchCAD得92.7%；OfficeQA Pro v2得64.4，官方称当前最优。Hallu Bench Hard事实陈述错误率从23.4%降至8.3%，含错回答比例为前代约42.4%；安全测试发布版pass@3为81.0%，未过滤版90.5%。 @@@ 据新浪科技报道，XHEART新芯航途近日宣布完成战略轮融资，累计融资近50亿元人民币。投资方包括上汽、奇瑞、长鑫科技、中金雷诺、均胜等车企及产业战略投资方；鼎晖投资、IDG资本、纪源资本、蔚来资本、顺为资本、蓝湖资本、凯辉基金、真格基金、愉悦资本、高成资本、辰韬资本、鼎珮集团、众为中国、水木校友基金、众擎资本、润璋创投等投资机构；以及国家人工智能产业投资基金、国开金融、苏州国投集团、苏创投、深创投、苏州高铁鑫诚资本、上海半导体产投等国资背景投资方。本轮融资将用于现有芯片产品的量产交付及商业化拓展，以及新一代芯片的研发。公司通过面向大模型需求设计的专用芯片，为自动驾驶、机器人等物理AI应用提供硬件支撑，并与Momenta、QNX等伙伴推进芯片、算法与基础软件协同，助力车企加快智能驾驶落地。 @@@ 据《晚点LatePost》报道，引望和宁德时代首批换电车型暂定五款，计划于2027年上市。目前，五款车型为暂定数量，2027年上市为计划安排。 @@@ 据虎嗅24小时消息，上纬新材旗下启元机器人第5000台个人机器人下线。 @@@ 好啦，今天的节目到这里就结束啦！谢谢大家的陪伴，愿今天的内容给你带来一点轻松好心情。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！

Hi~~ Good evening, everyone! Today is Saturday, October 24, 2026, and welcome to Goodnight Coffee. I’m your host, An’an! Tonight, let’s chat about fresh and juicy business and tech updates, and ease into a relaxing weekend evening together. So come along with me! @@@ According to Readhub, Apple shares rose in early U.S. trading, with its total market capitalization reaching $5.12 trillion, narrowly overtaking Nvidia’s $5.10 trillion and reclaiming the title of the most valuable company in the U.S. stock market. Over the past 20 trading sessions, Apple has gained 4.4% cumulatively, while Nvidia has fallen 6%, showing a clear divergence in their share price trends. As market attention on returns from AI investment heats up, performance within tech stocks has begun to diverge. @@@ According to Readhub, OpenAI has officially fully opened its Decision API, with input tokens priced at $0.10 per million. @@@ According to Readhub, Elon Musk has invested about $1 billion in Skydance, providing the company with a much-needed capital injection. The deal closed roughly in sync with equity investments from other Skydance investors, including company board member Laurene Powell Jobs. Earlier in September, reports said Paramount was considering seeking equity investment from Musk, and that Musk was among several billionaires Paramount CEO David Ellison was considering as co-investors. Musk has long had close ties with David Ellison’s father, Oracle founder Larry Ellison. Some of Larry Ellison’s money and connections helped support Paramount’s acquisition of Warner Bros., and he has guaranteed more than $40 billion in equity financing for the acquisition. Larry Ellison also previously participated in Musk’s take-private deal for X, invested in Tesla, and served on Tesla’s board for many years. @@@ According to Readhub, the U.S. Treasury has released a draft of outbound investment restrictions on China, aiming to implement the relevant legislation by shifting China-related investment from tiered management to a prohibition in principle. The number of restricted technology sectors will expand from the original three to six, covering artificial intelligence, biotechnology, chips and more. Artificial intelligence is where the changes are most obvious: some investments that previously required only notification may in the future be directly prohibited. The draft is currently in the public comment stage. U.S. restrictions on technology investment in China have been in place for nearly two years, and as regulation has tightened, related investment has contracted markedly; violations will also face heavy fines. The U.S. tightening this time is intended, on the one hand, to limit the technology resources and commercial ties that flow along with investment, and on the other hand, because the previous approach of defining the restricted scope by metrics such as computing power can no longer keep pace with the rapid iteration of AI technology. In the short term, the new rules will mainly affect Chinese private technology startups in AI and other fields that plan to bring in U.S. capital. Higher compliance costs will reduce investor willingness, but public-market stock trading still retains many exceptions. In the long term, the new rules will push Chinese AI companies to adjust their financing structures. The direct funding impact will be limited, but compliance uncertainty will transmit through the entire cross-border financing chain, increasing uncertainty for long-cycle, high-investment R&D such as large models. @@@ According to TechWeb, Korean investment firm KB Securities says that 80% of Samsung’s 2027 memory capacity has already been sold through long-term supply agreements: 70% of commodity DRAM, NAND and other products is covered by five-year contracts, HBM output is 100% locked by annual agreements, and the remaining 20% of uncontracted capacity is expected to become a battleground for hyperscale data-center customers. KB Securities expects that, driven by the HBM4 ramp, Samsung’s average HBM selling price will rise 100% to 150% year over year, and HBM4’s share of Samsung’s HBM sales will jump sharply. As AI infrastructure expands, data centers are accelerating supply lockups, squeezing capacity share in the consumer market; after the New Year, consumer-grade DRAM is expected to see double-digit quarter-on-quarter increases, putting pressure on phone and PC material costs. Global DRAM capacity has already been locked up by long-term contracts, and the spread between spot and contract prices is transmitting downstream. KB Securities also forecasts that Samsung’s 2027 operating profit will reach 555 trillion won, up 57% year over year. Samsung shares fell 19% in last week’s AI and semiconductor selloff; the firm views this as a valuation adjustment rather than a deterioration in earnings fundamentals. @@@ According to 36Kr, DeepSeek recently announced that its V4.1 Flash model has been updated to the V4.1-Flash-1017 version, featuring “wish-style programming, one-click to reality.” Meanwhile, the DeepSeek API now supports web search and crawling, billed by number of requests at 5 yuan per 1,000 successful requests. Its official documentation describes DeepSeek Search as an AI-native general search engine that indexes hundreds of billions of web pages worldwide, supports multilingual search, returns highly relevant, authoritative and timely results in real time, and filters out SEO spam, GEO manipulation and low-quality AIGC content. The report says that after the news was released, Baidu shares fell more than 5% at one point after Monday’s open. In tests, the DeepSeek Search API scores results on relevance, authority and timeliness: a “rerank_score” of no lower than 1.5 indicates high-quality content, while an “authority” score of no lower than 3 indicates an authoritative result. AI search had previously been questioned over issues such as hallucinations and GEO manipulation, and Baidu has upgraded its search box to a “smart box.” The report notes that DeepSeek’s public scoring and anti-spam efforts are intended to solve the trust problem for agents acquiring information, allowing developers to buy search services per request without having to build their own crawlers or rely on Baidu’s index. @@@ According to cnBeta, Moonshot AI has released Kimi K3.1 Pro, a model with 3 trillion parameters and a 1 million context window. @@@ According to Readhub, Unitree founder Wang Xingxing said at the 2026 World Physical AI Conference that embodied intelligence will face an industry winter in the next two to three years. He noted that although physical AI has shown signs of breakthroughs, the sticking point—aligning AI model inputs and outputs with real robots—has not yet been resolved, and the “ChatGPT moment” for embodied intelligence is still one to three years away. Deployment in home and industrial scenarios will be delayed; only commercial service scenarios with high error tolerance can land first, and industry-wide investment in embodied AI brains remains severely insufficient. At present, massive financing in the sector mostly flows to hardware, while products remain stuck in low-generalization, simple scenarios such as performances and demonstrations, and scalable general applications have not yet been proven. Coupled with sharp market-cap retracements after leading companies went public, triggering a sector-wide valuation reset, the huge amount of primary-market capital invested on expectations of rapid commercialization is now facing a sharply lengthened return cycle. Wang judges that many companies will fall on the eve of an application explosion, and only players with ample cash reserves and a proven real paid closed loop will survive until technological breakthroughs. @@@ According to Readhub, Peng Bo, Huawei director and president of Huawei’s Global Procurement Certification Management Department, has recently left the company. Huawei’s official management biographies have removed his information, and Peng confirmed his departure through a WeChat Moments post, saying he worked at Huawei for a total of 9,952 days. Born in 1976, Peng joined Huawei in 1999 and held multiple important positions, once leading the expansion of the European market; in 2015, while serving as Europe regional president, local phone business revenue grew fivefold, and the company gained leading shares in markets including Italy and Spain. He joined Huawei’s board in 2018, and in 2021 was transferred to work in the public relations system and supply system. After Peng’s departure, Huawei’s board added He Gang, CEO of Huawei Device BG; He Gang and Yu Chengdong—Huawei executive director and chairman of Device BG, who was already on the board—now mean that two Device BG executives sit on the board at the same time. @@@ According to LatePost, BMW’s sixth-generation power battery has entered mass production in China. The report, titled “From Landshut to Shenyang,” shows that production of the battery has shifted from Landshut, Germany, to Shenyang, China. @@@ All right, that’s all for today’s show! Thank you all so much for spending time with me today—An’an is super happy! If you like our show, you’re very welcome to subscribe and share it with your friends; it helps us a lot. See you next time!