AI周报第541期:科技巨头周三公布AI支出,CoreWeave债务陷入困境

内容来源:https://aiweekly.co/issues/big-tech-reports-ai-spending-wednesday-as-coreweave-debt
内容总结:
【本周焦点:AI投入能否见效?科技巨头财报与信贷市场信号受关注】
本周,投资者将密切关注人工智能(AI)投资回报情况。美联储将于周三公布利率决议,微软和Alphabet将在美股收盘后发布财报,亚马逊紧随其后于周四公布业绩。市场关注的核心问题是:AI支出是否仍在产生回报。与此同时,信贷市场已显现紧张迹象——为CoreWeave违约投保的年化成本现已超过10%。Anthropic预计最快将于近日公开提交IPO文件。Moonshot表示Kimi K3.1 Pro的开放权重将于11月4日发布。以色列将于周二举行大选,竞选期间AI深度伪造内容泛滥;美国中期选举则将于下周二举行。截至目前,AI带来的回报主要流向了出售AI技术与服务的公司,而非采购这些技术的企业。
【赞助内容】
您的团队在客户项目中使用AI。为他们提供一台不存储任何数据的服务器。
为团队执行研究、起草和分析任务的AI代理,运行在仅您可访问的隔离服务器上,采用Anthropic的Claude模型。数据仅在内存中处理,绝不存储。几分钟内即可完成设置,每月起价120美元。
【行业动态】
当前AI领域热门趋势一览,涵盖应用榜单与社区动态。完整汇总见“In the Wild”栏目。
- 一款AI房间改造应用在图形与设计榜单上快速攀升。Maison AI允许用户选择房间、风格和配色方案,生成室内、室外或花园的改造效果图,可免费下载。
- xAI的Grok Bot在生产力榜单上持续攀升。该应用允许用户像给同事发消息一样,通过手机向AI代理发送指令。
- 一款面向自建AI服务器用户的应用在生产力榜单上上升。Conduit是Open WebUI(自托管聊天界面)的手机端应用,本身不运行模型,用户可将其连接至自己的服务器或模型提供商。
- 一款将笔记转化为测验的学习应用在教育榜单上攀升。Gizmo可将扫描笔记和导入材料转化为闪卡和测验,并通过间隔重复帮助记忆。
- 嵌入短信中的AI助手数量正在增多。TechCrunch汇总了20余款此类产品。Instinct在9月融资10亿美元后估值达100亿美元;Poke则是苹果Messages for Business平台批准的首个AI代理。
【快讯】
实验室角斗士时代
- 周三是AI支出数据的关键日。据CNBC报道,Alphabet和微软将于10月28日(周三)发布财报,亚马逊于周四收盘后发布。美联储为期两天的会议也将于周三结束。重点关注各公司对明年数据中心支出的表态,以及云计算增长中来自AI的比例。
- Anthropic的公开IPO文件可能于本周落地。彭博社10月22日报道称,Anthropic预计将在未来数日内公开提交IPO文件。该公司已于6月1日秘密提交。根据美国证券交易委员会规则,公开文件须在公司开始向投资者路演前至少15天发布,因此文件日期可透露上市时间表。
- Replit将OpenAI设为默认模型,并从Anthropic转移。Replit CEO Amjad Masad表示“我认为ChatGPT现在更好”,并将OpenAI设为Replit Agent的默认选项,理由涉及成本、合作关系和性能。客户仍可选择Claude,但多数用户会沿用默认设置,因此Replit在Anthropic上的支出可能大幅下降。
- Kimi K3.1 Pro的开放权重定于11月4日发布。据VentureBeat报道,Moonshot表示将以开放许可证发布权重。Artificial Analysis将该模型综合排名第四,智能指数得分为55,高于K3的44分。X平台上发布的截图似乎要求网红不要标注Kimi付费推广内容。目前无人确认截图作者,Moonshot亦未回应。
AI资本开支税
- 信贷市场正在为出租AI芯片的公司定价真实风险。据CNBC报道,英伟达对AI云服务提供商的直接投资为70亿美元,担保和承诺金额为1645亿美元,合计1715亿美元。为CoreWeave违约投保的五年期成本约为1082个基点,而超过1000个基点通常被视为困境水平。CoreWeave、Nebius和IREN的股价本月均下跌至少30%。
- 更复杂的编程代理工具可能使成本翻三倍,却未带来更好结果。据VentureBeat报道,山东大学研究人员使用五个模型在近500项SWE-Bench Verified任务上运行了Claude Code、OpenCode和mini-SWE-agent。各方案成功率在统计上相同,但最复杂配置的每任务成本最高可达三倍。
政府认真对待AI的一年
- 以色列将于周二举行大选,竞选期间AI视频泛滥。据半岛电视台报道,内塔尼亚胡分享了一段针对对手Gadi Eisenkot的AI视频,财政部长Smotrich发布了一段显示最高法院院长被拖离抗议现场的AI视频。《国土报》发现数十个原本用于攻击伊朗政府的机器人账号已被转而用于支持利库德集团。
- 在被要求必须选择一位候选人时,Grok和GPT选择了相反的政党。据TechCrunch报道,在一项针对近500场美国中期选举的研究中,Grok 4.7有92%的时间选择共和党,GPT-6.1 Sol有97%的时间选择民主党。研究人员使用的是模型API。消费端应用拒绝做出选择,OpenAI表示API并不反映用户实际使用ChatGPT的方式。中期选举将于11月3日举行。
【AI回报流向卖方而非买方】
阿波罗全球管理首席经济学家Torsten Slok全年都在强调同一观点。他援引彭博社和Macrobond数据对《财富》表示,2023年初至2026年初,七大科技公司的利润率从约15%升至约25%,而标普500其余493家公司的利润率维持在约10%。“关键问题在于科技行业之外的投资回报周期有多长,”他说,如果企业不能迅速看到回报,将放缓AI支出。
这正是本周财报背后的核心问题。微软、Alphabet和亚马逊将公布强劲的AI收入,因为它们出售芯片、云服务和模型。但这些数字并不能说明其客户是否获得了回报。真正需要关注的是签署合同的企业。
选民同样持怀疑态度。CNBC 10月24日发布的民调显示,63%的登记选民反对在住所10英里范围内建设AI数据中心,仅13%表示支持。
如果超大规模云服务商本周再次上调支出,风险将落在借款出租芯片的公司(如CoreWeave)以及必须让支出产生回报的客户身上。
【关键要点】
- 关注资本开支指引,而非仅看AI收入。微软和Alphabet将于周三美联储决议后发布财报,亚马逊于周四发布。
- 报价现在决定了你默认获得哪个AI模型。Replit将OpenAI设为默认,客户需主动选择Claude。
- 关注信贷价格,作为AI基础设施问题的早期信号。CoreWeave债务的保险成本已处于困境水平。
- 将AI的候选人选择视为模型特征的反映,而非建议。Grok和GPT在超过90%的情况下选择了相反政党。
【首发研究】
AI Weekly在媒体报道前发布的一手研究。
- 一种新方法通过“边说话边思考”将语音助手等待时间缩短36%。Heartbeat Thinking在用户仍在说话时即运行推理和工具调用,而非等待静默。在FDBench v3基准测试中,其覆盖所需工具调用的比例高出16个百分点,最终答案平均等待时间缩短36.4%。该方法在另外两项基准测试中亦有提升。详见我们的论文页面。
【值得阅读】
- 被解雇的OpenAI研究员首次接受采访:Mikita Balesni和Tomek Korbak对NPR表示,OpenAI从未以书面形式告知他们所称泄露的具体信息。
- Character.AI破产前乱象丛生:Futurism发现该平台向未成年人提供涉及自杀、自残和大规模枪击的机器人。截至8月,该公司每月亏损约800万至900万美元,迪士尼正领衔竞购。
- AI墓地:TechCrunch持续更新的2026年关停AI产品和初创公司名单,包括OpenAI的ChatGPT Atlas浏览器和Sora社交应用。
【等等,什么?】
- ChatGPT在生成的漫画上签署真实漫画家姓名。Nieman Lab发现ChatGPT在其生成的漫画上添加了真实《纽约客》漫画家(包括Harry Bliss和Emily Flake)的签名。
【值得观看】
AI从业者当前正在转发的视频——在AI TV上精选。
| 《AI热潮对企业而言过于昂贵》404 Media | |
| Black Hat USA 2026 | 《“突发”新闻:OpenAI与Hugging Face事件》Black Hat |
【本周投票】
本周什么最能说明AI热潮的走向?
上周,325位读者投票:
如果一款AI工具明天在你公司软件中发现100个漏洞,会发生什么?
本周什么最能说明AI热潮的走向?
下期再见,
Alexis
中文翻译:
本期前瞻。周三美联储公布利率决议,随后微软和Alphabet在收盘后发布财报。亚马逊周四跟进。投资者想知道AI支出是否仍有回报,而信贷市场已经紧张起来:针对CoreWeave违约的保险成本现在每年超过10%。Anthropic预计随时公开提交IPO文件。Moonshot称Kimi K3.1 Pro的开放权重将于11月4日发布。以色列周二投票,此前一场竞选充斥着AI深度伪造视频,而美国中期选举距周二还有一周。在这一切之中:到目前为止,AI让卖它的公司赚到了钱,而不是买它的公司。
赞助商
你的团队在客户工作上使用AI。给他们一台什么都不留存的服务器。
为你的团队做研究、起草和分析的AI智能体,在只有你能访问的隔离服务器上运行Anthropic的Claude。在内存中处理,永不存储。几分钟内完成设置,每月120美元起。
in the Wild
当下AI领域的热门趋势,从应用排行榜到社区动态。完整汇总见In the Wild。
- 一款AI房间重新设计应用在图形与设计榜上快速攀升。Maison AI让你选择房间、风格和配色方案,然后为你的室内、室外或花园生成重新设计方案。免费下载。
- xAI的Grok Bot仍在生产力榜上攀升。Grok Bot让你像给同事发消息一样,从手机上给AI智能体发消息。
- 一款面向自建AI服务器用户的应用正在生产力榜上上升。Conduit是Open WebUI(自托管聊天界面)的手机应用。它本身不运行模型。你将它连接到自己的服务器或模型提供商。
- 一款把笔记变成测验的学习应用正在教育榜上攀升。Gizmo将扫描的笔记和导入的材料转化为抽认卡和测验,并用间隔重复帮助你记忆。
- 存在于你短信中的AI助手正在成倍增加。TechCrunch汇总了20多款。Instinct在9月融资10亿美元后估值达100亿美元,而Poke是苹果为其Messages for Business平台批准的首个AI智能体。
快讯
实验室角斗士时代
- 周三是AI支出数据的大日子。据CNBC报道,Alphabet和微软于10月28日周三发布财报,亚马逊于周四收盘后发布。美联储为期两天的会议也于周三结束。关注他们对明年数据中心支出规模的表述,以及其云增长中有多少来自AI。
- Anthropic的公开IPO文件可能本周落地。彭博社10月22日报道,Anthropic预计在未来几天公开提交IPO文件。Anthropic于6月1日秘密提交。根据SEC规则,公开文件必须在公司开始向投资者路演前至少15天发布,因此其日期能告诉你上市最快何时发生。
- Replit将OpenAI设为默认,并远离Anthropic。Replit CEO Amjad Masad表示“我认为ChatGPT现在更好”,并将OpenAI设为Replit Agent的默认,理由是成本、合作关系和性能的综合考量。客户仍可选择Claude,但大多数用户会沿用默认设置,因此Replit在Anthropic上的支出可能大幅下降。
- Kimi K3.1 Pro的开放权重定于11月4日发布。据VentureBeat报道,Moonshot称将在开放许可下发布权重。Artificial Analysis将该模型排名第四,其智能指数得分为55,高于K3的44。X上发布的截图似乎告诉网红不要标注付费的Kimi帖子。无人确认是谁写的,Moonshot也未回应。
AI资本开支税
- 信贷市场正在为出租AI芯片的公司定价真实风险。据CNBC报道,英伟达在AI云提供商中直接投资70亿美元,并有1,645亿美元的担保和承诺,合计1,715亿美元。据CNBC报道,针对CoreWeave违约的五年期保险成本约为1,082个基点,任何超过1,000的水平通常被视为困境。CoreWeave、Nebius和IREN的股价本月均下跌至少30%。
- 更花哨的编程智能体工具可能让成本翻三倍,却没有更好的结果。据VentureBeat报道,山东大学的研究人员在近500个SWE-Bench Verified任务上,用五个模型运行了Claude Code、OpenCode和mini-SWE-agent。成功率在统计上相同,但最复杂的设置每个任务成本高达三倍。
政府认真对待AI的一年
- 以色列周二投票,此前一场竞选充斥着AI视频。据半岛电视台报道,内塔尼亚胡分享了一段其对手Gadi Eisenkot的AI视频,财政部长Smotrich发布了一段显示最高法院院长被从抗议现场拖走的视频。Haaretz发现数十个旨在攻击伊朗政府的机器人已被转而支持利库德集团。
- 被迫选择一个候选人时,Grok和GPT选择了相反的政党。据TechCrunch报道,在一项针对近500场美国中期选举竞选的研究中,Grok 4.7有92%的时间选择共和党人,GPT-6.1 Sol有97%的时间选择民主党人。研究人员使用了这些模型的API。消费者应用拒绝选择,OpenAI表示API不反映人们如何使用ChatGPT。中期选举于11月3日举行。
AI让卖家受益,买家尚未受益
阿波罗首席经济学家Torsten Slok全年一直在强调同一观点。他援引彭博社和Macrobond数据对《财富》表示,七大科技公司的利润率从2023年初到2026年初从约15%升至约25%。标普500其余493家公司的利润率维持在约10%。“关键问题是科技行业之外的投资回报跑道有多长,”他说,如果企业看不到快速回报,就会放缓AI支出。
这就是本周财报背后的问题。微软、Alphabet和亚马逊将报告强劲的AI收入,因为它们卖芯片、云和模型。它们的数字并不能表明客户是否获得了回报。签合同的公司才是值得关注的。
选民也持怀疑态度。CNBC 10月24日发布的一项民调发现,63%的注册选民反对在其住所10英里范围内建设AI数据中心,仅13%支持。
如果超大规模云厂商本周再次提高支出,风险在于那些借钱出租芯片的公司,如CoreWeave,以及那些必须让支出产生回报的客户。
关键要点
- 阅读资本开支指引,而不只是AI收入。微软和Alphabet在周三美联储决议后发布财报,亚马逊周四发布。
- 报价现在决定你默认获得哪个AI模型。Replit将OpenAI设为其默认,其客户必须刻意选择Claude。
- 关注信贷价格,以捕捉AI基础设施麻烦的早期迹象。CoreWeave债务的保险定价处于困境水平。
- 将AI的候选人选择视为模型的反映,而非建议。Grok和GPT在超过90%的情况下选择了相反政党。
率先发现
AI Weekly在媒体覆盖之前发布的一手研究。
- 一种新方法通过在你说话时思考,将语音助手等待时间缩短36%。Heartbeat Thinking在用户仍在说话时运行推理和工具调用,而不是等待沉默。在FDBench v3基准上,它多覆盖了16个百分点的所需工具调用,并将最终答案的平均等待时间缩短了36.4%。它还在另外两个基准上有所改善。阅读我们关于该论文的页面。
值得一读
- 被解雇的OpenAI研究人员在解雇后首次接受采访:Mikita Balesni和Tomek Korbak告诉NPR,OpenAI从未以书面形式告知他们所称泄露的信息是什么。
- Character.AI在破产前是一场恐怖秀:Futurism发现关于自杀、自残和大规模枪击的机器人被提供给未成年人。截至8月,该公司每月亏损约800万至900万美元,迪士尼正牵头提出收购的初始报价。
- AI墓地:TechCrunch持续更新的2026年关停的AI产品和初创公司名单,包括OpenAI的ChatGPT Atlas浏览器和Sora社交应用。
等等,什么?
- ChatGPT在用真实漫画家的名字给假《纽约客》漫画署名。Nieman Lab发现ChatGPT在其生成的漫画上添加了真实《纽约客》漫画家的签名,包括Harry Bliss和Emily Flake。
值得一看
AI从业者当下正在转发的视频——在AI TV上精选。
| AI热潮对公司来说太昂贵了 404 Media | |
| Black Hat USA 2026 | “爆炸性”新闻:OpenAI与Hugging Face事件 Black Hat |
本周投票
本周什么最能告诉你AI热潮的走向?
上周,325位读者投票:
如果明天一个AI工具在你公司的软件中发现了100个漏洞,会发生什么?
本周什么最能告诉你AI热潮的走向?
下次见,
Alexis
英文来源:
This issue looks ahead. On Wednesday the Fed announces its rate decision, then Microsoft and Alphabet report after the close. Amazon follows on Thursday. Investors want to know whether AI spending is still paying off, and the credit market is already nervous: insurance against a CoreWeave default now costs more than 10% a year. Anthropic is expected to file its IPO paperwork publicly any day. Moonshot says Kimi K3.1 Pro's open weights arrive November 4. Israel votes on Tuesday after a campaign full of AI deepfakes, and the US midterms are a week from Tuesday. In the middle: so far, AI has paid off for the companies that sell it, not the ones that buy it.
Sponsor
Your team uses AI on client work. Give them a server that keeps nothing.
AI agents that research, draft, and analyze for your team, running Anthropic's Claude on an isolated server only you can reach. Processed in memory, never stored. Set up in minutes, from $120 a month.
In the Wild
What's trending in AI right now, from the app charts to the community feeds. The full roundup is in In the Wild.
- An AI room redesign app is climbing fast on the Graphics & Design chart. Maison AI lets you pick a room, a style and a color palette, and it generates a redesign of your interior, exterior or garden. It is free to download.
- xAI's Grok Bot is still climbing the Productivity chart. Grok Bot lets you message AI agents from your phone like coworkers.
- An app for people who run their own AI server is moving up the Productivity chart. Conduit is a phone app for Open WebUI, the self-hosted chat interface. It does not run models itself. You connect it to your own server or to a model provider.
- A study app that turns your notes into quizzes is climbing in Education. Gizmo turns scanned notes and imported material into flashcards and quizzes, with spaced repetition to help you remember.
- AI assistants that live in your text messages are multiplying. TechCrunch rounded up more than 20 of them. Instinct is valued at $10 billion after raising $1 billion in September, and Poke was the first AI agent Apple approved for its Messages for Business platform.
Quick Hits
The Lab Gladiator Era - Wednesday is the big day for AI spending numbers. Alphabet and Microsoft report on Wednesday, October 28, and Amazon reports after the close on Thursday, CNBC reported. The Fed's two-day meeting also ends on Wednesday. Watch what they say about how much they will spend on data centers next year, and how much of their cloud growth comes from AI.
- Anthropic's public IPO filing could land this week. Bloomberg reported on October 22 that Anthropic is expected to file its IPO paperwork publicly in the coming days. Anthropic filed confidentially on June 1. Under SEC rules, the public filing has to be out at least 15 days before the company starts pitching to investors, so its date tells you how soon a listing can happen.
- Replit made OpenAI its default and moved away from Anthropic. Replit CEO Amjad Masad said "I think ChatGPT is better now" and made OpenAI the default for Replit Agent, citing cost, partnership and performance together. Customers can still pick Claude, but most users stick with the default, so Replit's spending on Anthropic could fall sharply.
- Kimi K3.1 Pro's open weights are due November 4. Moonshot says it will release the weights under an open license, VentureBeat reported. Artificial Analysis ranks the model fourth overall, with a score of 55 on its Intelligence Index, up from 44 for K3. Screenshots posted on X appear to tell influencers not to label paid Kimi posts. No one has confirmed who wrote them, and Moonshot has not responded.
The AI Capex Tax - The credit market is pricing real risk into the companies that rent out AI chips. CNBC reported that Nvidia has $7 billion invested directly in AI cloud providers and $164.5 billion in guarantees and commitments, $171.5 billion in total. Five-year insurance against a CoreWeave default costs about 1,082 basis points, CNBC reported, and anything above 1,000 is usually read as distress. CoreWeave, Nebius and IREN shares are each down at least 30% this month.
- Fancier coding agent tools can triple the cost without better results. Shandong University researchers ran Claude Code, OpenCode and mini-SWE-agent with five models on nearly 500 SWE-Bench Verified tasks, VentureBeat reported. Success rates were statistically the same, but the most elaborate setup cost up to three times as much per task.
The Year Governments Got Serious - Israel votes on Tuesday after a campaign full of AI videos. Netanyahu shared an AI video of his rival Gadi Eisenkot, and Finance Minister Smotrich posted one showing the Supreme Court president being dragged from a protest, Al Jazeera reported. Haaretz found dozens of bots built to attack Iran's government had been switched to backing Likud.
- Forced to pick one candidate, Grok and GPT picked opposite parties. In a study of nearly 500 US midterm races, Grok 4.7 picked the Republican 92% of the time and GPT-6.1 Sol picked the Democrat 97% of the time, TechCrunch reported. The researchers used the models' APIs. The consumer apps refused to choose, and OpenAI says the API does not reflect how people use ChatGPT. The midterms are on November 3.
AI has paid off for the sellers, not yet the buyers
Apollo's chief economist Torsten Slok has been making the same point all year. Citing Bloomberg and Macrobond data, he told Fortune that profit margins at the seven biggest tech companies rose from about 15% to about 25% between early 2023 and early 2026. Margins at the other 493 companies in the S&P 500 stayed around 10%. "The key issue is the length of the ROI runway outside the tech sector," he said, and companies will slow their AI spending if they don't see a return quickly.
That is the question behind this week's earnings. Microsoft, Alphabet and Amazon will report strong AI revenue, because they sell the chips, the cloud and the models. Their numbers do not show whether their customers are getting a return. The companies that sign the contracts are the ones to watch.
Voters are skeptical too. A CNBC poll published October 24 found 63% of registered voters oppose an AI data center within 10 miles of their home, and only 13% support one.
If the hyperscalers raise spending again this week, the risk sits with the companies that borrow to rent out the chips, like CoreWeave, and with the customers who have to make the spending pay.
Key Takeaways - Read the capex guidance, not just the AI revenue. Microsoft and Alphabet report on Wednesday after the Fed decision, and Amazon reports on Thursday.
- Price quotes now decide which AI model you get by default. Replit moved to OpenAI as its default, and its customers have to choose Claude on purpose.
- Watch credit prices for early signs of trouble in AI infrastructure. Insurance on CoreWeave's debt is priced at distress levels.
- Treat AI candidate picks as a reflection of the model, not advice. Grok and GPT picked opposite parties more than 90% of the time.
Found First
Primary research AI Weekly published before the press covered it. - A new method cuts voice assistant wait times 36% by thinking while you talk. Heartbeat Thinking runs reasoning and tool calls while the user is still speaking, instead of waiting for silence. On the FDBench v3 benchmark it covered 16 percentage points more of the needed tool calls and cut the average wait for the final answer by 36.4%. It also improved on two more benchmarks. Read our page on the paper.
Worth Reading - Fired OpenAI researchers give their first interview since their dismissal: Mikita Balesni and Tomek Korbak told NPR that OpenAI never told them in writing what information it says they leaked.
- Character.AI was a horror show before it went bankrupt: Futurism found bots about suicide, self-harm and mass shootings served to minors. The company was losing about $8 million to $9 million a month as of August, and Disney is leading the opening bid to buy it.
- The AI graveyard: TechCrunch's running list of AI products and startups that shut down in 2026, including OpenAI's ChatGPT Atlas browser and the Sora social app.
Wait, What? - ChatGPT is signing fake New Yorker cartoons with real cartoonists' names. Nieman Lab found ChatGPT adding the signatures of real New Yorker cartoonists, including Harry Bliss and Emily Flake, to cartoons it generated.
Worth Watching
The videos AI practitioners are passing around right now — curated on AI TV.
| The AI Boom Is Too Expensive For Companies 404 Media | |
| Black Hat USA 2026 | The 'Breaking' News: The OpenAI–Hugging Face Incident Black Hat |
This week's poll
What will tell you most about the AI boom this week?
Last week, 325 of you voted:
If an AI tool found 100 flaws in your company's software tomorrow, what would happen?
What will tell you most about the AI boom this week?
Until next time,
Alexis
文章标题:AI周报第541期:科技巨头周三公布AI支出,CoreWeave债务陷入困境
文章链接:https://news.qimuai.cn/?post=5403
本站文章均为原创,未经授权请勿用于任何商业用途