内容来源：https://aiweekly.co/issues/microsoft-proves-it-doesnt-need-openai-alphabet-raises-85b

内容总结：

微软“脱钩”OpenAI：自研模型亮相，AI信任危机与资本狂热并行

在微软年度开发者大会Build上，该公司正式推出自研AI模型系列，包括推理模型MAI-Thinking-1，旨在与OpenAI和Anthropic竞争。这一举动标志着微软不再完全依赖OpenAI，两者关系从最大投资者转向公开竞争者。与此同时，佛罗里达州总检察长对OpenAI及CEO山姆·奥特曼本人提起诉讼，指控其将利润置于安全之上，并要求奥特曼承担个人责任。这是美国首例州级针对OpenAI的诉讼，涉及ChatGPT与多起暴力事件的关联。

AI信任问题持续发酵。研究人员披露的“TrustFall”漏洞显示，Claude Code、Gemini CLI、Cursor和GitHub Copilot等四款AI编程助手可被诱导执行恶意代码，部分场景甚至无需用户交互。为此，Workday发布“Agent Passport”产品，用于对AI代理进行背景审查与风险监控，以防提示注入、越狱攻击和目标劫持。该产品的诞生标志着企业已不再默认信任AI代理的安全性。

资本层面，Alphabet创纪录融资850亿美元用于AI基础设施建设，这是历史上最大规模的股权融资。五大美国超大规模云服务商今年预计向AI基础设施投入约7250亿美元，同比增长77%。美联储已正式将AI列为系统性金融风险之一。然而，公众对AI数据中心的反对率升至71%，最年轻的软件开发者就业率下降近20%，AI正在取代初级工程师从事的模板化工作。

行业分析师指出，AI投资规模已大到不容忽视。摩根士丹利警告AI“芯片通胀”正传导至消费价格，软银因债务融资押注OpenAI引发流动性担忧，布鲁克菲尔德则推出500亿美元AI基础设施计划。当前的核心问题不再是市场崩盘与否，而是首席财务官们开始追问：如果巨额投入无法带来相应回报，哪项赌注将被最先削减？

中文翻译：

微软利用自家开发者大会证明其可以脱离OpenAI独立运作，佛罗里达州总检察长对OpenAI提起诉讼并针对萨姆·奥尔特曼个人追责，研究人员及Workday新产品明确表明目前无人信任AI智能体，而就在美联储将AI列为系统性风险的同一周，Alphabet创纪录地筹集了850亿美元。资金流动的速度远快于信任的建立。

更多来自AI周报
更多信号，更少噪音——自主选择频道。
您正在阅读的是每周简报。以下是其他追踪报道的方式——所有渠道均免费，且可随时退出。

-
→ 浏览16项深度分析
每周专题通讯：生成式AI、机器学习、商业AI、机器人技术、前沿研究、地缘政治、医疗健康等。浏览全部16项深度分析 →

-
→ AI重大突发警报
当重大事件发生（如600亿美元收购、监管机构紧急会议、前沿模型泄露），订阅者会在数小时内收到通知。通常每天0-2封邮件。订阅突发警报 →

-
→ AI今日新闻（实时）
实时仪表盘随新闻扫描更新：过去48小时的评分报道、每周实体动向、以及涵盖113家AI公司、人物和话题的季度趋势线。打开AI今日新闻 →

速览

实验室角斗士时代

• 微软利用其Build大会证明可以脱离OpenAI。它推出了一个自研AI模型系列，旨在与OpenAI和Anthropic竞争，由推理模型MAI-Thinking-1领衔，据称该模型从零开始训练，未使用任何蒸馏技术。OpenAI的最大投资者如今正公开构建其替代方案。
• Suno在持续应对版权诉讼之际融资约4亿美元。这家AI音乐初创公司新一轮估值达54亿美元，尽管针对它的版权诉讼仍在进行。投资者正直接无视诉讼进行注资。

各国政府认真对待AI之年

• 佛罗里达州起诉OpenAI并针对萨姆·奥尔特曼个人。该州总检察长提起一项史无前例的案件，指控该公司将利润置于安全之上，并寻求追究奥尔特曼的个人责任，声称ChatGPT与多起暴力事件有关。这是美国首个起诉OpenAI的州。
• 欧盟推动将美国云服务排除出其公共部门。欧盟委员会公布了一项技术主权计划，旨在减少公共部门对美国云提供商的依赖，这是欧洲争取掌控自身AI和数字基础设施的广泛努力的一部分。

AI供应链受困

• 信任对话框漏洞暴露四款AI编程智能体。研究人员的TrustFall报告显示，Claude Code、Gemini CLI、Cursor和GitHub Copilot在处理文件夹信任时可能被诱骗运行恶意代码，而在CI运行环境中，Claude Code甚至可在无需任何交互的情况下执行代码。
• Workday现推出一项为AI智能体做背景调查的服务。它本周推出了Agent Passport，以思科为测试合作伙伴，在生产前对每个智能体（自研或第三方）进行测试和持续监控，防范提示注入、越狱、目标劫持等风险，并依据MITRE ATLAS威胁框架评分。早期版本将于今年晚些时候推出。该产品的存在是因为企业已不再假设其智能体是安全的。

AI资本支出税

• Alphabet创纪录地筹集850亿美元用于AI基础设施建设。此次股权发行是史上规模最大的，超额认购的批次被指定用于Alphabet AI战略背后的数据中心和算力。
• 最年轻软件开发者就业率下降约20%。斯坦福2026年AI指数报告显示，22至25岁软件开发者的岗位自2024年以来下降近20%，而同一公司30岁以上开发者人数却在增长。AI正在取代初级员工原本负责的模板化、明确指定的工作。

基建规模已庞大到不容忽视
回顾本周的动态，关键数字在于支出。Alphabet的850亿美元融资是最响亮的信号，但它只是众多信号之一。美国五大超大规模云服务商今年将向AI基础设施投入约7250亿美元，较2025年增长约77%。美联储现已将AI列为其金融稳定的首要风险之一。
仅本周，摩根士丹利警告AI“芯片通胀”正开始传导至消费价格；软银以债务为赌注押注OpenAI引发了新的流动性担忧；Brookfield推出了500亿美元的AI基础设施计划。甚至公众态度也在转变：对新数据中心的反对率已跃升至71%。
正是这种支出将所有问题联系在一起。微软之所以能负担自有模型，是因为同样的繁荣为其提供了资金。企业之所以购买治理AI智能体的工具，是因为它们正以繁荣所要求的速度部署这些智能体。账单最先落到的地方是薪资单，最年轻工程师招聘放缓已可量化。
这些都不需要市场崩盘才显重要。它只需要每位财务官现在都在大声问的那个问题：如果支出开始不能产生超过成本的回报，哪个赌注会先被砍掉？

核心要点

• 前沿：微软现在可以在同一平台上向您销售其自有模型或OpenAI的模型。这使得其最大的合作伙伴也成为其最新的竞争对手，并结束了OpenAI模型曾是Azure上唯一前沿选择的时代。
• 政策：州总检察长对CEO个人追责，将AI责任从公司资产负债表转移到了创始人身上。每个前沿实验室的法律团队本周都将重新审阅其安全披露文件。
• 安全：企业在信任智能体之前正购买智能体治理方案，因为研究人员不断展示智能体可在极少或无需交互的情况下运行恶意代码。将每个拥有代码仓库访问权限的编程智能体视为已暴露，直至其通过审查。
• 资本支出：基建的第一张账单并非GPU账单，而是初级工程岗位，并且日益是公众的耐心。关注初级岗位招聘和本地数据中心争端，而不仅仅是财报。

值得一读

• 谷歌新款开放权重Gemma 4：一款12B参数模型，旨在单块16GB GPU上运行。
• Meta的Business Agent在WhatsApp、Instagram和Messenger上推出，Meta正推动智能体面向小企业。
• Anthropic关于AI驱动的网络威胁：该实验室关于攻击者如何使用前沿模型的数据。
• 微软量子计算关于Majorana 2：将其实用量子计算时间线提前至2029年的芯片。

值得一看
AI从业者目前正在传阅的视频——由AI TV策划。
| Ed Zitron在彭博播客中表示：Anthropic，OpenAI不应被允许IPO | |
| AI泡沫：‘商业白痴’终于看到了无上限AI的负面影响 | Ed Zitron，The Tech Report | |
| ChatGPT的“杀人许可证” | Caelan Conrad |

本周投票
微软的自有模型真能让它脱离OpenAI吗？
上周，172位读者参与了投票：
Anthropic刚刚启动了首个前沿实验室IPO倒计时。美国证券交易委员会应该强制公开披露的最重要事项是什么？
微软的自有模型真能让它脱离OpenAI吗？
请回复告诉我，本周你是否会让AI智能体接近你的代码仓库。
——亚历克西斯

英文来源：

Microsoft used its own developer conference to show it can live without OpenAI, Florida's attorney general sued OpenAI and went after Sam Altman personally, researchers and a new Workday product made plain that nobody trusts AI agents yet, and Alphabet raised a record $85 billion the same week the Fed flagged AI as a systemic risk. The money is moving faster than the trust.
Get more from AI Weekly
More signal, less noise — pick your channels.
You're reading the weekly brief. Below are the other ways to follow the story — every channel free, easy to leave.

→ Explore 16 deep divesWeekly topic-specific newsletters: Generative AI, Machine Learning, AI in Business, Robotics, Frontier Research, Geopolitics, Healthcare, and more.Browse all 16 deep dives →

→ Breaking AI alertsWhen something major breaks (a $60B acquisition, a regulator's emergency meeting, a frontier model leak), alert subscribers know within hours. Typically 0-2 emails per day.Get breaking alerts →

→ AI News Today (live)Live dashboard updated as the scanner finds news: scored stories from the last 48 hours, weekly entity movers, and quarterly trend lines across 113 AI companies, people, and topics.Open AI News Today →
Quick Hits
The Lab Gladiator Era

• Microsoft used its own Build conference to show it can live without OpenAI. It launched a family of in-house AI models built to take on OpenAI and Anthropic, led by the reasoning model MAI-Thinking-1, which it says was trained from scratch with zero distillation. OpenAI's largest backer is now openly building the alternative to it.
• Suno raised about $400 million while still fighting copyright lawsuits. The AI-music startup's new round values it at $5.4 billion, even as the copyright lawsuits against it continue. Investors are funding straight through the litigation.
  The Year Governments Got Serious
• Florida sued OpenAI and went after Sam Altman personally. The state attorney general filed a first-of-its-kind case accusing the company of putting profit over safety and seeking to hold Altman personally liable, alleging ChatGPT was tied to several violent incidents. It is the first US state to sue OpenAI.
• The EU moved to push US cloud out of its public sector. The European Commission unveiled a tech-sovereignty package aimed at cutting the public sector's reliance on American cloud providers, part of a broader European push to control its own AI and digital infrastructure.
  AI Supply Chain Under Siege
• A trust-dialog flaw exposes four AI coding agents. Researchers' TrustFall disclosure shows Claude Code, Gemini CLI, Cursor, and GitHub Copilot can be tricked into running a malicious repository's code through how they handle folder trust, and on CI runners Claude Code runs it with no interaction at all.
• Workday now sells a way to background-check your AI agents. It unveiled Agent Passport this week, with Cisco as a testing partner, to test and continuously monitor every agent (its own or third-party) for risks like prompt injection, jailbreak, and goal hijacking before production, scored against the MITRE ATLAS threat framework. Early access lands later this year. The product exists because companies have stopped assuming their agents are safe.
  The AI Capex Tax
• Alphabet raised a record $85 billion to fund its AI build-out. The equity offering is the largest in history, sold in oversubscribed tranches and earmarked for the data centers and compute behind Alphabet's AI push.
• Employment for the youngest software developers fell about 20%. Stanford's 2026 AI Index reports software-developer jobs for workers aged 22 to 25 down nearly 20% since 2024, while headcount for developers over 30 at the same firms grew. AI is taking the boilerplate and well-specified work junior hires used to do.
  The build-out is now too big to look away from
  Step back from the week's launches and the number that matters is the spend. Alphabet's $85 billion raise was the loudest signal, but it sits inside a wall of them. The five biggest US hyperscalers are on track to pour roughly $725 billion into AI infrastructure this year, up about 77% from 2025. The Federal Reserve now lists AI among its top risks to financial stability.
  This week alone, Morgan Stanley warned that AI "chipflation" is starting to reach consumer prices, SoftBank's debt-fueled bet on OpenAI drew fresh liquidity-crunch concern, and Brookfield rolled out a $50 billion AI-infrastructure plan. Even the public is turning: opposition to new data centers jumped to 71%.
  That spend is what ties this issue together. Microsoft can afford its own models because the same boom funds them. Enterprises are buying tools to police AI agents because they are deploying those agents at the pace the boom demands. And the first place the bill lands is payroll, where a hiring slowdown for the youngest engineers is already measurable.
  None of this needs a crash to matter. It only needs the question every finance chief is now asking out loud. If the spending does not start returning more than it costs, which bet gets cut first?
  Key Takeaways
• Frontier: Microsoft can now sell you its own model or OpenAI's on the same platform. That makes its biggest partner also its newest rival, and it ends the era when OpenAI's models were the only frontier option on Azure.
• Policy: A state attorney general going after a CEO personally moves AI liability from the corporate balance sheet to the founder. Every frontier lab's legal team will re-read its safety disclosures this week.
• Security: Enterprises are buying agent-governance before they trust agents, because researchers keep showing the agents run hostile code with little or no interaction. Treat every coding agent with repository access as exposed until it is vetted.
• Capex: The build-out's first invoice is not a GPU bill, it is the entry-level engineering job and, increasingly, the public's patience. Watch junior hiring and local data-center fights, not just earnings.
  Worth Reading
• Google's new open-weight Gemma 4: a 12B model aimed at running on a single 16GB GPU.
• Meta's Business Agent rolls out across WhatsApp, Instagram, and Messenger as Meta pushes agents at small businesses.
• Anthropic on AI-enabled cyber threats: the lab's own data on how attackers are using frontier models.
• Microsoft Quantum on Majorana 2: the chip that pulled its useful-quantum timeline in to 2029.
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  | Anthopic, OpenAI Should Not Be Allowed to IPO, Says Ed Zitron Bloomberg Podcasts | |
  | AI Bubble: ‘Business idiots’ are finally seeing the downside of uncapped AI | Ed Zitron The Tech Report | |
  | chatgpt's license to kill you Caelan Conrad |
  This week's poll
  Will Microsoft's own models actually pull it away from OpenAI?
  Last week, 172 of you voted:
  Anthropic just started the first frontier-lab IPO clock. What's the most important thing the SEC should force out into public disclosure?
  Will Microsoft's own models actually pull it away from OpenAI?
  Reply and tell me whether you would let an AI agent near your repo this week.
  — Alexis