内容来源：https://aiweekly.co/issues/openai-lost-three-things-in-five-days

内容总结：

快讯速览

• 马斯克就OpenAI案出庭作证：“如果我们允许公然掠夺……”——陪审团遴选工作于周一完成，马斯克周二就这起价值1340亿美元的诉讼出庭，要求将OpenAI资产返还给非营利实体并罢免奥特曼。法庭直播每时每刻都在跟进。此案将为未来十年AI领域何为“营利性转化”确立法律模板。（来源：CNBC）
• OpenAI未达营收目标——拖累甲骨文及芯片股下跌——其公布的业绩低于甲骨文3000亿美元算力合约所依据的预测值。消息传出后，甲骨文及AI芯片板块应声下挫。“OpenAI营收为所有人资本开支背书”的交易逻辑已变为双向博弈。（来源：CNBC）
• OpenAI将模型引入AWS，终结微软独家协议——悄无声息的官宣。此前，Azure是唯一可运行OpenAI模型的生产环境。如今AWS也已加入。“微软掌控OpenAI分销渠道”的论断就此瓦解——而萨蒂亚甚至未获得一份联合声明。（来源：CNBC）
• DeepSeek大幅降价，中国价格战升级——DeepSeek再次下调API定价，这已是本季度第三轮调整。中国超大规模云厂商竞相将推理成本压向零。支撑西方AI估值体系的定价地板已向东移动数千公里。（来源：彭博社）
• 《华尔街日报》：AI巨额支出正蚕食大科技公司员工规模——资本开支正通过缩减人头来筹集。微软、Meta、亚马逊和Alphabet在过去六个月内均进行了裁员，同时上调了AI支出指引。“AI资本开支将带动招聘增长”的说法已被数据证伪。（来源：华尔街日报）
• 谷歌员工敦促皮查伊拒绝涉密五角大楼AI项目——数百名谷歌员工签署公开信，要求CEO放弃一项机密军事AI合同。这是自“Maven项目”以来谷歌员工在国防AI领域的首次大规模抗议。谷歌原以为已在2018年平息的内部分歧与五角大楼之间的紧张关系再度浮现。（来源：彭博社）
• 白宫起草指南，拟绕过Anthropic对新AI模型的风险警告——据Axios报道。该政策将允许联邦机构采购前沿模型，即便开发者自身的安全评估将其标记为高风险。Anthropic的“负责任的扩展政策”本是首个具有监管约束力的自愿安全框架，此举实则将其架空。（来源：WSAU/Axios）
• 科罗拉多州全美首部AI法：法官裁定暂不可执行——2024年科罗拉多州AI法案原定对开发者及部署者施加基于风险的义务，现已被搁置至2027年实施审查。这个率先出台AI监管的州已悄然退步。（来源：科罗拉多政治报）

当定价者沦为被定价者

五天内三条OpenAI重磅新闻，连起来读，景象截然不同。

庭审。 马斯克要求将OpenAI资产返还非营利实体的1340亿美元索赔，远不止私人恩怨——这是特拉华州衡平法院必须回答的问题：当原始使命是“为了让全人类拥有通用人工智能”时，从501(c)(3)非营利组织向公益公司转化在法律上需要满足什么条件？无论最终判决如何，已展开的证据开示程序将为所有在成长过程中接触过非营利实体的AI公司定下披露标准。

营收未达标。 OpenAI公布的业绩低于甲骨文3000亿美元算力合约所依据的预测值。甲骨文股价下跌，芯片板块随之下挫。市场正在发现：“OpenAI营收→甲骨文资本开支→英伟达/AMD营收→AI基础设施叙事”是一条链条，而链条总在最薄弱处断裂。两年来，OpenAI一直是拉动所有人业绩的需求引擎。本周，市场已开始为另一种可能性定价。

AWS交易。 直到周一，能在生产环境中运行OpenAI模型的唯一平台是Azure。那是微软价值130亿美元的保险单。AWS交易——悄无声息地宣布，没有联合新闻发布会——终结了这一切。OpenAI如今拥有两个分销渠道，而微软必须为其中一个渠道展开竞争。曾让萨蒂亚成为AI领域权力掮客的杠杆已不复存在。

三件事的关联所在。 三年来，OpenAI一直是那个设定价格的企业——为人才定价、为算力合约定价、为分销条款定价、为AI构建模式定价。而本周，这三项定价权在法庭、市场和云采购中分别受到挑战。没有一项尘埃落定，但所有环节都已变得可谈判。

DeepSeek的价格战和《华尔街日报》关于员工规模的报道是二阶信号。中国推理市场正向免费狂奔；西方资本开支则靠裁减人头筹集。定价权的时代结束，并非因为一场官司——而是因为投入要素（算力供给、人才供给、分销合同）同时变得更便宜或更具竞争性。

核心要点

• OpenAI的定价权已是2024年的故事。任何基于OpenAI营收线性增长以支撑甲骨文/英伟达/微软资本开支承诺的AI资本开支模型，都需要准备B计划。市场首次为此情景进行了定价。
• 微软独家协议是保险，而非战略。AWS交易证明，OpenAI一旦有能力就会走向多云架构。未来18个月内宣布的所有云-AI合作伙伴关系，都应带着这个注释来解读。
• Anthropic安全审查被架空是监管层面的重头戏。白宫愿意推翻开发者自身制定的负责任的扩展政策，表明华盛顿已将“自愿AI安全框架”视为采购障碍而非优势。关注Anthropic是公开抵制还是悄然妥协。
• 科罗拉多州退步+白宫绕过安全框架=监管钟摆在一个季度内完成转向。六个月前，舆论基调是“各州将填补联邦真空”。本周则恰恰相反。依靠监管护城河的企业需要寻找新模式。

值得一读

• 亚马逊Kiro：自主前沿智能体——亚马逊在AI智能体层面的低调押注。值得与OpenAI/AWS协议一同阅读——AWS如今既拥有前沿模型合作伙伴，也拥有自研智能体运行时环境。
• Anthropic：面向创意工作的Claude——一家即将被解除监管武器的公司推出的垂直产品。其战略时机本身就是一种表态。
• 英伟达：Nemotron-3 Nano Omni多模态智能体——来自英伟达研究部门的小模型多模态发布。“Nano”尺寸意味深长：这针对的是中国价格战所瞄准的边缘设备。
• 中国暂停发放新的L4级自动驾驶许可——在百度Apollo Go超过100辆无人驾驶出租车在武汉因系统故障被困之后。中国首次对大规模自动驾驶部署实施硬性监管暂停。
• Hugging Face推理服务器关键漏洞——如果你运行HF推理服务，本周请即刻修补漏洞。模型服务供应链正开始显现2017年包管理器供应链的影子。

本周投票
OpenAI的定价权：被击碎、被削弱，还是未受影响？

感谢阅读《AI周刊》。请将本期内容转发给一位需要了解OpenAI事件全貌的朋友。

中文翻译：

快速要闻

• 马斯克在OpenAI庭审中作证：“如果我们允许公然掠夺……”——陪审团遴选周一结束，马斯克周二就这起1340亿美元诉讼出庭作证，要求将OpenAI资产归还非营利机构并罢免阿尔特曼。庭审直播每小时更新。此案将为未来十年AI领域“营利性转型”的法律界定树立模板。CNBC
• OpenAI未达收入目标——拖累甲骨文及芯片股下挫——其公布的数据低于甲骨文3000亿美元算力合约所依赖的预测值。甲骨文及AI芯片板块受此消息影响同步下跌。“OpenAI收入支撑所有人资本支出”的交易逻辑如今已变为双向博弈。CNBC
• OpenAI将模型引入AWS，终结微软独占期——低调官宣。此前，Azure是唯一可运行OpenAI模型的生产环境。如今AWS同样可用。“微软掌控OpenAI分发渠道”的论断宣告终结——而萨提亚甚至没收到一份新闻稿。CNBC
• DeepSeek再度降价，中国价格战升级——DeepSeek再次下调API定价，为本季度第三次。中国超大规模云厂商正竞相将推理成本压至零。曾支撑西方AI估值的定价底线，已向东迁移数千里。彭博社
• 《华尔街日报》：AI豪赌正让科技巨头付出裁员代价——资本支出由员工薪资填补。微软、Meta、亚马逊和Alphabet在过去六个月均缩减人力，同时上调AI支出指引。“AI资本支出将带动招聘”的说法已被数据证伪。华尔街日报
• 谷歌员工敦促皮查伊拒绝五角大楼机密AI项目——数百名谷歌员工签署公开信，要求CEO退出机密军事AI合同。这是自“完美风暴”项目以来谷歌员工首次因国防AI发起重大抗议。谷歌以为在2018年已压制的内部与五角大楼矛盾再度爆发。彭博社
• 白宫起草指导方针，绕过Anthropic对新型AI模型的风险标记——据Axios报道。该政策将允许联邦机构采购前沿模型，即使开发者自身安全评估将其标记为高风险。Anthropic的“负责任扩展政策”曾是首个具有监管分量的自愿性安全框架，此举将使其失效。WSAU/Axios
• 科罗拉多州首部AI法案：法官裁定暂不可执行——2024年科罗拉多州AI法案原拟对开发者和部署者施加基于风险的义务，现已搁置至2027年实施评估。该州率先推动AI监管的步伐已悄然放缓。科罗拉多政治报

当定价者成为被定价者
五天内发生三件OpenAI大事。合并来看，图景已然不同。
庭审。马斯克要求将OpenAI资产归还非营利机构的1340亿美元诉讼，并非私人恩怨——这是特拉华州衡平法院必须回答的问题：当原始使命是“为人类打造AGI”时，从501(c)(3)非营利组织转型为公益公司需满足何种法律要求。无论判决如何，已启动的证据开示程序将为每一家在成长过程中接触过非营利机制的AI公司界定信息披露标准。
收入未达标。OpenAI公布的数据低于甲骨文3000亿美元算力合约所依据的预测。甲骨文股价下跌，芯片板块同步走低。市场正在发现“OpenAI收入→甲骨文资本支出→英伟达/AMD收入→AI基础设施叙事”是一条链条，而链条会从最薄弱环节断裂。两年来，OpenAI一直是拉动其他公司业绩的需求引擎。本周市场已为另一种可能性定价。
AWS协议。截至周一，运行OpenAI模型的生产环境仅有Azure。那是微软130亿美元的保险单。AWS协议——低调宣布，无联合发布会——终结了这一切。OpenAI现拥有两个分销渠道，而微软必须为其中之一展开竞争。曾使萨提亚成为AI造王者的话语权已不复存在。
三者共同点。三年来，OpenAI一直是那个为人才、算力合约、分销条款以及AI构建模式设定价格的公司。本周，这三项定价权分别在法庭、市场和云计算采购中受到挑战。无一尘埃落定，所有都已成为可议价项。
DeepSeek的价格战和《华尔街日报》关于人力裁减的报道是二阶信号。中国推理服务正向免费冲刺；西方资本支出由人力成本填补。定价权时代的终结并非源于某一场审判——而是因为所有要素（算力供应、人才储备、分销合同）同时变得更廉价或更具争议。

核心要点

• OpenAI的定价权已是2024年的故事。任何基于OpenAI收入线性增长以支撑甲骨文/英伟达/微软承诺的AI资本支出模型，都需要备选方案。市场本周首次为这一情景定价。
• 微软独占权曾是保险，而非战略。AWS协议证明OpenAI一旦具备条件，必然会走向多云。未来18个月内宣布的任何云-AI合作，都应带上这个星号看待。
• Anthropic安全评估被绕过是监管层面的焦点。白宫愿推翻开发者自身RSP评估，表明华盛顿已认定“自愿性AI安全框架”是采购障碍而非优势。需观察Anthropic是会公开抵制还是低调执行。
• 科罗拉多州退缩+白宫绕行=监管钟摆在短短一个季度内摇摆。六个月前的舆论还是“各州将填补联邦监管真空”。本周情况恰恰相反。依赖监管护城河的公司需要新的商业模式。

值得一读

• 亚马逊Kiro：前沿智能体实现自主化——亚马逊在AI智能体层的低调布局。值得与OpenAI/AWS协议一同研读——AWS如今同时拥有前沿模型合作伙伴和自研智能体运行环境。
• Anthropic：面向创意工作的Claude——这家即将被解除监管武装的公司推出垂直产品。战略时机本身就是一种表态。
• 英伟达：Nemotron-3 Nano Omni多模态智能体——英伟达研究部门推出的小型多模态模型。“Nano”规格至关重要：这针对的是中国价格战瞄准的边缘设备。
• 中国暂停发放新L4级自动驾驶许可——起因是百度Apollo Go超100辆无人出租车在武汉因系统故障被困。中国对主要自动驾驶部署实施的首次严格监管暂停。
• Hugging Face推理服务器出现严重CVE漏洞——若您运行HF推理服务，请于本周完成补丁更新。模型服务供应链正显现出类似2017年包管理器供应链的隐患。

本周投票
OpenAI的定价权：出现裂痕、遭受重创、还是保持不变？

感谢阅读《AI周报》。请将本期内容转发给一位需要了解OpenAI事件集群的读者。

英文来源：

Quick Hits

• Musk takes the stand in the OpenAI trial: "If we make it OK to loot..." — Jury selection wrapped Monday; Musk testified Tuesday in the $134B suit to return OpenAI's assets to the nonprofit and oust Altman. Live coverage tracked every hour. The case will set the legal template for what counts as a "for-profit conversion" in AI for the next decade. CNBC
• OpenAI missed revenue targets — Oracle and chip stocks dragged with it — Reported numbers came in below the forecasts that Oracle's $300B compute contract is collateralized against. Oracle and the AI-chip basket sold off in sympathy on the news. The "OpenAI revenue underwrites everyone's capex" trade is now a two-way street. CNBC
• OpenAI brings models to AWS, ending Microsoft exclusivity — Quiet announcement. Until this week, Azure was the only place to run OpenAI models in production. Now AWS too. The "Microsoft owns OpenAI's distribution" thesis is dead — and Satya didn't even get a press release. CNBC
• DeepSeek slashes fees in escalating Chinese price war — DeepSeek cut its API pricing again, the third round this quarter. Chinese hyperscalers are racing each other to zero on inference. The pricing floor that justified Western AI valuations has just moved several thousand miles east. Bloomberg
• WSJ: the AI splurge is now costing Big Tech its workforce — Capex is being funded by headcount. Microsoft, Meta, Amazon and Alphabet have all run reductions in the last six months while raising AI spending guidance. The narrative that AI capex would be additive to hiring is dead in the data. WSJ
• Google staff urge Pichai to refuse classified Pentagon AI work — Open letter signed by hundreds of Googlers asking the CEO to walk away from a classified military AI contract. First major employee revolt at Google over defense AI since Project Maven. The internal-vs-pentagon tension Google thought it had buried in 2018 is back. Bloomberg
• White House drafts guidance to bypass Anthropic's risk flag for new AI models — Per Axios. The policy would let federal agencies procure frontier models even if the developer's own safety evaluation flags them as high-risk. Anthropic's Responsible Scaling Policy was the first voluntary safety framework with regulatory weight; this strips it. WSAU/Axios
• Colorado's first-in-the-nation AI law: judge rules it can't be enforced yet — The 2024 Colorado AI Act, which would have imposed risk-based obligations on developers and deployers, is on hold pending a 2027 implementation review. The state that moved first on AI regulation has quietly stepped back. Colorado Politics
  When the price-setter becomes the price-taker
  Three OpenAI stories landed in five days. Read them together and the picture changes.
  The trial. Musk's $134B claim to return OpenAI's assets to the nonprofit isn't just a feud — it's a question Delaware Chancery will have to answer about what a 501(c)(3)-to-PBC conversion legally requires when the original mission was "AGI for humanity." Whatever the verdict, the discovery already happening will define the disclosure standard for every AI company that touched a nonprofit on the way up.
  The revenue miss. OpenAI's reported numbers undershot the forecasts that Oracle's $300B compute deal is structured against. Oracle sold off. The chip basket sold off. The market is discovering that "OpenAI revenue → Oracle capex → Nvidia/AMD revenue → AI infrastructure narrative" is a chain, and chains break at the weakest link. For two years OpenAI was the demand pull that made everyone else's numbers work. This week the market priced in the alternative.
  The AWS deal. Until Monday, the only place to run OpenAI models in production was Azure. That was Microsoft's $13B insurance policy. The AWS deal — announced quietly, no joint press conference — ends it. OpenAI now has two distribution channels and Microsoft has to compete for one of them. The leverage that made Satya the kingmaker of AI is gone.
  What ties them together. OpenAI spent three years being the company that set the price — for talent, for compute deals, for distribution terms, for the model of how AI gets built. This week each of those three pricing powers got contested in court, in markets, and in cloud procurement. None of them are settled. All of them are now negotiable.
  DeepSeek's price war and the WSJ workforce piece are the second-order signals. Chinese inference is racing to free; Western capex is funded out of headcount. The price-setting era isn't ending because of one trial — it's ending because the inputs (compute supply, talent supply, distribution contracts) all got cheaper or more contested at once.
  Key Takeaways
• OpenAI's pricing power is a 2024 story. Anyone modeling the AI capex stack on the assumption that OpenAI revenue grows linearly to underwrite Oracle / Nvidia / Microsoft commitments needs a Plan B. The market just priced in that scenario for the first time.
• Microsoft exclusivity was insurance, not strategy. The AWS deal proves OpenAI was always going to multi-cloud the moment it could. Every cloud-AI partnership announced in the next 18 months should be read with this asterisk.
• The Anthropic safety bypass is the regulatory story. A White House willing to override a developer's own RSP signals the Beltway has decided "voluntary AI safety frameworks" are a procurement obstacle, not a feature. Watch whether Anthropic publicly resists or quietly ships.
• Colorado retreating + White House bypass = the regulatory pendulum swung in one quarter. Six months ago the conversation was "states will fill the federal vacuum." This week says the opposite. Companies banking on regulatory moats need a new model.
  Worth Reading
• Amazon Kiro: Frontier Agents Go Autonomous — Amazon's quiet bet on the agent layer of the stack. Worth reading alongside the OpenAI/AWS deal — AWS now has both a frontier-model partner and a homegrown agent runtime.
• Anthropic: Claude for Creative Work — A vertical product launch from the company about to be regulatorily disarmed. The strategic timing is its own statement.
• Nvidia: Nemotron-3 Nano Omni multimodal agents — Small-model multimodal release from Nvidia's research arm. The "Nano" sizing matters: this is for the edge devices the China price war is targeting.
• China suspends new Level-4 autonomous driving permits — After 100+ Baidu Apollo Go robotaxis stranded in Wuhan during an outage. The first hard regulatory pause on a major Chinese AV deployment.
• Critical CVE in Hugging Face inference servers — Patch this week if you run HF inference. The model-serving supply chain is starting to look like the package manager supply chain did in 2017.
  This week's poll
  OpenAI's pricing power: cracked, dented, or unchanged?
  Thanks for reading AI Weekly. Forward this to one person who needs to see the OpenAI cluster.