内容来源：https://aiweekly.co/issues/481

内容总结：

【AI行业周报：格局剧变的一周，三大转折点重塑赛道】

过去七天，人工智能行业经历了三场地震级变动：Anthropic年度营收预期首次超越OpenAI，Meta放弃开源路线转向闭源，以及行业法律战与地缘冲突全面升级。这标志着AI竞争已从技术竞赛进入商业实力、战略路线与全球治理多维对抗的新阶段。

核心动态：

1. 营收王座易主：Anthropic凭借企业级市场爆发，年度经常性收入预期突破300亿美元，反超OpenAI的240亿美元。其百万美元级客户在两个月内翻倍，并与谷歌、博通锁定3.5吉瓦新一代TPU算力，验证了“安全优先”策略的商业化潜力。
2. Meta开源时代终结：由Alexandr Wang领导的超级智能实验室推出首款闭源模型Muse Spark，宣告公司放弃Llama奠定的开源路线。新战略将仅开源小型模型，顶尖技术转为自用，迫使依赖Llama生态的开发者寻找备选方案。
3. OpenAI内忧外患：首席财务官公开质疑2026年第四季度IPO计划“过于激进”，随后被排除在关键财务会议之外。同时，马斯克诉诸法庭要求罢免奥特曼与布罗克曼，OpenAI则反诉其通过xAI、特斯拉和SpaceX进行“反竞争行为”。
4. 中国突破算力封锁：深度求索公司宣布DeepSeek V4将完全基于华为昇腾芯片训练，成为首个未使用英伟达硬件的万亿参数前沿模型。这证明美国芯片出口管制仅能争取时间，无法阻止中国发展自主AI能力。
5. AI基础设施成军事目标：伊朗革命卫队公开威胁打击OpenAI在阿布扎比建设的Stargate数据中心，首次将AI物理设施置于国家间冲突前沿。数据中心安全已从技术问题升级为国家安全议题。

行业影响：

• 开源生态重组：谷歌趁Meta转向之际，发布完全开源的Gemma 4系列模型，试图填补生态空缺。
• 法律与就业震荡：好莱坞编剧工会达成新协议，获AI训练数据知情权与授权保障；同时全美科技行业裁员潮中，25%岗位削减与AI直接相关。
• 产业资本转向：英特尔加入马斯克主导的200亿美元AI芯片制造联盟；专注物理AI与国防的Eclipse基金募资130亿美元，创领域纪录。

关键转折信号：
Anthropic CEO达里奥·阿莫迪近期公开表示“指数级增长已近尾声”，这为其300亿美元营收数据提供了注脚——行业正从野蛮生长进入差异化竞争与全球治理深水区。与此同时，中国在受限环境下实现技术突破、AI基础设施成为地缘博弈筹码等趋势，预示着行业将面临更复杂的多维挑战。

（本期周期：2026年4月3日-9日）

中文翻译：

一周之内，三场剧变。Anthropic的年度营收预期超越OpenAI——300亿美元对240亿美元——企业需求在不到两个月内使其百万美元级客户翻倍，成为主要推力。Meta在Alexandr Wang的超级智能实验室下推出首个专有模型，放弃了曾定义Llama的开源理念。而围绕AI的法律机器也全速运转：马斯克希望解雇阿尔特曼，OpenAI要求调查马斯克，好莱坞编剧们则刚刚锁定了为期四年的强化AI保护协议。

2026年4月3日 – 9日
Anthropic比OpenAI更赚钱。Meta不再分享其顶尖成果。马斯克和阿尔特曼都想解雇对方。而在阿布扎比的某个地方，有人正在重新评估数据中心的保险费用。

核心要点

• Anthropic现已成为营收最高的AI实验室。 300亿美元年度预期对比OpenAI的240亿美元。推动力来自企业需求，而非消费者。如果你基于Claude构建产品，就是在押注市场领导者。
• Meta扼杀了开源AI的最大倡导者。 Muse Spark转为闭源。混合策略意味着开源只能获得残羹剩饭；顶尖模型将保持封闭。基于Llama开放性构建的开发者需要备选方案。
• OpenAI的IPO陷入真正困境。 当自家首席财务官公开表示上市时间表过于激进，并因此被排除在关键会议之外时，市场将此视为警告，而非内部意见分歧。
• 中国无需美国芯片即可构建前沿AI。 基于华为芯片的DeepSeek V4证明，出口管制只是在争取时间，而非解决问题。
• AI基础设施现已成为军事目标。 伊朗威胁攻击Stargate数据中心意味着，在争议地区建设AI设施时，物理安全已非可选，而是必需。

快讯

• 英特尔加入马斯克的Terafab：与特斯拉、SpaceX及xAI共同投资200亿美元以上建设AI芯片工厂。
• Eclipse融资130亿美元：专注实体AI、机器人及国防领域的最大规模基金诞生。
• Flowise CVSS 10.0漏洞：开源AI智能体构建工具遭主动攻击，超1.2万实例暴露。
• 首例深度伪造定罪：联邦《下架法案》首次获判有罪裁决。
• UnitedHealth投入30亿美元押注AI：2.2万名工程师自动化处理理赔与临床文档。
• 三大巨头联手应对中国：共享情报以阻断通过2.4万个虚假Claude账户进行的模型克隆。
• OpenClaw遭封禁：Claude订阅不再支持第三方智能体，需使用API付费方案。
• AI医疗记录推高成本：保险公司与医疗机构一致认为AI编码导致账单上涨。

🎬 视听首选
达里奥·阿莫迪："指数级增长已近尾声" · Dwarkesh播客 · Spotify收听
→ Anthropic首席执行官阐述为何扩展假说可能已达极限、AI将如何渗透经济，以及监管会否扼杀其效益。理解本周300亿美元营收数据的重要背景。

OpenClaw：击垮互联网的病毒式AI智能体 · Lex Fridman第491期 · YouTube观看
→ Peter Steinberger讲述一个业余项目如何获得18万GitHub星标，并迫使Anthropic封禁第三方智能体访问。以上OpenClaw封禁事件的背景解读。

营收王座易主
Anthropic年度预期达300亿美元，超越OpenAI · 4月7日 · 彭博社
→ 四个月内从90亿跃升至300亿美元。企业需求翻倍——超1000家企业年付百万美元使用Claude——Anthropic通过谷歌和博通锁定了3.5吉瓦次世代TPU算力。这家押注安全的公司正在营收上取胜。

Claude Mythos仅限网络安全领域开放 · 4月7日 · TechCrunch
→ 全球最强大的AI模型已存在，但你无法使用。Mythos通过"玻璃翼项目"与12家合作机构发现数千个零日漏洞。前沿模型或将率先应用于专业领域——网络安全首当其冲，公众版本永无可能。

Meta转向闭源
Meta发布首款专有模型Muse Spark · 4月8日 · CNBC
→ Alexandr Wang的超级智能实验室推出闭源模型，以"低一个数量级的算力"达到Llama中等规模变体水平。将应用于Facebook、Instagram、WhatsApp及雷朋智能眼镜。这家以开放构建AI品牌的公司认定，顶尖模型价值过高不宜共享。

Meta将开源小型模型，封锁顶尖成果 · 4月6日 · Axios
→ 混合策略：开放小型成果，封锁核心资产。放弃2万亿参数的Behemoth后，Meta转向更小、高效、专有的模型。Llama开源社区应认清现实。

OpenAI对阵各方
首席财务官打破沉默：2026年IPO"过于激进" · 4月6日 · WinBuzzer
→ Sarah Friar告知同事第四季度IPO无法实现，理由包括140亿美元预期亏损与6000亿美元累计支出。此后她被排除在关键财务会议之外。当首席财务官在科技史上最大金融事件上公开反对首席执行官时，市场不会忽视。

马斯克诉请法院解雇阿尔特曼与布罗克曼 · 4月8日 · CNBC
→ 索赔790-970亿美元损害、要求恢复非营利性质并解雇两位联合创始人——4月27日庭审前的最新动作。OpenAI反击，提请两州总检察长调查马斯克通过xAI、特斯拉和SpaceX实施的"反竞争行为"。

OpenAI提议机器人税、公共财富基金与四天工作制 · 4月6日 · TechCrunch
→ 这家年亏140亿美元的公司提出13页政策愿景，建议政府如何重新分配AI创造的财富——同时竭力维持自身营利架构。

开源竞赛
谷歌以Apache 2.0协议发布Gemma 4 · 4月2日 · 谷歌博客
→ 从边缘计算到310亿参数的四款模型全部开放许可——首个零商业限制的Gemma系列。原生多模态、25.6万上下文窗口，以极低成本媲美闭源模型。当Meta转向专有，谷歌正在填补开源真空。

DeepSeek V4将完全运行于华为芯片 · 4月5日 · TechWire Asia
→ 首个完全未使用英伟达硬件训练的万亿参数前沿模型。若能在华为昇腾950PR芯片上实现规模化运行，美国芯片出口管制的基本前提需重新审视。中国已证明可低成本训练，现正证明可自主训练。

AI与现实世界交锋
伊朗威胁"彻底摧毁"Stargate数据中心 · 4月6日 · TechCrunch
→ 伊朗革命卫队公布OpenAI价值300亿美元阿布扎比设施的卫星坐标并威胁打击。首次有国家明确将AI基础设施列为攻击目标。数据中心安全已从IT问题升级为国家安全问题。

科技裁员突破16.5万人——3月裁员中25%归因于AI · 4月7日 · 彭博社
→ 较一年前个位数比例大幅上升。美联社向120余名转攻AI营收的记者提供买断方案。"AI是否会取代我的工作"对越来越多知识工作者已成为过去时。

美国编剧工会锁定四年协议，AI保护条款升级 · 4月4日 · Variety
→ 3.21亿美元健康基金、更高分成收益及新条款要求制片方在AI训练前披露并获授权使用编剧作品。制片方仍拒绝为训练数据付费——但编剧现在有权知晓自己的文字何时被用于喂养机器。投票将于4月16日至24日进行。

英文来源：

Three seismic shifts in one week. Anthropic's revenue run rate passed OpenAI's — $30 billion to $24 billion — powered by enterprise demand that doubled its million-dollar customers in under two months. Meta launched its first proprietary model under Alexandr Wang's Superintelligence Labs, abandoning the open-source ethos that defined Llama. And the legal apparatus around AI went into overdrive: Musk wants Altman fired, OpenAI wants Musk investigated, and Hollywood's writers just locked down four more years of AI protections.
April 3 – 9, 2026
Anthropic makes more money than OpenAI. Meta stopped sharing its best work. Musk and Altman want each other fired. And somewhere in Abu Dhabi, someone is re-evaluating data center insurance premiums.
Key Takeaways

• Anthropic is now the highest-revenue AI lab. $30B run rate vs. OpenAI's $24B. Enterprise demand — not consumer — drove it. If you're building on Claude, you're betting on the market leader.
• Meta killed open-source AI's biggest champion. Muse Spark is closed. The hybrid strategy means open-source gets the scraps; the best models stay locked. Developers who built on Llama's openness need a Plan B.
• OpenAI's IPO is in real trouble. When your own CFO says the timeline is too aggressive and gets excluded from meetings for saying it, the market reads that as a warning, not a disagreement.
• China can build frontier AI without US chips. DeepSeek V4 on Huawei silicon is the proof point that export controls were buying time, not solving the problem.
• AI infrastructure is now a military target. Iran threatening the Stargate data center means physical security is no longer optional for AI buildouts in contested regions.
  Quick Hits
• Intel joins Musk's Terafab: Signed on to the $20B+ AI chip fab alongside Tesla, SpaceX, and xAI.
• Eclipse raises $1.3B: Largest fund yet for physical AI, robotics, and defense.
• Flowise CVSS 10.0: Open-source AI agent builder under active exploit — 12K+ instances exposed.
• First deepfake conviction: Federal Take It Down Act gets its first guilty verdict.
• UnitedHealth's $3B AI bet: 22K engineers automating claims and clinical docs.
• Big Three vs. China: Sharing intelligence to block model cloning via 24K fake Claude accounts.
• OpenClaw blocked: Claude subs no longer cover third-party agents. API pricing required.
• AI scribes inflating costs: Insurers and providers agree AI coding drives up bills.
  🎬 Watch & Listen First
  Dario Amodei: "We Are Near the End of the Exponential" · Dwarkesh Podcast · Listen on Spotify
  → Anthropic's CEO on why the scaling hypothesis may be reaching its limits, how AI will diffuse through the economy, and whether regulation will destroy the benefits. Essential context for the $30B revenue number this week.
  OpenClaw: The Viral AI Agent That Broke the Internet · Lex Fridman #491 · Watch on YouTube
  → Peter Steinberger on how a side project hit 180K GitHub stars and forced Anthropic to block third-party agent access. Context for the OpenClaw crackdown above.
  The Revenue Crown Changes Hands
  Anthropic Hits $30B Run Rate, Passes OpenAI · Apr 7 · Bloomberg
  → From $9B to $30B in four months. Enterprise demand doubled — 1,000+ businesses now spend $1M+ annually on Claude — and Anthropic locked in 3.5 gigawatts of next-gen TPU compute via Google and Broadcom. The company that bet on safety is winning on revenue.
  Claude Mythos Unleashed for Cybersecurity Only · Apr 7 · TechCrunch
  → The most powerful AI model in the world exists and you can't use it. Mythos found thousands of zero-days under Project Glasswing with 12 partner orgs. Frontier models may debut in specialized domains before consumers — cybersecurity first, general release never.
  Meta Goes Closed-Source
  Meta Launches Muse Spark — First Proprietary Model · Apr 8 · CNBC
  → Alexandr Wang's Superintelligence Labs shipped a closed model that matches Llama's midsize variant "for an order of magnitude less compute." Coming to Facebook, Instagram, WhatsApp, and Ray-Ban glasses. The company that built its AI brand on openness just decided the best models are too valuable to share.
  Meta Will Open-Source Small Models, Keep the Best Locked · Apr 6 · Axios
  → The hybrid strategy: release the small stuff, gate the crown jewels. After abandoning the 2-trillion-parameter Behemoth, Meta pivoted to smaller, efficient, proprietary models. Llama's open-source community should read the writing on the wall.
  OpenAI vs. Everyone
  CFO Breaks Ranks: 2026 IPO "Too Aggressive" · Apr 6 · WinBuzzer
  → Sarah Friar told colleagues the Q4 IPO won't happen, citing a $14B projected loss and $600B in cumulative spending. She's been excluded from key financial meetings since. When your CFO publicly contradicts your CEO on the biggest financial event in tech history, the market notices.
  Musk Seeks Court-Ordered Removal of Altman and Brockman · Apr 8 · CNBC
  → $79–97 billion in damages, nonprofit reversion, and removal of both co-founders — ahead of the April 27 trial. OpenAI fired back by asking two state AGs to investigate Musk's "anti-competitive behavior" via xAI, Tesla, and SpaceX.
  OpenAI Proposes Robot Taxes, Public Wealth Fund, 4-Day Workweek · Apr 6 · TechCrunch
  → A 13-page policy vision from the company losing $14 billion a year. Proposing how governments should redistribute the wealth AI creates — while simultaneously fighting to keep its own structure for-profit.
  The Open-Source Race
  Google Releases Gemma 4 Under Apache 2.0 · Apr 2 · Google Blog
  → Four models from edge to 31B params, all fully permissive — the first Gemma with zero commercial restrictions. Native multimodal, 256K context, and competitive with closed models at a fraction of the cost. As Meta goes proprietary, Google fills the open-source vacuum.
  DeepSeek V4 to Run Entirely on Huawei Chips · Apr 5 · TechWire Asia
  → The first 1-trillion-parameter frontier model trained without any NVIDIA hardware. If it performs at scale on Huawei Ascend 950PR silicon, the entire premise of US chip export controls needs rethinking. China proved you can train cheap; now they're proving you can train sovereign.
  AI Meets the Real World
  Iran Threatens "Complete Annihilation" of Stargate Data Center · Apr 6 · TechCrunch
  → The IRGC published satellite coordinates of OpenAI's $30B Abu Dhabi facility and threatened strikes. First time a nation-state has explicitly threatened AI infrastructure. Data center security just went from IT problem to national security problem.
  Tech Layoffs Pass 165,000 — AI Cited in 25% of March Cuts · Apr 7 · Bloomberg
  → Up from single digits a year ago. The AP offered buyouts to 120+ journalists pivoting to AI revenue. "Will AI take my job" is becoming past tense for a growing share of knowledge workers.
  WGA Locks Down Four-Year Deal With Expanded AI Protections · Apr 4 · Variety
  → $321M health fund, higher residuals, and new language requiring studios to disclose and license writers' work before AI training. Studios still won't pay for training data — but writers now have the right to know when their words feed the machine. Vote runs April 16–24.