内容来源：https://aiweekly.co/issues/anthropic-files-for-an-ipo-nvidia-ships-its-stack

内容总结：

AI实验室加速竞赛，美国各州监管步伐滞后——本周AI行业要闻

Anthropic提交IPO申请，发布Claude Opus 4.8模型

人工智能初创公司Anthropic本周向美国证券交易委员会（SEC）秘密提交了S-1注册声明草案，拟进行首次公开募股（IPO）。这是前沿AI实验室首次公开启动上市进程，股票数量和定价尚未确定，最终上市取决于SEC审查和市场条件。与此同时，该公司还推出了Claude Opus 4.8模型，宣称新模型在代码审查中遗漏缺陷的可能性比前代降低了约四倍，并为智能体用户新增了动态工作流和努力控制功能。该模型定价保持不变，输入/输出每百万token分别为5美元和25美元。

英伟达全面铺开2026年路线图

英伟达在GTC台北大会上密集发布多项重磅产品。Cosmos 3作为物理AI基础模型正式开源，其Super和Nano版本已在HuggingFace和GitHub上线。新芯片Vera Rubin进入全面量产阶段，供应链规模是Grace Blackwell的两倍，覆盖30个国家的150家台湾合作伙伴和350多家工厂。此外，英伟达通过与联发科合作推出RTX Spark，将1 petaflop的AI算力引入轻薄Windows笔记本电脑；DGX Station则定位为企业级桌面超级计算机。黄仁勋在演讲中表示：“AI现在已成为利润和GDP的生成器。”

谷歌Gemini 2.0 Flash正式退役

谷歌按计划在本周日正式退役Gemini 2.0 Flash模型，推荐用户迁移至Gemini 2.5 Flash。任何仍路由至2.0 Flash的生产应用需尽快规划迁移。

美国各州监管出现分化

加利福尼亚州参议院通过SB 867法案，禁止在儿童玩具中使用AI陪伴聊天机器人。该法案由参议员帕迪拉牵头，目前已提交至众议院审议。

与之形成对比的是，伊利诺伊州旨在要求数据中心自行生产可再生能源、报告用水情况并签署社区福利协议的《电力法案》（POWER Act）在委员会搁浅，本春季无法通过。州议员加贝尔等民主党人转而推动暂停至少27个数据中心总计约9.83亿美元的税收减免优惠，相关讨论将持续至秋季否决会议。

一周总结：实验室在飞奔，各州在爬行

本周的清晰脉络是：AI模型端和算力端的发展速度远远超过了监管者的应对能力。Anthropic一边提交IPO文件，一边发布更强大的模型；英伟达通过一场主题演讲开出了几乎一整年的硬件路线图。而在政府层面，加州仅在儿童玩具聊天机器人这一狭窄领域取得了进展，伊利诺伊州更广泛的数据中心监管尝试则陷入停滞。谷歌将一代模型从前沿淘汰为退役的周期，仅仅相当于一个州议会通过一项聊天机器人法案所需的时间。

中文翻译：

Anthropic今日向美国证券交易委员会（SEC）秘密提交了S-1表格草案，计划进行公开募股。该公司上周还发布了Claude Opus 4.8，代码可靠性提升了4倍。英伟达（NVIDIA）利用GTC台北大会开放了Cosmos 3，将Vera Rubin投入量产，并在开发者笔记本电脑上配置了1千万亿次浮点运算能力的AI盒子。谷歌（Google）今日停用了Gemini 2.0 Flash。加利福尼亚州的SB 867法案——禁止在儿童玩具中使用AI陪伴聊天机器人——已通过参议院；伊利诺伊州的数据中心监管法案在委员会审议中受阻。实验室在狂奔，各州却在爬行。
获取更多来自《AI周刊》的内容
更多信号，更少噪音——选择你的频道。
您正在阅读本周简报。以下是关注报道的其他方式——所有频道均免费，可随时退出。

→ 探索16个深度专题
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、医疗保健等。浏览全部16个深度专题 →

→ AI突发警报
当重大事件发生时（一笔600亿美元的收购、监管机构的紧急会议、前沿模型泄露），订阅者会在几小时内收到通知。通常每天0-2封邮件。获取突发警报 →

→ AI今日新闻（实时）
扫描仪发现新闻时实时更新的动态面板：过去48小时内评分的故事、每周实体动态、以及涵盖113家AI公司、人物和主题的季度趋势线。打开AI今日新闻 →
速览
实验室角斗士时代

• Anthropic向SEC秘密提交S-1表格草案——Anthropic, PBC今日就拟议的首次公开募股向SEC提交了Form S-1草案。根据第135条规定，该公司可宣布提交文件，但这并不构成出售要约；股票数量和定价尚未确定；实际IPO取决于SEC审查和市场状况。首个前沿实验室的IPO倒计时已开始。[Anthropic]
• Anthropic推出Claude Opus 4.8——定价维持在每百万输入/输出代币5美元/25美元。Anthropic表示，该模型“让其编写代码中的缺陷被忽视的可能性比前代产品低约四倍”，并为代理用户推出了新的动态工作流和努力控制功能。[Anthropic]
  全自动模式
• 英伟达Cosmos 3开放其物理AI基础模型——一个模型处理视觉推理、世界生成和动作预测，Cosmos 3 Super和Nano版本已在HuggingFace和GitHub上线。Cosmos联盟创始成员包括Agile Robots、Black Forest Labs、Generalist、LTX、Runway和Skild AI。英伟达的宣传语：将物理AI训练“从数月缩短至数天”。[英伟达]
• 英伟达Vera Rubin全面投产——黄仁勋在GTC台北大会表示：Vera Rubin正在扩大生产，其供应链规模是Grace Blackwell的两倍——在30个国家拥有150家台湾合作伙伴，涉及超过350家工厂。该平台将88核的Vera CPU与Rubin GPU及Spectrum-X以太网光子技术配对，旨在打造百万GPU的AI工厂。[英伟达]
• 英伟达借助RTX Spark将1千万亿次浮点运算能力的AI引入Windows笔记本电脑——RTX Spark通过与联发科（MediaTek）的合作，将1千万亿次浮点运算的AI计算能力带入轻薄Windows笔记本电脑；DGX Station则被定位为面向企业代理部署的桌面级超级计算机。英伟达的宣传语：每个工作站都拥有一台个人AI计算机。[英伟达]
• Gemini 2.0 Flash今日迎来预定的退役日期——根据谷歌的弃用文档，Gemini 2.0 Flash计划于2026年6月1日退役；gemini-2.5-flash是推荐的替代品。任何仍路由到2.0 Flash的生产应用都需要规划迁移。[谷歌AI]
  政府认真对待之年
• 加州SB 867法案禁止AI陪伴聊天机器人玩具，已通过参议院——参议员帕迪拉（Padilla）是主要起草人；参议员卢比奥（Rubio）在列名的共同起草人中。该法案于5月28日通过参议院，现已提交至众议院，并于同一天进行一读。法案页面上列出的适用范围：玩具和陪伴聊天机器人。[加州立法信息网]
• 伊利诺伊州针对数据中心的POWER法案在委员会审议中受阻——该法案曾要求伊利诺伊州的数据中心自行生产可再生能源、向州政府报告用水情况并签署社区福利协议。该法案不会在今年春季通过。共和党众议员罗宾·加贝尔（Robyn Gabel）及其他民主党议员现在正推动暂停针对至少27个设施、预计总生命周期税收减免及福利价值达9.83亿美元的数据中心税收抵免，与此同时，相关讨论将持续到秋季否决会期。[伊利诺伊州议会新闻]
  本周：实验室跑在了各州前面
  抛开新闻标题，主线其实很简单：AI的模型侧和计算侧本周的发展速度，超过了所有正在处理该问题的监管机构。
  Anthropic当天开始便向SEC秘密提交了S-1表格草案——这是首个进入公开记录的前沿实验室IPO申请。该公司还在同一时间段推出了Claude Opus 4.8，声称该模型“让其编写代码中的缺陷被忽视的可能性比前代产品低约四倍”，并为代理用户配套了新的动态工作流和努力控制功能。S-1本身根据第135条规定保密——股票数量和定价尚未确定——但当一个前沿实验室同时也成为一家上市公司时，AI问责制将变成什么样子，倒计时已经开始。
  与此同时，英伟达利用GTC台北大会，在一次主题演讲中几乎发布了其整个2026年路线图。Cosmos 3作为物理AI的基础模型开源，Super和Nano变体已在HuggingFace和GitHub上线。Vera Rubin在150家台湾合作伙伴以及30个国家超过350家工厂全面投产。RTX Spark通过与联发科合作，将1千万亿次浮点运算的AI计算能力带入轻薄Windows笔记本电脑。DGX Station则作为面向企业代理的桌面级超级计算机问世。配备共封装光学的Spectrum-X以太网光子技术现已投入生产，用于百万GPU工厂。黄仁勋在台上的原话是：“AI现在是一个利润生成器。AI现在是一个GDP生成器。”
  相比之下，州一级的立法反应显得微不足道。加州的SB 867法案通过了参议院，旨在禁止面向儿童的陪伴聊天机器人玩具——这是一项实际、具体且重要的规定。伊利诺伊州要求数据中心自行购买可再生能源、核算用水量并签署社区福利协议的更广泛尝试，则在委员会审议中受阻，加贝尔众议员和一群民主党议员现在反而是在防守税收抵免。
  谷歌在同一天停用了Gemini 2.0 Flash，这是本周最小的信号，或许也最具说明性：一代模型从前沿走向退役，所花费的时间大致相当于一个州立法机构通过一项聊天机器人法案所需的时间。
  关键要点
• Anthropic秘密提交的S-1文件，开启了一家前沿AI实验室成为上市公司的首个时间表。该公司从现在到定价期间提交的披露信息，将成为SEC文件中AI风险、资本支出和收入披露的运营基准。
• 英伟达通过一次主题演讲，将Cosmos 3开源，将Vera Rubin推入超过350家工厂的生产线，并将千万亿次浮点运算能力带到了Windows笔记本电脑上。目前，“物理AI”栈——感知、仿真、硬件——已成为一家公司已发布的产品路线图。
• 美国州级监管正在针对具体的消费者保护案件（加州参议院通过的SB 867玩具聊天机器人禁令）采取行动，而更广泛的基建问题（伊利诺伊州POWER法案）则不断被搁置。联邦层面针对AI资本支出的行动仍然缺失。
  值得一读
• Netflix工程师的Headroom代理削减了代理代币费用，已开源——Tejas Chopra的开源代理位于代理代码与大语言模型API之间，在代币到达模型之前对其进行压缩。已被Netflix多个团队采用；迄今已为用户节省约70万美元。[The Register]
• 英特尔Xeon 6“Clearwater Forest”声称每线程性能比AMD EPYC 9965高出30%——采用18A制程，最高288核，旗舰版6990E拥有576 MB L3缓存。该性能对比是英特尔自述的；AMD尚未回应。[Tom's Hardware]
• 中国邮政人形机器人在广州枢纽每小时处理1200个包裹——江高物流场地（日均处理650万件包裹，峰值超过1000万件）现正使用人形机器人，配合机械臂和无人叉车。这是一个生动数据点，反驳了“人形机器人仍停留在演示阶段”的观点。[Interesting Engineering]
• 三星超越美光，成为全球最大汽车存储器供应商——2025年首次达成，得益于在中国的积极扩张和软件定义汽车市场。虽与AI领域相邻，但遵循相同的需求曲线。[KED Global]
  本周投票
  Anthropic刚刚启动了首个前沿实验室的IPO倒计时。您认为SEC最应该强制要求公开披露什么内容？
  上周，有154位读者参与了投票：
  当员工对AI部署提出反对时，正确的回应是什么？
  Anthropic刚刚启动了首个前沿实验室的IPO倒计时。您认为SEC最应该强制要求公开披露什么内容？
  ——亚历克西斯

英文来源：

Anthropic confidentially filed a draft S-1 with the SEC today for a proposed public offering. The company also shipped Claude Opus 4.8 last week with a 4x code-reliability gain. NVIDIA used GTC Taipei to open Cosmos 3, ramp Vera Rubin into production, and put a 1-petaflop AI box on developer laptops. Google retires Gemini 2.0 Flash today. California's SB 867 — banning AI companion chatbots in children's toys — cleared the Senate; Illinois's data-center regulation stalled in committee. The labs sprint. The states crawl.
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

• Anthropic confidentially files draft S-1 with the SEC — Anthropic, PBC submitted a draft Form S-1 to the SEC today for a proposed IPO. Per Rule 135, the company can announce the filing without it constituting an offer to sell; share count and pricing aren't set; the actual IPO depends on SEC review and market conditions. The first frontier-lab IPO clock just started. [Anthropic]
• Anthropic launches Claude Opus 4.8 — Pricing stays at $5/$25 per million input/output tokens. Anthropic says the model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked," and ships new Dynamic Workflows and Effort Control features for agent users. [Anthropic]
  Auto Mode Everything
• NVIDIA Cosmos 3 opens its physical-AI foundation models — One model handles vision reasoning, world generation and action prediction, with Cosmos 3 Super and Nano live on HuggingFace and GitHub. Founding Cosmos Coalition members include Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI. NVIDIA's pitch: physical AI training "from months to days." [NVIDIA]
• NVIDIA Vera Rubin ramps into full production — Jensen Huang at GTC Taipei: Vera Rubin is ramping with a supply chain twice the size of Grace Blackwell — 150 Taiwan-based partners across 350+ factories in 30 countries. The platform pairs an 88-core Vera CPU with Rubin GPUs and Spectrum-X Ethernet Photonics, targeting million-GPU AI factories. [NVIDIA]
• NVIDIA brings 1-petaflop AI to Windows laptops with RTX Spark — RTX Spark brings 1 petaflop of AI compute to slim Windows laptops via a MediaTek partnership; DGX Station is positioned as a deskside supercomputer for enterprise agent deployment. NVIDIA's pitch: a personal AI computer at every workstation. [NVIDIA]
• Gemini 2.0 Flash hits its scheduled retirement date today — Per Google's deprecation docs, Gemini 2.0 Flash is scheduled to be retired on June 1, 2026; gemini-2.5-flash is the recommended replacement. Any production app still routing to 2.0 Flash needs to plan its migration. [Google AI]
  The Year Governments Got Serious
• California's SB 867 to ban AI companion chatbot toys clears the Senate — Senator Padilla is the lead author; Senator Rubio is among the listed coauthors. The bill cleared the Senate on May 28 and is now in the Assembly, with first reading the same day. Subject as listed on the bill page: toys and companion chatbots. [California Legislative Info]
• Illinois POWER Act for data centers stalls in committee — The bill would have required Illinois data centers to generate their own renewable energy, report water consumption to the state, and sign community benefits agreements. It will not pass this spring. Rep. Robyn Gabel and other Democratic lawmakers are now pushing to pause data-center tax credits worth $983 million in estimated lifetime tax breaks and benefits across at least 27 facilities, while talks continue into the fall veto session. [Capitol News Illinois]
  The week the labs lapped the states
  Strip out the headlines and the through-line is simple: the model side and the compute side of AI moved faster this week than every regulator working on it.
  Anthropic started the day by confidentially filing a draft S-1 with the SEC — the first frontier-lab IPO filing on the public record. The company also shipped Claude Opus 4.8 in the same window, claiming the model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked," paired with new Dynamic Workflows and Effort Control features for agent users. The S-1 itself is confidential under Rule 135 — share count and pricing not yet set — but the clock is now running on what AI accountability looks like when a frontier lab is also a public company.
  NVIDIA, meanwhile, used GTC Taipei to roll out almost its whole 2026 roadmap in a single keynote. Cosmos 3 open-sourced as the foundation model for physical AI, with Super and Nano variants live on HuggingFace and GitHub. Vera Rubin entering full production across 150 Taiwan-based partners and 350+ factories in 30 countries. RTX Spark putting 1 petaflop of AI on slim Windows laptops via MediaTek. DGX Station landing as a deskside supercomputer for enterprise agents. Spectrum-X Ethernet Photonics with co-packaged optics now in production for million-GPU factories. Jensen's actual line on stage: "AI is now a profit generator. AI is now a GDP generator."
  The state-level legislative reaction looks small next to that. California's SB 867 cleared the Senate to ban companion-chatbot toys for kids — a real, narrow, important rule. Illinois's broader bid to make data centers buy their own renewable power, account for their water, and sign community benefits agreements just stalled in committee, with Rep. Gabel and a group of Democratic lawmakers now playing tax-credit defense instead.
  Google sunsetting Gemini 2.0 Flash on the same Sunday is the smallest signal of the week and possibly the most telling: a generation of model has gone from frontier to retired in roughly the time it takes a state legislature to pass a single chatbot bill.
  Key takeaways
• Anthropic's confidential S-1 starts the first public-company timeline for a frontier AI lab. The disclosures the company files between now and pricing will become the operational baseline for what AI risk, capex, and revenue disclosure look like in SEC filings.
• NVIDIA used a single keynote to make Cosmos 3 open-source, push Vera Rubin into production at 350+ factories, and put a petaflop on a Windows laptop. The "physical AI" stack — perception, simulation, hardware — is now one company's shipped roadmap.
• US state-level regulation is moving on narrow consumer-protection cases (SB 867's toy-chatbot ban clearing the California Senate) while the broader infrastructure questions (Illinois POWER Act) keep getting punted. Federal action on AI capex remains absent.
  Worth reading
• Netflix engineer's Headroom proxy cuts agent token bills, open-sourced — Tejas Chopra's open-source proxy sits between agent code and LLM APIs, compressing redundant tokens before they reach the model. Adopted by several Netflix teams; ~$700K saved across users so far. [The Register]
• Intel Xeon 6 "Clearwater Forest" claims 30% per-thread edge over AMD EPYC 9965 — 18A node, up to 288 cores, 576 MB L3 cache on the flagship 6990E. The performance comparison is Intel's own; AMD has not responded yet. [Tom's Hardware]
• China Post humanoid robots hit 1,200 parcels/hour at Guangzhou hub — Jianggao logistics site (6.5M parcels/day, peaks above 10M) is now using humanoids alongside robotic arms and unmanned forklifts. A live data point against "humanoids are still demos." [Interesting Engineering]
• Samsung surpasses Micron as the world's top automotive memory supplier — First time in 2025, driven by aggressive China expansion and the software-defined-vehicle market. Adjacent to AI but the same demand curve. [KED Global]
  This week's poll
  Anthropic just started the first frontier-lab IPO clock. What's the most important thing the SEC should force out into public disclosure?
  Last week, 154 of you voted:
  When workers push back on AI rollouts, what's the right response?
  Anthropic just started the first frontier-lab IPO clock. What's the most important thing the SEC should force out into public disclosure?
  — Alexis