🏆 Headline
OpenAI Cuts Funding Target by $10 Billion: The $1.4 Trillion Valuation Bends
OpenAI has lowered the target for its latest funding round from $30 billion to about $20 billion, with the pre-money valuation target cut from $1.4 trillion to $1.25 trillion, the Financial Times reported. Announced in late September as a key stockpile for OpenAI's run-up to an IPO, the round has now shrunk by a third. The FT attributed the cut to two forces: the broader AI stock pullback and Anthropic's imminent public listing — the two companies are courting the same investors. Reuters reported that OpenAI is keen to close the round before Anthropic starts trading, so investor commitments don't get locked up by its rival first. OpenAI is pursuing a listing in 2027.
Source: Financial Times/Reuters/CNBC | 2026-10-31
ChatGPT and Codex Will Share One Usage Allowance Starting December
OpenAI's help center now says that starting in December, usage allowances on Plus, Pro and Business plans will be shared across ChatGPT and Codex — every Codex token spent leaves less for ChatGPT, and vice versa. The stated reason: "as we bring the agentic capabilities of Work into the main ChatGPT experience." Instant chat won't count toward the limit. The change was first spotted by product researcher Tibor Blaho in new checkout-page wording, before the official help article went up. Notably, OpenAI had just doubled Pro-tier usage on October 30; with two allowances merged into one pool, that doubling needs to be revalued.
Source: OpenAI Help Center/Windowsreport | 2026-10-31
Anthropic Draws a Red Line for Claude: No Recreating Existing Games
On October 27, Claude Code's built-in rules were quietly updated: building original games, genre games and works from the user's own designs and assets is fine, but "Do not build a recreation of a specific existing game, even if it is renamed or called original." Users found the rule cannot be disabled. After Game Developer inquired, Anthropic technical staff member Addy Osmani confirmed the team added two sentences to the Claude Code prompt on Opus 5.5, because "using Claude to infringe IP is against our usage policy," pledging that original and genre games keep working. Backdrop: the recent wave of AI game remakes, which has spooked studios into canceling public playtests, and drew user complaints that the new rule is overly broad and catches legitimate requests.
Source: Game Developer/Mashable | 2026-10-31
DeepSeek Open-Sources dsh-superagent: A Coding Agent Largely Written by Agents
On October 31, DeepSeek open-sourced dsh-superagent-prototype on GitHub under the MIT license; the repo crossed 2,100 stars on release day. The repository is candid about itself: a research prototype, not an official product, grown out of the team's daily development work "with much of the engineering done by coding agents." The form is a self-hosted orchestration engine: running inside DeepSeek Harness, it turns external events from Feishu and GitHub into managed agent tasks, each request getting its own session and directory, working through multiple GitHub issues in parallel.
Source: GitHub/Tech Bytes | 2026-10-31
Apple CEO: No AI for AI's Sake — "When We're Doing It Right, People Won't Have to Think About It"
Apple CEO John Ternus told Wired (reported by TechCrunch on October 31) that Apple is "absolutely going to leverage AI" in new products, but "We don't want to ship AI for AI's sake. When we're doing it right, people won't have to think about whether it was AI that made this happen or more traditional software." He framed the test as value users can feel, not the technology behind it, and reaffirmed the privacy and on-device processing line. This is the Ternus era's first batch of AI positioning — this week's M6 put the 2nm process fully behind on-device AI, and the touchscreen MacBook Pro had just broken a decades-old taboo.
Source: TechCrunch/Wired | 2026-10-31
Nvidia Weighs Playing the Cosigner for a Client's Data Center Loan
According to TechCrunch, Nvidia is in talks to guarantee financing for Nscale's AI data center campus in West Virginia — effectively lending its balance sheet to a client's creditworthiness. The campus's first 460-megawatt building is contracted to Anthropic under a $44.6 billion deal, planned around Nvidia's next-generation Vera Rubin chips. The chipmaker has gone from selling cards, to investing in customers, to backstopping its customers' lenders — each layer ties its own balance sheet tighter to its clients' fate.
Source: TechCrunch | 2026-10-31
Iluvatar CoreX Unveils Tiangai 300: Beats H200 Live on Stage; Outside Test Found One 5% Lead
Chinese chipmaker Iluvatar CoreX lit up its new general-purpose GPU Tiangai 300 for the first time at a Shanghai media day on October 28, claiming real-world performance that reaches Nvidia's H200 level and even surpasses it on mixed text-image-video multimodal tasks, plus rack designs it says cut costs by 40%. The company called it the first domestic general-purpose GPU with measured performance at the H200 tier. But reviews found the only comparable outside test shows just a ~5% lead, with most other claims resting on vendor-staged live demos.
Source: South China Morning Post | 2026-10-30
🏆 今日头条
OpenAI 融资目标砍掉 100 亿美元：1.4 万亿估值松口了
据英国《金融时报》报道，OpenAI 已把本轮融资目标从 300 亿美元下调至约 200 亿美元，投前估值目标也从 1.4 万亿美元降至 1.25 万亿。9 月底公布的这轮融资原是 OpenAI 上市前的重要弹药储备，如今金额缩了三分之一。 《金融时报》将下调归因于两件事：近期 AI 概念股整体回调，以及 Anthropic 上市在即——两家公司正在争取同一批投资人。路透社的报道称，OpenAI 希望抢在 Anthropic 挂牌之前完成本轮，免得投资人的钱先被对手锁走。OpenAI 正在推进 2027 年上市。 > 💬 这不是收缩，是定价权的转移。当 Anthropic 拿着招股书进场，同一批投资人的钱就有了第二张报价单。1.25 万亿不是 OpenAI 变便宜了，而是市场头一回把万亿估值当成可以砍价的数字来对待——此前所有融资故事里，价格都是卖方写的。
来源：金融时报/路透社/CNBC | 2026-10-31
ChatGPT 和 Codex 的额度要合并了：12 月起共享一份用量
OpenAI 帮助中心更新措辞：从 12 月开始，Plus、Pro、Business 订阅的用量额度将在 ChatGPT 和 Codex 之间共享——用掉多少 Codex，ChatGPT 就少多少，反之亦然。官方说明是「随着 Work 的智能体能力并入主 ChatGPT 体验」，即时聊天不占用额度。这个变化最早被产品研究员 Tibor Blaho 在结账页新文案中发现，帮助中心随后挂出正式说明。值得注意的是，10 月 30 日 OpenAI 刚把 Pro 全线用量翻倍，如今两份额度并成一份，翻倍的含金量要重新算了。 > 💬 单独限额是产品部门的地盘，共享额度是财务报表的语言。当编码用量和聊天用量放进同一个池子，OpenAI 等于承认：智能体不再是赠品，是要和主力产品抢同一块蛋糕的成本项。
来源：OpenAI Help Center/Windowsreport | 2026-10-31
Anthropic 给 Claude 划红线：不帮重制现有游戏
10 月 27 日，Claude Code 的内置规则被悄悄更新：可以构建原创游戏、类型游戏和基于用户自己设计与素材的作品，但「不得重制任何一款具体存在的游戏，即使改名或声称原创」。有用户发现该规则无法关闭。Game Developer 向 Anthropic 求证后得到技术成员 Addy Osmani 的确认：团队在 Opus 5.5 的 Claude Code 提示词里加了两句话，因为「用 Claude 侵犯知识产权违反使用政策」，并承诺原创和类型游戏照常可用。背景是近几个月的 AI 游戏重制潮——不少工作室因担心「抢先复刻」抢跑而取消了公开试玩，也有用户抱怨新规则范围偏宽，误伤正当需求。 > 💬 两句 prompt 就是一套知识产权政策：没有法务流程，没有申诉渠道，模型自己当法官。上一轮大家还在争论「模型该不该有价值观」，这一轮直接进入了「提示词即执法」的时代——效率极高，边界全靠写提示词的人拿捏。
来源：Game Developer/Mashable | 2026-10-31
DeepSeek 开源 dsh-superagent：一个大部分由智能体自己写出来的编码智能体
10 月 31 日，DeepSeek 在 GitHub 开源 dsh-superagent-prototype，MIT 协议，发布当天冲过 2100 星。仓库自述很坦诚：这是研究原型而非官方产品，代码来自团队日常开发工作，「大部分工程由编码智能体完成」。它的形态是一个自托管编排引擎：跑在 DeepSeek Harness 里，把飞书、GitHub 等平台的外部事件转成受管理的智能体任务，每个请求开独立的会话和目录，可以并行处理多个 GitHub issue。 > 💬 真正值得注意的不是又一个编码智能体，而是那行自述——「大部分工程由智能体完成」。当开源的不再只是模型，而是「用模型造工具的工作流」，对手抄走的就不止是权重，还有生产线本身。
来源：GitHub/Tech Bytes | 2026-10-31
Apple CEO：不为 AI 而 AI，「做得对的时候用户根本感觉不到」
苹果 CEO John Ternus 接受 Wired 采访（TechCrunch 10 月 31 日报道）时表示，苹果「绝对会用 AI」武装新产品，但「我们不想为了 AI 而 AI。做得对的时候，用户根本不需要想这功能是 AI 实现的，还是传统软件实现的」。他强调评判标准是用户感知到的价值，而非技术本身，并重申隐私与端侧处理的路线。这是 Ternus 时代第一批 AI 相关表态——本周的 M6 芯片把 2 纳米制程全押端侧，触屏 MacBook Pro 也刚打破多年禁忌。 > 💬 在全行业比谁嗓门大的一周里，这句「感觉不到」其实是最狠的定位：别人卖的是 AI 功能，苹果想卖的是「AI 消失」。是清醒还是掉队，要看明年产品说话——但至少这份克制，和满屏的智能体发布会形成了刺眼的反差。
来源：TechCrunch/Wired | 2026-10-31
英伟达考虑给客户的数据中心贷款当「联署人」
据 TechCrunch 报道，英伟达正洽谈为 Nscale 位于西弗吉尼亚州的 AI 数据中心园区提供融资担保——相当于给客户的借款人资质背书。该园区首栋 460 兆瓦建筑已由 Anthropic 签下 446 亿美元的租约合同，计划采用英伟达下一代 Vera Rubin 芯片。芯片厂商从卖卡，到投资客户，再到给客户的债主兜底，每一层都把自己的资产负债表和客户的命运绑得更紧。 > 💬 卖铲人给挖金队担保贷款，这生意只有一种风险：金矿挖空那天，铲子、贷款和担保一起归零。英伟达正在把「循环融资」的争议从财报电话会里的提问，变成资产负债表上的实打实科目。
来源：TechCrunch | 2026-10-31
天数智芯发布天垓300：现场对标 H200，第三方实测只找到一项领先
中国芯片公司天数智芯 10 月 28 日在上海媒体开放日首次点亮新一代通用 GPU 天垓300，宣称其实测性能达到英伟达 H200 水平，在图文视频混合的多模态任务上甚至反超，并称机架方案可把成本降低 40%。公司称这是首款实测性能触及 H200 量级的国产通用 GPU。不过媒体梳理发现，目前唯一一项第三方可比测试只显示约 5% 的领先幅度，其余宣称大多来自厂商自设的现场演示。 > 💬 「对标旗舰」的话术要打折听：现场演示是厂商自己搭的台，裁判也是自己请的。5% 的第三方领先和「全面对标」之间，隔着整个供应链、软件栈和量产良率的距离——但能走到这一步，本身已是牌桌上的变化。
来源：南华早报 | 2026-10-30
Google 拟让 robots.txt 管住 AI 训练：一大排公司站在了同一张纸前
Google Search Central 博客宣布，计划支持网站通过 robots.txt 表达统一的 AI 训练退出偏好。The Verge 报道称，这项共享标准获得了 OpenAI、微软、Google 等多家主要 AI 公司的支持，出版商今后可以用一个控制点，向所有爬虫一次性声明「我的内容不许用于训练」。此前每家模型厂商各有各的退出协议，网站主只能逐家谈判或干脆封锁。 > 💬 robots.txt 是上世纪的君子协定，AI 时代它被推举成产权说明书。大公司们罕见地站到同一边，未必全是善意——统一的标准对巨头是合规成本低，对中小出版商则是「要么签字、要么被绕开」的新门槛。
来源：The Verge/Google Search Central | 2026-10-31