报告形态 · 五维分类 + 风险 + 行动 + 时间线 ← 回站点
WEEK 2026-10-27 · 46 STORIES · SKILLHUB WEEKLY

SkillHub.news 每周AI

10 月 21 日 ~ 2026 年 10 月 27 日 · 周报 · 素材 4
本期看点46 条 · 约18分钟
01
本周主题线WEEKLY THREADS24 条
OpenAI 30-90%折扣撬走Anthropic大客户
02
官方事件精选OFFICIAL EVENTS14 条
Ireland: Build the future of technology with Intel Foundry | Intel
03
榜单异动RANKING MOVES8 条
obra/superpowers 星标 +79,301
01本周主题线WEEKLY THREADS24 条

OpenAI 30-90%折扣撬走Anthropic大客户

报道台账2026-10-26
Replit获超30%折扣后转移年合同额超1亿美元AI支出;Rippling编码负载OpenAI份额16%→52%(Anthropic 57%→18%);某70亿美元资产银行每席不足7美元获500席;投资者备忘录显示Anthropic Q2→Q3增速放缓——恰逢递表日冲击IPO定价叙事(The Information 10-24/The Decoder, 高)

寒武纪Q3业绩翻倍+非GPU加速器出货首超GPU

报道台账2026-10-26
寒武纪Q3营收35.31亿+104.47%、净利14.04亿+147.74%(前三季营收95.26亿+106.76%);海光+57.58%、沐曦扭亏预告;IDC:2026H1中国加速器出货200.8万颗,非GPU 101.6万颗(+80%)首超GPU,英伟达份额降至34%、华为升至28%(57.2万颗)(三季报/IDC转引, 高/中)

Cerebras 59片晶圆级芯片跑通3T Kimi K3

Cerebras官方博客多源转引2026-10-26置信度 高
4-bit量化分布59片WSE-3T(44GB片上SRAM装下每层权重),推理1000-2000+ tok/s、功耗约1.18MW;非英伟达硅片TB2.1 88.3%/BrowseComp 91.2%逼近GPT-5.6 Sol——前沿推理瓶颈证实为内存与互连而非FLOPs(Cerebras官方博客多源转引, 高)

FTC拟双查OpenAI/Anthropic+Storm-0940实战

含METR,Axios独家,立案文书未公布2026-10-26置信度 中
FTC拟依FTC法第5条向两家前沿厂商发强制取证(含METR,Axios独家,立案文书未公布,中);微软披露Storm-0940用智能体框架ARTEX攻破2250+ M365租户(部分3小时夺特权)、x402协议25漏洞、五大行拦截AI代理登录(微软官方/The Hacker News, 高/部分待核)

TikTok USDS洽购4GW+词元水电费化

娱乐资本论/政采公告2026-10-26置信度 高
USDS(银湖/甲骨文/MGX各15%、字节保留19.9%)接洽算力供应商,目标2030年代初锁定4GW≈甲骨文近5个季度交付量(850-1340亿美元量级,CNBC双源、4GW单源,中);北京台300万单一来源采购火山引擎词元(12个月),259家广电144条招标、融媒数智化重点项目9228万(娱乐资本论/政采公告, 高)

vLLM整柜跑满GB300逼近DeepSeek自研栈

报道台账2026-10-26
P/D分离+Wide-EP将DeepSeek-V4.1-Flash跑满GB300 NVL72:单GPU吞吐95.4万tok/s(与DeepInfer差距<6%)、上下文128K→855K仅+7%延迟——开源推理栈首逼头部闭源自研栈(vLLM官方博客, 中·无第三方复测)

Vals首版Router Index+法律榜倒挂

Vals/AA转引2026-10-26置信度 中
5款模型路由器无一进前十(最好Switchyard 57.8%排13/27,openrouter/auto仅52.2%),Gemini 4 Argon 69.2%居首;Muse Wonder Full法律指数77.86%登顶(Opus 5.5仅第七)+AA Harvey通过率13.06% vs Argon 3.89%碾压——通用榜与垂直场景严重倒挂(Vals/AA转引, 中)

微软开源Agensh+Mercury 3扩散LLM

报道台账2026-10-26
Agensh(MIT)去中心化多智能体框架:去中央编排器、1024编码智能体自组织,ProgramBench五难任务19.31%→28.78%(+49%);Inception Mercury 3扩散LLM输出1068 tok/s(约Luna 8倍),GPQA Diamond 89.23反超Luna 86.36,$0.35/M(MarkTechPost/机器之心, 高/中)

上海银行给AI一人公司放贷:经营数据=信用

上海市政府官网10-252026-10-26置信度 高
静安支行×闹海AI把平台上OPC创作者接单量/交付质量/回款数据纳入授信依据,9月已通过政银担机制发300万创业担保贷——轻资产AI创业主体首获系统性金融配套(上海市政府官网10-25, 高)

资本快照:亿铸10亿/Enact 3000万/Lambda 40亿

综合公开报道2026-10-26置信度 中
亿铸科技(3D DRAM存算一体已流片)超10亿元新融资;Enact Intelligence 3000万美元Pre-seed(LeCun/Mensch个人跟投);Lambda拟IPO前筹至多40亿美元@估值145亿——资金转向已验证流片/收入标的与明星学者早期票(综合公开报道, 中)

暂缓池(未采信)

报道台账2026-10-26置信度 中
"AWS 2026-27产能售罄、积压4960亿美元"(今日非财报日,数字疑为财报预期口径混入,暂缓);"1Password for Agents驱动年营收5亿美元"(单源待核);旧金山法院受理OpenAI智能体未授权访问诉讼(中文编译信源,待核)

长期跟踪区拐点更新

报道台账2026-10-26
Anthropic IPO:今日10-26为S-1递交窗口+OpenAI折扣战直接冲击定价(投资者预期回落至1.5万亿区间);OpenAI 300亿融资签约窗口下周;两者为下期首要跟踪点

合规卡600万颗排队:产能只够5-10%

报道台账2026-10-24
中企(字节/阿里/腾讯等)订购超600万颗RTX Pro 5500约350亿美元(单价约1.14万美元),英伟达2027全球产能上限约300万颗、商业交付2027Q1,现阶段仅满足5-10%——需求2倍于全球产能,算力稀缺制度化(路透10-20首发/DW/TrendForce多源,高)

AI债市全面走阔+SpaceX 400亿举债+Meta承诺8250亿

彭博/FT/SEC文件2026-10-24置信度 高
全球IG利差129bp创18个月最宽(CDX IG 64bp),年内AI债融资超4500亿美元(Janus Henderson:明年或近1万亿);SpaceX拟400亿美元举债购卡(100亿贷款+300亿IG债,Apollo牵头,BBB——系10-07旧闻降为背景增量);Meta租赁与合同承诺7000亿→8250亿美元,明言「赶在竞争对手占满融资渠道之前」(彭博/FT/SEC文件,高)

Anthropic IPO倒计时+OpenAI 300亿签约(双王定价)

报道台账2026-10-24
Anthropic拟10-26公开递交S-1、11-9路演、次周定价,拟募至多1000亿美元@至多2万亿、算力承诺升至7500亿(WSJ),纽约市审计长致信呼吁暂缓,投资者估值预期回落至1.5万亿区间;OpenAI 300亿融资(阿联酋资本+Coatue领投)传下周签约@投前1.25万亿,定位2027 IPO前过桥(IFR/CNBC,高/中高)

Kimi K3.1 Pro发布+OpenRouter中国71%

财联社/OpenRouter2026-10-24置信度 高
3T参数/100万上下文,TB4.0 65.7%超GPT-6 Astra 58.2%、SWE-Marathon 51.7%超Opus 5.5(46.3%),另有Highspeed高速版,权重最迟11-4开放;OpenRouter当周中国模型承载71% tokens(年初仅7%)、开源权重77%、DeepSeek请求份额24.4%居首(财联社/OpenRouter,高)

国办AI就业友好型意见18项 vs 巴克莱裁5000人

新华社/Guardian2026-10-24置信度 高
国办发〔2026〕29号(10-23发布,成文10-16)6方面18项:失业缓冲期(表述见「加大转型转岗就业支持」)、算法就业歧视治理、岗位替代风险评估、AI新职业开发(数据标注/AI训练师)、中小企业「小快轻准」轻量化升级;巴克莱裁5000人归因AI(利润+12%);Gartner 2002名CIO仅13%见实质回报(新华社/Guardian,高)

规则双线:得州第10州起诉OpenAI+训练禁直连实时互联网

The Hill/官方博客/央视2026-10-24置信度 高
Paxton 10-23起诉OpenAI+Altman(消费者保护法+SCOPE未成年人法,寻求禁未成年使用、新模型第三方评估批准,每违规罚1万美元,此前8州+佛州=第10州);OpenAI官宣前沿模型训练禁直连实时互联网、改缓存网页沙箱(红队发现VM逃逸/侧信道,起因9-20 DNS漏洞短暂联网,10-1已恢复RL训练);商务部解读:开源出口「即报即出」、API/SaaS/端侧不算技术出口、DeepSeek/MiniMax已履行登记(The Hill/官方博客/央视,高)

谷歌TPU首进Meta生产环境+产业交付线

报道台账2026-10-24
PyTorch原生运行时免改写跑TPU推理,Meta数据中心实际部署——英伟达垄断首被生产级撬动;Vera Rubin NVL72批量交付(每百万token推理成本前代1/45);辉羲智能年内融资超25亿元(机器人芯片出货全球第二约1/5);宇树G1累计下线1-2万台(王兴兴:普及需年产千万台);Cybercab车队169→319辆、11月进达拉斯、无监督FSD推迟2027Q2;中国移动宁夏中卫37.1亿EPC(186MW,与10-11交换机集采为不同项目)(Tom's Hardware/澎湃/财报会/北极星,高/中高)

GPT-5.6 Sol挖出PHP十年高危漏洞+生态安全层

OpenAI/Wiz/GitHub/澎湃2026-10-24置信度 高
两个未授权RCE「PHP2Shell」CVSS 9.2波及PHP 5.4-8.5,Wiz独立复现、官方发4个安全更新——AI找洞从演示变交付;MCP参考实现v2.0.0+Tasks and Events工作组设立(Agents WG降级兴趣组);Skills捐AAIF生态数据(100万Skills/2.8亿安装,厂商自报);美DHS要求联邦机构清退DeepSeek(OpenAI/Wiz/GitHub/澎湃,高)

数据三线:token单价-41%用量+50%/ChatGPT广告83国/Muse千万装机

报道台账2026-10-24
Ramp企业AI指数:有效token单价3月$1.15→9月$0.68(-41%)、用量环比+50%创纪录、企业侧Anthropic 51% vs OpenAI 44.5%、开源/中国企业采用仅6.4%(补课空间);ChatGPT广告新增26国至83国、广告主超10万(8月仅2万,月增约5倍);Muse Wonder超1000万活跃装机/约250万DAU(Sensor Tower)(Ramp多源转引/中高)

榜单口径战:AA v5 TB Science首测+TB版本混淆警示

报道台账2026-10-24
v5新增Terminal-Bench Science首测(Astra 63.3%/Opus 5.5 61.9%)+私有防污染集,v4.3.2发布后不可比;DeepSeek 1017 checkpoint登HF agents-last-exam 35.4+deep-swe 77双榜;第三方对比:同一模型TB2.1约76-88% vs TB4.0严格口径仅3-28%——引用高分先核基准版本号(AlphaSignal/dev.to/BenchLeader,中高/中)

资本快照:DeepSeek交割细节/Rivian收购世界模型/微亿智造挂牌

彭博/TechCrunch/Crunchbase2026-10-24置信度 高
DeepSeek交割投前约700亿美元(宁德时代拟增资100亿翻番,新进君联/石溪/吉利/中芯聚源/横店);Rivian收购Third Dimension(前Tesla/Wayve世界模型团队,扩建伦敦AI中心);Flapping Airplanes 10亿美元A轮(Lux+Index领投,成立不足1年);Higgsfield获英伟达+TCV追加(8月B轮4亿/估值54亿扩展);微亿智造10-28港股挂牌「工业具身智能第一股」(彭博/TechCrunch/Crunchbase,高)

趋势:实验室中层密集改口AGI+灭绝警告升级

报道台账2026-10-24置信度 中
前DeepMind Alex Turner:10年内灭绝概率1/3、超级智能12-18个月;DeepMind安全负责人Irving呼吁现在停止进步(先单边后中美协议);Hinton称AI主观体验与人类「一样多」;Anthropic RL负责人Douglas称AGI一两年内(与Hassabis 2-3年同向);Muse Wonder定价$0.2/M vs $200/月(单源待核)(NDTV/agihunt/Dealroom,中高/中)
02官方事件精选OFFICIAL EVENTS14 条

Can Gemini transcribe a song?

YouTube·Google Cloud2026-10-27置信度 高
原文 ↗

Our newest AI Quest: Mapping the human brain

官方·Google AI Blog2026-10-27置信度 高
原文 ↗

The media industry is becoming a prediction economy

官方·Databricks2026-10-27置信度 高
原文 ↗

GPT-6.1 Terra is now in orbit

YouTube·OpenAI2026-10-27置信度 高
原文 ↗
03榜单异动RANKING MOVES8 条

obra/superpowers 星标 +79,301

GitHub2026-09-03 → 2026-09-10
原文 ↗

mattpocock/skills 星标 +71,905

GitHub2026-09-03 → 2026-09-10
原文 ↗

anthropics/skills 星标 +49,058

GitHub2026-09-03 → 2026-09-10
原文 ↗

farion1231/cc-switch 星标 +36,887

GitHub2026-09-03 → 2026-09-10
原文 ↗
生态脉搏站点数据自更新
2026-09-03 → 2026-09-10 星标涨幅:obra/superpowers +79,301★ · mattpocock/skills +71,905★ · multica-ai/andrej-karpathy-skills +59,162★ · anthropics/skills +49,058★ · Shubhamsaboo/awesome-llm-apps +38,266★ · farion1231/cc-switch +36,887★ · nextlevelbuilder/ui-ux-pro-max-skill +35,298★ · Graphify-Labs/graphify +32,474★
※风险分级灰犀牛 / 黑天鹅
灰犀牛 · 可预见
  • 折旧久期错配:GPU 经济寿命 3—5 年 vs 结构化债务 7—10 年,2027 年起 SPV 残值缺口暴露,第一留置权人(芯片厂/银行)先承压(中高,50-70%)
  • Token 单价下行 vs 算力租金刚性:中档模型价格腰斩后,按「卡·年」收租的存量合约变成负毛利来源,触发 Q4 重新议价潮(中,30-60%)
  • 900% 净收入留存回落:预算迁移的存量耗尽后 NDR 快速回归,上市后首份财报若跌破 300%,杀伤力大于任何估值下调(中,30-60%)
  • 微软 6 万人买断执行力:协调层被砍后产品迭代节奏下滑,AI 资本开支的「人力对价」变成「执行力漏水」(中,40-60%)
黑天鹅 · 低概率大冲击
  • 大客户预算骤停:单一超大客户贡献占比过高(S-1 未披露集中度),一次预算冻结即造成营收断崖与连锁违约(低,<30%)
  • 训练数据授权集体诉讼扩大化:猫王案(274 首 / 4100 万美元)被判支持后,版权方批量起诉形成行业级赔付口径(低,<30%)
  • 事实基座污染:AI 生成图/视频被当作公共事件证据广泛传播(苹果裂屏图为前例),监管与司法对「图像证据」采信规则突变(低,<30%)
✦行动建议3
1. 算力三年 TCO + 合约条款审计

给自建/租赁算力做一张「3 年总持有成本 ÷ 3 年有效吞吐」单页表:折旧年限、利用率、2000—4500 美元/GPU/年软件年金、网络侧增量一并入表。路径:免费单机位测算引流 → 按机柜数收费的年度审计 → 代谈合约条款(利用率下限、Token 降价触发重议、残值兜底归属)。

时间窗 即刻起 6 周(Q4 采购谈判窗口);判据:客户是否用「吞吐跑不满」推翻原有比价结论
2. 3 秒语音同传试点(私有化)

用 Hy-SimulMT 开源权重在本机/私有云跑会议实时字幕 + 纪要,指标只有三个:延迟 ≤3s、音色相似、术语命中。路径:单场景跑通(跨境客服或内部周会)→ 加声音授权留痕与确权链 → 按坐席/按时长订阅;风险位:音色克隆须先拿本人书面授权。

时间窗 3—6 个月(语音交互跨过阈值后的抢跑期)
3. 把「可保性」做成交付验收清单

对自有 AI 产品做一次投保前体检:失效模式是否可结构化描述、参数/场景/性能三因子能否出数、第三方损失边界是否写清。路径:自检表 → 与保司/经纪对表拿报价区间 → 把费率因子反向写成 SLA 与合同责任条款(对外当卖点)。

时间窗 3 个月内(全国首个方案刚发布、费率分布未定);判据:能否拿到第一份书面报价
⏱未来 3 个月时间线概率
时间窗事件概率依据
11/4 前Kimi K3.1 Pro 完整权重按期放出高(85%)月之暗面官方公告 / HuggingFace 权重页
10 月底—11 月初Mistral Large 4 完整权重真实可下载,许可含商用限制条款高(80%)HuggingFace 模型页 + LICENSE 原文
未来 3 个月Anthropic 公布发行定价区间(SEC 424B 文件落地)高(75%)SEC EDGAR 424B4 / 交易所公告
未来 3 个月出现首批 AI 应用风险保险的实际费率样本(可对表)中(50%)金融街论坛后续 / 保司产品页 / 经纪报价
未来 3 个月云厂商在 Q4 采购谈判中引用 Rubin 带宽文档压价中(45%)Q4 财报电话会 / 供应链报道
素材来源:报道台账(滚动查重):2026-10-21 ~ 2026-10-27 共 41 条主题记录 · AI 官方视频监控台账:7 天 · 645 条事件(精选 14 条带链接) · 站点榜单快照:2026-09-03 → 2026-09-10 星标涨幅 Top8 · 风险/行动/时间线:承接 2026-10-27 日报
SkillHub.news — 全球 Skill 榜单
WEEK 2026-10-27 · 46 STORIES · SKILLHUB WEEKLY
SkillHub.news Weekly AI
Oct 21 ~ Oct 27, 2026 · Weekly · sources 4
Highlights46 stories · approx.18min read
01
WEEKLY THREADS24 stories
OpenAI's 30-90% discounts poach Anthropic's major customers
02
OFFICIAL EVENTS14 stories
Ireland: Build the future of technology with Intel Foundry | Intel
03
RANKING MOVES8 stories
obra/superpowers stars +79,301
01WEEKLY THREADS24 stories

OpenAI's 30-90% discounts poach Anthropic's major customers

报道台账2026-10-26
After securing over 30% discounts, Replit shifted more than $100 million in annual contract value of AI spending; for Rippling's coding workloads, OpenAI's share went 16%→52% (Anthropic 57%→18%); a bank with $7 billion in assets got 500 seats at under $7 per seat; an investor memo shows Anthropic's Q2→Q3 growth slowed—coinciding with filing day and hitting the IPO pricing narrative (The Information 10-24/The Decoder, High)

Cambricon Q3 results double; non-GPU accelerator shipments surpass GPU for first time

报道台账2026-10-26
Cambricon Q3 revenue 35.31亿, +104.47%, net profit 14.04亿, +147.74% (first three quarters revenue 95.26亿, +106.76%); Hygon +57.58%, MetaX preannounces turnaround; IDC: 2026H1 China accelerator shipments 200.8万颗, non-GPU 101.6万颗 (+80%), surpassing GPU for first time; NVIDIA share down to 34%, Huawei up to 28% (57.2万颗) (Q3 report/IDC cited, high/medium)

Cerebras runs 3T Kimi K3 on 59 wafer-scale chips

Cerebras官方博客多源转引2026-10-26Confidence High
4-bit quantization distributed across 59 WSE-3T chips (44GB on-chip SRAM holds each layer's weights), inference 1000-2000+ tok/s, power about 1.18MW; on non-NVIDIA silicon, TB2.1 88.3%/BrowseComp 91.2% nears GPT-5.6 Sol—frontier inference bottleneck confirmed as memory and interconnect rather than FLOPs (Cerebras official blog via multiple sources, High)

FTC plans dual probes into OpenAI/Anthropic + Storm-0940 real-world operations

含METR,Axios独家,立案文书未公布2026-10-26Confidence Mid
FTC plans to issue compulsory evidence demands to two frontier vendors under Section 5 of the FTC Act (including METR; Axios exclusive; case filing not disclosed; Medium); Microsoft discloses Storm-0940 used agent framework ARTEX to breach 2250+ M365 tenants (some gained privileges in 3 hours), 25 vulnerabilities in the x402 protocol, and five major banks blocked AI agent logins (Microsoft official/The Hacker News, High/partly pending verification)

TikTok USDS in talks to buy 4GW+; tokens becoming utility-like

娱乐资本论/政采公告2026-10-26Confidence High
USDS (Silver Lake/Oracle/MGX 15% each, ByteDance retains 19.9%) in talks with compute suppliers, targeting 4GW by the early 2030s ≈ Oracle's deliveries over nearly 5 quarters ($85–134 billion scale; CNBC double-source, 4GW single-source, medium confidence); Beijing TV 3 million single-source procurement of Volcano Engine tokens (12 months), 259 broadcasters 144 tenders, key converged-media digital-intelligence projects 92.28 million (娱乐资本论/government procurement notices, high confidence)

vLLM saturates a full GB300 rack, approaching DeepSeek's in-house stack

报道台账2026-10-26
P/D separation + Wide-EP saturates GB300 NVL72 with DeepSeek-V4.1-Flash: per-GPU throughput 954k tok/s (gap vs DeepInfer <6%), context 128K→855K adds only +7% latency—the first open-source inference stack to close in on leading closed-source in-house stacks (vLLM official blog, medium confidence · no third-party replication)

Vals' first Router Index + legal leaderboard inversion

Vals/AA转引2026-10-26Confidence Mid
None of the 5 model routers made the top 10 (best Switchyard 57.8% ranks 13/27; openrouter/auto only 52.2%), Gemini 4 Argon tops at 69.2%; Muse Wonder Full tops legal index at 77.86% (Opus 5.5 only seventh); AA Harvey pass rate 13.06% vs Argon 3.89% crushes — severe inversion between general leaderboard and vertical scenarios (Vals/AA cited, Medium)

Microsoft open-sources Agensh + Mercury 3 diffusion LLM

报道台账2026-10-26
Agensh (MIT) decentralized multi-agent framework: no central orchestrator, 1024 coding agents self-organize; ProgramBench five hard tasks 19.31%→28.78% (+49%); Inception Mercury 3 diffusion LLM outputs 1068 tok/s (about 8x Luna), GPQA Diamond 89.23 overtakes Luna's 86.36, $0.35/M (MarkTechPost/机器之心, High/Medium)

上海银行 lends to AI one-person companies: operating data = credit

上海市政府官网10-252026-10-26Confidence High
静安支行 × 闹海AI includes platform OPC creators' order volume/delivery quality/payment collection data in credit-granting criteria; in 9月 it issued 300万 in entrepreneurship guarantee loans via the government-bank-guarantee mechanism—asset-light AI entrepreneurial entities receive systematic financial support for the first time (Shanghai Municipal Government official website, 10-25, High)

Capital snapshot: 亿铸 10亿 / Enact 3000万 / Lambda 40亿

综合公开报道2026-10-26Confidence Mid
亿铸科技 (3D DRAM compute-in-memory has taped out) raises over 10亿元 in new funding; Enact Intelligence 3000万美元 Pre-seed (personal follow-on by LeCun/Mensch); Lambda plans to raise up to 40亿美元 pre-IPO @ valuation 145亿—capital shifts to targets with verified tape-out/revenue and early bets on star scholars (synthesized public reports, Medium)

Deferred pool (not adopted)

报道台账2026-10-26Confidence Mid
"AWS 2026-27 capacity sold out, backlog 4960亿美元" (today is not an earnings-report day; the figure is suspected to be mixed in from earnings-expectation basis; deferred); "1Password for Agents drives annual revenue of 5亿美元" (single-source, pending verification); San Francisco court accepts OpenAI agent unauthorized-access lawsuit (Chinese compiled source, pending verification)

Long-term tracking section: inflection-point update

报道台账2026-10-26
Anthropic IPO: today 10-26 is the S-1 filing window + OpenAI discount war directly hits pricing (investor expectations fall back to 1.5万亿 range); OpenAI 300亿 funding signing window next week; both are primary tracking points for the next issue

6 million compliant cards in queue: capacity only enough for 5-10%

报道台账2026-10-24
Chinese firms (ByteDance/Alibaba/Tencent etc.) ordered over 6 million RTX Pro 5500 for about $35 billion (unit price about $11,400); NVIDIA 2027 global capacity cap about 3 million units, commercial delivery 2027Q1, currently only satisfying 5-10%—demand is 2x global capacity, compute scarcity becoming institutionalized (Reuters 10-20 first report/DW/TrendForce multi-source, High)

AI bond market spreads widen across the board + SpaceX $40 billion debt raise + Meta $825 billion commitments

彭博/FT/SEC文件2026-10-24Confidence High
Global IG spreads 129bp, widest in 18 months (CDX IG 64bp); AI debt financing exceeds $450 billion year-to-date (Janus Henderson: could be near $1 trillion next year); SpaceX plans $40 billion debt raise to buy cards ($10 billion loan + $30 billion IG bonds, Apollo-led, BBB—old news from 10-07 downgraded to background increment); Meta lease and contract commitments $700 billion→$825 billion, explicitly saying before competitors fill up financing channels (Bloomberg/FT/SEC filing, High)

Anthropic IPO countdown + OpenAI $30 billion signing (dual-king pricing)

报道台账2026-10-24
Anthropic plans to publicly file S-1 on 10-26, roadshow on 11-9, price the following week; aiming to raise up to $100 billion at up to $2 trillion, compute commitment raised to $750 billion (WSJ); New York City comptroller sends letter calling for delay; investor valuation expectations fall back to $1.5 trillion range; OpenAI $30 billion financing (UAE capital + Coatue lead) reportedly to sign next week at $1.25 trillion pre-money, positioned as a bridge before 2027 IPO (IFR/CNBC, High/Medium-High)

Kimi K3.1 Pro released + OpenRouter China 71%

财联社/OpenRouter2026-10-24Confidence High
3T parameters/1M context, TB4.0 65.7% beats GPT-6 Astra 58.2%, SWE-Marathon 51.7% beats Opus 5.5 (46.3%), plus Highspeed fast version, weights to open by 11-4 at latest; on OpenRouter that week Chinese models carried 71% of tokens (only 7% at start of year), open-source weights 77%, DeepSeek request share 24.4% ranked first (Cailian Press/OpenRouter, High)

State Council General Office AI employment-friendly guidelines: 18 items vs Barclays cutting 5000 jobs

新华社/Guardian2026-10-24Confidence High
State Council General Office Document [2026] No. 29 (issued 10-23, dated 10-16), 6 areas and 18 items: unemployment buffer period (wording under “increase support for transition and job-transfer employment”), algorithmic employment discrimination governance, job-displacement risk assessment, AI new-profession development (data annotation/AI trainer), and “small, fast, light, precise” lightweight upgrading for SMEs; Barclays’ 5000 layoffs attributed to AI (profit +12%); Gartner 2002 CIOs: only 13% see substantive returns (Xinhua/Guardian, high).

Two-track rules: Texas becomes 10th state to sue OpenAI; training banned from direct connection to the real-time internet

The Hill/官方博客/央视2026-10-24Confidence High
Paxton sued OpenAI + Altman on 10-23 (consumer protection law + SCOPE minors law; seeking to ban minors from use and require third-party evaluation and approval for new models, USD 10,000 per violation; with 8 prior states plus Florida, making it the 10th state); OpenAI announced that frontier model training is banned from direct connection to the real-time internet and switched to cached-web sandboxes (red team found VM escape/side channel; caused by a 9-20 DNS vulnerability that briefly connected to the internet; RL training resumed on 10-1); Ministry of Commerce interpretation: open-source exports “report and export immediately”; API/SaaS/device-side do not count as technology exports; DeepSeek/MiniMax have completed registration (The Hill/official blog/CCTV, high).

Google TPU first enters Meta production environment + industrial delivery line

报道台账2026-10-24
PyTorch native runtime runs TPU inference without code rewriting, with actual deployment in Meta data centers—NVIDIA's monopoly first pried open at production level; Vera Rubin NVL72 delivered in volume (inference cost per million tokens at 1/45 of the previous generation); HuiXi Intelligence raised over RMB 2.5 billion this year (robot chip shipments second globally at about 1/5); Unitree G1 cumulative production reached 10,000-20,000 units (Wang Xingxing: mass adoption requires annual output of 10 million units); Cybercab fleet 169→319 vehicles, entering Dallas in November, unsupervised FSD delayed to 2027Q2; China Mobile's RMB 3.71 billion EPC in Zhongwei, Ningxia (186MW, a different project from the 10-11 switch centralized procurement) (Tom's Hardware/The Paper/earnings call/Polaris, high/medium-high)

GPT-5.6 Sol uncovers a ten-year-old high-severity PHP vulnerability + ecosystem security layer

OpenAI/Wiz/GitHub/澎湃2026-10-24Confidence High
Two unauthenticated RCEs 'PHP2Shell' with CVSS 9.2 affect PHP 5.4-8.5; Wiz independently reproduced, and the vendor released 4 security updates—AI vulnerability discovery moves from demo to delivery; MCP reference implementation v2.0.0 + Tasks and Events working group established (Agents WG downgraded to interest group); Skills donates AAIF ecosystem data (1 million Skills/280 million installs, vendor self-reported); US DHS requires federal agencies to remove DeepSeek (OpenAI/Wiz/GitHub/澎湃, high)

Three data points: token unit price -41%, usage +50% / ChatGPT ads in 83 countries / Muse 10 million installs

报道台账2026-10-24
Ramp Enterprise AI Index: effective token unit price $1.15 in March → $0.68 in September (-41%), usage +50% MoM, a record; enterprise-side Anthropic 51% vs OpenAI 44.5%; open-source/Chinese enterprise adoption only 6.4% (room to catch up); ChatGPT ads added 26 countries to reach 83, advertisers over 100,000 (only 20,000 in August, about 5x monthly growth); Muse Wonder over 10 million active installs / about 2.5 million DAU (Sensor Tower) (Ramp multi-source citation/medium-high)

Leaderboard methodology battle: AA v5 TB Science first test + TB version confusion warning

报道台账2026-10-24
v5 adds first Terminal-Bench Science test (Astra 63.3%/Opus 5.5 61.9%) + private anti-contamination set; not comparable after v4.3.2 release; DeepSeek 1017 checkpoint lands on both HF agents-last-exam 35.4 and deep-swe 77 leaderboards; third-party comparison: same model about 76-88% on TB2.1 vs only 3-28% under TB4.0 strict criteria—check benchmark version number before citing high scores (AlphaSignal/dev.to/BenchLeader, medium-high/medium)

Capital snapshot: DeepSeek closing details / Rivian acquires world model / 微亿智造 listing

彭博/TechCrunch/Crunchbase2026-10-24Confidence High
DeepSeek deal pre-money about $70 billion (CATL plans to add 10 billion to double; new investors 君联/石溪/吉利/中芯聚源/横店); Rivian acquires Third Dimension (former Tesla/Wayve world-model team; expanding London AI center); Flapping Airplanes $1 billion Series A (led by Lux + Index; less than 1 year old); Higgsfield gets follow-on from NVIDIA + TCV (August Series B $400 million / $5.4 billion valuation extension); 微亿智造 to list in Hong Kong on 10-28 as the 'first industrial embodied intelligence stock' (Bloomberg/TechCrunch/Crunchbase, high)

Trend: lab mid-level staff shift stance en masse on AGI; extinction warnings escalate

报道台账2026-10-24Confidence Mid
Former DeepMind Alex Turner: 1/3 extinction probability within 10 years, superintelligence in 12-18 months; DeepMind safety lead Irving calls for stopping progress now (unilateral first, then US-China agreement); Hinton says AI subjective experience is 'as much' as humans; Anthropic RL lead Douglas says AGI within one or two years (same direction as Hassabis's 2-3 years); Muse Wonder pricing $0.2/M vs $200/month (single source pending verification) (NDTV/agihunt/Dealroom, medium-high/medium)
02OFFICIAL EVENTS14 stories

Can Gemini transcribe a song?

YouTube·Google Cloud2026-10-27Confidence High
Original ↗

Our newest AI Quest: Mapping the human brain

官方·Google AI Blog2026-10-27Confidence High
Original ↗

The media industry is becoming a prediction economy

官方·Databricks2026-10-27Confidence High
Original ↗

GPT-6.1 Terra is now in orbit

YouTube·OpenAI2026-10-27Confidence High
Original ↗
03RANKING MOVES8 stories

obra/superpowers stars +79,301

GitHub2026-09-03 → 2026-09-10
Original ↗

mattpocock/skills stars +71,905

GitHub2026-09-03 → 2026-09-10
Original ↗

anthropics/skills stars +49,058

GitHub2026-09-03 → 2026-09-10
Original ↗

farion1231/cc-switch stars +36,887

GitHub2026-09-03 → 2026-09-10
Original ↗
Ecosystem Pulseself-updating
2026-09-03 → 2026-09-10 Star growth:obra/superpowers +79,301★ · mattpocock/skills +71,905★ · multica-ai/andrej-karpathy-skills +59,162★ · anthropics/skills +49,058★ · Shubhamsaboo/awesome-llm-apps +38,266★ · farion1231/cc-switch +36,887★ · nextlevelbuilder/ui-ux-pro-max-skill +35,298★ · Graphify-Labs/graphify +32,474★
※Risk Levelsgray rhino / black swan
Gray rhino · foreseeable
  • Depreciation duration mismatch: GPU economic life 3—5 years vs structured debt 7—10 years; from 2027, SPV residual value shortfalls surface, with first-lien holders (chipmakers/banks) under pressure first (medium-high, 50-70%)
  • Token unit prices falling vs rigid compute rent: after mid-tier model prices halve, existing contracts charging rent per 'card-year' become a source of negative gross margin, triggering a Q4 renegotiation wave (medium, 30-60%)
  • 900% net revenue retention pullback: after the stock from budget migration is exhausted, NDR quickly reverts; if the first post-IPO earnings report falls below 300%, the damage will exceed any valuation cut (medium, 30-60%)
  • Microsoft's 60,000-person buyout execution: after coordination layers are cut, product iteration cadence slows, and AI capex's 'labor consideration' becomes 'execution leakage' (medium, 40-60%)
Black swan · low-probability
  • Major customer budget freeze: a single mega-customer accounts for an overly high share (S-1 does not disclose concentration); one budget freeze would cause a revenue cliff and cascading defaults (low, <30%)
  • Training-data licensing class actions expand: after the Elvis case (274 songs / $41 million) is upheld, rights holders sue en masse, forming an industry-wide compensation benchmark (low, <30%)
  • Fact-base contamination: AI-generated images/videos are widely circulated as evidence of public events (Apple cracked-screen image as a precedent), and regulatory/judicial rules for admitting image evidence shift abruptly (low, <30%).
✦Action Items3
1. 3-year compute TCO + contract terms audit

For self-built/leased compute, create a one-page table of “3-year total cost of ownership ÷ 3-year effective throughput”: depreciation period, utilization, 2000–4500 USD/GPU/year software annuity, and network-side increments all included. Path: free single-slot measurement for lead generation → annual audit charged by number of racks → negotiate contract terms on behalf (utilization floor, Token price cut triggers renegotiation, residual-value backstop ownership).

window Starting now, 6 weeks (Q4 procurement negotiation window); criterion: whether customers use “throughput cannot be fully utilized” to overturn the original price-comparison conclusion
2. 3-second voice simultaneous interpretation pilot (private deployment)

Use Hy-SimulMT open-source weights to run real-time meeting captions + minutes on local machines/private cloud; only three metrics: latency ≤3s, voice timbre similarity, terminology hit rate. Path: prove out a single scenario (cross-border customer service or internal weekly meetings) → add voice authorization trails and rights-confirmation chain → seat-based/duration-based subscription; risk point: voice cloning must first obtain the person’s written authorization.

window 3—6 months (early-mover period after voice interaction crosses the threshold)
3. Turn “insurability” into a delivery acceptance checklist

Conduct a pre-insurance checkup on in-house AI products: whether failure modes can be structurally described, whether the three factors of parameters/scenarios/performance can produce data, and whether the boundaries of third-party losses are clearly written. Path: self-check form → align with insurers/brokers to obtain a quote range → reverse-write rate factors into SLAs and contractual liability clauses (use externally as a selling point).

window Within 3 months (the first nationwide solution has just been released and rate distribution is undetermined); criterion: whether the first written quote can be obtained
⏱Next 3 MonthsProb.
WindowEventProb.Basis
11/4 前Kimi K3.1 Pro full weights released on scheduleHigh (85%)Moonshot AI official announcement / HuggingFace weights page
10 月底—11 月初Mistral Large 4 full weights are genuinely downloadable; license includes commercial-use restrictionsHigh (80%)HuggingFace model page + original LICENSE text
未来 3 个月Anthropic announces offering price range (SEC 424B filing finalized)High (75%)SEC EDGAR 424B4 / exchange announcement
未来 3 个月First actual rate samples for AI application risk insurance emerge (comparable)Medium (50%)Financial Street Forum follow-up / insurer product pages / broker quotes
未来 3 个月Cloud vendors cite Rubin bandwidth documents to push down prices in Q4 procurement negotiationsMedium (45%)Q4 earnings call / supply chain reports
Sources:Coverage ledger (rolling deduplication):2026-10-21 ~ 2026-10-27: 41 topic records total · AI Official Video Monitoring Log:7 days · 645 events (14 selected with links) · Site leaderboard snapshot:2026-09-03 → 2026-09-10 Top8 Star Growth · Risk/Action/Timeline:Continues from the 2026-10-27 daily report
SkillHub.news — Global Skill rankings