报告形态 · 五维分类 + 风险 + 行动 + 时间线 ← 回站点
MONTH 2026-10 · 52 STORIES · SKILLHUB MONTHLY

SkillHub.news 每月AI

2026-10 · 月报 · 素材 4
本期看点52 条 · 约21分钟
01
本月主题线MONTHLY THREADS24 条
英伟达30亿美元入股Meta定增(供应商入股客户)
02
官方事件精选OFFICIAL EVENTS20 条
My Favorite Night of the Year
03
榜单异动RANKING MOVES8 条
obra/superpowers 星标 +101,959
01本月主题线MONTHLY THREADS24 条

英伟达30亿美元入股Meta定增(供应商入股客户)

Reuters 10-272026-10-28置信度 高
Meta定增融资上调至975亿美元、获约5倍超额认购,英伟达以30亿美元入股( Reuters 10-27,高)——「卖水人把金水桶押进渠公所」:AMD/高通/维谛等供应链同步入局,双方协议涉及Vera Rubin平台与AI健康Rootly合作(高)

OpenAI官宣证伪指数时间假设(ETH)+外部学者把关

OpenAI官方2026-10-28置信度 高
OpenAI 10-26官宣Lean形式化证伪ETH(10-17起Lean官库落地,同步三层证据链:QCD冷核物质数据梳理、ETH三维预测主导、OPEC yong MTR石油行业模型),出资300万美元设Simons奖金池、外部学者委员会把关科学成果发布(OpenAI官方, 高)——千年老水则碑崩裂,新尺重测

Oracle首度引入谷歌TPU(谷歌外首家超大规模云)

Oracle/Google官宣10-272026-10-28置信度 高
Oracle Cloud World官宣第八代TPU(8版本)2027年上架OCI,容量增长714倍与zettascale/10k LPX、可用Cerebras/CambrianNova替代方案,与Google云竞争格局松动(Oracle/Google官宣10-27,高)——分渠浇田:多渠分水成新常态

Anthropic官宣收购Mithril Medicine+六大药企合作

Anthropic官方10-272026-10-28置信度 高
一人药企创始人GuC(收购)「首次智凉事件」后18个月,Mithril小团队芯片级药研管线并入Anthropic;同步与六家药企11家实验室合作(18个月内小疾病药物试验)、LIMA合作管线发布253万美元IND资料(8-49人)(Anthropic官方10-27,高)——AI自营药研管线开张

Cloudflare披露GLM-5.3跨租户KV缓存泄露

Cloudflare官方10-262026-10-28置信度 高
Workers AI的GLM-5.3共享SGLang缓存泄露155条跨租户响应(bypass bug,secoc提交,涉某不确定厂商数据;10-28已修复v0.6.22版本)(Cloudflare官方10-26,高)——千家共饮一池水,池壁缝里渗出别家的水:共享池验收合格才开闸

高盛×OpenRouter:72%流量只换27%收入

高盛报告转引10-272026-10-28置信度 中高
OpenRouter Token请求26个月增295倍、开放权重占77%;头部5%用户贡献72%流量但仅贡献27%收入,结构性「流量/收入倒挂」(高盛报告转引10-27,中高)——巨型天平:一端流量粒子山、一端金币高翘

存储三线财报超预期Q4更旺(涨价传导)

综合财报10-27/282026-10-28置信度 中高
铠侠FQ2业绩超预期(NAND涨价15-20%)、DRAM合约价2026年涨幅上修至+40%区间,AI服务器HBM挤压传统产能,三星/SK海力士/美光利润预期上修(综合财报10-27/28,中高)——温度计掉头:存储周期拐点确认

日本启动AI法:投资收增设规则

日经/共同社转引10-282026-10-28置信度 中
10-26-28日本AI法专家会议议,正式合法化2026年AI法:聚焦投资与监管规则并行(日经/共同社转引10-28,中)——富士山下青铜戒尺:投资不误立规

荷兰政府拟用AI替代2.5万个政府岗位

荷媒/路透转引10-282026-10-28置信度 中
荷兰政府预算案:2030年前用AI节省替代约2.5万个公务员岗位(占文职约16%),创荷兰首例国家级「AI替代岗位」量化目标(荷媒/路透转引10-28,中)——文件柜自动开合吐纸:第一个国家级量化替代目标

Guidewire 17亿美元收购Hyperexponential+创造科技13亿B轮

官方/36氪10-282026-10-28置信度 中高
Guidewire以17亿美元收购保险精算AI公司Hyperexponential(Agent并行开发模式);同期国内创造科技获13亿元B轮(投后估值翻倍,Agent开发平台)(官方/36氪10-28,中高)——缆绳打结:垂直行业AI并购加速

混元实时同传开源+谷歌Genie 4世界模型

腾讯/Google官方10-272026-10-28置信度 中高
腾讯混元开源Hy-SimuMT实时同传(3GB级端侧可跑、官方新5个AI场景API中4个);同场竞秀谷歌Genie 4世界模型(纯30秒生成开世界、向AI Ultra订阅者开放)(腾讯/Google官方10-27,中高)——石桥喊话声波成涟漪:多模态双线开源/商用同日

MIRI负责人Nate Soares:「毁灭概率超50%」

Time/播客转引10-282026-10-28置信度 中
Nate Soares 10-27访谈重申:AGI 1-3年内人类灭绝概率「超50%」(此前十年未变该立场)、Aster AI后续并明批「理性权衡」;MIRI内部称已放弃与实验室合作游说(Time/播客转引10-28,中)——悬崖星空:忧虑者坐得最稳的石头

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
02官方事件精选OFFICIAL EVENTS20 条

My Favorite Night of the Year

B站·可灵AI2026-10-31置信度 高
原文 ↗

Kling AI Originals|My Favorite Night of the Year

YouTube·Kling AI2026-10-31置信度 高
原文 ↗

What Will You Canva Today? | Erase Your Ex

YouTube·Canva2026-10-31置信度 高
原文 ↗

What Will You Canva Today? | Promposal

YouTube·Canva2026-10-31置信度 高
原文 ↗

What Will You Canva Today? | New Drink

YouTube·Canva2026-10-31置信度 高
原文 ↗
03榜单异动RANKING MOVES8 条

obra/superpowers 星标 +101,959

GitHub2026-09-01 → 2026-09-10
原文 ↗

mattpocock/skills 星标 +92,450

GitHub2026-09-01 → 2026-09-10
原文 ↗

anthropics/skills 星标 +63,075

GitHub2026-09-01 → 2026-09-10
原文 ↗

farion1231/cc-switch 星标 +47,426

GitHub2026-09-01 → 2026-09-10
原文 ↗
生态脉搏站点数据自更新
2026-09-01 → 2026-09-10 星标涨幅:obra/superpowers +101,959★ · mattpocock/skills +92,450★ · multica-ai/andrej-karpathy-skills +76,065★ · anthropics/skills +63,075★ · Shubhamsaboo/awesome-llm-apps +49,199★ · farion1231/cc-switch +47,426★ · nextlevelbuilder/ui-ux-pro-max-skill +45,383★ · Graphify-Labs/graphify +41,752★
※风险分级灰犀牛 / 黑天鹅
灰犀牛 · 可预见
  • AI 基建债务进入信用分层:Anthropic 预期垃圾级 + 630 亿美元芯片融资 + 英伟达替 Nscale 背书,评级坐实或利率上行会沿同一条链上移融资成本,直接压缩推理补贴空间(灰犀牛,中 30-60%,影响算力租赁报价、neocloud 估值、模型公司融资)
  • 「认购缩水/资方弃购」由个案变普遍:可灵比上限少 6.45 亿元是第一个可观测的负向偏差,若下一轮再现折价退出,一级市场对视频/基建类标的分化加速(灰犀牛,中 30-60%,影响视频生成赛道、AI 一级市场估值锚)
  • Agent 调用量 ≠ 结算量:90 亿次任务请求不披露付费转化与续费率,若续费在第二个季度掉头,Agent 交付团队的工单成本会先失控(灰犀牛,中 30-60%,影响企业 Agent 采购、服务商续约)
  • 具身 POC 单点化:98% 出自单厂单工位,跨产线/跨品类是否守住 95% 未知;按 (0.98)^n 衰减,累计良率比节拍更该进评估表(灰犀牛,中 30-60%,影响具身智能量产订单、产线采购合同)
黑天鹅 · 低概率大冲击
  • OpenAI 采购谷歌 TPU 被证实(目前单源·知情人士):英伟达「一芯独供」的定价权叙事重估,连带台积电、谷歌云预期与国内 GPU 替代叙事同时换挡(黑天鹅,低 <30%,影响算力板块、供应链定价、二级市场)
  • 系统级 Agent 收编第三方通用 Agent:超半数 C 端 AI 应用被「系统一句话」替代,估值一次性下修——Apple 与腾讯两侧同时在「替你调用别的 App」(黑天鹅,低 <30%,影响第三方 C 端 AI 应用、应用层融资)
  • 98% 口径若实为单工序成功率,全链一次通过率仅约 23%(0.98^72):返工吃掉节拍优势,出现产线退货或合同纠纷(黑天鹅,低 <30%,影响具身智能厂商、客户验收争议)
✦行动建议3
1. 按「每任务」重算成本与报价

把自家最贵、最不敢自动化的三道工作流拎出来,按 0.2 美元/任务重跑一遍 ROI;对外报价模板从「token 单价」改成「每任务成本 + SLA」。路径:先跑 200 例配对样本,错答率差超 3pp 即判负(序贯检验口径),通过后再全量切换;同步重算「检索外采」一项,否则报价系统性偏高。

时间窗 本周起 2 周内完成第一轮;盯 11 月上旬第二家是否跟进降价——两家以上才算行业换轨
2. Agent 交付「验收+返工」产品化

把交付看板固定三列:任务进入量/一次通过率/返工工时;合同级写清验收口径、返工承诺、责任边界。路径:免费交付盲区体检引流 → 订阅制返工托管 → 验收报告 + 复购锚(计量单位+计费方式+复购率)。厂商只报第一列,谁先报后两列谁拿续约钥匙。

时间窗 即刻起至 2026 Q4 合同重谈窗口;厂商若公开闭环率,报价结构会整体重构,先动手者定价
3. 算力备援清单 + 信用体检

给客户/自家做一份「算力备援清单」:主供与备份、切换窗口、迁移成本;标的估值先问担保结构与评级档位,不问签了多少 GW。把「产业资本占比/认购缩水/债务评级」列为固定观察字段,与到账金额并列记账;供应冗余自己建,不指望卖方替你买。

时间窗 Anthropic 评级正式落地前的 2–4 周;这段窗口内的再定价会先打到「单一供应商依赖」最重的客户身上
⏱未来 3 个月时间线概率
时间窗事件概率依据
至 2027-01 末11 月 4 日 Apple 智能国行 Beta 如期开放测试(千问在 Siri/写作/视觉三处的调用边界可实测)高(85%)iOS 开发者 Beta 实机、官方发布说明
至 2027-01 末12 月 iOS 27.3 正式推送,且首批仍不含 Siri AI 与 WWDC 2026 下一代功能高(80%)iOS 更新日志、Apple 官方页
至 2027-01 末Anthropic IPO 评级正式落地,档位落在垃圾级较高档(BB 类)中高(65%)评级机构正式文件、SEC 文件
至 2027-01 末DeepSeek 之后至少一家厂商跟进降价或改计价口径(跟「每任务」或对标 0.75 美元/千次检索价)中高(60%)厂商官方定价页、API 文档变更
至 2027-01 末可灵交割后披露资金用途或对赌/业绩承诺条款中(50%)工商变更、公司公告
至 2027-01 末硅基流动港股 IPO 进程推进(聆讯/招股书披露 Token 业务收入结构)中(50%)港交所披露易、证监会备案
至 2027-01 末OmniHand 在第二家工厂/第二品类复现,成功率仍守住 95% 以上中(45%)厂商公告、客户侧采访
至 2027-01 末OpenAI 采购谷歌 TPU 出现第二源证实或双方否认中(45%)双方公告、主流媒体二次确认
至 2027-01 末至少一家头部厂商公开 Agent「一次通过率/复购率」口径(而非只报任务量)中低(40%)白皮书更新、发布会实录
素材来源:报道台账(滚动查重):2026-10-01 ~ 2026-10-31 共 84 条主题记录 · AI 官方视频监控台账:31 天 · 2732 条事件(精选 20 条带链接) · 站点榜单快照:2026-09-01 → 2026-09-10 星标涨幅 Top8 · 风险/行动/时间线:承接 2026-10-31 日报
SkillHub.news — 全球 Skill 榜单
MONTH 2026-10 · 52 STORIES · SKILLHUB MONTHLY
SkillHub.news Monthly AI
2026-10 · Monthly · sources 4
Highlights52 stories · approx.21min read
01
MONTHLY THREADS24 stories
NVIDIA invests $3 billion in Meta's private placement (supplier takes stake in customer)
02
OFFICIAL EVENTS20 stories
My Favorite Night of the Year
03
RANKING MOVES8 stories
obra/superpowers stars +101,959
01MONTHLY THREADS24 stories

NVIDIA invests $3 billion in Meta's private placement (supplier takes stake in customer)

Reuters 10-272026-10-28Confidence High
Meta private placement financing raised to $97.5 billion, about 5x oversubscribed; NVIDIA takes a $3 billion stake (Reuters 10-27, high) — 'the water seller pledges the golden bucket at the canal office': AMD/Qualcomm/Vertiv and other supply chain players join simultaneously; the agreement involves the Vera Rubin platform and AI health Rootly collaboration (high)

OpenAI officially announces falsification of the Exponential Time Hypothesis (ETH) + external scholar review

OpenAI官方2026-10-28Confidence High
OpenAI on 10-26 officially announced the Lean-formalized falsification of ETH (Lean official library implemented since 10-17; in parallel, a three-layer evidence chain: QCD cold nuclear matter data compilation, ETH three-dimensional prediction dominance, OPEC yong MTR oil industry model), contributes $3 million to set up the Simons prize pool, and an external scholar committee vets scientific result publication (OpenAI official, high) — the millennium-old water gauge stele cracks; a new ruler remeasures

Oracle introduces Google TPU for the first time (first hyperscaler outside Google)

Oracle/Google官宣10-272026-10-28Confidence High
Oracle Cloud World officially announces the eighth-generation TPU (8 versions) to be available on OCI in 2027, capacity growth 714x and zettascale/10k LPX, with Cerebras/CambrianNova alternatives available, and the competitive landscape with Google Cloud loosening (Oracle/Google official 10-27, high) — irrigating fields via separate channels: multi-channel water distribution becomes the new normal

Anthropic officially announces acquisition of Mithril Medicine + partnership with six major pharma companies

Anthropic官方10-272026-10-28Confidence High
One-person pharma founder GuC (acquisition): 18 months after the 'first Zhiliang incident', Mithril's small-team chip-level drug R&D pipeline merges into Anthropic; concurrently partners with six pharma companies and 11 laboratories (small-disease drug trials within 18 months), and the LIMA collaboration pipeline releases $2.53 million in IND materials (8-49 people) (Anthropic official 10-27, high) — AI's self-operated drug R&D pipeline opens

Cloudflare discloses GLM-5.3 cross-tenant KV cache leak

Cloudflare官方10-262026-10-28Confidence High
Workers AI's GLM-5.3 shared SGLang cache leaked 155 cross-tenant responses (bypass bug, submitted by secoc, involving data from an uncertain vendor; fixed 10-28 in v0.6.22) (Cloudflare official 10-26, high)—A thousand households drink from one pool, but other households' water seeps through cracks in the pool wall: the shared pool opened only after passing acceptance

Goldman Sachs × OpenRouter: 72% of traffic generates only 27% of revenue

高盛报告转引10-272026-10-28Confidence Mid
OpenRouter Token requests up 295-fold in 26 months; open weights account for 77%; top 5% of users contribute 72% of traffic but only 27% of revenue, a structural “traffic/revenue inversion” (Goldman Sachs report cited 10-27, medium-high)—A giant balance: a mountain of traffic particles on one end, gold coins tilting high on the other

Three memory lines' earnings beat expectations; Q4 stronger (price-hike pass-through)

综合财报10-27/282026-10-28Confidence Mid
Kioxia FQ2 results beat expectations (NAND price increase 15-20%), 2026 DRAM contract price increase revised up to the +40% range, AI server HBM squeezes traditional capacity, profit forecasts for Samsung/SK Hynix/Micron revised up (combined earnings 10-27/28, medium-high)—Thermometer turns: memory cycle inflection confirmed

Japan launches AI law: investment receives additional rules

日经/共同社转引10-282026-10-28Confidence Mid
10-26-28 Japan AI law expert meeting formally legalizes the 2026 AI law: focus on investment and regulatory rules in parallel (Nikkei/Kyodo cited 10-28, medium)—Bronze ruler under Mount Fuji: investment does not delay rule-making

Dutch government plans to use AI to replace 25,000 government positions

荷媒/路透转引10-282026-10-28Confidence Mid
Dutch government budget proposal: use AI to replace about 25,000 civil service positions by 2030 for savings (about 16% of civil service roles), setting the Netherlands' first national-level quantified target for 'AI job replacement' (Dutch media/Reuters citing, 10-28, Chinese) — filing cabinet opens and closes automatically and spits out paper: the first national-level quantified replacement target

Guidewire acquires Hyperexponential for $1.7 billion + 创造科技's 1.3 billion Series B

官方/36氪10-282026-10-28Confidence Mid
Guidewire acquires insurance actuarial AI company Hyperexponential for $1.7 billion (Agent parallel development model); in the same period, domestic 创造科技 raises a RMB 1.3 billion Series B (post-money valuation doubles, Agent development platform) (official/36氪 10-28, medium-high) — rope knot: vertical-industry AI M&A accelerates

Hunyuan real-time simultaneous translation open-sourced + Google Genie 4 world model

腾讯/Google官方10-272026-10-28Confidence Mid
Tencent Hunyuan open-sources Hy-SimuMT real-time simultaneous translation (3GB-class on-device runnable; 4 of the official 5 new AI-scenario APIs); same event showcases Google Genie 4 world model (generates an open world in just 30 seconds, open to AI Ultra subscribers) (Tencent/Google official 10-27, medium-high) — Shiqiao's call turns sound waves into ripples: multimodal dual-track open source/commercialization on the same day

MIRI head Nate Soares: 'Doom probability over 50%'

Time/播客转引10-282026-10-28Confidence Mid
Nate Soares 10-27 interview reiterated: AGI human-extinction probability 'over 50%' within 1-3 years (stance unchanged over the previous ten years), Aster AI subsequently also openly criticized 'rational trade-offs'; MIRI internally says it has given up lobbying for cooperation with labs (Time/podcast cited 10-28, medium) — Cliff starry sky: the worrier sits on the steadiest stone

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
02OFFICIAL EVENTS20 stories

My Favorite Night of the Year

B站·可灵AI2026-10-31Confidence High
Original ↗

Kling AI Originals|My Favorite Night of the Year

YouTube·Kling AI2026-10-31Confidence High
Original ↗

What Will You Canva Today? | Erase Your Ex

YouTube·Canva2026-10-31Confidence High
Original ↗

What Will You Canva Today? | Promposal

YouTube·Canva2026-10-31Confidence High
Original ↗

What Will You Canva Today? | New Drink

YouTube·Canva2026-10-31Confidence High
Original ↗
03RANKING MOVES8 stories

obra/superpowers stars +101,959

GitHub2026-09-01 → 2026-09-10
Original ↗

mattpocock/skills stars +92,450

GitHub2026-09-01 → 2026-09-10
Original ↗

anthropics/skills stars +63,075

GitHub2026-09-01 → 2026-09-10
Original ↗

farion1231/cc-switch stars +47,426

GitHub2026-09-01 → 2026-09-10
Original ↗
Ecosystem Pulseself-updating
2026-09-01 → 2026-09-10 Star growth:obra/superpowers +101,959★ · mattpocock/skills +92,450★ · multica-ai/andrej-karpathy-skills +76,065★ · anthropics/skills +63,075★ · Shubhamsaboo/awesome-llm-apps +49,199★ · farion1231/cc-switch +47,426★ · nextlevelbuilder/ui-ux-pro-max-skill +45,383★ · Graphify-Labs/graphify +41,752★
※Risk Levelsgray rhino / black swan
Gray rhino · foreseeable
  • AI infrastructure debt enters credit stratification: Anthropic expected junk-grade + $63 billion in chip financing + NVIDIA backs Nscale; if the rating is confirmed or rates rise, financing costs will move up along the same chain, directly compressing room for inference subsidies (gray rhino, medium 30-60%, affecting compute rental quotes, neocloud valuations, and model-company financing)
  • ‘Subscription shrinkage / investors abandoning purchases’ goes from isolated cases to widespread: Kling being RMB 645 million below the upper limit is the first observable negative deviation; if discount exits reappear in the next round, primary-market differentiation accelerates for video/infrastructure targets (gray rhino, medium 30-60%, affecting the video-generation track and the AI primary-market valuation anchor)
  • Agent call volume ≠ settlement volume: 9 billion task requests without disclosure of paid conversion and renewal rates; if renewals turn down in the second quarter, the ticket cost of Agent delivery teams will spin out of control first (gray rhino, medium 30-60%, affecting enterprise Agent procurement and service-provider renewals)
  • Embodied AI POCs become single-point: 98% come from a single factory/single workstation; whether 95% holds across production lines/categories is unknown; decaying by (0.98)^n, cumulative yield should enter the evaluation sheet more than takt time (gray rhino, medium 30-60%, affecting embodied-intelligence mass-production orders and production-line procurement contracts)
Black swan · low-probability
  • OpenAI buying Google TPU is confirmed (currently single-source, people familiar with the matter): NVIDIA’s ‘single-chip sole supply’ pricing-power narrative is repriced, with TSMC, Google Cloud expectations, and the domestic GPU substitution narrative all shifting gears at the same time (black swan, low <30%, affecting the compute sector, supply-chain pricing, and the secondary market)
  • System-level Agents absorb third-party general Agents: more than half of C-end AI apps are replaced by ‘one sentence from the system,’ with one-time valuation write-downs—Apple and Tencent are simultaneously ‘calling other Apps on your behalf’ (black swan, low <30%, affecting third-party C-end AI apps and application-layer financing)
  • If the 98% figure actually refers to single-process success rate, full-chain first-pass yield is only about 23% (0.98^72): rework eats up the takt-time advantage, leading to production-line returns or contract disputes (black swan, low <30%, affecting embodied-intelligence manufacturers and customer acceptance disputes)
✦Action Items3
1. Recalculate costs and pricing on a per-task basis

Pull out your three most expensive workflows that you are most reluctant to automate, rerun ROI at 0.2 USD/task, and change the external pricing template from 'token unit price' to 'cost per task + SLA'. Approach: first run 200 paired samples; if the incorrect-answer rate difference exceeds 3pp, it fails (under sequential testing); after passing, switch over fully; also recalculate the 'outsourced retrieval' item, otherwise pricing will be systematically high.

window Complete the first round within 2 weeks starting this week; monitor whether a second company follows with price cuts in early November—only two or more counts as an industry shift
2. Productize 'acceptance + rework' in Agent delivery

Fix the delivery dashboard to three columns: task intake volume / first-pass rate / rework hours; at the contract level, specify acceptance criteria, rework commitments, and responsibility boundaries. Path: free delivery blind-spot health check for lead generation → subscription-based rework management → acceptance report + repurchase anchor (unit of measure + billing method + repurchase rate). Vendors report only the first column; whoever reports the latter two first gets the key to renewal.

window From now through the 2026 Q4 contract renegotiation window; if vendors disclose closed-loop rates, the pricing structure will be restructured as a whole, and first movers set pricing
3. Compute backup checklist + credit health check

Create a "compute backup checklist" for clients/yourself: primary and backup supply, switchover window, migration cost; for target valuation, first ask about guarantee structure and rating tier, not how many GW were signed. List "industrial capital share / subscription shrinkage / debt rating" as fixed observation fields and record them alongside the received amount; build supply redundancy yourself, do not expect the seller to buy it for you.

window 2–4 weeks before Anthropic's rating officially takes effect; repricing within this window will hit customers with the heaviest "single-supplier dependence" first
⏱Next 3 MonthsProb.
WindowEventProb.Basis
至 2027-01 末On November 4, Apple Intelligence China Beta opens for testing as scheduled (Qwen's invocation boundaries in Siri, writing, and visual can be tested)High (85%)iOS developer Beta on physical devices, official release notes
至 2027-01 末In December, iOS 27.3 will be officially rolled out, and the first batch still excludes Siri AI and WWDC 2026 next-generation featuresHigh (80%)iOS release notes, Apple official page
至 2027-01 末Anthropic IPO rating officially finalized, at the higher junk-grade tier (BB category)Medium-high (65%)Official rating agency documents, SEC filings
至 2027-01 末At least one vendor follows DeepSeek with price cuts or a change in pricing basis (following 'per task' or benchmarking the 0.75 USD/1,000 retrievals price)Medium-high (60%)Vendor official pricing page, API documentation changes
至 2027-01 末After the Kling transaction closes, disclose use of funds or VAM/performance commitment clausesMedium (50%)Business registration changes, company announcements
至 2027-01 末SiliconFlow's Hong Kong IPO process advances (hearing/prospectus discloses Token business revenue structure)Medium (50%)HKEXnews, CSRC filings
至 2027-01 末OmniHand replicated at a second factory/second product category, with success rate still above 95%Medium (45%)Vendor announcements, customer-side interviews
至 2027-01 末Second source confirms OpenAI's Google TPU purchase, or both parties deny itMedium (45%)Both parties' announcements, secondary confirmation by mainstream media
至 2027-01 末At least one leading vendor publicly discloses Agent 'first-pass rate/repurchase rate' metrics (rather than only task volume)Medium-low (40%)Whitepaper update, launch event transcript
Sources:Coverage ledger (rolling deduplication):2026-10-01 ~ 2026-10-31, 84 topic records in total · AI Official Video Monitoring Log:31 days · 2732 events (20 selected with links) · Site leaderboard snapshot:2026-09-01 → 2026-09-10 Star Growth Top8 · Risk/Action/Timeline:Follows the 2026-10-31 daily report
SkillHub.news — Global Skill rankings