Parrel (@ParrelWorld) header
Parrel (@ParrelWorld) profile picture

记录日常感悟、以及社会、经济和科技方面的观察

Germany
Joined November 2022
Photo by Parrel (@ParrelWorld)Photo by Parrel (@ParrelWorld)Photo by Parrel (@ParrelWorld)
The Optical Communication Ecosystem: The Infrastructure carrying the AI Data Revolution !! 👉 AI infrastructure is not just about GPUs, memory and data centres. It also depends on how efficiently data moves between chips, servers and networks. This is where Optical Communication becomes critical. 👉 Unlike conventional electrical signalling, optical communication uses light to transmit information at high speeds, over long distances and with lower signal loss. 👉 The ecosystem has four critical layers: 1. Lasers - Generate Light: Semiconductor lasers generate the optical signals used in data transmission. Key technologies: Edge-emitting lasers, DFB lasers, VCSELs and tunable lasers. 2. Optical Modulators - Encode Data: Convert electrical information into optical signals by changing light intensity, phase or polarisation. 3. Photodetectors - Receive Light: Convert incoming optical signals back into electrical signals. Key technologies: PIN photodiodes, APDs and germanium-on-silicon photodetectors. 4. Waveguides, Multiplexers & Optical Fibres - Route Data: Guide, combine, separate and transport optical signals across chips, servers and networks. 👉 Lasers generate light → Modulators encode information → Optical fibres transmit it → Photodetectors receive it → Computing systems process the data. 👉 As AI clusters expand, demand for high-bandwidth, low-latency connectivity increases. 👉 This creates opportunities across optical transceivers, silicon photonics, InP-based components, fibre-optic infrastructure and advanced optical packaging. 👉 The next phase of AI infrastructure will not be defined by compute alone. It will also depend on how efficiently that compute is connected. Disclaimer: Not a buy/sell recommendation. Do your own due diligence. @mehul_n_trivedi @Dynamicinvstr @DhawalDoshi5
2
8
1
31
1,436
미국 AI 인프라 투자 규모 2027년 말 연환산 기준 1.4조 달러(약 1,960조 원)를 돌파할 전망임 🤯 . . 골드만삭스 글로벌 리서치에서 미국 국민계정(US National Accounts) 데이터를 뜯어보고 내놓은 AI 인프라 설비투자(케펙스) 전망치가 나옴. 2022년 기준점 대비 증분만 따졌는데도 그래프 기울기가 그냥 수직 벽을 타는 수준임 ㅋㅋ 다들 AI 하면 모니터 속 챗봇이나 코드 짜주는 소프트웨어만 생각하는데, 막대그래프 구성을 까보면 사실상 인류 역사상 최대 규모의 반도체·전력·열역학 총력전이라는 게 숫자로 증명됨 ㄷㄷ • 장비(Equipment) 부문이 전체의 70% 이상 압도적 비중 차지 2027년 4분기 연환산 약 1.4조 달러 중 무려 1조 달러 이상이 순수 하드웨어 장비에 꽂힘. 그중 컴퓨터 및 서버(진녹색)가 약 6,000억 달러, 반도체(연두색)가 약 4,000억 달러 규모로 팽창함. GPU와 HBM 웨이퍼를 갈아 넣는 수준을 넘어, 수십만 장의 가속기를 초고속 네트워크로 묶어 단일 거대 슈퍼컴퓨터(클러스터)로 조립하는 서버 랙 단위의 물량 공세가 본게임이라는 뜻임. • 발열 잡으려는 공조 설비(HVAC)의 존재감 급부상 차트중간에 남색으로 낀 HVAC(냉난방공조) 비중이 2026~2027년 들어 눈에 띄게 두꺼워짐. 차세대 AI 칩 하나당 열설계전력(TDP)이 1,000W를 우습게 넘기면서 기존 데이터센터처럼 에어컨 바람(공랭식) 불어넣어서는 서버 뚝배기 다 녹아내림. 이제는 칩 표면에 냉각수를 직접 들이붓는 액체냉각(DLC)과 정밀 열교환 설비가 필수 스펙으로 자리 잡았다는 증거임. • 구조물(Structures) — 결국 AI는 전기를 지능으로 바꾸는 토목 공사임 하늘색(데이터센터 건설)과 파란색(전력 인프라 건설) 막대가 합쳐서 연환산 2,000억 달러 이상을 뚫고 올라감. 칩을 아무리 찍어내도 변전소 짓고 송배전망 깔아서 기가와트(GW)급 전기를 못 끌어오면 그냥 비싼 고철 덩어리임. 빅테크들이 원전과 가스터빈, 전력망 확충에 사활을 거는 이유가 이 그래프에 고스란히 찍혀 있음. • 지식재산권(IPP) — 하드웨어 위에 얹히는 모델과 소프트웨어 스택 맨 꼭대기 회색 영역인 IPP(연환산 약 1,000억~1,500억 달러 수준)는 파운데이션 모델 학습 알고리즘과 AI 프레임워크 개발 등 무형 기술 자산을 의미함. 물리적 인프라가 깔리는 속도에 맞춰 그 위에서 돌아가는 에이전트와 추론 최적화 소프트웨어 생태계도 같이 덩치를 키우는 중임. 요약하자면, 지금 벌어지는 AI 스케일링 법칙은 단순한 IT 트렌드가 아니라 반도체 공돌이부터 냉각 설비 엔지니어, 전력망 토목 기술자까지 영혼을 갈아 넣어 물리적 한계를 뚫어내는 거대한 인프라 혁명임. 2027년 말까지 이 속도면 컴퓨팅 파워 체급 자체가 아예 다른 차원으로 넘어갈 듯함 ㅎㄷㄷ ※ 본 게시물은 글로벌 IT/테크 산업 동향과 최신 기술 리뷰를 공유하기 위한 순수 정보 제공 목적이며, 특정 기업에 대한 매수·매도 추천이나 금융 및 투자 자문과 전혀 무관
5
17
1,586
Parrel retweeted
今天 TPU 帖里有个词值得单拎出来:KV 缓存。谷歌把这东西往片外挪,号称首 token 延迟砍九成。 模型每生成一个 token,注意力机制都要回头看一遍之前所有的 token。KV 缓存就是把前面算过的 Key 和 Value 存下来,下一个 token 直接复用,不用重算。 代价是缓存随 token 数和并发数线性长。Qwen3-235B 跑 13 万 token 的上下文,一个请求就要 23.5GB 缓存,8 个并发就是 188GB。 本地跑大模型,KV 缓存是显存的大头。上下文开得越大,并发一多,显存先扛不住。 上下文窗口能吹 1M,KV 缓存才是真账单。blog.google/innovation-and-a…
3
3
4
601
NVIDIA co-founder Jensen Huang resigned as CEO this week. Days earlier, Google announced two more TPU systems, 8tx and 8i Max. Both build on the TPU 8t and 8i shown in the card. So how different are a TPU and a GPU when the job is running an LLM? Start with what an LLM needs. Memory large enough to hold the model. Memory fast enough to feed it, because every new token goes back through the weights. And fast links between chips, because large models do not fit on one. The second need is the one people miss. When a model writes text, the chip is often waiting on memory, not on math. How fast the weights can be read puts a ceiling on tokens per second. Batching lets many users share each weight read. Total output rises, but each user's stream does not get proportionally faster. The two designs differ in where the strength sits. Rubin puts more memory bandwidth behind each GPU. Google connects more TPUs within a single pod. NVIDIA also scales past one rack, but the tightest wiring stops at the rack. Google's wiring also changes with the job. TPU 8t uses a grid suited to regular communication. TPU 8i uses shorter paths for traffic that crosses the pod, especially when tokens are routed between experts. Each choice has a cost. More bandwidth per chip means more power and cooling at each chip. A bigger pod puts more hops between chips and asks more of the software that splits the model. So the useful question is where your model spends its time: doing the math, reading memory, or moving data between chips.
2
10
1
38
2,290
Starlink's satellite bandwidth capacity has grown at an incredible pace • 2019: ~0 Tbps • 2020: 7 Tbps • 2021: 24 Tbps • 2022: 44 Tbps • 2023: 87 Tbps • 2024: 199 Tbps • July 2025: 450 Tbps • End of 2025: 612 Tbps • March 2026: >700 Tbps • November 2026: ~880 Tbps That's an incredible increase in just 7 years Elon Musk says future Starship flights will deploy over 50 Tbps of Starlink capacity and SpaceX is planning for 27 Starship launches per day by 2030 The future is going to be crazy 🚀
6
23
2
105
6,615
I guess I’m a paparazzi now, 😂
59
119
31
2,240
224,787
AI cost per token is collapsing faster than electricity, solar, computing or DNA sequencing ever did which I think is one of most bullish signals for the AI buildout. Every drop in price opens up new ways to use AI which is why demand keeps compounding and why I believe this buildout still has a long way to go.
67
121
17
1,042
82,600
SpaceX–Tesla 合并预测 英文区和韩文区出现了好几个高质量帖子,记录一下笔记: 1) 披萨的大小是固定的,SpaceX 发行 $SPCX 新股来收购特斯拉 2) 核心:交换比例 特斯拉 1 股拿到的不是“$510”,而是“SPCX 3.64 股” $TSLA 溢价越高, $SPCX 就会跌得越狠,最后是一个平衡 3) 特斯拉最可能的价位 $510~550, 对应的比例是 3.64-3.92 (按照 SpaceX 现在价格 140) 特斯拉市值略超 2 万亿美元,这是马斯克薪酬方案首阶段确定的线 溢价 28~38% 与华尔街预期相符, $SPCX 跌幅也在可承受范围内 马斯克目前拥有 SpaceX 约 48%、特斯拉约 20% 的股份 合并后马斯克占 33.6% 的股份,但有 70% 的投票权,还是牢牢控制公司(现在是 85% 投票权) 4) 合并带来的上涨空间有限 即使定在 $510(+28%), $TSLA 实际价值也仅约 $452(+14%),因为 $SPCX 会因为增发跌 11% 5) 合并带来的上涨空间有限 交易完成还需 2~3 年,因此公告后股价会因合并失败风险及最终成交的时长而打折反映,现在类似 $384~407 的水平可能性很大 换句话说, $TSLA 的现价已经 Price in 了合并溢价 6) 市场上已经有了合并套利交易 一些持有者在交易达成前卖出 $SPCX 并买入 $TSLA ,这也是近期特斯拉上涨的原因之一 7) 近年来,特斯拉散户持股比例大幅下降,因此机构投资者现主导投票,他们将同等重视治理条款和价格 $510 的价格,会让特斯拉股东合计持有合并公司 51.5% 的股份,更利于推动合并的通过 8) 比起溢价数字,更要看合并后的公司能壮大多少 以上推导假设 2 家合并后短期总市值不变,没有包含协同价值对总市值的贡献 如果假设合并可以增加约 12.8% 的协同价值,就能让原 $SPCX 股东在这个模型下回到收购前的每股 $139.95,每股 $TSLA 对应约 $542.50, 两边股东相对各自基准价都能增值
24
23
1
138
28,597
联发科拿下 Google TPU 大单,对博通影响有多大? 郭明錤今天表示,联发科有望拿下 Google TPU v9 及后续世代的重要设计订单,成为未来几年 ASIC 市场的大赢家。市场第一反应是:博通的护城河是不是要被打破了? 事情没那么简单。 先看高速互连。 博通的 200G SerDes 已经量产,102.4T 的 Tomahawk 6 交换芯片也已出货。联发科虽然完成了 448G SerDes 流片,但仍处于验证阶段。 两者不能只比较传输速率,真正重要的是功耗、稳定性以及大规模部署能力。 这方面,博通仍有明显的商业化优势。 再看先进封装。 博通与台积电合作开发的 3.5D XDSiP 平台,结合了台积电的先进制程、CoWoS 封装和 3D 堆叠技术。今年 2 月,首款采用该平台的 2nm 定制芯片已经开始出货。 联发科第一代 ASIC 同样依赖台积电 CoWoS,已经进入交付阶段。但下一代部分产品计划引入英特尔 EMIB-T,良率和规模化生产仍是挑战。 Google 选择这条路线并不难理解:台积电先进封装产能紧张,引入英特尔可以分散供应风险,也有机会降低成本。 但新技术从验证走向稳定量产,往往比预想中困难。联发科能参与解决工程问题,却无法独自决定英特尔的产能扩张速度。 真正值得关注的是 Google 的策略变化。 Google 自己掌握 TPU 核心架构,并不需要把所有设计和制造任务交给同一家供应商。 引入联发科,不仅可以分散供应链风险,还能增加议价能力。 但我不认为多一个供应商,就能轻易压低博通的长期利润率。 过去 Marvell 同样在定制 ASIC 市场与博通竞争,但博通的整体盈利能力并没有因此明显恶化。2026 财年第三季度,博通半导体收入同比增长约 127%,集团综合毛利率反而从去年同期的 67% 提升至 69%。当然,综合毛利率也受到软件业务和产品结构影响,不能直接代表 TPU 项目的利润率。 ASIC 不是可以随意替换的标准化芯片。对 Google 这样的大客户来说,交付时间、功耗、稳定性以及整个系统的性能,可能比节省几个百分点的设计成本更重要。 更何况,博通今年 4 月披露了与 Google 覆盖未来多代 TPU 的开发和供应协议,网络及其他组件的供应安排延续至 2031 年。 博通赚的不只是 TPU 设计费,还有交换芯片、高速互连、光通信和整个 AI 集群的网络基础设施。 我的判断: 联发科正在成为 Google TPU 供应链中越来越重要的玩家,未来几年有望迎来收入爆发。但拿到更多设计订单,不等于技术全面超越,更不等于博通被替代。 联发科需要证明的,是下一代复杂封装能否按计划提升良率、实现规模化交付,并最终转化为利润。 博通则需要证明,即使 Google 将更多 ASIC 设计项目交给其他供应商,它依然能够凭借高速互连和系统级网络能力,维持盈利优势。 联发科可能拿走更多 TPU 设计份额,博通却仍然可能赚到更多 AI 基础设施的钱。 这两件事并不矛盾。 $GOOGL $AVGO
5
26
7,373
Parrel retweeted
And remember …
4,429
5,853
980
62,961
3,784,367
Parrel retweeted
老黄给所有年轻人的建议 这段视频是黄仁勋在YC Startup School上对年轻人说了一番话,看完真的泪崩。加上这两天的事件~ 大家都猜测老黄会再次创业吧~ 视频里,老黄回忆创业初期的恐惧与无知,却强调:现在是过去60年最好的创业时机,AI把计算机彻底重置,机会前所未有,他甚至嫉妒年轻人。 核心心态是“这能有多难?”,别提前脑补所有困难让焦虑阻止行动,让痛苦一点一点来,靠学习与韧性一天一天熬过去。 AI 时代是历史级创业窗口,别想太多、先干起来,要用学习与韧性应对不确定。 如果你也在犹豫要不要开始,不妨先问问自己:这能有多难?
1
6
16
1,262
Parrel retweeted
Some phone videos I took from the tower
14,264
30,447
2,536
302,089
9,125,046
Parrel retweeted
ESTOS 39 MINUTOS DEL FUNDADOR DE KIMI SON ORO PURO explica cómo montó desde China una empresa de IA valorada en más de 50.000 millones hay másteres de 20.000 € que no enseñan lo que él cuenta en 10 minutos. 39 minutos, GRATIS y sin relleno. guárdalo y mira esto cuando tengas tiempo, me lo agradecerás
248
2,023
65
12,519
1,487,668
Parrel retweeted
Anthropic's agent team beat a single agent by 90%. Cursor's built a browser in a week. Other big labs are moving the same way. This is how they work, and how you can run one today.
🤝 Paid partnership
46
45
14
403
685,153
Parrel retweeted
同样站在 2009 年那个路口,有人只看见一块显卡, 而黄仁勋则看见了往后二十年整个人工智能算力产业的样子。 那时候 Nvidia 市值 40 亿,Intel 1000 亿,差了 25 倍。 所有人都笑黄仁勋不过是个卖游戏配件的。 他说了句在场没人当回事的话: PC 的杀手应用是 Word 和 Excel,所以串行架构的 CPU 是王者, 但未来电脑要处理的不是打字和算数,是图像、3D 虚拟世界、艺术表达。 这些东西全是并行任务,CPU 搞不定。 17 年后,Nvidia 5 万亿,Intel 五千多亿,25 倍的劣势,变成了接近 10 倍的反超。 但我看了两遍才发现,这条视频最狠的不是老黄预测对了 AI,他 2009 年根本没提 AI。 他预测对的是另一件事:异构计算的必然性。 CPU 管串行,GPU 管并行,两个都要,但 GPU 的相关性在上升, 这个判断后来成了现代计算的铁律——手机 SoC、AI PC、数据中心,全是这个逻辑。 而且他在 2005-2006 年就把 CUDA 押上去了,一个显卡公司搞通用计算平台,投资人觉得他疯了。 打个比方,就像在一片荒地上挖了口井,当年所有人都在笑,自来水不香吗你挖什么井, 但十几年后城市盖起来了,才发现只有你这口井挖到了最深的蓄水层——所有房子的水管都只能接你这一口。 CUDA 就是这口井,黄仁勋挖了二十年。 他没去追 Intel 的赛道,默默在在修自己的路,从图形到科学计算到深度学习到生成式 AI 到物理世界模拟——每一步迁移,这条路都在变宽,十七年后,所有的车都拐上了他修的这条路。 远见从来不稀缺,酒桌上人人都有。 稀缺的是认准之后,肯花十年时间,把一句没人信的判断,亲手浇筑成一条别人绕不过去的护城河。 今天这个路口也站满了人,有人在盯更强的模型,有人在看下一个计算平台长在哪。 我们最该盯的其实不是市值曲线,应该是创始人嘴里那个词,黄仁勋在访谈里反复说 relevance——他不纠结谁更大,只纠结自己做的事跟未来还相关吗。 我觉得这句话比任何技术判断都值钱。 2009 年人人都说 Nvidia 就是个做显卡的,跟今天有人说某家 AI 公司就是个做 XX 的一模一样。 但真正的 alpha,永远藏在对工作负载演进方向的预判里。
全盘开源!!我用腾讯 WorkBuddy 跑通了一套港美股 AI 投研工作流:复查持仓、拆解财报、死磕数据幻觉,附 4 步提示词,建议直接照着抄。 只要你平时看财报、管持仓, 以后做投研复盘可以直接拿这套人机协作当AI投研sop工作流:核心逻辑不是去让 AI 算命猜明天涨跌,而是把它当成冷酷的审计员,把每一笔数字背后的证据链彻底查清。 可能算过这笔账大家才知道这套体系有多重要: 平时看财报最怕的从来不是看错大趋势, 而是 AI 瞎编一个假数据, 或者把公司第二年重列的会计口径混着算, 你几万甚至几十万的实盘决策就当场踩雷; 而这套工作流把所有暗坑全部照亮, 亏了钱也能清清楚楚复盘! 我一直觉得AI 在投研里真正该干的就是最费人工的脏活累活。 我把它总结为给数字查户口, 每一笔数据都必须过三关: → 哪来的:出自年报哪一节,点开必须能追溯到原文 → 跟谁比:跨期会计口径一不一样,能不能直接相减 → 谁说的:是已经核实的事实,还是管理层自己的解释 户口查不清的数字,一个都不准进决策结论。 视频里是完整录屏(演示全部使用匿名合成数据), 整套实测分为极度干净的四步: 做持仓看板,没给成本坚决不瞎猜 丢入原始持仓和价格快照,按币种折算港币权重。遇到未录入成本的头寸,老老实实标出成本缺失,杜绝模型幻觉插值。 集中度动态风控 补录买入记录后自动重新折算市值。 单只持仓因加仓被稀释到 29.18% 时, 集中度预警自动撤销,调仓前后一目了然。 排查年报重列的口径陷阱 拆解业务分部增长, 遇到公司跨期重列口径时, 数字精准锚定在原始年报章节, 点开就能看到重列附注,避免假增量误导。 终极追问,三栏物理隔离 遇到应收账款异动, 强制要求它分三栏拆解: 已验证事实、公司自己的解释、还需人工核实的判断,把管理层画饼与客观事实彻底隔绝。 最后想跟大家说,投研工具从来不是水晶球,比亏钱更可怕的是没有按照自己的体系去亏,共勉! 4 步提示词原文与体验指引整理在评论区了,老规矩大家自取👇
45
281
19
1,379
424,099
所有人都在盯着财报找泡沫,结果美联储警告:负债大头根本没进表内。 刚刚发布的《金融稳定报告》里,美联储把亚马逊、Alphabet、博通、Meta、微软、英伟达、甲骨文这七家公司截至今年二季度的财报分析了一遍。除了账面上的债务,这七家还背着超过 1 万亿美元尚未起租的租约,外加超过 2 万亿美元的照付不议合同、残值担保等其他义务。 当然,这并不是财务造假。这两笔都是未来多年的未贴现金额,不能跟已经贴现的表内债务直接相加。但美联储专门算这笔账,是因为这些表外义务的增速,已经把表内债务远远甩在了后头。 以前云巨头搞基建,靠自家现金流就够了,犯不着借太多钱。直到今年,风向彻底变了;到了三季度,除微软外,亚马逊、Meta、谷歌和甲骨文的自由现金流全都转负。 今年前 10 个月,全行业为 AI 算力发债已超 6800 亿美元,是 2023 年全年(1300 亿美元)的五倍多。光公开市场上没还的算力债就超过 1.1 万亿美元,私募信贷里还有不少算不清的大单。 七家去掉博通、英伟达,剩下五家云巨头,2020 年以来的累计资本开支占 GDP 比重,眼看就要盖过 90 年代那场著名的电信宽带狂潮。 既要铺开这么大的摊子,又不想让财报太难看,大厂在亲自发债之外,还玩了两套表外财技:一是云厂商跟第三方开发商签长租约,让开发商揣着租约去借钱动工;二是芯片厂给买自家 GPU 的算力或大模型公司做残值担保,巨头再顺手投笔股权,凑成美联储所称的“循环融资”。 前阵子《华尔街日报》扒出的 630 亿美元交易就是活教材:一家壳公司出面借钱买博通和谷歌设计的 TPU 芯片,靠 Anthropic 交租金还贷,博通用自家的资产负债表给银行担保;因为借款方是壳公司,只有 Anthropic 付不出钱时博通才承担风险。 顺着资金链往下摸,巨头的表内债主要在共同基金、养老金和保险机构手里;表外交易搭起来的 SPE(特殊目的实体)则把债务切成优先和次级两档,风险更高的次级债流向了高收益债基金、寿险公司和私募信贷。 华尔街几家最大的银行,对科技巨头、数据中心开发商、供电企业,以及相关 SPE 和私募信贷的敞口,合计也已逼近 1 万亿美元。 美联储也是颇有点黑色幽默:它说七巨头替弱信用小弟兜底,是替金融体系挡了一刀,一旦行业下行,巨头可能把亏损扛下来,反而减轻对其他人的冲击;可同一段话里也承认,巨头真实的风险敞口,可能比账面负债大得多。
28
87
13
341
41,590
🇺🇸 Изучив финансовую отчетность Nvidia, журналисты The Wall Street Journal обнаружили, насколько рискованной была бизнес-стратегия Дженсена Хуанга. За последние несколько лет глава компании выстроил систему, в которой Nvidia сама финансировала спрос на собственную продукцию. Схема была довольно простой. Nvidia предоставляла многомиллиардные кредитные гарантии небольшим ИИ-компаниям, которые на эти деньги строили дата-центры и закупали ее чипы. Если стартапам не удавалось найти клиентов на свои вычислительные мощности, Nvidia гарантировала их выкуп. Таким образом, в рамках одной коммерческой сделки компания могла одновременно выступать инвестором, гарантом по кредиту, поставщиком оборудования и покупателем вычислительных мощностей. Nvidia фактически сама создавала рынок сбыта для своей продукции, обеспечивая рост выручки и собственной капитализации. Теперь эта конструкция оказалась под угрозой: крупнейшие технологические компании начали переходить на чипы собственной разработки, снижая свою зависимость от Nvidia. Между тем компания уже взяла на себя контрактные обязательства по закупкам у поставщиков на $279 млрд, а общий объем ее финансовых и контрактных обязательств, включая условные гарантии, превышает $530 млрд. Иными словами, Хуанг поставил будущее Nvidia в зависимость от непрерывного роста спроса на искусственный интеллект, который компания во многом сама и финансировала. Если этот спрос начнет сокращаться, расплачиваться за агрессивную стратегию бывшего руководителя придется уже его преемникам. Теперь Nvidia оказалась перед крайне неприятным выбором. Продолжение политики Хуанга грозит дальнейшим ростом финансовых обязательств, значительная часть которых при неблагоприятном развитии событий может превратиться в реальные долги. Отказ же от этой схемы способен спровоцировать цепочку банкротств зависимых от Nvidia ИИ-стартапов, падение спроса на ее чипы и обвал акций самой компании. Причем последствия такого обвала рискуют выйти далеко за пределы Nvidia, потянув за собой весь американский технологический сектор.
121
281
36
1,284
216,563
Parrel retweeted
Jensen Huang is the Steve Jobs of Nvidia. He is irreplaceable .
1,902
5,913
729
85,431
4,064,949
(2/2) Jensen Huang's resignation — what it means for NVDA and the AI industry The resignation — verified: Thursday, November 12, around 8:48 AM ET, Nvidia announced Jensen Huang resigned as President, CEO and board member, effective immediately, ending a 33-year tenure. He co-founded the company in 1993. Interim CEO: CFO Colette Kress (EVP + CFO since September 2013, keeping the CFO role); the permanent search is internal and external. Harvey C. Jones was appointed the company's first-ever Chair of the Board. The 8-K: the board accepted the resignation on November 11, effective immediately; the company gave no reason. "Firing" is Rep. Moolenaar's word, not the fact — his own statement's title says "Resignation." Market reaction: Wednesday's close of $173.73 → pre-market $157.54 (-9.32%, about $400 billion of market value erased) → Thursday's close $165.77 → Friday $162.85 (-2.01%). NVDA fell 12.75% on the week. Analysts: Ives $225 (without Huang) vs $300 (with him); Stifel $315 → $205; BofA $350 → $280; William Blair cut to Market Perform. Rotation: AMD +18.93% on the week (Friday -8.75% to $534.06), AVGO up about 19% on the week (Friday $347.01, -4.76%). The NVDA–AMD spread is about -$371. Red pen: AMD trades at 38.44x forward earnings vs NVDA at 13.72x; the scale gap is about 11–13x (NVDA's data-center quarter is about $75 billion, AMD's $6.7 billion). Three structural breaks for the AI industry: The founder premium is gone: "Jensen is Nvidia" no longer holds, and the market has to re-price succession uncertainty. A succession vacuum: Kress is an interim CEO from finance, with no product or visionary pull. In your framework: "NVDA doesn't need an operator, it needs a visionary." The scrutiny continues: the Senate and House will investigate the Trump administration's approval of Nvidia chip exports to China; Huang has agreed to testify before Congress on the Alice Huang matter before Thanksgiving (no date set). The Warren + Coons preservation letter requires confirmation by November 17. Rumors and investigations (reported, not established): The NYT reported an extramarital relationship with Kimberly Powell, his direct report; Reuters reported an intimate relationship with a second employee (two sources, unnamed). Both are allegations, not findings. The Alice Huang spy case: CNN reported that intelligence to date "does not prove" she spied for Beijing (not an exoneration); the FBI probe is ongoing, with no official finding. CNBC reported she owned stakes in Chinese businesses that provided computing power to Huawei (source reporting, not a finding). Your rule has fired: the November 11 rule — liquidate NVDA holdings once the CEO is replaced — has fired. The $200 calendar spread (long Jun 17, 2027 $200 calls / short Dec 24, 2026 $200 calls, $12.18 debit) is confirmed held, with a maximum loss of $1,218 per spread. The next test: NVDA earnings Tuesday after the close, Kress's first print facing analysts — watch data-center demand, H200 sales to China, and guidance. Earnings breakdown: CoreWeave (CRWV) — beat-and-raise, sold on leverage Q3 (Nov 10, after the close): revenue $3.661 billion, +168% YoY, beating consensus of about $3.5 billion and the company's $3.4–$3.6 billion guide; GAAP EPS -$1.27, beating Zacks' -$1.45 (Finnhub shows -$0.95, a different basis); net loss $709 million. Profits: adjusted operating income $379 million, beating the $200–$260 million guide; GAAP operating income $207 million. The Sankey: revenue $3.7B → gross profit $2.5B (68%) → operating profit $0.2B (6%) → interest $0.8B + tax and other $0.1B → net loss $0.7B. Red pen: Q3 net interest expense of $846 million is more than 4x operating profit; Q4 interest is guided at $1.05–$1.15 billion; debt is $46.9 billion. R&D of $2.0 billion is 55% of revenue. 98% of revenue comes from committed contracts ($3.6B) — revenue quality is high; the problem is cost and leverage. Guidance: backlog $132.5 billion (+138% YoY); FY26 revenue guided up to $13.0–$13.3 billion; capex $37–$39 billion; active power over 1.7 GW; operating cash flow $1.975 billion; cash $6.5 billion. The stock: Nov 11 close $60.77 → Nov 12 $56.60 → Friday intraday $52.27 (-7.66%, low $49.24). A beat-and-raise sold off — that's the leverage story, not the revenue story. The slide's numbers were checked against the official print; US $3.4B + rest of world $0.2B sums to $3.6B vs the $3.7B total — a $0.1B rounding difference. Earnings breakdown: Nebius (NBIS) — a big beat, and the stock fell Q3 (Nov 12, before the open): adjusted EPS $0.31, revenue $980.8 million; revenue +571% YoY (Q3 2025: $146.1 million) and +68% QoQ (Q2 2026: $582.3 million). Adjusted EBITDA $465.8 million. Slide check: EPS consensus -$0.63 vs actual $0.31 (a 149% beat); revenue consensus $907.17 million vs actual $980.8 million (an 8% beat). The revenue consensus varies by vendor — the slide's $907.17 million matches MarketBeat, Finnhub's $925.68 million is the high outlier; the beat is large either way. Red pen one: the slide's "EPS +178% YoY / +358% QoQ" uses the absolute-negative-base convention — the reality is a swing from loss to profit (YoY -$0.40 → +$0.31; QoQ -$0.12 → +$0.31). Don't say +178%. Red pen two: FY26 revenue guidance was cut to the low end of the $3.0–$3.4 billion range (FactSet $3.35 billion); 2026 capex was raised to about $27 billion (from $20–$25 billion); 2027 capex is $40–$45 billion, far above the expected $28.77 billion; $7 billion-plus ARR was pushed from December to February; about $325 million of Q3 revenue was non-recurring. Prepayments are the bright spot: Q3 new deals had prepayments covering over 100% of related capex (Q2: 50–60%); over 90% of Q3 contract value carries a prepayment (up about 20 points); deferred revenue on the balance sheet is $9.4 billion (from $1.6 billion at end-2025); payback on 1–3 year deals is 19 months (Q2: 22 months); 2027 prepayments are expected at $17–$20 billion (2026: $10 billion). Customers are paying to fund the build-out. The stock: Nov 11 close $154.69 → Nov 12 $128.64 (-16.84%) → Friday intraday about $117 (-8.92%, low $109), about -24% in two days; one month -37.01%; 52-week high $299.86. NBIS continued: 2027 revenue guidance is about 4x — but the market is pricing capex Revenue trajectory: 2024 $92 million → 2025 $530 million → 2026 guide about $3.0 billion (low end of the range; analysts expect $3.37 billion) → 2027 guide $12–$13 billion, about 4x 2026 (analysts expect $11.98 billion). Q3 was $981 million, +571% YoY, beating the expected $908 million. Quarterly outlook (VJN estimates): Q3 2026 about $0.9B, Q4 about $1.5B; 2027 Q1 about $2.2B, Q2 about $2.7B, Q3 about $3.2B, Q4 about $4.0B; 2028 Q1 about $4.6B, Q2 $5.2B. Everything after Q3 2026 is an estimate (e), not actuals. Why a big beat still fell: revenue and EPS both beat big, but guidance was cut, capex jumped, ARR was delayed, and $325 million of revenue was non-recurring — the market is pricing the speed of the cash burn, not the speed of revenue. Prepayments covering 100%+ of capex are a guardrail, not a guarantee. Against CRWV: both are "revenue exploding, interest and capex exploding faster" AI infrastructure stories; CRWV's interest is 4x its operating profit, and NBIS's capex guide is about 50% above expectations. This week the market chose to sell first and ask later. Anthropic IPO — the price range is out; the seven giants are putting in $26 billion Price range: $70–$78 (S-1/A No. 2, Nov 10); about 825 million shares offered, raising $57.75–$64.35 billion; on shares outstanding the valuation is about $916 billion–$1.02 trillion, midpoint about $968 billion — about equal to the May Series H mark of $965 billion. Fully diluted: at $78, up to about $1.404 trillion (TechCrunch) — a different yardstick from the $1.02 trillion on shares outstanding; both numbers are real. The $26B concurrent placement (Class N non-voting shares, at the IPO price): Alphabet $10 billion, NVIDIA $10 billion, AMD $2.5 billion, Broadcom $1 billion, Micron $1 billion, Microsoft $1 billion, TSMC $0.5 billion — exactly $26.0 billion in total. All seven buyers are also Anthropic's chip and cloud suppliers. Timing: pricing reportedly Nov 18, first trade Nov 19 — unconfirmed; the filing carries no date. The founders keep 50.1% voting control through a single Class F share; the lock-up releases in stages (some non-founder insiders can sell 10% on day one). Anthropic results: Q3 revenue $16.96 billion, compute ate 58% Q3 2026: revenue $16.956 billion, +1,127% YoY, +47% QoQ (Q2 was +144% — growth is slowing); compute and infrastructure cost $9.863 billion, 58% of revenue; GAAP operating loss $567 million; adjusted operating income $2.463 billion, a 15% margin. The Sankey: revenue $17.0B — consumption-based $13.2B (78%, API and app token usage), subscription $3.7B (22%, Claude fixed-fee plans); customers over $100K: 11,900+ (+1,153% YoY), net revenue retention 900%. Operating expenses $17.5B: compute $9.9B (58%), personnel $1.7B (10%), sales and marketing $2.4B (14%), other $3.6B (21%); operating loss $0.6B (-3% margin). GAAP net loss of $23.29 billion is mostly a $21.9 billion non-cash note revaluation — not cash burned. Nine-month revenue is $33.24 billion, net loss $104.04 billion. Long-term commitments are about $900.9 billion (about $70.7 billion due within a year); the annualized run-rate at Oct 31 is about $81 billion; net proceeds are $84.6–$93.7 billion (including the placement). Your position: you hold pre-IPO Anthropic shares (via Forge Global) and will not chase the IPO. On Nov 9, Claude Fable 5.2 + Mythos 5.2 launched ($8/$40 per million tokens, 20% cheaper than 5.1). Against OpenAI: the same quarter, OpenAI had revenue of $10 billion and an adjusted operating loss of $10.9 billion (the loss exceeded revenue); Anthropic had revenue of $17 billion and an operating loss of $567 million. Both are burning money — Anthropic's unit economics are closer to breakeven. Next week: NVDA and WMT earnings; the FOMC minutes are the key NVDA earnings (Tuesday after the close): Kress's first earnings as interim CEO, facing analysts. Watch three things: data-center demand, H200 sales to China, and guidance. The first print since the resignation — the market wants to know whether "NVDA without Jensen can still deliver." November 17 is also the date in the Warren + Coons preservation letter; the congressional probe is heating up. WMT earnings (Thursday before the open): Q3 FY27, released 6:00 AM CT Thursday, call at 7:00 AM CT (CEO John Furner + CFO John David Rainey). Consensus EPS $0.64 (Zacks, +3.2% YoY), revenue about $186.52 billion (Zacks) / $187.27 billion (TipRanks). Options imply about ±7%; last time (Aug 20) the stock moved just +0.10% the next day. Watch: high-income customers, e-commerce, and the holiday-season outlook. WMT rose 1.51% on Friday — one of the few green defensive names. Next-week macro calendar: Mon 11/16: Empire State manufacturing (Nov); Fed Chair Warsh speaks at 12:00 PM; Williams speaks. POWL earnings after the close; SGRX begins trading split-adjusted (1-for-8). Tue 11/17: retail sales (Oct) 8:30 AM, +0.2% expected; industrial production 9:15 AM; NAHB housing index. NVDA earnings after the close; HD and DIS before the open. Wed 11/18: housing starts and building permits (Oct) 8:30 AM; FOMC minutes 2:00 PM. Anthropic pricing reportedly due; TGT, TJX and LOW before the open; PANW after the close. Thu 11/19: initial jobless claims 8:30 AM; Philly Fed manufacturing; new home sales. WMT earnings before the open; PDD and BABA before the open; Anthropic first trade reportedly due. Fri 11/20: Michigan consumer sentiment (Nov, final) 10:00 AM. SPCX Nov 20 $144 calls expire; FOMC meets Dec 8–9. Note: the earnings-calendar graphic has two small errors (Monday after the close, "MONTAUK" is labeled BEKE; Wednesday before the open, ICCM appears twice) — neither affects the NVDA / WMT dates. This week's key economic data Two stories this week: inflation isn't cooling fast enough, and housing gave way first. Labor is still tight; long-end demand is weak. Tue 11/10: CPI MoM (Oct) +0.4% actual vs +0.3% expected (prior +0.6%) Tue 11/10: Core CPI MoM (Oct) +0.3% actual vs +0.2% expected (prior +0.3%) Thu 11/12: Initial jobless claims (week ended 11/7) 200,000 vs ~195,000 expected (prior 189,000) Thu 11/12: 30-year Treasury auction ($25B) tailed 0.4bp Thu 11/12: Existing-home sales (Oct, SAAR) 3.75 million vs 3.85 million expected (prior 3.97 million) Thu 11/12: UK GDP QoQ (Q3, first estimate) +0.7% vs +0.7% expected (prior +0.5%) Fri 11/13: PPI MoM (Oct) +0.3% actual vs +0.4% expected (prior revised up to +1.4%) CPI (Tuesday, hotter than expected): core is +0.3% for a third straight month, about 3.7% annualized — far above the 2% target; shelter +0.3%, food +0.5%, energy +0.7%. Market reaction: gold fell about 2.5%, the 10-year jumped 12–13bp to about 4.99%, briefly touching 5.01%. Waller: a +0.3% core would make him consider a 50bp hike in December. CPI year-over-year red pen: the BLS published no 12-month change for October — October 2025 prices were never collected because of the government shutdown, so there is nothing to compare against. The 3.7% / 2.6% on the calendar graphic are September's official numbers, not October's. The Cleveland Fed nowcasts October YoY at about 3.8% headline / 2.6% core; an independent BLS-method estimate is about 4.0% / 2.8%. Year-over-year returns with November CPI (published Dec 10). PPI (Friday, cooler on the surface): MoM +0.3%, below the +0.4% consensus — but September was revised up to +1.4% (from +1.3%), so the base is hotter; YoY +6.5%, the biggest since November 2022. After PPI, December hike pricing is about 88% for +25bp / 12% for +50bp; the long end did not rally on the PPI. Labor and the auction (Thursday): initial claims of 200,000 beat the ~195,000 consensus but are still below the 200,000 line (as they have been since mid-September); the prior 189,000 was the lowest since 1969. The $25B 30-year auction tailed 0.4bp (high yield 5.441%), with dealers taking 11.4% (highest since June) — weak long-end demand is part of the bear steepening. Housing (Thursday, a 16-year low): existing-home sales of 3.75 million SAAR, -5.5% MoM and -8.8% YoY, the lowest since August 2010 and below the 3.85 million consensus. Inventory is 1.59 million homes (5.1 months of supply); the median price is $416,700 (+0.4% YoY). NAR chief economist Yun: mortgage rates are the highest since 2000 (except one month in 2023); the October 30-year mortgage averaged 7.33% (September: 6.86%). October pending home sales (out Nov 18) are the next check; HD and LOW report next week. UK GDP (Thursday, as expected): Q3 first estimate +0.7% QoQ, matching the Bank of England's November forecast and above Q2's +0.5%; YoY +1.9%. Services +0.9% led, production +0.1%, construction -0.7%. September MoM +0.2%, August revised to 0.0%. First estimates get revised (mean absolute revision 0.27pp); the next estimate is Dec 22. Little direct impact on US stocks; the global growth backdrop is mildly positive. The week's verdict: inflation is not out of control, but it is not cooling enough to breathe easy either — core is +0.3% for a third straight month and a December hike is fully priced; the long end is steepening on term premium plus weak auction demand, not on growth panic; labor is still tight while housing is the first to break under high rates. Next week's data tests: Tuesday retail sales, Wednesday's FOMC minutes. Sources: the user's economic calendar graphic (Trading Economics basis); CPI / PPI from the BLS, existing-home sales from NAR, UK GDP from the ONS, initial claims and the 30-year auction from verified US Treasury / Labor Department figures. The CPI YoY column was corrected after cross-checking — see the red pen above. This week's personal trade log — 13 trades, all long calls NVDA calendar spread, 100 spreads: long Jun 17, 2027 $200 calls / short Dec 24, 2026 $200 calls — $12.18 debit. Short leg expires Dec 24 (41 days). Friday close $162.85. Target $200 (+22.8%). Confirmed held; max loss $1,218 per spread. AVGO Jan 21, 2028 $370 calls — $52.12. Expiry Jan 21, 2028 (434 days). Friday close $347.01. Breakeven $422.12 (+21.6%). 6.6% out of the money; 52-week high $495. MRNA Jun 17, 2027 $370 calls — $47.70. Expiry Jun 17, 2027 (216 days). Friday close $223.37. Breakeven $417.70 (+87.0%). 65.6% out of the money; the card's breakeven was $399.50 — differs from the log. POWL Dec 17, 2027 $180 calls — $31.50. Expiry Dec 17, 2027 (399 days). Friday close $142.72. Breakeven $211.50 (+48.2%). 26.1% out of the money; earnings Monday after the close. INTC Dec 11, 2026 $75 calls — $0.71. Expiry Dec 11, 2026 (28 days). Friday close $59.66. Breakeven $75.71 (+26.9%). 25.7% out of the money; +27% in 28 days — a lottery ticket. SKHY Feb 19, 2027 $140 calls — $15.00. Expiry Feb 19, 2027 (98 days). Friday close $131.79. Breakeven $155.00 (+17.6%). 6.2% out of the money; spans the Jan 28 earnings. MRNA Dec 18, 2026 $320 calls — $6.70. Expiry Dec 18, 2026 (35 days). Friday close $223.37. Breakeven $326.70 (+46.3%). 43.3% out of the money; +46% in 35 days — a lottery ticket. INO Jan 15, 2027 $1.50 calls — premium not provided. Expiry Jan 15, 2027 (63 days). Friday close $1.43. No breakeven (cost not provided). 4.9% out of the money; the closest to the money on the whole sheet. INTC Mar 19, 2027 $80 calls — $4.10. Expiry Mar 19, 2027 (126 days). Friday close $59.66. Breakeven $84.10 (+41.0%). 34.1% out of the money; spans the January earnings; matches the card's $4.09. TSLA Jan 15, 2027 $500 calls — $4.20. Expiry Jan 15, 2027 (63 days). Friday close $390.22. Breakeven $504.20 (+29.2%). 28.1% out of the money; expires before the Q4 print — not an earnings ticket. AMD Feb 19, 2027 $700 calls — $20.70. Expiry Feb 19, 2027 (98 days). Friday close $534.06. Breakeven $720.70 (+34.9%). 31.1% out of the money; breakeven above the 52-week high of $658.52. AMD Feb 19, 2027 $870 calls — $6.05. Expiry Feb 19, 2027 (98 days). Friday close $534.06. Breakeven $876.05 (+64.0%). 62.9% out of the money; your $6.05 is below the card's $7.00; the most extreme leg on the book. TSLA Dec 15, 2028 $900 calls — $23.35. Expiry Dec 15, 2028 (763 days). Friday close $390.22. Breakeven $923.35 (+136.6%). 130.6% out of the money; a LEAP spanning about 8 earnings cycles; the card printed $22.80. Overall read: All long options: 13 trades, and except for the NVDA calendar spread, all are long calls — no puts, no short sales. One direction, one kind of leverage. Concentrated in AI / semis: AVGO, INTC x2, SKHY, AMD x2, NVDA and POWL (data-center power equipment) are 9 of the 13; MRNA x2, INO and TSLA x2 are the satellites. Three time layers: near-month lottery tickets (Dec 11 INTC $75 and Dec 18 MRNA $320, needing +27% / +46% to break even); February earnings tickets (SKHY $140, AMD $700, AMD $870, spanning Q4 earnings); long-dated thesis tickets (AVGO $370, POWL $180, TSLA $900 LEAP). Card vs log differences: the MRNA $370 call card implies a premium of about $29.50 (breakeven $399.50) vs your logged $47.70 (breakeven $417.70) — go with your cost. AMD $870 calls: your $6.05 is below the card's $7.00; TSLA $900 calls: your $23.35 is above the card's $22.80; SKHY $15.00 is below the card's implied $17.20; POWL $31.50 is slightly above the card's implied $30.09. The NVDA calendar spread: your November 11 rule (liquidate NVDA once the CEO is replaced) has fired, and this spread falls inside it. The short $200 leg is 22.8% out of the money with 41 days left — time is on the short leg's side; the long leg has 216 days and wants NVDA to repair toward $200. Friday closes: Finnhub (Nov 13, 2026 close, cross-checked across sources). Premiums and strikes are quoted verbatim from your fill records; INO's cost was not provided, so there is no breakeven for it. Next-week battle checklist Sat 11/14, about 1:17 PM: Starship Ship catch attempt — the biggest remaining catalyst for the SPCX Nov 20 $144 calls (breakeven $150.00); the booster already missed, so a full double catch is off the table. Mon 11/16: POWL earnings after the close (your Dec 17, 2027 $180 calls are held); SGRX begins trading split-adjusted (1-for-8); Empire State manufacturing; Warsh speaks. Tue 11/17 after the close: NVDA earnings — the week's biggest event. Retail sales at 8:30 AM before the open (+0.2% expected). SPX key levels: up through 7,622.50, down through 7,500.12. Wed 11/18, 2:00 PM: FOMC minutes — the most yield-moving item of the week. Anthropic pricing reportedly due. 10-year: watch R1 51.80 / S1 51.21. Thu 11/19 before the open: WMT earnings (consensus EPS $0.64, revenue about $186.5B). Anthropic first trade reportedly due. Initial claims at 8:30 AM. Fri 11/20: Michigan sentiment (final) at 10:00 AM; SPCX Nov 20 $144 calls expire; FOMC meets Dec 8–9. Semis: SMH — up through R1 500.26, down through S1 473.25; SOXX — up through R1 154.23, down through S1 144.50; the 200-day (SMH 523.77 / SOXX 160.27) is the real resistance. Volatility: VIX — up through R1 20.93, down through S1 17.65; MOVE — up through R1 100.89, down through S1 97.56. "True fear" needs VIX above 20 AND MOVE above 100 at the same time. Rates: the 2-year reclaiming 4.868 or breaking 4.826 is the directional signal; the 5-year is about the 5.00% line; the long end is pressed against the upper Bollinger band — don't chase. Position discipline: all 13 trades are long calls — one direction. The two near-month lottery tickets (Dec 11 INTC $75, Dec 18 MRNA $320) have only 28 / 35 days left; the thing to manage is time, not direction. One line: the broad market sits on its pivot, semis sit below theirs, rates sit on the upper band, volatility sits at the door of 20 — next week's answer comes Tuesday after the close from NVDA, not from Monday's open. Appendix: this week's earnings calendar (week of Nov 9) Earnings recap: CRWV (Tue after the close): revenue $3.661B beat, adjusted operating income $379M beat guidance — but interest is 4x operating profit, and the stock was sold. NBIS (Thu before the open): revenue $980.8M (+571% YoY), EPS $0.31, a swing from loss to profit — but guidance was cut and capex jumped, and the stock fell about 24% in two days. Thursday after the close, all five graded: CSCO revenue $18.5B, non-GAAP EPS $1.40 beat, FY27 guidance raised; AMAT EPS $4.16, revenue $10.41B, a ninth straight beat, guidance raised; NU EPS $0.22 in line, revenue $5.82B a small miss, 90-day delinquency up to 7.3%; DLO EPS $0.23 beat, revenue $398M miss; CBRS GAAP revenue $210.8M miss, but core revenue $230.9M beat, FY26 core revenue guidance raised. Friday before the open: PAVM, SBC, RMX, MIST, NDLS, EOC, EIC, BRAG, BSP, WRD. Earnings Whispers calendar, week of Nov 9. The calendar graphic has two small errors (Monday after the close, "MONTAUK" is labeled BEKE; Wednesday before the open, ICCM appears twice) — neither affects the CRWV / NBIS dates. [NOT FINANCIAL ADVICE, DYOR!] from Humble Trader | Gemini Trading
2
6
7
1,442