Three claims I see on here 100x a day: 1) SpaceX has more compute than anyone. It should reach ~2 GW by year-end. AWS and Microsoft are each adding ~1 GW per quarter globally, with Google and Meta likely in the same range. 2) BTM can add 5–10+ GW annually. Yet SpaceX, supposedly the fastest BTM scaler, will reach ~2 GW after 1–2 years. 3) Anthropic is reliant on SpaceX. Anthropic disclosed >300 MW at Colossus I. SpaceX later said the agreement also covers Colossus II, but has not disclosed the total MW. Dylan Patel estimates Anthropic scales from <2 GW to >5 GW this year. These claims do not reconcile!

🚨 Cursor CEO response to OpenAI: >OpenAI models = 5% of Cursor traffic >translation: we don’t need you >“we’ve trusted [OpenAI] platform to be neutral infrastructure” >translation: OpenAI just proved they’re not neutral It’s OVER

I began managing people in my 20s. After 48 hours using Grok Bot, one thing is clear: access to a team of 24/7/365 digital employees has just been democratized from a usability standpoint. Next hurdle: price.

Wow, check this out! Ship 40 on FORTE has risen. Chris Bray made a sweet little timelapse. Looks so overkill for the upper stage, but I'll take it. 😆 Again, make sure you follow Chris as there are plenty more shots from today to check out. Fun times! 🫶

HBM越堆越高的逻辑，可能要反转了 瑞穗证券昨晚发了份研报——他们的分析师刚跟SemiAnalysis总裁Doug O‘Laughlin吃完晚饭，信息量很大。 美股投资网挑几个最核心的结论先说： 内存还是AI最大的瓶颈。但这个瓶颈的形状，正在变。 HBM层数可能不升反降。GPU出货量反而能增加20-30%。 听着矛盾。往下看你就明白了。 HBM层数：风向在变 瑞穗刚把美光目标价从1375美元砍到1300美元，闪迪从1900砍到1875。应用材料从650砍到590，泛林从370砍到365。 砍价理由不是需求不行——是 "DRAM降规风险" 。 Doug跟我们聊到一个关键判断：GPU厂商未来可能在HBM上只堆4-8层，12层的吸引力在下降。16层、20层HBM和混合键合，在未来的DRAM世代里充满不确定性。 市场一直预期HBM会越堆越高。 现在剧本可能反过来。 一个关键的逻辑反转：HBM少了，GPU反而更多 Doug的原话是：HBM转换比率在下降，从3:1降到5:1。 逻辑链条非常直接： 每颗GPU用的HBM层数减少 同样数量的DRAM晶圆，能造出更多GPU GPU/AI芯片出货量可能增加 20-30% 更多GPU出货 → 更多训练和推理需求 → KV Cache总DRAM位元需求反而增长 少即是多。 内存含量下降，出货量上升，总需求不降反升。 铜缆和光互联：正在抢跑 Doug对铜缆的态度很明确：AEC是AI机架内连接的核心方案。 CRDO走在最前面。 Lumentum那边，NPO（近封装光学）被管理层定位为"增量机会"，预计2027年底到2028年进入市场。InP衬底的需求已经严重超过供应。 CPO（共封装光学）呢？还得等2-3年。 短期赢家是铜缆AEC和NPO/SiPh/InP。CPO还早。 800V可能推迟 聊到数据中心电力方案的时候，Doug提了一个很有意思的观察：数据中心运营商对800V高压直流方案有抵触。 原因很实际——变化太大，员工需要重新培训。 数据中心正在转向±400V双极双路机架。在传统400V直流布局上爬坡更容易，对SiC和辅助设备的需求反而减少。 800V推迟 ≠ 电力需求减少。只是换了一条更慢、更稳的路。更多的电压调节模块会补上来，继续支持GPU出货。 2027年的两个大数字 Doug还提到了2027年的两个预期： XAI/USPCX的10GW电力 5000亿美元的NOCL资金 BIC 2027年会给这些项目带来顺风。 10GW是什么概念？ 一个典型核电机组约1GW。10GW就是10座核电站的电力，全砸进AI数据中心。 这顿饭的核心结论：内存还是瓶颈，但瓶颈在变形状——层数在降、出货量在升、总位元需求不降。 短期赢家： AEC铜缆（CRDO）、NPO/InP（LITE） 短期承压： WFE设备商（AMAT/MKSI/LRCX），目标价已被下调 需要观望： CPO（还要2-3年）、800V（可能延迟） 大方向没变： AI算力需求还在往上走。只是技术路线在微调，供应链的赢家和输家在重新洗牌。 美股投资网认为，这顿饭最有价值的信息不是某个目标价上调或下调，而是HBM层数可能见顶这个信号。 如果市场一直按"HBM越堆越高"的逻辑在给内存和设备商定价，那这个预期的松动，就是最大的风险点。 $CRDO $LITE $AMAT $MKSI $LRCX #美股#

BREAKING: TeraFab secures two major Texas agreements for Phases 2 and 3, representing nearly $48 billion in investment. The previously announced $16.8 billion figure applied only to TeraFab’s initial Phase 1. These new agreements cover the Anderson-Shiro portions of the next two phases. • The agreements cover the Anderson-Shiro CISD portions of TeraFab Phases 2 and 3. Both became effective on August 18 and were posted by the Texas Comptroller on August 21. • The official applications project $23,838,288,600 of investment for each, or $47,676,577,200 combined. That includes approximately $18.78 billion in buildings and real property plus $28.90 billion in machinery and equipment. • Phase 2 construction is scheduled for 2029-2031, followed by Phase 3 from 2032-2034. Each plans approximately $7.53 billion, $8.09 billion and $8.22 billion of investment across its three construction years. • The filings describe TeraFab as a SpaceX-planned, vertically integrated semiconductor facility led by affiliated technology companies including Tesla, SpaceX and xAI. The broader four-phase plan is estimated at approximately $119 billion. • TeraFab aims to bring chip design, photomask production, wafer fabrication, advanced packaging and system integration together on one Texas campus, including sub-2-nanometer-class technology, supporting power generation and a space-compute test facility. • Each of these two portions is projected to create 617 direct permanent jobs, bringing the combined projection to 1,234. Each is also estimated to support 2,185 total jobs statewide when indirect and induced employment is included. • Each three-year build is estimated to support an annual average of 11,316 direct construction jobs, or 14,562 jobs when indirect and induced work in Grimes County is included. • Projected average salaries are $168,924 when Phase 2 begins operations in 2032 and $179,264 when Phase 3 begins in 2035. Estimated direct annual payroll is $104.23 million and $110.61 million, respectively. • The two economic-impact studies collectively project approximately $5.36 billion in state and local tax revenue across their respective 38-year analysis periods. This includes about $2.49 billion for Anderson-Shiro CISD and $1.45 billion in property taxes for Grimes County itself. • For newly eligible property, the taxable value used to calculate Anderson-Shiro CISD maintenance and operations taxes will be zero through construction completion and 50% of market value during each 10-year incentive period: Phase 2: 2032-2041 Phase 3: 2035-2044 • This is not a blanket exemption from all taxes or a direct cash payment. County taxes, school debt-service taxes, land, inventory and existing property remain outside the limitation. • The Texas Comptroller concluded that each phase should generate enough state and local tax revenue over 20 years to offset the school M&O revenue lost under the agreement. The applications also state that the incentive was decisive in choosing Texas after locations in Arizona, New Mexico and Nevada were considered. • Each contract separately requires at least $20 million of eligible investment and 10 qualifying jobs. These are legal compliance minimums, not the much larger project forecasts. • The required jobs cannot be counted under both agreements. They must be new, primarily on-site, full-time positions of at least 1,600 hours annually, with average wages of at least $66,668. TeraFab must also offer and contribute to employee health coverage. • Each agreement is backed by a $33,583,540 performance bond or guarantee, totaling $67,167,080. Public compliance reports, audits and site inspections are required, while failures can trigger penalties, termination and recovery of lost tax revenue plus interest.

Grok @bot is a smash hit It will have 100m users this time next year It’s the ultimate agent onboarding tool for non-technical people @bot will have more users than @x in three years if the X team keeps improving it at this pace

🥷 New stealth model: Ox Alpha Ox Alpha is a frontier model built for efficient coding, sustained agentic work, and real-world production use. - 1M token context window - Text, image, and video input Try it now and share feedback to improve the model!

A 300 AGENT GROK SWARM JUST OBLITERATED THE WORK OF A HUNDRED HEDGE FUND PHDS IN 40 MINUTES amateurs are still asking AI for basic text summaries while pros force agent swarms to expose hidden market maker routes. reading three hundred isolated documents is a massive waste of time when you can visualize the exact single points of failure in your trades. the exact wallet tracking architecture: > deploy a Grok swarm to cross-reference multi-token balances from your external files and map thousands of asset connections > force the agents to query The Graph API and log every single wallet as a linked node so hidden liquidity routes instantly light up > persist the subgraph volume data across sessions so your local database gets continuously smarter instead of resetting when you close the tab the complete automation script 👇