Sources: 200+ tweets extracted Time window: 12 hours
---
Bull market officially up 100%, with S&P 500 hitting 9 new all-time highs in 2026 (matching 2007 and 1966 peaks). Market concentrated in megacap growth—Tech +2.5%, Cons Discr +1.3%, Comm Services supported indices while equal-weight SPX was -0.2%. SOXX semiconductor ETF up 18 consecutive days, hitting record RSI levels, up ~40% this month (strongest since Feb 2000). Retail investor sentiment turned bullish (AAII survey: 46% bullish, highest since mid-January, bull-bear spread +11.6ppt, ending 9-week bearish streak). AMD +13.9% Friday; $MU broke $500 psychological barrier; Qualcomm +11.1%. Despite this strength, Michael Burry disclosed short SOXX positions via puts. The "Sell in May" narrative challenged—May-October positive in nearly every year over past decade (avg +7%). Cathie Wood/ARK bought ~$73M Amazon and 280,450 shares Wednesday. Banks (TD, Societe Generale, Morgan Stanley) using Significant Risk Transfer deals to offload data center debt—$1B hedged by TD, hinting at institutional caution on AI infrastructure thesis.
---
Tech mega-caps driving market while breadth narrow. AI infrastructure ecosystem (Google TPU supply chain, CoWoS, HBM, optical interconnect) seeing multi-level opportunity beyond just $AVGO/$TSM. Chinese "teapot" refiners absorbing Iranian crude (record ~1.9M bbl/day), keeping Chinese crude inventories stable despite Middle East disruption. P&G warned of $1B hit from Iran war after failing to hedge oil exposure. US crude rig count fell to 407 (matching pre-war levels). Basel 3 pushing banks from paper to physical silver ($SLV). RealPage algorithmic rent coordination still running despite DOJ settlement (no fines). PBM reform passed—CVS/UNH/CI facing margin compression; Mark Cuban's Cost Plus Drugs as existential threat. Nokia cited as "market forgot to reprice" AI infrastructure stock (Citron target $24.50 based on 49x 2028 EPS vs Infinera acquisition giving chip capability).
---
### 📊 Google TPU Ecosystem — Major Supply Chain Restructuring
Google TPU shift recasting AI compute supply chain. TPU 8 first split into training 8t and inference 8i—Google重构 AI chip roadmap by use case. Virgo network connects 134,000+ TPU chips at 47 Pb/s bandwidth with ~40% latency reduction. Optical interconnect value significantly amplified. Google already has 70%+ code internally written by AI; Google Cloud Q4 2025 surged 48% YoY, doubling past Amazon Web Services. New 8th-gen TPU细分训练/推理双版本 directly challenging Nvidia compute dominance.
Key beneficiaries mapped:
Max constraint: TSMC CoWoS + HBM. 2026 CoWoS demand 450-500万 units, supply ~365万,缺口20-27%. Google TPU target下调至约3M (from 4M). Five nodes to watch: 1. Google TPU出货是否上修 2. $AVGO AI revenue超预期 trajectory 3. $TSM CoWoS产能是否缓解 4. HBM价格与客户认证进度 5. 800G/1.6T光模块订单兑现
Google deepening Anthropic bet with up to $40B investment. Google owns ~14% currently; new investment likely brings stake to 20-25%. May own more from Anthropic investment than Gemini subscriptions generate. Note: Amazon simultaneously invested up to $33B in Anthropic ($5B immediate, up to $20B follow-on, $8B existing, totaling $33B potential; Anthropic committed $100B+ spend on AWS over 10 years including Trainium/Graviton). Anthropic securing 5GW compute (equivalent to ~5 nuclear plants; MSFT total global DC footprint ~5-6GW).
---
### 📊 TSMC Advanced Packaging — CoWoS Capacity Crisis
TSMC's CoWoS advanced packaging tighter than leading-edge wafer fabrication. NVIDIA booked >50% of CoWoS lines for years ahead. By late 2026, TSMC plans quadruple CoWoS capacity to 130,000 wafers/month. Chiayi AP7 becoming world's largest packaging hub. P2 2026 production target, P1 2027. Fabs operating non-stop including holidays. A16 and N2P (H2 2026 production) promise lower power/higher density for AI accelerators.
GPTC (Grand Process) lawsuit: TSMC advanced packaging equipment supplier denied technology leak to China, filed lawsuit against former GM for suspected trade secret infringement, "fully cooperating with authorities."
NanoFlex Pro at A14 node: TSM's North America Technology Symposium revealed 1.5x tall cell design eliminates gaps between cell types, increasing layout density and OD efficiency. Coming to N2U process in 2028 for 2-3% logic density gain.
FUSO Chemical ($4368.T): Only silica abrasive producer approved for TSMC N2. Electronic materials segment ~40% OP margins even before price hikes. CMP slurry identified as choke point.
---
### 📊 Intel — 18A Progress + Packaging Play
Intel 18A and 14A process improvements speculated to come from ex-TSMC executive. Wei-Jen Lo joined Intel Oct 2024; faces 2 cases in Taiwan (TSMC civil lawsuit + government criminal case alleging National Security Act violation over N2, A16, A14 technology data). Taiwan court authorized seizure of NT$2 billion (US$60M) of Lo's Taiwan assets. Intel fully backed Lo. $INTC
Intel EMIB packaging orders advancing with two CSPs (likely Google + Amazon). Prepaid tens of billions for capacity; deliveries expected H2 2026. Intel the only supplier with comprehensive advanced packaging portfolio competing with TSMC. Intel CFO noted strong customer confidence in EMIB.
Deeply personal Intel bull thesis from engineer: "Lip Bu Tan and Intel team delivered on innovation... this is just the beginning." Referenced 18A, backside power delivery, advanced packaging breakthroughs as "new Moore's Law."
---
### 📊 Memory Supercycle — Samsung Strike Risk
Samsung Electronics union strike could disrupt memory mass production for at least 1 month. Union voted 18-day general strike starting May 21. Demanding 15% of annual operating profit as performance bonuses + transparent bonus system. ~40,000 gathered at Pyeongtaek resolution rally; >50% of 76,100 union members attended. Union Chairman warned: "If company treats devoted union members as mere numbers, we too will respond with numbers—by demonstrating massive production disruptions."
Production impact estimates: Setup/maintenance suspension on semiconductor equipment = >1 month to restore normal operation. Output could drop 10-20% immediately if personnel pulled from fab. Impact stretches to 2x+ period for mass production recovery. Estimated losses up to 30万亿 won. Industry concern: Samsung (world's largest production capacity) disruption = memory supply crunch worsens. Competitors (SK Hynix, Micron) enjoy windfall from steeper memory price rises. Volatility especially pronounced in server DRAM and eSSD segments.
SemiAnalysis CEO Patel on memory:
---
### 📊 Nvidia — DeepSeek V4 Paper Reveals Architecture Alignment
DeepSeek V4 paper analysis shows Nvidia moat = anticipating mainstream LLM trajectory 3-5 years out. 6144 FLOPs/Byte passage explains why Nvidia pushed HBM4 pin speeds aggressively—Rubin HBM4 speed isn't "overkill" for V4, it's precisely balanced design. Nvidia working to crank up Rubin Ultra bandwidth = HBM bandwidth could become bottleneck for MoE models like V4. NVL domain scaling + Kyre move = pushing bandwidth-density of interconnect fabric, rebalancing compute-communication.
Nvidia G3.5 creating new memory tier: New tier between local NVMe SSDs and object/file storage, exclusively for AI inference. DeepSeek V4 paper (§3.6.2) argues KV cache can break GPU HBM limits and permanently offload to NVMe. Maps exactly onto ICMS rack shown at CES—Nvidia called it precisely.
InferenceX Day 0 DeepSeek v4 support on Blackwell B300, 5x faster than Hopper on Day 0. Great work by sgl_project, radixark, NVIDIA AIDev, CoreWeave emergency Blackwell nodes.
---
### 📊 Qualcomm — Data Center CPU Coming June
Qualcomm stock +11.1% Friday on data center CPU announcement. Datacenter Nuvia CPU launching June; could reverse overly short-biased sentiment. Watchlist commentary: "If Qualcomm really launches data-center Nuvia CPU in June, it could potentially reverse current sentiment."
---
### 📊 Misc Semiconductor Signals
Wiwynn Texas investment: Taiwan AI server giant (non-Nvidia clients: Amazon, Google, others/ASICs) announced US$123M (NT$3.87B) investment in Socorro, Texas—880,000 sq ft factory space for AI server builds.
Advantest watchlist: As GPU architectures become more complex, testing steps could rise exponentially. Question: more upside ahead?
Delton Technology watchlist: Specialized CPU substrate manufacturer with meaningful CPU substrate market presence.
Montage watchlist: If CPU adoption increases, memory interface usage should surge. Prefers Montage over Rambus due to CXL exposure.
Intel told Chinese CSPs CPU supply extremely tight next two quarters, should stabilize toward year-end. However, Chinese CSPs view timeline as overly optimistic, preparing for prolonged CPU shortage.
Samsung 300mm silicon photonics platform unveiled at OFC 2026. CMOS-based, incorporates 305nm silicon layer, 400nm silicon nitride layer, PN-doped modulators, germanium waveguide photodiodes, thermo-optic phase shifters, TSVs, bump packaging. Addressing surging data center interconnect demand from AI/HPC.
---
### 🔮 Modeling Implications — Semiconductors
Direct impacts (revenue/cost/margins):
Indirect/strategic:
---
### 🤖 AI Models & Releases
GPT-5.5 launch is major capability jump ("近两年最大能力跃迁"):
ChatGPT Image 2: "Absurdly good" — generating fake AAA game screenshots, Silicon Valley themed. Lip sync and sound significantly improved.
Claude vs OpenAI = "Pepsi and Coca-Cola of this era": One Chinese poster: "Claude and OpenAI are this era's Pepsi and Coca-Cola."
DeepSeek V4:
Anthropic "Project Deal": Claude agents autonomously interviewed 69 employees, learned preferences, then bought/sold/haggled on their behalf. Executed 186 physical deals totaling $4,000+. One Claude deduced user preferences so perfectly it bought the exact snowboard employee already owned.
Gemini new version coming "very very soon": Thomas Kurian CEO Google Cloud.
Microsoft MAI models: MAI-Transcribe-1 (most accurate transcription globally on FLEURS WER benchmark across 25 languages), MAI-Voice-1 (new standard for natural speech), MAI-Image-2 (top 3 model family on Arena). All available on Microsoft Foundry.
Kling AI 4K: Native 4K now available.
---
### 🛠️ AI Coding Tools & Agents
Cursor 3.2: New /multitask command for asynchronous subagents running parallel coding requests. User received $10,000 credits from Cursor. Cursor spending heavily on marketing (狂给营销号打钱). Post-Cursor acquisition, Elon/X sending money around. User: "cursor被收购以后四处送钱，可惜我连客户端都删除了." One user: "后发优势都做不到，抄也抄不会—用 codex 一周搓出来的功能都比 xchat 丰富."
Codex: "ship." Users reporting excellent results combining GPT-5.5 with Codex. Codex as orchestration layer; Claude Code for specialized subagents. One workflow: research→scoping→implementation→review with manager agent handling up to 16+ subagents. "Codex is a smooth operator."
Claude Code/Hermes/OpenClaw: Running 5 AI agents daily: Claude Code, IDE+CC, Codex, OpenClaw, Hermes. User notes: "creating and using agents are a form of alchemy—one wrong word/prompt/step and output goes from God tier to dog shit."
acpx 0.6.0: Control Codex/Claude via agents. Features: Claude system-prompt controls, session pruning, embeddable turn handles, --no-terminal, persistent-session fixes, WSL cwd translation, queue hardening, clearer error hints.
OpenClaw: Chinese AI researcher 罗福莉 surprisingly推崇, despite not being a "product manager/CEO type." OpenClaw is open source, Chinese AI companies now replicating capabilities.
Claude Agent SDK observation: "If Anthropic called Agent SDK 'Claude Code SDK' (they did originally) it would be 10-100x more popular."
clawsweeper: Runs 50 Codex instances in parallel around the clock, scans issues/PRs deep, closes what's already implemented. Closed ~4,000 issues in one day, thousands in pipeline.
---
### 📱 Consumer Products & Platforms
XChat released: Standalone mobile app for platform's direct messaging ecosystem. Rolling out for iOS. Chinese reaction: "XChat 团队一定是用 Grok 写代码了，要不不会搞这么久." User: "现阶段可以说用处不大." Another: "稳如小红书把私信功能重发了个 APP."
NotebookLM: Won 3 Webby awards including People's Voice Award for best AI productivity/work tool. New automated organizational feature for heavy research projects—auto-categorizes/labels source documents when >5 files uploaded.
Grok Imagine new model: Much better lip sync and sound. Nothing in videos is real.
Google DESIGN.md workflow analysis (15 key insights): 1. Google wants shared design language for AI; DESIGN.md gives designers/developers/agents one standard 2. Standard makes design rules portable between Stitch, coding agents, repos, builders 3. Reusable project memory—if keep telling tools same colors/typography/spacing/preferences 4. Markdown is right middle layer: structured enough for agents, readable for humans 5. Start from real reference systems (Stitch, Neuform, Variant, getdesign.md) 6. Pair spec with visual context (DESIGN.md + screenshots + HTML) 7. Use Remix and Iterate intentionally—Iterate keeps design close, Remix opens exploration 8. Curation is part of design—favoriting strong generations, hiding noisy ones 9. Small prompts are design operations—language-expressed interface edits 10. Agents need guardrails, not just inspiration 11. Good hero should expand into system (pricing, testimonials, motion, mobile, branding, social) 12. Exploration tools vs build tools have different jobs 13. Workflow order matters: DESIGN.md → generate → remix → section variations → builder → full site 14. Standard raises floor for quality/accessibility—lintable for missing essentials, broken refs, contrast 15. DESIGN.md is foundation, not finish line
UPTHOS Relay中转: Stable for half month, now supporting GPT-5.5 and GPT Image 2.
Alma Activity Recorder: User recommending everyone try it.
Real-time AI hardware: Chinese company launching omnidirectional microphone + custom buttons for voice-controlling Claude/Codex ("你的 vibe coding 好搭子").
---
### 🏢 Corporate & Strategic Moves
Adyen first-ever acquisition: EUR 750M cash acquisition of Talon (Berlin-based customer engagement API platform, loyalty management, personalized promotions, incentive optimization). 300+ global merchants including Nordstrom, H&M. Talon expects EUR 60M ARR end-2026 (30-40% growth). Adyen expects 100-200bps incremental revenue growth 2027, 100bps modest margin dilution 2027, >55% EBITDA margin maintained 2028.
Cursor potential acquisition by Elon: One analysis: "Cursor is Elon's first purchase, which is a huge mistake. Coding agent harness now open source (Codex, Claude Code). Current design works virtually perfectly. Users switch between IDEs frictionlessly—no lock-in like Twitter. Cursor is emperor with no clothes." Counterpoint: Cursor sent $10,000 credits to influencer; other Chinese creators also receiving.
DeepSeek first fundraise: For 2 years turned away Tencent and Alibaba. Three structural pressures all intensified simultaneously—what changed?
Denso withdraws ROHM TOB: Japanese auto parts giant DENSO withdrew full acquisition proposal for chipmaker ROHM after failing to get ROHM agreement. DENSO holds <5% ROHM stake. ROHM now pursuing power semiconductor integration with Toshiba and Mitsubishi Electric—"罗姆・东芝・三菱" alliance as main axis.
Lachy Groom backing India startup Pronto at $200M valuation.
Oklo, NVIDIA, Los Alamos working on plutonium-powered AI data centers.
Ford, Geely held talks about bringing Chinese tech to America—Chinese cars in US "inevitable."
---
### 💬 AGI Belief as Religion
"Believing that AGI will be achieved is like believing in God." Arguing about this is useless. Accept that many believe in invisible guy who cares about them and listens to every nonsense, rewarding when feels like it or punishing disproportionately, sometimes children, without clear reason. Hard to understand but not everything is understandable.
---
### 💬 AI Model Interchangeability
"LLMs are absolutely interchangeable... if you don't care about output quality." One user: "After 1 day of use, GPT 5.5 XHIGH is really good." SemiAnalysis note: "Most of our engineers switch between Codex and Claude models depending on task and IDE preference."
---
### 💬 AI Coding Agent Moats
"Once models get good enough, no moat?" Debate: Does massive TAM mean multiple winners (OpenAI, Anthropic + others) or does all surplus go to consumers without accruing to anyone? Counter: "There will always be advantage for those who iterate and arb early access to best models."
"If Anthropic had called Agent SDK 'Claude Code SDK' it would be 10-100x more popular."
---
### 💬 Nvidia Valuation Cycle
"Semiconductor stocks approaching peak multiples on peak earnings soon?" Do not forget many semis are notoriously cyclical. SemiAnalysis CEO Patel: "Economic value best models can create growing faster than our ability to supply Tokens. Gap will keep widening. Model vendor margins will keep expanding until hardware supply chain folks start saying 'why aren't we pulling our margins up too.'"
---
### 💬 China AI Progress
DeepSeek V4 narrative split:
"Chinese labs must be 10x more data efficient if they can't close compute gap."
Chinese domestic chip (NEU): 上海中紫星 NEU chip targeting Q4 2026 tape-out.实测显示 AI推理任务速度达到传统GPU方案近100倍，能耗仅十分之一。团队成员来自Intel/Nvidia/海思/加拿大顶级AI lab.
---
### 💬 Iran Financial Warfare Strategy
Iran's strategy = attacking US monetary policy-vield curve. Oil shock → inflation expectations → force Fed to hike → pop market bubble → force Trump to give Iran favorable conditions. March 30 Powell speech excluded rate hikes before inflation expectations unanchor. Market pricing = "even if inflation, won't consider rate hikes" → long-end Treasury yields rise with inflation. Long-term path = sacrificing system vitality. Long Treasury holders (banks, financial institutions) holding large unrealized losses—similar logic to 2023 SVB. China using teapots to keep fuel output high, Iranian crude imports at record.
---
### 💬 Anthropic vs OpenAI Competition
"Sam Altman 1:0 Dario Amodei." GPT-5.5 viewed as significant capability jump. Anthropic securing massive compute deals ($33B Amazon, potentially $40B Google) while OpenAI shipping faster. One analysis: Anthropic's unusual moves around GPT-5.5 launch converge on securing compute—must bind more deeply to those with physical resources. Deal structure more favorable to Amazon than Anthropic: Amazon invested in both OpenAI (up to $50B) and Anthropic, collecting toll-gate revenue from both.
---
### 💬 Tesla Robotaxi Scale
"If Tesla continues increasing unsupervised Robotaxi fleet by 2.4x every month, they'd have 21,000 driverless Robotaxis by year-end. Waymo currently has 3,000." AI4 became unsupervised January 2026 (3 months ago). Fleet 2.4x larger than last month. FSD doing "superhuman" moves—slowing for pedestrians before visible to driver. HW3 cars outside US getting FSD; V14 lite needed. Robotaxi margins ~20% (assuming perfect competition), not 80%. Long-term vision: hundreds of millions of autonomous vehicles for North America alone.
---
Profit efficiency (profit/employee):
X earnings data:
Anthropic metrics:
Market data:
Supply chain data:
AI model benchmarks:
---
### Direct Impacts (Revenue/Cost/Multiples)
AI Infrastructure (beneficiaries):
AI Model Companies:
Semiconductor Equipment/Materials:
Memory:
Optical/Silicon Photonics:
Timing considerations:
### Indirect/Strategic Implications
Nvidia moat validation:
Intel turnaround:
Qualcomm:
Amazon:
Chinese AI:
Tesla:
Anthropic vs OpenAI:
Nokia:
---
### Operational & Organizational
Samsung Electronics union strike timeline:
Intel vs TSMC executive case:
RealPage still operating:
Tesla FSD progress:
### Job Market & Labor
Amazon India SDE3 work: Friend doing "coding equivalent of grunt work" vs US counterparts doing core projects. "More or less true for most FAANGs in India."
YC hosting Demis Hassabis (DeepMind founder).
Fundstrat Capital hiring.
Nous Research AMA: Wednesday 8-11AM PST on r/LocalLLaMA.
Hacker competition for AI Builder founders: Chinese founder recruiting without regard for school pedigree or 3-5 years experience; just performance in hackathon.
### Frameworks & Mental Models
Portfolio construction lesson: Being long 100%+ SOXX doesn't make someone a PM. Can just buy SOXX for nearly free. 300% leverage requires betting smaller on each theme (AI 30%, other volatile themes 15% each, lower-vol Japanese 60%). Shorts defacto more risk than longs—need double number of ideas. "Just being long SOXX does not someone a PM."
Notion CEO Ivan Zhao thesis: AI agents like steam engine. Real change isn't the tool—it's redesigning entire workflows. Once Agent connects all context, humans shift from搬运工 to决策者. Notion has 700+ agents internally.
Andy Grove "Only The Paranoid Survive": Described as "shockingly prescient guide for the AI age."
Biological intelligence origin: Result of death. Trillion-year training run where loss function is extinction. Failed to understand reality = deleted.
### International & Non-English
Chinese perspective on AI industry pace: "前年感觉是一年变化一次，去年感觉一个季度变化一次，今年感觉一个月一个样"—industry changes accelerating to monthly.
Chinese reaction to XChat: "稳如小红书把私信功能重发了个 APP."
Chinese observation on Cursor: "cursor 被收购以后四处送钱，可惜我连客户端都删除了" + "后发优势都做不到，抄也抄不会，用 codex 一周搓出来的功能都比 xchat 丰富."
Chinese view on Anthropic: "现在这个时代，就是模型合智能体双轮驱动，Anthropic就是绝对领先的."
罗福莉推崇 OpenClaw: Chinese researcher surprisingly endorsing despite not being "product manager/CEO type." OpenClaw open source = Chinese AI companies replicating.
DeepSeek V4 vs R1 market reaction: "2025年DeepSeek R1发布，美股纳指被砸出一个坑来。2026年DeepSeek V4发布，纳指历史新高了."
Chinese view on Alibaba: "错过这一波的第一波。我们不会错过第二波."
Japanese perspective on meaningful life (translated).
Korean workplace culture: 주급이 늘어난 느낌...?
### Memes, Jokes & Cultural Moments
Inside joke: "Guy under 6 foot who's really into fitness."
Average American: [meme context].
Nerd Snipe podcast: 75th biggest tech podcast worldwide (was 170th Wednesday).
Homestar Runner reference: Pod!
Human use of computers: "one day not so long from now human use of computers will be over and we can all go to the park."
AGI belief vs religion comparison: Elaborate theological analogy.
Real job vs AMD 2014: "Imagine working a real job instead of just having bought $AMD in 2014."
Jane Street mental math interview: "What is the next day after today in DD/MM/YYYY where all digits are unique?" Answered 0.95 confidence, corrected to 0.97, passed. "0.95 huh?" interviewer chuckled.
WTF kids/lunatics: Multiple variations.
Fed President Retard: [meme context].
Buying art vs watches: Best at auction (once you know) vs dealers for watches (can't judge what's inside).
### Questions, Concerns & Risks
"Is it too late to confess I barely used Bloomberg terminal?"
X earnings disparity: "为啥我的X工资只有几百，但是别人都是几千?" + Blue V ratio ~5% normal; high ratio indicates fake.
DeepSeek R2 question: "Will there be a 'Deepseek r2'?"
US stocks underperform 5-10 years: "Who has this scenario on their bingo card?"
Real job vs passive investing: "Imagine working a real job instead of just having bought AMD in 2014."
"Benchmarks are saturated and gamed. Real life quality and progress is what matters."
"Is it Lindy?" [re: tech durability question].
"Every time I see someone say they one-shot this with an XYZ model, my honest reaction is: did you?"
Intel 18A/14A intellectual property: Ex-TSMC executive case ongoing—Taiwan civil + criminal.
Samsung strike ripple effect: Memory supply crunch could widen; production recovery >2x strike duration.
Data center financing risk: Multiple major banks simultaneously transferring risk out via SRTs—TD hedging $1B with forward-flow; Societe Generale + Morgan Stanley same. Satellite imagery showed 40% data center projects missing completion dates; electrical component shortages documented; community opposition hardening.
"Do not forget many semis are notoriously cyclical businesses."
"Has Amazon's moment finally arrived?" (contrarian question re: market re-rating).
### Personal Reflections & Lifestyle
Weekend reflection: "I've been following Intel since childhood... they are the reason I fell in love with chips and computers. Today I was vindicated for everything I've been talking about for over a year."
Chinese creator on X revenue: All donated to cultural public welfare projects; believes wealth beyond certain point is about impact not accumulation. Subscription $1 vs market alternatives; shares all ideas for free. Goal: give retail same edge as Citadel. Planning to release custom LLMs for free subsidized by subscription revenue.
Personal growth reflection: "在纽约，第一次觉得野心不是需要藏起来的事。在伦敦，学会把问题慢慢想清楚。那几年自由探索了很多东西，它们不是我想回去的地方，却是我现在生命的底色."
Date night: "女儿学校组织露营...老两口去打了羽毛球."
Pickleball: "Played pickleball for over an hour today. With tennis background, picking it up pretty quickly. Skepticically avoided the craze but it's actually really fun."
Tesla self-driving: "I'm cool calm and collected sitting in this traffic with Tesla Self-Driving."
"Constraints are the mother of innovation."
" Constraints are the mother of innovation. It was such a good day!"
—————— 📅 Published: 2026-04-26 03:11 HKT ⏰ Next update: 2026-04-26 07:00 HKT
Full analysis, implications, trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.
No macro headline is dominating the tape this morning; it’s a stock-specific session led by semis. **INTC** is +25% to ~$83 after a 13-broker upgrade cycle, **MXL** +30% AH on data center optical inflection, **TXN** +23% this week confirming analog has turned, while **SAP** is bouncing maybe +5% off
Factual bullet layer:
What the market is reacting to: It’s not the beat. It’s the structural narrative that agentic AI is compressing the CPU:GPU ratio from 1:8 in pre-training to as tight as 1:2 in agentic workloads (per Intel’s own CEO and Northland). Intel is effectively the only scaled x86 server CPU source, and supply is constrained into 2027. The stock is ripping on forced re-rating after being left for dead.
The Bull Case: The bull stack is deep and numbers-heavy. HSBC (Buy, $100) models 2026/2027 server CPU shipments +20%/+21% YoY, with DCAI revenue of $23.2B/$29.4B — 17% and 31% above consensus ($19.9B/$22.4B). HSBC’s GM estimates of 41.3%/45.0% are also well above street (37.2%/41.3%), arguing the market underestimates ASP upside and foundry reallocation flexibility. KeyBanc ($110) highlights the 500bps GM expansion to 41% driven by price increases and 18A yield recovery, with advanced packaging backlog building into a multi-year 2027 ramp. Benchmark ($105, Buy) flags output improving faster across Intel 7/10, Intel 3/4, and 18A, with advanced packaging now a billions-dollar opportunity. Evercore ISI upgraded to Outperform with a $111 target, calling it a "CPU renaissance" driven by inference and agentic AI. Citi ($95, Buy) also upgraded. Roth/MKM ($100, Buy) credits CEO Lip-Bu Tan with improved execution and sees multi-year agreement interest from large customers. Northland ($92) notes the CPU:GPU ratio shift and a potential 14A foundry customer decision in 2H26 as a catalyst.
The Bear Case: The valuation pushback is violent. Rosenblatt maintains Sell with a $50 target, arguing even after raising FY27 EPS to $1.25 and using a 40x multiple, the stock at $83 is "far ahead of the fundamentals". RBC ($80, Sector Perform) calls the setup balanced at >50x CY27 EPS and warns PC faces memory shortages and pull-forward risks. Needham (Hold) says the turnaround is underway but "shares fully reflect the company's prospects". Cantor Fitzgerald ($90, Neutral) argues traditional valuation metrics are challenged and Intel’s value depends on being a fast or slow follower to TSMC; the bull case requires TSMC to slow capacity additions. TD Cowen ($75, Hold) and Stifel ($75, Hold) both cite valuation as keeping them on the sidelines, with Q2 GM headwinds from Panther Lake mix.
What resolves the debate: 14A foundry customer commitments in 2H26 and sustained server CPU ASPs through AMD’s next-gen ramp. If Intel lands a marquee foundry customer and Q3 guide shows GM re-expansion, the $110 targets are in play. If 18A yields stall and PC weakens in 2H, the $50 bear case closes fast.
Factual bullet layer:
What the market is reacting to: A relief rally that SAP didn’t blow up guidance. The stock entered the print -41% over six months and near 52-week lows ($160.66), pricing in a massive reset. The cCBO held 25%. But the 5-broker cluster is cutting PTs anyway because the Middle East conflict introduces growth uncertainty that is impossible to model, and the organic guide down implies the headline maintenance was cosmetic.
The Bull Case: Morgan Stanley (Overweight, EUR190) says "no major reset of SAP's outlook is the key takeaway and a positive for investors" (Article 17). cCBO better than feared, cloud revenue strong, and margin expansion on track. With a PEG of 0.19 and Piotroski Score of 9, the selloff is overdone. Evercore ISI (In Line, EUR200) notes cCBO at 25% cc exceeded the ~24% bogey and FY26 outlook reiterated. KeyBanc (Overweight, EUR235, cut from 280) called the quarter "better than feared despite imperfect performance" with strength in cloud contract billings and Cloud ERP. Citizens/JMP (Market Perform) highlighted the clean EPS and op profit beats. The mid-May analyst meeting on AI monetization is a near-term catalyst.
The Bear Case: BMO Capital (Outperform, $200, cut from $210) says sustained Middle East conflict "increases growth uncertainty for SAP" and could impact SAP "more than other companies in BMO Capital's coverage universe" (Article 29). They note 2026 revenue guide was guided modestly lower when accounting for M&A, implying organic deceleration. KeyBanc's own PT cut to EUR235 reflects mixed guidance and supply chain booking fears. The market is asking: if cloud backlog is so strong, why reduce organic revenue growth guidance? The answer is likely M&A mix and geopolitical prudence, but it gives bears fodder. Macquarie’s cut on NOW (to $109) frames the broader software worry: macro and geopolitical uncertainty and weak US federal procurement are sector-wide drags.
What resolves the debate: The mid-May analyst meeting and specific quantification of Middle East revenue exposure. If SAP can isolate the geopolitical risk and show AI monetization accelerating, the stock re-rates toward EUR200+. If Middle East sovereign deals freeze, the organic guide risks another cut.
TXN reported what BofA (Buy, $320) called a broader and higher-quality recovery than last year’s false start. Industrial revenue grew +30% YoY and +20% QoQ, with every industrial subsector growing both sequentially and annually. Industrial remains 15% below the 2022 peak, suggesting a multi-quarter runway. Management sees no evidence of pull-in and wants to let Q2 play out before calling the upturn durable — a prudent stance given TXN’s direct model and short lead times mean the company sees demand in real time. 2026 capex is guided at $11.4B, declining to $7.5-8B by 2028.
STM delivered its largest earnings beat in nearly three years per UBS (Buy, EUR49). Q1 revenue was $3.1B vs. $3.04B consensus, with Q2 guided to $3.45B vs. $3.2B street. The guide implies a 12% sequential increase, led by automotive up low-double-digits and industrial up mid-20%. GM guided to 35.2% (+110bps) as utilization ramps from 70% to 80%. The AI data center business is expected to grow >50% YoY to >$500M, and the low Earth orbit satellite franchise is projected to drive ~$3B cumulative revenue from 2026-2028 (per Mizuho, Outperform, $56).
NXPI is the laggard. Stifel (Hold, $250) expects a slight Q1 beat and the ability to raise Q2 estimates, but with 56% automotive exposure, the firm remains cautious. The TXN print validates analog recovery, but NXPI needs auto production to cooperate.
The mechanism here is inventory digestion completeness. TXN’s direct model + short lead times = no phantom demand. STM’s utilization ramp from 70% to 80% with above-seasonal revenue guide proves foundries are refilling, not restocking. The debate is whether this is durable or a Q1 snapback. Bulls note TXN industrial at 15% below peak with broad-based growth = runway into 2027. Bears note STM missed EPS ($0.13 vs. $0.17) despite revenue beat, implying pricing and mix remain soft. The forward read is simple: if TXN guides Q2 above seasonal and STM pushes utilization to 85% in 2H, the analog cycle is real and read-throughs to ON Semi and Microchip are bullish.
DLR started 2026 with a record quarter. The company signed its largest lease ever — a 200MW deployment in Charlotte for AI inference — and raised 2026 guidance for revenue, adjusted EBITDA, and CFFO/share. The development pipeline expanded from $10B to ~$16.5B, with >60% pre-leased and 1.2GW under construction. Backlog hit $1.0B. Bernstein SocGen (Outperform, $232) notes the land bank increased by 1GW in Q1 alone. Stifel (Buy, $235) sees the updated guidance implying ~9-10% revenue growth and ~11% EBITDA growth with 60bps margin expansion. TD Cowen (Hold, $192) acknowledges the momentum but argues it is priced in after a 30% YTD rally to ~$206.
VRT delivered a Q1 beat and raised 2026 EPS guidance despite closing down ~2% on the day. Americas organic growth was +44% (vs. BNP Paribas Exane estimate of +35%), with margins expanding >500bps to 27%. EMEA fell ~29% on timing, but management raised the region’s full-year guide to flat YoY, implying a strong 2H ramp as industry capacity issues subside. Orders are expected up YoY.
The mechanism: The AI capex debate has shifted from "will hyperscalers spend?" to "can the physical layer keep up?" DLR’s 60% pre-leased pipeline de-risks the buildout — this is not a Field of Dreams exercise. VRT’s Americas margin expansion proves pricing power in power/thermal. The bull case is that 2026 is the year AI infrastructure moves from speculative pre-leasing to contracted revenue. The bear case is valuation: DLR at 27x NTM AFFO and VRT at 292% over the past year leave no room for execution wobbles. The forward read is DLR’s land bank supports 2028 growth, while VRT’s EMEA rebound in 2H is the margin bogey.
BESI is entering a capacity buildout in advanced packaging that has no peer. Needham (Buy, EUR300, from EUR210) notes hybrid bonding is being adopted by all three HBM suppliers with a uniform transition from TCB to hybrid bonding around 2028, driven by SRAM stacking and co-packaged optics. BESI remains the only viable supplier for die-to-wafer hybrid bonding. The firm forecasts 47% revenue growth in FY2026 and calls the hybrid bonding revolution "just getting started." The stock has surged 165% over the past year.
MXL reported and guided above expectations, with Roth/MKM and Needham both upgrading to Buy with $60 targets. Q1 EPS was $0.22 vs. $0.18 consensus; revenue $137.2M vs. $134.6M. Infrastructure revenue has surpassed Broadband as the largest contributor, and the company is positioned for AI data center optical interconnect growth as XPU clusters scale. Needham’s target is based on 25x CY28 EPS of $2.35. The stock ripped ~30% AH and is up 255% over the past year.
CLS (reporting Apr 27) is the integration read-through. BMO (Outperform, $450) raised estimates ahead of the print, citing sustained hyperscaler capex and CLS’s market position in switching and Google TPU products.
The mechanism: As XPU clusters scale, the constraints move from transistors to bandwidth and packaging. BESI controls the only die-to-wafer hybrid bonding tool needed for HBM4 and 3D logic stacking. MXL is riding the same datacenter optical upgrade cycle. The bull case is a multi-year order ramp with no competition. The bear case is both stocks are priced for perfection: BESI is flagged as overvalued on Fair Value; MXL is up 255%. The forward read is BESI’s 2H26 order flow and MXL’s Q2 guidance trajectory.
AMD reports May 5, but DA Davidson upgraded to Buy with a $375 target, arguing Intel’s print indicates meaningful upside to AMD estimates. The firm raised 2026 estimates by +$2B revenue and +$1.5B gross profit, now materially above consensus. The core thesis: agentic AI shifts the GPU:CPU ratio from 8:1 in pre-training to near parity in agentic workloads (per Intel’s CEO). With Meta guiding 6GW of datacenter capacity and OpenAI raising $122B, AMD’s 35% CAGR trajectory is underpinned. The new target is 32x CY27 EPS.
GOOGL is executing on a full-stack vertical integration strategy. At Cloud Next, the company launched the Gemini Enterprise Agent Platform, next-gen TPUs, and disclosed processing 16 billion tokens per minute in Q1 — exceeding OpenAI’s recently disclosed 15 billion and up 60% sequentially (per BMO, Outperform, $410). Evercore ISI (Outperform, $400) views Street cloud revenue estimates of $18B (+47% YoY) as highly reasonable, with backlog up 55% sequentially to $240B. TD Cowen (Buy, $375) expects agentic workflows to steepen adoption curves across the enterprise stack.
The mechanism: The market is repricing CPU exposure higher as agentic AI requires more CPU orchestration, while simultaneously rewarding GOOGL’s vertical stack (TPU + agentic platform + distribution) as a moat against NVIDIA’s ecosystem. The bull case for AMD is supply/demand imbalance allowing price increases across the portfolio. The bull case for GOOGL is that 16B tokens/min proves AI consumption is accelerating, not plateauing. The bear case: AMD has to execute on May 5 and Intel’s narrative steal is real; GOOGL faces antitrust overhang (Brazil CADE expanded investigation) and token growth does not equal revenue if inference prices compress. The forward read: AMD’s May 5 print is the next bogey for the agentic thesis; GOOGL’s Q1 cloud backlog commentary is the check on whether token velocity converts to dollar retention.
| Date | Ticker | Event | What to Watch | |------|--------|-------|---------------| | Apr 27 | CDNS | Q1 earnings | Hardware contribution, Hexagon sim integration, backlog | | Apr 27 | CLS | Q1 earnings | Hyperscaler capex tracking, Google TPU/switching mix | | Apr 28 | NXPI | Q1 earnings | Auto vs. industrial mix, analog pricing | | Apr 29 | AMZN | Q1 earnings | AWS growth, Anthropic ROI, operating income margin | | Apr 30 | AAPL | Q1 earnings | iPhone 17 units, memory cost inflation impact | | May 4 | PLTR | Q1 earnings | Government vs. commercial AI revenue split | | May 5 | AMD | Q1 earnings | Data center CPU supply constraints, agentic AI commentary |
The non-obvious read today is that the 2026 earnings derisking is happening in the hard parts of AI — silicon, power, and real estate — while the soft parts remain trapped in a geopolitical discount. INTC’s $1B+ of unfulfilled CPU demand, BESI’s hybrid bonding monopoly, and DLR’s 60% pre-leased pipeline are all sending the same signal: the physical AI buildout is accelerating into a supply-constrained 2027. Meanwhile, SAP’s organic guide down and NOW’s cRPO deceleration show that enterprise software is where macro uncertainty and AI deflation are surfacing first. The market is effectively running a barbell: long contracted infrastructure, short discretionary software exposed to sovereign procurement. If Middle East tensions persist, that spread widens. If they abate, the software unwind could be violent — but the floor isn’t in yet.
Full bull/bear breakdown, broker targets, and catalyst calendar — delivered by email at 07:00 HKT daily.