William , Engineer, CFA, FRM

SK Hynix just found a way to cool HBM memory from inside the package by cutting thermal resistance by 30% & paving the way for HBM5 stacking. With SK Hynix already supplying a major share of next-gen HBM for $NVDA Rubin, this shows why memory is becoming a core AI performance.

If you haven't read this, you should! $GLW Corning is one of my largest positions It doesn't have the highest torque But I believe it will be a steady compounder over time NFA

CXL is going to be the biggest change in AI serving since NVlink

Samsung Electronics Achieves World's First "900-Layer V-NAND"… Countdown to the 1,000-Layer Era Samsung Electronics has become the world's first to successfully implement a prototype of 900-layer-class V-NAND technology, taking another step toward the era of "1,000-layer NAND." Amid intensifying layer-count competition with rivals, the achievement is being seen as Samsung securing a decisive technological lead in a single leap. According to the semiconductor industry on the 25th, Samsung recently realized an integrated 900-layer-class V-NAND system using "Cell Multi Bonding (CMB)" technology, which joins two 450-layer cell wafers into one. NAND flash, which stores data, is a core component in AI servers, smartphones, and data center storage (SSDs). Much like stacking floors in an apartment building, the higher the layer count, the more capacity can be packed into a limited chip footprint, allowing more data to be stored while maximizing power efficiency. It is regarded as a key technology for winning dominance in the AI server and on-device AI markets, where high-capacity, high-efficiency components are essential. In the current mass-production market, SK hynix holds the highest layer count with its 321-layer 4D NAND. However, Samsung—while preparing for mass production of its 10th-generation V-NAND (V10, 400+ layers) this year—has at the same time vaulted to the 900-layer mark at the research stage, positioning itself advantageously in the next-generation NAND market. Regarding these research results, Samsung stated that "normal cell operation characteristics were verified," emphasizing that this goes beyond mere theoretical stacking to demonstrate a level of capability that actually functions. Since commercializing 3D V-NAND for the first time in the world in 2013, Samsung has continuously evolved its processes to overcome stacking limits. In the past, it used a "single-stack" method of drilling and stacking fine holes in a single pass, but as layer counts rose, it ran into physical limitations such as wafer warpage and alignment errors. In implementing the 900-layer device, Samsung resolved wafer warpage—the biggest obstacle—by adopting an advanced upper chuck design. Misalignment errors that occur during bonding were overcome with its proprietary "new overlay correction" technology. Newly introduced bitline (BL) and wordline (WL) structures also delivered meaningful gains, simultaneously reducing power consumption and chip size. Globally, China—led by Yangtze Memory (YMTC)—is chasing Korean firms at the threshold of 300-layer-class NAND mass production. With government support and the localization of equipment, it is pursuing capacity expansion and technological advancement at the same time. If YMTC succeeds in mass-producing 300-plus layers within the year, price competition could intensify and pressure Korean firms' profitability. For this reason, Samsung's 900-layer achievement is highly regarded as a strategic move to build a technological barrier over the medium-to-long term. An industry official said, "900-layer NAND technology is not simply three times 300 layers—it is a technology that changes the paradigm of the stacking process," adding, "It sends global customers the message that Samsung remains the technology leader, and it will have the effect of limiting Chinese firms' volume and price offensive."

🚨 A new ETF just filed that’s a pure bet on the photonics buildout: $LYTE 📈Roundhill - Same ETF shop that ran $DRAM to $1B AUM at fastest clip. Now- Photonics & Optics ETF. Holdings not disclosed yet. But the prospectus tells you exactly what’s getting in — if you read the fine print, let me break it down for you🧵 Here’s the part that matters : there’s one hard rule in the filing: ⛔️A company needs ≥50% of revenue from photonics/optics to qualify. That single number tells you who’s in and who’s out. 👇 ✅ CLEARS the 50% bar (the pure-plays): $LITE — optical transceivers, 1060nm VCSELs $COHR — lasers + optical components $AAOI — datacenter optics $POET — silicon photonics interposer $IPGP — industrial fiber lasers $CIEN — optical interconnect systems $CRDO — AI interconnect $SIVEF — InP lasers & Optic components These are the core of the fund. The sleeper angle nobody’s talking about: 🇨🇳 The prospectus explicitly builds in Chinese exposure — China A-Shares + VIE structures. It says outright: the fund “invests significantly in the securities of Asian issuers.” A US-listed photonics ETF with deliberate mainland exposure. Unusual. Why that matters?? Some of the most aggressive optical-component scaling is happening in Asia. $LYTE isn’t ignoring it — it’s writing A-shares into the strategy doc. That’s a geographic edge most US investors won’t have on day one Bottom line: $LYTE = the first pure-play photonics ETF ↳ 50% revenue screen = real pure-plays only ↳ Pure-plays in, conglomerates out ↳ Built-in Asia/China exposure 👇

CoreWeave: “From what we're seeing, I mean, Hopper prices are going up. Ampere prices are going up. Blackwell prices are going up across the board.” $CRWV $NVDA

We’re proud to announce a milestone moment for U.S. memory manufacturing. Micron has officially started manufacturing 1α DRAM at our Manassas, VA fab. As the only U.S. manufacturer of memory, Micron is strengthening America’s domestic memory supply with the world’s most advanced DDR4 technology — advancing our more than $200B U.S. manufacturing and R&D investment plans. bit.ly/43kpRJ4 #MemoryMadeInTheUSA #USManufacturing #GoMicron #IntelligenceAccelerated

$MU $DRAM $SNDK Holy Shit 625x More Memory Demand by 2028 vs 2022 That is Dell's CEO projecting memory demand growth His warning: supply takes years to expand. Demand is not slowing.

🔥 谷歌 I/O 2026 全梳理（看了完整发布会，文章由CLAUDE整理），我认为的重点已经BOLD，你们怎么看？ 5 月 19-20 日发布，进入"Agentic 时代"。 一句话总结：谷歌正面接管你的购物车、收件箱、搜索框，并把 Anthropic / OpenAI 的成本优势直接打掉。 22 个核心产品 👇 1/ Gemini 3.5 Flash 编码、agentic、多模态全面超越 3.1 Pro 速度比其他前沿模型快 4 倍 成本不到一半 API 定价战已经开打 ⚔️ 2/ Gemini Omni 任意输入 → 任意输出（先从视频开始） 理解重力、动能、流体动力学等物理规律 生成内容自带 SynthID 水印 角色一致性大幅提升 3/ AI Mode 月活破 10 亿 25 年来最大的搜索框升级 支持文本、图像、文件、视频、Chrome 标签页一起搜 AI Overviews 和 AI Mode 已合并为无缝体验 4/ Information Agents 🤖 24/7 后台跑的信息代理 跨博客、新闻、社交、金融、购物、体育实时监控 主动推送综合更新 今夏对 AI Pro / Ultra 开放 5/ Generative UI 搜索现在会为你的问题"即时设计"页面 表格、图表、模拟器、迷你 App 实时生成 婚礼策划、搬家追踪都可以做成持续使用的仪表板 6/ Universal Cart（重磅）🛒 跨 Search、Gemini、YouTube、Gmail 加入购物车 自动追踪价格、库存、兼容性 基于 Google Wallet 推荐最优支付方式 UCP 协议 + Google Pay 一键结账 Shopify、PayPal、Amazon 需要紧张了 7/ Gemini Spark 24/7 个人 AI 代理 手机/电脑关机也能后台工作 执行重大操作前会与你确认 下周向美国 AI Ultra 订户开放 Beta 未来支持短信、邮件、子代理、授权支付 8/ Daily Brief 夜间分析你的收件箱、日历、任务 早上给你一份个性化简报 + 行动建议 随时间学习你的偏好 9/ Antigravity 2.0 💻 独立桌面 App，多代理并行编排 一个代理写代码，另一个生成品牌资产 配套 CLI 和 SDK Gemini CLI 已被引导迁移至此 直接对标 Claude Code、Cursor、Windsurf 10/ Managed Agents（Gemini API） 单次 API 调用即可启动远程 Linux 沙盒 代理能推理、调用工具、执行代码、浏览网页 用 AGENTS.md 和 SKILL.md 定义自定义指令 11/ Google AI Studio 移动 App 即将推出 直接构建原生 Android 应用 一键发布到 Google Play 测试通道 前两个应用免费部署到 GCP，无需信用卡 12/ AI Inbox（Gmail） 自动生成上下文相关草稿 任务旁直接展示相关 Doc/Sheet/Slide 一键批量已读 今夏推出 Gmail Live 语音查询 13/ Google Pics 🎨 基于 Nano Banana 模型 对象分割、文本编辑翻译 传单、信息图都能做 深度集成 Workspace 14/ Docs Live & Talk to Keep 用语音创建编辑文档 自动从 Gmail、Drive 拉取相关信息 Keep 可把"脑暴"自动整理成清单 15/ Google Flow + Agent + Tools 🎬 Omni Flash 加入 Flow Agent 可多步任务、剧情头脑风暴、批量编辑 Tools 让你用自然语言"vibe code"任何创作工具可与其他用户分享和混编 16/ Flow Music 对话式制作音乐视频 单独编辑特定片段的歌词、风格、乐器 歌词可翻译为其他语言 17/ Ask YouTube 跨长视频和 Shorts 的对话式搜索 复杂查询如"如何教孩子骑自行车" 结构化、交互式答案 本月在美国桌面端实验性推出 18/ Gemini for Science 🔬 三大实验工具： Hypothesis Generation（Co-Scientist 多代理"想法竞赛"） Computational Discovery（AlphaEvolve 并行测试千种代码） Literature Insights（NotebookLM 结构化文献） 集成 30+ 生命科学数据库 19/ Android XR 智能眼镜 👓 两类：音频眼镜 + 显示眼镜 合作品牌：Gentle Monster、Warby Parker、三星 今秋上市，兼容 Android 和 iOS EssilorLuxottica × Meta 的独家护城河被撕开 20/ SynthID + C2PA 扩展 水印验证已用 5000 万次 扩展至 Search 和 Chrome 新增 C2PA Content Credentials，识别相机原图是否被改 OpenAI、Kakao、ElevenLabs 加入采用 21/ 新 $100 AI Ultra 计划 为开发者和高级创作者定制 用量上限是 AI Pro 的 5 倍 20TB 云存储 AI Pro 现免费包含 YouTube Premium Lite（每月 $8.99）

$NVDA NVIDIA PRICE TARGETS ACROSS THE STREET FOLLOWING Q1 EARNINGS: Baird $300 —> $500 Evercore ISI $352 —> $413 Melius $380 —> $400 BofA Securities $320 —> $350 Benchmark $250 —> $335 Raymond James $323 —> $330 Wedbush $300 —> $330 Bernstein $300 —> $315 KeyBanc $300 —> $310 Truist Securities $287 —> $307 Jefferies $275 —> $300 Mizuho $275 —> $300 Fubon Securities $270 —> $300 Morgan Stanley $285 —> $288 Goldman Sachs $250 —> $285

Micron $MU is speaking today at a JPMorgan conference and just released this: “Our financial outlook has strengthened since our last earnings call. We're on track for another substantial record free cash flow in fiscal Q3.“

NEW: @roundhill files for yet another AI related ETF with concentrated holdings. This time its the Roundhill Photonics & Optics ETF -- Ticker will be $LYTE

FSD Supervised is now available in: - United States - Canada - Mexico - Puerto Rico - China - Australia - New Zealand - South Korea - The Netherlands - Lithuania

今天英伟达财报的最大亮点： “储存占物料清单的比例从 9.3% 增加到 25.6%” 智能体时代，你就细品吧

Memory went from 9.3% of BOM to 25.6% of BOM

看起来大肉还是在 HBM 上面。

Trailing its competitors in the #HBM4 race, #Micron is aggressively pushing to catch up, viewing #Samsung’s recent labor unrest as a strategic window to attract top-tier talent. Given that #HBM production is exceptionally complex and demands highly specialized expertise, #TrendForce considers Micron’s move to recruit experienced personnel directly in #Seoul to be a logical and necessary strategy to bolster its competitive position.

KIS comment on NVIDIA Vera CPU: On today’s conference call, NVIDIA stated that it expects the standalone Vera CPU market to reach $20 billion in FY2027. The unit price of Grace CPU is estimated at around $3,000–$5,000. Since Vera is the successor to Grace and is optimized for AI agentic workloads, we expect Vera to carry a higher ASP of roughly $5,000–$8,000 per unit. Assuming a Vera CPU ASP of $8,000, a $20 billion market would imply 2.5 million CPUs. It remains unclear whether standalone Vera CPU sales will include the same SoCAMM capacity as NVL72. However, assuming the same capacity is applied, each Vera CPU would have 8 SoCAMM slots. Assuming 192GB per module, SoCAMM capacity per Vera CPU would be 1,536GB. Therefore, FY2027 SoCAMM demand for Vera CPU would be: 2.5 million CPUs × 1,536GB = 3.84 billion GB, or 30.72 billion Gb. CY2026, which broadly overlaps with NVIDIA’s FY2027, SoCAMM supply from the three major DRAM makers is estimated at 30 billion Gb. Therefore, combined SoCAMM demand from NVIDIA’s standalone Vera CPU sales and VR NVL72 sales already appears likely to exceed the annual supply capacity of 30 billion Gb. In our previous in-depth report, assuming CY2027 VR NVL72 shipments of 100,000 servers, we estimated the SoCAMM TAM at 44 billion Gb based on 192GB modules. If additional SoCAMM demand from standalone Vera CPU sales is added, the CY2027 SoCAMM TAM could exceed 80 billion Gb. An annual 80 billion Gb of LPDDR5 would be nearly equivalent to the annual LPDDR5 TAM used for smartphones. The shortage of LPDDR5 — and of DRAM overall — is likely to intensify further over time. $MU $DRAM

$NVDA NVIDIA Q1 FY27 (April quarter). • Revenue +85% Y/Y to $81.6B ($2.6B beat). • Operating margin 66% (+16pp Y/Y). • Non-GAAP EPS $1.87 ($0.10 beat). Q2 FY27 guidance: • Revenue $91.0B ($3.0B beat).

SK Hynix Raises Q2 Forecast on Rising DRAM and NAND Prices $MU $DRAM $SNDK