Rohan Paul@rohanpaul_ai · 6月19日68

ASML just became the center of a US-China chip fight after Washington said it fears a banned EUV lithography tool may have reached China, while ASML says it never shipped one there. US Commerce Secretary Howard Lutnick outlined concerns to ASML’s senior leaders that one of its top-of-the-line machines may have made its way into China, in violation of US-led export restrictions. EUV uses 13.5nm light to print the smallest chip patterns, which lets foundries pack more transistors into AI processors and phones. ASML however, denies selling extreme ultraviolet chip tool to China. For context, ASML is the sole produce of EUV lithography machines, and each one costs roughly $250M today, with newer High-NA EUV versions closer to $ 400M. And ASML sold of only 48 EUV systems in 2025, and its market share in EUV is effectively 100%. There are 314 EUV machines in operation around the world, 26 that have been decommissioned — and none in China, according to Bloomberg. Senior US administration officials said that they have evidence of ASML shipping to China specialty equipment used to transport EUV machines, as well as other components that could be used in EUV systems. The proof of those exports is too sensitive to disclose, they said Reuters had already reported a Shenzhen EUV prototype built with former ASML engineers, but making EUV light is still far from running a factory-grade production tool. The hard part is repeatable nanometer precision across optics, masks, software, stages, wafers, maintenance, and defect control. Shares of ASML fell ~2.6% in morning trading after this report . China’s blocked access to EUV equipment may be Huawei Technologies Co’s biggest chipmaking hurdle, as it tries to rival Nvidia in AI chips. Huawei’s chip head recently highlighted how the company is building chips without ASML’s EUV machines. If any EUV tool reached China, it would likely rank as 1 of the biggest known breaches of US-led export controls meant to restrict Beijing’s access to military-useful AI. There are also gaps between what US and non-US companies can sell to certain Chinese customers on Washington’s entity list, which places stricter limits on named firms than on China overall. For Huawei, the US has basically blocked nearly all technology access and used rules beyond its borders to apply those limits to firms like ASML. But US rules still permit some overseas dealings with other listed Chinese companies, including Huawei partner SwaySure Technology Co, which joined the list in 2024. Washington does have the power to force ASML to fully cut ties with SwaySure, but President Donald Trump’s team has not used it. The Netherlands has already placed curbs on ASML selling EUV machines and some advanced immersion deep ultraviolet, or DUV, lithography tools to China. But the Biden administration grew frustrated when ASML increased shipments of gear facing upcoming bans before certain DUV controls were officially enforced. Now if indeed EUV export-control violation is proven, that would certainly not mean the US “bans ASML,” because ASML is too central to TSMC, Samsung, Intel, Nvidia, Apple, and US chip policy. The more realistic damage would be Dutch investigations, fines, forced compliance controls, tougher license rules, and pressure to cut ASML’s China service, spare parts, software updates, and DUV sales, which could hurt the roughly 20% of 2026 revenue ASML expects from China. --- reuters .com/world/china/us-tells-asml-it-is-concerned-china-may-have-top-chip-tool-bloomberg-news-2026-06-19/

译美国商务部长Lutnick向ASML高层提出担忧，指一台顶级EUV光刻机可能已流入中国，违反出口管制。ASML否认向中国出售EUV设备。EUV每台约2.5亿美元，全球运行314台，据彭博中国为零。若违规属实将是美国对华芯片管制重大突破。ASML股价早盘跌约2.6%。路透曾报道深圳用ASML前员工搭建原型机但离量产有差距。美国可强制ASML切断与中国客户合作，荷兰已限制EUV及部分DUV出口，进一步收紧可能影响ASML约20%的2026年中国区收入。

AYi@AYi_AInotes · 6月19日64

我终于明白现在随处的AI时代对普通人的善意在哪，普通人上班一个月赚的工资，够开一家跑通全商业闭环的互联网公司，而且能连开好几年， 以前创业要凑团队，起码融个几十万，现在你一个人花$20就能先跑起来去做验证。成本几乎可以忽略不计🚀 跟大家分享个我最近悟透的一个扎心真相： 我们上班一个月赚的工资， 够开一家跑通完整商业闭环的互联网公司，连开好几年。 咱们每天挤地铁赶打卡，看老板脸色，应付办公室内耗， 一个月辛辛苦苦赚的钱，扣完房租吃饭和娱乐开销，能存下的也不多， 可能很多人都想不到现在启动一家能赚钱的互联网公司， 每月固定成本其实只需要二十美元左右。 除了Claude Pro 或最近很火的 OpenAI Codex 辅助全栈开发与 agentic coding， 剩下的全链路基建，几乎都能用免费额度起步， 从后端数据库部署上线到用户认证数据分析支付变现的全链路基建， 几乎全有免费起步额度，域名平摊到每月才一美元， 加起来一杯咖啡的钱，就能搭起一整套过去几十人团队才能撑起来的商业底盘：• Supabase 做后端 + 数据库（含 pgvector 向量功能）（免费额度够早期使用，真实规模后 Pro 约 $25/月） • Vercel 做部署上线（Hobby 免费但仅限非商业验证，商业化需升级 Pro $20/月） • GitHub 做版本控制（免费） • Clerk 做用户认证（免费额度高，50k MRU 内免费，早期完全够用） • Cloudflare 做 DNS 解析（免费） • PostHog 做数据分析（免费额度很慷慨，多数项目可长期免费使用） • Sentry 做错误监控（免费额度 5k errors/月，早期够用） • Upstash 做 Redis 缓存（免费额度够原型和早期使用） • Resend 做邮件服务（免费 3000 封/月但每天限 100 封，早期够用，规模后 Pro $20/月） • Stripe 按交易抽成，没有固定成本 域名平摊下来每月只要一美元左右。 以前开公司做产品，得先组团队、融资金、搭基建， 开发、运维、产品、运营、财务，光人力成本一个月就要几十万， 这还还没算服务器和软件的固定开销， 普通人根本连入场的资格都没有， 但现在不一样了， 所有的公司职能，被 AI 和这些托管服务大幅简化， 你一个人，就能当 CEO、产品经理、全栈开发、运营、财务， 不需要融资，不用稀释股权，不用看投资人脸色， 不需要团队，不用搞管理内耗，不用应付办公室政治。 你只需要找到一个能解决用户真实痛点的需求， 用 AI 辅助把产品快速做出来，然后持续迭代和获取用户， 其实一人公司的核心， 从来不是一个人干所有的活，

译推文指出，普通人一个月工资足以启动一家跑通商业闭环的互联网公司，每月固定成本约$20，主要来自Claude Pro或OpenAI Codex辅助开发。其余后端数据库、部署、认证、分析、监控、缓存、邮件、支付等全链路基建均有免费起步额度，域名约$1/月。同时，引用指出当前AI范式（视频生成、VLA）仅学习统计相关性，缺乏因果理解；UCSD黄碧薇教授提出四代AI划分（相关性小模型→因果小模型→相关性大模型/LLM→因果大模型），其团队Aether AI今日完成首轮融资。

SemiAnalysis@SemiAnalysis_ · 6月19日46

100% of AI chip startups have slides/“simulated performance data” showing that their chip is way better, but 99% of custom ASICs fail. Why? The MATH isn’t MATH until you realize that AI chips are about Software. It is relatively easy to build a chip and put numbers onto slides; it is hard to build great software. That is why 99% of AI chip startups fail.

译100%的AI芯片初创公司都有幻灯片/“模拟性能数据”显示它们的芯片好得多，但99%的定制ASIC失败了。为什么？直到你意识到AI芯片是关于软件的，MATH才算是MATH。构建芯片并把数字放到幻灯片上相对容易；构建出色的软件很难。这就是为什么99%的AI芯片初创公司会失败。

AYi@AYi_AInotes · 6月19日53

说个暴论：AI算力的下一个战场不应该会从电网容量转移到海浪能上。 当所有人都在抢电网指标的时候，有人直接把AI计算搬去了海上，靠海浪发电完全离网运行。 Peter Thiel刚领投了1.4亿美元，押注一个完全离网的海上AI数据中心，单节点1MW，靠海浪发电，海水冷却，低轨卫星回传。 也就是说，AI算力的下一个解法，不在陆地上的电网里了，一群硅谷最聪明的人，正在把 AI 服务器从电网里拔出来，扔进太平洋。 第三代浮动节点Ocean-3，完全离网运行。 海浪起伏驱动内部涡轮发电，单节点功率可达1MW。 密封舱服务器靠海水直接冷却，不消耗淡水。 算力通过低轨卫星回传，多节点可组成分布式集群。 不用征地，不用接电网，不用等数年的变电站审批。 哪里波浪能量强，就自主航行到哪里。 所有人都在挤破头抢电力指标。 有人直接把计算，搬到了能源本来就在的地方。 项目刚完成1.4亿美元B轮融资， 首台试点节点预计今年8月在北太平洋下水。 海洋腐蚀，公海维护，卫星带宽限制，都是实打实的坎。 真正规模化商用，至少要等到2027年。 激进的方向从来不缺质疑， 但瓶颈永远会倒逼，跳出原有赛道的解法。

译Peter Thiel领投1.4亿美元B轮，押注完全离网的海上AI数据中心。第三代浮动节点Ocean-3单节点1MW，靠海浪发电、海水冷却、低轨卫星回传，无需征地接电网。首台试点预计今年8月在北太平洋下水，规模化商用至少等到2027年。同时，Aether AI官宣首轮融资，UCSD黄碧薇教授推出因果AI，将AI分为相关性小模型、因果小模型、LLM、因果大模型四代，旨在从视频中自动提取物理规律。

Berryxia.AI@berryxia · 6月19日42

兄弟们，Claude 是一点也不闲着啊！ 这不又又又整活了… Claude Code直接上线了Artifacts功能，让你的coding session变成能实时刷新的共享互动页面。 现在你在Claude Code里做的任何事，比如PR 演示、项目dashboard、调试过程，都能一键生成一个互动页面，通过私有链接分享给团队。 关键是它会随着你的session继续工作而自动刷新，大家看到的永远是最新的版本。 Artifacts能调用你整个session的上下文：代码库、插件、技能、已连接的工具。 分享完全在组织内，私密性有保障。目前在Team和Enterprise计划的beta版可用。 这其实把AI coding从“单人黑箱”变成了“团队实时工作台”。 以前你得手动截图、复制代码、写说明，现在直接把AI的思考过程和输出做成活的artifact，别人点开链接就能看到完整脉络，还能跟着一起迭代。 以前团队协作AI coding最麻烦的就是上下文传递和版本同步，现在Artifacts直接把这个痛点干掉了。 AI不再只是帮你写代码，它可以帮你把整个工作过程变成可共享、可进化的活文档。 这波更新把Claude Code从“个人生产力工具”往“团队协作平台”又推了一大步。

译Claude Code 新增 Artifacts 功能（Team 和 Enterprise 计划 beta 版）。用户可从 coding session 生成交互页面（如 PR 演示、项目 dashboard），通过私有链接分享给团队；页面随 session 自动刷新，调用代码库、插件、技能等全部上下文。该更新旨在将 Claude Code 从单人工具拓展为团队实时协作平台，解决上下文传递和版本同步痛点。

Jeff Dean@JeffDean · 6月19日49

My @Google colleagues @NormJouppi, Sridhar Lakshmanamurthy, Cliff Young, and David Patterson recently wrote a paper that will appear in the July/August 2026 edition of @ieeemicro titled "Google's Training Supercomputers from TPU v2 to Ironwood: Architectural Stability, Scale, Resilience, Power Efficiency, and Sustainability Across Five Generations". It's chock full of interesting data about the evolution of TPU chip generations, as well as how workloads at Google have transformed over time (hint: lots more transformer-based models!), and how the generations have gotten ~30X more energy efficient per flop. Lots of changes over these generations: Air cooling in TPUv2 to water cooling in TPUv3 onwards 2D to 3D torus-based interconnects 30X improvement TFLOPS/Watt 256 chips (TPUv2) to 9216 chips (Ironwood) per pod Read the full paper: https://arxiv.org/abs/2606.15870

译Jeff Dean 等 Google 同事发布论文，回顾 TPU v2 到 Ironwood 五代训练超算的演进，将于 2026 年 7/8 月发表于 IEEE Micro。关键变化：TPU v2 采用气冷，v3 起改为水冷；互联从 2D 升级为 3D torus；每 pod 芯片数从 256 增至 9216；每 flop 能效提升约 30 倍。此外，Google 内部工作负载已大幅转向基于 Transformer 的模型。

jason@jxnlco · 6月19日42

Looks like some very solid performance numbers from GPT-5.5 in real world application! https://www.techtimes.com/articles/318607/20260618/softbank-openai-launch-ai-patching-advisory-japan-internal-test-found-10500-flaws.htm

译看起来 GPT-5.5 在现实世界应用中有非常扎实的性能数据！

Rohan Paul@rohanpaul_ai · 6月19日52

US approves faster grid connections for data centers. The plan lets PJM review up to 10 projects a year, but only projects of at least 250MW qualify, i.e. this is aimed at big power plants, not small local upgrades. The bottleneck is interconnection, the slow engineering and approval process that decides whether a new generator can safely plug into the grid without overloading wires, transformers, or protection systems. PJM says the threat is a possible 60GW shortfall over the next decade, and the pressure is coming from data centers that need steady electricity for servers, cooling, networking, and backup systems. The new track runs from July-26 through 2027, and selected projects are supposed to get grid agreements within 10 months and start operating within 3 years. --- wric .com/news/politics/ap-federal-regulators-back-trumps-plan-to-speed-power-to-energy-hungry-ai-data-centers/

译美国批准加快数据中心电网接入计划。PJM每年最多审查10个项目，但每个项目容量须至少250MW，主要针对大型发电厂。瓶颈在于互联审批流程缓慢。PJM警告未来10年可能出现60GW电力缺口，压力来自数据中心对服务器、冷却等持续供电需求。新计划从2026年7月运行至2027年，入选项目预计在10个月内获得电网协议，3年内投入运营。

Elon Musk@elonmusk · 6月19日50

Grok now on Databricks

译Grok模型现已在Databricks Agent Bricks上可用，可将xAI最新模型引入企业数据以构建强大的AI智能体。Elon Musk宣布了这一消息。

SemiAnalysis@SemiAnalysis_ · 6月19日65

Stop Saying Half of 2026 US Datacenter Capacity Is Canceled: Don't Believe The Vibecoded Estimates, Go Through Every Individual Filing Instead https://newsletter.semianalysis.com/p/stop-saying-half-of-2026-us-datacenter

译停止说2026年美国数据中心一半容量被取消： 不要相信那些Vibecoded Estimates， 而是逐一审查每个申报文件。 https://newsletter.semianalysis.com/p/stop-saying-half-of-2026-us-datacenter

xAI@xai · 6月19日38

Grok models are now available on Databricks Agent Bricks. Bring SpaceXAI's latest models to your enterprise data to power capable AI agents. https://x.ai/news/grok-databricks

译Grok 模型现已可在 Databricks Agent Bricks 上使用。 将 SpaceXAI 的最新模型引入您的企业数据，以驱动强大的 AI 智能体。 https://x.ai/news/grok-databricks

AK@_akhaliq · 6月18日41

GLM-5.2 is free on Hugging Face Inference Providers through Zai, Together AI, Novita, Fireworks, DeepInfra for the next 6 hours Set it up with Pi, opencode, Codex, Claude Code or any coding agent

译GLM-5.2 在 Hugging Face Inference Providers 上通过 Zai、Together AI、Novita、Fireworks、DeepInfra 免费提供，持续 6 小时 可搭配 Pi、opencode、Codex、Claude Code 或任何编码智能体进行设置。

SenseTime@SenseTime_AI · 6月18日43

Speed matters — so we built an 𝟴-𝘀𝘁𝗲𝗽 𝗱𝗶𝘀𝘁𝗶𝗹𝗹𝗲𝗱 𝗟𝗼𝗥𝗔 of 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗨𝟭-𝟴𝗕-𝗠𝗼𝗧-𝗜𝗻𝗳𝗼𝗴𝗿𝗮𝗽𝗵𝗶𝗰 for you. ⚡️ 𝟭𝟮.𝟱𝘅 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝘀𝗽𝗲𝗲𝗱𝘂𝗽 🎨 Infographic quality mostly on par with the base model Get started with SenseNova-U1-8B-MoT-Infographic-LoRA-8step-V1.0: 💻Github: https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/base_vs_distill.md#run-base-and-distilled-model 🤗https://huggingface.co/sensenova/SenseNova-U1-8B-MoT-LoRAs/blob/main/SenseNova-U1-8B-MoT-Infographic-LoRA-8step-V1.0.safetensors 👾Discord: http://discord.gg/BuTXPHmQub

译商汤推出 SenseNova-U1-8B-MoT-Infographic 模型的 8-step 蒸馏 LoRA（SenseNova-U1-8B-MoT-Infographic-LoRA-8step-V1.0），实现 12.5 倍推理加速，信息图（infographic）生成质量基本与基模型持平。模型权重已开源至 HuggingFace，GitHub 提供使用文档。

OpenCode@opencode · 6月18日28

MiniMax M3 has been getting a lot of love We are extending the 3x limits till the end of the month on OpenCode Go

译MiniMax M3 收获了大量喜爱 我们在 OpenCode Go 上将 3 倍限制延长到月底。

X.PIN@thexpin · 6月18日47

Alibaba Cloud launched its first data centers in France this week, its third European hub after Germany and the UK. Agentic AI services are planned for the region in the second half of the year. The timing is pointed. The European Commission published its Cloud and AI Development Act on June 3rd, explicitly flagging "limited data center capacity" as a threat to European digital independence. The EU wants to be a global AI leader. It needs infrastructure to do it. Europe's data sovereignty push is largely about reducing dependence on US hyperscalers — AWS, Azure, Google Cloud. Whether European businesses and regulators will draw that distinction sharply remains to be seen .So far, Alibaba Cloud's European expansion has proceeded without significant political friction — Germany and the UK came first, France now.

译阿里云本周在法国启用了首个数据中心，这是继德国和英国之后的第三个欧洲节点，计划今年下半年在该地区推出智能体AI服务。此举正值欧盟委员会6月3日发布《云计算和AI发展法案》，该法案明确指出“有限的数据中心容量”威胁欧洲数字独立。欧洲的数据主权推动旨在减少对美国云巨头AWS、Azure、Google Cloud的依赖。目前阿里云在欧洲的扩张尚未遇到重大政治阻力。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6月18日61

重點解讀台積電的玻璃核心載板投影片 台積電在 6 月 11 日的日本 JPCA Show 2026，進行主題為「AIの進化に不可欠な先端パッケージング技術」的簡報，該份簡報約40頁，當中某張標題為「Glass Substrate Development for CoWoS」的投影片流出，引發廣泛關注。 以下是對該投影片（見附圖）的重點解讀，能網路科普查到的技術細節就不多說了，要注意的是投影片上的 COP 不是 Chip-on-Package 的縮寫，而是 Coplanarity、也就是平整度 / 共面度。 ▌重點結論： 1. 台積電正式宣布與 Ibiden 以及群創合作，開發玻璃核心載板（glass core substrate），結構為玻璃上下各黏合 ABF 的三層結構設計，該技術就是用於 CoPoS 的 oS。 2. 市場低估玻璃核心載板的重要性，該技術對台積電是「must have」，意即 CoPoS 中，oS 的重要性高於 CoP，這也是該技術進行測試時，先搭配既有的 CoW 而非 CoP 的原因。 3. 玻璃核心載板單價較既有 ABF 載板高出數倍，群創加工的玻璃單價非常高，為最核心的材料。除 Nvidia 外，目前已有兩家美系客戶同樣表達高度興趣。 ▌與本投影片相關的產業調查： 1. 本投影片提及的玻璃核心載板由 250x250mm 切割而來，ABF 增層主要採用 Ajinomoto 的 GL107 並混搭 ABF-GCP，以 2027–2028 主流 AI 晶片 ABF 規格的 24-28 層進行測試 2. 台積電實驗時的 CoW 是測試載具（test vehicle），足以驗證採用複合材料時最具挑戰性的機械結構問題。測試結果良好意味著台積電、Ibiden 與群創已合作突破關鍵技術瓶頸。 3. 目前是由 Ibiden 負責切割 250x250mm 的玻璃核心載板。待 2H27 採用 510x515mm 做量產前模擬時，若 Ibiden 仍想降低生產複雜性以維持超高毛利率，可能會改交由更熟悉玻璃特性的群創切割。 ▌流出的投影片內容是將 CoPoS 中的 oS、也就是玻璃核心載板（投影片中的 glass-SBT）與 CoW 搭配的技術驗證結果，這是為解決該投影片前一頁所提到的「Substrate mechanical and electrical Dilemma」，而這顯著凸顯了 CoPoS 中 oS 的重要性。 1. CoPoS 中，CoP 要解決的是生產效率 / 切割經濟性的問題，這與成本與售價有關；而 oS 要解決的是翹曲與耐用性問題，這牽涉到能否做出晶片，以及晶片能否運作。 2. CoP 與 oS 兩者整合相得益彰，但展望未來數年，兩者的技術定位還是有些差異。CoP 是可選的絕佳優化選項（very-nice-to-have），沒有它的代價就是晶片更貴；但 oS 是必需品（must-have）， 沒有它可能連能否做出可用晶片都是問題。 3. 比較定位差異不是為了捧 oS 貶 CoP，這牽涉到客戶願意為哪個技術環節付錢的現實問題，細節下面分析。 ▌投影片中含金量最高的是電源完整性（power integrity；PI）改善，這對客戶意義重大，這也代表玻璃核心載板生產穩定後，台積電獲利能力與競爭優勢可望同步提升。 1. 技術說明：玻璃核心載板薄 → TGV（through glass via）垂直導通路徑短 → 導通路徑電阻（R）跟迴路電感（L）同降 → PI 改善 2. 對客戶意義重大原因：PI 改善 → 供電更穩 → 釋出功率餘裕（power headroom）→ 可整合更多電晶體、或拉高運作時脈 → AI 晶片算力提升 3. 對客戶而言，生產效率是台積電的基本責任，客戶不會為此多付錢；但 AI 算力提升能直接轉化為客戶的競爭力與獲利，故客戶願意為此買單。這也是 Nvidia 積極看待玻璃核心載板的原因。 4. 對台積電而言，玻璃核心載板可提升良率並降低成本，同時提高 AI 晶片的算力與售價，既是降本工具，也是漲價籌碼，對獲利與競爭力都是加分。 5. 目前載板成本佔 AI 晶片 BOM 約低個位數，封裝良率造成的損失約載板成本的 5-10 倍，故即便未來玻璃核心載板成本高於目前的數倍以上，但佔 BOM 比重仍低，且可改善封裝良率造成的損失，故預期玻璃核心載板的高單價不會影響客戶採用意願。 ▌簡報後的問答環節，有聽眾提問關於玻璃核心載板的 TGV 細節，台積電當場拒絕回答，因為玻璃核心載板的關鍵技術就是 TGV，核心 know-how 目前掌握在台積電與群創手中。相較下，另一個提問者的問題是關於 IVR、eDTC、與 LSI 的整合，台積電就回答了不少。 ▌根據產業調查，若一切順利，台積電的目標是在 4Q28-1Q29 開始量產玻璃核心載板，以符合 Nvidia AI 晶片迭代節奏。順帶一提，許多人在傳的 Ibiden 的法說投影片，上面將玻璃核心載板時程列為 CY30，我對此的解讀是：對外向來保守謹慎的 Ibiden 將玻璃核心載板正式列為發展路線，這更確定了該技術長期趨勢。但從 Ibiden 投影片的其他細節與市場資訊不完全一致來看，例如 reticle 時程與台積電公開宣稱的差約一個世代、Rubin Ultra 載板尺寸明顯大於其在 CY26-27 標示的 90x90 等，這說明了在預測未來時，需隨時多方交叉驗證。

译台积电在6月11日日本JPCA Show 2026展示玻璃核心载板投影片，宣布与Ibiden及群创合作，用于CoPoS的oS，结构为玻璃上下各黏合ABF的三层设计。单价较现有ABF载板高出数倍，除Nvidia外另有2家美系客户表达高度兴趣。技术方面，由250x250mm切割，ABF增层采用GL107混搭ABF-GCP，以24–28层测试，电源完整性改善可提升AI算力。TGV为核心know-how，由台积电与群创掌握。目标4Q28–1Q29量产。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6月18日64

Breaking down TSMC's glass core substrate slide On June 11, at JPCA Show 2026 in Japan, TSMC gave a roughly 40-slide presentation titled "Advanced Packaging Technology Essential to the Evolution of AI" (AIの進化に不可欠な先端パッケージング技術). One slide from the deck, titled "Glass Substrate Development for CoWoS," has since leaked online and widespread attention. Here's a closer read of that slide (see attached image). I'll skip the technical background that is already widely available. One thing to flag: the "COP" on the slide does not stand for Chip-on-Package. It means Coplanarity. ▌ Key conclusions: 1. TSMC has officially announced a partnership with Ibiden and Innolux to develop a glass core substrate. The structure is a three-layer design, a glass core sandwiched between two ABF build-up layers. This is the "oS" in CoPoS. 2. The market underestimates how important the glass core substrate is. It's a must-have capability for TSMC. In other words, within CoPoS the "oS" matters more than the "CoP", which is also why, when it was tested, it was paired with the existing CoW rather than with CoP. 3. The glass core substrate costs several times more per unit than existing ABF substrates. The glass processed by Innolux is very expensive per unit and is the single most critical material. Besides Nvidia, two US-based customers have also expressed strong interest. ▌ Industry checks tied to this slide: 1. The glass core substrate shown on the slide is cut from a full-size 250×250mm one. The ABF build-up layers mainly use Ajinomoto's GL107, mixed with ABF-GCP, and were tested at 24–28 layers, which is the mainstream ABF spec for AI chips in 2027–2028. 2. The CoW used in TSMC's experiment is a test vehicle. It is sufficient to validate the most challenging mechanical-structure issues that arise when working with composite materials. Good results mean TSMC, Ibiden, and Innolux have together broken through the critical technical bottleneck. 3. Ibiden currently handles cutting the 250×250mm glass core substrate. When the 510×515mm format is used for pre-mass-production simulation in 2H27, if Ibiden still wants to reduce production complexity to protect its ultra-high gross margins, it may hand the cutting over to Innolux, which is more familiar with the properties of glass. ▌ The leaked slide shows the validation results of pairing CoW with the "oS" in CoPoS, i.e., the glass core substrate (labeled "glass-SBT" on the slide). This addresses the "Substrate mechanical and electrical Dilemma" raised on the previous slide, and it strongly underscores how important the "oS" is within CoPoS. 1. Within CoPoS, what CoP solves is production efficiency / cutting economics, which ties to cost and price. What the oS solves is warpage and durability, which determines whether the chip can be made at all, and whether it can work. 2. CoP and oS complement each other well when integrated, but looking out over the next few years their technical roles still differ. CoP is a very-nice-to-have optimization, and going without it simply means a more expensive chip. But the oS is a must-have. Without it, even being able to make a usable chip is in doubt. 3. Comparing their roles isn't about elevating oS at the expense of CoP. It comes down to the practical question of which technical piece customers are willing to pay for. Details below. ▌ The real gold here is the power integrity (PI) improvement shown on the slide. This matters a great deal to customers, and it means that once glass core substrate production stabilizes, TSMC's profitability and competitive edge should rise in tandem. 1. How it works: the glass core substrate is thin → the vertical conduction path through TGV (through-glass vias) is short → conduction-path resistance (R) and loop inductance (L) both drop → PI improves. 2. Why it matters to customers: better PI → more stable power delivery → frees up power headroom → room to integrate more transistors, or to push clock speeds higher → more AI compute. 3. For customers, production efficiency is TSMC's basic responsibility, so they won't pay extra for it. But gains in AI compute translate directly into the customer's own competitiveness and profit, so customers are willing to pay for that. This is why Nvidia is so positive on the glass core substrate. 4. For TSMC, the glass core substrate raises yield and lowers cost while also boosting both the compute and the selling price of AI chips. It's both a cost-cutting tool and a pricing lever, a plus for profitability and competitiveness alike. 5. Substrate cost currently accounts for a low single-digit percentage of an AI chip's BOM, while losses from packaging yield run roughly 5–10× the substrate cost. So even if the glass core substrate ends up costing several times more than today's, its share of the BOM stays low, and it can cut the losses from packaging yield. The high unit price is therefore not expected to dampen customers' willingness to adopt it. ▌ In the Q&A after the presentation, an audience member asked about TGV details for the glass core substrate. TSMC declined to answer on the spot, because TGV is the key technology behind the glass core substrate, and the core know-how currently sits with TSMC and Innolux. By contrast, when another attendee asked about integrating IVR, eDTC, and LSI, TSMC answered at length. ▌ According to industry checks, if all goes well, TSMC is aiming to start mass production of the glass core substrate in 4Q28–1Q29, to match the cadence of Nvidia's AI chip iterations. As a side note: the Ibiden earnings presentation slide that many people have been circulating lists the glass core substrate timeline as CY30. My read is this: Ibiden, which has always been conservative and cautious in public, has now formally put the glass core substrate on its roadmap, which further confirms the long-term trend for this technology. That said, some other details on Ibiden's slide don't fully line up with what's known in the market. For example, its reticle timeline is off from TSMC's public claims by about a generation, and the Rubin Ultra substrate size is clearly larger than the 90×90 it marked for CY26–27. It's a reminder to always cross-check across multiple sources when forecasting the future.

译台积电在JPCA Show 2026公开玻璃核心基板，与Ibiden、Innolux合作，采用玻璃芯夹两层ABF的三层结构，用于CoPoS的“oS”部分。基板单位成本数倍于现有ABF，Innolux处理的玻璃最贵。客户包括Nvidia和两家美国公司。验证显示其解决了力学与电学难题，显著改善电源完整性。目前切割自250×250mm基板，ABF层24-28层，对应2027–2028年AI芯片主流规格。CoW测试车辆已验证复合材料机械结构瓶颈已突破。

Ant Ling@AntLingAGI · 6月18日50

It has been a privilege to collaborate so closely with the SGLang team @lmsysorg on optimizing Ling-2.6-1T. 🥳 The resulting performance gains speak for themselves: -53% reduction in MoE pre-fill latency -Up to 1.77x higher decode throughput on a 16-chip TPU v7x slice compared to a similar H200 cluster A significant milestone in efficient MoE scaling and hardware utilization!

译蚂蚁百灵与 SGLang 团队合作，将 1T 参数的混合 MoE 模型 Ling-2.6-1T 通过 SGLang-JAX 部署至 TPU v7x。优化包括：升级 Fused MoE V2 内核（token 和累加器驻留 VMEM，双缓冲专家权重，隐藏路由与预取）；混合内存池（10 个全注意力层 per-token MLA KV + 70 个 GLA 层 per-request 循环状态）；GLA 线性注意力逐块并行预填充；单控制器 DP 保持分组 RMSNorm 芯片本地化。效果：MoE 预填充延迟降低 53%；在 16 芯片 TPU v7x 切片上，解码吞吐量比同类 H200 集群最高提升 1.77 倍。

Rohan Paul@rohanpaul_ai · 6月18日52

Large AI campuses are indeed strategic infrastructure of a nation. US Department of Justice just moved to protect xAI’s Mississippi AI data center because the government sees high-speed AI infrastructure as part of America’s national security base. DOJ’s argument treats Grok and the systems around it as critical technology, not just consumer software, because advanced models can support defense operations. The move definitely reflects a wider policy shift: the U.S. does not want strategic AI workloads to depend on fragile supply chains, slow permitting cycles, or foreign-controlled compute capacity. AI compute is becoming a defense resource, where available power, data center scale, model capability, and operational readiness are now tightly linked.

译美国司法部近日保护 xAI 位于密西西比州的 AI 数据中心，认为高速 AI 基础设施属于美国国家安全基础。司法部将 Grok 及其系统视为关键技术而非普通消费软件，因先进模型可支持国防行动。此举反映政策转变：美国不希望战略 AI 工作负载依赖脆弱供应链、缓慢审批流程或外国控制算力。AI 算力正成为国防资源，可用电力、数据中心规模、模型能力与运营准备度紧密关联。

meng shao@shao__meng · 6月18日82

Vercel 开源了他们的 Agent Framework「Eve」 Agent 即目录，生产级能力开箱即用，把反复出现的 Agent 形态抽象成框架，让开发者只写「做什么」，而不是「怎么跑起来」。 https://vercel.com/blog/introducing-eve 核心设计：Agent 是一个目录 agent/ agent.ts # 模型与配置 instructions.md # 系统提示 / 人格 tools/ # 可执行能力 skills/ # 领域知识（Markdown） subagents/ # 子 Agent 委托 channels/ # Slack、Discord 等入口 schedules/ # 定时任务 connections/ # MCP / OpenAPI 外部连接 内置的生产能力 · 持久会话：每轮对话是可 checkpoint 的 durable workflow（基于开源 Workflow SDK），可暂停、崩溃/部署后恢复 · 沙箱：Agent 生成代码与主应用隔离；本地 Docker/microsandbox，部署用 Vercel Sandbox，可写 adapter · Human-in-the-loop：工具上设 needsApproval，暂停不占算力，审批后从断点继续 · Connections：MCP / OpenAPI 以文件声明；鉴权由框架代理，模型不接触 URL/凭证；Vercel Connect 处理 OAuth · 多 Channel：同一 Agent 服务 HTTP、Slack、Discord、Teams 等；Channel 间可 handoff · Tracing & Evals：OpenTelemetry 标准 trace；eve eval 可本地或 CI 跑，作部署门禁 开发与部署流程 · 本地：eve dev → TUI 可见每步（load_skill、tool call 等）；底层是 HTTP API，CI/脚本也可驱动。 · 部署：vercel deploy，Agent 即普通 Vercel 项目；部署不中断进行中的会话（在启动版本上跑完）。沙箱等通过 adapter 切换，代码不变。 · 接入团队：eve channels add slack 生成 channel 文件；审批在 Slack 里点按钮；schedules/ 用 cron 定时触发（部署为 Vercel Cron Job）。 · 工程化：Agent 进 Git（prompt/工具/skill 都有 diff 和 review）；Preview 部署可提前测 Slack bot；eve eval 进 CI 防回归。 Vercel 内部验证 · d0：月 3 万+ 问数，权限与提问者对齐 · Lead Agent：自主 SDR，年成本约 $5k，回报约 32 倍 · Athena：RevOps 6 周无工程师搭建，接 Snowflake/Salesforce · Vertex：约 92% 工单自动解决 · draft0：内容审阅流水线 · V：路由 Agent，统一入口分发到百级 Agent 舰队

译Vercel 发布开源 Agent 框架 Eve，核心设计“Agent 即目录”：通过 agent.ts、instructions.md、tools、skills、subagents、channels、schedules、connections 等文件声明行为。内置持久会话（可 checkpoint）、沙箱隔离（本地 Docker/Vercel Sandbox）、Human-in-the-loop 审批（不占算力）、MCP/OpenAPI 连接（鉴权由框架代理）、多 Channel 支持（HTTP/Slack/Discord）、OpenTelemetry 追踪与 eve eval 门禁。本地 eve dev TUI，部署为普通 Vercel 项目，不中断进行中会话。内部已验证：d0 月 3 万+ 查询，Lead Agent 年成本约 $5k 回报 32 倍，Vertex 约 92% 工单自动解决。

SemiAnalysis@SemiAnalysis_ · 6月18日45

Wide Expert Parallelism increases the total memory bandwidth available per MoE deployment. This means the model distributes the MoE expert weights across multiple GPUs, so each GPU only needs to load a tiny fraction of the weights. This translates to higher throughput per GPU, increasing perf per dollar and perf per watt.

译Wide Expert Parallelism增加了每个MoE部署可用的总内存带宽。这意味着模型将MoE专家权重分布到多个GPU上，因此每个GPU只需加载一小部分权重。这转化为每个GPU更高的吞吐量，提升了每美元性能和每瓦性能。

elvis@omarsar0 · 6月18日70

You can only truly get this level of output when using orchestrator agents that can coordinate multiple agents across projects. Build your own orchestration layer now. And own it.

译金融科技公司 Block 自建内部 AI 系统 Builderbot，可跨整个代码库协调多个智能体。工程师在 Slack 中标记后，系统自动研究、规划并交付。当前日处理 20 万次操作，每周合并 1500 个 pull request，贡献了 Block 全部生产代码变更的 15%，将原需数月的流程缩短至数天。DAIR.AI 创始人 Elvis Saravia 强调，只有通过编排层协调多个智能体才能实现此类输出，建议团队自建编排层。

xAI@xai · 6月18日62

Use VMs with Grok Build preinstalled with one click

译一键使用预装 Grok Build 的虚拟机

AYi@AYi_AInotes · 6月18日59

NVIDIA锁 → AMD抢 → 独立产能稀缺 → 定价权/订单改善，类似之前GPU、电源或HBM的某些阶段， 我觉得这条逻辑链值钱就值钱在它清晰——英伟达锁产能、AMD 下场抢、独立产能快见底、谁有货谁定价，但还是希望大家明白，清晰不等于现在就该冲进去。 先讲积极的那一半：如果 CPO 采用加速、AMD 的长约落地、AAOI 德州扩产兑现、Sivers 的定位坐实、Luxnet 规模化出货，这几个节点打穿任意一个，相关公司就有量价支撑，链条是完整的，催化剂也是可追踪的。 但冷水必须浇一下， 1️⃣预期已经跑了一段，Sivers 这类近期涨幅不小。 2️⃣小盘股的波动不是一般人能扛的，AAOI 日内吞人不是一次两次。 3️⃣技术路线不是铁板钉钉——LPO 线性插拔如果跑通，对高功率 CW 的需求会被削一刀。 4️⃣扩产、良率、InP 基板紧缺，随便哪个卡住，时间表就往后推。 5️⃣最关键的：CPO 大规模放量可能要到 2026-2028 年。逻辑再顺，中间隔了两年，什么变数都能发生。 6️⃣地缘政治、AI capex 节奏、利率，这些宏观变量也悬在上面。 这故事有点像早期的 GPU、HBM——逻辑漂亮，波动巨大，看得懂牌的人不一定扛得住牌桌摇晃。 看来AI下一个卡脖子生死线，不是GPU也不是HBM，可能是没人盯的高功率CW激光器。 NVIDIA已经锁死两家头部供应商核心产能，同时AMD也正下场争抢仅剩的个位数独立供给。 CPO加硅光子已成AI大集群标配路线，InP基板从2024年起就进入持续紧缺状态。 AMD只是第一张多米诺骨牌，后续头部云厂商大概率会集体下场锁定产能。 西方可接单的独立激光产能所剩无几，连头部厂自身都要外包且已全线满产。

译Trendforce报告显示AMD正积极采购锁定CW激光器供应。此前NVIDIA已与Lumentum/Coherent签订多年协议锁死核心产能，西方独立产能仅剩Sivers、AAOI等少数厂商，Lumentum自身受限可能外包至日本产线，后者已满负荷。CPO+硅光子成AI标配，InP基板持续紧缺。AMD是第一张多米诺骨牌，Amazon、Microsoft等云厂商大概率跟进。但CPO大规模放量或要到2026-2028年，期间存在预期已涨、小盘波动大、LPO技术替代、扩产良率瓶颈及宏观不确定性等风险。

Rohan Paul@rohanpaul_ai · 6月18日51

Quite a massive inferencing rack breakthrough from @TensordyneInc . They just announced an AI-inference rack, claiming 13x the rack throughput of NVIDIA’s NVL72 GB300 in a DeepSeek-R1 comparison based on internal simulations. What makes this a big deal is that Tensordyne is attacking inference at the math level. AI chips spend huge amounts of energy moving and multiplying numbers. Napier (its AI inference racks) works in log space, where multiplication becomes addition, and addition is cheaper to build, switch, cool, and repeat billions of times per token. So instead of spending tons of transistor budget on heavy multiply circuits, Napier tries to shrink the math itself. So that means less chip area for compute and more for SRAM, resulting in less power per token and way more inference packed into the same rack. If they have made log math accurate and fast enough for real inference, then Napier is not just pushing more power into a rack, it is changing the cost of the basic operation behind model serving. AI inference is no longer just a FLOPS race. It is a rack-level fight over power, memory locality, interconnect latency, and how many paying tokens can be served before the economics break. They reported their TDN Rack reaches 363,000 tokens per second on DeepSeek-R1 at user speeds of 210 tokens per second per internal simulation, compared with 27,400 tokens per second for Nvidia’s NVL72 GB300. 🧵 1.

译TensorDyne 发布 AI 推理机架 Napier，声称在 DeepSeek-R1 上基于内部模拟达到 363,000 tokens/s（用户速度 210 tokens/s），是 NVIDIA NVL72 GB300（27,400 tokens/s）的 13 倍。Napier 在对数空间中运算，将乘法转为加法，从而降低芯片面积与功耗，更多晶体管用于 SRAM，每 token 能耗更低、推理密度更高。此举改变 AI 推理经济学，不再单纯比拼 FLOPS，而是转向功率、内存局部性、互连延迟与 token 服务成本。

elvis@omarsar0 · 6月18日60

Highly-recommended reading! After using /loops & /goal throughout my projects, I believe that verifiers and robust guardrails are imperative to get current/future coding agents to work right. You can't just YOLO your way with blind autonomous loops. It doesn't work!

译Rahul 指出，fable+ 类模型本质是英语→代码解释器，Fable 5 最差。diff 大小按风险管控：高风险区（身份/数据/网络/资金）用小 diff，可经验验证代码用大 diff。软件交付速度取决于审查/合并能力而非 PR 生成，瓶颈在 lint、测试、CI、影子验证。智能体需深度理解全栈，风险优先级为安全>正确性>性能。复杂性成本变化，可能值得多维护 50% 代码换 5% 性能提升。低风险时将代码块作黑盒仅做经验验证。逻辑逐行审查成本高，只用于关键处。更快迭代需借助权限 opt-in、影子模式等护栏。

Chubby♨️@kimmonismus · 6月17日45

1/ .@TensordyneInc just dropped their Napier gen. Claiming 10x better efficiency than the disaggregation approaches (Nvidia/AWS + Groq/Cerebras), while beating them in user speed. New silicon already taped out. Fundamental innovations in math and chip-to-chip scale up interconnect. a thread 🧵:

译1/ .@TensordyneInc 刚刚发布了他们的 Napier gen. 声称效率比分解方法（Nvidia/AWS + Groq/Cerebras）好10倍，同时在用户速度上超越它们。 新芯片已流片。在数学和芯片间扩展互连方面有根本性创新。 一个帖子 🧵:

OpenCode@opencode · 6月17日26

GLM-5.2 is getting 3x the demand If you are experiencing some problems, we are actively working on it

译GLM-5.2 的需求量正在增长 3 倍 如果你遇到一些问题，我们正在积极处理中

OpenRouter@OpenRouter · 6月17日63

New lab: Cost Simulator A simple tool for looking at your real-world traffic and estimating cost savings with different models, using median endpoint pricing

译新实验室：成本模拟器 一个简单的工具，用于查看你的真实流量并估算使用不同模型时的成本节省，使用中位数端点定价。

Andrew Milich@milichab · 6月17日47

Try Grok models on Amazon Bedrock

译AWS 开发者现可使用 Grok 4.3 构建应用，该模型在幻觉率和工具调用方面行业领先，由 Bedrock 的安全推理引擎驱动。

Baidu Inc.@Baidu_Inc · 6月17日26

Beach. Brunch. Boardwalk. Repeat. Apollo Go now covers the full Jumeirah area, connecting riders across beaches, hotels, malls, offices, and some of Dubai's most popular destinations. If Jumeirah is already on your Dubai itinerary, why not make the journey part of the experience? Book an Apollo Go ride and see what the future of travel feels like.

译海滩。早午餐。海滨步道。循环。 Apollo Go现已覆盖整个Jumeirah区域，连接海滩、酒店、商场、办公室以及迪拜一些最热门的目的地。 如果Jumeirah已经在你的迪拜行程中，何不让旅途本身也成为体验的一部分？ 预约一趟Apollo Go，感受未来出行的模样。

Google DeepMind@GoogleDeepMind · 6月17日31

We’re working with @SciTechgovuk, @mhclg and @i_dot_ai on a new AI housing application planning prototype. 🏡 By cutting down the time spent on repetitive tasks, it could help planning officers focus their attention on complex projects and reduce processing times by up to 50%. → https://goo.gle/4xzqMDs

译我们正与 @SciTechgovuk、@mhclg 和 @i_dot_ai 合作开发新的AI住房申请规划原型。🏡 通过减少在重复性任务上的时间，它可以帮助规划官员将注意力集中在复杂项目上，并将处理时间缩短最多50%。→ https://goo.gle/4xzqMDs

AYi@AYi_AInotes · 6月17日55

Anthropic居然放下身段了，Claude主动适配了苹果的接口标准！ 以前全行业都要适配它的API标准， 现在主动给苹果做了Swift包，老老实实穿上了人家的接口衣服🤣 全网都在调侃Anthropic终于向苹果低头了， 但其实这件事跟认怂半毛钱关系都没有，更核心的变化藏在水面底下，这背后的行业变化，比模型跑分有意思多了！ 大家看完这份文档就知道，Anthropic这帮人其实精得很， Anthropic素来以桀骜著称，安全标准API设计全是自己的一套规矩，从来都是别人适配它的接口， 这次却主动做了Swift包，老老实实穿上了苹果定义的LanguageModel协议。 我们算一笔账就知道了， 现在苹果生态躺着几千万开发者，想让Claude真正跑进每台iPhone和Mac里，就必须按人家的规矩来。 不得不说苹果这步棋走得极优雅也极狠， 它不跟任何人拼参数比跑分，就做了一套统一接口，开发者同一套代码，就能在本地模型Claude和Gemini之间随便切换， 表面是开放兼容，实则是把自己变成了AI模型的分发平台，所有想来这个生态分蛋糕的玩家，都得遵守它定的规矩。 这意味着AI行业的竞争已经悄悄换了赛道， 以前拼的是谁的模型能力更强，现在拼的是谁的规则能被更多人接受， 因为定义标准的人，永远比做产品的人站得更稳啊， 所以得认清现实，在苹果的地盘上，苹果就是话事人，认怂比硬扛划算得多，我觉得这也是Anthropic这家年轻公司成熟的一个表现 https://platform.claude.com/docs/en/cli-sdks-libraries/libraries/apple-foundation-models

译Anthropic发布了Swift包，使Claude模型适配苹果定义的LanguageModel协议，从而能在iPhone和Mac本地运行。这一转变打破了Anthropic一向要求别人适配其API的惯例。背后是苹果通过统一接口把自己变成AI模型分发平台，开发者用同一套代码可在本地模型和Gemini等之间切换。推文认为行业竞争已从模型能力转向规则标准，在苹果生态内顺应其规矩比硬扛更明智。

Alibaba Cloud@alibaba_cloud · 6月17日49

🇫🇷 Alibaba Cloud launches its new cloud region in France, featuring two availability zones in Paris. As our third infrastructure hub in Europe, alongside Germany and the UK, the new region helps businesses scale innovation with locally hosted cloud services—and paves the way for upcoming Agentic AI services. 🔗: https://int.alibabacloud.com/m/1000405058/ #AlibabaCloud #France #AI #CloudComputing #AgenticAI

译🇫🇷 阿里云在法国推出新云区域，设有巴黎双可用区。 作为我们在欧洲（继德国和英国之后）的第三个基础设施中心，新区域通过本地托管的云服务帮助企业扩展创新，并为即将推出的智能体AI服务铺平道路。 🔗: https://int.alibabacloud.com/m/1000405058/ #AlibabaCloud #France #AI #CloudComputing #AgenticAI

Artificial Analysis@ArtificialAnlys · 6月17日52

Following up on our Intelligence Index v4.1 release yesterday, in the video below, Daniel from our team shares a short overview of what's changed: 1. Three upgraded evaluations: Terminal-Bench 2.1, τ³-Bench Banking and GDPval-AA v2 2. Cost, time, and tokens per task: Understand the cost, time, and tokens of tasks across our Index and for individual evals, and how these trade off against Intelligence 3. Cached input token reporting: We now report the amount of cached tokens a particular model uses and how this influences cost

译Artificial Analysis 昨日发布 Intelligence Index v4.1 更新，主要变化有三项：升级的评测 Terminal-Bench 2.1、τ³-Bench Banking 和 GDPval-AA v2；提供每项任务的成本、时间与模型 token 消耗数据，并展示这些指标与智能水平的权衡；新增缓存输入 token 报告，显示特定模型使用的缓存 token 量及其对成本的影响。

SenseTime@SenseTime_AI · 6月17日45

Proud to join forces with the 𝗛𝗼𝗻𝗴 𝗞𝗼𝗻𝗴 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗧𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝘆 𝗣𝗮𝗿𝗸𝘀 𝗖𝗼𝗿𝗽𝗼𝗿𝗮𝘁𝗶𝗼𝗻 (#𝗛𝗞𝗦𝗧𝗣) to build 𝗛𝗼𝗻𝗴 𝗞𝗼𝗻𝗴'𝘀 𝗹𝗮𝗿𝗴𝗲𝘀𝘁 𝗱𝗼𝗺𝗲𝘀𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗱 #𝗔𝗜 𝗗𝗮𝘁𝗮 𝗖𝗲𝗻𝘁𝗿𝗲 — a milestone for the city's AI future. We signed a #𝗠𝗼𝗨 to co-develop an AIDC with a target computing capacity of 𝟰𝟬,𝟬𝟬𝟬𝗣+ (𝗣𝗲𝘁𝗮𝗙𝗟𝗢𝗣𝗦) 𝗯𝘆 𝟮𝟬𝟯𝟬, rolling out in phases. The centre will serve local, cross-border, and international clients across a full range of AI needs — from 𝗺𝗼𝗱𝗲𝗹 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴, 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 to 𝗹𝗮𝗿𝗴𝗲-𝘀𝗰𝗮𝗹𝗲 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁. Drawing on the proven experience from our #𝗔𝗜𝗗𝗖 in Shanghai, which is one of the largest AIDC in Asia, we will bring to Hong Kong 𝗺𝗮𝘁𝘂𝗿𝗲 𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗲𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲, 𝗱𝗼𝗺𝗲𝘀𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗱 𝗚𝗣𝗨 𝘁𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝘆, 𝗵𝗶𝗴𝗵-𝘀𝗽𝗲𝗲𝗱 𝗼𝗽𝘁𝗶𝗰𝗮𝗹 𝗶𝗻𝘁𝗲𝗿𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝗼𝗻, and 𝗿𝗲𝗻𝗲𝘄𝗮𝗯𝗹𝗲 𝗲𝗻𝗲𝗿𝗴𝘆-𝘀𝗮𝘃𝗶𝗻𝗴 𝗮𝗻𝗱 𝗲𝗻𝗲𝗿𝗴𝘆-𝘀𝘁𝗼𝗿𝗮𝗴𝗲 𝘁𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝗶𝗲𝘀. For SenseTime, Hong Kong is our headquarters and a key R&D base — and, this AIDC is our commitment to keeping the city at the forefront of global AI. With roots at Hong Kong Science Park, we are honoured to write this next chapter with HKSTP — leveraging its advanced infrastructure, global networks, and beyond.

译商汤与香港科技园签署谅解备忘录，合作建设香港最大规模本地自研AI数据中心，目标计算容量40,000 PetaFLOPS+（2030年达成），分阶段投产。该数据中心将服务本地、跨境及国际客户的模型训练、推理与大规模应用部署。商汤基于上海AIDC经验，带来成熟运营能力、自研GPU技术、高速光互连及可再生能源与储能技术。香港作为商汤总部与关键研发基地，此举旨在巩固其全球AI前沿地位。

Rohan Paul@rohanpaul_ai · 6月17日55

Tensordyne just announced a breakthrough Inference system. Logarithmic AI compute chips which is 17x more tokens per watt and 13x higher throughput than NVIDIA Blackwell. The main math advance they say they unlocked is efficient logarithmic math directly in hardware. In log space, multiplication turns into addition, which is much easier to build than multiplier circuits That allows smaller compute circuits on the chip than today’s FP8 and INT8 GPUs.With fewer transistors, the chips stay cooler and use less energy, while the extra die space can hold more tensor engines, additional high-bandwidth SRAM and HBM3e memory, plus a fast interconnect fabric. For DeepSeek-R1, Tensordyne claims 363K tokens/sec per rack versus 27.4K for Nvidia’s comparison system They have successfully completed tape-out of the Napier processor, which is now in production at TSMC on its 3nm process node.

译Tensordyne 发布突破性推理系统，采用对数 AI 计算芯片。相比 NVIDIA Blackwell，每瓦特 token 数提升 17 倍，吞吐量提升 13 倍。核心创新是在硬件中实现高效对数运算，将乘法转为加法，从而缩小计算电路、减少晶体管、降低功耗，释放芯片空间用于更多张量引擎、高带宽 SRAM 和 HBM3e 内存。针对 DeepSeek-R1，单机架可达 363K tokens/sec，对照系统仅 27.4K。Napier 处理器已完成流片，在台积电 3nm 制程生产。

Rohan Paul@rohanpaul_ai · 6月17日46

TokenPilot reduces LLM agent costs via ingestion-aware compaction and lifecycle-aware eviction. Achieves 61–87% cost reduction on PinchBench and Claw-Eval with competitive scores. Argues that cheaper AI agents need stable memory, not just shorter prompts. Older methods usually cut or summarize the history, but that can shift the text around and break the prompt cache, which is the system that reuses unchanged prompt text to save money. TokenPilot tries to fix both sides at once by cleaning new tool results before they enter the context and by keeping the early prompt layout stable across tasks. It also waits before deleting old task history, because finished work can still help later tasks that refer to the same files or goals. ---- Link – arxiv. org/abs/2606.17016v1 Title: "TokenPilot: Cache-Efficient Context Management for LLM Agents"

译TokenPilot 提出一种针对 LLM 智能体的缓存高效上下文管理方法，通过摄入感知压缩和生命周期感知驱逐两大机制，在 PinchBench 和 Claw-Eval 基准上实现 61–87% 的成本降低，同时保持有竞争力的分数。传统方法通常直接截断或摘要历史，容易导致文本偏移、破坏 prompt 缓存。TokenPilot 在工具结果进入上下文前进行清理，保持早期提示布局稳定；同时延迟删除旧任务历史，因为已完成的工作仍可能为引用相同文件或目标的后续任务提供帮助。

Epoch AI@EpochAIResearch · 6月17日53

The end of the self-funded AI buildout? Hyperscaler cash capex is growing much faster than cash inflows. On current trends, they will be unable to fully fund the AI infrastructure buildout with cash from operations by the end of this year.

译自筹资金的AI建设是否走向终结？ 超大规模企业的现金资本支出增速远超现金流入。按当前趋势，到今年年底，他们将无法完全依靠运营现金流来为AI基础设施建设提供资金。

SemiAnalysis@SemiAnalysis_ · 6月17日51

RL Systems Mind the Gap: Matching Trainer and Generator Throughput RL Training Infrastructure, GRPO, PipelineRL, Async RL, Policy Staleness, RL Sandbox Infra, CPU Requirements, TCO Analysis, Thinking Machines Tinker https://newsletter.semianalysis.com/p/rl-systems-mind-the-gap-matching

译RL系统 注意差距：匹配训练器与生成器吞吐量 RL训练基础设施，GRPO， PipelineRL，异步RL，策略陈旧性， RL沙箱基础设施，CPU需求， TCO分析，思考机器修补