智享 | 蓝鸟会🕊️

不是？现在推特涨粉的底层逻辑已经变成这样了吗？ 观察了一个几天破千粉的号，发现个残酷的真相： ❌ 辛辛苦苦做价值输出：3个月涨粉800 ✅ 靠老号互fo+骗fo：几天粉丝破千，直接坐等领老马低保 合着我特意充了新蓝V，反而被平台算法给“优化”了？ 表面上看是内容为王，本质上其实是： 权重号（老号）自带天然流量光环 互fo裂变的速度，远快于靠内容死磕 差的仅仅是一个曝光机会 这波操作直接把老实做内容的独立创作者看傻了。

The high cost of living in New York is essentially a "top resource". Admission fee . That kind of "one who can run into top-notch connections and enjoy world-class convenience as soon as they step out of their home" "Density" It is indeed something that other cities find hard to replicate. Once you have lived here, it is hard not to be attracted by this ultimate vitality.

Been living in New York City for around five years now. Yes, it’s expensive, taxes are high, and there are some interesting characters walking around. Those are some of the costs. Here’s what we get: Access to the best restaurants in America, no matter what type of food you’re in the mood for. Everything our family needs is just a short walk away. You constantly get to see friends in person, as they’re always passing through. Some of the best public and private schools in America. The network you build here, just by going about your day-to-day life, is incredible. You run into some of the most interesting people doing the most amazing things at the highest level. Access to the best doctors in the world. The career opportunities here are immense, no matter what you do. Central Park – my go-to spot – never gets old. If you’re a shopper, there’s nowhere better in America. If you’re an entrepreneur, this city forces you to think bigger on a daily basis. Broadway, sports, concerts, comedy – the highest level of entertainment, right in your backyard. The subway. Yes, the subway – and yes, millions of normal people take it every day – gets you around this place like a time machine. It’s a wonderful place to raise kids. Every kids' activity you can think of is just blocks away. Our son loves the Natural History Museum, and endless playdates are available either in our building or within a three-minute walk. Maybe all those folks who can afford to live anywhere in the world but choose to raise their families here aren’t so crazy after all.

For ordinary investors, the most attractive aspect of this SpaceX IPO is that retail investors can enter the market at the same IPO price as institutional investors, which is very rare in previous large-scale ipos. However, the entry requirements of various securities firms do vary greatly: Fidelity requires $500,000 in assets, Schwab requires $100,000, while Robinhood and SoFi have zero entry requirements, which gives more people the opportunity to witness history. It can be foreseen that the competition for subscription will be extremely fierce, and everyone is submitting ";" "Intention confirmation" Also, be mentally prepared for the possibility of not getting a quota at all or having a very small quota. This is definitely one of the most anticipated events in the global financial market this year!

I’ve been asked by many to create one comprehensive post explaining how to prepare for @SpaceX’s IPO if you use one of the brokerages listed in SpaceX’s S-1 filing to allocate IPO shares to retail investors. Here it is: Fidelity: 1) $500,000 minimum account balance required to participate (including IRAs, individual, etc, but excluding 401k). 2) Enter an indication of interest. The indication of interest provides Fidelity with the maximum number of shares a customer is interested in purchasing. 3) Confirm your indication of interest shares on Fidelity's website after the registration statement has been declared effective and the offering has priced, which is typically after 7 PM ET on the night of pricing. Indications of interest may not be confirmed prior to the registration statement being declared effective and the offering pricing established. By confirming your indication of interest, you are placing an order to buy shares at the offering price. If you do not confirm your indication of interest, you will not be eligible for an allocation of shares. 4) Allocation of shares will occur on the morning following pricing and is usually complete before 9:30 AM ET. An alert will be sent once allocations are complete, and you can check your account to determine whether you were allocated shares. If you receive an allocation of shares, you must have adequate funds available to settle the purchase in the settlement date which is typically the trade date plus one business day. 5) You may increase your indication of interest up through the close of the indication of interest period. You may decrease or cancel an indication of interest until share allocation takes place. Once share allocation takes place, your indication may not be canceled or modified. Charles Schwab: 1) $100,000 minimum account balance required to participate (including IRAs, individual, etc, but excluding 401k). 2) On Schwab's website, under the Trade tab, select the IPO page to view the Calendar of Offerings, a list of upcoming IPOs. Once the IPO offering window opens (expected first week of June), investors will have the ability to submit a Conditional Offer to Purchase (COTP), also known as an Indication of Interest, from this page. 3) During an IPO's open COTP window, select Start COTP to review offering details and the preliminary prospectus. Then select the green button to proceed to the Eligibility Questionnaire, which is required to confirm investors meet eligibility criteria and are not restricted (per FINRA rules) from participating. After completing the questionnaire, you'll be able to indicate how many shares you're interested in purchasing based on the price range provided. Select Confirm to submit the COTP. 4) After the COTP has been submitted, regularly monitor the IPO page, which will indicate the Status of Your Conditional Offers to Purchase (COTPs), the expected pricing date, and current pricing status, plus any changes in the prospectus. When the IPO has been priced, you will affirm your COTP. You must affirm your COTP once the effective price is established in order to be eligible to purchase shares. To do so, select Affirm Now to review and finalize the share quantity. Robinhood: 1) There's no minimum account size requirement, but you must have enough buying power to cover your requested shares if you are allocated any. You must have an individual brokerage account. Retirement, custodial, and multiple investing accounts are not eligible for IPO Access. 2) Make sure IPO Access is enabled in your Robinhood app. Turn on your IPO notifications so that Robinhood notifies you when the SpaceX IPO comes online. 3) Request Shares: Once the IPO is announced and available, you can request shares through the app or website. This is a request for IPO shares. By placing a conditional offer to buy (COB), you’re asking for the opportunity to purchase a quantity of shares at the IPO price. An investor may place, edit, or cancel a COB after the initial price range is published and before the confirmation period ends. 4) Allocation is random and not guaranteed. The number of shares you request factors into how many you actually get, but it doesn’t affect the likelihood that you’ll get any allocation. You may get all, some, or none of the IPO shares you request. E*Trade: 1) E*TRADE does not publicly list a specific minimum account size required to participate in IPOs, but contact them to double check. That said, allocation priority for “hot” IPOs may still favor larger or more active accounts in practice, even if there’s no official minimum balance requirement. 2) Be a U.S. resident, have an active E*TRADE account (Individual, Joint and IRAs are all eligible) and complete the investor profile questionnaire. 3) Sign up for IPO alerts. 4) Submit a conditional offer to buy ("COB"). As part of this submission, you specify the number of shares and the maximum price you are willing to pay per share. COBs can only be submitted via the New Issue Center. A COB may be submitted once an offering is listed as "open" up until the status is changed to "closed." COBs that have already been submitted may be amended or cancelled after an offering is "closed" up until the status is changed to "allocate." At this point, no further changes may be made to a COB and you are bound by the terms of your COB. If there is no material change in an offering, customers will not need to reconfirm their COBs. If you have submitted a conditional offer, you must have available buying power to cover the full amount of your conditional offer in the account through which you submitted the conditional offer. 5) Shares are allocated to eligible accounts as a proportion, or percentage, of the size of their COB. The percentage is based primarily on the number of shares provided to E*TRADE for sale to its customers and the size of the overall demand for shares from E*TRADE's customers. Given the expected high demand for this offering and the limited availability of shares available for sale to E*TRADE customers, many COBs may not be allocated shares (according to E*Trade). Additionally, in many instances, allocations will be significantly smaller than the size of shares requested in a customer's COB. 6) E*TRADE makes its allocations after the pricing of the overall offering but before the stock begins trading. E*TRADE will inform customers via alert or email whether they have been allocated shares. Any allocation should be reflected in the relevant customer account once that allocation has been processed by E*TRADE. Sofi: 1) There is no minimum account balance/size requirement. Have an active Self-Directed Invest account. 2) Go to the “IPO Investing” section in the app or website 3) Select the IPO 4) Complete the IPO suitability questionnaire 5) Submit an “Indication of Interest” (IOI), which is basically a non-binding request for shares. 6) When the IPO is officially priced, SoFi will notify you to confirm your order. NOTE: Don’t be surprised if you receive fewer IPO shares than you requested, or none at all. Demand for the limited number of IPO shares available to retail investors will likely be extremely high, and each participating brokerage will only receive a limited allocation of shares to distribute to retail investors. For our international friends, keep in mind that @SpaceX said in their S-1 filing that allocations will also be made to retail investors by the underwriters, which include: • Goldman Sachs • Morgan Stanley • Bank of America • Citigroup • J.P. Morgan • Barclays • Deutsche Bank Securities • RBC Capital Markets • UBS Investment Bank • Wells Fargo Securities • Allen & Company • Cantor • Needham & Company • Raymond James • Societe Generale • Stifel • William Blair • BTG Pactual • ING • Macquarie Capital • Mirae Asset Securities • Mizuho • Santander so you can try reaching out to one of these places if you have assets with them and you may be able to request an allocation of some shares. I've already seen that happen with some Goldman Sachs clients. Lastly, and I stated this in a previous post, @SpaceX specifically stated in their S-1 filing that any purchase of their Class A common stock in this offering through these platforms will be at the same IPO price, and at the same time, as any other purchases in this offering, including purchases by institutions and other large investors, which means any retail investors that are lucky enough to get allocated some SpaceX IPO shares will pay the same price as the big guys. This will likely be the largest retail IPO share allocation in history, by far. If you have more questions, reach out directly to your brokerage and/or bank. And no, this post wasn't written by AI lol. Not financial advice.

@0xNnami 用起来顺手多了，以前手动操作浪费好多时间，现在效率直接起飞👍

@CycleDecoded 这工具确实方便了不少

熬夜爆肝做PPT的时代结束了。 最近GitHub上一个叫 html-ppt-skill 的开源项目彻底火了。你只要给它一句话，比如“做一份AI趋势分享，小红书风格”，剩下的内容大纲、智能排版、动态特效它全全包办。 【2/🧵】 最让我震惊的是它的视觉表现，完全不是以前那种死板的PPT： 36套高质感主题 + 31种专业布局 自带粒子动画与动态转场，做出来直接就是苹果发布会的高级感 支持一键导出图片，公众号/小红书/朋友圈一条龙传播 【3/🧵】 AI正在把复杂的事情变成一句话的门槛。未来拉开差距的不是谁能熬夜排版，而是谁能用AI高效产出。 尝鲜命令甩这了，独立开发者和职场人必存👇 npx skills add [github.com/lewislulu/html…](github.com/lewislulu/html…)

@jukan05 I’m really looking forward to the new Kirin chip this fall! If Huawei can achieve a breakthrough and close the gap with TSMC without relying on the most advanced equipment, this "new path" is definitely worth watching. I hope to see a genuine leap in performance.

HUAWEI SAYS IT HAS COME UP WITH A NEW PATHWAY TO SHORTEN ITS GAP WITH INDUSTRY LEADER TSMC, POTENTIALLY ACHIEVING A BREAKTHROUGH IN MAKING ADVANCED SEMICONDUCTORS WITHOUT CUTTING-EDGE EQUIPMENT (PEOPLE’S DAILY) HUAWEI: THE KIRIN SMARTPHONE CHIP TO BE LAUNCHED THIS FALL WILL SEE A SIGNIFICANT PERFORMANCE IMPROVEMENT. (PEOPLE’S DAILY)

找不到好用的 Agent 上网插件？ 我把 browse.sh 接进 Hermes 测了一下，彻底傻眼。 一行 CLI 命令，直接给你的 Agent 换上“24小时冲浪脑”： 👉 自动化找房/订机票/刷工作机会 👉 24h监控 Polymarket 行情与早期机会 👉 把所有“人肉爬信息”的破事，全拆成一调就用的 Skill 搞AI和自动化的兄弟，现在不码住，明天想用时哭都来不及！

Indeed. When the scientific research paradigm shifts from" "Experiment-driven" "Turn to";" "Reasoning driven";" Traditional academic accumulation may be rapidly reconstructed in the face of violent computing. The key now is no longer how much knowledge you have, but whether your field is within the range of AI evolution.

academics are unprepared for the coming world where much scientific progress is majorly a function of inference compute. whether OpenAI points the Eye of Stargate at your particular field will decide its acceleration. talent will leach away into the labs. it's already begun

@RocketLab Witnessing the leap from "It's a Test" to high-frequency launches, Electron has truly redefined small satellite delivery. Looking forward to the 100th launch this year. Congratulations, Rocket Lab! 🚀🚀

Nine years ago today, we launched Electron for the first time. Later this year, Electron is expected to reach 100 launches, likely making it the fastest privately developed orbital rocket to hit that milestone. From “It’s a test” to one of the world’s highest cadence launch vehicles in under a decade. Here's a little walk down memory lane from that first historic launch.

Indeed. In the AI era, the "signal-to-noise ratio" of code is productivity. The conciseness of dynamic languages, paired with comprehensive test suites, not only ensures robust behavior but also significantly enhances inference efficiency by reducing token consumption. This is the inherent advantage that Ruby—a language dedicated to extreme simplicity—demonstrates in the AI wave.

Agents don't need types. They're perfectly capable of pulling off incredible refactorings without. Give them a linter and a test suite, and you have all you need. Token efficiency is where it's at.

The IBKR Performance report is indeed calculated using time-weighted return (TWR), eliminating the impact of fund inflows and outflows on percentage figures—a technically sound approach. However, a 38-fold return implies remarkably high leverage and risk exposure. While these numbers are impressive, it's crucial to recognize that such "god-of-war" performance typically results from extreme leverage combined with luck, representing the pinnacle of survivor bias. For ordinary investors, such returns offer limited reference value and are not replicable.

我不懂为什么很多人质疑 Serenity 的账户表现。可以看到他使用的是 IBKR，他今年的回报是 3800%，也就是盈利 37 倍多。而且图中显示的是他的 performance（业绩）而不是 value（价值）： 1. 如果是 value 的话，可以通过不断存入本金来提高百分比； 2. 但如果是 performance，存入本金并不会产生影响，因为这是一个加权平均的收益。 不管你喜不喜欢，我们要尊重数字。

很多城市有人拿个喇叭回收电脑 回收手机 回收冰箱 回收这么赚钱吗

@ArtofSpecuycky 这篇不是在讲 AI，是在讲“谁卡住水电网，谁就收过路费”啊 真·卖铲子 2.0，黄仁勋一句“能源永远稀缺”，把后面一串都点名了哈

AI 基建真正的大行情，可能才刚刚开始 纵观整个 AI 产业链，我认为市场依然没有 fully price in 的还有很多板块和个股。 当 Hyperscaler 真正开始一年几千亿 CapEx 往下砸后，AI 背后对应的是一整套庞大的基础设施体系：算力、内存、网络、光通信、电力、储能、冷却、数据中心。 未来几年，很多真正拥有收费站属性的底层基础设施公司，可能才会迎来持续性的估值重构。 算力层大家都看过了。GPU、HBM、CoWoS，第一阶段的故事相对成熟。但有几个方向，市场还没有充分定价。 第一个：Ethernet 和光通信 $ANET 之前被错杀，其实有两个原因。 第一个是市场误以为NVLink 和 InfiniBand 会替代 Ethernet。但这混淆了Scale-Up 和 Scale-Out。 NVLink 解决的是机柜内部互联。真正大规模 AI 集群最终拼的是：数万张 GPU 之间的 Scale-Out 网络。 AI 越大，Ethernet 的地位反而越强。 第二个则是短期财务层面的担忧。 随着 1.6T、AI switching 和新一代高速产品开始 ramp， 市场担心：早期 deployment 成本、供应链压力以及 optics BOM 成本上升，会导致毛利率短期承压。但很多人忽略了一点AI 网络现在正从可选配置变成核心瓶颈。当 AI 集群进入十万卡级别后，真正决定效率的是 GPU 之间的数据流动效率。 UEC 对 InfiniBand 的挑战，市场还没有开始充分定价。随着 Tail Latency、LLR、CBFC 等能力逐渐成熟，以太网第一次开始真正接近 InfiniBand 的性能，但成本更低、生态更开放。 而 1.6T 交换机和 Silicon Photonics，可能才刚进入超级周期前夜。 $ANET、 $MRVL、 $CRDO、 $ALAB、 $LITE 本质上都属于 AI 数据流动层的收费站。 第二个：AI-RAN和边缘AI基础设施（这个还会专门写一篇单独讲） 这个方向市场理解度还非常低。 黄仁勋已经把话说得很清楚：AI不会只待在hyperscaler的数据中心里。它会下沉到运营商网络、下沉到边缘节点、下沉到基站。 未来全球几百万个基站，可能都会变成边缘AI节点。真正的受益者不只是GPU，而是整个AI-RAN基础设施层。 Nokia是AI-RAN联盟的核心发起人，英伟达直接投资了Nokia。T-Mobile想在全美试点AI-RAN基站，必须在诺基亚的通信机柜上进行升级。这个物理底座的位置，别人很难绕过去。 $NOK 第四个：能源——黄仁勋说能源永远稀缺 这句话不是随口说的。 一台GB200满载机柜120-140kW，传统数据中心整排柜子才10kW。AI集群对电网的瞬时冲击极其剧烈，很多地方变压器排队2-3年，高压设备缺货，并网审批堵塞。 $CEG 和 $VST是目前最直接的受益者，目前经历过回调后，估值相对于ai基建其他板块就变得相对便宜了。核电提供稳定的基础负载，正好匹配AI数据中心7x24小时不间断的用电需求。这两家已经和多个hyperscaler签署了长期供电协议。这不是新能源叙事，是AI mandatory infrastructure。 第五个：储能 市场现在还把 $BE, $FLNC和 $EOSE 当新能源储能公司看。但未来真正的需求来源可能是AI数据中心的电网调频和负载平衡。未来hyperscaler很可能自己建Gas Turbine加Battery Storage加Microgrid。储能正在从新能源配件变成AI电网稳定层的核心组件。这个估值体系的重写还没有发生。 第六个：数据中心本身 市场现在还把很多IDC公司当传统机房出租来看。 但AI数据中心已经完全不是以前那套逻辑。未来真正值钱的是有没有电力、并网能力、冷却系统、网络带宽和可交付的GPU capacity。数据中心正在逐渐变成AI工业基础设施资产。 $ORCL 已经不只是数据库公司。OCI越来越像AI时代的算力地主。过去市场看云，看的是SaaS增长、云份额、企业客户。但AI时代真正稀缺的开始变成可交付的算力capacity。Oracle已经和多个hyperscaler、AI公司以及主权国家签署长期AI算力协议，订单积压规模远超市场之前的预期。市场目前依然更习惯用传统软件公司的框架给ORCL定价，但OCI的资产属性可能更接近AI基础设施运营商。 $CRWV 是类似逻辑。CoreWeave本质上已经不只是云服务商，更像AI算力shortage时代的租赁平台。如果AI需求持续超过供给，真正拥有GPU capacity的公司会天然拥有更强的议价权。谁手里有capacity，谁就能定价。 $NBIS 是另一个市场关注的方向。未来AI真正的大规模商业化一定会走向推理层。NBIS卡的位置恰好在模型部署、推理优化和实际落地之间。很多人还在看谁在训练模型，但未来持续消耗算力的可能是Inference。 $IREN 的核心逻辑早已不只是挖矿。真正值钱的是低成本电力加HPC基础设施加长期能源优势。目前IREN已经拥有约1.4GW的确定性电力资源，与英伟达签署了长期GPU cloud服务协议，英伟达还通过认股权证进行了深度绑定。这已经不只是合作，更像是用真金白银给AI电力资产定价。 未来AI数据中心最值钱的资产是便宜电。谁手里有稳定、低成本、可持续扩张的电力资源，谁就更有可能拥有AI基础设施时代的定价权。 总结 黄仁勋已经把答案说出来了：能源永远稀缺，AI会扩散到每一个边缘节点。真正绕不过去的，是网络、电力、储能、光通信、边缘基础设施这些底层物理瓶颈。最容易持续扩估值的，是那些卡住物理瓶颈、别人绕不过去的收费站。 #AI #人工智能 #AIInfrastructure #数据中心 #光通信 #核电 #储能 #边缘AI

@CryptoMaster_70 这波太离谱了啊，CEO直接来一口“我先替你尝尝” 夜市烤玉米：我到底招谁惹谁了哈

世界第一CEO搶台灣路人的烤玉米吃 剛剛真實發生 非AI 台灣女網友發文： 在夜市遇到黃仁勳不稀奇 被湊過來問好不好吃，然後手上這隻就⋯ No 我的烤玉米～～～ 真的太有趣太興奮了

@AntonLaVay 以前看MLCC像看路边小零件，现在看像在看AI服务器的氧气管啊。 电压往上走，电容往芯片边上挤，结构性紧缺这味儿太冲了。

为什么 MLCC 又重要了？ 本文专注于三个问题，大家各取所需： 1. 为什么现在MLCC变得重要了？ 2. 为什么是高端MLCC？ 3. 为什么本次更像是结构性短缺而非补库存周期。 请注意，本文的逻辑您可以直接复制给你们的AI，AI会告诉你基于本文描述的情况还能找到哪些其他的产业，或是在中国A股有什么标的。 本文不赘述此处，但是欢迎大家评论区留言讨论。 觉得大家有点价值，欢迎大家画一刀点个订阅。 ---------TL:DR--------- 1. 为什么现在MLCC变得重要了？ 过去看MLCC，会把它当成一个手机、PC、汽车电子周期品。 手机出货好，MLCC好；消费电子差，MLCC差。这个理解不能说错，但在AI服务器时代，它已经不够用了。 因为AI数据中心正在把MLCC从一个“普通被动元件”，重新推回到一个非常关键的位置：Power Delivery Network，也就是供电网络。 AI服务器的核心问题，不只是GPU够不够多，HBM够不够快，光模块够不够密。还有一个更底层、更物理的问题： 这么大的电流，如何稳定、低损耗、快速响应地送到GPU/ASIC核心？这就是MLCC重新变得重要的原因。 现在的数据中心供电架构正在发生变化。传统服务器时代，12V供电已经用了很多年。但AI rack功耗暴涨之后，行业正在往48V/54V，甚至±400VDC/800VDC演进。 Google、Meta、Microsoft推动OCP Diablo 400；NVIDIA也在推800VDC AI factory power stack；TI、Vertiv、ABB、Delta这些公司也都在围绕800VDC架构布局。 但这里有一个容易被误解的点： 高压供电解决的是远距离传输效率，不是芯片核心附近的供电问题。800V也好，48V也好，最终到GPU/ASIC核心，仍然要变成不到1V的核心电压。 而一个1000W级别的AI芯片，如果核心电压约1V，意味着它附近要处理的不是几十安培，而是数百到上千安培的瞬态电流。 这才是真正可怕的地方。 AI芯片不是一个稳定耗电的灯泡。它的负载会快速跳变。某个计算任务起来，电流需求瞬间拉高；电源网络如果响应不够快，电压就会下陷，也就是voltage droop。droop太大，轻则降频，重则错误、宕机、可靠性下降。 所以越靠近GPU/ASIC，越需要大量电容作为局部电荷缓冲，压低PDN阻抗，抑制噪声和电压波动。 这就是MLCC在AI服务器里的真实作用。 它不是“板子上随便贴一堆小电容”。它是在帮GPU/ASIC维持高速运行时的供电稳定性。 2. 为什么是高端MLCC？ 但这里必须强调：真正重要的不是所有MLCC，而是高端MLCC。 为什么？ 因为AI服务器需要的不是普通消费级规格。它要的是：高容量、小尺寸、低ESL、低高度、高可靠、高耐压、耐高温，甚至要能放在package附近、land-side、die-side，或者参与嵌入式PDN设计。 普通MLCC解决不了这个问题。因为在高频场景下，电容不是只看容量。ESL，也就是等效串联电感，会变得非常关键。ESL太高，电容在高频下就不像电容，反而会失去去耦效果。 所以AI服务器真正需要的是低ESL、短电流路径、大电流截面积、能贴近芯片的MLCC。 这就是为什么村田在AI服务器供电指南里，不是泛泛而谈“MLCC需求增加”，而是专门讲die-side、land-side、低ESL、低高度、小型高容量，以及PDN仿真和元件摆放。 这背后的意思是：高端MLCC已经不只是材料问题，而是供电架构问题。这也解释了为什么这轮更像“结构性短缺”，而不是普通周期补库存。 3. 为什么本次更像是结构性短缺而非补库存周期？ 普通MLCC并不一定短缺。手机、PC、一般消费电子需求并不强，很多标准规格并没有进入全面紧缺。 但AI服务器用的高端MLCC是另一回事。 它受限于几个东西： 第一，需求增长不是单纯来自AI服务器数量增加，而是每块AI baseboard、每个power module、每个GPU/ASIC附近的电容用量和规格都在上升。 第二，高端MLCC产线不是普通产线随便切一下就能做。小型化、高容量、低ESL、高耐压、高温可靠性，都涉及良率、工艺、材料和测试能力。 第三，AI服务器客户认证周期长。进入GPU/ASIC供电网络的元件，不是今天报价、明天替换。它要和主板、封装、电源模块、热设计、仿真模型一起验证。 第四，头部供应商不太可能为了短期需求疯狂扩普通产能。经历过多轮MLCC周期后 村田 (村田製作所, Murata 6981.jp) 太阳诱电(太陽誘電, Taiyo Yuden 6976.jp) 三星电机 (삼성전기，Samsung Electro-Mechanics 009150.kr) TDK (6762.jp) 这些厂商更倾向于把产能分配给高端、高可靠、高利润规格，而不是重走低端过剩路线。 所以我们看到的可能不是“MLCC全行业普涨”，而是： 低端松，高端紧。消费级松，AI服务器紧。普通规格松，高容量/高耐压/低ESL/低高度规格紧。 这就是结构性短缺。 还有一个问题：硅电容会不会替代MLCC？ 我的理解是，不是简单替代，而是分工。越靠近die、越高频的位置，硅电容会更有价值。它可以进入封装，interposer、die-side附近，处理极高频瞬态。但板级、power module、48V输入输出、land-side、中高频去耦，仍然需要大量高端MLCC。 所以硅电容的出现，并不是否定MLCC逻辑，反而说明同一个趋势： AI芯片附近的电源完整性，正在变成新的价值池。 未来不是某一种电容通吃，而是MLCC、硅电容、聚合物电容、嵌入式电容基板一起分工。 因此，MLCC这条线最重要的判断，不是“会不会像2018年那样全行业大缺货”。 我认为更正确的问题是： AI服务器高端MLCC会不会持续紧？ 我的答案是：大概率会。 因为AI rack功耗还在继续上升，48V/54V只是当前阶段，±400VDC/800VDC是下一阶段，但不管远端电压怎么升，最终芯片核心附近都必须面对低压、大电流、高瞬态、高热密度的问题。 只要这个问题存在，高端MLCC就会继续重要。 短缺也更可能出现在这些方向： 高容量、小尺寸MLCC 低ESL、低高度MLCC land-side / die-side 用MLCC 48V电源系统里的高耐压MLCC 高温、高可靠、服务器级认证规格 能参与PDN仿真和客户协同设计的高端料号 所以这不是简单的“被动元件涨价故事”。 更准确地说： MLCC正在从消费电子周期品的一部分，变成AI基础设施供电网络的一部分。 这也是为什么它值得重新研究。 AI产业链的利润池，不只在GPU、HBM、光模块。 当算力继续堆高，瓶颈会自然扩散到供电、散热、互联、存储这些底层物理环节。 而MLCC这一次站上的，正是“供电完整性”这个位置。 这才是这轮高端MLCC行情最值得重视的地方。

@yuyy614893671 这组合也太会整活了吧，Lisa+Labubu，直接把我看乐了哈哈 又潮又怪，莫名有种“压力先放一边”的松弛感啊

今天刚刚看到这个视频 Black Pink的成员之一Lisa携手中国潮玩 Labubu 一起亮相世界杯MV 好喜欢😍！好解压！ #AI时代活人感

@Balder13946731 这波不是模型差距，是“认真做事”和“随便糊弄”的差距啊。 一个在算题，一个在编剧本，离谱得很。

Mythos>ChatGPT-5.6>Opus 4.8 >>Gemini 3.5 Pro Google连Gemini 3.5 Flash这种货色都敢拿出来，可以说和工作级AI已经差得十万八千里了。 我让Opus 4.7和Gemini 3.5 Flash 分别生成两张图，可以说完全暴露出了这两家最大的鸿沟，就是可靠性。 Opus不仅完全按照指示做出了图表，还检验和计算了所有数据，而Gemini直接伪造了大量数据，生成的图片有严重幻觉.. Opus 4.7是具有自省能力的，他会不断的去检验自己做对了没有，换言之，他可以做一个稳定可靠的接口。但是Gemini从以前到现在，都没有一个self-improve 的inference loop，换言之Google对待AI依然像是在生产玩具的态度，他的主营业务依然是云。 之前觉得Google的TPU走在了前面，但是其实 $AMZN Titanium 可能已经完全追上了，而Deepmind做的模型水平现在更接近Meta的Muse Spark，而不是ChatGPT或者Claude。

@Areskapitalon 这话说得也太狠了哈，直接把人家经济模型按在地上摩擦。 AI这波不是来上班的，是来掀桌子的啊。

这个年代破产得最彻底的就是印度的经济模式，中产梦被AI一夕之间彻底干碎了

@CEOBriefing 12个月就把软件工程自动化？这速度也太离谱了吧，码农还没回过神呢哈。 关键不是AI会不会，是大家到底会不会用，很多人怕是连“说明书”都还没翻开。

最新：Anthropic首席执行官Dario Amodei表示，软件工程工作有望在12个月内实现自动化。AI正引发一场范式转变，从“组建团队”转向“部署提示词”。如仍使用像Claude这样LLM，来处理基础任务，说明远未充分发挥这一强大工具潜能。对工具的误用，不仅浪费时间与金钱，更会错失良机。

@ArtofSpecuycky 老黄一开口，市场就开始换赛道了哈 昨天还在炒GPU，今天就得盯着“钱往哪儿流”了，真是AI版赶集啊

跟着黄仁勋的时间线走，每一阶段的钱都在提前切换 过去一年，如果你认真拆老黄每次GTC、财报会、闭门峰会说了什么，会发现AI基建的主线其实一直在切换。真正厉害的资金从来不是等新闻出来再追，而是提前拆下一阶段CapEx会流向哪里。 整个时间线其实非常清晰。 第一阶段：GPU → HBM → CoWoS → Memory 最早被老黄反复强调的方向。 Blackwell开始，HBM已经从外围配件升级成AI架构核心。老黄在财报会上明确说了：Blackwell和Vera Rubin的性能瓶颈，越来越取决于HBM带宽与供应。 很多人还把 $MU当周期股。但老钱已经开始把存储当AI structural asset来定价。 第二阶段：Ethernet → 光通信 → Silicon Photonics 2026上半年真正爆发的第二条主线。 AI集群进入十万卡级别后，真正限制效率的已经不是单卡性能，而是GPU之间的数据流动效率。老黄开始疯狂强调Spectrum-X、Ethernet、Scale-Out、Silicon Photonics。 NVLink解决的是Scale-Up，柜内互联。真正的大规模AI集群最终拼的是Scale-Out，柜外组网。AI越大，Ethernet的地位反而越强。 $ANET $MRVL $CRDO $ALAB $LITE ——这些AI数据流动收费站开始进入估值扩张阶段。 第三阶段：Inference CPU / Agentic AI 2026 Q1财报会后华尔街开始重新定价的方向。 很多人之前觉得AI时代CPU会边缘化。但老黄在5月财报会明确：Inference爆发后，CPU反而重新成为瓶颈。Agentic AI、Inference、Orchestration都需要大量CPU调度。 $ARM $QCOM $ADI $TXN $MCHP开始明显走强。 很多人还在盯GPU，但老钱已经开始提前布局Inference infrastructure。 第四阶段：AI-RAN + Edge AI + Physical AI 我认为这是2026下半年到2027真正可能成为主场的方向。 AI第一阶段是Training，第二阶段是Scale-Out。接下来AI必须真正离开hyperscaler数据中心，进入现实世界。 老黄已经把话说得很清楚：未来AI会进入运营商网络、边缘节点、基站、机器人、工厂、汽车。AI不会永远只待在云里。 这一阶段分三类受益方向。 AI-RAN和边缘网络：未来全球几百万个基站可能都会变成边缘AI节点。真正值钱的是射频、DAC/ADC、FPGA、边缘推理网络。 $NOK是AI−RAN联盟的核心发起人，英伟达直接投资背书。 Edge AI和On-Device AI：Agentic AI开始长进PC、手机、汽车、边缘网关。设备端AI最大的问题是低功耗推理加本地memory加实时连接。 $QCOM $SIMO这些传统边缘芯片公司的估值可能会被重新定价。 Physical AI和Robotics：散户喜欢炒机器人概念，但真正决定机器人能不能大规模落地的是工业控制、PLC、伺服系统、自动化基础设施。未来真正持续扩估值的，可能不是最性感的人形机器人概念，而是卡住物理世界执行层的工业收费站。 $TSLA $ABB $SYM $ROK。 总结 AI的下一阶段已经不只是训练模型。而是让AI真正进入现实世界。每一次老黄开口，都是在提前告诉市场下一阶段的钱会流向哪里。 #PhysicalAI #EdgeAI #AIRAN #光通信 #机器人 #数据中心 #半导体 #AI算力