供应链分析
产业链结构、上下游与瓶颈环节研究 · 共 1309 条 · 滚到底自动加载更多
-
OpenAI独占Cerebras算力致非超大规模客户等待无限期
机器学习 $RDDT 子版块有一个相当有趣的轶事性投诉。 关于 $CBRS + OpenAI： - 需要 1-2K tokens/秒，所以尝试使用 Cerebras。 - 声称 OpenAI 预留了所有产能，导致非超大规模云厂商的等待名单无限长。 - 陈述 https://t.co/ZXPPg7v5Ch
英文原文
There was a pretty funny anecdotal complaint from the Machine Learning $RDDT subreddit. On $CBRS + OpenAI: - Needs 1-2K tokens/sec so tries to use Cerebras. - Claims OpenAI reserved all capacity, making waitlist infinite for everyone else not a hyperscaler. - Stated https://t.co/ZXPPg7v5Ch
-
质疑美国AI巨头资本支出大幅下滑的可能性，并提及中国AI网络威胁论。
很难想象一个世界，其中美国AI超大规模云服务商（Hyperscaler）的资本支出（Capex）从$GOOGL到$META会出现大幅下降。 当像360这样的中国实体声称： 他们现在拥有一种“AI网络核武器”，可以黑客攻击西方公司和政府。 （周声称Mythos就像AI时代的网络... https://t.co/pBbhuAfJk3）
英文原文
Hard to see a world where US AI hyperscaler capex drops dramatically from $GOOGL to $META. When China entities like 360 go and claim: They now have a "AI cyber nuclear weapon" that can hack Western companies and governments. (Zhou claimed Mythos was like an AI era cyber https://t.co/pBbhuAfJk3
-
乐天与ASTS合资建日本LEO卫星网，以应对Starlink竞争。
据 Digitimes/日经报道，日本乐天(Rakuten)将与 $ASTS 成立合资企业，在日本建设低地球轨道(LEO)卫星网络。这一举措被广泛视为对 $SPCX Starlink 在日本日益增长的影响力做出的战略回应。
英文原文
Japan's Rakuten to establish joint venture with $ASTS to build out LEO satellite networks for Japan per Digitimes/Nikkei. "The move is widely viewed as a strategic response to the growing influence of $SPCX Starlink in Japan" https://t.co/RDFvgQNVsp
-
GM裁员换机器人并与NVDA合作，验证机器人行业价值
$GM 裁员 1,000 人并替换为 50 台机器人（6 月 22 日报道） 报道还显示 $GM 正与 $NVDA 洽谈工厂机器人合作。 这是机器人行业的验证，因为它可能证明了员工运营支出(operex)利润率 + 效率的提升 https://t.co/wocndLGOcp
英文原文
$GM cuts 1,000 workers and replaces them with 50 robots (Jun 22nd report) Reports also show GM is working on a deal with $NVDA on factory robotics. This is industry validation for robotics, since it likely demonstrates increased opex margins + efficiency from employee https://t.co/wocndLGOcp
-
POET 暗示激光器短缺将延续至 2029 年，加剧供应紧张。
关于 $POET 年度股东大会及光通信市场阅读的一些笔记： - “前三大激光器供应商占据 68% 的市场份额，且未来两年完全售罄” $LITE 首席执行官表示短缺将持续至 2028 年，因此 POET 隐含确认激光器短缺将延续至 2029 年。 - https://t.co/Lx8IHQ9fyK
英文原文
Just some notes on $POET AGM and read through on optical markets: - "The top three laser suppliers control 68% of the market, and they’re completely sold out for the next two years" $LITE CEO said into 2028, so POET implicitly confirms laser shortage is going into 2029 now. - https://t.co/Lx8IHQ9fyK
-
谷歌因算力限制Meta，致Meta转向Neoclouds且Gemini被削弱。
看来我们终于找到为什么 $META 在三月与 $NBIS 等 Neoclouds 签署巨额协议的原因了…… 以及为什么 Gemini 被大幅削弱。 据报道，$GOOGL 因算力限制，在 2026 年 3 月限制了 Meta 的容量。 谷歌 CEO 在上次财报中关于计算 https://t.co/LXGg8md2ur
英文原文
Guess we finally found why $META signed massive agreements with Neoclouds like $NBIS back in March... And why Gemini got super nerfed. $GOOGL reportedly restricted Meta's capacity in March 2026 because of compute restraints. Google's CEO said from last earnings computing https://t.co/LXGg8md2ur
-
中国AI在网安基准追平Anthropic，博主认为政府应对正确。
《华尔街日报》：“中国在网络安全领域已追平 Anthropic，AI 竞赛格局重置” 中国的智谱 AI (Zhipu AI) 在某些基准测试中（例如利用网络安全漏洞）已追平 Anthropic 的 Mythos 模型。 嗯，这可不是什么好消息。 与其指责做出了正确决定的政府…… https://t.co/L2GFG32zsM
英文原文
WSJ: “China Has Matched Anthropic in Cybersecurity, Resetting AI Race” China’s Zhipu AI, has matched Anthropic’s Mythos in certain benchmarks, such as exploiting cybersecurity bugs. Well, that’s not good. Instead of blaming the administration which did the correct thing… https://t.co/L2GFG32zsM
-
上游MLCC瓶颈笔记（内容未完整显示）
随便记一些笔记： 上游MLCC（多层陶瓷电容器）瓶颈…（原文未完整显示）
英文原文
Just some random notes : The upstream MLCC bottl…
-
看好宇树国内前景，但认为地缘风险阻碍其进入美国市场。
@chengchushow 我实际上对宇树科技(Unitree)在中国国内市场非常看好(bullish)。然而，我认为中国的人形机器人很难在美国市场取得太多渗透。特别是考虑到DJI事件，黑客可以窥探每个人的家，这展示了外国公司所面临的（风险/限制）。
英文原文
@chengchushow I’m actually very bullish on Unitree for CN domestic markets. However I don’t think there’s going to be much US market penetration for Chinese humanoids. Especially with the DJI incident where a hacker could see into everyone’s house, and shows what foreign companies have
-
机器人投资飙升，AI数据中心供应链与人形机器人存在交叉敞口。
机器人是下一个风口。 根据 PitchBook 3月数据（来源：a16z），交易数量和投资金额均呈飙升态势。 好消息是：相同的AI数据中心(AI DC)敞口通常也涉及人形机器人量产的交叉敞口。 例如DRAM/NAND与存储（用于人形推理/存储）的关系，或DFB激光器与 https://t.co/7BrzCd3cWD 的关系。
英文原文
Robotics is next. Both deal count and investment amounts are skyrocketing per pitchbook March data (source: a16z) Good thing is: the same AI DC exposure often has cross-exposure to humanoid ramp. Like DRAM/NAND with memory (on humanoid inference/storage) or DFB lasers with https://t.co/7BrzCd3cWD
-
马斯克收购光网络初创Mesh，分析其光引擎自研及激光器外采供应链。
$SPCX / 埃隆·马斯克收购了 Mesh，一家光网络初创公司。 该公司正在研发 1.6T OSFP（可插拔）光模块。 看来他们拥有光引擎/封装方面的技术，但很可能从商业公司采购连续波分布反馈(CW DFB)激光器。 $SIVE 是对初创公司更友好的供应商之一 https://t.co/A1H7Bd6ZU4
英文原文
$SPCX / Elon Musk acquires Mesh, an optical networking startup. Which is working on 1.6T OSFP (pluggable). It’s seems they own the optical engine/packaging side of things, but likely sources CW DFB lasers off merchant companies. $SIVE is one of the more startup friendly https://t.co/A1H7Bd6ZU4
-
反对高端开源模型，认为RISC-V的教训表明开源可能让中国绕过技术封锁
我实际上反对开发极高端的开源模型。我们在RISC-V上做过这个实验，结果给中国提供了一条绕过ARM和其他公司的途径。
英文原文
I'm actually against making extreme high end open source models. We did that experiment with RISC-V and gave China a way around ARM and others.
-
承认DeepSeek的创新性，主张用银行级监管框架解决AI模型蒸馏问题。
不，DeepSeek实际上非常有创新性。所以我才说美国模型只需要做好就行。至于蒸馏，那是另一个可以通过银行级安全/认证加监管解决的问题，至少对前沿模型来说是这样。因为分享凭证（如Netflix账号）给那些代理网络虽然违反服务条款，但还不构成犯罪。应该像共享银行账户一样来处理。
英文原文
Nah DeepSeek is actually very innovative. That's why I'm saying US models just needs to get good. As for distillation that's another problem that can be solved with bank level security/onboarding + regulations, at least for the frontier models. Since it's against TOS but not a crime to share credentials like Netflix accounts yet to those proxy networks. Should be treated like sharing bank accounts.
-
美企因成本优势转向使用中国AI模型，印证瑞银报告。
随手抛出一个想法，瑞银(UBS)的报告在经验上是真实的，而且有趣的是，关于Anthropic蒸馏技术的中文报告更多了。许多使用AI的美国初创公司/企业：正在使用像DeepSeek这样的中国模型，因为它们比 $NVDA 便宜得多。
英文原文
just throwing out a random thought, the UBS report is anecdotally true + and it's interesting there's more Chinese reports on Anthropic distillation. Many US startups/companies using AI: Are using Chinese models like DeepSeek since they’re much cheaper than https://t.co/COEW0qbwBK
-
马斯克警示内存供需失衡及涨价，凸显其对该瓶颈的高可见度。
埃隆·马斯克(Elon Musk)只是随意地拉响了警报。关于$MU / SK海力士(SK Hynix) / 三星(Samsung) 内存相对于供应的巨大需求+价格上涨。他可能对内存瓶颈(memory bottleneck)的可见度比其他人更高……
英文原文
Just Elon Musk casually sounding the alarm. On the massive demand + price hikes for $MU / SK Hynix / Samsung memory relative to supply. He probably has higher visibility into the memory bottleneck than others… https://t.co/iszX0YJHuS
-
功率半导体涨价，利好美股相关标的，中国厂商亦跟涨。
看来功率半导体(Power Semis)已经开始涨价了。 在800V直流(800V DC)转换全面到来之前，这对从 $AOSL 到 $POWI 的美国功率半导体交易主题上是看涨的。 中国的扬杰科技(Yangjie Technology)宣布今年第二次涨价，全线产品提价 https://t.co/UrseTyBMFp
英文原文
Looks like Power Semis are already starting price hikes. Which is bullish thematically for US power semi trade from $AOSL to $POWI, before the 800V DC shift fully hits. China's Yangjie Technology, announced a 2nd price increase this year, raising prices across its full product https://t.co/UrseTyBMFp
-
内存与大盘半导体抗跌，光子、太空及热门AI软件股大幅回调。
感觉唯一没崩盘的…… 是像 $MU 这样的内存、指数，或者像英特尔这样的大盘半导体股。 - 光子学领域从 $AXTI 到 $SIVE 下跌 40%。 - 太空领域从 $ASTS 和 $RKLB 一个月下跌 40%。 - 热门 AI 股如 $PLTR 年初至今下跌约 35%。 - 软件股如 $CRM 下跌 https://t.co/aqC7wEgAKB
英文原文
Feels like the only thing that hasn’t crashed… Is memory like $MU, indexes, or large cap semis like Intel so far. - Photonics from $AXTI to $SIVE down 40%. - Space from $ASTS and $RKLB down 40% 1M. - Popular AI names like $PLTR is down ~35% YTD. - Software like $CRM down https://t.co/aqC7wEgAKB
-
对比日本AI瓶颈垄断企业与硅谷高估值种子轮
全球市场简而言之： 日本公司：在超大规模云服务商（Hyperscaler）AI建设的关键瓶颈环节拥有事实上的垄断地位。 估值：1.5亿-3.5亿美元 硅谷：来，给你2亿美元做“种子轮”。https://t.co/O99br8y1yr
英文原文
Global markets in a nutshell: Japanese Company: Active monopoly over random chokepoints in the hyperscaler AI buildout. Valuation: $150m-$350M Silicon Valley: Here's a $200m for a "Seed Round". https://t.co/O99br8y1yr
-
美光CEO预测人形机器人将开启长达数十年的内存需求超级周期。
今天有个很有趣的声明：$MU CEO 预测由人形机器人驱动的记忆体需求将开启一个持续数十年的周期。 “他说，人形机器人所需的内存量大约是当今 L2+ 级自动驾驶汽车的十倍。” “而这一需求浪潮预计将在 https://t.co/bVnvG1Ihit 之前开始”
英文原文
Very interesting statement today: $MU CEO predicts a multi-decade memory demand cycle driven by humanoid robots. "Humanoid robots, he says, will require roughly ten times more memory than today’s Level 2+ autonomous vehicles." "And that demand wave is set to begin before the https://t.co/bVnvG1Ihit
-
POET/SIVE有望为三大超大规模云服务商提供动力
来自 Poet OSINT 社区的有趣新信息发现： $POET / $SIVE 很可能将为顶级三大超大规模云服务商（亚马逊、微软、谷歌之一）提供动力。 鉴于 Lumilens 首席执行官 Ankur Singla 的 LinkedIn 更新。 他声明他们的客户是前三大之一 https://t.co/Rs0D8HE1T9
英文原文
Fun new information discovery from Poet OSINT community: Seems likely that $POET / $SIVE are going to power a Top-3 hyperscaler (either Amazon, Microsoft, Google). Given a Linkedin update from Ankur Singla (CEO of Lumilens). Who stated their customer is one of the top 3 https://t.co/Rs0D8HE1T9
-
梳理OpenLight在Advantest之外的公开供应链生态伙伴。
OpenLight（私有公司）每次我看它似乎都在变得越来越大。 如果你对他们 Advantest 之外的公开生态系统感兴趣： $JBL（$SIVE 合作伙伴）用于扩展光子集成电路(PICs) $MRVL 和 $MXL 用于数字信号处理器(DSPs) $TSEM 用于晶圆代工。 $300394.SZ / TFC Optical（封装测试/子组件）
英文原文
OpenLight (private) seems to get bigger and bigger every time I look at it. If you’re curious about their public ecosystem outside of Advantest: $JBL ( $SIVE partner ) for scaling PICs $MRVL and $MXL for DSPs $TSEM for the foundry. $300394.SZ / TFC Optical (OSAT/subassembly)
-
博主批评彭博社错误将台积电和英伟达供应链比作泡沫。
彭博社在3月照搬美银针对韩综指的剧本，简直蠢透了。他们将台湾$TSM ~23.6倍远期市盈率(P/E)和$NVDA上游供应链描述为“泡沫”。接着用“债务”而非“利润率”一词，并将台湾市场与互联网泡沫(.com)相提并论。
英文原文
Nothing is more stupid than Bloomberg copying BofA playbook with KOSPI back in March. And framing Taiwan’s $TSM ~23.6x forward p/e and $NVDA upstream supply chain as a “bubble”. Then using the words going into “debt” instead of “margin” + comparing Taiwan’s market to the .com https://t.co/DPXpK4zhzq
-
NVDA导致EML瓶颈迫使架构转向，IMO光子学及CW激光器成关键瓶颈。
集成模块光学(IMO)光子学主题+连续波(CW)激光器瓶颈是顶级的。市场真的像有短期记忆丧失症，忘了$LITE从2024年到现在是如何从30亿美元涨到650多亿美元的。因为$NVDA导致了外延激光管芯(EML)瓶颈，并迫使架构改变。我们实际上正在看到同样的情况
英文原文
IMO photonics theme + CW laser chokepoint is goated. It's legit like markets have short term memory loss and forgot how $LITE went from $3B -> $65B+ from 2024 to now. Because $NVDA caused EML bottlenecks, and forced architectural changes. We're literally seeing the same
-
回顾CXL内存池化概念股ALAB和MRVL过去4个月的大幅上涨。
做多CXL（Compute Express Link，计算快速链接）以实现内存池化的最佳时机是4个月前。 $ALAB 从 $117 涨至 $440，涨幅 276% $MRVL 从 $79 涨至 $309，涨幅 291% https://t.co/7XmXHyXgb1
英文原文
Best time to long CXL for memory pooling was 4 months ago. $ALAB up 276% from $117 -> $440 $MRVL up 291% from $79 -> $309 https://t.co/7XmXHyXgb1
-
特朗普应修复盟友关系，利用海外供应链优势赢得贸易战。
特朗普政府现在可能意识到，赢得贸易战的关键在于前沿供应链，如量子计算(Quantum)、人工智能(AI)和机器人(Robotics)，而非廉价的耐克袜子出口。鉴于$ASML、$TOWA、$LPK、味之素(Ajinomoto)等关键企业位于日本、欧盟、台湾、韩国等地，且各国掌握着使这些行业运转所需的垄断技术和瓶颈环节，而这些都恰好位于美国境外。对所有合作伙伴加征关税显然不是个好主意。但在我看来，现在修复关系并通过合作伙伴将全球供应链武器化以换取更好的贸易协议，为时未晚。
英文原文
Trump administration is prob realizing now that the key to winning Trade Wars: Is through frontier supply chains like Quantum, AI, Robotics. Not cheap Nike sock exports. With $ASML , $TOWA, $LPK, Ajinomoto, over in Japan, EU, TW, Korea, and others holding all the cards… Since each country holds different monopolies and chokepoints needed to make each industry work. Which all happens to be OUTSIDE the US. Tariffing all our partners turned out to be not a great idea. But not too late imo to repair relations and weaponize global supply chains through partners if Trump wants better trade deals.
-
分析多只AI供应链股票的NAV折价、控股结构及增长潜力，偏好独立增长标的。
个人笔记如下： 1. Priortech 持有 $CAMT 21% 的股份，市值约为其 1.35 倍，看起来更像是一家控股公司。 2. Bit Digital 的人一直在喊单，但 $WYFI 的净资产值(NAV)折价并不纯粹，存在大量稀释且并非 AI 母公司。 3. 纬创(3231)是我自 $SIVE 以来见过的最佳标的，加上 Ayar 和纬颖(Wiwynn)是我最喜欢的“三剑客”。 市值约 162 亿美元，第一季度收入增长惊人，同比增长 144%。 持有纬颖 35.46% 的股份，约为其市值的 0.66 倍。我预计纬颖将继续增长。 4. 中美硅科技(5483)持有全球晶圆(6488) 46.64% 的股份。 市值 35 亿美元，持股价值约 79 亿美元，但独立增长并不快，尽管其交易价格大幅低于净资产值(NAV)。 5. 一真控股(Iljin Holdings)持有一真电气(Iljin Electric) 42.99% 的股份，市值约 2.2 亿美元，而净资产值(NAV)持股约 11.3 亿美元。这与变压器/电缆直流相关。 但我不确定是否相信韩国股票能实现 NAV 解锁，可能更适合激进投资者(Activist Investors)。 6. Simmtech Holdings 市值约 1.7 亿美元，而母公司约 10 亿美元，看起来也很离谱。 即 6 倍以上的 NAV。这是 PCB 基板相关的名字。同样，另一个可能不信任韩国治理类型的名字。 我认为 $ACMR 和沪硅产业(WUS) 很快都会有 H 股子公司上市，且这两者拥有最大的 NAV 折价和独立增长能力。 此外，有一位非常成功的激进投资者正在针对沪硅产业，所以我预计那里会达成某种中间立场。 所以也许周一我会在此基础上增加这些股票的持仓集中度，我们拭目以待。纬创独立增长非常快。 我仍在进行研究，通常这些是“淋浴时的想法”，因为我尚未正式得出结论，但觉得其他人可能感兴趣。
英文原文
In terms of personal notes: 1. Priortech owns 21% of $CAMT , ~1.35x their MC, seems more like holding co. 2. Bit Digital people kept shouting, but NAV discount is not clean $WYFI , lot of dilution + not AI parent. 3. Wistron (3231) is the best one I've seen since $SIVE + Ayar + Wiwynn are one of my favorite trio. ~$16.2B MC, Q1 revenue up a ridiculous amount with 144% Y/Y growth. Owns 35.46% of Wiwynn, which is ~0.66x of MC. I'd expect Wiwynn to keep on growing. 4. Sin-American Silicon 5483 does own 46.64% of GlobalWafers (6488). MC is $3.5B, stake value is ~$7.9B, but not exactly growing too fast independently, though it's heavily trading heavily discount to NAV. 5. Iljin Holdings owns 42.99% of Iljin Electric, it's ~$0.22B MC vs. ~$1.13B NAV holding. This was transformer/cable DC related. But again idk if I trust korean stocks for NAV unlock, probably better for activist investors. 6. Simmtech Holdings was ~$0.17B MC vs. ~$1.0B of parent, which also looks ridiculous. So 6x+ NAV. PCB substrate related name. Also another prob don't trust korean governance type name. I think both $ACMR and WUS have H-Share subsidiary listing soon, and those were the biggest NAV discounts with independent growth. Then you have a highly successful activist going after WUS so I'd expect some middle ground there. So maybe I'll put more concentration into those on top of what I already own Monday, we'll see. And Wistron was growing very fast independently. I'm still doing research, normally this goes in shower thoughts since I haven't formally made a conclusion, but thought others might be interested.
-
质疑中国如何从ASML走私大型设备。
中国究竟是如何从 $ASML 走私进一辆校车大小的东西的？这才是真正的问题。
英文原文
How does China manage to smuggle in something the size of a school bus from $ASML ? Is the real question.
-
回顾存储暴涨，警示光通信预测过热，但看好光子学处于超周期早期。
提醒一下：3个月前，99%的X用户看空存储芯片... 自那以来：$MU $380 -> $1122 (+195.26%) $SNDK $565 -> $2155 (+281.42%) $EWY $132 -> $219 (+65.91%) SK海力士 849K -> 2.685M KRW (+215.5%) 三星 172K -> 362K KRW (+110.7%) 如果你看到大量预测称三星等存储公司在2028年将成为全球最赚钱的公司。 也许现在应该独立思考，跳出当时的叙事如“石油、LNG、氦气、伊朗等”。 现在对光通信(optical)及其2027、2028年的预测可能也是如此。 但我确实认为光子学(photonics)和存储是两大顶级主题，且光通信正处于超级周期的非常早期阶段。
英文原文
Just a reminder: 99% of X was bearish on memory 3 months ago... Since then: $MU $380 -> $1122 (+195.26%) $SNDK $565 -> $2155 (+281.42%) $EWY $132 -> $219 (+65.91%) SK Hynix 849K -> 2.685M KRW (+215.5%) Samsung 172K -> 362K KRW (+110.7%) If you see a bunch of projections on memory names like Samsung becoming the most profitable company in the world in 2028. Might be a good idea to think independently outside of the narratives at the time like "Oil, LNG, Helium, Iran, etc." Probably same thing now with optical names and their projections into 2027, 2028. I do think photonics and memory are the 2 top themes though, with optics being very early into the supercycle.
-
LPK获大客户认可，玻璃基板成封装新趋势，Absolics与三星电机将陆续量产。
我的意思是……如果非要我猜的话，$LPK 确实声称“全球主要玩家中 80% 的客户已选择 LPKF 设备”。 玻璃基板(Glass Substrates)显然是下一个封装技术的转变！ 从预估的时间表来看，主要玩家如 Absolics 将在 2026 年下半年(H2 2026)开始量产爬坡，三星电机(Samsung Electro Mechanics)则在 2027 年。
英文原文
I mean... if i had to guess, $LPK did claim "80% of customers among major global players have selected LPKF equipment". Kinda clear glass substrates is the next packaging shift! From est. timelines, major players like Absolics is ramping H2 2026, Samsung Electro Mechanics 2027
-
AMD抢购CW激光，西方供应链产能稀缺，瓶颈或刚开始。
Trendforce 报告称，$AMD 正积极通过多个主要采购订单来确保连续波激光(CW laser)的供应…… 这可能是瓶颈的开端吗？ 除了 $SIVE 或 $AAOI 以及可能的 Macom 之外，西方供应链中几乎没有多少独立产能了。
英文原文
Trendforce reports that $AMD is actively trying to secure CW laser supply with multiple major procurement orders... Is probably just the start of the bottleneck? There's not much independent capacity in Western supply chains left other than $SIVE or $AAOI and maybe Macom.