供应链分析
产业链结构、上下游与瓶颈环节研究 · 共 1413 条 · 滚到底自动加载更多
-
AI供应链瓶颈加剧，内存/基板/光学组件供应紧张，利好相关标的。
本周我发现的有趣内容摘要： - $INTC CEO 表示部分内存价格已上涨 5-7 倍，并警告短缺将在 2027 年恶化。 $SNDK / $MU / 三星兄弟听到这个会很开心……（来自圣克拉拉 2026 AI 基础设施峰会） - $GFS + $MRVL 签署延长 SiGe 产能协议，用于可插拔/NPO/CPO。"预计将增加显著产能以支持 Marvell 日益增长的需求。" 鉴于 $SIVE 是 SCALE 的参考激光器，我喜欢关注 Globalfoundries 的相关更新。 - $CIEN 预计至 FY2029 年收入年增约 30%，调整后毛利率 50%，自由现金流利润率 20%。 公司称组件可用性而非需求限制了增长……似乎从 $NVDA 到上游供应商，供应都是限制收入增长的因素。 - 华邦电子以 11.2 亿美元收购英飞凌的 NOR Flash 和 F-RAM 业务。 鉴于 NOR Flash ASP 和重要性持续上升，我很惊讶英飞凌会出售，但这有利于华邦的整合。 - 三星据报道扩大 DDR5 模块和 SSD 组装的外包。 我认为一些台湾本地报告将创见和宇瞻标记为潜在受益者。报告具体点名了 Dreamtech、Hanyang Digitech 和 SFA Semicon。 我注意到其他韩国公司进入新的三星电子合同，可能是为了传统内存模块，大概 1-2 个月前……所以认为一些媒体评论可能也错过了这一点。也许太小以至于他们没注意到？ - $NBIS 三个月内第二次提高 AI 云价格（~17-21%），对其他 Neoclouds 如 $CRWV 到 $IREN 的财报是利好解读。 - $AMZN 与 $GNRC 签署 24 亿美元协议，用于 2027-2028 年的发电机交付……并获得了与高达 80 亿美元总采购额挂钩的认股权证。 值得关注亚马逊与谁签署这些认股权证协议（例如 Alchip, $AAOI, $QCOM, $MRVL 等）以及他们想要什么样的基础设施采购。 - $TSEM 和 NewPhotonics 开始批量出货激光集成光学引擎。 不熟悉 NewPhotonics，所以这里不多评论。但今天会做点研究。 - 高盛据报道认为 ABF 缺口将从 14% 扩大到 51%。预测短缺为 2026 下半年 14%，2027 年 34%，2028 年 51%…… 更多的基板瓶颈内容，关于需求不平衡并不令人惊讶。 - Crusoe 以 309 亿美元估值融资 39 亿美元。 好吧，这个估值比 Instinct 以 10 万用户为基础以泡沫 100 亿美元估值融资要好…… - 日本和美国据报道讨论价值 129-193 亿美元的 $GFS 运行晶圆厂。 这相当于 Globalfoundries 市值的一半 lol。这里没有太多细节。 - $ARM CEO Rene Haas 表示他有信心确保足够的供应以满足 20 亿美元的客户订单（针对其 AGI CPU）。 可能还有更多需求，可能像往常一样是供应受限的增长，所以 $ARM 发出确保分配的信心信号是好事。 - 供应链消息人士称 Kioxia 约 50% 的 LTA 产能已被占用，而 SK Hynix/三星为 3-5 年的 LTA 预留约 70% 的产能。 不确定人们为何看空内存…… 我不喜欢像新闻播报员一样逐个发布更新，所以希望这能节省大家查看我喜欢关注的内容的时间。 顺便说一句，惊讶于 $AMD 到 $INTC 的市场在加息后今天反弹。认为高贝塔值在抛售中超调最多。
英文原文
TLDRs of things I found interesting this week: - $INTC CEO says some memory prices have risen 5-7× and warns the shortage will worsen in 2027. $SNDK / $MU / Samsung bros will be happy to hear this... (from the AI Infra Summit 2026 in Santa Clara) - $GFS + $MRVL sign extended SiGe capacity agreement for pluggable/NPO/CPO. "Expected to add significant capacity to support Marvell’s growing requirements." I like to keep an eye out on Globalfoundries related updates given $SIVE is their reference laser for SCALE. - $CIEN expects roughly 30% annual revenue growth, 50% adj. gross margin, 20% FCF margin, through FY2029. The company says component availability, not demand, is limiting growth... Seems common across $NVDA to upstream suppliers that supply is the limiting factor for revenue growth. - Winbond acquires Infineon's NOR Flash and F-RAM business for $1.12 billion. Very surprised Infineon sold considering NOR Flash ASP + importance just keep going up and up and up, but good for Winbond consolidation. - Samsung reportedly expands outsourcing of DDR5-module and SSD assembly. I think some local Taiwan reports flag Transcend and Apacer as possible beneficiaries. The report specifically identifies Dreamtech, Hanyang Digitech and SFA Semicon. I've noticed other Korean firms entering new Samsung Electronics contracts for likely legacy memory modules maybe like ~1-2m ago.. So think some media commentary probably missed this too. Maybe too small for them to notice? - $NBIS raises AI cloud prices for the 2nd time in three months (~17-21%), bullish read through for other Neoclouds like $CRWV to $IREN earnings. - $AMZN signs $2.4B agreement with $GNRC for generator deliveries throughout 2027-2028... and received warrants tied to as much as $8B of aggregate purchases Good to keep an eye out on who Amazon signs these warrant agreements with (eg. Alchip, $AAOI, $QCOM, $MRVL, etc) and what type of infra procurement they want. - $TSEM and NewPhotonics begin volume shipments of laser integrated optical engines. Not familiar with newphotonics, so not going to leave much commentary here. But will do some research today. - Goldman Sachs reportedly sees the ABF deficit expanding from 14% to 51%. Forecast shortages are 14% in 2H26, 34% in 2027 and 51% in 2028... More substrate bottleneck stuff, not surprising regarding demand imbalances. - Crusoe raises $3.9B at a $30.9B valuation. Okay this valuation is better than Instinct raising at a bubble $10B valuation off 100k users... - Japan and the U.S. reportedly discuss a $12.9–19.3B $GFS run fab. That's like half of Globalfoundries marketcap lol. Not much details here. - $ARM CEO Rene Haas said he is confident that they can secure enough supply to meet $2B in customer demand (for their AGI CPU). There's probably a lot more demand, probably supply constrained growth as usual so good $ARM signaled confidence in securing allocations. - Supply chain sources said Kioxia has about 50% of its LTA capacity tied up, while SK Hynix/Samsung reserve about 70% of capacity for LTAs for 3-5 years. Not sure how people are bearish on memory.... I don't like posting updates one-by-one like a news reporter, so hope this saves some time of stuff I like looking at. On a side note was surprised markets from $AMD to $INTC rallied today after rate hikes. Think high-beta overshot the selloff the most.
-
富士康拟联合投资以解决CW激光器等光子学瓶颈。
让人想起 $NVDA 与 $LITE + $COHR 的情况： 富士康（Foxconn）正在考虑联合投资，以避免上游材料短缺。 富士康互连技术董事长指出，连续波(CW)激光器和光纤是光子学领域最大的瓶颈。 看到这两个主题在行业评论中被反复提及，非常有趣。 几乎就像 X 上某个动漫角色关于 CW 激光器瓶颈的说法是对的。
英文原文
Reminiscent of $NVDA w/ $LITE + $COHR: Foxconn is now considering joint investments to avoid upstream material shortages. Foxconn Interconnect Chairman identifies CW lasers and fiber as the biggest bottleneck in photonics. Very interesting to see these two themes reiterated across industry commentary. Almost like a certain Anime character on X was right about CW laser bottlenecks.
-
AI算力需求长期锁定，短期恐慌不影响整体建设。
$IREN CEO 在 Anthropic 关于前沿模型放缓的通话中表示：“即使模型从此不再改进……仅部署现有能力所需的算力，也将超过世界未来数年能构建的总和。”我同意，当前叙事让人想起 $NVDA DeepSeek 抛售潮或 $GOOGL TurboQuant 恐慌。结果如何？这些每年发生一两次的恐慌最终都被证明对整体 AI 建设无关紧要。尤其是当我们的需求合同已涵盖从先进封装到内存采购直至 2030 年代的整个周期时。
英文原文
$IREN CEO on Anthropic call for frontier slowdown: “Even if models never improved from here… just rolling out what they can already do would take more compute than the world can build for years” I agree, and current narrative is reminiscent of the $NVDA DeepSeek selloff or the $GOOGL TurboQuant scare. Guess what happened? These scares once or twice a year all turn out to be immaterial to the overall AI buildout. Especially when we have contracted demand all the way into the 2030’s now from advanced packaging to memory procurement.
-
AI巨头不会放缓研发，监管实为遏制竞争。
我认为 Anthropic 或 OpenAI 有意义地放缓前沿模型开发(frontier development)的概率几乎为 0%。 而且他们试图推动的大多数“监管”都是监管俘获(regulatory capture)，旨在减缓竞争。 1. Anthropic/OpenAI 不会自愿交出他们的...
英文原文
I assign essentially 0% probability that Anthropic or OpenAI will slow down frontier development in any meaningful way. And that majority of the "regulation" they try making is regulatory capture to slow down competition. 1. Anthropic/OpenAI will not willingly hand over their
-
甲骨文GPU高溢价续租证伪快速折旧论，利好AI算力租赁及英伟达。
虽然是周末，但我惊讶于自己错过了 $ORCL 财报中的这一信息。这对 $NBIS 到 $IREN 的影响相当积极，因为：甲骨文表示所有用于续租的 GPU 容量均以高于此前合同 20% 的溢价被重新售出。大部分设备机龄在 4 年以上……因此，这再次证伪了 Burry 关于 GPU 快速折旧的论点。而 $NVDA 和 Neocloud 阵营则赢了。
英文原文
It’s the weekend but surprised I missed this from $ORCL earnings. Implications for $NBIS to $IREN are pretty positive since: Oracle said all GPU capacity for renewal were resold +20% premium above previous contracts. Majority of the equipment was 4Y+ and older… So another L for the Burry’s rapid GPU depreciation thesis. And W for $NVDA and the Neocloud party.
-
Anthropic报告预示AI将重塑经济结构，利好资本密集型AI硬件板块。
Anthropic 的《经济情景》报告读起来很有趣：1. Anthropic 的“极端”情景显示 2030 年 GDP 增长达 15.4%，约为当前增速（~2.1% Y/Y）的 7.3 倍，这与 Elon Musk 关于 AI 将超越国债的说法一致。2. 几乎所有分歧发生在 2027 年之后。模型显示 2026-2027 年分离度较小，随后曲线在 2028-2030 年急剧弯曲。我的解读是：2026-2027 为基建年，2028-2030 为经济加速期。这与 $AMZN / $META / $GOOGL 的资本支出及短期 FCF 逆风（但 2028 年后预计巨额盈利）相吻合。3. 极端情景下，知识工作者失业率升至 17.9%，整体失业率 11.9%。“个人层面意味着程序员和客服可能需转行做电工和护士，这些受 AI 影响较小。”计算机科学专业 GG。4. 非认知工资：极端情景下比基准高 33.6%。电工/建筑类劳动力变得极具价值？（但这尚未计入人形机器人...）而 2030 年知识工作者工资比基准低 11.5%。5. 总资本收入比无 AI 时高约 81%。生产力爆炸 -> 总劳动收入几乎不变 -> 增量 GDP 几乎全归资本。因此资本所有权而非劳动... 使得计算、电力、半导体、网络、机器人等股票变得非常重要... TLDR: - 经济增长起飞 - 需思考的工作 = 失业 + 降薪 - 体力劳动更重要，如电工 - AI 股票起飞。
英文原文
Anthropic's Economic Scenario publication was a pretty interesting read: 1. Anthropic's "extreme" scenario has GDP growth reaching 15.4% in 2030 So around 7.3x current rates (~2.1% Y/Y), which is consistent with what Elon Musk is saying about outgrowing the national debt with AI. 2. Almost all of the divergence happens after 2027. Their model shows relatively little separation in 2026-2027, then the curves bend sharply into 2028-2030. So my read is that it points to 2026-2027 as buildout years and 2028-2030 as economic acceleration. Kinda gets mapped as well when you look at $AMZN / $META / $GOOGL capex, and short term FCF headwinds (but massive projected profitability post 2028) 3. In an extreme scenario, unemployment of knowledge workers goes to 17.9%. With 11.9% economy-wide unemployment "At the individual level, it means coders and call service center agents may have to switch to jobs like electrician and nurse, which are less exposed to AI." GG computer science majors. 4. Non-cognitive wages: +33.6% vs. baseline in an extreme scenario. So your electricians/construction-type labor becomes extremely valuable? (but this isn't factoring in humanoids yet...) Whereas -11.5% for knowledge workers in 2030 vs. baseline. 5. Aggregate capital income is ~81% higher than it would have been without AI Productivity explodes -> aggregate labor income barely changes -> essentially all of the incremental GDP accrues to capital. So that makes ownership of capital rather than labor... so equities like compute, power, semis, networking, robotics, etc. become pretty important... TLDR: - Economic growth go brr - Jobs that require thinking = unemployment + lower wages - Physical labor = more important, like electricians. - AI equities go brrr.
-
博主看好光子学供应链，认为激光器光纤将重现内存超级周期。
花旗银行TMT团队：“激光器和光纤正被定位为‘下一个HBM（高带宽内存）’。我从2025年就开始谈论像 $LITE / $AAOI 这样的激光器公司，以及向连续波(CW)架构的转变，这领先于当前的行业共识。我完全同意Lightmatter和花旗的观点。仅供参考：激光器方面：$LITE / $SIVE / $AAOI / $COHR / $MTSI / 古河电工 / 住友电工是较知名的标的。光纤方面：$GLW / 古河电工 / 古河光缆 / 住友电工是较知名的敞口（显然供应链中还有更多）。*披露，个人对该主题有持仓。补充评论称：‘紧张的供应状况将支持这些资产的稀缺性溢价。’我们已经看到激光器价格上涨，正如今天在深圳CIOE 2026上看到的 $SIVE 渠道检查以及 $LITE 的财报评论所示。因此，我认为2027-2028年的光子学超级周期将非常类似于2025-2026年的内存周期。尤其是随着高盛修正后的估计显示，每GPU-ASIC的光学内容将显著增加……
英文原文
Citibank TMT: "lasers and optical fibers are being positioned as the "next HBM." I've been talking about lasers like $LITE / $AAOI since 2025... and architectural shifts toward CW, ahead of the current industry consensus. And fully agree with Lightmatter/Citi. Just for reference: For lasers: $LITE / $SIVE / $AAOI / $COHR / $MTSI / Furukawa / Sumitomo are your better known ones. For fiber: $GLW / Furukawa / Fujikura / Sumitomo are your more well-known exposure (obviously a lot more in the supply chain) *disclosure, personal exposure to theme With added commentary that: "Tight supply conditions will support a scarcity premium for these assets." We're already seeing price hikes with lasers as seen with $SIVE channel checks at CIOE 2026 Shenzhen today + $LITE commentary from earnings. So I think the 2027–2028 photonics supercycle will look a lot like memory did in 2025–2026. Especially as optical content/GPU-ASIC goes up significantly per Goldman Sachs revised estimates...
-
Astra需求激增致OpenAI或暂停新订阅，反映AI算力瓶颈。
对Astra的需求前所未有，以至于OpenAI可能会暂停新的Pro订阅。如果你需要翻译：AI基础设施建设+算力需求激增。
英文原文
Demand for Astra is unprecedented to the point OpenAI might pause new Pro subscriptions. If you want a translation: AI buildout + compute go brrr.
-
高盛在最近的一份报告中上调了全球光模块市场预测：- 2026年：677亿美元（上调33%）- 2027年：1314亿美元（上调81%）- 2028年：1485亿美元（上调115%）对于 $NVDA Rubin Ultra，高盛增加了其模型中每GPU的光模块数量：1.6T：0 -> 2；3.2T：3 -> 5。高盛还将2027年1.6T及以上产品的预测提高了61%……这也非常重大，预计硅光子( Silicon Photonics )渗透率为80%。因此，每芯片的光学内容更多 = 更多的连续波(CW)激光器，更多的绝缘体上硅(SOI)衬底，光学公司日子好过。
英文原文
Goldman Sachs in their recent note upgraded their global optical module market forecast to: - $67.7 billion (+33% revision) for 2026 - $131.4 billion (+81% revision) for 2027 - $148.5 billion (+115% revision) for 2028 For $NVDA Rubin Ultra, GS increased its modeled optical modules per GPU: 1.6T: 0 -> 2 3.2T: 3 -> 5 GS also raised their 1.6T and above forecast for 2027 by 61%... so that's also very material, with 80% silicon photonics penetration rate estimated. So more optical content per chip = a lot more CW lasers, more SOI substrates, happy optical companies.
-
存储瓶颈未变，供需失衡加剧，基本面与短期价格常背离。
既然 $MU 到 $SNDK 都实现了大幅反弹。很高兴能提醒一下，存储瓶颈（Memory Bottleneck）并没有改变！连续波激光器（CW Lasers）、衬底（Substrates）或其他领域也是如此。但短期情绪（取决于价格/宏观）往往会变化。我认为许多供需失衡（Demand Imbalances）会比人们预期的更严重：- 今天日经新闻提到一些日本分销商称存储需求缺口已达 40-60%（比供应高出 67-150%）。随着普遍价格到年底上涨 50%。- $SPCX 未包含在 $1.3T 超大规模云厂商资本支出（Capex）数据中（我认为 Wells Fargo 等估计 AI 资本支出约 2630 亿美元），因此总资本支出数据可能会令人惊讶。- $SNDK 预计 80% 的毛利率将持续到 2030 年……（欢迎来到标普 100 指数）- 再次强调，现在对 2031 年有了长期的可见性。像 Samsung Memory 这样的公司波动性非常大……我的一些持仓上涨了 270%+，所以我发现更容易在波动时期持有。但无论如何，运营基本面并不总是与短期价格走势一致。同样的概念也可以应用于其他板块。
英文原文
Now that $MU to $SNDK have pulled a major recovery. Good to have a reminder that the memory bottleneck hasn’t changed! Same with CW lasers… substrates… or others. But short term sentiment (depending on prices/macro) often does. I think a lot of demand imbalances get worse than people expect: - There’s some Japanese distributors in Nikkei today saying memory demand deficit has reached 40-60% (67-150% higher than supply). With general prices going up 50% by years end. - $SPCX isn’t included in the $1.3T hyperscaler capex figures (I think Wells Fargo ets. ~$263b AI capex), so total capex figures might be surprising. - $SNDK expects 80% gross margins to continue into 2030… (welcome to S&P 100) - and once again you have long term visibility into 2031 now with companies like Samsung Memory is very really volatile… some of my positions are up 270%+, so I find it a bit easier to hold through volatile periods. But regardless, operating fundamentals doesn’t always align with short term price movements. Same concept can be applied to other sectors.
-
NVDA/AVGO受供应限制，AI技术突破，巨头受益。
让我理清一下，各位……你们拥有： $NVDA 70% 的营收增长预测受限于供应瓶颈（若无限制将超 100%） $AVGO AI 营收未来两年同比增速超 100%，从约 500 亿美元增至 2300 亿美元 OpenAI Astra 大幅超越 AGI 基准并开始自我训练。 $MSFT / $GOOGL
英文原文
Let me get this straight chat… you have: $NVDA 70% supply constrained revenue growth projections (would be 100%+) $AVGO AI revenue growing 100%+ Y/Y for next 2 years from mid ~$50B -> $230B OpenAI Astra blowing away AGI benchmarks and starting to train itself. $MSFT / $GOOGL
-
韩国零售界因三星高杠杆前景，在代币化辩论中支持HOOD。
根据我的渠道调研，韩国零售市场似乎坚定地站在 $HOOD 一边，而非 $AMC 的 CEO。韩国方面认为，40 倍杠杆的三星 / $SKHY 这一前景，足以成为赢得代币化辩论的有力论据。
英文原文
From my channel checks, it appears Korean retail is firmly siding with $HOOD over the $AMC CEO. Korean sources find the prospect of 40x leveraged Samsung / $SKHY a powerful enough argument to win the tokenization debate.
-
分析美中博弈下，欧日企业在EUV及光子学领域的垄断地位。
所以……这是我对美中紧张局势的看法： 你拥有像这样的欧洲垄断企业： - $ASML（极紫外光刻机 EUV） - $SOI（光子学-硅基绝缘体 Photonics-SOI） - 蔡司 SMT（ASML EUV 供应商）、通快 Trumpf（ASML EUV 激光供应商） 日本拥有许多近乎/完全的垄断企业，例如： - TOK、JSR、信越化学 ShinEstu、富士胶片 Fujifilm 用于 EUV
英文原文
So... here's my take on US / China tensions: You have European monopolies like: - $ASML (EUV) - $SOI (Photonics-SOI) - Zeiss SMT (ASML EUV supplier), Trumpf (ASML EUV laser supplier) Japan has many near/complete monoplies like: - TOK, JSR, ShinEstu, Fujifilm for EUV
-
中国暂停部分稀土对美出口，日本依赖进口镝铽可作为重要风向标。
据报道，中国暂停向美国出口部分稀土。这不是全面禁令，也未提及具体稀土品种。但……如果让我猜测，日本通常是一个不错的风向标，因为日本：- “几乎完全依赖进口镝和铽 https://t.co/rjCGzBGTmD
英文原文
China reportedly halts some rare earth shipments to US. Not a new blanket ban, or any mention of specific rare earths. But… if I had to guess, Japan is usually a decent indicator of what’s actively important since Japan: - “sources virtually all of its dysprosium and terbium https://t.co/rjCGzBGTmD
-
解析SIVE披露的CW DFB产能及其在全球供应链中的占比。
在审视 $SIVE 披露的超 1 亿 CW DFB 非凡产能时，这是正确的回应。提供一些背景：Trendforce 7 月数据显示，全球 CW/EML 产能约为 6.084 亿/年（年化月产量约 5070 万单位）。其中 $AVGO、$LITE 和住友（Sumitomo）占据了 https://t.co/tx3kjhYoO1
英文原文
This is the correct response when looking at $SIVE >100m CW DFB extraordinary capacity disclosure. To give some context, Trendforce July data cites global CW/EML capacity is roughly 608.4M/year (~50.7M monthly units annualized). With $AVGO, $LITE, and Sumitomo making up https://t.co/tx3kjhYoO1
-
NVDA收购Hugging Face属防御性举措，因CRM等已接触。
@Shiboo9019 这 partly 是一次防御性收购。据报道，$CRM 和其他公司开始接触 Hugging Face 寻求收购后，$NVDA 才开始了谈判。
英文原文
@Shiboo9019 it's partly a defensive acquisition. $NVDA talks started after $CRM and other companies were reportedly approaching Hugging Face for an acquisition.
-
英伟达129亿美元收购Hugging Face，类比微软收购GitHub。
$NVDA 以 129.303 亿美元收购 Hugging Face。 Jensen 陈述收购原因为： - 1800 万+开发者使用 Hugging Face - 300 万+模型，50 万+数据集，100 万+应用 - 20 万+公司使用该平台 - 支持开放权重模型 这几乎让我想起 $MSFT 收购 Github？
英文原文
$NVDA acquires hugging face for $12,930,300,000. Jensen stated the reason being: - 18M+ devs use Hugging Face - 3M+ models, 500k+ datasets, 1M+ apps - 200k+ companies use the platformm - support of open weight models this almost reminds me of $MSFT buying Github? https://t.co/8FhE7b9mU0
-
披露信息对晶圆代工伙伴分配影响重大。
@Jayden11129225 是的，因此从这一披露来看，对其晶圆代工合作伙伴分配的影响要重要得多。
英文原文
@Jayden11129225 Yes, so implications on their foundry parter allocations are a lot more meaningful from this disclosure.
-
AVGO CEO称激光器需求远超供应，验证行业景气度。
$AVGO CEO：“无论是对电吸收调制激光器(EML)还是连续波(CW)激光器的需求，都远远超过行业内的供应。” 这来自全球最大激光器供应商之一的相当重要的信息……即使他们将激光器产能扩大了3倍。 这对激光器公司来说是令人兴奋的验证。 https://t.co/l4alSsexMT
英文原文
$AVGO CEO: "Demand for lasers, whether it is EML lasers, CW lasers. Is far surpassing supply out there in the industry" Pretty material coming from one of the largest laser suppliers in the world... Even while they 3x laser capacity. Exciting validation for laser companies. https://t.co/l4alSsexMT
-
分析MRVL光学项目激光采购地传闻，质疑其合理性。
@Mohsen20Es 供应商关系不会公开，这就是为什么研究这些线索很有趣。 我发现 $MRVL 的光学项目(CPO)负责人在激光供应采购上考虑欧洲，这非常有趣。 这就像他们随机在供应链中点名非洲一样，（暗示这种说法不合常理或令人惊讶）。
英文原文
@Mohsen20Es Supplier relationships are not going to be public, which is why it's fun looking into these clues. I just found it very interesting $MRVL optical program lead had Europe in mind for laser supply sourcing. It's like if they randomly named dropped Africa for supply chains,
-
Marvell高管指出欧洲/日本是SiPH激光器供应来源。
$MRVL 光学工程高级副总裁兼首席技术官 Radha Nagarajan 在台湾半导体展上表示： 1. 以欧洲/日本作为硅光混合组件(SiPH)激光器供应来源的示例。 Marvell 的高管在举例 SiPH 模块的激光器供应来源时，脑海中想到的是欧洲，这非常有趣。 https://t.co/zB5pDH1XVC
英文原文
$MRVL SVP and CTO of Optical Engineering Radha Nagarajan, at Semicon Taiwan: 1. Names Europe/Japan for laser supply sourcing as examples for SiPH modules. Very interesting that Marvell's SVP had Europe in mind when giving examples of laser supply sourcing for SiPH modules. https://t.co/zB5pDH1XVC
-
指出市场忽视SIVE在CPO激光器的地位，且过度聚焦可插拔模块。
@dockery58 是的，共封装光学(CPO)的出货量甚至还没开始……大家都忘了 $SIVE 是主要的 CPO 激光器供应商，因为管理层过于聚焦无线/5G/固定无线接入(FWA)/卫星通信。像 $JBL 这样的可插拔模块(Pluggables)本应是 2028 年 CPO 放量的过渡桥梁，但不知为何成了主导焦点。
英文原文
@dockery58 Yeah, CPO volumes haven’t even started… everyone forgot that $SIVE was a major CPO laser supplier because management focused so much around wireless/5g/fwa/satcom. Pluggables like $JBL were supposed to be the bridge for 2028 CPO ramp but somehow became the dominant focus.
-
台湾半导体表现优异，Foci 快速复苏，质疑财报无视宏观。
值得注意的是，台湾的表现远远超越世界其他地区，如美国/欧盟。 Foci（$NVDA / $TSM FAU 供应商）股价从 400 -> 900 -> 400 -> 800，且在 7 月后恢复得极快。 不确定像 Etron 或其他公司的财报是否荒谬到完全无视宏观环境？https://t.co/WHzwj9n6e8
英文原文
Very interesting to notice Taiwan wildly outperform the rest of the World like US/EU. Foci ($NVDA / $TSM FAU supplier) went from 400 -> 900 -> 400 -> 800, and recovered insanely fast from July. Not sure if earnings like Etron or others are just absurd to not care about macro? https://t.co/WHzwj9n6e8
-
美国拨款1.74亿保障镓供应链，利好AA及Sojitz。
美国战争部宣布提供1.74亿美元融资以保障镓(Gallium)供应链安全。受益方包括： - $AA (Alcoa Corp)，因其冶炼厂生产镓。 - Sojitz (2768)，作为日本合资的下游采购合作伙伴。 很高兴看到美国在关键矿物供应链上的投资。 https://t.co/fyi203MH2B
英文原文
US Department of War announces $174M financing for securing Gallium Supply Chains. Beneficiaries are: - $AA (Alcoa Corp) for Gallium production at its refinery. - Sojitz (2768) as the Japanese JV offtake partner Good to see US critical minerals supply chains investments. https://t.co/fyi203MH2B
-
博主对相机公司入局光模块市场及YouTuber持股表示困惑。
@GreekFreakHoops 啊……这让事情变得比那家“恐龙相机公司”进入1.6T光收发器市场还要令人困惑。我正在看新闻，看来是因为那位YouTuber持有8.5%的股份？
英文原文
@GreekFreakHoops ah... that makes things even more confusing than the dinosaur camera company entering the 1.6T optical transceiver market. I'm reading the news now, apparently it was because of the Youtuber's 8.5% stake?
-
GPRO拟并购光子公司进军AI光模块，博主对此跨界竞争表示意外。
什么……$GPRO 在宣布计划与光子学公司“Starman Optical”合并后上涨了 79.17%。 旨在为 AI 数据中心制造美国本土的 800G/1.6T 光收发器。 GoPro 试图与 $AAOI 竞争，这绝对不在我的雷达范围内。 https://t.co/5z2XXHaxvs
英文原文
What... $GPRO is up 79.17% after planning to merge with a photonics company "Starman Optical" To build US 800G/1.6T optical transceivers for AI data centers. GoPro trying to compete with $AAOI was definitely not on my radar. https://t.co/5z2XXHaxvs
-
分析英伟达投资偏好及半导体、机器人供应链关键瓶颈与趋势。
@Bejewele_din_Fr - 英伟达($NVDA) 喜欢投资专用集成电路(ASIC) 公司 - 内存/多层陶瓷电容器(MLCC)/ABF载板/绝缘体上硅(SOI) 因长期可见性而令人兴奋(brrr) - 传统DRAM 因中国内存向高端转移而令人兴奋(brrr) - 中欧就企业问题展开博弈 - 共封装光学(CPO) 时间线保持不变 - 机器人领域的行星滚柱丝杠很酷
英文原文
@Bejewele_din_Fr - Nvidia likes investing in ASIC companies - Memory/MLCC/ABF/SOI brrr from long term visibility - legacy dram to brrr from Chinese memory moving to higher end - China EU fighting over companies - cpo timeline in tact - planetary roller screws in robotics are cool
-
梳理AI供应链瓶颈：内存、ABF、CPO及地缘政治动态。
一些我感兴趣的TLDR新闻摘要： 1. $GOOGL（漏看了）也联合$NVDA投资了联发科，参与其39亿美元融资。Google/Nvidia在投资眼光上真是“山羊”（顶级），从$SPCX到Anthropic都是如此。 2. 三星电机披露了7.789亿美元的MLCC（多层陶瓷电容器）订单；$CAPA（新MLCC ETF）也为无法访问韩/日/台市场的投资者提供了敞口。但关键在于，每个加速器所需的MLCC数量增加，可能预示着未来比人们预期更极端的瓶颈。 3. 中国渠道调查显示：行星滚柱丝杠、6D力传感器和灵巧手难以实现国产替代。其中行星滚柱丝杠的国产替代率最低。例如$SHA0.DE做行星滚柱丝杠。只是听听护城河挺有意思。 4. 据报道，三星约70%的内存产能已锁定至2031年/商品DRAM也陷入短缺 - 首尔经济日报。TrendForce预计2026年Q3服务器DRAM合约价环比上涨13-18%。SK海力士也公开表示内存短缺将持续至2030年。 5. $CXMT已开始小批量生产HBM3E，并据称实现了LPDDR6的大规模量产。这对传统DRAM短缺的暗示更为积极。 6. 思科在台湾SEMICON表示共封装光学(CPO)已准备好大规模部署；可插拔模块仍将是市场主流。$MRVL也表示客户正积极规划早在2027年扩大光学部署。 7. 闻泰科技 -> 收购安世半导体 -> 荷兰政府实际上剥夺了闻泰的控制权 -> 闻泰正通过中国法院反击。但安世3亿美元资产被冻结。当然还有更多地缘政治戏剧，总是很娱乐。$AAPL也明确指控OpenAI销毁关键证据...（更多乐子） 8. 关于SOI（绝缘体上硅）衬底$SOI的ASP（平均售价）上涨描述不多，但他们现在有10+客户参与LTA（长期协议），$2亿营收被称为“底线”，评论称客户更关心获得晶圆而非制造地。所以基本上随着硅光子技术在2027年起飞，现在进入产能预留阶段。 9. 一些渠道调查显示高功率MOSFET和大面积TVS二极管的交货期高达~52周。（$STM据报道也在8月23日实施了2026年第三次涨价） 10. 传统封装产能可能在2027年面临超过20%的缺口，“引线键合设备”成为主要瓶颈，本周早些时候已覆盖过，受益者如... 11. 声称所有六大ABF供应商2026年订单已满，交货期12-14个月，部分供应商拍卖剩余产能；$NVDA、$AMD和超大规模客户锁定ABF产能至2028年，并推动供应商规划2029-30年扩产。像Kinsus (3189)、Unimicron (3037)、南亚PCB (8046)、Ibiden (4062)等名字大概都很开心。 12. CCL（覆铜板）主导的层压板产品再次涨价10%-20%，今年第7次涨价（如建滔）。行业估计HVLP4铜箔短缺在2026年约48%/2027年43%，而Low-Dk/T-Glass交货期已延长至30周以上。 13. “(GaAs)衬底材料和砷化镓外延源材料在2026年第三季度也面临新一轮涨价，据报道比限制初期高出5到6倍。”“中国最近放宽了一些衬底出口管制”，但据digitimes称“InP衬底严重短缺”。只是些$AXTI相关更新。 14. $AMD和$AVGO据报道预订了PowerTech计划的大部分FOPLP产能，早于2027年中大规模量产。 15. $INTC考虑使用$INTC代工生产HBM4E基底芯片。但$SKHY直接回应称“部分内容与事实不符”。 16. “OpenAI囤积数万台$AAPL Mac mini和Mac Studio”用于强化学习。Anthropic也从$AMZN租赁Mac mini单元。Tim Cook退休有点可惜，但老哥表现确实“山羊”（顶级）。只是些随机有趣的笔记。
英文原文
Just some TLDR news I found interesting: 1. $GOOGL (missed this) also invested in Mediatek alongside $NVDA for their $3.9B raise. Google/Nvidia are goats with investments you look at $SPCX to Anthropic. 2. Samsung Electro-Mechanics discloses $778.9M MLCC deal / $CAPA (new MLCC ETF) also gives exposure now for folks without Korea/Japan/Taiwan market access. But thing is more accelerators x MLCCs per accelerator might indicate a more future extreme bottleneck than people expect. 3. China channel checks state: planetary roller screws, 6D force sensors, and dexterous hands resist domestic substitution. Then planetary roller screw as the top with lowest domestic substitution rate. eg. $SHA0.DE for planetary. Just interesting to hear moats I guess. 4. Samsung reportedly has ~70% of memory production capacity committed through 2031 / commodity DRAM has also fallen into short supply - Seoul Economic Daily. TrendForce expects server DRAM contract prices to rise 13-18% QoQ in 3Q26 SK Hynix also publicly stated the memory shortage lasts through 2030. 5. $CXMT has begun producing HBM3E in small quantities and reportedly claimed mass production of LPDDR6. Implications are even more positive for legacy DRAM shortages. 6. Cisco at SEMICON Taiwan said CPO is ready for mass deployment; pluggable modules will remain the market mainstay. $MRVL also said customers are actively planning scale-up optics deployments as early as 2027 7. Wingtech -> bought Nexperia -> Netherlands effectively removed Wingtech’s control -> Wingtech is fighting back through Chinese courts. But Nexperia got $300m of their assets frozen. More geopolitical drama of course, always entertaining. $AAPl also explicitly accused OpenAI of destroying crucial evidence... (for more fun) 8. Not much description around ASP hikes regarding SOI substrates $SOI, but they now have 10+ customers involved with LTAs, $200m revenue described as a "floor", comments about customers more concerned about obtaining wafers than where they are manufactured. So basically moving into capacity reservation now stages as silicon photonics takes off in 2027. 9. Some channel checks put lead times for high power MOSFETs and large area TVS diodes as high as ~52 weeks. ($STM also reportedly implemented its 3rd 2026 price increase on August 23) 10. Traditional packaging capacity could face a deficit exceeding 20% by 2027" with "wire bonding equipment" emerging as the primary bottleneck", covered this earlier this week with beneficiaries like 11. Claims that all six major ABF suppliers are booked for 2026, with 12-14 month lead times and some suppliers auctioning remaining capacity; $NVDA, $AMD and hyperscalers locking ABF capacity through 2028 and pushing suppliers to plan 2029–30 expansions. Names like Kinsus (3189), Unimicron (3037), Nan Ya PCB (8046), Ibiden (4062), etc are all probably happy. 12. CCL leading laminated board products increased by another 10%-20%, 7th price hike this year (eg. Kingboard). Industry estimates put the HVLP4 copper-foil shortage at ~48% in 2026 / 43% in 2027, while Low-Dk/T-Glass lead times have stretched beyond 30 weeks. 13. "(GaAs) substrate materials and gallium epitaxial source materials are also facing another round of price increases in the third quarter of 2026, reportedly 5 to 6 times higher than at the early stage of the restrictions." "China has recently eased some substrate export controls" but there's a "severe InP substrate shortage" per digitimes. Just some $AXTI related updates 14. $AMD and $AVGO reportedly booked most of Powertech's planned FOPLP capacity ahead of mid-2027 mass production 15. $INTC considering $INTC foundry for HBM4E base dies. But $SKHY responded directly that "some of the content is different from the facts". 16. "OpenAI Hoarding Tens Of Thousands Of $AAPL Mac mini and Mac Studio" for reinforcement learning. Anthropic has been renting mac mini units from $AMZN as well. Kinda sad Tim Cook retired, but dude had a goated run. Just random interesting notes
-
Anthropic签350亿算力大单，前沿模型需求远超供给能力。
Anthropic 与由 $NVDA 支持的 Lambda 签署了价值 350 亿美元的 AI 算力协议。 $HUT / 数据中心，搭配 Lambda / GPU。 前沿模型的需求正在迅速超越生态系统构建容量的能力。 看起来 Anthropic 正在锁定更多算力……下一家被选中的新云/colo 玩家是谁？ https://t.co/49yvFKjr49
英文原文
Anthropic signs a $35B AI compute deal with $NVDA backed Lambda. $HUT / DC, with Lambda / GPUs. Frontier model demand is rapidly outpacing the ecosystem's ability to build capacity. It also looks like Anthropic is securing more compute... Which neocloud/colo player is next? https://t.co/49yvFKjr49
-
英伟达投资联发科巩固推理地位，或促使AMD与博通结盟。
没错，$NVDA 投资 35 亿美元进入联发科(Mediatek)。 这感觉像是英伟达(Nvidia)正在扮演下一代专用集成电路(ASIC)赢家如 $MRVL / 联发科(Mediatek)的“造王者”，其二级效应是侵蚀 $AVGO 的地位。 这一策略 playbook 类似于他们此前针对像 $NBIS 和 $CRWV 这样的新云厂商(Neoclouds)（相对于超大规模云项目）所采取的做法。 帮助培育下一代领导者，然后使它们在财务和战略上与英伟达保持一致。 我认为这很精彩，但此类举动可能会促使 $AMD、$AVGO 和某些超大规模云厂商(hyperscalers)走得更近。 无论如何，此举巩固了英伟达在推理(inference)领域的战略地位（此前超大规模云厂商的 ASIC 曾是长期的看空论点）。
英文原文
Yep, with $NVDA investing $3.5B into Mediatek. It does feel like Nvidia is playing kingmaker with the next generation of ASIC winners like $MRVL / Mediatek, with a second order effect of eroding $AVGO position. Playbook feels similar to what they did with Neoclouds like $NBIS and $CRWV (vs. hyperscaler cloud programs) Help create the next leaders, then make them both financially and strategically aligned with Nvidia. I do think it’s brilliant, but moves like this might push $AMD, $AVGO, and certain hyperscalers closer together. Regardless this move cements Nvidia’s strategic position in inference (where hyperscaler ASICs were a former long-term bear case).