供应链分析
产业链结构、上下游与瓶颈环节研究 · 共 1315 条 · 滚到底自动加载更多
-
解析InP供应链双瓶颈，指出AXTI因掌控全产业链而受益。
我认为细微差别在于存在两个不同的瓶颈。一个是原料/精炼加工，另一个是磷化铟(InP)衬底。住友/ $COHR 处于其中一个环节，但同样受到如7N非标准铟等原料定价的影响。我个人最看好 $AXTI 的原因是他们身处两个不同的瓶颈环节，因为他们拥有整个供应链并受益于价格上涨。全球大部分上游原料来自Vital或由AXT衍生。
英文原文
I think the nuance is that there’s two different bottlenecks. One is feedstock/refinery processing. The other is InP substrates. Sumitomo/ $COHR sits in one of them but are affected by feedstock pricing like 7n indium nonstandard. The reason I personally like $AXTI the most is because they sit in two different bottlenecks since they own the entire supply chain and benefit from price hikes. Most of the worlds upstream feedstock come from Vital or are derived from AXT
-
分析美伊冲突下开心果供应链垄断及石油国防股对冲机会
如果美国入侵伊朗，最有趣的受益者将是：开心果。绝非玩笑。你知道美国（加州）和伊朗在开心果生产上形成了虚拟双头垄断吗？这类似于 $AXTI 在磷化铟(InP)领域的情况，两者控制了全球约70-80%的供应。如果加州成为垄断者，$JBSS 等公司可能会因开心果价格上涨而受益。不幸的是，没有开心果期货，所以我没有建立任何头寸，但这有趣的事实让我觉得好笑，所以想分享出来。我稍后会写一篇关于更细微的二阶效应和潜在做多标的的文章。美国打击似乎很可能，因此标准的石油股如 $CVX 或 $XLS 以及国防股如 $AVAV 或 $ITA 可能是不错的做多/对冲选择。
英文原文
The funniest beneficary if US invades Iran was: Pistachios. Not even joking. Did you know US (California) and Iran operate a virtual duopoly in pistachio production? This is the $AXTI InP situation as the two control ~70-80% of the world’s supply. Companies like $JBSS might benefit from Pistachio prices going up (if California becomes a monopoly) Unfortunately there’s no Pistachio Futures, so I’m not taking any positions but I just found this fun fact amusing, so wanted to share. I’ll do another writeup on more nuanced second order effects and potential longs soon. US strikes seem likely so standard oil like $CVX or $XLS and Defence like $AVAV or $ITA might be good longs/hedges here.
-
建议OpenAI收购树莓派以通过预装API实现硬件级锁定。
说实话，@sama 应该收购树莓派(Raspberry Pi)，并在其硬件推广中销售预装 OpenClaw（使用 OpenAI API）的设备。 这是实现类似苹果(Apple)那样硬件级锁定效应的极低成本（<11亿美元）。 MiniMax 及其竞争对手将从树莓派和 Mac mini 上的 OpenClaw 类部署中获益最多。出厂即预集成 ChatGPT 将提高采用率，并增加其他竞争者的进入摩擦。
英文原文
Honestly, @sama should buy Raspberry Pi and sell predeployed OpenClaw using OpenAI APIs in their hardware push. Very cheap price to pay (<$1.1B) for hardware level lock-in like what Apple does. Minimax and their competitors benefiting the most from OpenClaw type deployments on Pi and Mac minis. Having a ChatGPT pre integrated off the bat would increase adoption and increase friction for others.
-
媒体引用其树莓派AI硬件论点，看好代理式AI本地编排趋势。
很酷，《卫报》引用了我关于树莓派(Raspberry Pi)的论点。 先是路透社，然后是彭博社、经济时报，现在轮到《卫报》了！ OpenClaw🦞真的病毒式传播，彻底改变了关于树莓派($RPI)和MiniMax(HKG: 0100)的叙事。 随着OpenAI的收购，看起来代理式AI(agentic AI)和本地编排(local orchestration)硬件只会从这里开始增长。 我只是在评论人们视而不见的结构性转变，正如我们进入AI应用的新前沿。
英文原文
It’s cool “The Guardian” cited my thesis on Raspberry Pi. First Reuters, then Bloomberg, Economic Times, and now The Guardian! OpenClaw🦞 really went viral and transformed the narrative on Raspberry Pi ( $RPI ) and Minimax (HKG: 0100). Following the OpenAI acquisition, it looks like agentic AI and local orchestration hardware is only growing from here. I’m just commentating on structural shifts people miss in plain sight, as we enter a new frontier in AI applications.
-
澄清AI代理编排需隔离硬件环境，反驳本地运行LLM的误解。
@coinsearch71105 实际上，市场对隔离硬件环境的需求非常高。具体来说，是AI代理编排(AI agent orchestration)。人们有一个误解，认为你在本地运行大语言模型(LLM)推理。OpenClaw是通过API与Minimax或Claude进行交互的。
英文原文
@coinsearch71105 There's actually very high demand for isolated hardware environments. In specifically, AI agent orchestration. People make the misconception that you run LLM inference locally. OpenClaw interfaces with an API to Minimax or Claude.
-
分析存储芯片期权隐含波动率被低估，预期将维持高位。
我指的是两年期期权，而非30天到期(30DTE)的期权。但正如你所指出的，短期内波动性普遍更高。我展示的最可能情景是：三星(Samsung)和SK海力士(SK Hynix)各自的隐含波动率(IV)在55-75之间，且占指数权重的50%。做市商(MMs)此前定价基于5-10年的历史数据，当时指数停滞且预期均值回归，而非三星/SK海力士在新存储交易中的波动性。我使用多个期权数据源，Robinhood因对散户最直观易引而常被提及。但如果IV回落至32%，我就错了；但在我看来，由于波动率和上行空间存在巨大定价偏差，IV很可能维持高位。
英文原文
I’m talking about 2 year dated options, not 30dte. But as you’re referencing, there’s more short term volatility in general. Most likely scenario I’m showing is Samsung/SK Hynix individually are around IV 55-75 and make up 50% of the index. MMs were pricing in 5-10 year historical values where the index was stagnant + expecting a reversion to mean, rather than the new memory trade volatility in Samsung/Sk Hynix. I use multiple option sources, Robinhood is the easiest to cite visually to retail. But if IV drops back to 32% I would be wrong but it will likely remain much higher since that’s a huge mispricing of volatility/upside from my view.
-
利用韩国指数期权定价错误，做多$EWY看涨期权以捕捉存储周期波动率扩张。
指数下跌-2.46%。整个期权链全线飘红+13-20%。 这就是当你发现做市商在期权链中存在定价错误时会发生的情况。 话虽如此，隐含波动率(IV)回升至更合理的38-39%，但个别组件（SK海力士、三星）的波动率可能仍有几个百分点的偏差。 （引用内容翻译）： 我发布在“淋浴思考”频道的交易思路： 韩国指数波动率套利并利用布莱克-斯科尔斯模型。 $EWY 看涨期权似乎存在定价错误。 这是贝莱德旗下的韩国指数，主要由存储芯片（三星电子、SK海力士）构成。 尽管该指数被定价为普通指数的隐含波动率，但个股每日波动2-5%+，且1年涨幅达136.25%。 三星波动剧烈。SK海力士波动剧烈（例如预估65%-80%）。 但通过指数组合后的定价远低于低贝塔值的 $GOOGL (37.33%) 和 $AMZN (39.12%)，隐含波动率仅约32%。 我观察 $EWY 一段时间，它确实看起来波动很大。 至于定价，我猜测做市商(MM)基于历史平均值（5-10年）定价隐含波动率，当时韩国指数完全持平。他们预期两年后的看涨期权会回归均值。 但这种波动率应成为新常态，因为市场正在定价新的存储超级周期（例如 $TSM 从30% IV升至46.2% IV）。 看涨期权将从三星+SK海力士带动指数上涨中受益。 主要好处是你能获得 $KORU 无法提供的 Vega 扩张。 你也无法像美国个股那样获得这种期权做市商的压盘效应，因为这是韩国国家指数且期限较长。 简而言之：个别组件SK海力士+三星具有高波动性。 它们基本占指数的一半，但指数期权以低波动率定价，或许是因为基于过去5-10年的历史数据。 看涨期权将从未正确定价的 Vega 扩张中受益，因为做市商的前瞻波动率估计过于锚定于历史已实现波动率，而过去5-10年 $EWY 的波动率很低。
英文原文
Index down -2.46%. The entire option chain green +13-20%. This is what happens when when you find mispricing in option chains by market makers. That being said it’s a more respectable 38-39% IV, but maybe few percent off (SK Hynix, Samsung) individual components volatility still.
-
树莓派在低成本智能体集群中具优势，企业更看重AI自动化而非个人助理。
不完全是，取决于具体用例。如果你在做智能体集群(agentic swarms)或智能体营销，树莓派(Raspberry Pis)是更优选择，因为它们更便宜。关键在于连接互联网。像回复iMessage这样的“个人助理”用例，潜在市场规模(TAM)较小。批量购买设备的人正在寻找更广泛的商业应用，即AI智能体可以在互联网上进行自动化操作。
英文原文
Not exactly, depends on use case. If you're doing agentic swarms, agentic marketing, Raspberry Pis are superior choice since they're cheaper. It's mainly connection to internet. The "Personal Assistant" use cases like responding to an imessage are low TAM. The people who bulk-buying devices are looking for broader business applications where AI agents can do automation across the internet.
-
OpenClaw模型在低成本硬件运行提升了树莓派实用性。
软银(Softbank) + $ARM 实际上是树莓派(Raspberry Pi)的最大股东之一！但正如你提到的，苹果(Apple)和树莓派都使用 ARM 指令集架构(ISA)。大多数人起初都在囤积苹果设备，但自从 OpenClaw 模型能够在低成本硬件上运行以来，树莓派的实用性开始增加。
英文原文
Softbank + $ARM is one of Raspberry Pi's largest owner actually! But yeah as you mentioned, Apple and Raspberry Pi use arm isa. Majority of people started off hoarding Apple devices, but ever since openclaw models were able to be run on lower cost hardware, raspberry pi started having more utility.
-
云环境限制AI智能体应用，建议本地裸金属部署及环境隔离。
几点说明： - 像 AWS 这样的云租户操作系统（Cloud TOS）会阻止许多智能体（Agentic）相关应用，因此你需要在本地运行服务器。 - 平台能识别虚拟机（VMs）并检测虚拟机管理程序（Hypervisor）指纹。对于 AI 自动化而言，裸金属（Baremetal）是绕过检测的痛点所在。 - 每个智能体需要隔离环境。给予 openclaw 根（root）权限存在风险，可能会摧毁多个环境。
英文原文
Few things: - Cloud TOS like AWS prevents many agentic related applications, so you need to run servers locally - Platforms know VMs and detect hypervisor fingerprinting. Baremetal pis for AI automation bypass detection - isolated environments for each agent. there's dangers in giving openclaw root access nuking multiple environments
-
OpenClaw通过API编排，无需在隔离设备上运行本地LLM。
@VJNCapital OpenClaw 通过 API 进行编排，你不需要在这些隔离设备上运行推理或本地大语言模型(LLM)模型。
英文原文
@VJNCapital OpenClaw does orchestration through APIs, you don't run inference or local LLM models on these isolated devices.
-
AI智能体爆发将推动树莓派TAM从教育向AI领域激增。
@dirtyculture OpenAI 刚刚收购了 OpenClaw……我不太认为会是这样。 从 $META 的 Manus 到 OpenAI 的 OpenClaw、Picoclaw 等，将出现大量其他智能体(agentic)变体。 树莓派(Raspberry Pi)的潜在市场规模(TAM)正因从教育领域扩展到 AI 智能体而飙升。
英文原文
@dirtyculture OpenAI just bought OpenClaw... Don't quite think it will. There's going to be a ton of other agentic variants from $META Manus to OpenAI OpenClaw, Picoclaw, and others. TAM of Raspberry Pi just shot through the roof from education to AI agents.
-
树莓派因AI代理需求被囤积，类比英伟达GPU转型。
@glr_1990 树莓派(Raspberry Pi)从用于测试的1-2台教育设备，变成了被囤积用于AI代理部署。这让我想起$NVDA GPU从“游戏玩家”转向AI训练/推理的过程。
英文原文
@glr_1990 Raspberry Pi went from 1-2 educational devices for testing to hoarded for AI agent deployments. Reminds me of $NVDA GPUs going from "gamers" to AI training/inference.
-
DeepSeek架构成LLM范式存疑，若性能获证将利好美股AI企业。
@beauty_oe DeepSeek 的架构是否会成为大语言模型(LLM)的下一代范式，仍然是一个巨大的不确定因素。近期，市场对中国 AI 模型的期待高涨，如果其在现实社会中的性能得到验证，可能对包括 $ALAB 在内的美国企业产生积极影响。
英文原文
@beauty_oe DeepSeekのアーキテクチャがLLMの次世代パラダイムとなるかは、依然として大きな不確定要素です。 昨今、中国のAIモデルへの期待が高まっており、実社会での性能が証明されれば、$ALABをはじめとする米国企業にとってもプラスに働く可能性があります。
-
看好DeepSeek带动CXL需求，推荐关注ALAB、AVGO和MRVL。
如果你看好带有 Engram (v4) 的 DeepSeek 大语言模型，做多 CXL (Compute Express Link) 似乎是个不错的想法。 - $ALAB (Leo) - $AVGO - $MRVL 是美国市场的三大巨头。 尤其是 Astera。 如果它成为一种新的架构范式，像 Astera 这样的公司可能会销售数百万个全新的 CXL 控制器来管理外部内存池。 话虽如此，许多关于 DeepSeek 的报道通常夸大其词（内部基准测试与实际表现不同），它可能仅被用于离线批处理。 近期的全面下跌为这三家公司提供了良好的机会。
英文原文
If you're bullish on DeepSeek LLMs with Engram (v4), going long on CXL seems like a idea. - $ALAB (Leo) - $AVGO - $MRVL are your big three in the US. Especially Astera. If it becomes a new architectural paradigm, companies like Astera maybe would sell millions of net-new CXL controllers to manage the external memory pools. That being said, lot of reports are usually sensational around DeepSeek (internal benchmarks perform different than IRL), and might just be relegated to offline batch processing Recent drop across the board presents a good opportunity for these three.
-
解析Picoclow使用树莓派绕过检测的技术细节及风险。
是的，这是一个合理的观点，我在发帖前已经研究过了。 新的 Picoclaw/压缩变体使用的内存少于 10 MB，而且不在本地运行推理。 这就是为什么我上周五才注意到树莓派（Raspberry Pi）的角度，当时 Picoclaw 的帖子正在病毒式传播。 这里有几个细微差别。平台知道虚拟机（VM）并检测虚拟机监控程序指纹。裸机树莓派用于自动化以绕过检测，因为它们看起来只是 Arm 设备，这很可能是一个主要原因。 此外，据我所知，当你给予 Openclaw AI root shell 访问权限时，也存在可能的交叉污染。
英文原文
Yep that's a fair point that i looked into before posting. New picoclaw/compressed variants use less than 10 mb of ram + doesnt run inference locally though. Hence why I only saw the raspberry pi angle last Friday when there were viral posts from picoclaw. So few nuances. Platforms know VMs and detect hypervisor fingerprinting. Baremetal pis for automation bypass detection since they just look like Arm devices, which is likely a big reason. There's also possible cross contamination when you give openclaw ai root shell access from what i've heard
-
群联CEO驳斥中国存储倾销论，看好AI存储结构性需求。
群联电子(Phison) CEO 采访中的一个有趣评论。针对中国存储厂商（长江存储 YMTC 和长鑫存储 CXMT）将倾销市场的谣言，CEO 潘汉德表示：“说这话的人是在说梦话”。$SNDK、$MU 以及存储板块似乎看不到供应短缺的迹象。AI 对存储的需求看起来是结构性的且呈指数级增长。
英文原文
Interesting comment from Phison's CEO interview was this. When addressing the rumor that Chinese memory makers (YMTC and CXMT) will flood the market: Pua says: "The people saying that are talking in their sleep/dreaming". $SNDK, $MU, and the memory trade appear to have no shortage in sight. The memory demand for AI looks structural and exponentially growing.
-
存储需求结构性短缺，推理瓶颈在存储，中国产能无外溢。
群联电子(Phison) CEO关于存储与投资框架的访谈摘要： “收过路费者”（Toll Collectors）： - 美光(Micron) ($MU) - SK海力士(000660.KS) - 三星电子 - 西部数据(Western Digital) ($WDC) - $SNDK T2层级： - $MRVL - $SIMO - 群联电子(Phison Electronics) 随着AI向边缘端迁移，设计连接存储与计算逻辑/软件控制器的公司将捕获巨大价值。 T3层级： - 纯存储(Pure Storage) ($PSTG) - NetApp ($NTAP) - 希捷(Seagate) ($STX) 随着Vera Rubin推理服务器推出，键值缓存(KV Cache)和数据生成的爆发将触发针对数据中心存储密度和高容量企业级固态硬盘(Enterprise SSDs)的硬件升级周期。 有趣的是：$EBAY（翻新电子产品）可能成为受益者。 - 做空/规避低毛利消费硬件。 - 做空/规避未对冲的汽车/IoT制造商。 主要Alpha观点： - “三年预付”现金流：存储晶圆厂要求3年现金预付款以保障供应。 - 推理瓶颈在于存储而非GPU：单批次1000万台$NVDA Vera Rubin平台需每台20+TB SSD，仅此项就将消耗去年全球NAND产能的20%。 - “中国供应过剩”看空论调已死： Pan完全驳斥了关于长江存储(YMTC)和长鑫存储(CXMT)的观点。中国内部AI需求巨大，将瞬间消化100%国内产量。不会有廉价中国存储流入全球市场来拯救西方硬件OEM。 访谈TLDR： 存储需求是结构性的。供应端无结束迹象。$INTC CEO上月已确认此点。
英文原文
TLDR of Phison CEO interview on Memory and Investment Framework: "Toll Collectors": - Micron ( $MU ) - SK Hynix (000660.KS) - Samsung Electronics, - Western Digital ( $WDC ) - $SNDK. T2: - $MRVL - $SIMO - Phison Electronics Companies that design the logic/software controllers connecting memory to compute will capture massive value as AI moves to the edge. T3: - Pure Storage ( $PSTG ) - NetApp ( $NTAP ) - Seagate ( $STX) As Vera Rubin inference servers roll out, the explosion in KV Cache and data generation will trigger a massive hardware upgrade cycle specifically focused on data center storage density and high-capacity Enterprise SSDs. Hilariously: $EBAY (refurbished electronics), might be a beneficiary. - Short / Avoid Low-Margin Consumer Hardware. - Short / Avoid Unhedged Auto/IoT Makers Main alpha points: - The "3-Year Prepayment" Cash Flow. Memory foundries are demanding 3 years of cash prepayments to guarantee supply. - The Inference Bottleneck is Storage, Not GPUs. A single 10-million-unit run of $NVDA Vera Rubin platform requires 20+TB of SSD per unit, which alone would consume 20% of last year's global NAND capacity. - The "Chinese Supply Glut" Bear Thesis is Dead: Pan entirely dismisses this point around YMTC and CXMT. China’s internal AI demand is so massive that it will instantly swallow 100% of its domestic production. No cheap Chinese memory will leak into the global market to rescue western hardware OEMs. TLDR from the interview: Memory demand is structural. No supply end in sight. $INTC CEO confirmed this last month.
-
业内认为长存长鑫扩产不会导致存储价格崩盘，反驳媒体观点。
写得很好！我最喜欢这条评论： 针对长存(YMTC)和长鑫(CXMT)扩产可能冲击市场并压低价格的担忧，Pua用一句话回应：“说这话的人在做梦。” 有趣的是，英特尔(Intel)CEO和行业内部人士都持相同观点，而媒体却在呼吁存储价格崩盘。
英文原文
Great writeup! This was my favorite comment: In response to concerns that capacity expansion by YMTC and CXMT could shock the market and drive down prices, Pua answered in one line: "The people saying that are dreaming." Interesting how Intel’s CEO and industry insiders are all saying the same thing, while media are calling for a memory price crash.
-
Clarity Act 禁止稳定币收益，保护银行低息模式，或致 USDC 流动性枯竭。
这是 Genius 上 Clarity Act 扩展带来的稳定币相关收益。这是银行的“毒丸计划”，因为它从法律上禁止了任何竞争。这意味着将 USDC 存入稳定币新银行（Stablecoin neobank）-> 公司提供的 3-5% 收益被禁止。如果稳定币本身以及新银行/交易所/金融科技都没有收益，可能会导致 USDC 存款外流并耗尽市场流动性。银行支票账户平均利率约为 0.39%。正如你提到的，像 $SOFI 这样的新金融科技是例外。银行提供“高收益储蓄”，但很多时候，他们隐藏条款，6 个月后降低利率 -> 降至 0.1% 并迫使你开设新账户。如果其他银行/交易所等提供真正的百分之几的 USDC 存款国债利率并保持恒定，将颠覆银行现有的低/零费用支票账户模式。消费者没有获得真正的利率，所有利润都流向了银行。
英文原文
This is stablecoin related yield with the Clarity Act expansion on Genius. This is the Bank's poison pill as it just legally bans any competition. Meaning depositing USDC into a Stablecoin neobank -> the company giving 3-5% yield off that is banned. Having no yield from both the stablecoin itself + neobanks/exchanges/fintechs would likely cause a deposit flight out of USDC and drain liquidity from the market. Bank checking accounts are in average are ~.39%. There are exceptions with new fintechs like $SOFI as you mentioned. Banks offer "high yield savings" but a lot of times, they hide hidden clauses where they lower those after 6 months -> goes to 0.1% and make you open new accounts. If every other Bank/Exchange or others offered true few percent treasury rates from USDC deposits and made it constant, it would disrupt the low/zero fee checking account models banks have. Consumers aren't getting true rates, and any profits just go to banks.
-
OpenAI警告国会：电力供应是AI竞争核心瓶颈，需投资美国能源。
OpenAI 就 Deepseek 蒸馏问题向国会发送了一份备忘录： “维持美国在人工智能领域的优势，取决于我们能否可靠地大规模生成和输送电力。” 电力输送 - $VRT, $ETN, $PWR, $WMB, $KMI 一级能源供应商：$CEG, $VST, $TLN, $GEV, $NEE, $BEPC, $D 电网能源/储能 - $TSLA, $FLNC, $NRGV, $BE 能源：$TE, $FSLR, $NRG 这对这些公司来说是一个被重申的顺风因素。 而对于那些已经锁定吉瓦(GW)级容量的公司，如 $IREN, $NBIS, $WULF 和 $CIFR，则存在二级顺风效应。 备忘录的核心议题围绕知识产权盗窃和国家安全问题。但关于维持优势的最大警告在于能源。 OpenAI 警告国会，2024年中国新增了429吉瓦(GW)的电力容量，这超过了美国整个电网的三分之一，也超过了全球电力增长的一半。他们认为，如果没有对美国电网进行激进扩张，中国“蛮力”式的能源建设最终将使其超越西方的AI能力。 光子学(Photonics)、先进封装(Advanced Packaging)和存储(Memory)是目前增长最快的三个瓶颈。然而，OpenAI 明确警告美国政府，谁能产生最多的电力，谁就能赢得AI竞赛。 他们的信息是：投资美国能源。
英文原文
OpenAI sent a memo to congress regarding Deepseek distillation: "Sustaining the American advantage on AI depends on depends on whether we can reliably generate and deliver power at scale." Power Delivery - $VRT, $ETN, $PWR, $WMB, $KMI Tier-1 Energy Providers: $CEG, $VST, $TLN, $GEV, $NEE, $BEPC, $D Grid-Energy / Storage - $TSLA, $FLNC, $NRGV, $BE Energy: $TE, $FSLR, $NRG This is a tailwind reiterated for these companies. And there's a second-order tailwind for companies that already secured GW capacity like $IREN, $NBIS, $WULF, and $CIFR. The core issue of the memo was around IP theft and national security issues. But the largest warning about sustaining an advantage was Energy. OpenAI warning Congress that in 2024, China added 429 Gigawatts (GW) of new power capacity, which was more than a third of the entire US grid and more than half of global electricity growth. Without a radical expansion of the American power grid, they believe China’s "brute force" energy buildout will eventually allow them to surpass Western AI capabilities. Photonics, Advanced Packaging, and Memory are three fastest growing bottlenecks right now. However, OpenAI explicitly warned the U.S. government that whoever generates the most power wins the AI race. Their message: Invest in American Energy.
-
科技巨头负债投入AI基建，资金流向英伟达等上游供应商。
世界上最富有的公司正在为AI基础设施建设而负债。 尽管拥有数千亿美元的净利润，从 $AMZN 到 $GOOGL 的公司都大幅增加了资本支出(capex)： 以至于部分公司预计将在2026年出现净现金为负的情况。 以下是结果及受益者： 亚马逊 ( $AMZN ) 2025年：+$460亿 2026年（预估）：+$110亿 Alphabet ( $GOOGL ) 2025年：+$803亿 2026年（预估）：+$130亿 Meta Platforms ( $META ) - 净债务(Net Debt) 2025年：+$229亿 2026年（预估）：-$70亿（预计转为净债务状态） 微软 ( $MSFT ) 2025年：+$492亿 2026年（预估）：+$590亿 甲骨文 ( $ORCL ) 2025年：-$980亿（净债务） 2026年（预估）：-$1150亿 微软似乎处于最安全的位置。虽然亚马逊和谷歌主要依靠运营收入来资助AI基础设施建设。 然而，那巨大的现金流缓冲已经消失。甲骨文和Meta似乎正在通过负债来推动建设，尽管 $META 实现了惊人的运营收入数字。 现在，资金流向了哪里？ 运营收入预测： 1. $NVDA - 美国 $1350亿 -> $1865亿 -> $2401亿 运营收入 2. 三星 - 韩国 ~$302亿 -> ~$1700亿 -> ~$2267亿 3. SK海力士 - 韩国 ~$327亿 -> ~$1240亿 -> ~$1610亿 4. $TSM - 台湾 ~$622亿 -> ~$863亿 - ~$1033亿 – $1053亿 5. $AVGO - 美国 ~$415亿 -> ~$628亿-$651亿 - ~$847亿-$933亿 当然，这些是基于分析师预测的粗略估计。 然而，从总体趋势来看，这看起来像是一种杠杆式赌注，即AI基础设施建设完成后将带来自由现金流(FCF)的红利。 但明显的赢家似乎是英伟达、三星、SK海力士、博通和台积电。 随着超大规模云服务商(hyperscalers)中部分公司陷入债务，它们将资产负债表转移给这些供应商，期望从AI支出中获得长期投资回报率(ROI)。
英文原文
The richest companies in the world are going into debt for the AI buildout. Despite hundreds of billions in net income, companies from $AMZN to $GOOGL have increased capex so much: That some are projected to be net negative cash in 2026. Here are the results and who profits: Amazon ( $AMZN ) 2025: +$46.0 billion 2026 (Est): +$11.0 billion Alphabet ( $GOOGL ) 2025: +$80.3 billion 2026 (Est): +$13.0 billion Meta Platforms ( $META ) - Net Debt 2025: +$22.9 billion 2026 (Est): -$7.0 billion (Expected to swing into Net Debt) Microsoft ( $MSFT ) 2025: +$49.2 billion 2026 (Est): +$59.0 billion Oracle ( $ORCL ) 2025: -$98.0 billion (Net Debt) 2026 (Est): -$115.0 billion Microsoft appears to be in the safest position. While Amazon and Google have been largely funding the AI buildout with operating income. However, that large cashflow buffer has vanished. Oracle and Meta appear to be in debt to fuel the buildout, despite $META achieving staggering operating income numbers. Now, where does the money flow into? Operating Income Projections: 1. $NVDA - USA $135.0B -> $186.5B -> $240.1 Billion Operating Income 2. Samsung - Korea ~$30.2B -> ~$170B -> ~$226.7 Billion 3. SK Hynix - Korea ~$32.7B -> ~$124B -> ~$161.0 Billion 4. $TSM - Taiwan ~$62.2B -> ~$86.3B - ~$103.3B – $105.3 Billion 5. $AVGO - America ~$41.5B -> ~$62.8B-$65.1B - ~$84.7B-$93.3B Of course, these are rough estimates based on analyst projections. However, from the general trend, this looks like a leveraged bet that the AI buildout will pay off dividends in FCF after they're finished. But the clear winners appear to be Nvidia, Samsung, SK Hynix, Broadcom, and TSMC. This comes as hyperscalers, with some going into debt, transfer over their balance sheets to them, expecting a long term ROI from their AI spend.
-
指出存储芯片进入新囤货周期，关注WDC等标的。
@elicapitalgroup 很棒的公司。但这看起来实际上是一个新囤货周期的开始，我之前没考虑到 $WDC、$STX 或 $SNDK。
英文原文
@elicapitalgroup Great companies. But it's actually appears to be the start of a new hoarding cycle, wasn't thinking about $WDC, $STX, or $SNDK.
-
列出台积电营收预测，并提及特斯拉营业利润差异。
@XDJoeLee 这是不言而喻的。 $TSM（市值约1.9万亿美元）根据预测： ~622亿美元（2025年） -> ~858亿 – 874亿美元（2026年） -> ~1033亿 – 1053亿美元（2027年）。 差点没把 $TSLA 算进去，因为它的营业利润与其他公司差距太大，但想了想大家应该会有兴趣。
英文原文
@XDJoeLee That's a given. $TSM (~$1.9T MC) from projections: ~$62.2 Billion (2025) -> ~$85.8B – $87.4 Billion (2026) -> ~$103.3B – $105.3 Billion (2027). Almost didn't include $TSLA due to gap between operating income with the rest but thought people would be interested.
-
AI 驱动下，三星和海力士营业利润增速惊人，有望在 2027 年追平或超越美股巨头。
全球最盈利公司排名（Mag 7 vs. 世界） 2025->2026->2027 年营业利润(Operating Income)预测。 #1: $NVDA (美国, 4.4T 市值) 🇺🇸 ~1350 亿 -> 1865 亿 -> 2401 亿美元 #2 三星电子 (韩国, 8200 亿市值) 🇰🇷 ~302 亿 -> ~1700 亿 -> ~2267 亿美元 #3 $MSFT (美国, 2.9T 市值) ~1285 亿 -> 1530 亿 -> 1815 亿美元 #4 $GOOGL (美国, 3.7T 市值) ~1290 亿 -> 1420 亿 -> 1730 亿美元 #4 海力士 (韩国, 4100 亿市值) ~327 亿 -> ~1240 亿 -> ~1610 亿美元 #5 $APPL (美国, 3.76T 市值) 1331 亿 -> 1460 亿 -> 1605 亿美元 #6 $AMZN (美国, 2.13T 市值) 800 亿 -> 1050 亿 -> 1365 亿美元 #7 $Meta (美国, 1.62T 市值) 833 亿 -> 970 亿 -> 1215 亿美元 #8 $TSLA (美国, 1.31T 市值) 44 亿 -> 80 亿 -> 240 亿美元 韩国 8200 亿市值的三星电子预计将在 2027 年在营业利润上追平 $NVDA。 同时，海力士预计将在 2027 年在营业利润上超越 $APPL 和 $AMZN。 主要结论是，由于人工智能(AI)的加速部署，美国超大规模云服务商和韩国股票的增长令人惊叹。
英文原文
Global ranking of the most profitable companies in the world (Mag 7 vs. World) Projections for 2025->2026->2027 (Operating Income). #1: $NVDA (USA, 4.4T MC) 🇺🇸 ~$135.0B -> $186.5B -> $240.1 Billion #2: Samsung Electronics (Korea, $820B MC) 🇰🇷 ~$30.2B -> ~$170B -> ~$226.7 Billion #3 $MSFT (USA, $2.9T MC) ~$128.5B -> 153.0B -> $181.5 Billion #4 $GOOGL (USA, $3.7T MC) ~$129.0B -> $142.0B -> $173.0B #4 Sk Hynix (Korea, $410B MC) ~$32.7B -> ~$124B -> ~$161.0 Billion #5 $APPL (USA, $3.76T MC) $133.1B -> $146.0B -> $160.5B #6 $AMZN (USA, $2.13T MC) $80.0B -> $105.0B -> $136.5B #7 $Meta (USA, $1.62T MC) $83.3 -> $97.0B -> $121.5B #8 $TSLA (USA, $1.31T MC) $4.4B -> $8.0B -> $24.0B Samsung Electronics, a $820B company in Korea is projected to catch up to $NVDA in 2027 in operating income. Meanwhile Sk Hynix is projected to overtake both $APPL and $AMZN in operating income in 2027. The main takeaway is that the growth of both US hyperscalers and South Korean equities is astounding due to Artificial Intelligence ramp.
-
硅基金刚石散热技术预计2027-2028年成熟，政府拨款指引技术趋势。
@zephyr_z9 是的，硅基金刚石(On Silicon Diamond)加上用于散热的合成金刚石，大概要等到 2027 年底到 2028 年？ 但你可以从政府拨款的资金流向中，看出哪些技术是即将兴起的。
英文原文
@zephyr_z9 yeah, diamond on silicon + synthetic diamonds for cooling is prob like late 2027 into 2028? But you can see what's up and coming from where the money flows from gov grants.
-
微软AI需求推高存储波动预期，远期隐含波动率或存在定价偏差。
我的假设是MMS使用了GARCH模型或基于长期平均值的波动率预测。韩国指数在过去10-20年间处于区间震荡。如果回顾往年，模型通常假设波动率在2年内会均值回归。但如果我们看微软对三星/SK海力士陡峭的营业利润预测，加上$MU/$SNDK等存储芯片固有的波动性，看涨期权将从波动率扩张中受益。因此，在我看来，远期隐含波动率存在定价偏差。
英文原文
My assumption is MMS used garchr models or volatility forecasting projections against long-term averages. Korean indexes were range bound for 10-20 years. If it looks at previous years, models typically assume volatility might re-revert to mean in 2 years. But if we look at MS's steep operating income projections on Samsung/Sk Hynix + inherent volatility with memory like $MU / $SNDK, it looks calls will benefit from vega expsnion. So further out IV looks to be mispriced from my view.
-
微软前沿模型开发属实，需重视关键矿产供应链。
@ak_dfranco 路透社报道关键矿产，认真对待此事很有道理。话虽如此，$MSFT 开发前沿模型并加以使用这一基础事实是成立的。这自去年起就已知处于开发阶段。https://t.co/KDOhV1pl9V
英文原文
@ak_dfranco Makes sense to take this seriously as a Reuters reports on critical minerals. That being said, some basis holds with $MSFT developing frontier models and using them. This has been known to be in development since last year. https://t.co/KDOhV1pl9V
-
韩系存储巨头盈利或超美科技巨头，AI内存短缺或成常态，建议关注亚裔半导体股。
这些数据令人震惊： 预计三星和SK海力士将在2027年成为全球最盈利的公司。 它们的预测值超过了$APPL和$GOOGL，这两家公司的营业利润均约为4万亿美元级别。 作为参考，三星的估值约为8200亿美元，SK海力士的估值约为4100亿美元。 这意味着到2027年，一家估值约4100亿美元的SK海力士，其盈利能力将超过$GOOGL（3.7万亿美元）。 根据摩根士丹利此前的估算，SK海力士和三星预计将带来： 合计约3877亿美元的营业利润。 美国两家最盈利的公司$APPL和$GOOGL在2025年的合计营业利润为2630亿美元。（谷歌1290-1320亿美元，苹果1331亿美元） 2027年预测： 三星电子：约2267亿美元 SK海力士：约1610亿美元 2027年预测： 苹果：约1560-1650亿美元 谷歌：约1680-1780亿美元 较小的韩国股票在盈利能力上超过数万亿美元美国超大规模云服务商的统计数据令人震惊。 一个真正有趣的观点是，一家4100亿美元的公司其盈利能力超过了4万亿美元以上的超大规模云服务商。 但市场正在定价的更大问题是，内存短缺是暂时的，还是它们会成为AI基础设施建设中像GPU一样必要的“石油”。 如果你对这个问题的回答是“很可能”，那么获得韩国、日本或台湾股票的敞口可能是一个不错的选择。
英文原文
These numbers are staggering: Samsung and SK Hynix are projected to become the most profitable companies in the world by 2027. Their projections exceed $APPL and $GOOGL, both ~$4T companies in operating profit. For reference, Samsung is valued at ~$820B and SK Hynix is valued at ~$410B. That would make a ~$410B company in SK Hynix more profitable than $GOOGL ($3.7T) in 2027. By Morgan Stanley estimates earlier, SK Hynix and Samsung are est. to bring in: ~$387.7 Billion USD combined operating income. America’s two most profitable companies $APPL and $GOOGL combined brought in $263 Billion USD for 2025. (Google $129-132B, Apple $133.1B) 2027 est. Samsung Electronics: ~$226.7 Billion Sk Hynix: ~$161.0 Billion 2027 est: Apple: ~$156B-$165B Google: ~$168B-178B The statistics of smaller Korean equities exceeding multi trillion dollar US hyperscalers in profitability is staggering. It’s a genuinely interesting point, that a $410B company exceeds $4T+ hyperscalers in profitability. But the bigger question markets are pricing in is if the memory shortage is ephemeral, or if they become a necessary “Oil” like GPUs for the AI buildout. If your answer to that is “likely, might be good to get exposure to Korean, Japanese, or Taiwanese equities.
-
LPTH订单积压1亿，等待客户从锗供应链转向黑钻技术。
@Ren_aramb 是的，$LPTH 积压订单约 1 亿美元。需求确实存在，看起来现在是在等待企业从外国拥有的锗（Germanium）供应链转向黑钻（Black Diamond）技术。
英文原文
@Ren_aramb Yeah $LPTH is backlogged with ~$100m. Demand is there, looks like a waiting game for companies to switch over to black diamond from foreign owned germanium supply chains https://t.co/sgVpFZFwMa