概念词典 · 技术
HBM 高带宽内存
通过 TSV 硅通孔把多片 DRAM 垂直堆叠在 GPU/AI 加速器旁边的高带宽内存方案，AI 训练算力的真正瓶颈。
别名： HBM／ 高带宽内存／ high bandwidth memory／ HBM3／ HBM3E／ HBM4
是什么
High Bandwidth Memory (HBM) 是把 8-16 片 DRAM 通过 TSV (硅通孔) 垂直堆叠成一柱，再通过 interposer 紧贴 GPU 芯片放置的内存方案。相比传统 DDR / GDDR：
- 带宽提升 10x+（HBM3E 单 stack 1.2 TB/s）
- 功耗下降约 40%
- 但单位价格高 5-10 倍，且全球只有三家能做：SK Hynix、Micron、Samsung
为什么是 AI 时代真正的瓶颈
GPU 算力按摩尔定律继续扩张，但内存带宽如果跟不上，算力就被”喂不饱”。LLM 训练里，70% 的时间是在等内存，不是在算。NVIDIA H200/B100/B200/B300 一代比一代堆更多 HBM，HBM 的产能直接决定 GPU 的产能。
Serenity 的视角
他对 HBM 的核心论点是 “内存周期被 AI 永久改变”：
- 传统 DRAM 是 18-24 个月一个周期（涨价→产能扩张→过剩→跌价）
- HBM 因为产能切换成本高、且 AI 需求是结构性的，这一轮涨价可能持续到 2028 年才转折
代表标的：
- $MU Micron：纯美股 HBM 标的
- SK Hynix（韩股）：HBM3E 良率第一
- Samsung（韩股）：追赶者，HBM4 是翻盘机会
- $SNDK SanDisk：周期类标的，被他归类为”周期 + AI 双 buff”
看图理解
他常引用一张”内存周期图”，2028 年前持续提价、需求永久增加、毛利率扩张三件事叠加，等于”高额利润周期”。
提到此概念的推文 (89)
按时间倒序 · 滚到底自动加载更多
-
内存大厂2027产能售罄，反驳供过于求论调。
SK海力士、$MU 和三星已售罄2027年的DRAM/HBM产能。 $SNDK、三星、美光当前的年度NAND产能已售罄，铠侠和SK海力士预计将于2026年8月前敲定分配方案。 - 客户仅能获得其初始请求量60-70%的配额 - 业内人士指出，2027年将迎来最严重的内存短缺时刻 - 分配数量已基本确定，但最终交货价格将在临近交付时确定 来源：Digitimes，引用行业消息。 当产能早已售罄时，很难看到那些关于2027年早中期内存“供过于求”的说法。
英文原文
SK Hynix, $MU , and Samsung have sold out of 2027 capacity for DRAM/HBM. $SNDK, Samsung, Micron current annual NAND capacity has been sold out, with Kioxia and SK Hynix expected to finalize allocations by August 2026. - Customers are being allocated only 60–70% of the volumes they initially requested - industry insiders pointed out that 2027 will enter the most severe moment of memory shortage - Allocation amounts are mostly decided, but final shipment pricing will be determined closer to delivery Source: Digitimes, citing industry sources. Hard to see those memory "oversupply" claims in early-mid 2027 when they're all sold out of capacity already.
-
博主认为AI需求加速信号明确，短期波动属去杠杆，看好光互连及存储板块。
我个人认为这是极短期的去杠杆行为，导致许多个股超跌。 - $BE 和 $TER 公布了超预期的业绩，Bloom 报告营收同比增长 166% + 利润率扩张 + 上调 2026 年指引... Teradyne 报告营收同比增长 104%，每股收益增长 300%+。 - 中国领先的光学玩家公布了惊人的初步财报，对西方光子学板块有极强的映射效应。 - 我认为财报季倾向于提醒市场 AI 加速仍在继续。我预计 $LITE、$SNDK / SK Hynix 以及板块龙头 ERs 将继续这一趋势。 - 但感觉市场更多是在奖励现有的加速，而非基于短期修正（如 $AMKR 所见）的未来一年增长。 - 总之，对我跟踪的主题来说看起来非常积极。 - 然后你有 $GOOGL 资本支出上调至 1950-2050 亿美元，这通常是超大规模云厂商最大的 AI 需求信号。 - 预测市场显示下次 FOMC 会议利率不变的概率为 77%。特朗普政府周一呼吁美联储降息。 - 所以今年三次加息的恐惧似乎有些夸大。 - 从 DUV 到 $CXMT 产能过剩的中国恐惧也被极度夸大。我们在每个行业都见过这种情况，直到人们记住激光或 HBM 瓶颈存在是有原因的。 - 就个人想法而言，看到所有东西一起涨跌是不健康的，尤其是这些主题。但我个人并不担心，因为我对我的持仓有充分信心，特别是像 $SIVE 或 $AAOI 这样的光互连领域。 - $META 计算和 $GOOGL GCP 利润率/增长显示出对 $NBIS 及类似新云业务模式的需求增加。 - $META 及其他超大规模云厂商关于与 $SNDK 和三星/SK Hynix、$MU 的长期协议（LTAs）的笔记指向结构性内存需求。 - 可以一直列举下去... - 但当 Jim Cramer 告诉每个保证金交易员在开盘时“全部卖出”，而新闻说存在 AI 泡沫时。 - 我讨厌别人说“它在跌所以卖出”。我相信重要的是看加速的营收/每股收益增长。如果营收加速与超大规模云厂商资本支出保持一致，那么该逻辑依然成立。 - 总之：我看到了 AI 需求/营收加速的迹象，超大规模云厂商资本支出是最重要的跟踪指标之一。 - 我猜我们会看到该主题广泛复苏。但我无法预测具体发生在哪一小时、哪一天或哪一周。
英文原文
I personally see it as extreme short term deleveraging that overshot many individual names. - $BE to $TER reported blowout results, with Bloom reporting 166% Y/Y rev growth + expanding margins + raised 2026 guidance... Teradyne reported 104% Y/Y revenue growth with 300%+ EPS growth. Your leading optical players over in China reported amazing preliminary earnings + extremely strong read through for the Western photonics sector. Think earnings season tend to remind markets about continued AI acceleration. I'm expecting $LITE, $SNDK / SK Hynix, and sector leader ERs to continue that trend. But it does feel like markets are rewarding existing acceleration more rather than future growth a year out over near-term revisions as seen with $AMKR. But as a TLDR, looks very positive for the themes I'm tracking. - Then you have $GOOGL capex raised to $195-$205B, which is typically the biggest AI demand signal from hyperscalers. - 77% 0bps change est. next FOMC decision from prediction markets. Trump administration on Monday called for the Fed to lower interest rates. So your 3x rate hikes fears this year seem kinda overblown. - Chinese fears from DUV to $CXMT overflooding are extremely overblown as well. We see this with every sector from time to time until people remember that lasers or HBM bottlenecks for a reason. As for personal thoughts, it was unhealthy seeing everything rise up or down together, especially with these themes. But I'm personally not worried since I have full conviction in my names, especially optical interconnects like $SIVE or $AAOI. $META compute and $GOOGL GCP margins/growth shows increasing demand for $NBIS and similar neocloud business models. $META + other hyperscaler notes around LTAs with $SNDK and Samsung/SK Hynix, $MU point to structural memory demand. Can go on and on... But when you have Jim Cramer telling every margined DC trader to "sell everything" at market open and news saying there's an AI bubble. I do hate it when others say "it's dropping so sell". My belief is that the important thing to look is accelerating revenue/EPS growth. And if that thesis stays in-tact with revenue acceleration in line with hyperscaler capex. TLDR: I see signs of AI demand / revenue acceleration, hyperscaler capex is one of the most important things to track. And my guess is that we'll see the theme recover broadly. But I can't predict what hour, day, or week that happens.
-
AI供应链多环节瓶颈与涨价，AMD光互连及存储需求强劲。
一些新闻摘要(TLDR)： - 如果你看好中国存储，$CXMT 明天 IPO。 - 三星电子据报在获取大型 FC-BGA 基板方面遇到困难。伊比登(4062)据报要求长期协议(LTA)担保，三星电机寻求预付款。 - 三星 + $AVGO 签署至 2030 年的存储 + 代工 AI 框架，预计超过 2000 亿美元。 - $SOI 预计 2027 财年硅光子学收入将比上一年翻倍，超越摩根士丹利 60% 的增长预测。光子学逻辑起飞。 - SK 集团 + $NVDA 签署 5000 亿美元+ 合作伙伴关系，建设 AI 数据中心和 HBM4 内存。 - 据报 SKC Absolics 玻璃芯延迟至 2027 年。目标年底完成最终可靠性测试，明年量产。所以如果你好奇，这确实推迟了 $LPK 和其他人的爬坡（因此财报后下跌）。 - $QCOM 智能手机处理器价格因上游供应商涨价而两位数上涨。 - $META 预计发行 120 亿美元项目级/SPV 融资用于德克萨斯州埃尔帕索 AI 数据中心扩建。 - Naver 宣布 $NVDA 和 Brookfield 投资 100 亿美元建设 1GW 级 AI 工厂。近期计划是 2028 年达到 200MW。（英伟达投资 10 亿美元，Brookfield 非约束性 90 亿美元） - 64GB DDR5 服务器模块较 6 月底合约价上涨 146% - $INTC 将 14A 工艺量产提前一年，2027 年下半年风险生产，2028 年量产，相比此前 2029 年 HVM 预期。 - 三星电机赢得 2 亿美元 MLCC 订单（现有瓶颈）。 - $AMD（班达纳瓶颈），宣布 Helios 全面生产，2026 年 Q3 出货。包括与 Anthropic 部署高达 2GW 的 MI455X GPU 和与 OpenAI 的 6GW 基础设施。给出新的 >50% CAGR TAM，从 2025 年的 260 亿美元增至 2030 年的 2200 亿美元 CPU 市场。 - 2027 年为 Mi500 推出光互连。据渠道检查，AMD 正走向 CPO 路线（似乎很可能使用 Ayar）。 - $ORCL 赢得 70 亿美元国防部企业软件合同 - JX 金属将其韩国子公司半导体目标产能翻倍，投资 40 亿日元，2027 年下半年开始运营，应对三星电子和 SK 海力士等主要客户需求激增 - 钨六氟化物现货价格同比飙升 2.1-2.5 倍，继日本关东电化和中央气体宣布永久停产之后。随着 3M 计划退出 PFAS 生产，氟化液体供应面临真空。 - 从四家光芯片制造商财报来看，源杰/Eoptolink/TFC Optical/东山精密：无重大订单削减，1.6T 出货预计加速进入 2027 年。光芯片供应商预计将从短缺中获取超额利润。 - 宇树科技目标 2026 年人形机器人产能 30,000 台。阅读人形 TAM 像 $CCXI 等一样扩展。 - 储能锂电池订单上半年在中国增加 2900%。不太熟悉 EVE Energy 和其他电池制造商。
英文原文
Just some TLDR news: - $CXMT IPO tomorrow if you like Chinese memory. - Samsung Electronics reportedly struggling to secure large FC-BGA substrates. Ibiden (4062) reportedly requested LTA guarantees, Samsung Electro-Mechanics sought prepayments. - Samsung + $AVGO sign memory + foundry AI framework through 2030, expected to exceed $200b - $SOI expects FY2027 silicon photonics revenue to double compared to the previous year, surpassing Morgan Stanley's 60% growth projection. Photonics thesis go brrr. - SK Group + $NVDA sign $500B+ partnership to build out AI DCs and HBM4 memory. - SKC Absolics glass core delay to 2027 from reports. Targeting final reliability testing EOY, mass production next year. So if you're curious, this does push back some ramps from $LPK and others (hence drop on ER). - $QCOM price hikes by double digits for smartphone processors, due to upstream supplier hike pricing. - $META expected to issue $12B in project-level/SPV financing for El Paso, Texas AI DC expansion - Naver announced a $10 billion investment from $NVDA and Brookfield to construct a 1GW-scale AI Factory. Near term plan is 200MW by 2028. (Nvidia $1B investment, Brookfield nonbinding $9B) - 64GB DDR5 server modules rises 146% versus end-June contract pricing - $INTC brought forward 14A process mass production by a year, risk production H2 2027 and volume production in 2028, vs. 2029 HVM expectations. - Samsung Electro-Mechanics wins $200m MLCC order (existing bottleneck). - $AMD (the Bandana bottleneck), announces Helios is in full production with shipments Q3 2026. Including an up to 2GW MI455X GPU deployment with Anthropic and a 6GW infrastructure rollout with OpenAI. Gave new >50% CAGR TAM to $220B by 2030 from $26b in 2025 for CPU market. Rolls out optical interconnects for Mi500 in 2027. From channel checks, AMD is heading down to the CPO route (seems likely to use Ayar). - $ORCL wins $7B Department of War enterprise software contract - JX metal doubles semi target capacity at its KR subsidary with a 4B yen investment, with operations to begin H2 2027, amid surging demand from major customers Samsung Electronics and SK Hynix - Tungsten hexafluoride spot prices surged 2.1-2.5x Y/Y following Japan's Kanto Denka and Chuo Gas announcing permanent production halts. Fluorinated liquid supply faces a vacuum as 3M plans to exit PFAS production. - From the four optical chipmaker earnings, Yuanjie/Eoptolink/TFC Optical/Dongshan Precision: no major order cuts, 1.6T shipments expected to accelerate into 2027. Optical chip suppliers expected to capture outsized margins from shortages. - Unitree Robotics Targets 30,000 Humanoid Robot Production Capacity by 2026. Read through on humanoid TAM scaling like $CCXI and others. - Energy storage lithium batteries orders increase first half orders by 2,900% apparently in China. Not as familiar with EVE Energy and other battery makers.
-
质疑市场客观性，建议直接参考英伟达声明而非过度解读。
@j1nv3st 不妨说，我认为市场不再视它们为客观。人们倾向于发布过度延伸的主张，而忽略了细微的技术细节。比如几个月前关于 $MU 在 $NVDA 的 HBM4 共封装光学(CPO) 供应链中毫无份额的说法。我个人只是直接听取 $NVDA 的声明，
英文原文
@j1nv3st Let’s just say I don’t think markets view them as objective anymore. There’s been tendencies to post overreaching claims over a nuanced technical details. Like how $MU had no HBM4 share with $NVDA few months back. I’m personally just listening to $NVDA direct statements and
-
指出HBM相关热压键合设备竞争开始升温。
看起来 HBF（存储）大战正在打响： - Hanmi Semiconductor（先发者） - Hanwha Semiconductor（先发者） - ASMPT（进入者） - K&S（进入者） 这些在报告中都有提到，其中 Hanmi 领先。 “如果有人想研究这个细分领域，TC bonders（热压键合设备）预计会立即受益。”
英文原文
Looks like HBF (memory) wars are kicking off: - Hanmi Semiconductor (first mover) - Hanwha Semiconductor (first mover) -ASMPT (entry) - K&S (entry) Flagged in reports, with Hanmi leading. “TC bonders are expected to benefit immediately” if people wanna research that segment. https://t.co/so1qyyBG2d
-
列出2026下半年四大做多主题：CPO、800VDC功率半导体、物理AI机器人、玻璃基板/先进封装。
如果你想找多头想法，2026年下半年有四个关键主题： 1. CPO，我最喜欢的主题，没有之一。我无法强调市场对它有多误解。 我在硅谷聊过的每个人现在都在私下资助光学初创公司，这是未来的重大信号。 2. 800VDC 功率半导体。我认为其他分析师某种程度上说服了我。 我曾小仓位持有 $NVTS，但令我惊讶的是它几乎翻倍了。进一步研究后，这在“哪些玩家变重要”上是一次相当剧烈的转变。 3. 物理 AI、人形机器人/机器人。你知道 $VPG 基本涨了三倍，我一直在找其他可做多玩家。 最有意思的下游多在私营部门，然后零部件主要在中国/日本。 4. 也许玻璃基板/先进封装仍然有机会，例如 $LPK。由于很多存储玩家最近已经重估，存储在列表中往后排了，不过小众 HBM4 玩家仍有很多机会。
英文原文
Four critical themes for h2 2026 if you want long ideas: 1. CPO, #1 #1 #1 #1 favorite theme. I cannot stress how much markets misunderstand this. Everyone I've talked in SV are privately funding optical startups right now and it's a big signal of what's to come. 2. 800 VDC power semis. I think other analysts kinda convinced me on this. I was in $NVTS for a tiny bit, low concentration. But this almost doubled to my surprise. But after doing more research, it's a pretty seismic shift in terms of what players become important 3. Physical AI Humanoids/Robotics. As you know $VPG basically tripled, and there's a lot of other players I've been trying to find longs on. Most downstream of interest are private sector... Then components are mainly in China/Japan. 4. Maybe Glass Substrate / advanced packaging stuff still eg. $LPK. Memory got pushed down on the list given how much players have re-rated recently, lot of opportunity still with niche hbm4 players.
-
梳理玻璃基板光刻胶、ABF瓶颈和关键矿物供应链。
一些随机笔记： 关于玻璃基板：“YC Chem 据称成为业内首家供应玻璃基板光刻胶的公司”（KR 112290）。 - 目标年底量产，通常量产 + 首发对股价有利。 “目前正与三家以上公司洽谈玻璃基板材料供应”，所以还有更多认证。 “Samyang NC Chem 也在开发玻璃基板光刻胶材料”（482630）。 - 正在送样，明年量产，是另一个玻璃基板名字。 “$NVDA 可能通过长期协议、预付款或战略合作提前锁定高端基板产能，目标是避免 CoWoS 和 HBM 过去出现的供应限制重演。”这与 $TSM COUPE 基板有关。 “Commercial Times 强调，AI GPU、ASIC 和 CPO 需求可能在2027年前触发另一轮高端 ABF 基板短缺。” 所以还有更多基板瓶颈。 另外还有：“中国同意解决美国对稀土和关键矿物短缺的担忧，特别是钇、钪、钕和铟。” 如果你想找多头想法，可以研究钇、钪、钕和铟供应链。 现在一切都在回调，但如果你对更多随机想法感兴趣，这是值得关注的东西。
英文原文
Just some random notes: For glass substrates: "YC Chem has reportedly become the first in the industry to supply photoresists for glass substrates" (KR 112290) - targeting EOY mass production, usually mass production + first to market is good for stock prices. "is currently in discussions with more than three companies regarding the supply of glass substrate materials" so more qualifications. "Samyang NC Chem is also developing photoresist materials for glass substrates" (482630) - Sampling, mass production next year another glass substrate name " $NVDA may seek to secure high-end substrate capacity in advance through long-term agreements, prepayments, or strategic partnerships, aiming to avoid a repeat of the supply constraints previously seen in CoWoS and HBM." This is in relation to $TSM COUPE on Substrate. "As Commercial Times highlights, demand from AI GPUs, ASICs, and CPO could trigger another supply shortage of high-end ABF substrates by 2027." So more substrate bottleneck stuff. Then there's: "China's agreement to address US concerns over rare earth and critical mineral shortages specifically yttrium, scandium, neodymium and indium" So good to look into yttrium, scandium, neodymium and indium supply chains if you want long ideas. Everything is kinda correcting right now, but just something to keep a lookout for if ur interested in more random ideas.
-
Towa财报短期因会计幻象被误读，长期H2布局极好，利润率拐点已至
关于Towa (6315)： 财报非常复杂。我原本的论点（短期）中对当前财报的"Beat"判断是错误的，这是我的失误。 但对于H2 2026来说，这是一个极好的结构性多头布局，而非H1。 TLDR：短期由于算法交易无法理解这些细微差别而看跌，长期非常积极（市场是前瞻性的） 营收 ¥54.36B（同比+1.7%）， 净利润 ¥4.59B（同比-43.4%）。 营业利润 ¥6.91B（同比-22.1%） 新闻标题写着"EPS暴跌-43.4%"或"订单不及预期"，这可能会触发算法交易抛售。 1. 正如你们在美国Towa交易中看到的，"EPS暴跌-43.4%"，算法可能已经根据标题卖出了。 -> 但这是由于去年有一笔一次性"损坏赔偿补偿"支出，以及去年价值13亿日元的一次性股票销售。 -> 此外，"新客户压缩设备的初始成本增加"也导致了本季度盈利能力下降。 所以这是一次会计幻象，并非真正的盈利能力问题+新订单规模化，是一次性的。 然而： 利润率拐点已经到来。这是最重要的信号。 全年营业利润率为12.7%。要使年平均达到12.7%，而Q1-Q3徘徊在10%左右，Q4应该需要达到约18.4%左右？ 这对未来盈利能力来说极其看涨，表明HBM压缩机正在发挥作用，与传统设备相比。 2. 订单"不及预期"： "随着前端工艺产能扩张，预计在2027年3月财年下半财年加速增长。" 这表明 $MU、Sk Hynix和其他公司尚未准备好空间，因为他们仍在建设前端晶圆厂生产线。 但H2，为大幅超预期做好了准备。需求可见度是存在的......只是营收/利润攀升被推迟到H2。 _ Towa前瞻指引预测销售额¥64.0B（同比+17.7%），营业利润¥10.24B（同比+48.0%），这标志着未来将创历史最高盈利能力。 IMO他们还压低了营收指引，因为他们明确表示下半年加速，但没有给出太多预期超预期。 日本公司也倾向于极度保守，他们的股息提高也是一个重要信号。 无论如何，我认为美国抛售可能只是由于负面会计标题+流动性较低。另外，我早了大约4-6个月。 但Towa在H2的布局非常积极。只是不是那种像 $SNDK $AAOI 那样单日10%+的爆发式公司。 TLDR：长期看涨，短期很多细微差别被标题误导了。 只需要再有耐心等待几个月。利润率正在上升，营收/订单被推迟到H2，股息提高，等等。
英文原文
For Towa (6315): Earnings are very nuanced. I got the current ER "Beat" wrong in my original thesis (short term), my bad. But it's an amazing structural long and setup for H2 2026 rather than H1. TLDR: Near term bearish algorithmically since they miss the nuance, very positive H2 (markets are forward looking) Revenue ¥54.36B (+1.7% YoY), Net Profit at ¥4.59B (-43.4% YoY). Operating Profit ¥6.91B (-22.1% YoY) News headlines say "EPS crashed -43.4%" or "order miss" that might trigger an algorithmic selloff. 1. As you've seen with US Towa trading, "EPS crashing -43.4%", and algos might have sold headlines. -> But this was due to last year, of one-off "compensation for damage" payout and one-time ¥1.3 billion yen stock sales from last year. -> Also "increased initial costs for new customers in compression equipment." for new equipment is caused profitability losses this quarter. So a bit of an accounting mirage, not really any profitability issues + scaling new orders, so one-off. However: Margin inflection point is already here. Which is the biggest signal. Full year operating margin was 12.7%. For the annual average to reach 12.7% when Q1-Q3 was hovering around ~10%, Q4 should have been around ~18.4% or something? Which is extremely bullish moving forward for profitability, and shows HBM compression machines are doing work, compared to legacy equipment. 2. Order book "miss": "Acceleration expected during the second half of fiscal year 3/27 as front-end process production capacity expands." This signals $MU, Sk Hynix, and others don't have the floor space ready yet, since they're still building front-end wafer fab lines. But H2, is setup for massive beat. Demand visibility is there... just revenue/profit ramp deferred to H2. _ Towa forward guidance forcasted ¥64.0B in Sales (+17.7% YoY) and ¥10.24B in Operating Profit (+48.0% YoY), which signals all time high profitability in the future. IMO they also sandbagged revenue guidance, since they literally said acceleration second-half but didn't give much of a projection beat. Japanese companies tend to be ultra conservative too, their dividend hike is a large signal too Regardless, I think the US selloff was probably just due to negative accounting headlines + lower liquidity. Also I was a bit too early by like 4-6 months. But Towa is an extremely positive setup for H2. Just not your explosive $SNDK $AAOI 10%+ a day type company. TLDR: Long term bullish, short term lot of nuances missed with headlines. Just need to wait a few more months if you have patience. Margins are increasing, revenue/orders deferred H2, dividend hikes, etc.
-
发现一个做按摩椅和人形机器人副业的 Nvidia CPO 供应商
今天才知道，居然有一家 $NVDA 的 CPO 供应商，副业还做按摩椅和美国人形机器人。 Toto 马桶和 HBM 的梗到处都在冒出来。
英文原文
Today I learned there’s a $NVDA CPO supplier that builds massage chairs and US Humanoids on the side. The Toto toilet HBM meme keeps showing up everywhere.
-
比较 Auros 和 Towa 的不同风险画像
@2victorspoils 我更喜欢 Auros，因为它们和 SK Hynix / Samsung 的 HBM4 hybrid bonding 路线相关，估值大约 2.8 亿美元。 日本那边的 Towa 就更像是一家无聊但有垄断属性的复利型公司。 这只是我目前发过的其中一些名字，后面我还会分享更多。
英文原文
@2victorspoils I liked Auros for higher upside since they’re part of SK Hynix/Samsung Hbm4 hybrid bonding related roadmaps at ~$280mz Over in Japan, Towa was a bit of a more boring but monopolistic compounder. As the ones I’ve posted about so far. There’s a lot more, I’ll share a few later
-
@marv_jones6 第 1 是 CPO，第 2 还是 CPO，第 3 也是 CPO，第 4 是 HBM4，第 5 是玻璃基板。
英文原文
@marv_jones6 #1 CPO #2 CPO #3 CPO #4 HBM4 #5 glass substrates
-
回顾只做 photonics 和 memory 时的惊人 YTD
如果你只盯着 photonics + memory，那今年本来会是一个史诗级收益年。 从 $EWY 到 $AAOI 的所有顶级板块长仓都涨了几百个百分点。 我一直说，Photonics + Memory 在主题上是最好的。 现在还有一些细微的架构变化： photonics 这边是 pluggables -> CPO memory 这边是 hbm3e -> hbm4 本质上还是同一对超级周期。 只是上游瓶颈和受益者不一样。
英文原文
Would have been a generational YTD if you focused on photonics + memory anon. Every top sector long from $EWY leaps to $AAOI is up multiple hundreds of percent. I’ve always said Photonics + Memory was the best thematically. Now there’s subtle architecture changes: On photonics: pluggables -> CPO On memory: hbm3e -> hbm4 The same two supercycles. Different upstream bottlenecks and beneficiaries.
-
英文原文
@fable_7887 Upstream CPO and HBM4
-
Auros是三星/SK海力士供应商，受益于HBM4混合键合计量和薄膜厚度测量两大产品线的量产增长。
所以我确实认为Auros (322310)在韩国是一家市值2.69亿美元、非常有趣的公司。 我从研究中发现，他们是一家相对不知名的三星/SK海力士供应商。 而且他们应该会从HBM4资本支出的混合键合计量(Hybrid Bonding Metrology)中获得高产量增长。 基本上是纯粹专注于两种产品： -> HBM4 / HBM4e / HBM5周期，之前$KLA在红外(IR)计量领域拥有垄断地位。 ---> 现在可能在三星工厂获得资格，下半年产能增长预计。SK海力士在升级到混合键合时也可能获得资格。 关于SK海氨酸的注意事项：我以为会是HBM4e，但显然Trendforce今天报道SK海力士已完成12层混合键合HBM验证，并正在提高量产良率。所以除了三星之外，可能还有一个催化剂。 -> 薄膜厚度测量(Thin-film thickness measurement)。 ---> 现在正在获得资格，与"主要国内芯片制造商"(可能是三星/SK海力士)合作，目标是今年实现批量供应。 免责声明(我持有仓位，不构成投资建议，请自己做尽职调查，这只是我的思考过程) 无论如何，他们在过去十年一直在开发，终于要从多年资格认证中将两种核心产品投入量产，可能是今年下半年。存在风险，例如不进入大批量生产(HVM)，但我个人承担这个风险。 所以我发现这家公司非常有趣，只是想分享一些关于这家公司的想法。
英文原文
So I do think Auros (322310) is very interesting at $269M MC over in Korea. What I've found from research was that they're a relatively unknown Samsung / SK Hynix supplier. And they should high volume ramp from HBM4 capex for Hybrid Bonding Metrology. Basically pure play on two products: -> HBM4 / HBM4e / HBM5 cycles, that $KLA had a monoply over for IR metrology. ---> Getting qualified now likely in Samsung factories, H2 volume ramp est. Sk Hynix likely qualifying too when they upgrade to hybrid bonding. Note on SK Hynix: I thought it would be HBM4e but apparently Trendforce reported today SK Hynix has completed 12-high hybrid bonding HBM validation and is raising yields for mass production. So maybe there's a catalyst there too aside from Samsung. -> Thin-film thickness measurement. ---> Getting qualified now, with "major domestic chipmaker" (likely Samsung/Sk hynix), targets mass supply this year. Disclosure (I do have positions, NFA, please do your own DD, this is just my thought process) Regardless, they've been developing for the past decade, only to volume ramp two core products from years of qualification likely H2 this year. There are risks involved, such as not proceeding to HVM, but I'm personally taking that risk. So I found this company very interesting, just wanted to publish some thoughts on this company.
-
分析CPO供应链中三家最有可能实现3倍涨幅的公司：SIVE(MC约$1.3B)、MSSCORP($1.2B)、Auros($210M)。
至于3倍涨幅的股票，在这些水平上我认为： 1. SIVE 2. MSSCORP (6830) 3. Auros (322310) 是我最看好的猜测。以下是我的思路： 1. SIVE：我真心认为他们明年能达到100亿美元以上的市值，他们是CPO激光器领域的绝对前沿，与LITE和COHR并列。 按130亿美元市值计算...关于可能的版图映射： 光子学领域：AMD CPO、MRVL Celestial CPO、JBL 1.6T、Lightmatter、Ayar、ALChip、GUC、O-Net (ELS)、POET。 太空与国防领域：通过YSS、RTX、ERIC、Bae Systems实现的Golden Dome项目。 硅光子学领域：AAPL（Apple Watch）。 这已经是非常多的客户了，而且还在持续增加。 他们总是可以通过IP收购向下游拓展TAM，或垂直整合来快速追赶LITE的600亿美元市值，一旦获得更多融资的话。 2. MSSCORP (6830)：CPO检测领域的垄断者，约12亿美元市值。 对CPO良率的100%垄断，TSM、AMAT、NVDA、LCRX、INTC等很可能都是客户。 "公司的目标是占据CPO检测市场90%的份额" 这基本意味着100%，只是他们不想引发反垄断担忧。如果他们守住垄断地位且CPO产能增加，很容易从12亿美元涨到50-90亿美元。 3. Auros (322310)：三星/SK海力士的供应商，约2.1亿美元市值，专注于混合键合(Hybrid Bonding)量测。 基本上专注于两类产品： -> HBM4 / HBM4e / HBM5周期，这是KLA此前在IR量测领域拥有垄断地位的市场。 ---> 目前正在三星工厂进行认证，下半年预计量能爬坡。SK海力士在升级到混合键合时很可能也会进行认证。 -> 薄膜厚度测量。 ---> 正在获得"主要本土芯片制造商"（三星或SK海力士之一）的认证，目标今年实现批量供应。 他们已经开发了十多年，终于要在下半年从数年的认证期转向两类产品的量能爬坡。 如果他们转型到量能爬坡阶段，似乎非常有希望在3倍至6.3亿美元左右，这在韩国市场中像是一颗未被发现的宝石。 当然，我不确定这些公司最终会如何发展，这全都是推测性的，但这是高置信度的供应链映射。 但凭我的直觉，这三只我自己持有的股票在这个水平上最有可能。
英文原文
As for 3x brrrs these levels: 1. $SIVE 2. MSSCORP (6830) 3. Auros (322310) Are my best guesses. Here's my thought process: 1. $SIVE: I genuinely do see them being $10B+ next year, they're the literal bleeding edge for CPO lasers alongside $LITE and $COHR. At a $1.3B MC... For likely mapping: Photonics: $AMD CPO, $MRVL Celestial CPO, $JBL 1.6T, Lightmatter, Ayar, ALChip, GUC, O-Net (ELS), $POET. For Space + Defense: Golden Dome via $YSS, $RTX / $ERIC / Bae Systems. Silicon Photonics: $AAPL (Apple Watches). This is just a stupid amount of customers and it's still increasing. They can always TAM expansion downstream through IP acquisitions or vertically integrate to speedrun $LITE's $60B MC one day once they get more funding. 2. MSSCORP (6830): CPO monopoly over inspection at ~$1.2B. 100% monopoly over CPO yields, $TSM, $AMAT, $NVDA, $LCRX, $INTC, and others are all likely customers. "The company’s goal is to seize a 90 percent share of the CPO inspection market" This basically means 100%, they just don't want antitrust. If they defend their monopoly and CPO ramps, can easily see this worth ~$5B-$9B from $1.2B 3. Auros (322310): Samsung / SK Hynix supplier at ~$210M for Hybrid Bonding Metrology. Basically pure play on two products: -> HBM4 / HBM4e / HBM5 cycles, that $KLA had a monoply over for IR metrology. ---> Getting qualified now likely in Samsung factories, H2 volume ramp est. Sk Hynix likely qualifying too when they upgrade to hybrid bonding. -> Thin-film thickness measurement. ---> Getting qualified now, with "major domestic chipmaker" (either Samsung/Sk hynix), targets mass supply this year. They've been developing for the past decade, only to volume ramp two products from years of qualification H2 this year. Seems extremely likely to 3x to $630M if they switch to volume ramp, feels like an undiscovered gem in the Korean market? Of course, not sure how they play out and this is all speculative but high confidence supply chain mapping. But off the top of my head these three that I own are the most likely ones at this level.
-
TOWA 或成 HBM4 上游受益者
是啊……我越看越觉得，市值约 13.4 亿美元的 $TOWA / $TOWCF（6315），可能会成为 SK 海力士、三星、$MU 在 HBM4 支出中的一个“捡漏”受益者。 它高度周期化，跟支出周期强相关。 但我们看到存储厂商的资本开支在全面上调。 这倒不是那种能 4 倍涨的标的…… 但我的猜测是，新一轮 HBM4 存储周期开始后，必须的资金流入激增会让它被显著重估。 尽管它在这轮新周期里似乎已经被大多数人忘了，但它在压缩建模流程上有垄断地位。 混合键合虽然只是去掉了一个步骤，但仍然需要 Towa（垄断企业往往会拿到更高溢价）。 我也不觉得这轮周期和上一轮一样，因为最近存储公司赚钱的速度实在太夸张了？ 只是我的想法，但如果我的判断没错……等大家都在建 HBM4 产线时，它应该会在 11 日周一（东京时间）的财报里体现出来？
英文原文
Yeah... the more I look into things, $TOWA / $TOWCF (6315) at ~$1.34b looks like a sleeper beneficiary for HBM4 spend from SK Hynix, Samsung, $MU. It’s hyper cyclical with spending cycles. But we’re seeing hikes in capex spend across the board from memory makers. Not exactly a 4x type play... But my guess is that it should get materially re-rated from a spike in unavoidable inflows for the start of the new HBM4 memory cycle anyway. And they appear to be a largely forgotten beneficiary for this new cycle despite their monopoly in compression modeling processes. Hybrid bonding just removes one step in the process but requires Towa anyway (monopolies tend to get higher premiums) I also don't think this cycle is like the last due to how much memory companies have been printing recently? Just my thoughts, but if my guess is right... it should show up in earnings on the Monday 11th (Tokyo Time) as everyone’s building out HBM4 lines?
-
测试/良率链条普遍利多
我帮你盯着这件事（测试/良率篇）： $VIAV 和 $FORM 的财报：极度看多 这意味着什么？ 像 $ONTO / $CAMT 这样的名字会起飞。再加上 $TOWA（6315），因为有迹象显示存储产量正在激进爬坡。 像 Msscorps / $KEYS 这样的名字也会起飞。 更广泛来看，存储和光学生态的上游良率、测试、验证和检测都会大幅受益。而且如果它们正在扩产，这对 $COHR、$FN、$LITE 等也都是领先指标。 对 $VIAV： -> 4.068 亿美元 vs. 3.93 亿美元（超预期）同比增长 42.8%。 -> 每股收益 0.27 美元 vs. 0.2-0.24 美元 指引为 4.27 亿至 4.37 亿美元，显示加速。 对 $FORM： -> 2.26 亿美元，按年增长 32%，每股收益 0.56 美元 vs. 0.45 美元 -> 毛利率大幅提升到 49%（这说明有定价权）。 -> 指引为每股收益 0.61 美元，中值营收约 2.4 亿美元。 “高带宽存储（HBM）需求创纪录，以及代工和逻辑网络应用更强劲” 基本上就是更小的良率/测试生态整体都在起飞。BRRR。
英文原文
Monitoring the situation for you (testing/yields edition): $VIAV and $FORM earnings: Extremely Bullish So what does this mean? Names like $ONTO / $CAMT go brr. Throw in $TOWA (6315), since there's indication of aggressive memory production ramp. Names like Msscorps / $KEYS should go brrr. Broader upstream yields, test, validation, and inspection for both memory + optical ecosystem go heavily BRRR. And it's a leading indicator for $COHR, $FN, $LITE, and others if they're ramping up production. For $VIAV: -> $406.8M vs. $393M (beat) 42.8% Y/Y growth. -> $.27 EPS vs $0.2-$0.24 Guidance was $427m-$437m, indicating acceleration. For $FORM: -> $226M, 32% Y/Y, $.56 EPS vs. $.45 -> margins increased a TON to 49% (which indicates pricing power). -> Guidance was $.61 EPS, midpoint ~$240m revenue. "Record demand for High Bandwidth Memory (HBM) and stronger "Foundry & Logic networking applications" Basically the smaller yields/test ecosystem in general. BRRR.
-
上游设备商TOWA受益于三大存储厂HBM4资本支出扩张，CPO封装尚未量产需观察ASX等信号，整体由超大规模数据中心资本支出驱动。
$TOWA最大的信号来自SK Hynix、Micron和Samsung，因为这是上游...三大厂商增加了资本支出预测+建设HBM4，所以资本支出确实会流向Towa。 MSSCORP更微妙。这是用于共封装光学(CPO)的，还没到量产的阶段...你可能需要关注$ASX和其他信号。 但它们都主要从超大规模数据中心资本支出作为总的框架来获得推导，今天这个框架基本是看涨的。
英文原文
$TOWA biggest signal is SK Hynix, Micron, and Samsung since this is upstream... and the big 3 increased capex projections + building out HMB4 so that capex does flow to Towa. MSSCORP more nuanced. This is for CPO, which is not volume ramped yet... You'd probably want to look at like $ASX and other signals. But they all get derived mainly from hyperscaler capex as an overarching umbrella, which was largely bullish today.
-
Win Semi 是 HBM4 量产核心
所以 $SIVE 不负责产能扩张，真正负责量产爬坡的是 Win Semi。 Win 的体量非常大。$AVGO、$LITE、$QCOM、联发科、$MTSI、$NXPI 全都在用他们。 他们也在 $AAPL 和 SpaceX 的供应链里。所以只要你锁定 Win 的产能，超大规模云厂商的量产放量就是可去风险的。
英文原文
So $SIVE doesn't handle capacity scaling, Win Semi does volume ramp. Win is massive. $AVGO, $LITE, $QCOM, Mediatek, $MTSI, $NXPI, all use them. They're in $AAPL, SpaceX supply chains too. So volume scaling for hyperscalers is derisked if you secure allocation and do it through Win.
-
所以我就直说了：市值 13.5 亿美元的 Towa（6315）…… 它是 HBM4（压缩成型）上很罕见、活生生的垄断定义。 虽然今年到现在几乎没怎么涨，但像 $MU、SK 海力士、三星这些存储厂商都是他们的客户。而且每一家存储厂商的财报都在指向大幅增加 capex。 就连从 $TSM 的财报也能看出来，所有主要半导体公司都在经历一轮巨额资本开支周期，以满足 AI 需求。 三大存储厂商也都从 HBM3E 和 NAND 里赚得很猛……所以下一轮 capex 周期大概不会像上一轮那样了（意思就是花得会更多）。 像 $ASML 一样，这种东西周期性极强，但我想在未来几个月的 HBM4 capex 爆发里拿到敞口。 我想把这个名字和 $LPK（玻璃核心基板）一起放到大家视野里，因为它们都是功能性垄断。 但我确实觉得现在这个时间点刚刚好，毕竟所有机器供应商都在大幅重估，而它相对还是平的。 这也不算什么新发现，因为之前还有别的分析师和一些随机关注者提过这个名字，但希望我这次把时点抓对了。 （免责声明：我确实持仓，这只是我自己想法的 TLDR，请不要跟单）
英文原文
So just putting it out there: Towa (6315), at $1.35B... Is a rare, living definition of monopoly over HBM4 (compression molding). It's been kinda flat YTD, but every memory company like $MU, Sk Hynix, Samsung are their customers. And each of the memory company earnings signaled massive capex increases. Even as seen with $TSM earnings, every major semi is going through a massive capex cycle to meet AI demand. And all three memory makers have printed from hbm3e and nand... so the next capex cycle is probably not like the last (meaning a lot more spend). Like $ASML, this is hyper-cyclical but I wanted exposure to the upcoming HBM4 capex ramp over these next few months. Thought I'd put this name on people's radar alongside $LPK (glass core substrates) as a functional monopoly. But I do feel like this timing is about right while every machine supplier is having a massive re-rating yet this was relatively flat. Not exactly a new find, since a few other analysts + random followers had this name but hope I get the timing right. (Disclaimer: I do hold positions, this just TLDR of my own thoughts, please don’t copy trade)
-
存储链条继续起飞
从 $MU 到 $EWY（SK 海力士、三星）再到 $SNDK，存储相关名字最近都在疯狂起飞。 除了三大厂之外，还有 NAND 或台湾的名字，比如南亚科。 现在也许是看看 Towa（6315）的好时机，用来捕捉 HBM4 上游垄断的 capex 周期？ https://t.co/VZPwzB0SFK
英文原文
Memory names from $MU to $EWY (SK Hynix, Samsung) and $SNDK are going brrrr lately. Aside from the big three, then your NAND or TW names like Nanya. Might be a good time to look at names like Towa (6315) to capture upstream monopoly capex cycles for hbm4? https://t.co/VZPwzB0SFK
-
欧洲和日本的垄断都在起飞
像 $LPK 这种欧洲“垄断”（玻璃核心基板）最近也在一路起飞。 $ALRIB 技术上是双寡头，但也在这条线上……（量子 / MBE） 日本那边也有像 Towa 这种（HBM4 / 压缩成型）我持有的名字。 通常垄断会拿到更高倍数。 我之后会慢慢提到更多，只是得先把它们都找出来。
英文原文
European “monopolies” like $LPK (Glass Core Substrates) have also been going brrr lately. $ALRIB is technically a duopoly but it’s there too… (Quantum / MBE) Then there’s some over in Japan like Towa for (HBM4/compression molding) too that I own. Usually monopolies get higher multiples. I'll mention some more over time, just gotta find them all.
-
CPU 热度盖过了存储
现在市场都在盯着 CPU，差点把存储给忘了。 但……SK 海力士、$MU、三星最近都在提高 capex，而且三星最近还开始了 HBM4 生产。 $TSM 也释放了全行业创纪录 capex 的信号。 但就像 $LPK 在玻璃核心基板上的逻辑一样…… HBM 阵营里其实也有不少结构性垄断，是市场可能忘掉了的，尤其是在日本这些地方。 它们大概率会从当前的 HBM4 capex 周期里受益。 韩国那边，像 Hanmi Semi（KRX: 042700）今天都涨了 27.6%+，所以我猜全球其他公司也很快会跟上。
英文原文
Markets are looking at CPUs right now and kinda forgot about memory. But... there's increased capex spend with Sk Hynix, $MU, Samsung around now, with Samsung starting HMB4 production recently. $TSM also signaled record capex across the board. But just like $LPK in glass core substrates... There's a decent amount of structural monopolies over in the HBM camp markets that I'm thinking about in places like Japan. That markets may have forgotten? They would largely benefit from current HBM4 capex cycles. Over in Korea, things like Hanmi Semi (KRX: 042700) have been taking off, up 27.6%+ today, so I'd guess the other companies around the world might play catchup soon.
-
博主分享其对AI供应链瓶颈轮动的精准判断，从存储到光子学再到CPO/ELS，并引用研报分析AI投资超级周期的瓶颈解决序列和CPO商业化拐点
我认为我精准把握了机构层面的瓶颈轮动。 -> 在存储股的上涨尾声捕捉到了 $SNDK、三星、SK海力士、$MU -> 在光子学领域领先机构布局了 $AAOI、$AXTI、$LITE、$COHR -> 现在再次行动，大量加仓SiPh、ELS和CPO：$SIVE、$TSEM、$SOI、$AEHR、Win Semi以及其他 当然如果你想稳健操作： $MRVL（产能自用）、$AVGO（产能自用）、$TSM和$NOK也都有涉及，但他们是更大的玩家。直接曝险下一个超级周期才是理想的。 我仍然认为从存储到EML光收发器的增长空间巨大，但最大的繁荣出现在新架构周期起点/拐点处。 除了宏观因素干扰了一些交易外，预计资本将很快轮动到CPO/ELS供应链。 券商研报的摘要：AI基础设施投资超级周期遵循严格的"瓶颈解决"序列： 算力/GPU（2023）-> 存储/HBM（2024）-> 互联/网络（2025+） 翻译： "我们判断AI价值链的第三轮投资周期已正式启动。继GPU（2023）和HBM（2024）之后，2025年后，光互联将成为增长最快的核心环节。 2027-2028年将是关键拐点，CPO商业化、1.6T标准化和Scale-Up光互联将形成共振，结构性扩大相关行业TAM。" 一些要点： 1. CPO（共封装光学）正从实验室走向商业化量产 2. 2027-2028年是重大结构性拐点（现在在2026年提前布局是个好主意，从测试端$AEHR到ELS端$SIVE或封装端$POET） 3. 光学部件、材料和测试设备的市场总规模（TAM）预计将结构性扩大3到5倍（我认为这还是保守估计）。
英文原文
I think I nailed the institutional bottleneck rotation. -> Caught the tail end of memory name rise with $SNDK, Samsung, SK Hynix, $MU -> Frontran institutions with photonics with names like $AAOI, $AXTI $LITE, $COHR. -> Doing it again now by adding heavily toward SiPh, ELS, and CPO: $SIVE, $TSEM, $SOI, $AEHR, Win Semi, and others. Of course if you want to play it safe: $MRVL (captive), $AVGO (captive), $TSM, and $NOK all do it as well, but they're larger players. And getting direct exposure to the next supercycle is ideal. I still think there's tons of room to grow for memory to EML optical transceivers, but the largest boom is at the start/inflection point of a new architectural cycle. Macro messing up some trades aside, expecting capital rotation soon toward CPO / ELS supply chains. TLDR from analyst note: The AI infrastructure investment supercycle follows a strict "bottleneck resolution" sequence: Compute/GPUs (2023) -> Memory/HBM (2024) -> Interconnect/Networking (2025+). Translation: "We judge that the third investment cycle of the AI value chain has officially begun. Following GPUs (2023) and HBM (2024), post-2025, optical interconnects will become the fastest-growing core segment. 2027–2028 will be a critical inflection point where CPO commercialization. 1.6T standardization, and Scale-Up optical transitions align, structurally expanding the TAM of related industries" Some takeaways: 1. CPO (Co-Packaged Optics) is moving from the lab to commercial mass production 2. 2027-2028 is the major structural inflection point (good idea to frontrun this now in 2026 from testing with $AEHR to ELS with $SIVE or packaging with $POET). 3. The Total Addressable Market (TAM) for optical components, materials, and testing equipment is expected to structurally expand 3x to 5x (I think this is sandbagging a bit).
-
提醒 AI 供应链里很多瓶颈并不在美国，IBKR 也比 HOOD 覆盖更多标的。
@afasano23 AI 供应链里很多乱七八糟的瓶颈 / 断点都不在美国……比如日本那边的 HBM 和 Toto 马桶。 和 $HOOD 等平台比起来，$IBKR 能买到的股票范围要大得多。
英文原文
@afasano23 Lot of the random bottlenecks/chokepoints in the AI supply chain are not US based… as seen with HBM and Toto toilets over in Japan. $IBKR has access to most stocks compared to $HOOD and others.
-
指出全球 AI 供应链瓶颈遍布各地，SIVE 这类光源公司只是众多关键断点之一。
AI 供应链里的瓶颈和公司遍布全球。 正如你从 HBM 和日本的 Toto 马桶上看到的那样，很多都非常随机。 恰好，像 $LITE 这种给 Ayar、$POET 以及硅光 / CPO 公司供货的 cw dfb 激光供应商，也只是瑞典那边的一家公司。
英文原文
Bottlenecks and companies in AI supply chains are around the world. As you've seen with HBM with Toto toilets, lot of them are very random. Just so happens the $LITE type cw dfb laser supplier to Ayar, $POET, and silicon photonic/cpo companies is some random company Sweden.
-
AXTI与住友垄断InP衬底，成AI光子化关键瓶颈。
匿名者，你听了我关于 $AXTI 的论点吗？ 警告：整个 AI 行业可能会受到两家公司的瓶颈制约： 1. $AXTI（市值 7 亿美元） 2. $SMTOY（市值 317 亿美元） 这两家公司控制了全球 60–70% 以上的磷化铟(InP)衬底市场。 未来的 $NVDA、$GOOGL TPU v7 集群、$META、$MSFT、$AMZN 超大规模计算集群都需要基于磷化铟(InP)的激光器和接收器。 $AVGO、$LITE、$COHR 使用磷化铟(InP)用于 800G/1.6T 光收发器中的电吸收调制激光器(EML)、分布反馈激光器(DFB)及其他光基础设施。 如果没有磷化铟(InP)衬底，供应链就会停滞。 在查看了从 TPU 到 Maia 的物料清单(BOM)后，看起来未来的专用集成电路(ASIC)、GPU 和超大规模部署严重依赖光子学。 而这两家供应商可能会冻结全球磷化铟(InP)衬底市场，涵盖几乎所有领域： - 超大规模计算光学（TPU 集群等） - 光收发器（5G、数据） - 激光雷达(LiDAR)（Robotaxi、无人机、军事） - 光模块（互连集群） - 硅光子学激光芯片（Nvidia 未来的共封装光学(CPO)以及 Intel/Broadcom 硅光子学引擎使用磷化铟(InP)连续波激光阵列。） 由于这些公司占据了市场供应的大多数： - AXTI（估计 ~30–35%） - 住友（估计 ~30%） - JX 日本（估计 10-15%） 仅此而已。（例如，Yole 2021 年的行业报告指出“住友电工 + AXT 共同拥有‘超过 75%’的磷化铟(InP)衬底市场”） 超大规模计算/AI 正在向光子学转变，但整个 AI 行业是脆弱的。 如果 $AXTI 或 $SMTOY 中的任何一家停止供应材料，整个未来的 AI 建设就会瘫痪。更疯狂的是，一家 7 亿美元的公司可能成为这一切的中心。 随着 AI 行业向光子学转变，磷化铟(InP)衬底可能会与高带宽内存(HBM)一起成为最大的瓶颈之一。
英文原文
Hope you listened to my $AXTI thesis anon? https://t.co/d5eg0BIhlG
-
SK海力士与三星氦气供应安全，市场恐慌错判，属买入机会。
氦气恐慌为 $EWY、KOSPI、台湾和日本制造了引人注目的标题。 这与之前发现韩国巨头安然无恙后的LNG恐慌类似。 然而，底层机制显示，SK海力士和三星等存储厂商（Memory Players）对潜在的氦气中断具有高度免疫力。 SK海力士向路透社发表声明： “已长期确保多元化的供应链和充足的氦气库存。” “因此，公司几乎不可能受到影响。” 三星尚未发表声明，但他们在第四季度财报电话会议中表示： “我们自豪地成为行业内首个开发并部署半导体制造氦气回收系统的公司。” “这使我们能够回收和纯化氦气以供重新部署，每年减少约4.7吨的消耗，并实现约19%的重复使用率。” 三星确认该系统已在生产线上运行，可能保护了其供应链。 当投资者因金永培（Kim Young-bae）关于90%依赖中东进口将导致HBM生产停摆的说法而恐慌时， Volza贸易数据反驳了这一点，但差异可能源于气态与液态氦的分类、转运，或金出于政治效果而夸大其词。 无论如何，这两大韩国巨头似乎已很大程度上确保了其氦气供应链免受中东冲突的影响。 市场抛售的核心盲点在于： 美国不会让其中东冲突阻碍其AI建设。SK海力士和三星正处于这一核心。 美韩技术繁荣协议（US/Korea Technology Prosperity Deal）旨在保护先进技术供应链。 尤其是当像SK海力士这样的公司发表明确声明时： 声明公司不可能受氦气影响。 市场正在为氦气/LNG中断的最坏情况定价，但查看这些声明表明这是一个买入机会。
英文原文
The helium panic makes for a compelling headline for $EWY, KOSPI, Taiwan, and Japan. Similar to the LNG panic before finding out the Korean giants were fine. However, the underlying mechanics show the memory players in SK Hynix and Samsung in specific are highly insulated from potential Helium disruption. SK Hynix put out a statement to Reuters: "Long secured diverse supply chains and sufficient inventory" of helium. "Therefore there is almost no chance that the company will be affected." Samsung have not put out statements yet, but from their Q4 earnings call they stated: "We are proud to be the first in the industry to develop and deploy a helium reuse system for semiconductor manufacturing" This "enables us to recover and purify helium for redeployment, cutting annual consumption by approximately 4.7 tons and achieving a reuse rate of around 19%" Samsung confirmed that this system is already active on production lines and likely protects their supply chain. When investors panic from headlines around structural exposure due to Kim Young-bae's claims about 90% reliance on Middle Eastern imports shutting down HBM production. The Volza trade data contradicts it, but the discrepancy could be gaseous vs. liquid helium categorization, transshipment, or Kim rounding up for political effect. Regardless, it looks like the two large Korean giants have largely secured their Helium supply chain from conflicts in the Middle East. The core point of the selloff markets are missing: United States will not let their AI buildout stall from Middle East Conflicts. SK Hynix and Samsung are in the center of it all. US/Korea Technology Prosperity Deal to protect advanced technology supply chains as well. And especially so, when companies like SK Hynix put out definitive statements: Stating there's no chance the company is affected from helium. Markets are pricing in the worst case scenarios for Helium/LNG disruption but looking at the statements suggest it's a buying opportunity.
-
NVDA锁定EML产能，光子学上游材料受制于中国，COWOS/HBM/光子学成瓶颈。
$NVDA 去年锁定了大部分光模块(EML)产能，导致其他超大规模云服务商面临巨大瓶颈。然而，光子学整体是一个材料问题，最上游的材料由中国控制。这并非英伟达能单方面决定的。不过，他们通常是最早布局的，这给 $META、$MSFT、$AMZN、$GOOGL 留下了更多问题，因此本文指出共封装光学(COWOS)、高带宽内存(HBM)和光子学可能是瓶颈。
英文原文
$NVDA secured most of the EML capacity last year and caused a huge bottleneck for other hyperscalers. However, photonics as a whole is a materials problem, the most upstream materials controlled by China. Not exactly something Nvidia can decide. However, they're usually the first to things and that leaves more problems to $META, $MSFT, $AMZN, $GOOGL afterward, hence the point of the post signaling cowos, hbm, and photonics are probably the bottlenecks.
-
光子学成AI新瓶颈，激光与InP紧缺，CPO变革预计2028年。
光子学是下一个主要瓶颈。 $NVDA 提前发出了每一个瓶颈的信号：从 HBM（与三星/SK海力士合作）到 CoWoS，再到如今对 $LITE 和 $COHR 的投资： 激光晶圆厂、共封装光学(CPO) 和磷化铟(InP)。 针对每个瓶颈中最具非对称性的多头标的： 1. InP 衬底：$AXTI，住友，JX 2. InP 上游原料+加工：$AXTI 3. 激光器：$AAOI（自研），$AVGO，$COHR，$LITE 4. CPO：$TSEM，Soitec。 $AAOI 财报电话会确认了激光器瓶颈，三家超大规模云厂商希望买断其能生产的所有光收发器。 $AXTI 的积压订单确认了 InP 衬底瓶颈。（玩家图片来源：IndexBox） CPO 瓶颈被广泛预期将在 2027-2028 年末发生。 $AVGO 关于“CPO”的评论带来了短期波动。但这不同于正在发生的激光->收发器和 InP 瓶颈。 关于时间框架： $AAOI，$LITE，$COHR 和激光收发器瓶颈正在实时发生（并预计像内存一样在 2028 年前恶化）。 $AXTI，住友和 InP 衬底瓶颈正在实时发生（并预计只要 AI 在未来多年使用光子学，情况就会恶化）。 由 $NVDA 引领的向 CPO 的大型架构转变可能发生在 2028 年。 这些对于 AI 的下一个范式转变来说似乎是不可避免的。
英文原文
Photonics is the next major bottleneck. $NVDA has signaled each one ahead of time from: HBM (with Samsung/Sk Hynix) to CoWoS and now with the $LITE and $COHR investment: Laser Fab, CPO, and InP. For the most asymmetrical longs in each bottleneck: 1. InP Substrates: $AXTI, Sumitomo, JX 2. InP Upstream Feedstock + Processing: $AXTI 3. Lasers: $AAOI (internal), $AVGO, $COHR, $LITE 4. CPO: $TSEM, Soitec. The laser bottleneck was confirmed from the $AAOI earnings call when three different hyperscaler wanted to buy out any optical transceiver they can produce. The InP substrate bottleneck was confirmed with the backlog from $AXTI. (Image source of players: IndexBox) And the CPO bottleneck is widely expected to happen later in late 2027-2028. There's short term volatility from $AVGO comments around "CPO" in specific. But that's different than the laser -> transceiver and InP bottlenecks happening now. For timeframes: $AAOI, $LITE, $COHR and the laser transceiver bottleneck is happening real time (and is expected to get worse like memory into 2028). $AXTI, Sumitomo and the InP substrate bottleneck is happening real time (and is expected to get worse as long as AI uses photonics for the many years to come). And the larger architectural shift to CPO led by $NVDA will likely happen in 2028. These feel inevitable for the next paradigm shift in AI.