Chris Lee

Keep Walking

[Unclestocknotes @Unclestocknotes]

CPO 真的要來了？我認為 2027 是轉折年，2028 才是財報驗證年。
野村整理了全球主要光學廠商對 CPO 的產品與量產時程。看完之後，大叔感覺不是「某一家會噴」，而是整條供應鏈的時間表，開始慢慢對上了。

統整後結論：
2026 是認證與小量生產，2027 是 design win 開始轉成商業出貨，2028 才比較像進入大規模量產與財務兌現。
所以我不會把 2027 寫成「CPO 全面普及元年」，但它很可能是第一個商業化轉折年。

1. Corning $GLW ：Scale-out 先走，Scale-up 後面接
野村認為 Corning 的 Scale-out 市場會在 2027 年啟動、2028 年成熟。
這個方向我認為合理，因為 AI 叢集越做越大，機架內與機架間的光纖密度只會越來越高。
有一個地方要修正：
市場常看到「2030 年 100 億美元」這個數字，但 Corning 官方比較接近的說法，是希望到 2030 年建立約 100 億美元的 Photonics revenue stream。裡面還包含光纖、線纜、連接器、fiber management 等很多東西。
所以 Corning 的機會是把更多高密度、高精度、高附加價值的光連接一路送到晶片附近。

2. $AAOI：ELSFP 是最大槓桿，但別把產能直接當營收
AAOI 已經在 2026 年開始小量生產，預計 2026 下半年到 2027 年逐步放量。
公司目標是在 2028 年，把包含高功率雷射的 ELSFP 模組產能提升到每月約 40 萬件。
但大叔要提醒 Installed capacity ≠ qualified capacity ≠ actual shipment。有工廠、有設備，不代表客人已經下單。
有月產 40 萬件的能力，也不代表一開始利用率就是 100%。AAOI 的優勢是營運槓桿很大，但風險也最大。
最後要看的是：客戶認證、良率、實際出貨、毛利率，四個能不能一起上來。

3. Coherent $COHR ：垂直整合能力最完整
Coherent 的優勢不是只有雷射。他從
InP、CW laser
EML、photodiode、silicon photonics、optical components
module，一路都有布局。
野村預估：
Scale-out 銷售 2026 下半年啟動
Scale-up 2027 下半年開始
高功率 CW-LD 在 2026 Q4 放量，這個節奏在產業上是合理的。
但我要提醒，這些精確到季度的時間點，更像是券商供應鏈訪查或管理層交流口徑，不是都已經出現在公司正式財測裡。
所以方向可信，時間點還要繼續驗證。

4. Lumentum $LITE：吃到的是「供應稀缺」
野村提到 Lumentum 現在高功率雷射供應非常緊，甚至無法完全滿足 NVIDIA 單一客戶需求。
比較可以確認的是，Lumentum 正在擴大高功率雷射與美國 6 吋 InP 製造能力，而且新產能要到 2028 年才會陸續開出。
所以高功率 CW laser 不是你今天買設備，下季就能量產。
它需要晶圓、epitaxy、製程、封裝、burn-in、reliability、客戶qualification、yield ramp，這也是為什麼現在能認證、能量產、能穩定出貨的公司，有機會在 2027–2028 吃到比較好的價格與議價權。

5. 住友電工：野村認為住友電工的高功率 CW-LD 正在進行最終認證，並與 NVIDIA 和多家 hyperscaler 接洽。
試點出貨預估 2027 年 3 月底開始，2028 年 3 月進入正式量產。
這個時間表很漂亮，但大叔還是要提醒，這些日期比較像野村供應鏈預估，所以住友電工要追的是認證有沒有過、誰是客戶、有沒有拿到正式 design win、量產良率能不能跟上。

6. 我認為 2027 是CPO 開始從技術驗證，走向 design-win-to-volume。
2026：樣品、認證、小量 production。
2027：第一批商業出貨，Scale-out 先走，Scale-up 開始進來。
2028：新產能開出，量產放大，營收與毛利反映在財報上。

7. 高功率雷射的量產準備期很長。所以「先進去的人，會有一段供應稀缺紅利。」所以護城河是我已經認證、良率穩定，而且客戶現在就需要我。

8. 大家都在看：CW laser、ELSFP
CPO engine、InP，但還有一塊被低估，Data center 內部的 fiber management 和物理層佈線。
CPO 把光纖一路拉到 switch ASIC 或 GPU 附近，需要更多Fiber Array Unit
、Fiber harness、高密度連接器
、optical trays發、PM fiber、chip-level fiber management，所以未來是光纖離晶片更近、密度更高、精度更高、單位價值更高。
這也是為什麼 大叔特別愛康寧Corning 這類公司會有機會，不是因為他們碗盤不會破🤣

大叔觀點
野村整理很棒是五家供應商的時間表，開始慢慢重疊在 2027–2028。
CPO 現在進入的是認證→ 產能→ 良率→ 量產→ 財報，比賽才要開始。
不要太早起跑，但也不要太慢跑。大叔會把幾家公司分成這樣看：
Lumentum→ 高功率 CW laser 的供應稀缺性最直接。
Coherent→ 垂直整合能力最完整
AAOI→ ELSFP 營運槓桿最大，但執行風險也最高。
Corning→ 被動光學與 fiber management 的確定性相對高。
住友電工→ 關鍵在認證與實際 design win。
現在先不猜誰會是最後贏家。應該先看誰先認證、先量產、先把良率做好，然後在客戶還找不到第二家的時候，把錢收進來。

[Arvind Srinivas @glocalinvestor]

Goldman: AI Optical networking to grow 10x in 3 yrs

The report has some very interesting views:

- Goldman just broke down the BoM. the switch ASIC is 16% of the bill of materials. The other 84% is light.
- $75.8K of BoM sells for $130K. ~42% gross margin at the box. Its the optical engine.
- Silicon photonics penetration in datacom goes from 6% (1Q24) to 46% (4Q28E)
- Light source supply is "very tight" through all of 2026, "tight" through 2027, and does not reach balance until 2H28.
- 800G adopted across every hyperscaler. 1.6T live at $NVDA and $GOOG now, META 2026, $MSFT / $AMZN / $ORCL 2027. 3.2T at Nvidia and Google in 2028
- CPO scale-out on the roadmap at Nvidia, Google and $META .

Names by layer (my mapping):
- ASIC: $AVGO $MRVL.
- Lasers / light source / Optical engine: $COHR $LITE $AAOI
- Substrate: $AXTI $IQE $SOI.PA
- SiPh foundries: $GFS $TSEM $TSM $INTC
- Analog front-end: $CRDO $SMTC $MXL
- Connectivity and passives: $GLW $APH US Conec
- Optical assembly / OSATs: $FN $AMKR $ASX
- System: $CLS $ANET
- China optical-engine side: InnoLight , Eoptolink

Meanwhile, credit where due @GoldmanSachs has made this report completely free & public unlike some blood thirsty "analysts" suing people. Check it out. Link 👇

[💎 大阪NEW LIFE @newlifeosaka]

真牛逼！
疯了！一个人2万元手搓94分钟AI科幻电影，刘慈欣《山》居然真被拍出来了 #刘慈欣 #AI电影 #科幻电影 #人工智能 #山

[Chris Lee @ROE2017]

诺德股份HVLP 4小道消息通过台厂认证

[ImNotHarsh | 📈💸 @imnotharsh]

$INTC Intel EMIB is now being outright mentioned by $SKHY SK Hynix in their materials - at Hot Chips 2026

It’s happening…

[每日资讯 @XiaoM17]

微软今天收到了英伟达第一批正式量产的Vera Rubin，

拆解微软最新的超级AI机柜

看看各个零部件对应的A股供应商 有哪些

详情见图片，兄弟们。

[Chris Lee @ROE2017]

@AYi_AInotes 片尾曲《雨后轻风有香》非常好听

[Alex @Alex_Intel_]

Bernstein Taiwan analysts have 2028 EMIB production when converted to a CoWOS equivalent (modeling purposes) as greater than all 2025 CoWoS, and around 1/3 market share in 2028

EMIB is very efficient, and Intel Malaysia is a very large facility. In addition Intel has Amkor Korea to help output

$INTC

[P Equity Research 📰 @pequityresearch]

InP shortage is getting worse as expected.

> A further price increase is anticipated in the fourth quarter, with the increase expected to exceed 10%, marking the largest increase in history. Suppliers have stated that "even with money, you may not be able to buy them."

> Industry insiders revealed that the price of indium phosphide substrates started to rise in the fourth quarter of last year and has been raised three times since then, now reaching a "fourth consecutive increase"; the epitaxial wafers made from the substrates have also increased in price twice, and are heading towards a "third consecutive increase" in the fourth quarter, which means that the market supply shortage is more severe than expected.

> Indium phosphide (IP) epitaxial wafers must be manufactured from IPP substrates, but upstream IPP substrate suppliers are constrained by policies and production capacity, and have so far been unable to meet market demand. Due to this supply-demand imbalance, prices have been rising steadily, with increases from the initial 3% to 5%, and the latest news indicates a potential surge of over 10%, the largest increase in history.

We need more InP...

[Lip-Bu Tan @LipBuTan1]

Following Intel’s $23 billion oversubscribed capital raise, I’m grateful to everyone who has reached out with messages of support and confidence in our future. Thank you, @adenatfriedman and @Nasdaq, for helping us mark this important moment for Intel. We are building a new @intel - more focused, more disciplined and moving with greater speed. The work continues, and I’m excited about what lies ahead!

[Kwesi Kay @EiiKwesiKay]

Chinese aesthetics at their finest.

[Hardik Shah @AIStockSavvy]

$SNDK | Sandisk Corporation Ratings & PT

[P Equity Research 📰 @pequityresearch]

UBS: Vera Rubin BOM & Server Costs

Conventional Server BOM & Memory Cost Trends

> Surging Memory Share: Memory now accounts for roughly 69% of the Bill of Materials (BOM) cost for a conventional data center server (loaded with 16 DIMMs of 64GB DRAM and 3.84TB NAND), a sharp rise from 40% in 2025.

> Cost Escalation: Driven by rising DDR and NAND contract pricing, incremental memory costs alone are estimated to drive a 127% increase in overall conventional server costs (totaling $26,077 up from $11,513).

> Component Breakdown:

DRAM costs show a 315% YoY surge ($14,624 total, representing 56% of cost).

NAND NVMe Gen4 costs show a 227% YoY increase ($3,379 total, representing 13% of cost).

AI Server Superchip Memory Projections (GB300, VR200, VR300)

> GB300 (2026): Total HBM3E and LPDDR5X costs per superchip are estimated at $9.7k (53% to 58% of total cost), translating to approximately $350k per rack.

> VR200 (2026–2027): Combined HBM4 and SOCAMM2 costs sit at $24.3k (62% of cost) in 2026, climbing to 70% of the total in 2027. Per-rack memory costs move from $875k in 2026 to $1.143m in 2027.

> VR300 (2027): Assuming 756GB HBM4E 12-Hi and SOCAMM2, superchip memory costs could reach $35.5k (77% of total superchip cost), driving per-rack memory costs to a massive $1.273m.

Macro Rack-Level Trajectory

> Milestone Threshold: Total memory costs are projected to comfortably exceed US$1 million per Vera Rubin rack by 2027.

> Structural Shift: Across the AI server lineup, total memory costs as a percentage of the overall stack continue a heavy upward trajectory, moving from an already substantial 58% for the GG300 up to 77% for the VR300.

For reference, this is from late May.

$NVDA $MU $DRAM $EWY

[Eamon涷渔 @Eamonyiu]

三十三重天外天，九霄云外有神仙。
仙人都是凡人变，只怕凡人志不坚 。
南天门？白玉京？我已傻傻分不清！

[Athu @athuinvests]

Photonics starting to show real momentum heading into the August earnings wave.

Here are stocks I'm watching closely and why:

1. Lasers & optical components: $LITE $COHR $AAOI $SIVE

The backdrop: Lumentum's CEO said a few days ago that they're falling further behind demand and could be sold out of 2028 capacity within two quarters. He's also warned InP, the material every optical laser depends on, is now more supply constrained than memory chips.

Three of the four in this group print back to back over the same stretch, $AAOI Aug 6, $LITE Aug 11, $COHR Aug 12.

> While Lumentum & Coherent are obviously expected to show more than strong numbers in their print (both with a double beat), I'm actually more keen on seeing how their guidance looks, and if they'll raise it once again.

> With AAOI, the question is more on execution against the guidance they gave, and whether they're heading to beat that $1B target for FY2026. I think they'll probably make a case for $1.2B+ instead, given the continuous signals of 1.6T+ TAM expansion across the industry (photonics in general too).

> On Sivers, expected to release earnings around the end of this month, it's more about how their opportunity pipeline looks and whether the backlog is growing, plus any added signs of them heading into 2027 with CPO scaling ramps that show they're well positioned to capture and execute on it.

So, $LITE & $COHR are more established and proven on execution, but $AAOI & $SIVE definitely have a lot of eyes on whether they can do what the two photonics giants have already shown.

2. Foundry: $TSEM

Tower doesn't sell chips to consumers, it sells wafer capacity to the companies building the chips. That makes its order book a leading indicator for the whole group.

They already signed $1.3B in silicon photonics contracts for 2027 revenue with their largest customers. Their named collaborators on the silicon photonics side include $NVDA and $MRVL. When a foundry is getting prepaid years out by the biggest names in the space, that's a clean signal for where real demand sits, not projected demand.

$TSEM reports Aug 4. Their guidance becomes a two way signal for both upstream and downstream players, since orders today become shipped product tomorrow, but it also means demand for materials (like $IQE, with whom they have a partnership) increases along the same line. That in turn affects test system players like $AEHR, sitting at the wafer level of the game.

3. Connectivity silicon: $MRVL $CRDO $ALAB

These three are turning that optical hardware into usable bandwidth inside the data center.

> $MRVL reports Aug 20, and its DC revenue is guided up 50% this fiscal year. This name is tied to the "$1T AI infrastructure" framing from Matt Murphy and Jensen Huang. Watching their optics commentary closely, since that segment's growth guide just got raised from 50% to 70%+ last quarter.

> ASICs & memory businesses are also key contributors to overall revenue, which diversifies them and sets them apart from the other pure play names in this post. Keen to see that side too.

$CRDO and $ALAB are the emerging & proven hyperscaler deal players.

> Credo just closed its DustPhotonics acquisition to build out its own optical stack, guiding 80%+ revenue growth this fiscal year on AECs, optical DSPs and retimers, the connectivity layer for scale out and scale up. It reports Sept 2, a little outside the main wave but worth watching for confirmation of the trend.

> $ALAB reports Aug 4. They're the PCIe/CXL connectivity name for rack scale AI systems, and their Scorpio fabric switches are seeing broader hyperscaler rollout in the back half of the year. Q2 guide was $355 to 365M, up mid teens sequentially.

[The Kobeissi Letter @KobeissiLetter]

BREAKING: Fed Chair Warsh is considering reducing the amount of times that the Fed meets in a year, per NYT.

Details include:

1. The FOMC has met eight times annually since 1981, with Federal law requiring at least four meetings per year

2. Fed Chair Warsh has reportedly not proposed a specific number of meetings

3. Fewer meetings would mean fewer scheduled rate votes and less frequent policy decisions

A major shift is underway at the Fed.

[*Walter Bloomberg @DeItaone]

U.S., ISRAEL WEIGH STRIKES ON IRAN ENERGY SITES

The U.S. and Israel are preparing a potential joint campaign targeting Iranian energy infrastructure, with strikes possible this weekend, according to CBS.

President Trump has not given final approval. Officials are reportedly weighing the economic impact, with discussions including completing operations before Monday’s market open. White House and Israeli officials are coordinating plans.

[Hardik Shah @AIStockSavvy]

$SKHY | UBS 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗲𝘀 𝗕𝘂𝘆 on 𝗦𝗞 𝗛𝘆𝗻𝗶𝘅, PT $𝟮𝟬𝟰

Analyst sees shares not fully reflecting structurally higher memory profitability and stronger FCF generation driven by an AI-led demand inflection.

[루팡 @DrNHJ]

https://t.co/4VaPjuVinR

[John Intel @intelfabs]

And we have Bernstein estimates on Intel $INTC

Easy enough!

We model Q226 at $14.3B/ $0.20, a touch below consensus at $14.4B/$0.21.

For Q326, we model $14.9B/$0.28, below to in-line compared to consensus at $15.1B/$0.28.

For full-year 2026, we model $57.6B/$1.07, below consensus at $58.6B/$1.11.

For full-year 2027, we model $63.3B/$1.50, a bit below consensus at $66.0B/$1.57.