Monday Sleeper

[Azeem Azhar @azeem]

When you run an AI agent today, more than half of what you pay for is the model re-reading the context.

Analysis: https://t.co/GgF01HSxdO

[SemiVision @ JPCA👁️👁️ @semivision_tw]

2026 Jan to April…..
Who is Bottleneck ? Easy to understand

[The Orca Fin @axbkol]

@Arm Nvidia N1X ARM based-SoC (System-on-Chip) with customized Windows 11 laptops.

> CPU: 20 core based on ARM v9.2 cores.
> GPU: Blackwell architecture with 6,144 CUDA cores
> Unified LPDDR5X memory architecture
> Those are coordinates for Taipei, Taiwan
> Computex 2026 next week

[Sean @sean_________]

Rosenblatt InP lasers checks:
According to our checks, NVIDIA, which is the driving force behind scale up CPO, asked the supply chain to increase InP laser capacity by ~20x from 2025-2030. The vendors appear to have taken a more conservative stance, agreeing to an ~12x increase ( $AAOI / $LITE etc)

[Irrational Analysis @insane_analyst]

Does this table look painful to make?

It was painful.

[Lip-Bu Tan @LipBuTan1]

Congratulations to my good friend Jensen Huang on being awarded an Honorary Doctorate in Science and Technology from @CarnegieMellon University for his outstanding contributions to accelerated computing and Artificial Intelligence. It was my honor to place upon him his doctoral hood this morning. @intel and @nvidia are collaborating to develop exciting new products!

[AYi @AYi_AInotes]

Claude团队的工程师，已经彻底抛弃Markdown了。

不是Markdown不好用，
是AI变得太快，它已经跟不上了。

以前AI写10行笔记，Markdown刚刚好，
现在AI能一次性输出1000行计划、复杂流程图、完整代码审查，
密密麻麻的纯文字墙谁有耐心看得完？

作者自己都说，他从来没完整读完过100行以上的AI生成MD文件。

更要命的是：现在都是AI写，我们只看不改。

Markdown最大的优点“易手动编辑”，现在已经彻底没用了。

而HTML，才是AI时代真正的沟通语言，
它能做到的事，Markdown想都不敢想：

• 直接生成带颜色的表格、SVG流程图、可点击的原型

• 加滑块调参数、拖拽排序任务、实时预览Prompt效果

• 改完一键导出成代码或Prompt，喂回给AI继续迭代

• 发个链接别人点开就能看，不用下载任何工具

作者直接放出了20个现成示例：

从代码审查的彩色diff，
到可拖拽的任务看板，
从动画参数调试器，
到一键生成的幻灯片。
每一个都是能直接用的生产力工具。

最爽的三个用法，现在就能抄：

1. 代码审查：让AI把PR生成带注释的彩色diff+模块调用图

2. 做计划：生成带时间线、风险表、流程图的交互式项目页

3. 临时工具：让AI写一个Prompt调参器，改完直接复制结果

当然它也有缺点：
多花一点token，生成时间长2-4倍，版本控制不如MD干净。
但作者说：体验提升了10倍，这些代价完全值得。

本质上不是格式之争，而是人机协作方式的升级。
因为Markdown是给人写给人看的，
而HTML是给AI写给人用的。

随着当AI越来越聪明，我们需要的不再是文字墙，而是能互动、能操作、能思考的界面。

现在打开Claude，输入“帮我做一个HTML文件……”，你会打开一个全新的世界。

[Jeff Pu @sssjeffpu]

# 3D Packaging & **Onto Innovation (ONTO TP $397) Note

* **Outlook:** 2026 revenue guided at **$1.3bn**; capacity supports **$2bn** by 2027.

### **3D Packaging Trend** We expect the shift to 3D for Nvidia Feynman, AWS Trainium, and Google V9 (Broadcom).

* **Capacity Surge:** We forecast TSMC SoIC capacity to be 15KPM and 35KPM by end of 2026 and 2027.

* **Tools:** **G5/3Di** tool shipments expected to jump from ~30 units in 2026 to >100 units in 2027.

#ONTO

[Counterpoint Research @CounterPointTR]

𝗔𝗿𝗺-𝗯𝗮𝘀𝗲𝗱 𝗖𝗣𝗨𝘀 𝘁𝗼 𝗖𝗮𝗽𝘁𝘂𝗿𝗲 𝟵𝟬% 𝗼𝗳 𝗔𝗜 𝗔𝗦𝗜𝗖 𝗦𝗲𝗿𝘃𝗲𝗿 𝗖𝗣𝗨 𝗠𝗮𝗿𝗸𝗲𝘁 𝗯𝘆 𝟮𝟬𝟮𝟵

𝖳𝗁𝖾 𝖠𝖨 𝗂𝗇𝖿𝗋𝖺𝗌𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖾 𝗌𝗁𝗂𝖿𝗍 𝗂𝗌𝗇’𝗍 𝗃𝗎𝗌𝗍 𝖺𝖻𝗈𝗎𝗍 𝖺𝖼𝖼𝖾𝗅𝖾𝗋𝖺𝗍𝗈𝗋𝗌—𝗍𝗁𝖾 𝗁𝗈𝗌𝗍 𝖢𝖯𝖴 𝗅𝖺𝗒𝖾𝗋 𝗂𝗌 𝗎𝗇𝖽𝖾𝗋𝗀𝗈𝗂𝗇𝗀 𝖺 𝗌𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖺𝗅 𝗍𝗋𝖺𝗇𝗌𝗂𝗍𝗂𝗈𝗇 𝖿𝗋𝗈𝗆 𝗑𝟪𝟨 𝗍𝗈 𝖠𝗋𝗆 𝖺𝖼𝗋𝗈𝗌𝗌 𝗁𝗒𝗉𝖾𝗋𝗌𝖼𝖺𝗅𝖾𝗋𝗌. 𝖣𝗋𝗂𝗏𝖾𝗇 𝖻𝗒 𝖼𝗈𝗌𝗍, 𝖾𝖿𝖿𝗂𝖼𝗂𝖾𝗇𝖼𝗒, 𝖺𝗇𝖽 𝗀𝗋𝖾𝖺𝗍𝖾𝗋 𝖼𝗈𝗇𝗍𝗋𝗈𝗅 𝗈𝗏𝖾𝗋 𝗌𝗂𝗅𝗂𝖼𝗈𝗇, 𝖼𝗈𝗆𝗉𝖺𝗇𝗂𝖾𝗌 𝗅𝗂𝗄𝖾 @Google, @awscloud, @Microsoft, 𝖺𝗇𝖽 @Meta 𝖺𝗋𝖾 𝗂𝗇𝖼𝗋𝖾𝖺𝗌𝗂𝗇𝗀𝗅𝗒 𝖽𝖾𝗉𝗅𝗈𝗒𝗂𝗇𝗀 𝗂𝗇-𝗁𝗈𝗎𝗌𝖾 𝖠𝗋𝗆-𝖻𝖺𝗌𝖾𝖽 𝖢𝖯𝖴𝗌 𝖺𝗅𝗈𝗇𝗀𝗌𝗂𝖽𝖾 𝖼𝗎𝗌𝗍𝗈𝗆 𝖠𝖨 𝖠𝖲𝖨𝖢𝗌. 𝖶𝗂𝗍𝗁 @Arm 𝖾𝗑𝗉𝖾𝖼𝗍𝖾𝖽 𝗍𝗈 𝗉𝗈𝗐𝖾𝗋 around 𝟫𝟢% 𝗈𝖿 𝗁𝗈𝗌𝗍 𝖢𝖯𝖴𝗌 𝗂𝗇 𝖠𝖨 𝗌𝖾𝗋𝗏𝖾𝗋𝗌 𝖻𝗒 𝟤𝟢𝟤𝟫, 𝗍𝗁𝗂𝗌 𝗆𝖺𝗋𝗄𝗌 𝖺 𝖿𝗎𝗇𝖽𝖺𝗆𝖾𝗇𝗍𝖺𝗅 𝗌𝗁𝗂𝖿𝗍 𝗂𝗇 𝖽𝖺𝗍𝖺 𝖼𝖾𝗇𝗍𝖾𝗋 𝖺𝗋𝖼𝗁𝗂𝗍𝖾𝖼𝗍𝗎𝗋𝖾 𝖺𝗇𝖽 𝗌𝖾𝗆𝗂𝖼𝗈𝗇𝖽𝗎𝖼𝗍𝗈𝗋 𝖽𝖾𝗆𝖺𝗇𝖽.

𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀:
  • 𝖣𝗋𝗂𝗏𝖾𝗇 𝖻𝗒 𝗌𝖼𝖺𝗅𝖾𝖽 𝖺𝖽𝗈𝗉𝗍𝗂𝗈𝗇 𝖺𝗇𝖽 𝗌𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖺𝗅 𝗆𝗈𝗆𝖾𝗇𝗍𝗎𝗆, 𝖠𝗋𝗆-𝖻𝖺𝗌𝖾𝖽 𝖢𝖯𝖴𝗌 𝖺𝗋𝖾 𝗈𝗇 𝗍𝗋𝖺𝖼𝗄 𝗍𝗈 𝗌𝗎𝗋𝗉𝖺𝗌𝗌 𝗅𝖾𝗀𝖺𝖼𝗒 𝗑𝟪𝟨 𝖽𝖾𝗉𝗅𝗈𝗒𝗆𝖾𝗇𝗍𝗌 𝗐𝗂𝗍𝗁 𝗆𝖺𝗃𝗈𝗋 𝗁𝗒𝗉𝖾𝗋𝗌𝖼𝖺𝗅𝖾𝗋𝗌’ 𝖠𝖨 𝖠𝖲𝖨𝖢 𝗌𝖾𝗋𝗏𝖾𝗋 𝗉𝗅𝖺𝗍𝖿𝗈𝗋𝗆𝗌.

  • 𝖡𝗒 𝖾𝖺𝗋𝗅𝗒 𝟤𝟢𝟤𝟫, 𝖠𝗋𝗆 𝖺𝗋𝖼𝗁𝗂𝗍𝖾𝖼𝗍𝗎𝗋𝖾𝗌 𝖺𝗋𝖾 𝗉𝗋𝗈𝗃𝖾𝖼𝗍𝖾𝖽 𝗍𝗈 𝖽𝗈𝗆𝗂𝗇𝖺𝗍𝖾 𝗍𝗁𝖾 𝖠𝖨 𝖠𝖲𝖨𝖢 𝗌𝖾𝗋𝗏𝖾𝗋 𝖢𝖯𝖴 𝗆𝖺𝗋𝗄𝖾𝗍, 𝗉𝗋𝗈𝗉𝖾𝗅𝗅𝖾𝖽 𝖻𝗒 𝗍𝗐𝗈 𝗉𝗈𝗐𝖾𝗋𝖿𝗎𝗅 𝖼𝖺𝗍𝖺𝗅𝗒𝗌𝗍𝗌 – 𝖺𝗀𝗀𝗋𝖾𝗌𝗌𝗂𝗏𝖾 𝗌𝖼𝖺𝗅𝗂𝗇𝗀 𝗈𝖿 𝖠𝗋𝗆 𝖺𝗋𝖼𝗁𝗂𝗍𝖾𝖼𝗍𝗎𝗋𝖾 𝗅𝗂𝖼𝖾𝗇𝗌𝗂𝗇𝗀 𝖿𝗈𝗋 𝗉𝗋𝗈𝗉𝗋𝗂𝖾𝗍𝖺𝗋𝗒 𝗁𝗒𝗉𝖾𝗋𝗌𝖼𝖺𝗅𝖾𝗋 𝗂𝗇-𝗁𝗈𝗎𝗌𝖾 𝖢𝖯𝖴 𝗌𝗂𝗅𝗂𝖼𝗈𝗇, 𝖺𝗇𝖽 𝗅𝖺𝗎𝗇𝖼𝗁 𝗈𝖿 𝗍𝗁𝖾 𝗍𝗎𝗋𝗇𝗄𝖾𝗒 𝖠𝗋𝗆 𝖠𝖦𝖨 𝖢𝖯𝖴 𝗉𝗅𝖺𝗍𝖿𝗈𝗋𝗆.

  • 𝖠𝗋𝗆’𝗌 𝖿𝗈𝗋𝖺𝗒 𝗂𝗇𝗍𝗈 𝗍𝗁𝖾 𝖽𝖺𝗍𝖺 𝖼𝖾𝗇𝗍𝖾𝗋 𝖢𝖯𝖴 𝗆𝖺𝗋𝗄𝖾𝗍 𝗐𝗂𝗍𝗁 𝖠𝖦𝖨 𝖢𝖯𝖴, 𝗐𝗂𝗍𝗁 𝖬𝖾𝗍𝖺 𝖺𝗌 𝗈𝗇𝖾 𝗈𝖿 𝗍𝗁𝖾 𝗆𝖺𝗋𝗊𝗎𝖾𝖾 𝖼𝗎𝗌𝗍𝗈𝗆𝖾𝗋𝗌, 𝖿𝗎𝗋𝗍𝗁𝖾𝗋 𝗏𝖺𝗅𝗂𝖽𝖺𝗍𝖾𝗌 𝗍𝗁𝖾 𝗌𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖺𝗅 𝗌𝗁𝗂𝖿𝗍 𝖺𝗐𝖺𝗒 𝖿𝗋𝗈𝗆 𝗆𝖾𝗋𝖼𝗁𝖺𝗇𝗍 𝗑𝟪𝟨-𝖻𝖺𝗌𝖾𝖽 𝖢𝖯𝖴𝗌 𝗂𝗇 𝗍𝗁𝖾 𝗀𝗋𝗈𝗐𝗂𝗇𝗀 𝗁𝗒𝗉𝖾𝗋𝗌𝖼𝖺𝗅𝖾 𝗌𝖾𝗆𝗂-𝖼𝗎𝗌𝗍𝗈𝗆 𝖠𝖨 𝗂𝗇𝖿𝗋𝖺𝗌𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖾.

𝗠𝗼𝗿𝗲 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀 𝗶𝗻 𝗮 𝗱𝗲𝘁𝗮𝗶𝗹𝗲𝗱 𝗽𝗼𝘀𝘁 𝗵𝗲𝗿𝗲:
https://t.co/cULsizbTu1

𝗔𝗰𝗰𝗲𝘀𝘀 𝗼𝘂𝗿 “𝗔𝗜 𝗔𝗦𝗜𝗖𝘀 𝗦𝗵𝗶𝗽𝗺𝗲𝗻𝘁 𝗙𝗼𝗿𝗲𝗰𝗮𝘀𝘁, 𝗗𝗲𝗰 𝟮𝟬𝟮𝟱” 𝗵𝗲𝗿𝗲:
https://t.co/At5MTNUB4j

𝗔𝗹𝘀𝗼 𝗿𝗲𝗮𝗱 "𝗔𝗿𝗺 𝟮.𝟬: 𝗚𝗼𝗶𝗻𝗴 𝗕𝗲𝘆𝗼𝗻𝗱 𝗟𝗶𝗰𝗲𝗻𝘀𝗶𝗻𝗴 𝘁𝗼 𝗖𝗮𝗽𝘁𝘂𝗿𝗲 𝗦𝗶𝗹𝗶𝗰𝗼𝗻 𝗩𝗮𝗹𝘂𝗲 𝘄𝗶𝘁𝗵 𝗔𝗚𝗜 𝗔𝗜 𝗖𝗣𝗨" 𝗵𝗲𝗿𝗲:
https://t.co/SZwsKKc5jm

#AIInfrastructure #DataCenters #Arm #x86 #Hyperscalers #CloudComputing #Semiconductors #AIChips #AI @cerebras @OpenAI @FujitsuOfficial @SoftBank #ASIC $ARM

[Monday Sleeper @monday_sleeper]

$TER

[SemiAnalysis @SemiAnalysis_]

CPO (Co-Packaged Optics) testing is critical because the cost of failure is enormous. Once an optical engine is attached to the switch package, a defect can jeopardize the entire assembly rather than a replaceable module. The CPO test flow follows four sequential phases, each with distinct technical challenges, equipment requirements, and vendor ecosystems. (1/5)

[Monday Sleeper @monday_sleeper]

Mgmt sandbagged FY27 revenue big time. $PSTG

[The Earnings Correspondent @earnings_guy]

$PSTG (Pure Storage) #earnings are out:

[Quartr @Quartr_App]

We've just created this infographic visualizing 10-year EPS CAGRs among Food, Beverage, and Tobacco companies.

Here are the eight most impressive names:
$MGPI +49%
$CELH +34%
$COKE 32%
$PODR +28%
$DOLE +27%
$POST +26%
$AGRO +24%
$MOWI +22%

[Monday Sleeper @monday_sleeper]

$TER $ONTO $KLIC

[Ryan Shrout @ryanshrout]

Arm just put up one of those earnings quarters where nearly every metric points the same direction: up and to the right.

I had a chance to speak directly with Jason Child, @Arm CFO, after the print, and the confidence level matched the numbers. Q3 revenue came in at $1.242B (up 26% YoY) with royalties at $737M (up 27% YoY) and license and other at $505M (up 25% YoY), plus non-GAAP operating margin at 40.7%.

Arm also raised Q4 guidance to $1.470B and raised non-GAAP EPS guidance to $0.58, even while much of the industry is talking about memory-driven constraints.

A few takeaways from my conversation that felt especially important:

1⃣ The market may be choppy, but Arm exposure is broad enough to keep compounding.
Jason framed Arm strength as “general exposure” across client, handsets, cloud AI, edge, and automotive. That diversity matters when one segment is noisy. It also shows up in the royalty line, which is being pulled by multiple end markets at once.

2⃣ Smartphones are still the largest business, but the mix is the story.
He acknowledged unit headwinds, with the expectation that OEMs protect the high-end and constrain the low-end. Net result, he sees only a modest impact to Arm smartphone revenue for the full year and an even smaller impact when blended into the whole company.

3⃣ Cloud AI is accelerating fast, and it is not subtle.
Arm calls out data center royalty revenue growth of 100% YoY, and positions the shift toward agent-based inference as a structural tailwind because it increases the need for power-efficient CPU cores.
Jason echoed that point in a memorable way: as inference and agentic workloads scale, more of the workload lands on the CPU, and in his view, Arm is the CPU for AI.

4⃣ Watch the hyperscalers, and possible “value capture” shift.
Beyond the usual names (Graviton, Cobalt, NVIDIA), Jason called out the Google TPU path as particularly interesting, paired with Google Axion CPUs.
And on the handset side, he pointed to something I had not considered: if phone vendors move from a Qualcomm core to Arm CSS-based designs, Arm can make materially more per design. Arm itself highlights CSS momentum with 21 CSS licenses across 12 companies, with multiple customers already shipping CSS-based chips.

My bottom line: Arm is executing like a platform company that is getting pulled forward by AI, not pushed around by cycles. Record quarter, raised guide, and a CFO message that the compute foundation is shifting in Arm's direction across cloud, edge, and physical AI.

If you are watching the AI stack, it is getting harder to treat Arm as anything but a clear leader. The next few quarters should be fun.

[Monday Sleeper @monday_sleeper]

$TER Mgmt sandbagged BIG time

[Sravan Kundojjala @SKundojjala]

Teradyne sees mobile not recovering to previous peak despite 2nm/WMCM. Huge installed base in mobile limits incremental growth. Merchant DC GPU qual in process and revenue likely from 2H.

[ariel reyez romero @ReyezAriel]

写在sndk新高时刻的一点历史：

SanDisk 与 Kioxia（原东芝存储）两家的合作可以追溯到 2000 年。
这是一段从一开始就围绕技术路径、资本强度与长期供给稳定性设计的深度合资。

当时 NAND 还远未成为今天这种资本密集、技术高度复杂的产业，但即便在早期，存储芯片就已经呈现出一个明显特征：制程升级速度快、试错成本高、规模效应强。SanDisk 在闪存架构、控制器和系统层面拥有明显优势，而东芝则在制造、设备整合和大规模量产方面经验深厚。双方很早就意识到，单靠任何一方，都难以在 NAND 这条赛道上长期领先。

真正让这段合作“定型”的，是随后十多年 NAND 技术路线的演进。随着行业从平面 NAND 转向 3D NAND，研发和资本门槛出现了结构性跃升。堆叠层数、良率爬坡、设备选型，每一步都意味着巨额投入和高度不确定性。在这种背景下，合资的意义不再是分利润，而是分风险：制程失败的成本对半切，资本开支不再集中压在单一公司的资产负债表上。

这也是为什么，这种合作在行业下行周期中显得格外“抗打”。无论是 2018–2019 年，还是 2022–2023 年，当 NAND 价格深度下探、现金流承压时，SanDisk 与凯侠的联合工厂都能相对从容地调整节奏、延后扩产、控制供给。外界看到的是“共渡难关”，但真正的原因在于：合资结构本身就是为了应对长期波动而存在的。

需要强调的是，这并不是传统意义上的利润共享模式。双方联合投资晶圆厂、共同推进制程研发，但晶圆按既定比例分配，之后各自定价、各自销售、各自承担库存和市场风险。这种安排的核心目标只有一个：长期、可预测的 NAND 晶圆供给。在企业级 SSD 和 hyperscaler 合同中，供给稳定性本身就是竞争力。

从更宏观的角度看，这段合作其实是 NAND 行业走向寡头化的一个缩影。当技术复杂度和资本强度同时抬升，产业自然会从“多家竞争”走向“少数联盟”。SanDisk 与凯侠的合资，并非周期产物，而是这种产业逻辑下的必然选择。

也正因为如此，今天再回看这段历史，会发现它解释了很多现实问题：为什么在需求的情况下，NAND 价格快速反弹，但厂商对扩产保持高度克制；以及为什么头部玩家在定价和供给节奏上，拥有越来越强的话语权。

这不是一段为了活过冬天的合作，而是一段为了在未来十年持续参与牌局而设计的长期同盟。

[Monday Sleeper @monday_sleeper]

$TER

[SemiVision @ JPCA👁️👁️ @semivision_tw]

This year, whether from keynote talks by Micron, Marvell, Advantest, Teradyne, MPI and other leading companies, or from technical discussions throughout the forum, one theme was consistent: testing is now essential for high-speed SerDes, KGD validation, multi-stacked memory, system-level reliability, and thermal integrity. As AI/HPC chips increase in package size and push signaling frequency ever higher, test strategies must evolve beyond traditional wafer sort toward a new electrical × optical × thermal tri-domain framework. This shift is gradually turning testing into a critical engineering platform—one that will shape the next decade of semiconductor processes and advanced packaging architectures.

#SemiVision #Semiconductor #SiPh #SiliconPhotonics #CPO #Photonics #AI #Insight #Technology

[Monday Sleeper @monday_sleeper]

TSMC's huge capex trend is proven by KYEC's capex plan. Teradyne is a big upcoming player in GPU & ASIC testing; Advantest won't be the only beneficiary. This AI testing cake is big enough for everybody. $TER

[Dan Nystedt @dnystedt]

Nvidia’s main chip testing supplier, Taiwan’s KYEC, will spend a record high NT$39.4 billion (US$1.3 billion) on capex next year to meet strong customer demand, media report, as KYEC’s board voted to approve the plan. KYEC has also signed an NT$35 billion 5-year syndicated loan with a consortium of Taiwan banks led by Mega International. $NVDA $TSM $TER #KYEC #Advantest #Semiconductors https://t.co/LommBFKFTg