SamLee9355

[Azeem Azhar @azeem]

The GenAI economy has generated $110 billion in sales over the past 12 months. It is growing fast. On an annualized basis, the revenue run rate exceeds $175 billion.

These numbers took us several months to construct, and as far as we know, it’s the first bottom-up, deduplicated measure of consumer and enterprise AI spending across the full stack.

We are releasing this research today in our first The State of the AI Economy report.

https://t.co/cJwZb0T99C

[Jukan @ICML @jukan05]

Many people point to CSPs already running negative cash flow and ask how on earth CAPEX can keep growing from here.

It is a foolish question.

Look at what Arvind Krishna (IBM CEO) said in a podcast conversation.

“Year one of enterprise AI is a net loss. That is exactly why 80 percent of companies right now look like they have failed. It comes down to more engineers, higher token costs, and opportunity cost.

Year two is precisely when the 10x return kicks in. Most companies are not failing at AI. They are simply at the worst point on the curve, right before ROI compresses and improves.”

Ultimately, at this 10x return inflection point, the CSPs’ own revenue base will be expanded by AI, so even as CAPEX grows in absolute terms, the CAPEX to revenue ratio will decelerate, eventually setting off a virtuous cycle, a flywheel effect.

Why do people look to the future when judging CAPEX, but look only at the present when judging cash flow? Cash flow should also be judged by pulling forward the future cash flow.

[Elon Musk @elonmusk]

Tim Cook, who told The Wall Street Journal that the jump in costs was unlike anything he had seen “in any area in over 40 years.”

Biggest price jump in anything I’ve ever seen too. https://t.co/aypJGgssnN

[bubble boi @bubbleboi]

Ok ok I hear it’s anthropic.

@itsclivetime don’t confirm or deny. But just tell me your favorite color.

[Daniel Chang @Daniel_Chang411]

<台積電全新 CoPoS 方形封裝確實用了「玻璃」，但成品裡卻找不到？>

台積電目前最先進、也是市場上高階 AI 晶片所採用的三種封裝架構演進：CoWoS-S ➔ CoWoS-L ➔ CoPoS。

本質上，這場技術革命就是中間那層「中介層（Interposer）」的變革！ 👇

1️⃣ CoWoS-S（經典版）
中介層： 整片純矽（Silicon）
特性： 技術最成熟、良率最高，但成本最貴且尺寸容易受限。

2️⃣ CoWoS-L（主流版）
中介層： 有機 RDL + 局部矽橋（LSI）
特性： 成本比 S 版低、尺寸可放大。

3️⃣ CoPoS（創新方形版）
中介層： 大面積有機 RDL + 多個 LSI 矽橋與特殊深溝槽電容（DTC）
特性： 改為「方形」利用率極高，能大幅減少圓形晶圓的邊緣損耗。

💡 關鍵澄清：台積電到底有沒有用玻璃？
最近很多人在討論台積電方形封�����入「方形玻璃（310*310mm）」，這到底在哪裡？
答案是：它只是個「陪跑」的生產工具，成品中完全不存在！
製程中（借地施工）： 因為有機中介層太薄、太軟，機器在上面做線路和放晶片時會晃動，所以台積電「暫時」將其鋪在方形玻璃載具（Glass Carrier）上施工。
完成後（功成身退）： 當結構固定硬化後，台積電會利用雷射把這片方形玻璃「完全剝離、移除並回收」。

⚠️ 這與 Intel 陣營完全不同！Intel 研發的是將玻璃留在晶片內部的「玻璃核心載板（Glass Core Substrate）」，而台積電的玻璃自始至終都沒有進入成品。

若有錯誤請產業專家指正了。

[SamLee9355 @SLee9355]

@BabybusFL BYOC 也要更新GPU吗，看起来是一个向byoc结构转型的print

[郭明錤｜Ming-Chi Kuo @mingchikuo]

關於台積電的次世代先進封裝 CoPoS 的幾個關鍵（省略可��詢到的技術細節）：

1. 預計 2H28 量產，目標提升 9.5 倍光罩尺寸以上的封裝之量產經濟性，Nvidia 的 AI 晶片 Feynman 可能將首度採用。

2. 根據產業調查，兩個不同的地方會用到玻璃（尺寸 mm）：
→ 310 x 310 的臨時玻璃載具（glass carrier）
→ 250 x 250（測試）/ 510 x 515（量產）的玻璃面板，加工後切割為玻璃核心載板（glass core substrate）

3. 玻璃核心載板的架構主要分成三層：玻璃作為核心層，上下以 ABF（ABF-GCP）增層包覆。玻璃加工的挑戰，像是TGV（through glass via）、填銅 / 金屬化（metallization）等，指的都是這個階段。

4. CoPoS 常見的錯誤論述：
→ ❌ 錯誤 1：採用玻璃中介層（interposer）。⭕️ 應修正為：玻璃非中介層，其互連角色由晶片側 RDL 與玻璃核心載板側 TGV / ABF 增層分別承接。
→ ❌ 錯誤 2：玻璃取代 ABF。⭕️ 應修正為：如前述的玻璃核心載板架構，玻璃與 ABF 並存。
→ ❌ 錯誤 3：晶片放在玻璃上。⭕️ 應修正為：晶片���附於玻璃核心載板的 ABF 增層表面。

5. CoPoS 將延續並強化台積電先進封裝的優勢，預期讓該優勢能見度可達約2032年。

[SamLee9355 @SLee9355]

@AIInvestorHQ what about a comparison between spacex🤣

[法克魷吉米 Jimmy | Real TSMC Insight | 純愛を学び中 💜 @jimmy_yoasobi]

你好👋

由於先進製程與CPO相關技術細節都屬於公司機密 （security A等級）

沒辦法直接透露實際良率情況

不過可以從從目前客戶（Broadcom、NVIDIA等）已經公開的新聞稿和量產進度來看
台積電在CPO的時程上肯定是on schedule

$TSM

[SamLee9355 @SLee9355]

@BigbirdflyChan 确实，他们不应该不知道这一点啊，甚至COHR还是pluggable光模块占比更大，太奇怪了

[SamLee9355 @SLee9355]

@ShanghaoJin @jiangchenglu_ 是啊，不太理解市场简直在乱砸，做EML也够够的，产能那么紧加上NV take or pay contract，NPO也要光源啊，确实不太理解😂

[SamLee9355 @SLee9355]

@jiangchenglu_ @ShanghaoJin 确实很怪，但semianalysis 文章里点名受CPO影响最大的是Lite、COHR、HIMX、AAOI，就没提sive这种蹭的CW

[SamLee9355 @SLee9355]

Agreed, but way more different situation compared with Samsung, US gov standing back of the Intel.

[Jukan @ICML @jukan05]

TSMC was too naive from the start.

Intel is emerging as a serious competitor. TSMC has been extremely strict about preventing Samsung-designed chips from being manufactured at TSMC, so why did it allow Intel to manufacture chips at TSMC?

At this point, TSMC obviously has no reason to allocate CPU capacity to Intel anymore. Instead, it will—and should—actively allocate wafers to Intel’s CPU competitors.

[bubble boi @bubbleboi]

I ran the calculations once:

If Nvidia just uses Intel for packaging the I/O die and other peripherals it’s about 4.8 billion in revenue.

If they use it for all of Feynman it’s 60-80 billion in revenue.

Now imagine TPU ontop of that. We are looking at over 100 billion in just packaging revenue not to mention TPU will most likely be on 14A & Nvidia will use 14A for I/O dies and maybe more.

[Jeff Pu @sssjeffpu]

Intel (INTC) Note Takeaways:

•Big Raise for Foundry Capacity
◦Intel 3: +80% (mainly Ireland) by end-2028 (vs end-2026)
◦18A: +100% (Arizona) by end-2028 (vs end-2026)
◦14A: Strong ramp in Oregon

•Capacity Drivers:
◦Robust server CPU demand
◦18A yields improving to ~80%, EMIB to 90-95% in recent weeks
◦Major U.S. fabless on pipeline
◦Major smartphone maker to start with 18A-P (tablets/PCs) + committed to 14A capacity
◦TeraFab project late-2028

•18A Capacity Shift:
Pathfinder Lake volume wafer capacity → Clearwater Forest (CWF) starts 3Q26, with full volume 1Q27 (expected to be ~1/3 of total 18A wafers)

•Financial Matrix:
◦IFS to turn profitable in 2H27
◦DCAI revenue +39% YoY in 2026E / +30% in 2027E

•Nvidia Collaboration: Likely includes x86 CPUs
→ PC (RTX + Nova Lake in 2027)
→ Server CPUs (Intel x86 IP + foundry) in late 2028

•TP Raised to $135

[SamLee9355 @SLee9355]

@asianinvestors Thanks！agree AI RAN would be huge potential for NOK, full of catalyst, love this. NOK should be rerating by the market of its own InP fab. Ciena is limited by supply as a fabless DCI supplier brings NOK at a strong position

[Brian Armstrong @brian_armstrong]

Good take

My guess is
- demand for intelligence is near infinite
- but 80% of workloads will be running on 99% cheaper models within 12-18 months
- 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level ochestrator agents?)
- rough analogy might be what % of macbooks or gaming PCs sold have the maxed out specs for CPU/GPU, prices are falling much faster than Moore's law here though
- this leads me to think the limiting factor will be energy and compute, not better models

At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.

[Freda Duan @FredaDuan]

RSI, and whether “SOTA rotation” is about to break.

We’re used to SOTA rotating every few months. One lab pulls ahead, another catches up. Gemini was SOTA literally two quarters ago - feels like ages.

But RSI - recursive self-improvement - could change the competitive dynamics.

Dario on Dwarkesh’s podcast back in Feb: Dwarkesh challenged him - if recursive improvement is real, why does SOTA still keep rotating between labs? Dario’s answer was that, until very recently, the compounding advantage from AI-assisted AI research was still too small to really matter (but it's changing).

Since Jan/Feb, I’ve heard more and more researchers talk about RSI. The idea is simple: you build a better AI, that AI helps you build the next better AI, and the loop starts compounding.

It also explains part of the current “token-maxxing” dynamic - big co. spending $ billions on Claude Code just to keep up. And it helps explain the rumors that $OpenAI / $Google are reorganized around coding as the top priority.

Once that loop crosses a certain threshold, it starts to look like a true industrial revolution. Horses were never going to catch up with cars.

[SamLee9355 @SLee9355]

@SemiAnalysis_ So Virgo uses 400G/800G instead of 800G/1.6T？