AI capex · X 热门 · 2026-09-24 15:04 UTC

# AI capex — X 热门讨论 (2026-09-24 15:04 UTC)

## @HyperSharkk (Hyper Shark!) · 09-24 13:09 · ♥37 ↻14 💬0 Framework ที่ผมใช้วิเคราะห์หาโอกาสลงทุยใหม่ๆ ในช่วงนี้หน้าตาเป็นไง มาดู!!!

1. New BOM Item - Architecture ใหม่บังคับให้ต้องซื้ออะไรที่เมื่อก่อนแทบไม่ต้องมี?

2. Qualification Bottleneck - ของที่มีอยู่แล้ว แต่ตรงไหนที่ยังติด reliability / yield / permit / validation / deployment?

3. CapEx Beneficiary - คนที่กำลังแก้คอขวดต้องซื้อเครื่องจักร/อุปกรณ์จากใคร?

4. Attach-rate Expansion - AI รุ่นใหม่ต้องใช้ของชิ้นนั้น “มากขึ้นต่อ GPU / Rack / MW” เท่าไร?

แค่ 4 คำถามนี้ ผมก็เจอหุ้นเป็นกอบเป็นกำแล้ว ลองดูนะครับ อาจจะงงๆ หน่อย แต่อ่านเยอะๆ มันจะเริ่มเข้าใจมากขึ้นๆ > 引用 @HyperSharkk: $GLW มีของชื่อ GlassBridge เป็น Fiber-to-PIC Connector

เผื่อใครตามไม่ทัน ง่ายๆ คือ ยิ่ง Optical ขยับเข้าใกล้ Chip มากขึ้นๆ ปัญหามันไม่ได้มีแค่ “Laser พอมั้ย?”

ปัญหามันจะเป็น จะเอา Fiber จำนวนมหาศาลนั้นอะ ไปต่อกับ PIC ยังไง? ให้เร็ว แม่น และผลิตได้เยอะๆ?

GlassBridge ใช้ Wafer-based Glass + Passive Alignment รองรับ >24 Optical Channels เพื่อลดความซับซ้อนตอนประกอบ CPO/NPO

ตรงนี้เหมือนกับที่เราเพิ่งเจอจาก AFL + Alcyon เป๊ะเลยครับ

แต่ๆๆ $GLW น่าสนใจขึ้นไปอีก เพราะมันไม่ได้มีแค่ Tech

$NVDA + $GLW มีดีลร่วมกันสัญญาหลายปี และ $GLW กำลังเพิ่ม U.S. Optical Connectivity Capacity 10 เท่า พร้อมเพิ่ม Fiber Capacity >50%

ฝั่ง $META + $GLW ก็มีดีล $6B สำหรับ Fiber / Cable / Connectivity ใน Data Center

มันเลยเชื่อมโยงกันแบบนี้...

AI Cluster ใหญ่ขึ้น → Fiber Density เพิ่ม → Fiber-to-PIC ยากขึ้น → CPO/NPO ต้องการ Connector แบบใหม่ → GLW มี GlassBridge → ลูกค้าเจ้าใหญ่ๆ กำลังให้มันเพิ่ม Capacity

เนี้ยแหละสิ่งที่ผมกำลังหาเลย

- New BOM ✅ - Qualification Bottleneck ✅ - CapEx Expansion ✅ - Attach-rate Expansion ❓

เหลือแค่ข้อสุดท้าย… คือ AI System รุ่นใหม่ๆ ต้องใช้ของพวกนี้ “มากขึ้นกี่เท่า” ต่อ GPU / Rack

ถ้าผมหาคำตอบตรงนี้เจอ ผมว่า Thesis ของ $GLW จะน่าสนใจขึ้นอีกเยอะเลย https://x.com/HyperSharkk/status/2103109628901155001

## @youngstockuser (美股投资young) · 09-24 11:37 · ♥31 ↻5 💬19 $AVGO最近确实弱得离谱。

6月最高495美元，现在只剩355美元附近，回撤接近30%。更难受的是，这几天$AMD、$INTC、$ARM甚至Memory都狠狠干了一轮，AVGO却明显跑输。

但把这轮下跌拆开看，Broadcom基本面其实没有出现对应级别的恶化。

第一刀来自预期太高。Q3营收295.9亿美元，同比增长86%，EPS 3.32美元，AI半导体收入167亿美元，同比增长221%。问题是Q4营收指引348亿美元，略低于当时华尔街约350亿美元的预期。对普通公司这是小Miss，对前面已经涨到495美元的AVGO，市场要的根本不是“很好”，而是每个季度都必须炸。

第二刀是竞争。Google把部分Custom Chip订单分给Marvell以后，市场开始担心Broadcom的ASIC份额会被慢慢吃掉。

第三刀就是这两天中国开始调查国有数据中心里的Broadcom交换设备。昨天AVGO又跌2.6%到354.99美元，这条消息直接把Networking的中国风险重新摆到桌面上。

再叠加10Y重新站上5%、AI债券利差扩大、市场开始挑剔AI CapEx，AVGO基本把最近所有利空一起吃了。

但我为什么还是看它？

因为Broadcom在AI里卖的东西并不只有一颗芯片。

Custom ASIC是一条线，AI Networking又是一条线。Q3 AI半导体收入167亿美元，Q4公司直接看到217亿美元，同比增长236%；整个半导体业务上季度收入208亿美元，同比增长127%。更夸张的是，公司已经把FY2027 AI芯片收入预期提高到约1150亿美元。

这个统治力没有因为股价从495跌到355就突然消失。

这让我想到前段时间的$META。

Meta之前同样被市场狠狠干：CapEx太高、AI投入回报不确定、Llama竞争力被质疑。结果Muse一出来，12天280万下载，资金马上重新计算Meta的AI价值，股价随后涨了超过20%。

AVGO缺的也可能只是一个重新让市场兴奋的催化。

下一份超预期AI订单、ASIC客户进展、Networking继续加速，或者中国风险最终没有市场想得那么严重，都可能让资金重新回来。

现在354.99美元回420，也就是大约18%的空间，而且420甚至仍然比6月份495美元的高点低约15%。

所以420对AVGO来说，在我的框架里根本不是什么疯狂目标。

短线什么时候到没人知道，10Y继续冲、AI CapEx降温或者Custom ASIC份额被抢，都会拖慢修复。

但只要AI ASIC和Networking这两个基本盘没有坏，我更愿意把350附近看成一次预期被狠狠干下来的阶段，而不是Broadcom统治力结束。

META已经演过一次：故事没坏的时候，股价可以因为情绪跌得很深，也可以因为一个催化把估值重新打回来。

AVGO下一次市场重新愿意给溢价的时候，420我一点都不会觉得奇怪。 https://x.com/youngstockuser/status/2103086293022572936

## @pepemoonboy (Pepe Invests) · 09-24 13:52 · ♥41 ↻0 💬7 $META is having its iPhone moment.

The market still looks at Meta as a social media company that is spending way too much on AI.

I think the market has been historically focused on their CapEx number and completely missed what the money was actually building.

The iPhone mattered because it took the internet, the camera and the phone and put them into one device that stayed with you all day.

Meta is trying to do the same thing with AI, and they are one of the only companies that actually has all the pieces.

The way I see it, the future most likely looks like a world where you just talk to your AI and it goes and does the work for you.

It books the trip, answers the email, runs the errands and keeps your life organized.

That is exactly what Muse is.

It is a personal assistant that can shop, book travel, schedule appointments and send emails on your behalf.

You can use it through its own app on iOS, Android and the web, or just message it inside WhatsApp like you would a friend.

The Ray-Ban glasses are next, which means it will eventually see what you see and act on it in real time.

Muse will eventually sit on top of everything Meta already owns, and that is where I think this gets super interesting.

Now think about what each piece gives Meta.

The glasses show them what people see throughout the day.

Facebook, Instagram, WhatsApp and Threads show them how people talk and interact with each other.

Muse shows them what people want done, how they make decisions and how they plan their lives.

What we see, how we interact and how we think.

In my opinion that is the closest thing to a full loop of human intelligence that any company owns today.

Other companies have one or two of those.

I cannot name another company that has all three at this scale.

That is why I think humanoid robots are the logical next step.

A robot working in your home needs to understand what it sees, how people behave around it and what you actually want from it.

That is the exact data Meta is collecting right now.

As I have said since day one, I truly believe Meta will one day power the brains of humanoid robots.

The next decade is about to be unreal… > 引用 @pepemoonboy: This. Is. Nuts.

$META https://x.com/pepemoonboy/status/2103120400041267347

## @zerohedge (zerohedge) · 09-24 14:20 · ♥32 ↻5 💬4 What T. Boone Pickens Would Ask About AI CapEx https://t.co/sMNYmlUYnb https://x.com/zerohedge/status/2103127339840274443

## @Abbycadabby87 (Abby) · 09-24 14:37 · ♥35 ↻3 💬2 The semiconductor sector is cooling off across the board, yet $NBIS is still holding up very well.

I think capital is starting to reselect a certain type of company in this cycle: it doesn’t matter how much they spend on AI — what matters is whether they can monetize it quickly.

Nebius has recently raised prices several times. Some on-demand GPU prices increased by around 17%–21%, some CPU instances by 25%, and some Memory products by as much as 41%.

The most important thing is that demand is still there even after the price increases. That doesn’t happen very often.

It’s like buying those damn luxury goods — they’re already expensive, yet they’re still in short supply.

If we bring it back to the stock market itself, this tells us AI compute is still tight, and Nebius has already started turning that supply-demand advantage directly into pricing power. The four large customer contracts signed in Q2 each averaged more than $1 billion in size. What customers are buying is the compute itself. They don’t need to wait for a traditional business to slowly complete its AI transformation.

That’s also the biggest difference between Nebius and the large Hyperscalers.

Microsoft, AWS, and Google constantly need to prove how much incremental profit their massive AI CapEx will eventually generate. Nebius has a much simpler revenue logic: more compute capacity coming online directly means more billable revenue.

So when looking at NBIS now, there are only two things we need to watch:

Whether demand drops after the price increases, and whether the new compute capacity can continue to sell out quickly.

As long as those two things remain intact, the valuation logic for $NBIS will continue to strengthen. https://x.com/Abbycadabby87/status/2103131653182799915

## @VanquishTrader (VanquishTrader) · 09-24 13:12 · ♥34 ↻1 💬4 MUSE COULD TURN $META INTO A $5 TRILLION COMPANY

Meta Connect made the path to a $5 trillion company much clearer to me because Meta is no longer just trying to build the best AI model, it is trying to own the interface through which billions of people actually use AI. Muse Charm is probably the clearest example because it is a 5G connected device that can reach Muse without going through iOS or Android, while glasses and eventually the $1,299 VR Glasses give Meta additional hardware surfaces it controls directly. Meta has spent its entire history distributing through Apple and Google, so if personal agents become the next major computing interface, owning more of that distribution layer becomes incredibly valuable.

But the hardware only matters if Muse can actually do things, and that is why I thought the commerce announcements were probably the most financially important part of Connect. Stripe, Shopify, PayPal, Walmart and Instacart give Muse real places to transact, which starts moving Meta from monetizing the ad that creates intent toward monetizing the transaction that completes it. Zuckerberg has already said Meta ultimately expects to take a small fee from transactions, so the opportunity is no longer just a $20 or $100 subscription because Muse can potentially sit across discovery, recommendations, payments and checkout while Meta’s existing advertising engine continues feeding users into that flow.

The glasses are what can make that agent layer much harder to replicate because Meta and EssilorLuxottica have already sold more than 7 million units over the past year, entry pricing starts around $249 and Meta is now adding use cases such as an FDA cleared over the counter hearing aid that give consumers a reason to buy the hardware even if they do not care about AI. That is the part I think people may be underestimating because hardware that succeeds for reasons beyond AI still puts Muse directly in front of the user, and once the agent can see what you see, hear what you hear and connect into commerce, payments and messaging, Meta starts owning much more of the consumer journey than it does through a social feed alone.

There is still plenty Meta has to prove because no new frontier model actually shipped at Connect, the VR Glasses are still well over a year away, Muse Charm does not have a price yet and the company is still spending $130 billion to $145 billion on CapEx, so I would not pretend the entire thesis was de-risked in one keynote. What mattered more to me was seeing over 1,500 connector applications in less than a week and the pieces of the ecosystem beginning to connect because if Muse becomes the intelligence layer, Meta’s apps provide billions of users, the glasses and Charm provide owned distribution and the commerce partners provide somewhere for the agent to actually transact. That combination is why I think the path toward a $5 trillion Meta increasingly comes from stacking several new businesses on top of the advertising machine rather than needing any single AI product to carry the valuation by itself. https://x.com/VanquishTrader/status/2103110220075729139