# AI capex — X 热门讨论 (2026-09-20 08:28 UTC)
## @sleepy0x13 (sleepy.md) · 09-20 03:28 · ♥58 ↻9 💬5 一个美国数据中心工程师，每天要走 3 万步，工作内容包括闻有没有焦味、摸管道有没有裂缝、上屋顶看鸟有没有把冷凝器啄坏。
《华尔街日报》今天这篇我挺喜欢，它没有再讲几千亿美元 CapEx、多少 GW 电力，而是直接钻进了弗吉尼亚 Ashburn 的一个数据中心，看看里面的人每天到底在干什么。
33 岁的 James Waddy 在 Digital Realty 工作。他管的这座数据中心接近 100 万平方英尺，每天光巡检就能走到 3 万步。
他的工作方式甚至有点像人肉传感器。
走进机房先听，有没有不该出现的声音；再闻，有没有设备过热烧出来的味道；管道要用手摸，看有没有裂缝；灯坏没坏、变压器是不是正常，也都得一路看过去。
有一次他走过架空地板，下面就是给服务器送冷风的空间，他停下来听了一会儿，然后说，没有异响，也没有味道，一切正常。
园区走廊又长又空，有些技术人员干脆骑滑板车移动。里面甚至准备了床和淋浴间，碰上极端天气，员工可能直接住下来守服务器。楼顶则是一排排冷凝器，巡检还得检查鸟有没有啄坏设备。
我觉得这些细节比再看一张 AI 数据中心投资图有意思多了。
过去一年我们聊「AI 基建」，很容易把它理解成 GPU、电厂、变压器、融资和土地的组合。
真正走进去之后，它其实是一套必须 24 小时不能停的巨大物理系统。美国现在已经有大约 2700 座运营中的数据中心，新的还在继续盖，但很多设施常驻员工可能也就几十到几百人。
于是会出现一种很具体的工作，一个人每天在几十亿美元设备之间走三万步，靠眼睛、耳朵、鼻子和手，尽量在某台机器真正坏掉以前发现异常。
最有意思的是，Waddy 自己平时甚至不怎么用 AI。
但现在美国 AI 狂潮里越来越多的钱、GPU 和电，最后都得经过这些他每天巡检的机器。 > 引用 @WSJ: Powering the country’s massive AI build-out requires a lot of walking, checking for burning smells and monitoring for bird damage. https://t.co/Ie4qIoFpy8 https://x.com/sleepy0x13/status/2101513755238564266
## @TejaswiPalam (Tejaswi | AI Industrial Shift 🇮🇳) · 09-20 04:19 · ♥43 ↻1 💬2 I thought ACE was just another crane company
Then I looked at what it’s quietly building
Cranes Material handling equipment Construction equipment Agriculture equipment
Now
Heavy cranes Defence Exports
That changed the picture India is building more roads
Railways Ports Factories Power projects Defence infrastructure
Machines needed to build them are getting bigger and more specialised
That’s where ACE gets interesting
The bottleneck isn’t just construction demand
It’s having the right equipment at the right scale
ACE is moving into heavier cranes through its KATO partnership
Building a dedicated defence facility expanding exports 👀
Q1 FY27 revenue hit ₹786 Cr Up 20% PAT reached ₹119 Cr Up 22%
FY27 capex is planned at ₹200–250 Cr
But here’s the part that caught my attention
New defence facility could eventually support ₹500 Cr of revenue capacity
Existing platform New capacity New products New end markets
ACE isn’t just trying to sell more cranes
It’s trying to capture a bigger share of the equipment needed to build India’s next infrastructure cycle https://x.com/TejaswiPalam/status/2101526678044577915
## @Ajay_Bagga (Ajay Bagga) · 09-19 06:46 · ♥30 ↻2 💬3 Astronomical Cash Burn: Projections indicate that OpenAI expects cumulative cash burn/losses to reach approximately $280 billion by 2030, driven primarily by massive data center infrastructure, specialized GPU compute capacity, and aggressive R&D spending.
Massive Capex & Infrastructure Commitments: To support its long-term roadmap (targeting up to 10 gigawatts of compute capacity), the company faces compute, cloud rental, and infrastructure costs running into hundreds of billions of dollars.
Capital Intensity vs. Revenue Trajectory: Even as OpenAI targets significant revenue expansion ($280 billion by 2030 through subscriptions, enterprise deployment, and advertising), compute and operational costs are scaling faster than organic cash flows—requiring recurring, massive equity infusions from strategic partners.
The "Revenue per GW" Bottleneck:
Financial analysts note that breaking even depends heavily on increasing monetizable output per gigawatt of data center capacity, making capital efficiency and model unit economics the single most critical metric for the frontier AI sector. https://x.com/Ajay_Bagga/status/2101201128918745547