# AI capex — X 热门讨论 (2026-09-22 09:21 UTC)

## @dexteryy (Dexter Yang ᯅノ🌐🔗🧙🏻‍♂️👾🦋) · 09-22 07:10 · ♥92 ↻11 💬5 骂得好脏 亚历山大王是个内行

Muse是个透明化云端沙盒工作环境（Manus和Grok bot的模式）和Agentic Web生态优先的、全场景的、更主动的美版豆包，比止步于chatbot（云电脑和专家模式的层级太深不够优先）的手机豆包（豆包手机则走偏了）和本地桌面工作环境/办公场景优先的workbuddy能更进一步让 Agentic AI 渗透到下沉市场

Meta不愧是美版字节+腾讯，是能让市场相信AI的完整体，既有字节的下沉市场和商业模式/赚钱能力（同样做免费C端AI应用和做AI capex，市场不会怀疑Meta的货币化能力，而OpenAI、Google和腾讯做再好也会被怀疑AI capex不可持续），又有腾讯的社交场景卡位生活场景卡位（但同样不具备对长期稳定的真正的平台/入口的卡位，所以Meta、字节和OpenAI都在努力做OS和个人计算设备）

感谢小扎，我的美股从7年半23恢复到26倍了（即将超越三个月前最高的27倍） > 引用 @dexteryy: 为了解释为什么字节的商业模式是最好的那种、比腾讯好的多，对腾讯和字节的主营业务做了一下battle，见图1，图2-4应该有助于理解

关于背后反映的事实、趋势和逻辑，我也强调几条：

1. 从「主动查找/关注/订阅/安装、人工维护关系、人为分发、主动召回和打开」到「基于各种context做算法化/AI化的 自动创造兴趣/需求、自动分发、自动行动」是大势所趋

2. 跟包括微信、淘宝、Google 在内的其他平台相比，抖音/TikTok 是更前沿、更有势头的平台，不是抖音昙花一现朝不保夕，而是微信原本可以新增的和越来越多存量的使用时长、日活和用户场景（比如外部动态、内容发现、内容消费、商品发现、购买决策）被转移到抖音

3. 用户在内容上能花费的有限时间/注意力，大部分从游戏、聊天、传统/新媒体转移到了算法推荐内容，短视频只是算法推荐内容中最下沉、规模最大的内容形式而已

4. 受益于技术不是「五环内用户」的特权，这世界上大多数人属于「下沉市场」，他们同样有资格利用技术改善生活/人生、进入更大的世界拥有更多可能，字节这种基于客观数据而非主观价值判断的产品方法论正是把技术普及到下沉市场和更广大人类群体的必要途径 > 顺带一提：这些下沉市场用户愿意花费时间的游戏、聊天、自媒体/主流媒体，其实同样属于很多人眼里的「精神毒品」、「跟用户价值有冲突的货币化」

5. 微信从来都没法把在中国这种不成熟市场上做出的社交产品形态（其中很多是扭曲的，比如小程序这种非标准非开放 web）复制到全球，抖音/TikTok 则不同，背后也反映了谁更前沿、谁更有可持续性

6. AI 时代里，微信比抖音更危险：

抖音的核心在于后端的多模态内容沉淀、算法/AI自主能力，以及前端的新一代人机交互模式和多模态交互能力，趋向拥有越来越多的context，受益于 Agentic AI、Generative AI 乃至 Physics AI 的发展和需求

微信的核心在于人与人的社交关系沉淀和社交场景，以及围绕这些context自然衍生或强行捆绑的「上网需求」和「生活需求」乃至它们进一步产生的context

在 AI 时代，很多过去好像「离不开」的人与人社交/通信/协作需求和社交关系，会被发现其实是「泡沫」或「摩擦」，被挤干或消解，很多用户场景会从社交context和为社交优化的人机交互中解绑出来，把自己的context融入到其他更通用的context和更直觉/高效的人机交互中

AI 容易带来的不是「新聊天应用」，而是： - 更少的「聊天」（现代媒体出现前，人们在聊天上花费的时间远远超过今天，参考欧洲那些贵族宅邸的设计） - 更多的context（跟桌面互联网转换到移动互联网时一样，context的增量远大于存量） - 更解绑和可组合的软件需求（移动应用只是过渡形态） - 对人类的两大影响外部世界能力（使用更多更复杂的工具，跟更多人做更复杂的语言沟通）都要求更低（门槛/成本更低+杠杆更大）

这很可能让「聊天」的需求和场景逐渐不再能支撑起超级应用（更不用提微信这种想架空操作系统的超级应用），不再是分发入口，进一步加速减少人们投入在上面的时间/注意力 https://x.com/dexteryy/status/2102294520683749579

## @CNBCTV18Live (CNBC-TV18) · 09-22 03:23 · ♥32 ↻6 💬1 #CNBCTV18Exclusive | “Harder & harder for India to remain non-aligned,” says #JamieDimon, #JPMorgan Chase, on CNBC-TV18

Jamie Dimon Says tariffs on #India with respect to #Russian oil purchases are not a good idea; sees Russia-Ukraine war potentially prolonging for another 5 years. Opposes new #Fed task forces; says #Fed should stick to 2% inflation target. On #AI capex: “It’s not just #RoI, sometimes it’s just table stakes.”

#CNBCTV18 #JamieDimon #JPMorgan #India #RussiaUkraineWar #USFed #AI #Markets https://x.com/CNBCTV18Live/status/2102237402714751004

## @pepe_maltese (Pepe Maltese) · 09-21 22:15 · ♥31 ↻1 💬2 $IREN - Is Chanos wrong?

Yes, getting $IREN anywhere near the Bernstein type of outcome will require an absurd amount of gross capital. Tens of billions. There is no version of this where AI infrastructure suddenly becomes cheap.

But I think the bear case keeps throwing “gross capex required” into “shareholder capital required.” that is not the same thing.

My thesis is that $IREN becomes a capital recycler, not a serial equity issuer.

The early MWs are expensive because you are financing development risk. You are funding power, construction, GPUs, integration all before the asset has really proved itself. But once the asset is live, contracted, accepted and cash flowing, it becomes something else, because it becomes financeable.

A stabilized Microsoft backed facility says here is the contract, here are the cash flows, here is the operating history, here is the collateral. Capital likes this

It all hinges on how much common equity is required per incremental MW. If that number stays high, the bull case breaks, if it falls as the platform scales, then the capex number becomes much less scary. Because then the company doesn’t need to self fund 5.8GW. It only needs to fund the equity slice that unlocks it. And I think that slice should shrink with execution.

Because the current network reduce risk somewhere in the capital stack, because Microsoft reduces credit risk, NVIDIA reduces technology and supply risk, Mirantis reduces operational risk.

Successful commissioning reduces execution risk. Customer prepayments reduce funding risk. And if each one lowers perceived risk, the cost of the next dollar of capital should fall.

That is the hidden compounding mechanism.

I actually think one of the most underappreciated things in this whole story is that $IREN’s real product may eventually become its cost of capital. Sounds boring, but that is how infrastructure empires are built. The issue is value created per share.

If share count goes up 15% but durable cash-generating capacity goes up 40%, that can still work beautifully. If share count goes up 15% and returns on new capital are mediocre, then shareholders are just financing empire building.

So I would not judge $IREN by “did they dilute?” I would judge it by whether equity intensity per MW is falling. And this is also why I don’t want to get hypnotized by $13B of adjusted EBITDA.

The real destination has to be high returns on incremental equity capital after refresh costs and financing costs. If they get to $13B EBITDA but needed $30B of new common equity to get there, Chanos wins.

If they get anywhere close while the majority of growth is funded through prepayments, debt, asset backed structures, GPU financing and retained cash flow, then the whole business has transformed. Then $IREN is no longer just consuming capital, it is recycling it.

$IREN will need enormous gross capital to build out 5.8GW. I do not think shareholders will need to provide most of it because execution changes the cost of the next dollar of capital The bull case is that $IREN becomes progressively less equity intensive.

I’m planting the flag there. > 引用 @RealJimChanos: And how much capital will $IREN have to invest to bring in that $13B of Adjusted EBITDA in 2030…?! https://x.com/pepe_maltese/status/2102159837610098736