# AI capex — X 热门讨论 (2026-10-11 03:05 UTC)

## @KelseyDingMD (Galaxy_Chloe) · 10-10 22:13 · ♥130 ↻16 💬7 李录这篇，分析水准很高，对半导体行业的未来走势其实给了很清晰的框架。

“上游资本开支对应的订单增长，并不意味着下游能够同步获得相同规模的利润”。反过来看，这恰恰是半导体的位置红利：AI产业链里，钱先流进芯片厂的口袋。再加上“新增信息分发的边际成本可以压得很低，但AI每生成一次回答、执行一次Agent任务，都需要继续消耗GPU算力、电力和存储资源”。算力需求不是一锤子买卖，是持续抽水。只要AI调用量还在涨，半导体的需求曲线就比市场想象的更长。

但李录的提醒，同样是写给半导体行业的：“AI的发展可能经历不止一轮繁荣与崩溃”。这一点非常符合我对半导体未来走势的判断，不单单是说半导体强周期，历史上一轮轮boom-bust就没停过，这一轮AI capex的繁荣期，订单爆不等于永远爆。“技术进步可以持续很多年，资本市场却未必会沿着同一条曲线定价”。技术叙事不能直接翻译成股价，巨额资本投入最终要靠下游的收入、利润、现金流来消化。

看半导体，别只看技术多性感，要看capex能不能持续、看下游客户的钱是不是真能赚回来。繁荣期吃肉，萧条期活下来。这才是这个行业的完整剧本。技术曲线向上，定价曲线震荡，中间每一轮崩溃都是对半导体持仓的极大考验。 > 引用 @JohnsonZ91127: 喜马拉雅资本创始人李录在哥伦比亚商学院第36届格雷厄姆与多德年度论坛上谈了不少AI。他认可这场技术革命的长期价值，但认为AI的发展可能经历不止一轮繁荣与崩溃。技术能走多远，除了取决于算力和算法，还受到资本投入、商业回报以及社会治理的约束。

李录对生成式AI的理解相当克制。他认为，AI利用庞大的算力和算法，从人类创造的数据中识别复杂关联，逐渐具备推理、编程和知识处理能力。它可以大幅降低知识和专业服务的获取门槛，但目前仍主要建立在人类积累的知识体系之上。能源、交通、制造业这些物理世界的约束不会因此消失。对于AI脱离人类控制、持续自我改进的可能性，他尤其警惕，认为人类必须保留监督和干预能力。

商业层面有个很现实的问题：AI与传统互联网的成本结构差别很大。 互联网基础设施建好之后，新增信息分发的边际成本可以压得很低，但AI每生成一次回答、执行一次Agent任务，都需要继续消耗GPU算力、电力和存储资源。模型训练需要大量前期投入，推理服务还会随着调用量增长产生持续成本。即使单次Token价格不断下降，任务复杂度和调用规模上升后，整体算力支出依然可能增加。

这也解释了AI产业链不同环节的盈利差异。英伟达、博通、台积电等上游供应商，可以直接从芯片采购和制造订单中获得收入；微软、谷歌、Meta等大型科技公司则需要承担数据中心建设、设备折旧和运营成本，再通过云服务、广告效率提升和企业客户付费收回投资。模型厂商还要面对推理成本、价格竞争和持续研发投入。上游资本开支对应的订单增长，并不意味着下游能够同步获得相同规模的利润。

李录还谈到资本过度集中的问题。他举标普500指数为例，2010年以来约16年的年化回报达到14%—15%，但2000—2010年几乎没有回报。把过去26年放在一起，长期年化回报又回到高个位数。市场可以在相当长的时间里偏离合理估值，但资本不可能无限增加。当大量资金集中追逐少数热门公司，另一边往往会出现一些经营质量不错、价格却受到冷落的企业。他认为，这种分化在美国、中国及其他市场都存在。

对于价值投资者而言，技术创新和股东回报始终需要分别衡量。AI能够提高生产率，扩大知识服务的供给，但巨额资本投入最终仍要通过收入、利润和现金流来消化。技术进步可以持续很多年，资本市场却未必会沿着同一条曲线定价。 这也是李录判断AI可能经历多轮繁荣与崩溃的重要原因。 https://x.com/KelseyDingMD/status/2109044801148793275

## @ChairmansLedger (The Chairman's Ledger) · 10-10 17:09 · ♥126 ↻4 💬14 I'm genuinely struggling to read this market.

On one side, the AI buildout looks as compelling as ever. Demand for compute is enormous and obvious, capacity is scarce and the scale of the contracts being signed is truly insane.

On the other, the market is hammering many of the cos expected to deliver that capacity. Rates are high, financing is expensive and suddenly everyone is asking whether the economics justify the capex.

I understand both sides. A co can have incredible demand and still be a terrible investment if the cost of delivering it under delivers returns. Equally, the market can price in a financing disaster that never shows up - it in of is already with a lot of circular financing (much of the economy is like this anyway but it's harder to see).

What's strangest to me is the broader market sitting near ATHs while parts of AI infra feel like a full blown bear market. A bunch of names down 50%+ easily.

My conviction in the LT buildout hasn't changed. I think the market is badly mispricing some of these businesses and that several will be worth substantially more in a few yrs, like orders of magnitude more.

But here's something important about how I approach risk.

I'm generally very bullish and willing to push the limits on position sizing. The single most important thing for me is that I cannot be forced to sell.

I don't use margin. Most of my wealth sits outside this public equity book in a safety sleeve of fixed income, bonds, uncorrelated HFs and other investments.

That structure affords me the luxury of being wrong. I can be wrong on timing for yrs, or even wrong outright on individual investments, without jeopardizing my financial position or being forced to liquidate at the worst possible time.

That's how I reconcile my appetite for risk with my approach to managing it. I'm willing to tolerate enormous volatility in this part of my portfolio precisely bc I've structured the rest of my wealth to withstand it.

I don't know when this market will turn or what finally changes sentiment. But I remain extremely optimistic on what's being built, and I've put myself in a position where I can afford to wait and my hunch is that that's a good call. https://x.com/ChairmansLedger/status/2108968243507364067

## @rev_cap (Rev Cap) · 10-10 23:26 · ♥59 ↻1 💬36 If AI capex hits a wall, do non AI stocks blow up or are we in a bit of a technical recession and it all goes down? https://x.com/rev_cap/status/2109063074053529933