下跌倒计时.. 260513

注：本文为原发于260512的星球desk note，部分敏感内容删减。

图：自4月1日入场做多SOXX以来，至今+55%，我们已进一步收紧止盈

The AI trade is starting to r hyme with the dot-com bubble.

AI 交易开始与互联网泡沫有些相似了。

Not because AI is fake. Not because every infrastructure dollar will be wasted. And definitely not because this is another Pets.com story.

并不是因为 AI 是假的。也不是因为每一笔基础设施投入都会被浪费。当然更不是因为这又是一个 Pets.com 式的故事。

The real comparison is much more specific: capital formation is running ahead of validated end-demand.

真正的相似之处要具体得多：资本形成的速度正跑在已被验证的终端需求前面。

That was the problem in the late 1990s. The internet was real. Adoption was real. The mistake was that telecom, fiber, data centers, and hardware suppliers built for an economic future that had not arrived yet. The technology survived. Many of the capital structures did not.

这正是 20 世纪 90 年代末的问题。互联网是真实的。普及也是真实的。错误在于，电信、光纤、数据中心以及硬件供应商，是在为一个尚未到来的经济未来进行建设。技术活了下来，但许多资本结构没有。

The same risk is building again.

同样的风险正在再次形成。

AI demand is real. Enterprise adoption is real. Consumer usage is real. The models are improving, and the product surface area is expanding. But the infrastructure layer is now being financed as if the end-state economics are already proven.

对 AI 的需求是真实存在的。企业采用是真实存在的。消费者使用是真实存在的。模型正在改进，产品的应用边界也在扩大。但基础设施层如今获得融资的方式，仿佛其最终经济效益已经得到验证。

They are not.

并非如此。

That distinction matters. A real technology boom can still create a bubble in the assets used to finance it.

这种区别至关重要。一场真正的技术繁荣，仍然可能在为其融资所使用的资产中制造泡沫。

OpenAI has disclosed more than $20 billion of ARR for 2025. Google has reported hundreds of millions of paid subscriptions. Microsoft has reported tens of millions of paid Copilot seats and broad GitHub Copilot adoption. These are not small numbers. They prove demand exists.

OpenAI 已披露其 2025 年年度经常性收入超过 200 亿美元。Google 已报告拥有数亿付费订阅用户。Microsoft 已报告拥有数千万付费 Copilot 席位，以及 GitHub Copilot 的广泛采用。这些都不是小数目。它们证明了需求的存在。

But they do not yet prove that the entire infrastructure stack has earned the hundreds of billions being committed ahead of it.

但它们还不能证明，整个基础设施堆栈已经赚到了为其提前投入的数千亿美元。

That is the core issue.

这才是核心问题。

The market is not only pricing AI adoption. It is pricing the assumption that AI adoption will grow fast enough, at high enough margins, for long enough, to justify an industrial-scale buildout across GPUs, power, cloud, networking, storage, and data centers.

市场定价的不只是 AI 的采用情况。它定价的还有这样一种假设：AI 的采用将以足够快的速度增长，并在足够长的时间内维持足够高的利润率，从而足以支撑在 GPU、电力、云、网络、存储和数据中心等领域进行工业规模的建设。

That is a much bigger claim.

这是一个大得多的论断。

The financing structure is where the risk becomes visible.

融资结构正是风险显现的地方。

图：The AI Financing Loop

Demand is no longer moving in a straight line from end customers to suppliers. It is moving through cloud commitments, supplier financing, warrants, leases, and strategic capital. That is how revenue can look clean before final demand is fully proven.

需求不再沿着从终端客户到供应商的直线传导。它正在通过云承诺、供应商融资、认股权证、租赁和战略资本传导。这就是为什么在最终需求尚未得到充分验证之前，收入看起来也能很干净。

We are no longer just talking about normal capex. We are talking about long-duration cloud commitments, supplier-linked financing, warrants, GPU-backed debt, customer contracts used to support infrastructure funding, and hyperscalers pulling forward years of capacity.

我们谈论的已不再只是普通的资本支出。我们谈论的是长期云服务承诺、与供应商挂钩的融资、认股权证、以 GPU 为担保的债务、用于支撑基础设施融资的客户合同，以及超大规模云服务商将未来数年的产能提前兑现。

That does not mean every deal is reckless. It means the system has crossed into a regime where growth is being coordinated by balance sheets, financing structures, and strategic commitments.

这并不意味着每一笔交易都是鲁莽的。这意味着系统已经进入一种新的阶段：增长正在由资产负债表、融资结构和战略承诺共同协调。

Once that happens, revenue can accelerate before end-demand is fully validated.

一旦这种情况发生，在终端需求得到充分验证之前，营收就可能加速增长。

This is how late-cycle infrastructure bubbles usually work. The top line looks clean. The order book looks strong. Every supplier points to committed demand. Every customer points to future usage. The entire chain starts validating itself.

这就是后周期基础设施泡沫通常的运作方式。营收看起来很漂亮。订单簿看起来很强劲。每一家供应商都指向已承诺的需求。每一位客户都指向未来的使用量。整个链条开始自我验证。

But the real test is not revenue.

但真正的考验不是营收。

The real test is cash conversion.

真正的考验是现金转化。

Free cash flow is already after capex. That is why this chart matters. The issue is not whether hyperscalers have money today. The issue is whether more of their operating cash flow is being recycled into infrastructure before the final demand curve is fully proven.

自由现金流已经扣除了资本开支。这就是为什么这张图很重要。问题不在于超大规模云服务商今天有没有钱。问题在于，在最终需求曲线尚未得到充分验证之前，他们是否正将更多的经营现金流重新投入到基础设施中。

This is where the AI cycle looks less comfortable. The hyperscalers are much stronger than dot-com companies ever were, but that strength is exactly why the numbers are so large. Amazon’s free cash flow has been compressed by a massive surge in property and equipment purchases. Alphabet still generates enormous free cash flow, but capex is rising rapidly. Microsoft, Meta, Alphabet, Oracle, and Amazon are all pushing infrastructure spending to levels that would have been unthinkable a few years ago.

这正是 AI 周期看起来不那么令人安心的地方。超大规模云服务商远比当年的互联网公司强大，但这种强大恰恰也是数字如此庞大的原因。Amazon 的自由现金流因房地产和设备采购的大幅激增而受到压缩。Alphabet 仍然能产生巨额自由现金流，但资本支出正在迅速上升。Microsoft、Meta、Alphabet、Oracle 和 Amazon 都在将基础设施支出推高到几年前还难以想象的水平。

So the question is not whether these companies can afford to spend.

所以问题不在于这些公司是否花得起钱。

The question is whether the returns on that spending arrive fast enough to prevent the market from changing the discount rate on the whole trade.

问题在于，这些支出的回报是否来得足够快，从而防止市场改变对整个交易的贴现率。

That is where bubbles usually turn.

那通常是泡沫转向的地方。

They do not turn when people suddenly decide the technology is useless. They turn when investors move from “how big can this get?” to “when do we get paid?”

当人们突然认定这项技术毫无用处时，它们并不会转向。它们转向，是在投资者从“这能做多大？”变成“我们什么时候能拿到回报？”的时候。

The dot-com analogy is useful for that reason. In 2000, internet usage was growing. Broadband lines were growing. Online advertising was growing. The problem was that monetization was too small relative to the capital formation built around it.

出于这个原因，互联网泡沫的类比很有用。2000 年，互联网使用量在增长，宽带线路在增长，在线广告也在增长。问题在于，相对于围绕它所形成的资本积累，变现规模太小了。

Today, AI monetization is much larger than early internet monetization. But the infrastructure commitments are also vastly larger.

如今，AI 的变现规模远超早期互联网的变现规模。但基础设施投入也大得多。

That is the part investors need to respect.

这是投资者需要尊重的部分。

The bullish case says AI usage will eventually grow into the infrastructure. That may be right. But “eventually” is not a valuation input by itself. The market still has to bridge the gap between current monetization and current capital intensity.

看涨的观点认为，AI 的使用最终会融入基础设施之中。这或许是对的。但“最终”本身并不是估值的依据。市场仍然必须弥合当前变现能力与当前资本密集度之间的差距。

If that gap narrows, the cycle survives.

如果这个差距缩小，循环就会继续存在。

If that gap widens, the weakest part of the trade breaks first.

如果这种差距扩大，交易中最脆弱的部分会最先崩溃。

This is where storage becomes interesting.

这就是储存变得有意思的地方。

Memory is carrying a large part of the semiconductor acceleration. That is why storage is not the original cause of the AI bubble. It is the late-cycle amplifier: fundamentally real, but tactically dangerous.

存储器承载了半导体加速增长的很大一部分。这就是为什么存储并不是 AI 泡沫的原始成因。它是周期后段的放大器：从根本上说是真实的，但在战术层面上却很危险。

Storage sits downstream from GPU clusters and AI workloads, but it is still several steps away from the final customer monetization event. That means storage demand can look strongest exactly when the cycle is most extended. As long as clusters are being built, storage demand follows. As long as training and inference workloads keep expanding, storage looks indispensable.

存储位于 GPU 集群和 AI 工作负载的下游，但距离最终面向客户的变现环节仍有好几步之遥。这意味着，恰恰在周期被拉得最长的时候，存储需求看起来可能最为强劲。只要集群还在建设，存储需求就会随之而来。只要训练和推理工作负载持续扩张，存储看起来就不可或缺。

But if capex gets questioned, storage can reprice before the platforms do.

但如果资本开支受到质疑，存储可能会先于平台重新定价。

That is why the storage trade can feel both fundamentally justified and tactically dangerous at the same time. The demand is real. The cyclicality is also real.

这就是为什么储能交易会同时让人感觉在基本面上站得住脚，但在战术上又充满危险。需求是真实存在的，周期性也同样真实存在。

In late-cycle infrastructure trades, the best numbers often appear near the top. Revenue acceleration, backlog strength, margin expansion, and management confidence can all peak right before investors start questioning the durability of the order book.

在周期后段的基础设施交易中，最漂亮的数据往往出现在顶部附近。营收加速、积压订单强劲、利润率扩张以及管理层信心，都可能在投资者开始质疑订单簿可持续性之前见顶。

The trigger is unlikely to be a weak chatbot demo.

触发因素不太可能是一场表现糟糕的聊天机器人演示。

The trigger is more likely to come from the financing layer.

触发因素更可能来自融资层面。

Watch free cash flow. Watch supplier-backed credit. Watch customer concentration. Watch whether capex guidance keeps rising without a matching improvement in monetization. Watch whether management language shifts from “capturing demand” to “earning returns.”

关注自由现金流。关注由供应商支持的信贷。关注客户集中度。关注资本开支指引是否在持续上升，却没有伴随变现能力的相应改善。关注管理层的措辞是否从“抓住需求”转向“获得回报”。

That language shift matters.

这种措辞上的转变很重要。

When companies talk about capturing demand, investors hear growth.

当公司谈论抓住需求时，投资者听到的是增长。

When companies start talking about returns on invested capital, investors hear discipline.

当企业开始谈论投资资本回报率时，投资者听到的是纪律性。

And discipline is not bullish for every supplier in the chain.

而这种纪律性并不利好产业链上的每一家供应商。

This can break tomorrow, or it can squeeze higher for a few more days or weeks. That is not the point. When positioning, headlines, and momentum are all pointing in the same direction, late-cycle trades can still extend beyond what looks reasonable.

这明天就可能崩掉，也可能再被挤高几天或几周。但重点不在这里。当持仓、头条消息和动能都指向同一个方向时，周期后段的交易仍然可能延续到超出看起来合理的程度。

But I would not confuse that with a healthy setup. The longer this leg extends, the more fragile the trade becomes, because the market is no longer just paying for growth. It is paying for the assumption that the financing loop will keep converting into real cash flow.

但我不会把这和健康的格局混为一谈。这一波走势拖得越久，这笔交易就越脆弱，因为市场如今买单的已不只是增长。它买单的还是这样一种假设：融资循环会持续转化为真实的现金流。

AI can still change everything.

AI 仍然可以改变一切。

The winners can still be real.

赢家依然可能是真实存在的。

The technology can still be one of the most important platform shifts in decades.

这项技术仍然可能是数十年来最重要的平台变革之一。

But none of that eliminates the bubble risk.

但这一切都不能消除泡沫风险。

The bubble is not in the idea that AI matters.

泡沫并不在于认为 AI 很重要这一观点。

The bubble is in the assumption that every layer of the infrastructure stack will be paid for on time, at attractive returns, without a capex pause, financing accident, or monetization gap.

泡沫在于这样一种假设：基础设施堆栈的每一层都会被按时买单，并且都能获得可观回报，不会出现资本开支暂停、融资意外或变现缺口。

That is the part of the dot-com analogy that matters.

这才是类比互联网泡沫中真正重要的部分。

Real technology.

真实的技术。

Real adoption.

真实的采用。

Real overbuild risk.

真正的过度建设风险。

It only needs the market to realize that the financing loop got ahead of the cash.

它只需要市场意识到，融资循环已经跑在现金流前面了。

Tick Tock... Tick Tock...

滴答……滴答……

完整内容见星球：https://t.zsxq.com/aUUD2

本文所涉及一切内容仅供分享参考，接收人不因阅读本文而被视为我方客户，也不构成向任何人发出出售或购买证券或其他投资标的的邀请，更不构成交易依据。市场有风险，投资需谨慎，本文所有信息及观点均不构成任何投资建议；过往业绩不代表未来表现，投资者应依据自身风险承受能力进行审慎评估、独立决策并自行承担全部责任。任何情况下，我方均不对因使用本文内容而产生的任何损失承担责任。相关数据可能存在滞后或误差，我方不对信息的完整性及准确性承担法律责任。