人工智能泡沫担忧是确有依据还是夸大其词？

是否存在AI泡沫？

Is there an AI bubble?

我们都听过AI的利好观点，认为我们正处于一场将改变世界的技术革命的早期阶段，而引领这场革命的公司将为投资者带来巨大回报。

We've all heard the bull case for AI, that we're in the early innings of a technological revolution that will change the world, and the companies leading this revolution will generate tremendous returns for their investors.

但经过多年的大规模投入和股价上涨，我们开始听到越来越多的质疑声。

But after years of heavy spending and rising stock valuations, we're starting to hear a lot more skepticism.

那么是否存在泡沫的迹象呢？

So are there signs of a bubble?

如果我们确实身处泡沫之中，这对投资者意味着什么？

And if we are in a bubble, what does it mean for investors?

我是艾莉森·内森，这里是高盛交易所。

I'm Alison Nathan, and this is Goldman Sachs Exchanges.

每个月，我都会与投资者、政策制定者和学者们讨论市场最紧迫的动向问题，为高盛研究部的《焦点报告》提供见解。

Each month, I speak with investors, policymakers and academics about the most pressing market moving issues for our top of mind report from Goldman Sachs Research.

本月，我与高盛研究部的两位同事进行了交流，他们来自我们的美国分部。

This month, I spoke with two of my colleagues here in Goldman Sachs Research, our U.

分部。

S.

互联网股票研究分析师埃里克·谢里丹与我们美国的

Internet equity research analyst, Eric Sheridan, and our U.

团队

S.

软件行业股票研究分析师，卡什·兰根。

Software equity research analyst, Kash Rangan.

我首先询问了Eric和Kash关于AI建设当前进展的情况，以及与预期相比如何。

I started by asking Eric and Kash about where the AI build out stands today and how that compares to expectations.

在基础设施层面，资本投入和支出规模超出了预期。

On the infrastructure layer, the amount of capital and the amount of spend has surprised to the upside.

短期内这种支出水平主要与服务需求相关，由查询GPT产生的计算需求超出了现有容量。

That level of spend short term is mostly tied to the fact that demand for these services, the compute need generated by you querying GPT is outstripping the available capacity.

因此基础设施领域在满足服务需求的资本投入方面超出了预期。

So infrastructure has surprised to the upside on the need for capital to meet the demand for services.

平台层可以说是少数几家公司，它们正从单纯运行基础模型转向构建API解决方案或在基础模型之上开发应用。

The platform layer arguably is the handful of companies that are transitioning from just running foundational models to either building API solutions or applications on top of the foundational model.

目前只有少数几家公司开始崭露头角，它们拥有执行这一转型所需的资本规模和人才储备。

And there's only a handful of those companies that have started to emerge that have the scale of capital and talent to execute against that.

我们看到更多消费级应用涌现，主要是通过消费者使用ChatGPT和Google Gemini，企业级应用方面我将留给Cash来讨论。

We've seen more applications emerge on the consumer side mostly through the usage of ChatGPT and Google Gemini by consumers, and I'll leave it to Cash to talk about the application side for enterprise.

没错。

Right.

好的。

Okay.

Kash，我想听听你的看法。

Kash, I do wanna get your take.

基础设施建设持续的时间比任何人预期的都要长得多。

The infrastructure build out has gone on a lot longer than anybody expected.

但与此同时，当我观察我的研究领域时，基础设施层的这种活动如何渗透到平台层正变得更加明显。

But at the same time, when I look at my coverage area, how this activity at the infrastructure layer is percolating up to the platform layer is starting to become more discernible.

平台层如今的状态比一年前、两年前要好得多。

The platform layer is in a much better position today than it was a year ago, two years ago.

令人失望的是应用层。

The disappointment has been at the application layer.

正如埃里克所说，许多消费级应用正在体现AI的价值，无论是聊天GPT还是面向消费者的云端应用。

To Eric's point, a lot of consumer applications, they are exemplifying the value of AI, whether it's chat GPT or cloud application at the consumer level.

但在企业级和终端用户层面，虽然出现了一些生机，但尚未达到我们的预期。

But at the enterprise level, end user level, there are some signs of life, but we're not where we expected.

如果让我猜测这些公司今天能达到什么水平，我本会给出一个数字，而我们现在远低于这个数字。

If you asked me to guess at where these companies would be today, I would have given you a number, and we would be well below the number.

但我们正在接近目标。

But we're getting there.

虽然没达到我们一两年前的预期，但相比六个月、九个月前，我们已开始看到这些应用的雏形。

It's not where we expected it to be a year ago, two years ago, but relative to where we were six months ago, nine months ago, we're starting to see these applications.

但归根结底，埃里克，资本支出投入如此之大——甚至超过去年我们面对惊人高数字时的预期——回报潜力...

But ultimately, Eric, there is so much CapEx going in so much more than we even thought when there was amazingly high numbers last year, the return potential.

这个潜力也增长了吗？

Has that grown too?

我要说明几点。

So I'll make a couple of points.

总体支出的增长是个非常庞大的数字，这导致许多投资者开始认真追问投资回报率的问题。

The rise in spending in aggregate is a very large number, and it has resulted in a number of investors really asking the ROI question.

这个质疑声只会越来越大，不会减弱。

That question is only going to build in scale, not abate in scale.

我真的不认为现在有任何动力能让人停止采取攻势。

And I really don't think there's any incentive for anyone to stop playing offense today.

英伟达最近公布了一个数字：从现在到本年代末，投入将在3到4万亿美元之间。

NVIDIA recently put out a number of 3 to $4,000,000,000,000 between now and the end of the decade.

我认为我们接触的大多数投资者都难以证明3到4万亿美元累计支出的回报合理性，除非AI成为社会某种终极状态下经济产出的主要驱动力。

I think most investors we talk to would struggle to justify the return profile on 3 to 4,000,000,000,000 of cumulative spend unless AI is the main driving factor in an enormous amount of the economic output of society in some sort of end state.

话虽如此，在过去的计算周期中，像我这样的人往往无法展望六、七、八、九、十年后并断言未来图景。

That being said, there have been computing cycles where folks like me have not been able to look out six, seven, eight, nine, ten years and say, this is what it's gonna look like.

在不同时间节点，比如我们过度建设桌面计算时，Netflix就应运而生了。

At different points in time, it was we overbuilt to desktop computing, and then Netflix was created.

接着浏览器大战爆发，门户网站大战上演，所有这些都极大地推动了桌面端使用率。

And the browser wars happened, and the portal wars happened, and all of those things drove a lot more desktop usage.

当年建设频谱、基站和无线网络时，没人预料到30亿人将拥有具备这种电池续航和运算能力的智能手机。

There was nobody when Spectrum and towers and wireless was being built that thought 3,000,000,000 people would have smartphones with this battery capability and power capability.

事物确实在变化，我认为最前沿的企业会基于长期视野进行信心投资。

Things do change, and I think the most leading edge companies invest with confidence against the long term time horizon.

来自投资者的质疑将持续增加。

The questions from investors will continue to rise.

坦白说，如果资金规模继续膨胀，以我们现有的认知将难以应对这些质疑。

And I think if the dollars keep rising, I'll be brutally frank, we'll struggle to answer them with what we know today.

根据我分析过的每个计算周期规律，这通常会导致某个阶段出现幻灭低谷——要么是支出问题，要么是采用率问题，或是两者共同导致无法在六到十二个月内给出令人满意的答案。

And, typically, through every computing cycle I've ever analyzed, that leads to a trough of disillusionment at some point where either the spend or the adoption rate or some combination of the two don't give a satisfactory six, twelve month answer.

而这次我们能否避免这种情况呢？

And are we going to avoid one this time?

如果我们能避开泡沫，我会非常震惊。

I would be shocked if we avoided one.

在我们研究过的任何技术周期中，通常只有两到三家公司能覆盖资本成本并获得超额回报。

In any technology cycle that we've covered, there are typically two to three companies that earn their cost of capital and earn an excess return on that cost of capital.

一个行业里有四五六七家公司都在做同样的事，却都能在同一垂直领域或产品上获得超额回报，这种情况并不常见。

The idea that there are four, five, six, seven companies in an industry that do something and they all earn an excess return in the same vertical or the same product is is not typical.

我不明白为什么这次会有所不同。

And I don't know why this should be any different.

显然我们进行这场对话，是因为看到科技巨头的表现和当前估值水平，市场对泡沫形成的担忧日益加剧。

Obviously, we're having this conversation because the concerns that there is a bubble forming in the markets when you see how mega cap tech has performed, the valuations where they are today.

你认为当前是否已进入泡沫区域？你在关注哪些指标来评估这一点？

Do you think concerns that we are now in bubble territory are there, and what are you watching to assess that?

我并非想轻率地回答这个问题。

I'm not trying to be flip about this question.

三年来关于泡沫的讨论一直持续不断。

There's been more talk of a bubble in a continuous nature for three years.

我经历过的另外两次泡沫时期，身处泡沫中时反而没这么多讨论。

Two other bubbles I lived through, there wasn't this much talk of the bubble while we were in the bubble.

那么，我是否看到了类似九十年代末或零七年时期的迹象？

So, look, do I see signs that point me back to the late nineties or the o seven time period?

确实有。

Sure.

私募市场的估值比公开市场高出许多。

Private market valuations are well ahead of public market valuations.

公开市场估值高于历史常态，但仍低于1999年和2000年所见的峰值水平。

Public market valuations are above historical norms, but they're also below the peak of public market valuations that you saw in '99 and 2000.

资本市场活动仍远低于我们在2020、2021、2007、2008、1998和1999年看到的水平。

Capital market activity is still well below the levels that we're seeing in 2020, 2021, 2007, 2008, 1998, 1999.

因此我认为存在过度繁荣的迹象。

So I would argue there are signs of exuberance.

有些迹象与历史时期相似，但我不认为它完全符合我们从前几轮周期中学到的教训，至少目前还没有。

There are signs that rhyme with past periods of time, but I wouldn't necessarily align it perfectly with some of the lessons we've learned in prior periods, at least not yet.

可以说这是个时间维度的问题——我的意思是现在还没到那个阶段。

Now that's arguably a duration narrative that I'm saying that we're just not there yet.

这是我给出的一个分析框架。

That would be one framing I would give.

另一个与以往时期不同的是，通常那些不盈利的公司...

The other that is a little bit different than prior periods, typically, the companies that generate no profits.

而在1999年，是那些没有收入的公司推动了市场上最狂热的估值。

And in 1999, it was companies that generated no revenue are the ones that were driving the most exuberant valuations in the market.

'七巨头'中的大多数公司都创造了超常水平的自由现金流，并回购股票和支付股息。

The magnificent seven, most of those companies generate outsized levels of free cash flow and buy back their stock and pay dividends.

1999年几乎没有公司回购股票和支付股息。

There were very few companies buying back their stocks and paying dividends in 1999.

所以除了简单类比外，还需要注意这些关键区别。

So there are some key distinctions I would keep in mind as opposed to just drawing a straight analogy.

纳什，你对泡沫担忧和我们是否可能身处其中有什么看法？

Nash, do you have any thoughts about bubble concerns and signs that we may or may not be in one?

在我所关注的股票中肯定没有这种情况。

Certainly not in the stocks that I cover.

许多软件股正因AI可能对其终端市场造成的影响而估值低迷——无论是就业岗位转移影响消费，还是AI能以更经济高效的方式编写软件（比如通过管道编码等技术），这显然降低了对Salesforce等公司应用软件包的购买需求。

A lot of software stocks are trading at depressed valuations because of what AI could presumably do to their end markets, whether it's job dislocation towards consumption or AI enables you to write software in a cost effective manner because you got pipe coding and other things that obvious the need to buy an application software package from the likes of Salesforce, etcetera.

因此我们正经历一场生存危机，目前市场给予的是确定性折价而非溢价。

So we're going through an existential crisis right now that there is a decided discount rather than a premium.

本轮周期最与众不同的特点（这甚至会让通常谨慎的人也稍减戒心）在于：这些资本究竟从何而来？

And something that is so different about this cycle, which should make even somebody that's normally a bit circumspect a little bit less cautious is where is this capital coming from?

埃里克和我经历过九十年代末，那时投入高风险项目的资金主要来自风投和私募资本。

Eric and I lived through the late nineties, and we saw the capital that was deployed in high risk projects was venture capital and private capital.

而这次，资本来自资金雄厚、财力充沛的超大规模科技巨头。

This time, the capital is coming from the well heeled, deep pocketed hyperscaler giants.

它们的资金成本相当低廉。

Their cost of capital is quite low.

它们用这笔巨资承担风险的能力非常强。

Their ability to take risk with this amount of money is is quite high.

即便意味着要经历几轮试错才能找到AI价值变现的最佳路径，它们也完全能承受。

And even if it means a couple of cycles of trying and finding out what the best way to extract value from AI, they can do it.

因此，来自完全不同投资群体的巨额资本供给，使得本轮周期更具容错性——在我们最终确立AI商业模式前，必然会出现大量失误和纰漏。

So the availability of so much capital coming from a very different audience, a different set of investors makes the cycle a little bit different and more tolerant of SNAFUS and missteps, which we will have plenty of these things before we nail the ultimate AI business model.

但我们开始看到新趋势：有些实体正以80%债务和20%股权的融资结构组建。

But we're starting to see a new development whereby entities are being put together that are funded with 80% debt, 20% equity.

即便在这20%的股权部分中，也有由发起实体提供担保的抵押品。

And within even within the equity piece, there is a collateral that's backed to a sponsoring entity.

这些实体能够以极低的资本成本发行债务。

And these entities are able to issue debt at very low cost of capital.

但我们需要警惕的一个风险因素是，系统中的杠杆效应正在显现。

But something we should be aware of as a risk factor is that leverage in the system is starting to emerge.

我们必须确保系统正常运转，最终这些推动资本需求的公司——无论是基础模型公司还是其他类型——都能达成它们的财务目标。

We have to make sure that the the system works, that ultimately these companies that are driving the need for capital, whether it's the foundation model companies, whatnot, are able to hit their numbers.

我们需要谨慎监控这种财务模式是否可行，因为在一个低毛利率的商业模式上叠加了多重杠杆。

We need to be carefully monitoring if this math works because there's a lot of leverage upon leverage on top of a low gross margin business model.

因此需要积极监控。

So active monitoring needed.

没错。

Right.

让我把这个问题抛回给你，埃里克。因为归根结底，英伟达向OpenAI投资了1000亿美元，而OpenAI又承诺向甲骨文购买3000亿美元的云计算服务——尽管他们实际上并没有这笔资金。

And let me put that back to you, Eric, Because, ultimately, at the end of the day, we have NVIDIA investing a 100,000,000,000 in OpenAI, and then we have OpenAI pledging 300,000,000,000, which they don't really have on cloud compute from Oracle.

然后甲骨文又在购买英伟达的芯片，而英伟达又在投资英特尔。

And then Oracle's buying NVIDIA chips, and then NVIDIA is investing in Intel.

我是说，这种循环性难道不会让你感到不安吗？

I mean, the circularity of all this, does that make you nervous at all?

确实会。

It does.

正如我所说，我认为当前正在发生的一些事情与1998、1999年的情况有相似之处。

As I said, I think there are things you can point to going on right now that rhyme with 1998, 1999.

这就是一个例子。

That would be an example of it.

我在九十年代末和二十一世纪初以电信分析师和电信投资者的身份开始了职业生涯。

I started my career as a telecom analyst and a telecom investor in the late nineties and early o o's.

那是环球电讯、Level 3和Qwest的时代，这些公司用债务相互交易带宽容量，一个人的收入就是另一个人的容量，反之亦然。

The era of Global Crossing and level three and Quest, and these companies were trading capacity with each other using debt, and one person's revenue was another person's capacity and vice versa.

当债务曲线过高时，整个体系就崩溃了。

And when the debt curve got too high, it all fell apart.

当你理清所有收入时，会发现彩虹尽头远没有你想象的那么多收入。

And when you untangled all the revenue, there was nowhere near the revenue you thought there was at the end of that rainbow.

这些带宽容量最终在2003至2005年间被市场消化吸收。

That capacity was eventually absorbed by 2003 and 2004 and 2005.

但根据你在市场中的入场时机不同，你的回报可能在很长时间内都相当低。

But depending on where your entry point was in the market, your return could have been quite low for very long periods of time.

当你开始看到债务出现，看到公司相互投资，看到供应商投资于提供带宽的公司时，投资者自然会——也理应——质疑这些关联能否理清，并确保系统不会产生复合效应。

When you start seeing debt, when you start seeing companies investing in other companies, when you start seeing suppliers invest in a company that delivers capacity, investors are naturally and correctly asking questions about untangling at all and then making sure that there isn't compounded effects that can be built up in the system.

最后一个问题：Eric，什么情况会让你变得不那么乐观？

Very last question, which is, Eric, what would make you less optimistic?

我持续关注使用率增长、采用率提升、货币化进程和自由现金流的状况。

I continue to monitor rise in utility, rise in adoption, monetization, and free cash flow.

说到底，如果企业的支出方式最终危及其自由现金流生成能力，这将很难自圆其说。

I mean, at the end of the day, it'll be very hard to argue if companies spend in a way that eventually puts their free cash flow generation in real risk.

这通常可能成为市场的转折点。

And that typically can be a tipping point in the market.

比如削减股息、减少股票回购这类行为，以及过度使用债务。

People cutting dividends, cutting buybacks, things like that, excessive use of debt.

这些正是我密切关注的事项。

Those are things that I am watching for.

那么，凯什，什么因素会让你变得不那么乐观呢？

And, Kesh, what would make you less optimistic?

我想回到之前提出的观点。

I'd go back to the point that I made earlier.

在信贷周期初期，新实体主要通过大量举债而非资产负债表中的现金获得融资。

At the beginnings of a credit cycle where new entities are being funded with a lot of debt as opposed to cash from the balance sheet.

对超大规模企业的担忧较少，更担心的是若信贷周期不配合可能引发的连带损害。

Less worried about the hyperscaler, but more worried about the collateral damage in case credit cycle does not really cooperate.

这将是我重点关注的方面。

That's what I'd be watching.

当这个周期不配合时，将会在整个科技生态系统中产生连锁反应。

Where that cycle does not cooperate, it will have a ripple effect through the rest the tech ecosystem.

我们就此打住吧。

Let's leave it there.

感谢埃里克·谢里丹和卡什·兰根，也感谢大家收听本期高盛交易所节目。

My thanks to Eric Sheridan and Kash Rangan, and thank you for listening to this episode of Goldman Sachs Exchanges.

我是艾莉森·内森。

I'm Alison Nathan.

本文所述观点截至发布之日，可能随时变更且不另行通知，未必反映高盛集团或其关联机构的整体观点。

The opinions and views expressed herein are as of the date of publication subject to change without notice and may not necessarily reflect the institutional views of Goldman Sachs or its affiliates.

本资料仅供信息参考，不构成投资建议、高盛任何实体采取特定行动之推荐，亦非购买或出售任何证券或金融产品的要约或招揽。

The material provided is intended for informational purposes only and does not constitute investment advice, a recommendation from any Goldman Sachs entity to take any particular action, or an offer or solicitation to purchase or sell any securities or financial products.

本材料可能包含前瞻性声明。

This material may contain forward looking statements.

过往表现并不预示未来结果。

Past performance is not indicative of future results.

高盛集团及其任何关联公司均未就本材料所含声明或信息的准确性或完整性作出任何明示或暗示的陈述或保证，并对任何目的下依赖该信息的行为概不承担任何责任。

Neither Goldman Sachs nor any of its affiliates make any representations or warranties expressed or implied as to the accuracy or completeness of the statements or information contained herein and disclaim any liability whatsoever for reliance on such information for any purpose.

提及的第三方机构名称均属相关公司所有，此处使用仅严格出于信息识别目的，并非暗示任何此类公司与高盛之间存在所有权或授权关系。

Each name of a third party organization mentioned is the property of the company to which it relates, is used here strictly for informational and identification purposes only and is not used to imply any ownership or license rights between any such company and Goldman Sachs.

为方便起见提供文字记录，可能与原始音视频内容存在差异。

A transcript is provided for convenience and may differ from the original video or audio content.

文字记录。

Transcript.

未经高盛书面明确同意，任何接收者不得以任何形式复制、分发、出版、全部或部分转载本材料，亦不得向任何他人披露。

This material should not be copied, distributed, published, or reproduced in whole or in part or disclosed by any recipient to any other person without the express written consent of Goldman Sachs.

适用于本材料提及发行人（如有）的研究披露信息，可通过您的高盛代表或访问www.gs.com/research/hedge.html获取。

Disclosures applicable to research with respect to issuers, if any, mentioned herein are available through your Goldman Sachs representative or at www.gs.com/research/hedge.html.

高盛不认可任何候选人或政党。

Goldman Sachs does not endorse any candidate or any political party.

©2025年，高盛集团。

Copyright 2025, Goldman Sachs.

保留所有权利。

All rights reserved.