最令人担忧的AI泡沫信号

随着Anthropic的IPO越来越近，关于AI泡沫的讨论也越来越热烈。

As the Anthropic IPO inches closer, discussions of an AI bubble are also getting a lot louder.

这些担忧包括闲置在亚马逊仓库里的GPU、新型云债务的质量，以及企业AI收入的集中度。

The concerns range from GPUs sitting in Amazon warehouses unused, to the quality of neocloud debt, to the concentration of enterprise AI revenue.

本期节目，我们将逐一分析这些担忧，讨论哪些实际上最令人担忧，而对于那些不那么令人担忧的，未来什么情况会让它们变得更令人担忧。

In this episode, we're going to go through all of those concerns, talking about which are in fact most concerning, and for those that aren't, what would make them more concerning in the future.

《AI每日简报》是一档每日更新的播客和视频节目，内容涵盖AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了，朋友们，在开始之前，先快速说几个通知。

All right, friends, quick announcements before we dive in.

首先，感谢今天的赞助商：毕马威、Robots & Pencils、CData和Blitzy。

First of all, thank you to today's sponsors, KPMG, Robots & Pencils, CData, and Blitzy.

想要收听无广告版本，请访问patreon.com/aidailybrief，或者也可以在Apple Podcasts上订阅。

To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts.

如果你想了解更多关于赞助节目的信息，请发送邮件至sponsors@aidailybrief.ai。

And if you want to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai.

我们这次要聊一个好久没聊的话题，就是AI泡沫。

We're doing something we haven't done for a while, which is an AI bubble episode.

最近，越来越多的人在说AI泡沫已经走到头了。

Now, recently, the calls have been getting louder that the AI bubble has reached the end of the road.

周三，美国金融改革组织发布了一份报告，警告说泡沫用他们的话讲就是“随时会爆”。

On Wednesday, Americans for Financial Reform produced a report warning that the bubble is, in their words, "primed to pop."

高级政策分析师玛雅·詹金斯警告说：“如果情况不变，普通劳动者将为AI崩盘付出高昂代价。”

Senior policy analyst Maya Jenkins warned, "If nothing changes, working people will pay a high price for an AI crash.

政策制定者应该赶紧采取行动来降低风险。”

Policy makers should be taking urgent action to reduce the risks."

当然，关于泡沫的讨论自从去年成为焦点以来，已经变了不少，过去一个月里，大家也指出了很多警示信号。

The bubble discourse has, of course, evolved quite a bit since its dominant place in the narrative last year, and there have been many things that people have been pointing to as warning signs over the past month.

我们今天要聊的是，这些信号里哪些最重要，有哪些理由让我们乐观，或者至少不那么担心，以及如果看到哪些情况发生，会让我更担心。

What we're going to talk about today is which of these signs matter most, what the reasons for optimism, or at least reduced fear are, and what things in each of those cases, if we saw them happen, would make me more concerned.

不过，先简单介绍一下背景。

First, though, a bit of background.

科技股已经连续下跌三周了，周四纳斯达克指数下跌了3%，而之前一周基本持平。

We're coming up on a three-week downtrend for tech stocks, with the NASDAQ losing 3% on Thursday following a relatively flat week.

有人把原因归咎于Meta周三晚上的财报电话会议，会上明确表示，他们下一阶段的AI建设将部分通过增发股票来融资。

Some blamed Meta's Wednesday night earnings call, which made it clear that the next leg of their AI build-out would be partially funded by equity dilution.

不过，这波抛售已经持续了好几周了。

This sell-off, though, has been happening for weeks.

高盛在最新的交易流向报告中指出，上周对冲基金以至少十年来最快的速度抛售科技股。

In their latest trading flow report, Goldman Sachs reported that hedge funds sold tech stocks at their most rapid pace in at least a decade last week.

周四的市场表现也出现了一些有趣的变化。

We also got an interesting change in behavior during the Thursday session.

过去几个月，国债收益率一直在飙升，推高了商业债券的利率，也让数据中心融资变得更贵。

For the past few months, treasury bond rates have been surging, dragging up interest rates on commercial bonds and making data center financing more expensive.

周四，超长期30年期美国国债收益率跌至一个月低点5.38%。

On Thursday, the ultra-long 30-year US Treasury bond dropped to a one-month low of 5.38%.

现在说这个还为时过早，目前短期债券的收益率还没有下降，但这表明市场出现了典型的避险反应。

Now, this is early, and none of the shorter-dated bonds have seen falling rates so far, but this is indicative of a classic flight to safety reaction in the markets.

换句话说，就是更多资金从AI这类高风险资产中撤出，转向国债这种安全港，这可能意味着后面还会有更多麻烦。

In other words, more money flowing out of risky assets like the AI trade and into the safe havens of treasuries, which could spell more trouble ahead.

当然，过去几周也出现了不少看空的分析。

Certainly, the past few weeks have also seen quite a bit of bearish analysis.

一个典型的例子就是Gartner，他们抓住新云领域的失误，声称到2030年底80%的新云公司会破产。

One big example was Gartner, who latched onto stumbles in the neocloud sector to claim that 80% of neoclouds will go bankrupt by the end of 2030.

他们的报告认为，到那时，3.5万亿美元的AI基础设施支出会因为需求不足而面临减值。

Their report argued that 3.5 trillion in AI infrastructure spending would be subject to an impairment by then due to a lack of demand.

Gartner的首席预测师John Lovelock周一告诉CNBC：“目前，我们的计算能力超过了需求，而且很多代币和处理能力都在免费增值方案中被赠送出去。”

Gartner's chief forecaster, John Lovelock, told CNBC on Monday, "At the moment, we have more compute capacity than we have demand, and we have a lot of the tokens and processing power being given away in freemium solutions.

AI竞赛中的每个人都在争夺最后幸存者的位置。

Everyone in the AI race is running for a last man standing position.

Gartner追踪的AI竞赛中，几乎每个市场都可能收缩到只剩两家主导企业。”

Pretty much every market that Gartner is tracking in our AI race is likely to collapse down to a duopoly."

为了量化一下，Gartner估计到2030年，这些数据中心的拥有者需要产生8.5万亿美元的收入，才能覆盖他们已公布的资本支出、融资成本以及投资资本回报。

To put some numbers around this, Gartner estimated that through 2030, the owners of these data centers will need to generate 8.5 trillion in revenue to cover what they have announced around capital spending, financing costs, and a return on invested capital.

不过，他们又说，从另一方面来看，他们预计企业和消费者的AI总收入到那时也不会超过3.6万亿美元。

On the other side of the equation, they argued, though, they believe that total enterprise and consumer AI revenue is forecasted to not reach more than 3.6 trillion.

如果Gartner说得对，那这个差距可不小。

Quite a big gap if Gartner are right.

与此同时，这引发了一整周关于AI泡沫担忧的疯狂新闻热潮。

That, meanwhile, kicked off a frenetic week-long news cycle around AI bubble fears.

Citadel的首席执行官肯·格里芬也是众多发表看法的人之一。

Citadel CEO Ken Griffin was one of many offering his opinion.

在周二的一场华尔街会议上，他宣称我们“几乎肯定处于泡沫之中”。

At a Wall Street conference on Tuesday, he proclaimed that we are, quote, "almost certainly in the middle of a bubble."

他的论点主要围绕投资回报的不确定性，这让人们对AI支出重新思考。

His argument hinged on uncertain ROI, causing a rethink on AI spending.

格里芬评论说：“美国企业界在从AI模型中获取价值的能力上参差不齐。”

Griffin commented, "Corporate America has been very uneven in its ability to extract value from its AI models.

我觉得所有美国企业都在看着他们付给Anthropic和OpenAI的账单，心里想：‘好吧，我们得想想怎么确保从中获得生产力。’”

I think all of corporate America is looking at their bills from Anthropic and OpenAI and going, 'Okay, we're going to need to think about how we're sure we're getting productivity out of this.'"

但这并不是说泡沫论已经成了共识。

Which is not to say that bubble calls have been consensus.

在另一边，私募股权公司黑石的领导们这周可是频频发声。

On the other side of the argument, leaders from private equity firm Blackstone were out in force this week.

当然，这里得打个折扣，因为黑石在新云领域有大量的股权和债务投资，包括很早就参与了CoreWeave和Firmless。

Now, the grain of salt here is, of course, that Blackstone has massive equity and debt positions in the neocloud sector, including extremely early involvement in CoreWeave and Firmless.

但上周，资深高管Jas Kaira表示，他并不担心数据中心的债务问题。

But last week, veteran executive Jas Kaira said that he isn't concerned about data center debt.

因为超大规模云服务商是租户，Kaira声称这些债务最终由世界上最信誉良好的公司来背书。

Because hyperscalers are the tenants, Kaira claimed that this debt is ultimately backed by the most creditworthy companies in the world.

这是他的原话。

His phrase.

在周四接受彭博社采访时，黑石总裁Jon Gray直接回应了Ken Griffin关于泡沫的评论。

In a Bloomberg interview on Thursday, Blackstone president Jon Gray spoke directly to Ken Griffin's bubble comments.

Gray说：“说到泡沫，你想到的是先建好产能，然后寄希望于需求会来。”

Gray said, "When you think about a bubble, you think about building capacity and then hoping the demand comes.

对于AI，你是在建设产能，同时希望自己能跟上需求。

With AI, you're building capacity and you're hoping you can keep up with demand."

现在，这里需要保持怀疑的是，你看到的每个分析市场的专家都有自己的利益牵扯。

Now, the thing to be skeptical here is that every talking head you see analyzing markets has a dog in the fight.

我提到过黑石在CoreWeave和Firmless这样的公司有持仓，但反过来，肯·格里芬在市场波动时获利最多。

I mentioned that Blackstone has positions in companies like CoreWeave and Firmless, but on the flip side, Ken Griffin profits most when there's volatility in the market.

但这场讨论中有一条主线，你也会看到，我认为这最终是唯一真正重要的东西，那就是需求问题。

But there is a through line to this discourse that, as you'll see, I think is pretty much the only thing ultimately that will matter, which is the question of demand.

在很多方面，其他一切都只是围绕一个大问题的代理之争：AI需求会有多大，以及这些需求的价值最终是否高于在基础设施上的投入。

In many ways, everything else is just a proxy fight for the big question of how much AI demand is there going to be and whether the value of that demand is ultimately higher than the amount being spent on the infrastructure.

现在，当我们深入探讨过去几周影响市场的重大事件，并试图从无关紧要的信号中筛选出重要的信号时，我认为有必要在术语上稍微精确一些。

Now, as we dig deeper into some of the big events that have shaped markets over the past couple of weeks and try to sort through the signals that matter from those that don't, I think it is worth being somewhat precise about our terms.

当我们使用像“AI泡沫”这样的大词时，会失去细微差别。谈论那种让股市下跌20%或30%的熊市的人，和谈论像互联网泡沫破裂那样、纳斯达克两年内蒸发超过75%的世代级泡沫破裂的人，说的其实是完全不同的东西。

When we use big apocalyptic terms like AI bubble, we lose nuance, and people who are talking about a bear market in stocks of the kind that takes stock markets down 20 or 30% are talking about something quite different than a generational bubble burst, like the dot-com crash, which saw over 75% wiped off the NASDAQ across two years.

换句话说，股市崩盘和信贷或金融危机不是一回事。

Stock market crashes, in other words, are not the same as a credit or financial crisis.

这两者可以单独发生。

You can have one without the other.

而且在这两者中，股市崩盘的破坏性要小得多。

And of the two, a stock market crash is far less damaging.

虽然股市崩盘非常痛苦，有时对投资市场的人来说尤其痛苦，特别是那些依赖投资收入生活的人，比如靠固定收益退休账户生活的人，但如果银行和整个金融体系没有系统性问题，市场是可以自然恢复的。

While absolutely painful, sometimes extremely so for those invested in the markets, and especially those who rely on their investments in markets like those living off fixed income retirement accounts, if there's not a systemic problem for banks in the broader financial system, markets can recover naturally.

而信贷危机则完全不同，大规模违约和连锁反应会在整个经济中造成损害，这种问题可能需要央行介入。

A credit crisis, on the other hand, where mass defaults and contagion create impairments throughout the economy, is quite a different problem, and one that can require central bank intervention.

如果说这个月AI交易中哪个部分最让人担忧，那就是新云服务商。

Now, if there's one segment of the AI trade that's been the locus of concern this month, it's the Neoclouds.

Firmus在澳大利亚的IPO被取消了，因为投资银行找不到愿意以他们期望价格购买的足够买家。

We had the Firmus IPO in Australia getting canceled after investment banks couldn't find enough buyers at their desired price.

然后CoreWeave的一部分债务跌到了不良水平，这意味着违约的可能性相当大。

Then a portion of CoreWeave debt traded down to distressed levels, implying a reasonable chance of default.

微软随后因为电力延迟退出了英国的一个数据中心计划，让nScale陷入了困境。

Microsoft then pulled out of a planned data center in the UK because of power delays, leaving nScale high and dry.

而且最近，我们开始看到AI领域的大玩家们出手相救。

And more recently, we're starting to see the big players in AI ride to the rescue.

有传言说Anthropic正在洽谈与nScale签订一份450亿美元的新合同。

Anthropic is rumored to be in talks for a new $45 billion contract with nScale.

高盛已经把CoreWeave的股票评级上调为买入，说市场对他们债务的担忧有些过头了，而且股票已经被超卖了。

Goldman Sachs has upgraded CoreWeave stock to a buy, saying that some concerns over their debt are overblown and the stock has become oversold.

还有传言说NVIDIA正在洽谈向Firmus投资数亿美元，帮助他们在IPO失败后获得资金支持。

And NVIDIA is rumored to be in talks to invest hundreds of millions into Firmus to help finance them after that failed IPO.

当然，要说对Neocloud的担忧完全是小题大做，那也太过分了。

It would certainly be going too far to say that the Neocloud jitters are overblown.

这绝对是市场上值得关注的一个领域，投资者们都在密切关注。

It is absolutely an area of the market to watch, and investors are keeping a very close eye on it.

但同样重要的是，我们要正确看待这些数字。

But it's also important to keep the numbers we're talking about in perspective.

nScale的目标是进行300亿美元的IPO。

nScale is targeting a $30 billion IPO.

在这个行业里，主要公司加起来价值大约10万亿美元，一个300亿美元的公司相对而言就是小角色。

In an industry where the key firms are worth around $10 trillion as a collective, a $30 billion company is a relative minnow.

从纯粹的经济角度看，整个Neocloud行业就算全归零，可能影响也不大，当然，除了叙事上的影响，那会是巨大的。

From strictly an economic perspective, it's likely the entire Neocloud sector could go to zero without it having much of an effect, other than, of course, the narrative effect, which would be gigantic.

这就是为什么大玩家们会联合起来给这些小公司兜底。

And that's why the big players are coming together to backstop the small guys.

根据信息，NVIDIA已经通过信贷担保和数据中心租赁兜底，为Neoclouds提供了1700亿美元的支持。

According to the information, NVIDIA has already provided $170 billion in support for the Neoclouds through credit guarantees and data center rental backstops.

前沿实验室和超大规模云服务商也表明，如果谁有点不稳，他们愿意把数十亿美元的合同摆到台面上来。

The frontier labs and hyperscalers have also shown they're willing to put multi-billion dollar contracts in the headlines if anyone gets too shaky.

整个行业都极其希望这些公司能存活下来，数据中心能建起来，因为这关系到他们的切身利益。

The entire industry has a huge vested interest in making sure these companies survive and their data centers get built.

那么，最值得警惕的担忧会是什么呢？

So what, then, would be the most worrying concerns to watch for?

如果一家Neocloud陷入困境，却没人出手相救，那才是更令人担忧的信号。

If a Neocloud gets into trouble and nobody rides to the rescue, that would be a more concerning signal.

如果困境从CoreWeave自己的债券蔓延到租给Meta、微软或谷歌的数据中心项目债务，那可能表明市场怀疑的是租户，而不仅仅是Neocloud。

If distress started to spread from CoreWeave's own bonds to project debt for data centers leased to Meta, Microsoft or Google, that might suggest that the markets doubt the tenant, not just the Neocloud.

最后，如果某个超大规模云服务商或实验室因为不需要而放弃容量，那将是巨大的信号。

Finally, if a hyperscaler or lab walks away from capacity because it doesn't need it, that would be huge.

记住，微软退出英国被明确归咎于电力延迟。

Remember, Microsoft's UK exit was blamed very specifically on power delays.

如果取消是因为需求不足，那完全是另一回事。

A cancellation blamed on demand would be a totally different story.

这周某些批评者特别关注的一个故事是，亚马逊有一大堆GPU在仓库里积灰。

One story that's getting a ton of attention from certain critics this week is that Amazon has a huge pile of GPUs gathering dust in a warehouse.

周二，Business Insider写道：“在新的早期机架接收流程下，AWS预计在最终部署前，将在美国设施接收约50亿美元的NVIDIA GB300服务器机架。”

On Tuesday, Business Insider wrote, "Under a new early rack receipt process, AWS expects to receive roughly $5 billion in NVIDIA's GB300 server racks at US facilities before final deployment.

公司预计在可预见的未来，每月还会收到约30亿美元的GB300机架。”

The company projected roughly 3 billion more in GB300 racks each month for the foreseeable future."

GB300是当前最高端的AI芯片，是Blackwell系列的最新款。

GB300s are the current top-of-the-range AI chips, the latest in the Blackwell series.

有消息人士说，亚马逊有地方放更多的机架，但没有电力来启动它们。

Sources said that Amazon has room for more racks, but they don't have the power to turn them on.

根据Business Insider看到的邮件，AWS至少有一个团队没有，用他们的话说，一条“清晰的预期路径”来安装这些新机架。

According to emails viewed by Business Insider, at least one team at AWS did not have a, quote, "clear forecasted path" to getting the new racks installed.

亚马逊发言人表示，在建设之前预先存放硬件是常规操作，而且一旦电力到位就能立即接入GPU，这是最佳商业实践。

An Amazon spokesperson said it's routine to stage hardware ahead of construction, and that being prepared to plug in GPUs once the power is available is best business practice.

文章还指出，亚马逊正试图在竞争中抢先锁定稀缺的GPU。

The article also noted that Amazon is attempting to secure scarce GPUs ahead of the competition.

这周批评者的观点是，NVIDIA似乎在强迫超大规模云服务商接收GPU，即使他们没有电力来使用，因为担心未来芯片的分配会减少。

The critics take here this week has been the idea that NVIDIA is somehow forcing hyperscalers to take delivery of GPUs, even if they don't have the power to use them for fear of losing allocation to future chips.

有人直接声称NVIDIA在渠道压货，或者强迫销售以更快地确认收入。

There have been direct claims that NVIDIA is channel stuffing or forcing sales to realize revenue more quickly.

基本上就是说，强势而有权势的黄仁勋把所有的超大规模云服务商都牢牢掌控在手里。

Basically, that the ruthless and powerful Jensen Huang has all the hyperscalers under his thumb.

按照这些批评者的逻辑，这意味着NVIDIA的一部分收入是虚假的，对GPU的需求，也就是对AI的需求，比表面看起来要疲软，整个泡沫即将破裂。

As the logic among those critics goes, this means that some amount of NVIDIA's revenue is fake, demand for GPUs, and therefore AI, is softer than it appears, and the whole bubble is getting ready to fall apart.

我很难得出同样的结论。

I have a hard time coming to the same conclusion.

没有理由怀疑商业内幕的报道。

There's no reason to doubt Business Insider's reporting.

他们在AWS的消息来源非常可靠，这显然是一篇历时数月的调查报道，目的是弄清楚实际安装了多少GPU。

They're very well sourced at AWS, and this has clearly been a months-long investigative piece to find out how many GPUs are actually being installed.

但奥卡姆剃刀式的解释并不是AI需求不足。

But the Occam's razor explanation here is not a lack of AI demand.

而是缺乏通电的数据中心。

It's that there's a lack of powered up data centers.

即使黄仁勋是个冷酷无情的商人，如果长期计划只是无限期地囤积GPU，亚马逊也不会订购数百亿的GPU。

Amazon, even if Jensen was a ruthless businessman, would not be ordering tens of billions of GPUs if the long-term plan is just to warehouse them indefinitely.

更合理的解释是，亚马逊确实担心未来芯片的供应。

It would make much more sense that Amazon is genuinely concerned about future access to chips.

目前，GPU仍然以极高的溢价出售，旧硬件的售价甚至超过了原价。

Currently, GPUs are still being priced at a gigantic premium, with old hardware selling above its original price.

与此同时，旧硬件仍在积极使用中，所以把GPU存放几个月也没什么大碍。

Meanwhile, old hardware is still in active use, reducing the harm in storing GPUs for a few months.

说实话，今天我们聊的很多事情里，这个故事感觉像是在抓救命稻草，拼命找任何迹象来证明AI基础设施的建设正在崩溃。

Frankly, among a lot of things that we'll talk about today, this is one story that feels like gasping at straws, looking for any possible sign that the AI infrastructure build-out is falling apart.

它有点过于依赖与互联网泡沫时代暗光纤的对比，但那是已安装基础设施需求不足的问题。

It's leaning a bit too hard on the comparison to dark fiber during the dot-com era, but that was an issue of too little demand for already installed infrastructure.

而现在是基础设施不足，无法满足需求。

This is too little infrastructure to service demand.

建设数据中心和接入电力确实面临很多挑战，也带来不少影响和问题。

There are plenty of implications and issues with the challenges of getting data centers built and power connected.

但这些问题最终并不能说明GPU或整个AI的需求有什么问题。

But those issues don't have anything to say ultimately about the demand for GPUs or AI in general.

再说一次，按照最简单的解释，也就是奥卡姆剃刀原则，如果AWS不得不存放GPU，那更可能是一个积极信号，说明需求极高，而不是什么为了抬高NVIDIA股价的阴谋。

Again, just going by the simplest available explanation, Occam's razor, if AWS has to warehouse GPUs, it seems more likely to be a positive sign that demand is sky high than a nefarious plot to boost NVIDIA earnings.

而且说实话，即使真是那样，也只能说明NVIDIA依然拥有全球最强的定价权。

And frankly, even if that is the case, it just demonstrates that NVIDIA still has all the pricing power in the world.

不过，大家心里都在悄悄琢磨一个问题：这些数据中心债务到底都堆在哪儿了。

Quietly bubbling away in the back of everyone's mind, though, is the question of where all the data center debt is sitting.

这就回到了股市崩盘和系统性风险之间的区别。特别是经历过2008年那次危机之后，每次市场一有风吹草动，大家自然会问，这种风险会不会传染，怎么传染。

This gets back to that idea of the difference between a stock market crash and a more systemic issue Especially after the experience that people lived through in 2008, every time there's any sort of market wobbles, the natural question is to ask if and how contagion can flow.

在这种情况下，如果AI泡沫破了，会不会像金融危机时那样，把养老金体系或者大型保险公司给炸了？

In this case, if the AI bubble bursts, is it going to blow up the pension system or major insurers like we saw during the GFC?

这个问题很难给出确切的答案，而且在这个领域，我们确实应该保持谦逊。

That is a notoriously difficult question to answer definitively, and this is an area where we should absolutely tread with humility.

我们确定的是，大量数据中心债务已经被证券化，卖给了保险公司和养老基金。

What we know for sure is that a significant amount of data center debt has been collateralized and sold to insurers and pension funds.

据《The Information》报道，CoreWeave上次发行债券时，一批大型保险公司购买了超过7.4亿美元的债务。

The Information reports that a collection of giant insurance firms bought more than $740 million worth of CoreWeave debt the last time they sold bonds.

他们提到了美国通用人寿保险公司、富达、担保人寿和丹佛安全人寿，这些公司的投资组合都由黑石集团管理。

They listed American General Life Insurance, Fidelity, and Guaranty Life and Security Life of Denver, all of which have portfolios managed by Blackstone.

从表面上看，这个数字挺大的，但实际上，7.4亿美元分散在这些大型投资组合里，也就是个零头。

In the abstract, that seems like a big number, but realistically, $740 million spread across these large portfolios is a rounding error.

他们持有数据中心债务这件事本身还是有意义的，如果这些债务违约，确实会留下一个窟窿，但到目前为止，看起来还不像2008年金融危机后那种大规模资产减值的风险。

It is still meaningful that they're carrying data center debt, and it would leave a hole if that debt defaults, but it certainly so far doesn't look like the potential of a mass impairment we saw following the GFC.

毕马威和德克萨斯大学奥斯汀分校的一项新研究发现，当人们使用AI工作时，相似的技能并不保证相似的结果。

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

研究人员研究了500多名早期职业人士，发现表现最好的人通过引导、评估和优化AI的输出，持续放大了AI的价值。

Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

这些顶尖表现者被称为“AI放大器”，他们的优势不在于知识本身，而在于他们如何与AI协作。

These top performers, called AI Amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

想了解更多关于AI放大器与其他人的区别，请访问kpmg.com/us/aiamplifiers。

Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers.

到了这个阶段，问题已经不再是公司是否在积极使用AI了。

At this point, it's no longer a question of whether companies are actively using AI.

但能不能用好AI，那就是另一回事了。

Using it well, on the other hand, is a whole different story.

不过，Robots & Pencils这家公司，我觉得就是为这个时代而生的。

Robots & Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用AI工程公司，直接与客户合作解决对业务真正重要的问题，而不是那些只存在于幻灯片里的实验。

They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每次合作都从客户真正需要的结果出发，倒推整个方案。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域里分辨真正的AI工程和那些虚头巴脑的东西，这就是关键所在。

If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker.

去robotsandpencils.com看看吧。

Head to robotsandpencils.com.

本期节目由CData赞助播出。

This episode is brought to you by CData.

推出AI的公司们一直在撞同一堵墙。

Companies rolling out AI keep hitting the same wall.

代币花费不断攀升，安全问题拖住了AI代理的后腿，而且没人知道AI到底是怎么被使用的。

Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used.

根本原因在于，AI与业务所依赖的系统和上下文是脱节的。

The root cause, AI is cut off from the systems and context your business runs on.

所以CData打造了Connect AI Gateway，一个连接你的AI和系统的统一控制点。

That's why CData built Connect AI Gateway, one control point between your AI and your systems.

它基于支持Palantir、Google、Salesforce等数百家公司的连接层构建，每个智能体都能获得它所需的确切数据和工具。

Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs.

它能精细到记录级别的治理，让你终于能在同一个地方看到你组织的完整AI使用情况。

It governs down to the record and lets you finally see your org's complete AI usage in one place.

准确的答案，更低的token成本，以及一个每次交互都在学习你业务的网关。

Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction.

详情请访问cdata.ai。

See how at cdata.ai.

Blitzy对大型代码库的理解，能够实现自主安全修复、现代化改造和新功能开发。

Blitzy's understanding of massive code bases unlocks autonomous security fixes, modernization, and new features.

那么，当完全没有遗留代码时，会发生什么呢？

So what happens when there's no legacy code at all?

绿地项目本应是容易的部分。

Greenfield is supposed to be the easy part.

全新开始，没有技术债务。

Clean slate, no technical debt.

但即便是全新项目，也只能按人的速度推进，一个冲刺接一个冲刺。

But even Greenfield moves at human speed, one sprint at a time.

Blitzy 把工作单元从开发者变成了整个项目，自主规划、构建、测试并验证整个应用，从零开始。

Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.

数十万行可直接投入生产的代码。

Hundreds of thousands of lines of production-ready code.

有一位 Blitzy 客户搭建了一个全新应用，共 53.4 万行代码，把原本 65 周的路线图压缩到了两周。

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into two weeks.

另一位客户在没有前端工程师的情况下，交付了整个应用。

Another shipped an entire application with no front-end engineer.

无论是遗留系统还是全新项目，答案都一样。

Legacy or greenfield, the answer is the same.

软件以计算的速度运行。

Software at the speed of compute.

在blitzy.com上打造未来。

Build what's next at blitzy.com.

网址是B-L-I-T-Z-Y.com。

That's B-L-I-T-Z-Y.com.

另一个需要注意的风险点是保险公司和私募股权公司日益紧密的合作。

Another pocket of risk to be aware of is the growing pairing of insurance firms with private equity.

最典型的例子就是Apollo Capital，它在2022年与一家名为Athene的保险公司合并了。

The canonical example is Apollo Capital, which merged with an insurance firm called Athene in 2022.

这是沃伦·巴菲特开创的一种经典融资策略。

This is a classic financing strategy pioneered by Warren Buffett.

保险公司能享受税收优惠，投资公司则能获得稳定的资金来源。

The insurance company gets favorable tax treatment, and the investment firm gets a steady flow of capital.

这是金融领域一个复杂的角落，肯定超出了我们节目的讨论范围。

This is a complex corner of finance that is certainly beyond the scope of this show.

但有一点很重要，这些并不是像State Farm那样提供车险和家财险的大型保险公司。

But one important thing to understand is that these aren't major insurers like State Farm providing auto and home insurance.

相反，它们是离岸保险公司，主要销售人寿保险和年金。

Instead, they're offshore insurers that mostly sell life insurance and annuities.

这确实是个风险，在AI信用违约时它们可能会受损，但主要风险还是落在投资年金的富人身上。

It's a real risk, and they could definitely become impaired during an AI credit default, but they're mostly a risk for the wealthy people who invest in annuities.

最后是养老基金，它们肯定对AI行业有敞口。

Finally, there's the pension funds, which definitely have exposure to the AI sector.

怎么可能没有呢？

How could they not?

它们的工作就是投资股票和债券，而过去四年里，大部分回报都来自AI。

Their job is to invest in stocks and bonds, and for the past four years, the bulk of the returns have come from AI.

据《华尔街日报》报道，自2022年以来，AI为13个公共养老基金带来了1380亿美元的收益。

The Wall Street Journal reports that AI drove 138 billion in gains across 13 public pension funds since 2022.

上周，美国金融改革组织的一份报告指出：“投机性的数据中心债务正被重新打包，出售给养老基金、退休基金、大学捐赠基金和其他机构投资者，这种情形令人不安地联想到次贷泡沫。”

Last week, a report from Americans for Financial Reform stated, "Speculative data center debt is being repackaged and sold to pension and retirement funds, university endowments, and other institutional investors in a manner eerily reminiscent of the subprime bubble.

私募股权公司收购了人寿保险公司，并让它们背负高风险债务，使投保人和公众面临损失，并增加了金融传染的威胁。”

Private equity firms have purchased life insurers and loaded them up with risky debt, exposing policyholders and the public to losses and raising the threat of financial contagion."

那份报告主要是想反对AI救助计划，很多人都确信特朗普政府会一有机会就出手救助。

That report was largely interested in arguing against an AI bailout, which many are certain the Trump administration will indulge in at the earliest opportunity.

但关键是要理解它和次贷泡沫的相似之处，以及不同之处。

But understanding not just how it is similar to the subprime bubble, but how it is different, is essential here.

大家都记得《大空头》里那副Jenga塔，就是次贷危机时把房贷切成小块，打包成CLO卖给养老基金。

Everyone remembers the slices of mortgages chopped up and sold to pension funds as CLOs during subprime, depicted as a Jenga stack in "The Big Short."

但关于为什么这实际上是个问题，讨论得不多。

There is less discussion about why that was actually a problem.

问题不只是把债务切碎再组合成CLO。

It wasn't just about chopping up the debt and combining it into a CLO.

问题在于这些CLO被评为投资级债务。

The problem was that those CLOs were rated investment-grade debt.

它们被当作和国债一样安全。

They were treated as being just as safe as Treasury bonds.

这次，数据中心债务被当作公司债券处理，有时是投资级，有时不是，但绝对远不能和国债相提并论。

This time around, data center debt is being treated as corporate bonds, sometimes investment grade, sometimes not, and definitely nowhere near the same as Treasury bonds.

它们和成千上万的其他债券一样，都躺在养老金和保险公司的资产负债表上。

They're sitting on pension and insurer balance sheets just the same as thousands of other bonds.

在次贷危机期间，拆分债务明确且有意地被用来掩盖其风险程度。

During subprime, chopping up the debt was explicitly and intentionally used to obfuscate how risky it was.

而这一次，没有人假装数据中心债务像房子一样安全。

This time around, no one is pretending that data center debt is as safe as houses.

但再说一次，我不想只关注为什么现在可能值得稍作停顿和感到压力，还想关注未来可能更令人担忧的事情。一个令人担忧的点是，如果我们看到披露的风险敞口数量级飙升，或者集中在少数几家私募股权所有的保险公司手中，那么740亿美元分散在大型投资组合中其实很单薄。

But again, I want to not just focus on why things are maybe worth a little pause and the stress right now, but what could in the future be more cause for concern One cause for concern would be that if we saw a disclosed exposure jump by orders of magnitude or concentrating in a handful of PE-owned insurers, 740 million spread across big portfolios is thin.

但如果这个数字飙升到数百亿美元，集中在两三家保险公司手里，那情况就完全不同了。

But if that jumps to tens of billions sitting in two or three insurers, that could be a different story.

另一个担忧是，如果数据中心债务开始被重新打包成评级高于其底层贷款的证券。

Another concern would be if data center debt started to get repackaged into securities that were rated higher than the loans underneath them.

再说一次，重点不是说这些事情不可能发生，而是它们目前还没有发生。

And again, the point here is not that these things are impossible, it's that they're not the things that are happening yet.

现在，回到救助这个话题，最近另一个热议的话题是，大型AI公司正试图让自己变得“大到不能倒”。

Now, going back to the idea of a bailout, another big topic of discussion of late is the idea that the large AI players are looking to make themselves too big to fail.

这个话题在Anthropic和埃隆·马斯克宣布将通过两种截然不同的机制向公众赠送股权之后，又变得异常激烈。

This discussion came back with a vengeance after Anthropic and then Elon Musk said that they would gift equity to the public through two very different mechanisms.

现在，关于这在民主社会中是否合适、是否是重新分配收益的好方法，或者是否是一个明显的信号表明亿万富翁和大型科技公司需要缴纳更多税款，有一个非常重要的政策讨论。

Now, there is a whole very important policy discussion about whether this is actually appropriate in a democratic society, whether it's a good way to redistribute gains, or possibly a glaring sign that billionaires and mega cap companies need to be paying more in taxes.

但这不是我们今天要讨论的重点。

But that's not the particular can of worms that we're opening today.

显然，有一个有力的论点认为，让公众直接持有公司股份可能会使这些公司变得大到不能倒。

There is clearly a solid argument that giving the public a direct stake in the companies risks making them too big to fail.

需要明确的是，当我们说“大到不能倒”时，我们使用的是过去15年中的一个特定术语，指的是政府觉得有义务介入保护这些机构，无论它们在自由市场中原本会面临什么。

And to be clear, when we say too big to fail, we're using a specific term from the last 15 years that refers to institutions that the government feels obligated to come in and protect, regardless of what would otherwise happen to them in the free markets.

在这个问题上，我可能比你们中的一些人更悲观，如果是这样，请原谅我。

This is an area where I am perhaps more cynical than some of you, and forgive me if so.

实际上，我认为如果你担心的是我们会再出现一批大到不能倒的公司，在我看来，不管他们是否向公众赠送股份，这个局面已经形成了。

Realistically, I think if your concern is that we're going to get another group of too big to fail companies, in my opinion, that ship has already sailed, whether or not they give shares to the public.

在美政府与市场关系的历史现阶段，在我看来，政府极不可能坐视一家市值万亿美元的公司在其经济中炸出一个大洞。

At this stage in the history of the relationship between the US government and markets, it seems massively unlikely to me that the government would just stand around while a trillion-dollar company blew a giant hole in their economy.

2008年政府没有让通用汽车倒闭，而且从那以后，政府跟市场打交道的方式肯定变得奇怪多了。

The government didn't let it happen with GM in 2008, and certainly the norms of how the government engages with markets have gotten a lot weirder since then.

另一方面，至于说这些公司是在市场崩盘前策划救助，这个说法我实在不信。

On the other hand, as for the argument that these companies are engineering a bailout ahead of a market crash, that's something I just don't buy.

不管别人怎么说Anthropic，现在很明显他们对自己的使命有真正的信念，最简单的解释还是他们想建立一个巨大的公益基金来推进这个使命。

Regardless of what anyone wants to say about Anthropic, it's pretty clear at this point that they have a genuine belief in their mission, and the simplest explanation, once again, is that they want to establish a giant public benefit fund to further that mission.

至于埃隆·马斯克，如果他真的相信这种捐赠是影响公共利益的最佳方式，那也完全符合他一贯的做法。

As for Elon Musk, you have the convergence of the fact that it would be entirely consistent if he did believe that this sort of giving is the best way to impact public benefit.

但更 cynical 一点看，他可能是在观察形势，觉得某种再分配迟早会发生，他想抢在前面，这样也许能掌控局面，而不是被动接受。

But also there's the fact that more cynically, he might be reading the room that some sort of redistribution is going to happen, and he thinks if he gets out ahead of it, he might be able to control it rather than be subject to it.

还有一种更业余的心理分析，就是这个人喜欢被爱，而没有什么比给美国公众110亿美元更容易赢得爱了。

And there's also the more armchair psychologist assessment that the man loves being loved, and there's no easier way to being loved than to give $11 billion to the American public.

这个目前我觉得不值得花太多时间讨论，但为了完整起见，值得留意的是，如果你开始看到AI公司在危机出现之前就要求联邦明确担保、贷款计划或融资支持，那可能会引发另一场对话。

This is one that at the moment I don't think is worth spending a ton of additional time on, but for the sake of completeness in terms of what to watch for, if you start to see AI companies begin to ask for explicit federal guarantees or loan programs or financing backstops before there's a crisis, that could cause, let's say, a different conversation here.

在所有人们提到的警告信号中，我觉得最有趣、最值得深思的是收入质量的问题。

I think of all the warning signs that people are bringing up, the one that is most interesting and deserving of deep consideration is the question of revenue quality.

不过就算在这个话题里，也有好几种不同的情况，我觉得每种情况背后的逻辑都不一样。

Although even within this, there's a bunch of different versions of it that I think have very different dynamics.

一方面，英伟达报告说，第二季度营收里，光两个客户就占了39%。

On the one hand, you have NVIDIA, who reported that just two customers made up 39% of their Q2 revenue.

另一方面，《华尔街日报》最近泄露了Anthropic的IPO招股书，里面提到上半年营收的五分之一来自前两大客户。

On the other is the Wall Street Journal's recent leak of Anthropic's IPO prospectus, which noted that a fifth of Anthropic's revenue in the first half came from their top two customers.

还有，Ramp的数据显示，在OpenAI和Anthropic通过Ramp平台的消费总额里，前1%的企业占了80%。

Relatedly, Ramp data has shown that the top 1% of businesses account for 80% of spending volume for OpenAI and Anthropic through Ramp's platform.

那这些情况有什么不同呢？

So how are these different?

英伟达的客户集中度高，是因为全世界能建几十亿美元数据中心的企业就那么几家。

Well, NVIDIA has a highly concentrated customer base because there's only a handful of companies in the world that are building multi-billion dollar data centers.

这个逻辑可能也部分适用于Anthropic的情况。

And some of this might even apply to the Anthropic story.

他们很大一部分收入很可能来自那些超大规模云服务商，因为这种关系更像是批发商和分销商，而不是普通的供应商和客户。

A large chunk of their revenue is likely coming from the hyperscalers because that relationship is that of a wholesaler and a distributor rather than a vendor and customer.

当你的最大客户之一有很大风险可能撤资时，收入质量就成了问题。

Revenue quality becomes an issue when there's a big risk that one of your largest customers could pull the plug.

对于转售Claude代币的超大规模云服务商来说，短期内这种情况似乎不太可能发生。

For hyperscalers reselling Claude tokens, that seems very unlikely to happen in the short term.

我觉得更大的担忧是，少数重度用户占据了企业AI使用的大部分，这种分布太集中了。

A much bigger concern, I think, is this narrow distribution of power users making up the bulk of enterprise AI use.

和往常一样，我们需要结合Ramp的数据背景来看，他们服务的只是市场的一小部分，都是非常前瞻性的公司，技术非常先进，很多来自科技行业。

As always, we need to contextualize Ramp data, recognizing that they service a very, very slim part of the market, extremely forward-looking companies, extremely tech-advanced companies, many of them from the tech sector.

但值得注意的是，Ramp发现，AI客户中前1%的80%是风投支持的初创公司。

And yet it's still worth noting that Ramp found that 80% of the top 1% of AI customers are VC-backed startups.

其中一部分交易量是把前沿代币包装成服务，并提供风投补贴的折扣，这当然是一些智能体初创公司的商业模式。

Some amount of that volume is wrapping frontier tokens in a service and providing a VC-subsidized discount, which is, of course, the business model for some of the agent startups.

我们也可以推测，很多代币交易量是用于内部使用，这些也是由风投资金资助的。

And we can also presume that a lot of the token volume is for internal use, which is also funded by VC dollars.

如果风投资金枯竭，这些初创公司资金耗尽，问题就来了。

Where this becomes a problem is if VC money dries up and these startups hit the end of their runway.

目前，AI风投市场依然非常火爆，但这种情况随时可能改变。

At the moment, AI venture is still a raging bull market, but that can always change.

更大的风险，也是真正关键的问题，是AI支出中有多少会长期遵循幂律分布，而不会扩散开来？

The bigger risk, and really the big question here, is how much of AI spending is permanently going to be subject to power laws without broadening out?

换句话说，这种高度集中是否可能持续，即使整个AI行业的需求结构在成熟？

In other words, is this incredible concentration likely to hold as the demand profile for the entire AI industry, even as buyers mature?

还是说，这仅仅反映了处于不同采用阶段的企业之间的差异？

Or is it simply an indicator of the difference between companies at different stages in their adoption cycle?

如果现在初创公司每月每位员工花费约9000美元，处于图表顶端，那么十年后，是否所有企业都会达到每月每位员工9000美元的水平？

If the startup spending around $9,000 per employee per month are the top of the chart now, at some point in a decade, are all enterprises spending $9,000 per employee per month?

如果不是，它们又能接近这个数字多少呢？

And if not, how close to those sort of numbers are they going to get?

Ramp观察到，自夏季以来，所有群体的企业每位员工的支出都翻了三倍。

Ramp has seen businesses across all cohorts triple their spend per employee since the summer.

不在前1%但在前10%的企业，每位员工的支出从约200美元增长到约800美元，增长速度超过了中位数。

Businesses not in the 1% but in the 10% went from around spending $200 per employee to around $800 per employee, growing faster than the median.

这说明采用曲线正在向整个经济领域扩展，但当然也不能保证一定会这样。

That suggests an adoption curve that is expanding out across the economy, but there's certainly no guarantee there.

另一个变数，也是我们接下来要讨论的最后一个话题，就是当大型实验室上市后，这个风险会变得更加具体。

Another curveball, which gets into our next and final line of discussion, is that this particular risk gets far more tangible after the major labs go public.

目前，这个风险是基于Ramp的数据、经常被误解的硅谷数据代币指数，以及其他各种非官方来源，这些来源只掌握了整个情况的一小部分。

At the moment, this risk is based on data from Ramp, the often misunderstood Silicon Data Token Index, and various other unofficial sources that have a tiny sliver of the full picture.

一旦Anthropic和OpenAI上市，它们就无处可藏了。

Once Anthropic and OpenAI go public, they'll have nowhere to hide.

它们经过全面审计的财务数据将按季度公布，它们的CEO将在财报电话会议上接受股票研究员的质询。

Their fully audited financials will be reported quarterly, and their CEOs will be grilled by equity researchers on earnings calls.

如果它们没有增长，没有改善收入质量，所有人都会知道。

If they're not growing and improving their revenue quality, everyone is going to know about it.

当然，很明显，Anthropic即将进行的IPO是AI泡沫重新成为讨论焦点的一个重要原因。

And of course, clearly, the rapidly approaching Anthropic IPO is a big part of the reason that the AI bubble is back in the discourse.

泡沫市场往往会在某个催化剂打破人们的幻想时见顶，而对许多人来说，Anthropic的IPO似乎很可能就是那个催化剂。

Frothy markets often top on a catalyst that breaks the suspension of disbelief, and for many, the Anthropic IPO seems likely to fit that bill.

到目前为止，我们了解到的Anthropic的财务状况，至少和很多人原本预期的相比，看起来并不那么亮眼。

What we've learned so far about Anthropic's financials have, compared to at least what many people were thinking, looked less than stellar.

Anthropic上半年亏损800亿美元的消息在社交媒体上疯传，尽管这只是会计亏损，实际现金亏损是87亿美元。

The headline that Anthropic made an $80 billion loss in the first half went viral on social media, even though this was an accounting loss and their true cash loss was $8.7 billion.

不过，随着Anthropic进入最后阶段，我们看到估值在走软。

Still, we're seeing valuations soften as Anthropic enters the final stretch.

OpenAI的融资预计很快完成，据《The Information》报道，估值将达到1.25万亿美元。

OpenAI's fundraising is expected to close soon, and "The Information" reports that the valuation will come in at $1.25 trillion.

早前的报道说是1.4万亿美元。

Earlier reporting had it at $1.4 trillion.

我不知道你能不能把早期泄露的数字和最终实际数字之间的差异，明确视为估值下调，而不是单纯的报道错误，但投资者确实是这么看的。

Now, I don't know if you can consider the difference in an early leaked reported number and the actual factual final number as being clearly a valuation downgrade as opposed to just misreporting, but that's certainly how investors are looking at it.

虽然一些Anthropic投资者最初在媒体上推动2万亿美元的估值，但最近有报道称，另一批投资者在推动1.5万亿美元。

While some Anthropic investors were originally pushing for a $2 trillion valuation in the press, more recently, there's been reporting that another group is pushing for $1.5 trillion.

与此同时，IPO市场其他部分完全冻结，本月至今募资不到6亿美元。

Meanwhile, the rest of the IPO market is completely frozen, with less than $600 million raised so far this month.

Truist Financial的股票资本市场主管West Riggs说：“人们真的被吓住了。

West Riggs, the head of equity capital markets at Truist Financial, said, "People are truly paralyzed.

市场必须自己理清头绪。”

The market has to sort itself out."

德勤会计师事务所的美国资本市场交易负责人说：“估值上陷入了僵局，要弥合这个差距，市场才能继续往前走。”

The US capital markets transaction leader at Deloitte & Touche said, "There is a stalemate on valuation, and rationalizing that gap is what needs to happen before the broader market can move forward."

基本上，想上市的公司要价太高，投行给不了那么多。

Essentially, companies looking to IPO are asking for more than investment banks will give them.

这可能意味着新云公司无法上市，也就拿不到下一轮融资，最后只能让私募股权来兜底。

This could mean neo clouds are unable to go public and consequently unable to access their next tranche of funding, leaving private equity holding the bag.

但说到底，就像任何市场或市场心理问题一样，这个问题只能靠接下来发生的事来解决。

But ultimately, as with any market or market psychology issue, this one is only resolved by the next thing happening.

本周的《彭博观点》中，Chris Hughes写了AI交易的一长串问题，但总结说：“目前压在市场心头的一些疑问会随着时间得到解决。

In "Bloomberg Opinion" this week, Chris Hughes wrote about the long list of problems with the AI trade but concluded, "Some of the questions currently weighing on the market will be resolved over time.

但如果Anthropic能赶紧完成IPO，大家就能少操一份心了。”

But if Anthropic could just get on with its IPO, that would give everyone one less thing to worry about."

而且我觉得这里确实存在风险，不是那种系统性的传染风险之类的，而是对AI交易信心的风险。

And I think that there is real risk here, not for some sort of systemic contagion or anything like that, but from the standpoint of confidence in the AI trade.

如果Anthropic以1.5万亿美元的估值定价IPO，而投资者几个月来一直在谈论2万亿美元的数字，那虽然不太理想，但最终可能也不会太让人担心。

If Anthropic prices their IPO at $1.5 trillion after investors splashed around the $2 trillion number for months, that's not great, but ultimately probably not all that concerning.

但另一方面，如果Anthropic以1.5万亿美元定价，而市场反应依然冷淡，那就会证实很多人对AI交易状况的最大担忧。

If, on the other hand, Anthropic prices at $1.5 trillion and we still see a weak reception in the market, that will confirm a lot of people's worst fears about the state of the AI trade.

今天这期节目的重点，既不是要彻底否认市场担忧的理由，也不是要让你大吃一惊，说AI泡沫破裂是不可避免的。

The point of today's episode is neither to definitively try to deny any causes for market concern, nor to surprise you and come out saying that some AI bubble burst is inevitable.

相反，我试图逐一审视这个叙事中的各个子部分，谈谈我认为它们目前的状况，以及未来什么会让我更加担忧。

Instead, what I've tried to do is look at each of the different subcomponents of the narrative in turn and talk about where I think they are today and what would make me more concerned going forward.

不可否认的是，AI泡沫即将破裂的叙事一直是这个市场周期中反复出现的主题，可以追溯到生成式AI时代刚开始的时候。

One thing that's undeniable is that the narrative that the AI bubble is about to burst has been a recurring feature of this market cycle, going back all the way to the beginning of the generative AI era.

不管出于什么原因，它似乎每年这个时候都会出现。

For whatever reason, it seems to come up around this time of year.

2024年，在多个实验室报告模型改进速度放缓之后，我们看到了对AI扩展遇到瓶颈的巨大担忧。

In 2024, we saw massive concern about AI scaling hitting a wall after reports that multiple labs were seeing a slowdown in model improvement.

但就在那之后不久，我们迎来了第一批推理模型，再往后，芯片内存的进步使得训练数万亿参数的模型成为可能。

But then just shortly after that, we got the first reasoning models, and farther after that, advances in chips with more memory that allowed multi-trillion parameter models to be trained.

然后，当然，去年有人担心企业没有从AI中看到回报，不会再继续投入。

Then, of course, last year, there was a concern that enterprises weren't seeing returns from AI and wouldn't continue paying.

这当然体现在让我头疼的MIT报告中，说95%的生成式AI试点项目都失败了。

This, of course, embodied in the bane of my existence, the MIT report about the 95% of generative AI pilots failing.

剧透一下，企业支出并没有枯竭。

Spoiler alert, enterprise spend didn't dry up.

相反，它蓬勃发展，而且每个季度还在以大约50%的速度增长。

Instead, it boomed and is still growing at around 50% a quarter.

现在，之前的泡沫预测错了，并不意味着未来所有的泡沫预测也都会错。

Now, previous bubble calls being wrong does not mean that all bubble calls in the future will be wrong as well.

这仅仅意味着，我们最好对所有这些夸张的大标题保持怀疑态度。

It just means we'd all be better served carrying around bags of salt at this point to take with all of these big, bombastic headlines.

随着Anthropic的IPO临近，我们很快就会得到更明确的答案。

With the Anthropic IPO coming up, we will get a much more definitive answer soon.

无论结果如何，影响都还在未来，既让人期待又让人无奈，而现在我们只能猜测和讨论。

The impacts of whatever that shows remain tantalizingly and frustratingly in the future, and for now, all we can do is speculate and debate.

希望这些内容能给你提供更多谈资，今天的《AI每日简报》就到这里。

Hopefully, you've got a bit more fodder for that now, and that's going to do it for today's "AI Daily Brief."

感谢你的收听或收看，一如既往，下次再见，祝一切顺利。

Appreciate you listening or watching, as always, and until next time, peace.