What a $30B Hedge Fund Implosion Really Means for AI — 文稿

M10:00

今天这期 AI Daily Brief，要聊的是夸张到离谱的营收增长，不过同时还有一家对冲基金爆雷？

Today on the AI Daily Brief, insane revenue growth, but also a hedge fund blowup?

AI 和市场到底发生了什么？

What is going on with AI and markets?

AI Daily Brief 是一档每天更新的播客和视频节目，聊的是 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，各位朋友，在我们开始之前先快速说几个通知。

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

M10:35

在正式开始之前，再补充两个很快的小提示。

Two more quick notes before we dive in.

首先，今天这期节目属于那种， headline 里的所有新闻也都刚好契合主话题，所以这一期就只有 main。

First of all, today is one of those episodes where all of the stories in the headlines also fit the theme of the main, so it's gonna be a main only.

M10:54

我很期待看看你们这个周末会做些什么。

I'm excited to see what you do this weekend.

不过说到这儿，我们来聊点数字。

But with that, let's talk some numbers.

今天我们有两条新闻，乍一看像是在讲 AI 周边市场里完全不同的两种故事。

Today we have two stories that feel on first glance like they're telling totally different stories about the markets surrounding AI.

一方面，我们看到了 AI lab 营收方面完全离谱的预估和真实数字，在很多层面上都让人很难真正消化。

On the one hand, we have just absolutely bonkers estimates and real numbers for AI lab revenue, which are in many ways genuinely hard to wrap your head around.

而另一方面，则是一家由神童式人物掌舵的对冲基金彻底崩盘，这也再次引发了大家对 AI 市场持续性的质疑。

And on the other side, we have the utter implosion of a wunderkind-led hedge fund that is spurring renewed questions about the durability of AI markets.

所以我们来看看，这两件非常不同的事件到底各自在说明什么，以及它们对接下来 AI 市场的走向意味着什么。

So let's figure out what stories these two very different events are telling and where they point for AI markets next.

我们先从营收这边说起，因为 OpenAI 和 Anthropic 看起来都正在迎来一波营收增长的回升。

We're going to start on the revenue side, where both OpenAI and Anthropic appear to be having a resurgence in revenue growth.

CNBC 报道说，在最近一次全员会上，OpenAI 的 CFO Sarah Friar 告诉员工，七月份的 ARR，也就是 annualized recurring revenue，已经超过了整个第二季度，她还补了一句，第二季度本身也一点都不差。

CNBC reported that during a recent all-hands, OpenAI CFO Sarah Friar told staff that ARR, annualized recurring revenue, for July had exceeded the entire second quarter, adding, and Q2 was no slouch.

不过，如果没有完整上下文，其实不太清楚 Friar 这句话具体是什么意思，而且那些报道也没有把这件事解释得更明白。

Now, without the full context, it is not exactly clear what Friar meant, and the articles didn't do a lot to clear that up.

但可以确定的是，大家传达出来的结论就是，OpenAI 那一个月简直是大丰收。

But the takeaway certainly was that OpenAI had an absolute bonanza of a month.

接着到了周四，Axios 报道说 Anthropic 的营收也在飞速飙升。

Then on Thursday, Axios reported that Anthropic was also seeing revenue skyrocket.

确实，粗略算一算，Anthropic 的 run rate 大概到了七百一十亿美元，而他们上一次谈到营收是在五月，当时是四百七十亿美元。

Indeed, back-of-the-napkin math put Anthropic at a $71 billion run rate, up from $47 billion in May when they last discussed revenue.

这个数字是基于 Tae Kim 的一篇帖子，他当时引用的是 AI 投资研究平台 FundA 的数据。

This figure was based on a post from Tae Kim, who was referencing data from AI investment research platform FundA.

与此同时，他们的数据还显示，OpenAI 的 ARR 也已经接近五百亿美元。

Meanwhile, their data also showed that OpenAI was sitting just shy of $50 billion in ARR.

当然，很明显，这些数据只能当成非常粗略的估算来看，但结合 Friar 的说法，大方向上看起来应该是对的。

Now, obviously, this data should be treated as a very rough estimate, but it seems directionally correct based on Friar's comments.

这也和 SemiAnalysis 的估计一致。他们在这个月月初写道，Anthropic 目前的 ARR 已经高于六百亿美元，而且看起来这个季度结束时会有十亿美元利润。

It also lines up with estimates from SemiAnalysis, who at the beginning of the month wrote that Anthropic is currently operating above $60 billion in ARR and looks set to end the quarter with $1 billion in profit.

而且有些人觉得，他们还会继续往上冲。

And some believe they will just keep going.

在最近的一篇博客里，Dwarkesh Patel 写道，Anthropic 很可能会以一千亿到一千五百亿美元的营收结束今年。

In a recent blog post, Dwarkesh Patel wrote, Anthropic likely ends the year with $100 to $150 billion in revenue.

一方面，这些数字看起来确实夸张得惊人，但如果 Anthropic 真的在光是七月份就增加了一百亿美元 ARR，那也不是完全不可能。

Now, on the one hand, those numbers seem absolutely gobsmacking, but if Anthropic really jumped $10 billion in ARR in July alone, it doesn't seem impossible.

前 Atlantic 作者 Derek Thompson 指出，Dwarkesh 这个人有机会接触很多幕后的交流，所以如果 Anthropic 真能达到这个水平，那它的创收能力就会超过 Tesla 和 SpaceX 两家加起来。

Pointing out that Dwarkesh is in a position to have a lot of behind-the-scenes conversations, former Atlantic author Derek Thompson noted that if Anthropic can hit this mark, they will have eclipsed the revenue-generating capacity of Tesla and SpaceX combined.

对于那些没有一直密切关注的人来说，这一波暴涨也像是凭空冒出来的一样。

For those not paying close attention, the surge also felt like it came out of absolutely nowhere.

就在一周前，The Wall Street Journal 还发过一篇文章，说美国企业界突然决定不再往 AI 上大把烧钱了。

Just a week ago, The Wall Street Journal wrote an article about how corporate America had suddenly decided to stop blowing money on AI.

当然，那是他们的原话，不是我说的。

That is obviously their words, not mine.

而且表面上看，这个说法似乎也挺合理。

And on the face of it, it seemed reasonable.

过去几个月，媒体关于 enterprise AI 的很多叙事，都是 CFO 们在想办法收紧 token 预算，用更便宜的开源 Chinese models 去替代昂贵的 frontier AI。

So much of the media's story around enterprise AI for the past few months is CFOs trying to rein in token budgets and substituting expensive frontier AI for open-source Chinese models.

那这两家公司到底是怎么做到，反而创下自己最好月份之一的呢？

And so how the heck did these 2 companies have one of their best months yet?

我反复提过的两个点，我就借这个机会再强调一遍。

2 points that I've made repeatedly that I will use this as a chance to reinforce.

说实话，这其实就是同一个点的两个部分。

Which are honestly actually just part and parcel of the same point.

那就是，我们现在消耗掉的，只是 AI 对 intelligence 总需求里极其小、甚至小到几乎可以忽略不计的一部分。

And that is we are currently consuming a tiny, even vanishingly small percentage of the total possible demand for intelligence from AI.

是的，确实有非常非常有限的一小撮公司，它们的一部分用户已经用得够深了，以至于它们真的不得不去设定 token 限额之类的东西。

Yes, we have a very, very limited handful of companies who have a portion of their users that are deep enough that they actually have to do things like impose token limits.

但绝大多数用户群体，仍然在往上走，头顶上还有很长很长的空间。

But the vast majority of the user base remains on the upswing with miles and miles of air above them.

企业里正在发生的，不是公司决定不花钱了，而是它们看到了早期的警报：等 AI 真正发展到完全成熟的大规模阶段时，它们最终不能做的事情，就是把 Claude Opus 5 直接拿来解决自己每一个问题。

What's happening in the enterprise is not that companies have decided to stop spending, it's that they are seeing the early warning shots that what they can't do ultimately as AI gets to full mature scale is simply deploy Claude Opus 5 for every single problem they have.

换句话说，它们是在未来这个问题真正爆发之前先往前预防，通过建立更复杂的 AI 使用架构来应对，这些架构会包含多个不同模型，以及更聪明的 caching 和 provisioning 安排。

They are, in other words, looking to get out ahead of a problem which is primarily the domain of the future, by creating more complex AI usage architectures that involve multiple different models and smart caching and provisioning arrangements.

但几乎没有任何情况是，公司突然就在更少地使用 AI 了。

But in almost no cases are companies all of a sudden using less AI.

现在，我个人的看法是，未来好几年，在 total intelligence demand 这件事上，我们都会处在那种上升通道里，前方还有很长的空间。

Now, it's my personal opinion that we are going to be on that upswing with miles of air ahead of us in terms of total intelligence demand for years to come.

原因很简单，就是物理规律。

And the reason for that is simply physics.

需求增长的速度，会比我们把更多 intelligence 上线的能力更快，而且会一直更快。

The speed at which demand will increase is faster and will be faster than our ability to bring more intelligence online.

事实证明，在这个规模下，把更多 intelligence 上线所需的整套基础设施都建出来，花的时间就是比需求增长更久。

It turns out that building the entire slate of infrastructure needed to bring more intelligence online at this scale just takes longer than the demand grows.

这就引出了我要讲的第二个小点，我认为，OpenAI 或 Anthropic 在它们所在区间里，几乎以任何成本产出的每一个 token，都会有人买单。

Which brings me to the second subpoint that I'll make, which is that I believe that every single token that OpenAI or Anthropic produce at almost any cost within the bands of where they are will be bought.

更低价模型的加入，以及那些做出不同取舍的模型的存在，只会增加营收，不会整体上减掉营收。

The addition and presence of lower-priced models and models that make different trade-offs will be adding to the top line, not subtracting overall.

这并不是说这些公司会坐着不动，只靠高价产品慢慢发酵。

Which is not to say that these companies are going to sit back and marinate only in their high-priced offerings.

事实上，我们也刚刚听到消息，OpenAI 在降价。

In fact, we also just got news that OpenAI is slashing prices.

从周四开始，GPT-5 的两个小版本都降了价，Nano 版本降了 80%，到每百万个 output tokens 1.20 美元；Mini 版本降了 20%，到每百万个 output tokens 2 美元。

Effective Thursday, the 2 smaller versions of GPT-5 saw a price cut, with Nano down 80% to $1.20 per million output tokens, and Mini down 20% to $2 per million output tokens.

Pro 的价格保持不变，不过 OpenAI 推出了一个新的 fast mode，速度提升 2.5 倍。

Pro prices were held steady, but OpenAI introduced a new fast mode with a 2.5x speed boost.

这很明显说明，他们已经意识到，成本是他们必须去竞争的一个维度，而且更高效的模型，竞争格局也会越来越复杂。

Now, this shows some clear awareness that cost is a vector that they need to compete on and that more efficient models are going to increasingly have a more complicated competitive landscape.

但别搞错了，这可不是因为 OpenAI 卖 intelligence 卖不动了，才搞的打折促销。

But make no mistake, this is not a discount sale because OpenAI is having any problem selling intelligence.

这其实是在战略性地利用他们自己的位置，试图巩固市场里某一块未来可能会变成弱点的部分。

It's strategically leveraging their position to try to shore up a part of the market that could increasingly be a weakness.

现在，为什么这些收入数字不只是播客里好玩的标题，而是真的重要，是因为对 OpenAI 和 Anthropic 模型的需求，处在 AI economy 里其他一切东西的上游。

Now, why all of these revenue numbers matter as more than just interesting headlines for podcasts is that demand for OpenAI and Anthropic models is upstream of everything else in the AI economy.

这个逻辑很简单：OpenAI 和 Anthropic 的收入要上涨到足以证明继续投入越来越大规模的新 capex 开发是合理的，也就是 data centers，当然还包括为这轮扩张所需的融资；所以说，真正让人担心的，其实是相反的情况。

The equation pretty simply is OpenAI and Anthropic revenue going up enough to justify increasingly large expenditures on new capex development, i.e., data centers, and the financing that's requiring for that buildout, meaning of course that the concern is on the inverse.

如果需求下滑，那这些事情就都不再有财务上的意义，一切都可能轰然崩塌。

If demand were to fall, none of those things would make financial sense and it could all come crashing down.

但就目前来看，说到收入，往上看真的只有一片开阔天空。

But at the moment, when it comes to the revenue, it is just nothing but air up there.

现在，仔细听我节目的朋友应该记得，我以前说过，这些巨大的收入数字，以及它们所代表的从把 AI business models 想成 seats，转而去想 tokens 的 total addressable market 的转变，是怎么让很多人不再担心去年那么火的 Q4 AI bubble 的。

Now, close listeners will have heard me talk about before the fact that these massive revenue numbers and the shift that they represented from thinking about AI business models in terms of seats to instead thinking about it in the total addressable market for tokens is what got a lot of folks off of their Q4 AI bubble worries that were such a big story last year.

那今年，AI bears 把担忧点集中在哪儿了呢？

So where have the AI bears focused their concerns this year?

在某种程度上，bear case 其实并没有变。

To some extent, the bear case hasn't changed.

这背后还是基于一些担忧，比如说对 Nvidia 循环交易的担忧。尤其是最近又在传，Nvidia 可能要为 OpenAI 两千五百亿美元的数据中心需求兜底，所以这种声音最近又变得更响了。

It's still based on things like concerns about Nvidia's circular deals, which, given recent talks for Nvidia to backstop $250 billion of OpenAI's data center demand, have come back louder recently.

说到 semiconductor 这条交易线，很多分析师——甚至可能是大多数——还是把这个行业看成周期性的，觉得需求注定会崩掉。虽然就这个具体情况来说，我真不觉得以前有什么例子比现在更适合说明“过去的结果不保证未来的表现”这句话。也就是说，AI boom 之后的 semiconductor 交易，现在跟以前已经完全、从根本上不一样了。

When it comes to the semiconductor trade, many, if not most analysts, still view the industry as cyclical, with demand destined to crash, even though in that particular case, I don't know that there's ever been quite as good an example of past results do not guarantee future performance, i.e., the semiconductor trade now is totally and fundamentally different to what it was before the AI boom.

当然，先说清楚，这不代表它就不可能出问题。

Which to be clear doesn't mean that it can't go badly.

只是说，就算出问题，也不会再是以前那种周期性原因导致的。

It's just not going to go badly for the same cyclical reasons that it did before.

而且，市场里始终有一种说法：AI 的需求根本不可能一直增长下去，尤其是随着更便宜的 Chinese models 变得已经足够好、可以替代，它可能已经见顶了。

And indeed, there's the ever-present argument that AI demand simply can't continue to grow and has perhaps reached the high water mark as cheaper Chinese models become good enough to substitute.

那这个观点，显然也正是我刚刚回应过的。

Now, obviously, that's just what I addressed.

我在这个问题上的立场其实非常明确。

My position on that is pretty clear.

不过，非常重要的一点是，我们任何人如果主要只接触市场里的某一个板块、或者某一个主题，就很容易觉得那就是市场里最重要的事。

However, very importantly, it is very easy for any of us who intersect primarily with one part or one theme of the market to think that that's the most important thing in markets.

但如果你稍微往 AI ecosystem 外面看一眼，就会发现，现在很多看空 AI 的情绪，跟 AI 的基本面其实完全没关系。

However, if you take a quick peek outside of the AI ecosystem, you'll notice that a lot of the bearish sentiment on AI right now has absolutely nothing to do with AI fundamentals.

这周，Federal Reserve 没有加息，但现在很多分析师都认为，只要通胀再抬头，很快就会迎来一次加息。

This week, the Federal Reserve declined to raise interest rates, but many analysts now believe a rate hike is coming as soon as inflation picks up.

Iran 战争持续不断的混乱局面，也让投资者非常紧张。

The ongoing fracas of the Iran war is making investors extremely jittery.

而且从年初那种宏观环境很适合投机性科技股大涨的状态，到过去这几个月，市场已经明显转向了一种 risk-off 的情绪。

And from the beginning of the year, where we saw macro conditions ripe for a boom in speculative tech stocks, over the past couple of months, the market has transitioned into a decidedly risk-off mood.

对我们来说，重要的是要能分辨：AI 相关资产价格下跌，到底是因为大家对 AI 的看法变了，还是只是因为它们作为最大的一批股票，受到了更广泛市场担忧的影响。

It is important for us to be able to distinguish when AI-related market prices going down has to do with some change in belief about AI, or whether they are just the biggest stocks being influenced by broader concerns.

总体来看，Nasdaq 今年到现在是小幅下跌。尽管七月份有过一波强劲反弹，但现在又一次站在了进入回调的边缘。

Overall, the Nasdaq is currently down slightly on the year, and despite a strong recovery in July, is once again on the brink of a correction.

过去一个月里，Mag 7 基本走平，而 semiconductor index 则从六月份高点回撤了百分之二十三。

The Mag 7 is basically flat over the past month, and the semiconductor index has taken a 23% drawdown from June highs.

software index 这个月上涨了百分之三，但这其实代表的是资金从 AI trade 里轮动出来，转去了那些在 SaaSpocalypse 期间被狠狠打压的公司。

The software index is up 3% for the month, but that represents a rotation out of the AI trade and into the names that were beat up during the SaaSpocalypse.

而这还只是美国市场。我们也总是不断惊讶地发现，世界上显然不只有这一个市场。

And that's just the US market, which we are continuously surprised to discover is not the only market in the world.

如果把视线放到海外，拿全球联动性很强的 South Korea 举个例子，你会发现那边的市场简直是一团糟。

If you look abroad to South Korea for one globally interconnected example, that market is in absolute shambles.

韩国的主要指数 KOSPI 在一个月里下跌了百分之四十，成了韩国历史上最严重的一次股灾。

The major Korean index known as the KOSPI is down 40% in a month, making it the worst stock crash in Korean history.

比九十年代的 Asian financial crisis 还惨，比二零零八年的 GFC 还惨。

Worse than the Asian financial crisis in the '90s, worse than the GFC in 2008.

当然，韩国股市的构成方式跟美国很不一样。美国的大盘指数追踪的是五百只股票，而韩国这个主要指数追踪的是一百只，而且其中真正大到足以影响全局的，其实只有极少数几家公司。

Now, the Korean stock market is composed very differently to the U.S. Rather than the major index tracking 500 stocks, it tracks 100, and only a tiny handful of them are large enough to matter.

排在最前面的就是 Samsung 和 SK Hynix，这两大 memory 生产商加起来，自己就大约占了整个指数的百分之五十。

Right up the top of that list are Samsung and SK Hynix, the two major memory producers that make up around 50% of the index by themselves.

对比一下的话，Mag 7 大约占 S&P 500 的百分之三十。

By way of comparison, the Mag 7 are about 30% of the S&P 500.

另一个很大的区别，是韩国散户交易者的行为方式。

The other big difference is the behavior of Korean retail traders.

韩国大概有百分之三十的人口会积极参与交易，而美国大概只有百分之零点二。

Around 30% of the population actively trades compared to around 0.2% in the US.

而且他们出了名地喜欢加杠杆。leverage 这个词，我们在节目后半段还会反复提到。韩国散户爱杠杆爱到什么程度呢？韩国监管机构最近甚至禁止推出新的 leveraged ETF，来维护金融稳定。

They also notoriously love leverage, a word that we are going to talk about a lot in the latter part of this show, to the point that Korean regulators recently banned new leveraged ETFs to protect financial stability.

根据 Goldman Sachs 的说法，这周大约有一百二十万个韩国账户收到了 margin call，也就是说，他们必须再补钱进去，才能撑住仓位。

According to Goldman Sachs, around 1.2 million Korean accounts were margin called this week, meaning that they had to deposit more money to stay afloat.

这差不多相当于韩国成年人口的百分之三点四。

That's around 3.4% of the adult population.

其中多达三十六万个账户被强制平仓，也就是说，因为没能满足 margin call 的要求，他们被迫卖出了资产。

As many as 360,000 accounts were liquidated, meaning they were forced to sell after failing to meet a margin call.

想把市场波动简单归结成只是这个因素或者那个因素，永远都是很危险的。

It's always dangerous to try to say market moves are only about this thing versus that thing.

而且现在，你会看到大家在围绕这些具体动作争夺不同的解读。

And right now you're seeing a competition for interpretations around these particular moves.

对一些人来说，这是出了名有周期性的 semiconductor 行业，在长期 AI 需求受到质疑的时候，经历一次可预期的暴跌。

To some, it's a notoriously cyclical semiconductor industry experiencing a predictable crash as long-term AI demand is questioned.

但另一方面，眼下其实还有一个非常机械性的故事正在上演。

But on the other hand, there is a very mechanical story playing out.

因为相当大一部分韩国人口，被迫在价格暴跌时卖出。

As a sizable chunk of the Korean population is forced to sell into a price crash.

这两种解读，对这一切到底说明了整体 AI 市场处于什么状态，有着天差地别的含义。

Those two interpretations have wildly different implications for what this all says about the state of AI markets in general.

另一个在看空 AI 的人当中广泛流传的大故事，是表外债务的问题。

Another big story doing the rounds with AI bears is the problem of off-balance-sheet debt.

正如我这几个月一直在讨论的，我们现在已经到了这样一个阶段：hyperscalers 越来越没办法继续自己掏钱来支撑基础设施扩建，转而越来越多地依靠债务来为这轮建设融资。

As I've been discussing for months, we're at the phase where the hyperscalers are increasingly having to turn away from funding the infrastructure buildout themselves towards instead turning to debt to fund that buildout.

上周，Nikkei Asia 稍微深挖了一下，整理出了一些相关数字。

Last week, Nikkei Asia did a little digging and came up with some numbers around that.

他们认为，hyperscalers 持有一点六五万亿美元的数据中心债务，而这些债务很大程度上并不是放在他们自己的资产负债表上，而是放在 special purpose vehicles 里。这个东西本质上就是为了给单个项目融资而专门设立的空壳公司。

They believe that hyperscalers are holding $1.65 trillion in data center debt, debt which is largely being held by the companies not on their balance sheets, but in special purpose vehicles, which are effectively shell companies spun up for the sole purpose of financing individual projects.

很震惊，很可怕，很吓人，对吧？

Shocking, horrifying, terrifying, right?

嗯，如果你去听那些看空 AI 的人的说法，这里面确实有 subprime 的影子。

Well, if you listen to the AI bears, this has shadows of subprime.

hyperscalers 没有把债务放在自己的资产负债表上、按规定向市场披露，而是把它藏进了空壳公司里。

Rather than putting debt on their own balance sheet where they have to disclose it to the market, the hyperscalers are hiding it in shell companies.

这些债务又被切分、重组，包装成 structured credit products，卖给保险公司、private credit firms，还有养老金。

This debt is sliced and diced into structured credit products and sold off to insurance companies, private credit firms, and pensions.

如果你是 California 的老师，那你的养老金可能就暴露在这种危险又隐蔽的资产上。

If you're a California teacher, your pension is exposed to this dangerous and shadowy asset.

但这其实并不是完整的故事，而且拿它跟 subprime 相比，往往也只是停留在表面。

But that's not really the full story, and the comparisons to subprime tend not to make it past the surface level.

在他那份非常出色的 newsletter《Notes on the Crises》里，Nathan Tankus 非常细致地对比了隐藏的数据中心债务，以及二零零八年让整个金融系统几乎瘫痪的 CDOs。

In his fantastic newsletter Notes on the Crises, Nathan Tankus wrote an extremely detailed comparison between hidden data center debt and the CDOs that brought the financial system to its knees in 2008.

而且在这个具体问题上，我确实觉得有必要提醒一下信息来源是谁。

And in this particular instance, I do think it is worth noting the source.

Nathan 不是那种像 David Sacks 一样、在 Silicon Valley 的 venture capitalist，一边替自己站台一边发言的人。

Nathan is not a David Sacks Silicon Valley venture capitalist talking his own book.

他是一个立场非常偏左、也非常真诚的市场评论员，从 COVID 以来就一直在写这份 newsletter。过去大概五年里，如果你像我一样一直关注他，你会发现，从性格上说，他基本不可能去做那种不是建立在自己能找到的最佳信息之上的论证，不管这个论证最后恰好支持了谁的 talking points。

He's an extremely left-leaning, extremely earnest markets commentator who's been writing this newsletter since COVID who, if you have watched for the last five years or so as I have, is basically temperamentally incapable of making an argument that isn't just based on the best information that he can find, regardless of whose talking points it reinforces.

总之，说到这种把隐藏的数据中心债务和 collateralized debt obligations 作比较的说法，Nathan 的核心观点是，hyperscalers 和金融危机时期那些 subprime 借款人，在根本上就是两类完全不同的借款人。

In any case, when it comes to this comparison between the hidden data center debt and collateralized debt obligations, Nathan's core argument is that the hyperscalers are fundamentally a different type of borrower to these subprime borrowers during the financial crisis.

如果你相信这种隐藏债务会再次把整个经济拖垮，那你就得赌至少有几家 hyperscalers 会破产，不只是现金流吃紧，也不只是股价腰斩。

Believing that hidden debt will once again break the economy requires you to bet that at least a couple of the hyperscalers go bankrupt, not struggle to maintain cash flows, not see their stock price cut in half.

而是真的还不上债，发生违约。

But actually default on their debts.

考虑到这些借款人整体上的实力，这其实是一个门槛非常高的判断。

That is quite a burden of belief given the overall strength of these borrowers.

然后，另一个重大区别在于，这些债务到底是怎么被使用的。

Now, the other big difference is how this debt is being used.

在金融危机爆发前的那段时间里，subprime mortgages 被重新打包成了 investment-grade 的 CDOs。

In the lead-up to the financial crisis, subprime mortgages were repackaged as investment-grade CDOs.

在银行间清算系统里，它们被当成基本上和美国政府发行的 Treasury bills 一样的东西来使用。

They were accepted as basically the same as U.S. government-issued Treasury bills for use in the interbank settlement system.

而二零零八年这个体系一旦断裂，真正造成最大破坏的恰恰是这部分，远远超过了“房价下跌”本身，或者 Bear Stearns、AIG 的倒下。

This was the part of the system that did the most damage when it broke in 2008, much more so than the crisis being about home prices falling or the collapse of Bear Stearns or AIG.

那些都只是金融系统地基开始崩塌之后表现出来的症状。

Those were symptoms of the foundations of the financial system falling apart.

但这一次，没有人假装数据中心债务和 Treasury bills 是一回事。

This time around, no one is pretending that data center debt is the same as Treasury bills.

这些债务主要是卖给 private credit firms，然后塞到保险公司和养老金的资产负债表上。

The debt is largely being sold to private credit firms and shoved onto the balance sheets of insurance companies and pension funds.

先说明一下，要是 data center 行业崩盘、这些债务又发生违约，那对这些机构肯定会非常糟糕。但至少在当下，就算真的出现这些违约——虽然看起来概率不大——也没有那种会把事情演变成系统性危机的机制，更不太可能像二零零八年金融危机那样蔓延。

Now, to be clear, it would be very bad for these organizations if the data center industry collapses and these debts default, but there, at this moment at least, are not the same sort of mechanisms for even those defaults, as improbable as they seem, to cause a systemic crisis in any sort of manner similar to the financial crisis of 2008.

当然，没有系统性风险，并不代表谁就可以对当前的债务状况，或者债务继续上升的趋势，掉以轻心。但我觉得，我们还是应该按它真实的样子来处理，而不是硬往某种历史类比上套，那种类比大多也就是为了吸眼球、抢标题。

Now, just because there isn't systemic risk doesn't mean anyone should be flippant around either the current state of debt or the trajectory of debt, but I think that we should deal with it as it actually is rather than plumbing for this sort of historical analogy, which is mostly interesting for grabbing headlines.

M117:41

好，那在进入这周科技公司财报季之前，以上这些就是整个大背景。

Now, all of this was the background as we entered this week of tech earnings.

大家脑子里都在想一个问题：这些 hyperscalers 里，会不会有人开始缩减 CapEx，并宣布 AI 支出并没有带来足够的回报。

The question on everyone's mind was whether any of the hyperscalers would pull back on CapEx and announce that AI spending was failing to deliver sufficient returns.

Google 上周三已经第一个交卷了。

Google had already gone first last Wednesday.

虽然增长很强，但他们也宣布，随着 CapEx 超过利润，他们出现了多年来第一个现金流为负的季度。

While growth was strong, they announced their first cash flow-negative quarter in years as CapEx overtook profits.

而投资者听到的就只有这一条消息，结果股价一路大跌。

And that was the only message investors heard, sending the stock price tumbling.

这周三，Meta 和 Microsoft 在同一个晚上公布了财报。

This Wednesday saw Meta and Microsoft report on the same night.

Meta 的财报反响很差，分析师还是没搞清楚这家公司的 AI 战略到底是什么。

Meta's earnings were very poorly received, with analysts remaining unclear on what the company's AI strategy actually was.

Mark Zuckerberg 小心翼翼地提到过把多余算力租出去这个想法，但他说得很保守，意思更像是 Meta 还是自己留着用比较划算。

Mark Zuckerberg tiptoed around the idea of renting out spare capacity, suggesting Meta was still better off keeping it for their own use.

至于 AI 销售，Zuckerberg 拿出来的是一堆混杂在一起的选项，在很多人看来，根本没法真正讲通。

And as for AI sales, Zuckerberg presented a melange of options that to many didn't seem to really stack up.

很明显，他还是一心想把 AI agents 卖给消费者，也就是他所谓的 personal superintelligence。但问题也很明显，光靠消费者本身，不可能支撑起几千亿美元的 CapEx。换句话说，这类收入更像是靠 seats，而不是靠总体 token 用量。

He is clearly still set on selling AI agents to consumers, what he calls personal superintelligence, but it's obvious that consumers by themselves can't justify hundreds of billions of dollars in CapEx, i.e., those revenues are much more about seats than about aggregate tokens.

接着，Zuckerberg 又列了一些选项，包括 enterprise AI、把他们内部的生产力工具拆分出来做成独立产品，以及一系列新的 vibe-coded apps。

Zuckerberg then listed options including enterprise AI, spinning off their internal productivity tools as separate products, and a constellation of new vibe-coded apps.

不过他也承认，这些方向都要求 Meta 练出一套很不一样的能力，而股价也立刻应声下跌。

However, he acknowledged that these would all require Meta to flex a quite different muscle, and the stock immediately sank.

相比之下，Microsoft 的财报就形成了非常鲜明的对比。

Microsoft's earnings were a stark contrast.

CFO Amy Hood 传达的信息非常明确：资本纪律。

CFO Amy Hood presented a clear message of capital discipline.

财报本身很强，但最大的重点是，Hood 预测 Microsoft 至少在未来一年里会继续保持正现金流。

Earnings were strong, but the big takeaway was Hood's forecast that Microsoft would continue to be cash flow positive for at least the next year.

这对 Microsoft 来说，其实是个相当大的取舍。

Now, this is a pretty big trade-off for Microsoft.

Azure 正在快速增长，年化经常性收入，也就是 ARR，第一次达到一千亿美元，现在大约占到未来预期收入的四分之一。

Azure is booming, reaching $100 billion in ARR for the first time and now representing around a quarter of forward revenue.

限制 CapEx，也就意味着他们同时在限制增长。但很明显，他们准确读懂了市场情绪：现在市场根本不想要那种不计代价的增长，所以也非常慷慨地奖励了 Microsoft 的这种谨慎。

By limiting their CapEx, they are also limiting growth, but it is very clear that they read the room correctly and that the market is not in the mood to want growth at all costs and handsomely rewarded Microsoft for their prudence.

现在，对所有这些公司来说，最大的挑战依然是：怎么让那些只看很短期的投资者，至少足够站在你这边，好让你继续做那些对长期必须做的事；同时，又不能为了迎合短期，牺牲掉太多你在长期竞争中真正需要的东西。

Now, the great challenge for all of these companies continues to be how to get very short-term investors on side enough to be able to continue to do what you need to do for the long term while also not making too many concessions to what you need in the long term to compete.

这周最后登场的是 Amazon，他们公布了强劲的财报，同时 CapEx 也小幅上调。

Rounding out the week was Amazon, who reported strong earnings and a slight CapEx bump.

AWS 销售额同比增长了百分之三十七，而这看起来已经足够让投资者接受今年的 CapEx 预测从两千亿美元上调到两千两百亿美元。

AWS sales are up 37% year over year, and that seems good enough for investors to endorse this year's CapEx forecast rising from $200 billion to $220 billion.

CEO Andy Jassy 把这次增加解释为成本上升，而不是规模扩张，不过他也为这笔支出做了辩护，表示这显然是必要的。

CEO Andy Jassy explained the increase as increased costs rather than expanded scope, but he also defended the spending as clearly necessary.

Jassy 说，就算花到这个数，我们到二零二六年还是没有足够的产能来满足现有的全部需求，而且我相信这种情况到了二零二七年也依然会存在。

Jassy said, even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027.

事实上，我们现在已经看到的二零二八年需求，都非常惊人。

In fact, the demand we already have for 2028 is striking.

所以朋友们，这也再次说明，Anthropic 和 OpenAI 收入加速增长这件事为什么那么重要。

And this again, friends, is why the acceleration of revenue for Anthropic and OpenAI matters so much.

它带来的影响之一，就是给了这些 hyperscalers 足够的腾挪空间，让他们可以提高支出，去匹配不断上升的成本，并让这波 AI 交易继续滚下去。

Among other effects, it gives the hyperscalers the latitude they need to hike spending to match rising costs and keep the AI trade rolling.

这些 hyperscalers 的说法是，需求还在，而且还在继续上升。

The hyperscalers say the demand is still there and it's still rising.

这也就把我们带到了可能是本周最疯狂的一条故事。

Which brings us to perhaps the wildest story of the week.

科技财报很快就被放到次要位置了。

Tech earnings quickly took a backseat.

因为最有名的那个 AI 对冲基金爆了。

As the most famous AI hedge fund blew up.

周五早上有消息说，Leopold Aschenbrenner 的基金 Situational Awareness 已经被清盘，并且由 Citadel 接手。

On Friday morning, it was reported that Leopold Aschenbrenner's fund, Situational Awareness, had been liquidated and taken over by Citadel.

现在回头看，Leopold 的经历一直都很传奇。

Now, Leopold has had a crazy story.

他跟我一样，都是 FTX 那场疯狂风波里的难民；也跟我一样，后来都进入了 AI，不过他切入这个行业的位置和我很不一样。

He, like me, was a refugee of FTX madness, and he, like me, found his way into AI, although intersecting with the industry in a very different spot.

他原本在 OpenAI，但后来离开了，外界一直有不少疑问，不确定到底是因为政策分歧，还是因为他对内部事务说得太多。

While he was originally at OpenAI, he left under some cloud of questions around whether it was disagreements in policy or him speaking too freely about internal matters.

不过，他的故事真正开始迅速发酵，是在二零二四年夏天。当时他发出了那篇同名文章《Situational Awareness》，这是一篇长达一百六十五页的大部头，让华尔街很多人第一次意识到 AI 很可能会从根本上改变世界。

But where his story really picked up steam was the summer of 2024 when he dropped that namesake essay, Situational Awareness, a 165-page tome that woke many on Wall Street up to how fundamentally AI would likely change the world.

Aschenbrenner 借着新获得的名气，募到了几亿美元，开始成立一个对冲基金，押注 AI 的崛起。

Aschenbrenner used his newfound reputation to raise several hundred million dollars to start a hedge fund to bet on the rise of AI.

接下来的几个月里，又有几十亿美元流入这个基金，Situational Awareness 也开始交出行业领先的回报。

Over the coming months, several billion poured into the fund, and Situational Awareness started posting sector-leading returns.

到了二零二五年年中，很多人都把 Situational Awareness 看成史上最成功的对冲基金故事之一，而把 Aschenbrenner 视为真正的天才少年。

By the middle of 2025, many were viewing Situational Awareness as one of the most successful hedge fund stories ever, and Aschenbrenner as a genuine wunderkind.

就连到今年第一季度末，Situational Awareness 看起来都还是一路高歌猛进。

As recently as the end of Q1, Situational Awareness was still riding high.

基金当时报出了百分之四百三十九的净回报。用一个非常“专业”的说法，这对一家对冲基金来说简直离谱。

The fund reported 439% net returns, which is, to use a highly technical phrase, absolutely insane for a hedge fund.

其他基金也在复制他们的交易，这意味着它在下游产生的影响，成了 neo-cloud 和 semiconductors 这波上涨中的一个巨大推动力。

Other funds copied their trades, meaning its downstream impact was a huge driver in the run-up in neo-clouds and semiconductors.

它身上很大一部分神秘色彩，就在于 Aschenbrenner 才二十出头，而且完全没有金融背景。

A big part of the mystique was that Aschenbrenner was in his early 20s with no finance background.

但结果证明了，就算是对未来判断惊人、眼光超强的 AI 神童，也一样会被杠杆狠狠干翻。

And yet, it turns out even AI prodigies with incredible foresight can get smashed by leverage as well.

而就在本周早些时候，裂缝开始显现了。

Earlier this week, the cracks started to show.

最开始只是一些传言，说因为 semis 大幅下跌，这只基金已经收到了 margin call。

It began with whispers that the fund had been margin-called due to the sharp decline in semis.

接着到了周三，Financial Times 报道说，这只基金正在寻求额外资金。

Then on Wednesday, the Financial Times reported that the fund was seeking additional capital.

到了周四早上，一切都结束了。

By Thursday morning, it was all over.

全球最大的交易公司之一 Citadel Securities，实际上已经把 Situational Awareness 买下来了。

Citadel Securities, one of the largest trading firms in the world, had effectively bought out Situational Awareness.

现在，考虑到这件事接下来一定会被各种叙事大书特书，所以这里很有必要从技术层面拆解一下，到底发生了什么。

Now, given how much this is going to be seized in the narrative storytelling, it is important here to get technical in unpacking what really happened.

到了今年，Situational Awareness 的体量已经膨胀得非常巨大。

By this year, Situational Awareness had grown gigantic.

消息人士说，他们大概拿到了约一百亿美元的投资资金，并把这部分资本增长到了大约三百亿美元的 equity。

Sources say that they took in about $10 billion in investment capital and had grown that to around $30 billion in equity.

但是，Aschenbrenner 让这只基金跑在四倍杠杆上，这意味着基金持有的头寸规模大约有一千二百亿美元。

However, Aschenbrenner was running the fund at 4x leverage, meaning that the fund had around $120 billion in positions.

即便放在对冲基金世界里，这也算相当激进了，也让这只基金在面对回撤时暴露得极其严重。

Even in the hedge fund world, that is pretty extreme and left the fund massively exposed to a drawdown.

所谓 leverage，就是用借来的钱，去下一个比你自有资本本来能承受的更大的注。

Leverage means using borrowed money to make a bigger bet than your capital would allow.

所以拿 Situational Awareness 来举例，这只基金有三百亿美元自己的钱，但在四倍杠杆下，它控制的是一千二百亿美元的投资，剩下那九百亿美元本质上就是借来的。

So in the example of Situational Awareness, the fund has $30 billion of its own money, but at 4x leverage, it controls $120 billion of investments, making the other $90 billion effectively borrowed.

因为这些投资规模是基金实际 equity 的四倍，所以市场每动一下，影响都会被放大四倍。

Because the investments are 4 times larger than the fund's actual equity, every market move is amplified fourfold.

投资组合如果波动百分之五，体现在这只三百亿美元基金上，就是百分之二十的收益；但如果下跌百分之二十五，这只基金就会被百分之百抹掉。

A 5% portfolio move shows up as 20% gains on the fund's $30 billion, but a 25% down move makes the fund 100% wiped out.

但问题不只是市场本身在发生什么，更关键的是，站在抵押品的角度，放贷方会提出什么要求。

But the question isn't just what's happening in the markets, but what the lenders expect from a collateral standpoint.

所谓 margin call，就是出借方在说，用来给债务兜底的安全垫变得太薄了，并且要求借款方要么追加更多现金、提高抵押品水平，要么就被迫卖出他们的投资。

A margin call is the lender saying that the cushion backstopping debt has gotten too small, and a demand that the borrower either puts in more cash to increase that collateral or be a forced seller of their investments.

所以继续用 Situational Awareness 这个例子来说，在它那一千两百亿美元的仓位里，其中三百亿美元是自己的权益资本，九百亿美元是加杠杆借来的，而银行这九百亿美元，是靠这个基金那三百亿美元的缓冲垫来保护的。

So continuing with the Situational Awareness example, at their $120 billion position with $30 billion of their own equity and $90 billion borrowed on leverage, the bank's $90 billion is protected by the fund's $30 billion cushion.

现在，如果这些仓位下跌百分之十，整个投资组合就会亏掉一百二十亿美元。

Now, if the positions fall 10%, the portfolio loses $12 billion.

那这个基金现在就只剩下一百八十亿美元的权益，来给银行那九百亿美元贷款提供保护了。

The fund now only has $18 billion of equity protecting the bank's $90 billion loan.

银行这时候可能会觉得这个缓冲太薄，于是要求追加数十亿美元的抵押品。

The banks may decide that that is too thin and demand billions in additional collateral.

如果基金拿不出这笔现金，他们就得卖出仓位，并偿还一部分借款。

If the fund can't provide that cash, they have to sell positions and repay some borrowing.

这里关键的一点是，这会迫使借款方卖掉那些他们其实依然看好的投资，而且很可能是在非常糟糕的价格上卖出，原因仅仅是银行不愿意继续给他们融资了。

Now, critically, this can force the borrower to sell investments that they still believe in, possibly at terrible prices, simply because the bank will not keep financing them.

你大概已经能看出来，这为什么会变成一个恶性循环。

And you can probably see how this becomes a vicious cycle.

仓位失利，银行要求更多抵押品，基金为了筹现金而卖出资产，而这些被迫抛售又会把价格进一步压低，价格越低就带来更多亏损和更多 margin call。

Positions fail, banks demand more collateral, the fund sells to raise cash, those forced sales push prices even lower, prices going lower produces further losses and more margin calls.

你很快就可能亏掉超过自己全部的权益，因为价格下跌的速度，可能比你卖出仓位的速度还要快。

You can very quickly lose more than your equity because prices may move faster than positions can be sold.

比如说，如果那一千两百亿美元的投资组合暴跌百分之三十，就会亏掉三百六十亿美元，比基金原本三百亿美元的权益还多出六十亿美元。

If that $120 billion portfolio plunges 30%, for example, it loses $36 billion, $6 billion more than the fund's $30 billion in equity.

那样一来，银行就会面临基金已经拿不出来、但仍然欠着的钱。

The banks would then be owed money that the fund no longer has.

所以在实际操作中，银行会更早就尝试清算，目的就是防止出现这种结果。

So in practice, they try to liquidate much earlier precisely to prevent that outcome.

这也就是为什么你会看到这个故事发展得这么快。

And this is why you see this story happening so quickly.

这不是什么拖了整整一个月的漫长回撤。

This wasn't some protracted month-long drawdown.

Leopold 重点押注领域里的那些股票，跌幅已经大到让出借方开始要求平仓这些仓位了。

Stocks in Leopold's key area of bets had fallen by enough that lenders started calling in the positions.

而且就像以前已经发生过很多次那样，Leopold 没办法找到足够多的投资人来填上这个缺口，于是成了被迫卖出的一方。

And as has happened so many times before, Leopold couldn't find enough investors to cover the difference and became a forced seller.

现在，虽然从某些角度看，这件事很像是最基础的 Markets 101，但也有人认为，市场里其他参与者确实很可能在推动这个结果发生。

Now, while in some ways this has the look of Markets 101, there are some who believe that it's realistic to think that others in the market were pushing for this outcome.

SEC 注册基金是必须披露自己仓位的，所以所有人都知道 Situational Awareness 在 AI 这笔交易上做了巨额多头。

SEC-registered funds are required to report their positions, so everyone knew Situational Awareness was massively long the AI trade.

而且也知道，哪些具体股票会对他们伤害最大。

And knew the specific stocks that would hurt them most.

在接受 TBPN 采访时，Martin Shkreli 就专门谈到了这件事到底是怎么运作的。

In an interview with TBPN, Martin Shkreli discussed exactly how this works.

稍微意译一下，他的意思是，我觉得这周早些时候就已经有一些玩家在提前布局，想做他们所说的 shooting against a fund。

Paraphrasing just a little bit, he said, I think some players were already positioning earlier in the week, looking to do what they call shooting against a fund.

如果你知道某个人一定得清仓，那对你来说最好的做法，就是先把你们共同持有的仓位全都卖掉，然后再开始做空他们持有的所有东西。

If you know somebody has to liquidate, the best thing for you to do is sell all the positions you have in common and then start shorting everything they have.

这会加速它的崩盘。

It accelerates the downfall.

这非常常见，也很可悲，而且非常 Darwinian。

Very common and sadly very Darwinian.

如果你相信这至少是故事的一部分，那这又一次说明，我们看到的 AI 概念股走弱，背后推动因素可能和 AI 基本面一样多，甚至更多是市场结构本身的问题——在这个案例里，还包括市场里的博弈和算计。

Now, if one believes that this is at least part of the story, it is again another way in which weakness that we've seen in AI names may be driven as much by the structure of markets, and in this case gamesmanship in the markets, as opposed to AI fundamentals.

接下来会发生什么呢，据传最终接手 Situational Awareness 投资组合的买家 Citadel，是以比原仓位低百分之二十到百分之五十的折扣拿下这些资产的。

Now, in terms of what happens next, the ultimate buyer of the Situational Awareness portfolio, Citadel, is rumored to have gotten a 20 to 50% discount on the positions.

这就意味着，他们其实并没有特别大的压力，要马上把这些东西重新卖回市场。

That means there's not really a huge rush to sell it into the market.

当然，世界上没有哪个基金能无限期持有一千两百亿美元的风险敞口，但 Citadel 肯定算得上是最有能力慢慢把它卸掉的机构之一。

Now, no fund in the world has the ability to hold $120 billion in exposure indefinitely, but Citadel is certainly among the best places to sell it off slowly.

而且 Citadel 也不是那种纯粹押方向的玩家。

Citadel also isn't a directional player.

他们靠频繁交易赚钱，而且会一直尽量把市场仓位维持在中性。

They make money by trading often and always try to maintain a neutral market position.

这就意味着，他们本来就对这次回撤做了自然对冲，所以大概率不像 Situational Awareness 那样，处在同样脆弱的位置。

That means that they were naturally hedged against this drawdown and likely aren't in the same vulnerable position as Situational Awareness was.

另外，也很有必要把这次爆仓跟其他一些著名的 hedge fund 失败案例对比一下。

Now, it is also worth contrasting this blowup with other famous hedge fund failures.

1998 年 LTCM 崩盘的时候，他们是在用极高杠杆交易货币和债券。

When LTCM collapsed in 1998, they were trading currency and bonds on massive leverage.

那件事之所以会演变成亚洲金融危机，不是因为有一家 hedge fund 爆了，而是因为外汇市场根本吸收不了他们的清算。

The reason that metastasized into the Asian financial crisis was not because a hedge fund blew up, but because the currency market couldn't absorb their liquidation.

2008 年 Bear Stearns 的崩溃也是类似的问题。

The Bear Stearns collapse in 2008 was a similar issue.

他们交易的是次级贷款 CDO，但真正导致风险传染的核心原因，是他们的公司债在整个 Wall Street 都被拿来当抵押品。

They were trading subprime CDOs, but the core reason that caused contagion was because their corporate debt was being used as collateral across Wall Street.

这两种情况下，这些基金都碰到了金融系统里很关键的部分，所以它们的清算才会演变成系统性冲击。

In both cases, the funds were touching important parts of the financial system, which allowed their liquidation to become a systemic shock.

而这次爆仓，看起来更像 2021 年 Archegos 的失败。除非你在金融行业，或者你当时也正好在做另一档日更播客，不然你大概率都没听过。

This blowup, on the other hand, looks more like the Archegos failure in 2021, which you've likely not heard of unless you're in the financial industry or were also doing a different daily podcast back then.

那次是一百亿美元级别的爆仓，所以没有 Citadel 这次这么大，但相似之处在于，它也是在用巨额杠杆交易科技股。

That was a $10 billion blowup, so not quite as large as Citadel, but it was similar in that it was trading tech stocks on massive leverage.

那次爆仓确实很惨，而且很可能也推动了 2022 年的熊市，但它并没有引发系统性危机。

The blowup was painful and likely contributed to the bear market of 2022, but it didn't cause a systemic crisis.

不过它还是有下游影响，其中一家给他们做 prime broker 的 Credit Suisse，因为吃下了巨大亏损，后来不得不并入 UBS。

It did have downstream effects, with one of their prime brokers, Credit Suisse, having to merge into UBS after taking a massive loss.

这次不一样的一个重要原因，是 Citadel 这笔收购进行得非常顺利。

One of the big reasons that this time is different is how smoothly the Citadel purchase went.

从目前各方面的信息来看，这个仓位听起来现在还没有亏到水面以下。

By all accounts, it sounds like the position is still above water.

而且 Citadel 很可能也很有条件把这事扛过去。

And Citadel is likely well-placed to ride it out.

Ken Griffin 看起来像是接过了当年 Warren Buffett 在 2008 年、或者更早期 JPMorgan 在 20 世纪初扮演的那个角色：接手困境资产，吸收更大事故的风险。

Ken Griffin has seemingly stepped into the role that Warren Buffett played in 2008, or that JPMorgan played in the early 1900s, taking over distressed assets and absorbing the risk of a larger incident.

说白了，Wall Street 现在在这种私下接盘的操作上，已经熟练太多了。

Frankly, Wall Street has just become much better at these private market takeovers.

市场是在周三第一次公开传出他们陷入困境的消息，而到周四早上，交易就已经敲定了。

The distress was first publicly reported on Wednesday, and by Thursday morning, the deal was done.

归根结底，金融危机从来都不是单纯因为股票回撤。

Ultimately, financial crises are never about equity drawdowns by themselves.

真正的大问题，是违约开始沿着抵押品体系层层传导、不断蔓延的时候。

It's when defaults start cascading through the collateral system that there's a major problem.

也就是说，股市暴跌当然不好受，但它本质上跟全面性的金融危机是完全不同的两回事。

I.e., stock market crashes are not fun, but they are fundamentally a very different thing to a full-on financial crisis.

而从市场接下来怎么反应的早期迹象来看，这些判断某种程度上也得到了印证。

Kind of confirming all that is the early indications we have around how the market reacts from here.

现在清算已经结束了，做空这些股票也就没有那种机械性的驱动了。

Now that the liquidation is over, there's no mechanical incentive to short the stocks.

有些机构也许会想试着给 Citadel 施压测试一下，但过去所有这么干的人，结果都很惨。

Some firms could decide to pressure test Citadel, but that's gone very poorly for everyone who's tried in the past.

他们实在是太大了。

They're just far too big.

这周早些时候，就已经有一些分析师在说市场见底了。

Earlier in the week, some analysts were already calling the bottom.

JPMorgan 在周一写道，市场已经在发出买入信号，准备开始反弹。

JPMorgan wrote on Monday that the market was flashing buy signals and was ready to rally.

而现在随着局面解决，市场也确实强劲反弹了。

And the market has bounced hard now that the situation is resolved.

Nasdaq 在周四上涨了百分之二点八，这是它一个月以来表现最强的交易日之一。

The Nasdaq was up 2.8% on Thursday, one of its strongest days in a month.

Situational Awareness 重仓的韩国市场，现在涨势非常猛。

The Korean market, where Situational Awareness had heavy positions, is absolutely ripping.

KOSPI Index 在周五收盘时上涨了百分之十五，盘中早些时候一度涨到百分之十七。

The KOSPI Index ended the Friday session up 15% after cruising up as much as 17% earlier in the day.

当然了，现阶段这还只是一次重大事件之后的技术性反弹，不过，Situational Awareness 的爆雷，完全有可能真的标志着这轮 AI 回调在局部范围内已经见底。

Now, of course, at this stage, this is just a relief rally after a major event, but it is entirely possible that the blowup of Situational Awareness could have actually marked a local bottom for the AI drawdown.

有一句很有名的话，你以前可能听过某个版本。

There's a famous phrase that you might have heard some version of before.

你可能对政治不感兴趣，但政治对你感兴趣。

You may not be interested in politics, but politics is interested in you.

对于我们很多关注 AI Daily Brief 的人来说，我对 AI 市场差不多就是这种感觉。

That's kind of how I feel about AI markets for many of us who are here around the AI Daily Brief.

当然，你们当中有些人是专业投资者，对你们来说，这些事情显然都非常关键，也非常重要。

Certainly some of you are professional investors for whom this is all obviously very mission critical and important.

但对于那些只是想弄明白 AI 会怎样影响自己的一般人，我的观点是，到现在这个阶段，AI 已经和经济中的很多部分，以及当前经济结构中的很大一部分，紧密地绑在了一起。所以，就算你本身并不是以投资为主，了解现在发生了什么，也还是很值得。

For the average folks who are just trying to understand how AI is going to affect them, my argument is that at this point, AI is so integrally tied to so many parts of the economy and so much of the current structure of the economy that even if you are not primarily an investor, it is worth understanding what's going on.

如果现在要把这一切总结一下，那故事大概就是：第一，市场对循环融资的持续担忧，以及更广义上对支撑这轮 AI 建设背后债务性质的担忧。正如我之前说过的，这些其实是判断 AI 泡沫会不会真正彻底成形时，最重要的一些泄压阀。

And right now, if we have to sum this all up, the story is, one, continued concerns around circular financing and just the nature of the debt in general behind the AI buildout, which, as I've said in the past, are some of the best pressure release valves when it comes to whether an AI bubble would fully form.

一家非常知名的 hedge fund 爆雷了，而主要原因并不是 AI 基本面，更多还是老生常谈、而且屡试不爽的杠杆问题。

A very notable hedge fund that blew up, not mostly because of AI fundamentals, but instead because of tried and true issues of leverage.

但在这一切背后，AI 的需求故事还在继续，而且只是在不断增长。

But behind it all, a demand story for AI, which does nothing but continue to grow.

希望你现在对外面到底发生了什么，已经有了更清楚的感觉。

Hopefully you feel like you have a better sense of what's going on out there now.

好，那今天的 AI Daily Brief 就先到这里。

For now, that's going to do it for today's AI Daily Brief.

一如既往，感谢你的收听或收看。

Appreciate you listening or watching as always.

那我们下次见，peace。

And until next time, peace.

已剔除 3 处广告（点击展开查看）

M10:20广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Retool, and Airtable.

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

To learn more about sponsoring the show, send us a note at [email protected].

M10:42广告 · 已剔除

And second, your reminder to come check out the AI Summer Adventure.

You can find it at summeradventure.ai.

It's a choose-your-own-adventure-style program where you can do projects at basically any level of AI learning.

Go check it out.

M114:45广告 · 已剔除

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising.

The highest impact users aren't better prompt engineers.

They treat AI like a reasoning partner.

They frame problems, guide thinking, iterate, and push for better answers.

And the good news?

These behaviors are teachable at scale.

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

Learn more at kpmg.com/us/sophisticated.

That's kpmg.com/us/sophisticated.

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

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

Greenfield is supposed to be the easy part.

Clean slate, no technical debt.

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

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

Hundreds of thousands of lines of production-ready code.

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.

Another shipped an entire application with no frontend engineer.

Legacy or greenfield, the answer is the same: software at the speed of compute.

Build what's next at blitzy.com.

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

This episode is brought to you by Retool.

Generating a working app now takes about 5 minutes thanks to AI.

Getting it safely into production—with auth, permissions, audit logs, security reviews—that part still takes time.

That's the gap Retool closes.

Build apps however you want: natively in Retool, with Claude Code, Codex, or any coding agent, or by importing React you've already built.

It all deploys into Retool and picks up governance automatically.

Security lives in the platform, not in whatever the AI wrote.

So your team ships at AI speed without the shadow IT and the endless reviews that stall out most vibe-coded projects.

It's why teams at Amazon, Stripe, and Brex build on Retool.

And new enterprise customers who sign by September 30th get up to $10,000 in AI credits per year.

Start building at retool.com/AIDaily.

This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

New users get $1,000 in inference.

Forget local agents and chat workflows waiting on your laptop to be prompted.

HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing's agent turns competitor moves into landing pages.

Sales' agent enriches leads, drafts emails, and updates the CRM.

Ops' agent chases the paperwork and tracks the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you add agents that feel like teammates.

Hire yours at HyperAgent, built by the team at Airtable.

Claim your $1,000 in inference at hyperagent.com/AIDailyBrief.