如果人工智能失败了会怎样？次级数据中心危机，SpaceX还能有多糟？

如果AI的建设出现延误，整个经济会不会跟着受影响？

If the AI build out stumbles, does the rest of the economy go with it?

数据中心的扩张，是不是一颗定时炸弹？

Is the data center build out a ticking time bomb?

SpaceX和特斯拉会不会合并？

And will SpaceX and Tesla merge?

接下来，在周五的科技深度播客里，我们会全面聊聊这些话题。

We'll cover it all on a big technology podcast Friday edition right after this.

面对持续的变革与机遇，TMT领域的领导者必须拿出实际成果，而不仅仅是空谈想法。

In the face of ongoing disruption and opportunity, TMT leaders need to deliver tangible results, not just ideas.

在速度和表现至关重要的时候，普华永道结合市场洞察、行业专长，以及AI、云计算和前沿技术，帮你加速转型，从战略到执行，实现可衡量的投资回报。

When pace and performance matter most, PwC combines market insights and deep sector experience with AI, cloud, and emerging tech to accelerate your transformation and drive measurable ROI from strategy to execution.

普华永道能帮你预判未来，赶在变革之前行动，赢得竞争。

PwC can help you anticipate what's next, outpace disruption, and compete.

想了解更多信息，请访问pwc.com。

For more information, visit pwc.com.

欢迎收看《大科技》周五特别节目，我们将以一贯冷静、细致的方式为您解析最新动态。

Welcome to Big Technology podcast Friday edition where we break down the news in our traditional cool headed and nuanced format.

今天这期节目内容非常精彩。

We have a great show for you today.

我们会深入探讨，如果人工智能失败了，哪怕只是其中某个关键环节出问题，会带来怎样的连锁影响。

We're gonna talk all about what would happen if AI fails, if even one leg on the AI stool starts to crumble, what the implications will be.

我们还会讨论，数据中心会不会陷入类似次贷危机的局面，SpaceX的处境到底有多严峻，以及它和特斯拉合并的计划眼下是否真的可行。

We're gonna talk about whether we're heading towards a subprime data center crisis and how bad SpaceX can get, and whether its plans to merge with Tesla are currently even feasible.

和往常一样，周五我们邀请到Margins的Ranjan Roy和我们一起分析。

Joining us as always on Friday to do it is Ranjan Roy of Margins.

Ranjan，很高兴见到你。

Ranjan, good to see you.

欢迎回来。

Welcome back.

我也很高兴见到你，Alex。

Good to see you, Alex.

这周我整理节目时，感觉比平时低落了一些，说不准是自己的情绪影响了节目的基调，还是最近的世事真的让咱们对AI的讨论变得不一样了。

This week, I have to say, as I was putting together the show, I definitely felt a little bit more downtrodden than usual, and I can't tell whether that's my mood influencing the way the show is being put together or the world events influencing how we're gonna talk about AI this week.

我觉得应该是后者。

I think probably the second.

其他方面都挺好的。

Everything else is going well.

外面天气真美。

It's beautiful outside.

现在正是盛夏时节。

We're full swing in summer.

你我都快去度假了，外面一切都那么惬意，可AI圈里却隐隐透着不安，甚至可能要出问题。

You and I are both about to take some vacation and and right now, you know, things couldn't be better on the outside, but on the inside in AI, things look kind of rough or potentially rough.

说到AI这件事，我们总是在技术前景光明和经济前景不明之间来回摇摆。

Because I think with this AI story, we always go back and forth between the technology being very promising, which it is, and the economic story being somewhat in question.

确实值得怀疑。

Question.

这一点一直萦绕在我们心头，我相信听众们也一样。

That's always been in the back of our minds and I'm sure of our listeners' minds.

这周，仿佛到处都能看到警示和担忧的迹象。

And this week, seemingly everywhere, there were certainly signs for caution and concern.

首先，我想提一下《纽约时报》的一篇文章，它问：如果AI热潮降温，经济会怎样？

So let me, to begin with, just cite this New York Times story that asks what will happen to the economy if the AI boom starts to tamp down.

文章指出，如果投资者对AI的信心动摇，那些完全建立在AI之上的经济体系可能会彻底崩塌。

The story says if investor confidence in AI falters, the economy actively built atop it could come crashing down.

今年标普500指数的涨幅中，相关股票贡献了大约一半。

Related stocks account for roughly half of the rise in the S and P 500 this year.

股价下跌，哪怕跌得很厉害，通常也只影响金融圈，不会波及更广。

Declines in stock prices, even drastic ones, don't necessarily have much impact beyond the world of finance.

但这次可能不一样，因为股市的规模实在太大了。

But what could make this time different is the sheer scale of the stock market.

经济研究发现，投资者每在股票上赚到100美元，就会多花大约3美元在商品和服务上，这就是所谓的财富效应。

Economic research has found that for every $100 investors gain in their stock portfolios, they spend about $3 more on goods and services, a phenomenon known as the wealth effect.

但财富效应也会反过来起作用。

But the wealth effect also operates in reverse.

当股价下跌时，投资者就更不愿意花钱了。以当前的估值来看，股市下跌30%可能导致消费者支出减少近7000亿美元。

When stock prices fall, investors become less willing to spend and at present valuation a 30% decline in the stock market could lead to a nearly $700,000,000,000 pullback in consumer spending.

光是这一项，就足以引发一场经济衰退。

That could be enough to set off a recession on its own.

所以，说白了，这些事目前还没发生，但文章认为，如果AI股票的热潮退去，可能会带来实实在在的问题，甚至引发衰退——因为人们现在花的钱，很大程度上是因为看到投资组合在增值，而这些增值的很大一部分，归根结底都跟AI股票脱不了干系。

So basically, you know, obviously none of this has begun yet, but the the story argues that if we were to see a pullback on the AI stock bonanza that could leave to lead to real world problems potentially even a recession because people are spending so much because they see their portfolios rise and a large part of the reason why their portfolios are rising are AI stocks one way or another.

拉詹，你觉得这是个合理的担忧吗？你对AI泡沫破裂可能带来的负面影响怎么看？

Legitimate area of concern, Ranjan, or what do you think about the the potential negative effects of an AI bust here?

我觉得，从财富效应的角度来看，要准确归因其实一直挺难的。

Well, I think when we examine it in terms of the wealth effect, I think it's always a little difficult of what you can actually directly attribute.

不过，也许我们最近都感觉有点低落，因为世界杯结束了。

I will say maybe we're all feeling a bit down because the World Cup is over.

但我也想说，任何在那期间试图买过球票的人都能切身体会到财富效应——普通人在买票时花两千、三千美元，居然都觉得理所当然。

But I would also say anyone who tried to get tickets during that could feel the wealth effect just front and center when people normal people you're talking to were spending two grand, $3 on a ticket, and it just seemed normal.

所以我觉得，尤其是在各种奢侈消费方面，我们确实亲眼看到了财富效应的存在，而且非常直观。

So I think like in all kinds of especially luxury spending, we've seen the wealth effect, you know, like live and very present.

所以我觉得，这其中有多少能真正归因于人们账面上的收益，这还挺有意思的。

So I think how much of that can actually be attributed to people's paper gains is it's interesting.

要真的做出直接的相关性分析总是挺难的，但我确实觉得，我是说，尤其是最近这几天，出现了一些回调。

It's always difficult to actually do, like, you know, a direct correlation there, but I do think, like, I mean, some kind of pullback, especially in the last few days.

像‘七巨头’那几只股票，我觉得昨天——我们今天是周五录制的——

Mag seven, I think yesterday, we're recording on Friday here.

周四那天，七巨头股票出现了五年来最大的整体跌幅，我记得当时看到的数据是这样的。

On Thursday, it was the largest aggregate decline in Mag seven in five years, I believe I saw it was.

而且这还没算上SpaceX，我们待会儿会聊到它。

So this stuff and that's not counting SpaceX, which we can we're gonna talk about later.

所以我觉得这确实是会发生的现实情况。

So I think it's going to be real.

对我来说，总觉得这种事迟早会发生。

To me, it's always like it was going to happen at some point.

唯一的问题是，这会对整体经济造成剧烈的负面影响，还是说只是让世界杯门票从两万美元降到一万美元而已？

The only question is, is this gonna be, like, a dramatic negative effect on the economy overall, or does it just mean if the World Cup was today, you could get tickets for, like, a grand instead of $2?

没错。

Right.

有些消费实在太离谱了，哪怕稍微回落一点，也能让局面回归正常。

Some of the spending is so outrageous that, you know, even a small pullback will sort of bring it back to normalcy.

对吧？

Right?

消费已经疯狂到极致了，但你也能看到，就像文章提到的，有些支出其实并不属于极端奢侈，对吧？

Like things have gone crazy to the extreme, but you could also see some of the spending, you know, as the article indicates, in non extreme luxury purchases, right?

人们出去吃饭更频繁了，旅行也更多了，很可能是因为他们的投资组合表现太好了。

People are going out to dinner more often, they're going on trips more often, likely because their portfolios are doing really well, right?

我们连续两年标普指数涨了220%以上，今年才刚过一半，标普又涨了8%左右。

We're basically on the back of 220% plus gains in the S and P years in a row and then this year we're a little bit more than halfway through and the S and P is up about 8%.

所以人们资产的增长推动了更多消费，而市场集中度又都在人工智能领域。

So the increases in people's portfolios are leading to more spending and the concentration in the market is in AI, right?

所以有25%集中在那七家科技巨头，现在还有一些内存公司也在推高人们的资产价值。

So 25% is in the Mag-seven and then there's now these additional memory companies that are also pushing up the value of people's portfolios.

这篇文章就是这么分析的，说这是可能引发连锁反应的一种情况，你觉得这种可能性大吗？

So this is how the article argues that, you know, one scenario for how things might end up unraveling and you let me know if you think this is feasible.

他们认为，如果企业发现投资回报没预期的那么快，就可能削减开支，迫使AI实验室和供应商下调增长预期。

They say if companies find that their investments aren't paying off as quickly as they hope, they might pull back their spending, forcing the AI labs and their suppliers to trim their growth projections.

这种落差一旦出现，可能会引发市场抛售，让企业更难、更贵地融到发展AI所需的资本。

Such a disappointment could incite a market sell off, which would make it more difficult or more expensive for companies to raise capital they need to fund the AI buildout.

这样一来，企业就可能推迟或取消建设数据中心、发电厂及相关基础设施的计划，进而导致建筑业裁员。

That in turn could lead companies to delay or cancel plans to build data centers, power plants, and related infrastructure, giving way to layoffs in the construction industry.

与此同时，市场下跌会让高净值人群缩减消费，进一步引发更多岗位流失，最终拖累经济陷入衰退。

And at the same time, the drop in the market would push wealthy consumers to pair their spendings, their spending leading to wider job losses and ultimately a recession.

你怎么看？

What do you think?

你觉得这合理吗？

Plausible?

我觉得这确实很有可能。

I think certainly plausible.

我的意思是，这跟任何一次经济下滑都差不多。

I mean, it's like any downturn.

真正致命的不是最初的下滑。

The it's it's not the initial downturn that's gonna kill you.

而是它引发的连锁反应，也就是第二、第三波的影响。

It's the secondary and second and third order effects of it.

不过我也不太确定。

But I I don't know.

我还是觉得，我们最近看到的股市表现，特别是这周的情况，到底只是轻微的回调，还是意味着更深层的问题。

I still feel what we've seen in the market, especially this week, and whether that's just kind of like a slight correction or whether it actually signifies something much deeper.

接下来我们会聊聊谷歌的支出和自由现金流，我觉得这点其实挺值得注意的。

And we're gonna get into Google spending and free cash flow and which I think is actually really notable.

把它放在整个AI讨论的时间框架里来看，我一直说，时间跨度才是整件事最有趣的地方。因为如果谷歌现在投资是为了两三年后的回报，那我们可能根本看不到什么效果，甚至可能要等上一年。

To to me, time frame, it with with the whole AI discussion, I've said this regularly, like, the time frame is the most interesting part of the entire thing because, you know, if Google is investing now for two to three years down the road and you're really not gonna see anything or maybe even a year down the road.

但在短期内，这会带来更多问题。

But in the near term, it causes more issues.

这种影响在近期内——比如明天和下个月——会有多大？

How much of that affects spending in the near like, tomorrow and next month?

我觉得这种影响肯定是存在的。

I think it's definitely gonna be there.

但人们是否还相信这个中期前景，可能正是区分我们看到的只是表面波动，还是真正更深层、更严重问题的关键。

But whether people continue to believe in that medium term story, I think could be the difference whether or not we're actually seeing something much more systemic and worse.

是的。

Yeah.

这个故事最后有一句话说得特别好，说的是：目前，怀疑论者需要提供证据；但一旦开始出现令人失望的信息，哪怕只是缓慢地 trickle 下来，责任就转到乐观派身上了。

There's a great line that ends this story, which is a quote, right now the burden of proof is on the skeptics, but once you have the slow trickle of disappointing information, then the burden starts to be on the optimists.

这话说得太妙了。

Oh, that's beautiful.

对，确实很精彩。

Yeah, is beautiful.

是的。

Yeah.

没错，毕竟到现在为止，大家都相信多花钱最终是值得的，因为AGI迟早会带来回报。

Right, because ultimately, like everyone has gone along, you know, with the story of more spending will inevitably be worth it because AGI.

对吧？

Right?

这基本上就是一直以来的逻辑，但如果像谷歌、微软、Meta这样的大公司突然觉得，也许我们没必要再砸那么多钱建基础设施，也能从其他渠道获得AI的好处，这个逻辑可能很快就会崩塌。

Like that's basically been the logic and this could really unravel very quickly if the big tech companies like Google, Microsoft, like Meta determine, hey, maybe we don't need to, you know, make these massive infrastructure spends and we can have some of the benefits of AI from from, you know, other sources.

你觉得现在我们看到的，是不是Kimi K3对七巨头产生的影响？

Do you think this is we're seeing the Kimmy k three effect here on mag seven?

你觉得人们是不是已经直接把这两件事联系起来了？这样一来，更高效的支出可能让杰文斯悖论来得慢一点。

Do you think people have already kind of directly connected the two and suddenly more efficient spend, maybe Jevan's paradox will take a little longer for us to get to.

所以现在大家都觉得，其实现在能用更便宜、更高效的方式完成这些事了。

So so everyone's like, actually, you will be able to do things more efficiently and cheaply now.

所以那种大规模资本支出的逻辑，就没那么站得住脚了。

So the that whole capex story doesn't quite make as much sense.

还是你觉得这只是涨得太多，现在涨势放缓了？

Or do you think this is just a little bit of a it's gone up a lot and it's not going up as much?

是的。

Yeah.

我不觉得Kimi K3能直接让人觉得能更省钱。

I I don't think Kimi k three I think I don't think that is, you know, a direct lead into people saying you can do it more cheaply.

我们上周其实聊过这个，但Kimi K3其实并不怎么节省token。

Again, if we spoke about this a little bit last week, but Kimi k three actually is is not very token efficient.

就算价格低了，你还是得建数据中心、租云服务，跑一个特别耗token的模型，未必能省下你预期的那些成本。

So even if the prices are cheaper, you still need the data centers, you still need the cloud hosting and you're going to run a very token hungry model and not necessarily get the savings that you expected.

但我确实觉得，这是个累积效应——模型越来越便宜，而且证明了OpenAI和Anthropic之外的模型也能行。

But I do think that this is a cumulative situation of the models getting cheaper, people showing that models outside of OpenAI and Anthropic can work.

OpenAI和Anthropic至少在他们的预测里，离长期盈利还很远。

OpenAI and Anthropic not really being close, at least in their projections, to long term profitability.

像Meta和Grok这样的模型越来越便宜，但真正的收益其实并没有出现在OpenAI和Anthropic之外。

More and more models like Meta and Grok being cheaper and effectively the gains not really being found outside of OpenAI and Anthropic.

所以这两家公司，我们之前一直说它们是整个行业里的双重短板，如果你不是其中之一，现在确实开始让人担心了——你到底怎么靠这个赚钱？

So those two companies which we've talked about as the sort of dual points of failure in this whole thing, I think that that is starting to be a cause of concern if you're not one of them in terms of how are you going make money off of this?

另外，你刚才提到谷歌的财报，对吧？

And then I think, by the way, you mentioned Google earnings, right?

这正好引出了谷歌的一个大问题。

That leads right into this big problem with Google.

这周谷歌股价跌了将近8%，财报发布后立刻跳水，原因是它的资本支出从原先的数字飙升到了2050亿美元，用于AI基础设施建设。

So Google this week is down like 8% and it dropped immediately after its earnings after its capex went from went up to two zero five more than 200,000,000,000 in terms of its AI infrastructure spend.

对吧？

Right?

根据《华尔街日报》的报道，这家市值4万亿美元的公司在公布第二季度财报时，将预计资本支出上调至一个更高的区间，结果投资者用脚投票，股价盘后一度下跌超过4%，后来才小幅回升。

So this is from the Wall Street Journal after the $4,000,000,000,000 company raised its estimated capital expenditures to a new range that extends past figures while reporting on its second quarter earnings, investors punished the stock which fell more than 4% and after hours trading before recovering slightly.

说白了，分析师们认为，一年2000亿美元这个数字，是绝对不能突破的红线。

So basically what's going on is, you know, the analysts are saying this 200,000,000,000 in a year was a do not cross line.

之前大家预期的数字是1800亿到1900亿美元，对吧？

Previously it was supposed to be 180,000,000,000, 190,000,000,000, right?

现在它的预算范围变成了1950亿到2050亿美元，回到《纽约时报》提到的那个情景，大家都在问：这投入到底什么时候能有回报？

And now it's going to be 195,000,000,000 to $2.00 $5,000,000,000 and you know going back to that New York Times scenario, people are asking, Hey, where's the payoff here?

对吧？

Right?

我们期待的那些增长到底在哪？

Like, where is the associated growth that we're supposed to see?

我觉得谷歌之所以成了关键问题，是因为它不像OpenAI和Anthropic那样取得模型上的突破，也没像Cloud Code那样展现出明显的增长。

And Google became this linchpin, I think, because it's not having the model success as OpenAI and Anthropic, it's not having the growth that we're seeing in something like Cloud Code.

现在市场终于在问：如果你不拿出实际成果，光靠乐观预期砸钱是行不通的。

And these questions are finally market saying, listen, you can't just spend based off of this optimistic scenario if you don't show us the results.

谷歌的云业务确实在增长，但我们已经开始看到反噬了。就像我说的，如果市场不买账，这种情况可能很快就会崩盘，你觉得呢？

Now Cloud is growing for Google, but it's we're beginning to see the pushback and this, like I said, this could unravel very quickly, don't you think, if if the market does not allow for it?

是的。

Yeah.

但我还是不太喜欢把这说成是‘崩盘’这种说法。

But I it's still I don't love this idea that, like, looking at it as unraveling.

我觉得，现在责任开始转移到乐观派身上，而不是怀疑者了。

I guess, going back to the idea that the burden starts to be on the optimist rather than the skeptics.

让我觉得不可思议的是，2025年9月，市场居然奖励那些说要花更多钱在还没实现经济回报的技术上的人。

Like, it was crazy to me, September 2025, where you were rewarded by the market for saying you're going to spend a lot more money on a technology that has not had the economics of it actually worked out yet.

Oracle的股价从那年9月的高点已经跌了65%。

Again, Oracle is down 65% from the peak up in back in September.

在任何市场环境下，如果你说要砸大钱在没验证过的东西上，还能得到奖励，这本身就有点离谱。

That was when like, any market environment where you are rewarded for saying I'm gonna spend a ton more money on something that's unproven is already a bit crazy to me.

所以我还是觉得，这更多是市场在自我合理化，而不是正在崩盘。

So I'm still gonna look at all of this as rational like, the rationalization of the markets rather than the unraveling of the markets so far.

我是说，确实如此。

I mean Right.

咱们再等一周看看吧，不过

Let's wait a week but

是的。

Yeah.

但这个情况有点不一样。

But but this okay.

那理性预期会不会最终变成崩溃呢？

So can't rationalization lead to unraveling?

我觉得迟早会有一个节点，投资者会重新审视这项新兴技术——到目前为止，他们愿意押注，就像我们前面说的，大家投资组合上涨，根本原因就是大家对AI、对AI故事有共同的信念。

Like, there's gonna be a point, don't you think, where investors are gonna look at this, you know, this emerging technology and so far they've been willing to bet on it and again, like going back to the story at the top here, the reason why people's portfolios are up is because there's been this collective belief in AI, in the AI story.

一开始是作为一项技术故事，后来又像硅谷一贯的做法那样，蔓延到整个经济，人们还相信他们最终能找到盈利模式。

First as a technology story, but then in like traditional and typical Silicon Valley mode of spreading to the whole economy, there's also been this belief that they'll figure out the business model.

我觉得现在市场传递的信息是：也许他们根本找不到盈利模式。

I think what we're seeing with the market right now is, hey, maybe they won't figure out the business model.

距离ChatGPT发布已经四年了，花钱的势头却丝毫没有减弱。

We're four years past CHET GPT, the spending continues seemingly unimpeded.

盈利模式可能正在浮现，但我也不确定。

The business models maybe are emerging, I don't know.

目前最靠谱的盈利模式，其实就是为这些技术提供基础设施。

The best business model is just providing the infrastructure for these things.

产品确实是在进步，这点很明显，但并不是所有方面都在提升。

And the products are are, you know, obviously improving, but not across the board.

目前看来，只有 OpenAI 和 Anthropic 的产品势头比较猛。接下来我想引用 Ranjan Roy 先生本周发的一条关于 Google 产品的推文。

And the only ones that seem to have the product momentum are OpenAI and Anthropic and I'll go to a tweet from Mr.

他说 Gemini 正在迅速成为类似 Copilot 那样的存在。

Ranjan Roy this week about the Google product.

Gemini 正快速确立其作为副驾驶（Copilot）的地位。

Gemini is quickly achieving co pilot status.

那边到底在搞什么，完全摸不着头脑。

No idea what's going on over there.

所以我的意思大概是，市场这张空白支票可能得面临个抉择：要么继续投，要么干脆别投了，中间好像没那么多妥协的余地。

So basically my point is maybe this blank check from the market goes from, you know, hey we're gonna you know it doesn't seem like there's gonna be this middle ground between blank check and you can't spend on AI anymore.

我就是觉得，他们不可能说只允许你花五百亿美元就完事了，对吧？

I just don't see them saying you can spend $50,000,000,000 right?

我感觉要么就全押上，去搞通用人工智能，那样能赚大钱；要么就不搞。

I feel like it's either, you know, you go all in and you're gonna build AGI and that'll be worth a ton of money or you don't.

我看不出有什么中间地带。

I don't see the

中间地带？

middle ground.

你看到中间地带了吗？

You see the middle ground?

没有，亚历克斯。

No, Alex.

我喜欢健康的市场。

I like my markets healthy.

我不知道你怎么想，但我喜欢市场保持健康，我觉得这其实就是一种‘空白支票’，而且一直以来都是这样。

I don't know about you, but I like my markets healthy and I think this is just a little bit of again, you you you use the word correctly blank check, and that's what it's been for so long.

尤其是过去这一年多，出现了太多这样的时刻，比如今年二月，高管们还在炫耀自己在AI上的巨额开支。

And, like, again, in the last, call it twelve months more than anything, there have just been so many of these little moments where, again, back in February, like senior executives bragging about their clawed code spend.

这些时刻回头一看，你就会觉得当时有点不对劲。

Those are the moments that you always look back on and you're like, something was a little bit off.

市场竟然奖励甲骨文，因为它们背上了大量债务，还跟OpenAI做了这么多疯狂的承诺，这些交易都建立在极其乐观、甚至不切实际的假设上，结果还得到了回报。

The market rewarding Oracle for taking on a bunch of debt to and all these crazy commits from OpenAI and these deals being structured around like these very optimistic, if not crazy scenarios and being rewarded for it.

所以我觉得，这对市场来说是好事，这种局面迟早得来。

So so I still think this is good for the market and the this had to happen.

而且，与其让事情失控，不如现在这样有节制地发生。像我们常听节目的人应该知道，我对代理式AI的中长期前景非常乐观。

And it it like, it's better that this happens in a kinda controlled way where people start and again, I as regular listeners know, I'm very optimistic about at least the medium term and the world of agentic AI.

我觉得，与其让市场一路直线上升，直到某天突然崩盘，不如现在就适度调整一下。

And I think, like, it's better that this happens in some way rather than it's just a straight line up forever until it's not.

那投资者们是怎么给自己讲一个合理的理由，说我们愿意让这些大科技公司每年花上一千亿美元呢？

So what is the healthy market story that investors tell themselves to be like, we're we are gonna spend like, we're gonna give these big tech companies the leeway to spend, let's say, 100,000,000,000 a year.

与其把所有自由现金流都花光，不如拿一半出来投AI，这样不是更合理吗？

And, you know, instead of flushing in their entire all their free cash flow, maybe use half of it on AI.

是的。

Yes.

没错。

Yes.

说得太对了。

That's exactly it.

说实话，这次引发担忧的原因是，我认为这是谷歌历史上第一次预计出现负自由现金流。

Like, again, the the reason this raised alarm bells is I think this was Google's first quarter of projected negative free cash flow ever.

我的意思是，从来都没有过，这简直不可思议。

So I mean ever, like that's crazy.

所以也许别让数字直接变负，别这么极端。

So the like maybe just don't make it negative.

不如直接说，我们打算投入一千亿美元，这可不是小数目。

Like maybe just say we're going in, we're going in, you know, a $100,000,000,000 is a lot of money.

现在我们好像都忘了，私营公司动不动就值上万亿美元。

Like, we we seem to I feel forget that now when private companies are valued at a trillion.

但我们不能把赖以赚钱的现金奶牛全押上去，得用一种更谨慎但依然积极的方式推进。

But going in where we are not gonna fundamentally put our entire cash cow that has made all of us so much money at risk for this and we're gonna do it in a slightly more measured but aggressive way.

在我看来，这种说法比Meta、谷歌这些公司互相盯着对方要健康得多。

To me that's a healthier story versus Meta and Google and everyone just looking at each other.

再说，微软根本没这么做。

And again Microsoft is not doing this.

大家都在互相盯着，心里想着：行，你先上吧。

Like all of them looking each other in the eye and being like, alright, you're going for it.

我也要上了。

I'm going for it.

咱们全都要赌一把，要么一起成功，要么一起完蛋，先把钱砸进去再说。

We're all gonna like sink or swim on this and just shovel money in and see what happens.

这才是我觉得健康的市场。

So that's my healthy market.

没错。

Right.

但我想知道，市场对人工智能到底怎么看？

But I'm I'm asking like what is the narrative about AI?

我的意思是，大家普遍相信AI会怎样？让我先把问题说清楚。

Like what does the market believe about let me just get the question out.

是的，没错，当然。

Yeah, yeah, yeah.

现在的情况是，市场根本不管不顾地砸钱，就赌一定能出通用人工智能。

Like right now, like basically the market is investing blank check, there's gonna be AGI.

如果真有大撤退，那说明这技术在经济上根本行不通。

You know, if there's a complete pullback, basically it means economically feasible technology.

所以市场会不会转变思路，觉得AI可能不会像下一个iPhone或者下一个互联网，但会是一种全新的辅助计算方式，这样投资就说得通了，我们可以慢慢加码，一步步下注。

So does the market thesis then sort of become, you know, AI might not be like the next, I don't know, iPhone or the next internet, but it will be this helpful new mode of computing and we will actually this might make sense and we will just keep investing in it to grow incrementally and sort of make our bets that way.

让大科技公司这么去下注，而不是——我们稍后会和保罗·库德罗夫斯基聊这个——干脆买个通用人工智能的看涨期权。

You know, have big tech make their bets that way versus let's say, and we're gonna talk about it with Paul Kudrowski in a couple weeks, but a call option on AGI.

这正是我想要的，我太喜欢这个思路了。

That's what to me, I love it.

我就想要这个。

That's all I want.

我一直希望市场能这么想：那个通用人工智能的看涨期权，把所有钱押上去，要么大获成功，要么血本无归。

That's what I wanted the market to actually think for a long time that and I love that call option on AGI that you just put all your money in and either it works or it doesn't.

我觉得这样不健康。

That's not healthy to me.

这就像经济在Robinhood上像赌徒一样炒作，而不是真想把这事做成。

That's like the economy trading on Robinhood as like a degen versus let's make this work.

咱们应该把它逐步扩大规模。

Let's scale it out.

也许它根本就不是传统意义上的渐进式发展。

And maybe it's not even incremental in the traditional sense.

它依然激进，但不是那种整个经济现在像一张看涨期权的感觉。

It's still aggressive, but it's not that kind of like the economy is a call option right now.

好吧，为了论证一下。

Okay, but then just for the sake of argument

请说。

Please.

一个看涨AI的人可能会说，这正是我最欣赏的——假设我是某家公司，我坚信这一点。

An AI bull might say this is the best thing that I like, let's let's say I'm one company and I'm a believer in this.

你可能会说，这正是我最想要的东西。

You might say this is the best thing that I want.

假设你是OpenAI，毕竟它是最有可能的那一个。

Let's say you're OpenAI because OpenAI is the most likely candidate.

我希望谷歌、微软、Meta和亚马逊都以为这只是个渐进式的技术。

I want Google, Microsoft, Meta, Amazon to believe this is an incremental technology.

我赌它是指数级的突破，而大公司不就是这么倒下的吗？

I'm going to make the bet that it is an exponential technology And isn't this the way that big tech companies lose?

它们没在真正能颠覆自己的技术上押重注，结果让初创公司趁虚而入，对吧？

Is that they don't bet big enough on the thing that will effectively disrupt them and an upstart comes in, right?

再往深一层说，如果你是桑达尔，你其实已经起步晚了。

And so just to take it one level deeper, if you're Sundar, you already sort of started late.

如果你听市场的风声，反而减少投入，这不等于把战场直接让给OpenAI了吗？

Isn't this you just seeding the battle to OpenAI if you listen to the market and spend less?

因为如果这真成了颠覆性的技术，那你等于已经输了。

Because if this does become this ultra powerful technology, you are effectively losing it.

你身为CEO，如果哪怕有10%的可能，这项技术真会像OpenAI所相信的那样颠覆一切，你还能淡定地松油门吗？

It's almost like can you rightfully with your CEO brain take your foot off the gas pedal if you believe there's even a 10% chance that this becomes the technology that OpenAI believes it will be.

但至少扎克伯格、孙达拉这些人，一直以来就是这么想的。

But that that's the way they have been thinking, at least Zuckerberg, Sundar and others.

扎克伯格就说过很多次，哪怕砸进去几百亿，跟它的潜力比起来，还是远远不够。

I mean, I think Zuckerberg made lots of comments around that around like even whatever billions of dollars is not enough given the potential.

对我来说，我还是更欣赏这种健康的竞争和市场机制，这本来就是亘古不变的规律——一个市值四万亿美元的行业巨头，跟初创公司怎么可能用同样的方式做事？

To me, again, I like this theme of healthy competition and economy because that's kinda normal for time immemorium, like, in large incumbent, not just large, $4,000,000,000,000 incumbent giant of the entire economy does not approach things the same way as upstart company does.

长期以来，这些大公司都活得不错，但说到底，问题还是在于：如果有一家公司率先实现了通用人工智能，是不是就意味着其他所有公司都输了，只有它赢？

And incumbents have done very, very well for a long time and like but but it all I guess it all comes back to that question of if one company reaches AGI before others, does everyone else lose and only they win?

而我觉得，这可能是眼下最大的转变。

And to actually maybe that's the biggest shift right now.

这种心态，我从来就没有过。

To me that mentality, I have not I've never had that.

现在，随着Kimi K3的讨论，还有模型路由这些话题，我感觉越来越多的人开始有这种感觉了。

I feel more and more people with the Kimi k three conversation and everything else and model routing.

你有没有听说，Stripe可能花一百亿美元收购OpenRoute？

I don't know if you saw Stripe might buy, like, OpenRoute for 10,000,000,000.

对，确实如此。

Like Mhmm.

我觉得这事儿大家都在谈，我也一直在这么想。

Like, everyone and and I've been thinking saying this.

我长期在做的，就是模型之间的互通性。

This is what I work on for a long time, model interoperability.

关键在于底层架构和产品体验。

It's about the harness and the product.

现在这种趋势，感觉真的在发生重大转变。

Like, the more it feels like that is a giant vibe shift right now.

那种‘要么实现通用人工智能，要么一无所有’的想法——赢家通吃。

And that idea that AGI or bust, whoever wins, owns the entire thing.

截至今天，7月24号星期五，你还认为除了OpenAI和Anthropic之外，有谁真这么想吗？

Do do you think anyone other than OpenAI and Anthropic still believe that as of Friday, July 24 today?

没有。

No.

但我认为这是因为到目前为止，这个游戏一直按某种方式在玩。

But I think that's because the game has been played a certain way up until this point.

等等，先听我说完，好吗？

Wait, just hear me out here, right?

也就是说，OpenAI 和 Anthropic 一直以来卖的其实是智能，而不是产品。

Which is that OpenAI and Anthropic have been sellers of intelligence, not products effectively.

没错，他们有 ChatGPT 和 Cloud Code，但那只是他们一直附带做的 API 业务，对他们来说非常重要。

I Yes, they have the ChatGPT and the Cloud Code, but that's like they've always had this API business on the side that's been very important to them.

我觉得这些公司可能觉得，API 业务将来会成为他们的负担，因为唯一一种你实现了 AGI 却还是输了的情况，就是把 AGI 开放给了别人。

I will posit that there's a chance that these companies think the API business is going to be a liability to them because the only way you reach AGI and still lose is if you make that AGI available to others.

而你赢的方式，是把 AGI 把握在自己手里，说：我之前确实开放了很多模型，但以后只给想做东西的人提供能力弱一点的版本。

And the way that you reach AGI and win is if you hoard it and basically say, I've been making all these models available, I'll still make less powerful models available to people that want to build things.

我刚发布了 GPT-10，这才是真正的重磅产品，其实 GPT-8、9 和 10，我都不打算对外开放了。

I just made OpenAI I just made GPT-ten and GPT-ten is the big one and actually you know GPT-eight and nine and ten I'm not going make available to people.

我只会把这用在自己的专有产品里，然后往上游走，推出OpenAI的客户关系管理系统、OpenAI的设计工具、OpenAI的客服系统。

I'm going to make that only available for use in my proprietary products and I'm going to go upstream and make, you know, OpenAI CRM, OpenAI design, OpenAI customer service.

你想获得AGI的访问权限？想用上它的强大能力？

You want the access, you want the power of AGI?

那就用在我的产品上吧。

You use it on my products.

顺便说一句，这样也能防止别人做模型蒸馏。

And by the way, that also prevents some distillation as well.

这才是把你的成果变成经济收益，同时把别人挡在外面的正确方式。

And that is how you turn what you've done into something that you can reap economic benefit from while keeping everybody else at bay.

我其实特别认同这个观点，如果你相信AGI或者商业化的逻辑，这确实是正确的策略。

So I love that actually and I think if you believe it that the AGI or Buss story, that actually is the right strategy.

那既然这么合理，他们为什么还不这么做呢？

So why aren't they doing it if they really believe it?

因为啊，Alex，他们其实并没有真的这么想。

Because they didn't Alex I don't

没有任何迹象表明他们会关闭自己的API。

have any information to suggest that they're gonna shut off their APIs.

我只是说，我们可能已经看到了一些苗头，比如那个Mythos项目，还有GPT 5.6的缓慢发布。你等着瞧吧，他们可能最终别无选择，只能关掉这些API。

I'm just saying we might have seen the beginning of it as well with the mythos thing, with the slow rollout of GPT 5.6, just wait because they may have no choice but to shut those APIs off.

说到健康的竞争，我其实挺欣赏这种做法的。

I mean, I going back to what is healthy competition, I kinda like this.

如果你认同前沿实验室应该走的路，那我们的策略一直很明确：要么全力押注AGI，掌控智能，把它变成产品。

If you have very distinct strategies for the Frontier Labs, our strategy is it's always been about AGI or bust owning the intelligence, turning that into products.

这将是我们整个业务的核心。

This is gonna be our entire business.

他们就该这么干。

They should go for it.

让其他人走另一条路好了，就是过去一周大家一直在讨论的那些：模型互操作性、基础设施、集成框架之类的。

And then let everyone else take the other route, which is what a lot of the world has been talking about over the last week of model interoperability, infrastructure, harnesses, all these kind of other things.

这样一来，竞争就更清晰了——到底谁才是最终的赢家。

And you have two then it's like then it's like a cleaner competition who's gonna win.

是的。

Yeah.

他们就该这么做。

They should do it.

听听这个。

Listen to it.

顺便说一下，也许可以把 API 限制在你们信任的前线工程师版本里使用。

And by the way, yeah, the other or could even make API sort of accessible only in your, like, trusted forward deployed engineer version.

比如，如果你是摩根大通，想用我们的 GPT-10，我们可以派顾问过去，确保你碰不到代码，直接帮你开发产品。

Like, if you're, for instance, if you're JP Morgan and want to build with our, you know, GPT-ten, we'll send some consultants over, make sure you don't get to touch the code and we'll build products for you.

但如果你是 Salesforce，我觉得我们可能会成为竞争对手。

But if you're Salesforce, I think that, you know, maybe we're gonna compete with you.

所以很抱歉，你只能用到 GPT-6，再往上就不行了。

So sorry, you can't have access to anything beyond GPT-six.

但问题就在于，克莱德设计那会儿，到底是什么时候的事了？

But that's where, I mean, there was, how long ago was Claude Design?

就三个月前吧，对吧？

Only like three months ago, right?

是的。

Like Yeah.

当你给Anthropic开放系统访问权限时，他们很可能直接复制你的产品，这种担忧依然存在。

You know, there's that moment of giving anthropic access to your systems and they're just gonna copy your products, which I do feel is still there.

也就是说，你刚才说的这种模式，本质上就是：只要你掌握了智能，就能做这类事情。

Like, I mean, that's still which is kind of the model you're talking about here, which is in if you own intelligence, you can just do these kind of things.

但我也不确定。

But I don't know.

你觉得未来几个月这事会怎么发展？

How do you how do you actually see this playing out over the next few months?

对。

Yeah.

就是这样。

Like that.

就是那样。

Like that.

我觉得可以。

I I think okay.

我看看。

I I look.

我们已经见过不少了，听起来是有点荒谬。

Just just, you know, we've seen so it sounds crazy.

我这么说可能觉得离谱，但到目前为止，AI领域已经发生过太多离谱的事了。

The way I'm saying it, it sounds crazy, but 've seen some crazy stuff happen in AI so far.

我觉得离谱的事还没完。

I don't think we're done seeing crazy things.

所以我觉得吧，你最好做好心理准备，说不定哪天API就没了。

So I think that, you know, just just, I don't know, prepare for the API to go away.

我的建议是，如果你真的在搞AGI或者BUS，那就听亚历克斯的。

Well, to to to to to say, I think if you if you are AGI or BUS, follow Alex's advice.

不对。

No.

我的意思是，我觉得这样

I mean, I think that that's

我不希望他们这么做。

I don't want them to do this.

我的意思是，

I mean,

我觉得最好还是

I think it'd better to

让每个人都能用上更多AI，但他们其实别无选择。

have more AI for everybody, but it's like they won't have a choice.

但这一点我同意。

But that's the only I agree.

这才是唯一真正合乎逻辑的做法，而不是现在这种零敲碎打、根本说不通的方式，我们已经看出来了。

That is the only actual logical approach rather than the current kind of piecemeal one that doesn't really say like and we're seeing that.

我们亲眼看着你推出神话，金米也出来了，这一切都在实时上演。

We're seeing you come out with mythos and Kimmy comes out and, like, we're seeing it in real time.

然后你又抱怨起模型压缩之类的问题，没完没了。

And then you're complaining about distillation, etcetera, etcetera.

如果你真的快接近通用人工智能了，那就坦然承认吧。

Like, if you're if you you're close to AGI, own it.

那样的话，你上市时也有个好故事可讲了。

Then you got a story too for the IPO.

这确实是个好故事。

That is a story.

明白了。

Okay.

那后来怎么样了？

So what happened?

咱们继续聊聊本期的主题吧——可能出什么问题。

Let's keep going with our, know, sort of what could go wrong theme this episode.

我想让你关注一下埃德·齐特隆的一篇文章。

I want to point you to a piece from Ed Zitron.

这周他把这篇文章命名为《次级数据中心危机》。

This week he called it the Subprime Data Center Crisis.

你本人当年在交易大厅亲历过金融危机，我想跟你聊聊这个情景，看看你觉得有没有相似之处，咱们再往下讨论。

Now, as somebody yourself who was on the floor of a trading desk during the financial crisis, I'd like to run this scenario by you and you can tell us whether you think it has any parallels and we can go from there.

埃德在这篇文章里详细分析了，当有人想建一个AI数据中心时，背后那些复杂的金融操作。

So so this is kind of Ed outlying what happens when someone wants to build an AI data center and sort of the financial machinations around that.

他说，一旦有人决定建AI数据中心，就会成立一个特殊目的实体，然后通过这个实体举债，有时还会把债务分成不同层级，大多数情况下卖给机构投资者、资产管理公司或者银行。

He says when somebody decides to build an AI data center, they form a special purpose vehicle which then raises debt, in some cases slices it into tranches, and in most cases sells them to institutional investors, asset managers, or banks.

你可以把SPV看作一个独立的小公司，当有人跟AI数据中心公司——比如OpenAI——签合同的时候，实际上签的是跟这个SPV的协议。

Think of a SPV as its own little company, and when somebody signs a contract with an AI data center company, let's say OpenAI, they're actually signing a deal with the SPV.

SPV拿到融资后，会把钱付给承包商和供应商，比如NVIDIA的GPU，同时也会从客户合同里收钱，前提是客户真的有钱支付。

When the SPV receives funds from the debt it raises, it makes payments to contractors and suppliers like NVIDIA for GPUs and receives the revenue from the customer contract assuming said customer is paying or has anything to pay for.

在建设期间，利息支出会从一个预先存好的利息储备账户里扣，等客户付款后才补上。

During construction, interest payments are taken out of the SPV from a prefunded interest reserve account when a customer pays.

这些资金用来支付数据中心的运营成本，然后按债务优先级偿还债权人。

The SPV uses those funds to pay for operating expenses of the data centers and then creditors based on their seniority and debt.

如果还有剩余，这些钱就归控股公司所有，算作收入。

Then if anything is left, the holding company, all this money counts as revenue.

所以，我刚才说的其实就是，这种建设模式风险挺大的。

So basically what I was just saying is, you know, this is this is a risky build out.

整个结构是由这些特殊目的公司搭建的，你投资时希望本金能拿回来，但其实没把握。

It's being put together by these SPVs that you know you would hope you would get the money back if you invest but you're not sure.

而要还上这些钱，必须靠这些AI数据中心产生收入。

And there's got to be revenue from these AI dentists data centers to pay it back.

你对这种安排怎么看？

What do you think about about this setup?

我们之前简单聊过，但让我直接说说埃德的核心观点。

We've talked about it a little bit, but let me just give the bottom line from Ed.

简单来说，每建一个数据中心，就成立一个完全独立的实体，专门持有芯片、承担债务，很多时候也承担大部分风险。

Put simply, every time somebody builds a data center, form a completely separate entity that owns the chips, owns the debt, and in many cases owns most of the risk.

从风险程度来看，你怎么看这个安排？

What do you think about this in terms of a riskiness quotient?

我觉得挺有意思的。

I think it's interesting.

再说一次，我很欣赏我们现在开始更细致地对比这个情况和当年的金融危机、抵押贷款证券——因为本质上，故事真的很像：过去是房地产作为底层资产，现在变成了数字基础设施、数据中心这些资产。

And again, like, I I actually really like how we're starting to get much more granular about how how this compares to the financial crisis and mortgage backed securities and like because again, the the story is actually very similar in that you have an underlying asset, whether in the past it was real estate, now it's digital infrastructure, data center infrastructure.

过去，房地产的价值是通过市场定价来衡量的，而且价格一直在涨。

You have, in the past, the value the the kind of marked value of real estate was how it was measured and it kept going up.

现在呢，一方面公司市值在涨，另一方面还有这种循环融资，我们之前也聊过很久了，它正在推高价格。

Now you have both kind of the marked value of companies, but also you have this kind of circular financing, and we've talked about this for a long time, that is inflating the price.

以前是房子或者整个房地产市场的估值在支撑。

Before it was the mark of a house or entire housing market.

现在变成看收入怎么确认，但风险早就被分拆转手了。

Now it's like how that revenue is being recognized, but then the risk has already been passed off and sliced up.

如果市场出现下滑，回到我们一开始讨论的问题，谁拥有什么、谁该承担多少，现在变得非常模糊。

And then any kind of downturn going back to where we started this conversation, who owns what, who pays what is very murky.

而且，想拿回自己的钱，哪怕只是其中一部分，都变得特别困难。

And then actually trying to get your money back, even a piece of your money back becomes very difficult.

所以这些相似之处，确实非常明显。

So all of those parallels are, I mean, very similar.

这么说吧，别再提世界杯没足球了，免得更沮丧，但这个论点逻辑上是自洽的。

Think like, not to not to bring us down even further after no more soccer in the World Cup, but it definitely that it's it's a logically consistent argument.

这一点我先承认。

I'll give it that to start for now.

好的。

Okay.

那咱们继续看看埃德是怎么说的。

So now let's let's keep going with what Ed Ed is arguing here.

对吧？

Right?

他主要的意思是，很多债务其实看不见，因为它们被藏在特殊目的实体里，而不是直接由公司持有。

So basically, what he's saying is a lot of this debt is you can't really see it because it's in these special purpose vehicles as opposed to held by the companies.

然后这些债务被藏在这些不透明的特殊目的载体里，被切得七零八落，分散到整个经济体系中。

And then it's in these opaque opaque SPVs and chopped up and farmed out through the economy.

那到底有多少债务呢？

So how much debt is there actually?

这是埃德·斯托里提供的数据。

Well, this is from Ed Story.

彭博社估计，目前人工智能数据中心的未偿还债务超过5000亿美元，其中至少有2000亿美元来自私人信贷，占私人信贷贷款总额的8%左右。

Bloomberg estimates there's over $500,000,000,000 in outstanding AI data center debt, at least $200,000,000,000 of it held by private credit making up roughly 8% of outstanding private credit loans.

不过，实际数字很可能还要高得多。

That being said, the number is likely much higher.

《日经亚洲》本周报道，过去五年里，Meta、谷歌、亚马逊、微软和甲骨文累计的债务总额已达1.65万亿美元，此外还有数千亿美元的表外债务——这意味着企业通过结构设计，可以把这些债务排除在负债报表之外。

Nikkei Asia reported this week that MetaGoogle, Amazon, Microsoft, and Oracle have accrued around $1,650,000,000,000 in outstanding debt over the last five years with an additional hundreds of billions of dollars worth of off balance sheet debt, meaning that the corporate structure allows the company to not include it as lie a part of its liabilities.

你怎么看这个情况？

What do you think about this?

这情况可真够严重的。

That's pretty pretty bad.

不会。

Would No.

不会。

No.

我的意思是，现在我们讨论的这些资金进入特殊目的载体后，并不算作资本支出。

I mean, now that we're and then I mean, and also I wanna note the piece continues around that the money that goes into these SPVs doesn't count as a capital expenditure.

Meta报告的资本支出是886亿美元，但这不包括Hyperion特殊目的载体，它另有460亿美元的敞口。

So Meta recorded 88,600,000,000 in capital expenditures, but that doesn't include the Hyperion SPV, which had its own $46,000,000,000 of exposure.

这个过程非常相似。

And this process is very similar.

抱歉。

I mean, sorry.

我正在试着找找和2007到2008年之间的相似之处，但说实话，

I'm I'm, like, trying to see how many parallels I can draw to 2007, 2008, but like

已经太多了，老兄。

It's already too many, man.