← 返回任务列表

A Field Guide to AI Market Freakouts

225 段 · 1 位说话人 · 原片 24:53
M1
M10:00

今天这期 AI Daily Brief,我们来聊一份 AI 市场恐慌实地指南。

Today on the AI Daily Brief, a field guide to AI market freakouts.

AI Daily Brief 是一档每天更新的 podcast 和视频节目,内容聚焦 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,朋友们,开始之前先快速说几个通知。

All right, friends, quick announcements before we dive in.
M1
M11:21

不过总的来说,就像我说的,我最开心的还是有你们在这儿,跟我一起探索这个疯狂的世界。

Mostly though, like I said, I am just excited to have you here as we explore this insane world.

今天还有一点要说明,就是今天就是那种连所有标题新闻都完全能归进主线主题里的一天,所以这一整期我们就只有一个延展的大主题。

The one other note today is that it was one of those days where even all of the headlines fit inside the main theme, so we just have one extended theme for the entire episode.

明天我们会回到平常的结构,把标题新闻和主话题分开,不过现在,先来聊聊 AI 市场周期性恐慌这件事,以及它对 AI 未来意味着什么。

Tomorrow we will be back with our normal division between headlines and main, but for now, let's talk about the periodicity in AI market freakouts and what they mean for the future of AI.

欢迎回到 aidailybrief。

Welcome back to the aidailybrief.

我们现在正处在新一轮 AI FUD 和担忧情绪的当中。

We are in the midst right now of our latest round of AI FUD and concern.

这一次具体讨论的是 Chinese AI,以及它可能会怎么影响 OpenAI 和 Anthropic 这类公司的营收潜力,特别是当它们正准备在今年晚些时候或者明年年初上市的时候。

Now, this one specifically is about Chinese AI and how it might impact the revenue potential of companies like OpenAI and Anthropic, especially as they position to go public later this year or early next year.

这几年我一直非常密切地观察这件事,我觉得投资人对 AI 感到紧张的方式,其实已经呈现出一些非常清晰的模式。

Having watched this very closely now for the last few years, I think there are some pretty clear patterns in the specific ways in which investors get stressed about AI.

所以,这就是我们今天要聊的内容。

And so that's what we're going to talk about today.

不过在进入这个话题之前,我们得先更新一下这一轮担忧的最新进展,因为 Trump 政府对 Moonshot 提出了新的指控。

However, to get into it, we first need to do an update in this particular round of concern as the Trump administration levies new allegations against Moonshot.

本周早些时候,财政部长 Scott Bessent 提议,作为对 distillation 攻击的回应,Chinese AI 公司可能会面临制裁。

Earlier this week, Treasury Secretary Scott Bessent proposed that Chinese AI companies could face sanctions as a response to distillation attacks.

虽然在 Chinese 模型表现这个问题上,model distillation 一年多来一直都是热门话题,但这是第一次有行政部门成员暗示,政府可能会对此采取行动。

And while model distillation has been a hot topic for over a year when it comes to Chinese model performance, this was the first time a member of the executive suggested the government might do something about it.

Bessent 在周三进一步强化了自己的政策立场,他在 X 上发帖写道:我们支持 open source AI,以及它所释放出的创新潜力,但 open source 不等于可以随意掠夺 American IP。

Bessent doubled down on his policy view on Wednesday, writing in a post on X, we support open source AI and the innovation it unlocks, but open source is not open season on American IP.

当 PRC 公司进行秘密的、工业级的大规模 distillation 攻击,而且这种行为已经越过界线、构成 IP theft 时,制裁以及列入 entity list 都会被摆上桌面。

When PRC firms conduct covert industrial-scale distillation attacks that cross the line into IP theft, sanctions and entity list designations will be on the table.

而这条帖子发出几小时后,白宫 Office of Science and Technology Policy 主任 Michael Kratsios 就点名指控了 Moonshot,也就是 Kimi K3 的开发者。Kratsios 写道:我们掌握的信息显示,Moonshot AI 在开发 K3 模型时,对 Anthropic 的 Claude 进行了 distillation。

Now that post came hours after White House Office of Science and Technology Policy Director Michael Kratsios made specific allegations against Moonshot, the creators of Kimi K3, wrote Kratsios, we have information that Moonshot AI distilled Anthropic's Claude for the development of its K3 model.

为此,他们开发了一个复杂的内部平台,用来对美国模型实施大规模 distillation,并且可以在多种访问方式之间快速切换,以躲避侦测。

To do this, they developed a sophisticated internal platform to conduct large-scale distillation against US models, allowing them to quickly switch between multiple methods of access to avoid detection.

Moonshot AI 还获取了配备 GB300 的服务器,并且在 Thailand 使用过 GB300,很可能是用来训练他们的 AI 模型。

Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.

美国坚定支持 AI 的自由和公平发展,包括一个繁荣的竞争生态系统,覆盖 frontier models、专用系统、open-source frameworks 以及 open-weight models。

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks and open-weight models.

合法的 AI distillation,也就是用来创建更小、更高效模型的那种,在这个开放创新生态里发挥着非常关键的作用。

Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem.

但是,以窃取美国专有技术、削弱美国研究为目的的那种大规模、秘密、工业化 distillation,是不可接受的。

However, large-scale covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.

这里要再说一次,虽然一直以来都有传言,说 Chinese labs 通过 Southeast Asia 的非法渠道拿到 NVIDIA GPU,但这是政府官员第一次就这种做法提出明确指控。

Now, here once again, while there have long been whispers of Chinese labs getting access to NVIDIA GPUs through illicit pathways across Southeast Asia, this is the first time a government official has made specific claims about the practice.

Anthropic 公共政策负责人、同时也是前国务院官员的 Sarah Heck 证实,Anthropic 正在就这件事与政府合作。Heck 写道:非法的、对抗性的 distillation 属于 IP theft 和 industrial espionage,它会增强对手方的军事和情报能力。

Anthropic Head of Public Policy and former State Department official Sarah Heck confirmed that Anthropic is working with the administration on this matter, with Heck writing, illicit adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities.

这是一个国家层面的挑战,会给美国以及民主盟友带来严重的国家安全风险。

It is a national challenge that creates serious national security risks for the United States and democratic allies.

在这些帖子发出之后,The Information 报道称,商务部正在积极调查 Moonshot 和其他 Chinese labs,原因是它们绕过出口管制来获取 NVIDIA GPU。

Following the posts, the information reported that the Commerce Department has an active investigation into Moonshot and other Chinese labs for circumventing export controls to access NVIDIA GPUs.

Signal 很好地总结了很多人的感受,他们写道:哇,我这样理解对吗?

Signal summed up the feeling of many when they wrote, wow, am I understanding this correctly?

第一,把 Chinese 模型列入 entity list。

One, put Chinese models on the entity list.

第二,迫使 American 公司去购买 American labs 更贵的 AI。

Two, force American companies to buy more expensive AI from American labs.

第三,然后眼看着世界其他地方去用那些更便宜、而且智能水平相当甚至更强的模型。

Three, watch the rest of the world use cheaper models with equal or better intelligence.

第四,你会发现,distillation 并不会因为 Washington 发了个清单,就神奇地停止。

Four, discover that distillation did not magically stop because Washington published a list.

第五,这会让 American 企业在全球范围内更没有竞争力,同时还会抬高普通 American 的生活成本。

Five, make American businesses less competitive globally while raising prices for ordinary Americans.

第六,与此同时,这一切根本挡不住 Chinese AI labs 变得越来越强。

Six, meanwhile, none of this stops Chinese AI labs from becoming better and better.

恭喜你,你这是靠给整个国家的竞争力加税,来保护国内的 AI 公司。

Congratulations, you've protected domestic AI companies by taxing the competitiveness of the entire country.

真是一团巨大的烂摊子。

What a colossal mess.

不过话说回来,Bessent 和 Kratsios 在 White House 内部到底有多少支持,其实都还不是特别清楚。

Now, for what it's worth, it's not even particularly clear how much support Bessent and Kratsios have within the White House.

Wired 在周三写道,Trump administration 内部对于该怎么回应 China 领先 AI models 的快速崛起,意见是分裂的。

Wired wrote on Wednesday, the Trump administration is split over how to respond to the rapid rise of China's leading AI models.

这场争论大体上分成两派:一派是 White House 的部分人,他们主张对 Chinese AI 采取更严格的管制,因为这些模型可能很快就会和强大的 U.S. models 分庭抗礼;另一派是 Commerce Department,他们认为这些限制根本行不通。

The debate is broadly divided between parts of the White House, which has pushed for stricter controls on Chinese AI that may soon rival the powerful U.S. models, and the Commerce Department, which has viewed those restrictions as unworkable.

消息人士说,Commerce Secretary Howard Lutnick 更倾向于用竞争来对抗 Chinese AI。

Sources said that Commerce Secretary Howard Lutnick would prefer to combat Chinese AI with competition.

他提出过给 U.S. labs 激励,鼓励他们把自己的 models open-source,而且最近几周也已经和多家 labs 谈过这个想法。

He has pitched incentives for U.S. labs to open-source their models and spoken with multiple labs about the idea in recent weeks.

Andrew Curran 写道,Commerce Department 显然正在强烈反对监管,因为他们觉得这根本不可行。

Writes Andrew Curran, the Commerce Department is apparently arguing strongly against regulation, which they see as unworkable.

David Sacks 肯定会站在他们这边。

They are certain to have David Sacks on their side.

American open source 也许会迎来一个出人意料的盟友,那就是 Howard Lutnick。

American open source may have a surprise ally in Howard Lutnick.

根据这篇报道,他主张直接给 U.S. open-source software 提供激励,加快开发速度,这样他们才能赶上 Chinese OSS。

According to the report, he's arguing for direct incentives to U.S. open-source software to accelerate development so they can match Chinese OSS.

Commerce 认为,这是一个可以避免去监管任何人的办法。

Commerce sees this as a way to avoid regulating anyone.

而且据说,Lutnick 甚至还直接和一些没被点名的 American labs 负责人见过面,讨论到底该怎么最好地推进这件事。

Apparently, Lutnick has even met directly with leaders at unnamed American labs to discuss how best to accomplish this.

这会不会是带着联邦资金支持的 open AI OSS 复兴?

Perhaps a rebirth of open AI OSS with federal funding?

我觉得,他应该去见见 Nous、Prime Intellect,还有其他一些团队。

He should meet with Nous and Prime Intellect, among others, in my opinion.

这进展很让人兴奋。

Exciting developments.

AEI 的 Ryan Fedasiuk 写道,这可能是截至目前为止,二零二六年对 U.S. AI policy 来说最重要的一周。

Writes AEI's Ryan Fedasiuk, this might be the most important week for U.S. AI policy so far in 2026.

现在,有一类人正在非常密切地盯着整件事,那就是投资人。

Now one group that's watching this whole thing very closely is investors.

最近大家越来越担心一件事——这也是 AI 市场一长串担忧中的最新一个——那就是,如果企业买家突然之间开始特别在意成本,转头去找更便宜的替代方案,那他们很可能就会转向这些 open China models,从而压缩流向 American 头部 labs 的收入,并且会对整个市场带来重大影响。

There has been a growing concern, the latest in a long line of AI market concerns, that if all of a sudden, enterprise buyers get all cost conscious and they look to cheaper alternatives, they're going to find themselves moving towards these open China models, undercutting the revenue coming to the major American labs with big implications across the market.

而对那些一直密切关注的人来说,这一切其实都是一个更大市场叙事的一部分,这个故事到现在已经演了快四年了。

And for people who are watching closely, all of this is part of a broader market narrative story that's been happening now for coming up on four years.

现在 AI 最重要的问题之一,不是谁在用 AI,而是谁把它用得好。

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

KPMG 和 University of Texas at Austin 刚刚分析了一百四十万次真实职场里的 AI 互动,结果发现了一个很让人意外的现象。

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

影响最大的那批用户,并不是更会做 prompt engineering 的人,而是把 AI 当成推理伙伴来用的人。

The highest impact users aren't better prompt engineers, they treat AI like a reasoning partner.

他们会先界定问题,引导思考,反复迭代,还会不断推动 AI 给出更好的答案。

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

而好消息是,这些行为是可以大规模教会的。

And the good news, these behaviors are teachable at scale.

如果你正试着从只是能接触到 AI,走向真正具备能力,那 KPMG 关于高级 AI 协作的这份研究很值得你花时间看看。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.
M1
M19:46

从二零二二年底 ChatGPT 发布开始——那时候正好处在 Fed 加息周期的中段——市场里基本就是 AI 对抗其他一切。

Since the end of 2022 when ChatGPT was released, which was right in the midst of the Fed rate hiking cycle, it has been AI versus everything in markets.

确实,回到二零二三年初那波 AI 热潮刚开始的时候,大家当时很困惑,为什么在 Fed 还在加息的情况下,市场却还在一路猛涨。

Indeed, back in the early days of the AI boom at the beginning of 2023, there was a lot of confusion on why markets were ripping while the Fed was still raising interest rates.

Fed 一直到七月才停止加息,而那时候也已经很明显了,利率政策根本挡不住 AI 投资。

The Fed didn't stop hiking until July, and it became obvious that rate policy would do nothing to deter AI investment.

到了现在,AI 已经成了美国经济里的一个结构性组成部分,而且可以说很可能还是最重要的一部分。

At this point, AI has become a structural part of the US economy and arguably the most important part.

根据 Bloomberg 的说法,AI 投资现在占了美国 GDP 增长的百分之二十五,是历史上所有行业里最大的单一贡献来源。

According to Bloomberg, AI investment now represents 25% of US GDP growth, the largest single contribution of any sector in history.

JP Morgan 的分析师说,自从 ChatGPT 在二零二二年底发布以来,AI 推动了 S&P 500 百分之七十五的回报、百分之八十的盈利增长,以及百分之九十的资本开支增长。

JP Morgan analysts claim that since the release of ChatGPT in late 2022, AI drove 75% of returns in the S&P 500, 80% of earnings growth, and 90% of capital spending growth.

所以很现实地说,现在 AI 产业已经成了美国经济和美国市场的核心驱动力。

In a very real way then, the AI industry is now the driving force in the US economy and US markets.

也正因为这样,市场会格外害怕 AI 出问题,或者势头反转,因为他们觉得太多东西都押在这上面了。

As a consequence of this, the market is extra terrified of AI going badly or reversing because they feel like so much is riding on it.

在这种背景下,所谓 AI 泡沫的讨论,就不只是一个吸引眼球的标题了,而是市场分析里一个持续存在、而且非常核心的部分。

And in this context, AI bubble talk isn't just a catchy headline, but is a persistent and central piece of market analysis.

虽然这么说确实有点过度简化,但到了这个阶段,如果有人把美国经济基本看成是在赌 AI 会不会带来足够颠覆性的变化、能不能创造出足够的收入来证明这波基础设施扩张是合理的、从而让这场狂欢继续下去,那也不算离谱。

While acknowledging that this is a bit reductive, at this point, one could be forgiven for looking at the American economy as basically a bet on AI being transformative enough to drive enough revenue to make the infrastructure buildout justified so the party gets to continue.

顺便说一句,就算你平时对市场并不怎么感兴趣,现在 AI 热潮已经发展到一个所有人都得关注、都得理解这套叙事的阶段了,因为现在 S&P 500 里大概有百分之五十都是 AI 股票,或者是受 AI 影响很大的股票。

And by the way, even if you're not typically interested in markets, we're now at a point in the AI boom where everyone needs to pay attention and understand the narrative, given that around 50% of the S&P 500 is now AI or AI-exposed stocks.

这就意味着,就算你只是被动投资指数,或者只是往 401(k) 里存钱,我的朋友,你也已经有相当大的 AI 敞口了。

That means that even if you are just a passive indexer or contribute to a 401(k), you, my friend, have significant AI exposure.

好,有了这些背景,我们接下来就来说说那些会把投资者吓到的事。

Now, with this background, let's talk about the stuff that freaks investors out.

我们现在正在经历的这一种,当然就是便宜模型这件事——它会削弱 Anthropic 和 OpenAI 收高溢价、并继续扩大营收的能力。

The one that we are living through right now, of course, is cheap models, undercutting Anthropic and OpenAI's ability to charge a premium and continue to grow their revenue.

这是当下最新的这波恐慌,不过它其实有很多不同版本,一路都可以追溯到二零二五年初的那个 DeepSeek 时刻。当时市场相信,这家中国实验室只花了几百万美元,就做出了前沿 AI,而美国那些实验室在训练上花的却是几十亿美元。

This is the freakout du jour, but it's had various versions going all the way back to the DeepSeek moment at the beginning of 2025 when the market believed that the Chinese lab had produced frontier AI for a few million bucks when the US labs were spending billions on training runs.

当然,DeepSeek 当时被呈现出来的方式,本身就有很多问题。

Now, there were a ton of things wrong with how DeepSeek had been presented.

大家后来也开始弄明白,training compute 和 inference compute 是两回事。

People learned about the difference between compute for training and compute for inference.

DeepSeek 的一些说法,后来也都被打上了很大的问号。

Some of DeepSeek's claims started to have big asterisks around them.

而且最后,市场也发现 DeepSeek R1 实际上并没有领先美国训练出来的那些最强模型。

And finally, markets also ultimately discovered that DeepSeek R1 wasn't actually ahead of the best models trained in the US.

它只不过是当时市场上第一个、也是唯一一个免费的 reasoning model,而 OpenAI 的 o1 还被锁在订阅后面。

It was simply the first and only free reasoning model on the market, with OpenAI's o1 still locked behind subscription.

现在,随着 Kimi K3 的发布,我们又来到了另一个 DeepSeek 时刻;而在很多方面,这件事其实早就被 GLM 5.2 先铺垫过了,算是一道开胃小菜——那时候 Fable 因为美国政府的缘故下线了。

We now find ourselves at another DeepSeek moment with the release of Kimi K3, which in many ways was teed up with the amuse-bouche of GLM 5.2 when Fable was offline, courtesy of the US government.

现在,对很多投资者来说,他们对这件事的过度简化理解,就是一种几乎刻在脑子里的信念:觉得中国 AI 很便宜,而美国 AI 贵得吓人。

Now, to many investors, the oversimplified understanding of the situation is a hardwired belief that Chinese AI is cheap while US AI is eye-wateringly expensive.

更雪上加霜的是,K3 出现的时候,前面我们刚刚经历了一个持续好几个月的叙事,说美国公司都在大砍 token 预算。

And it didn't help that K3 arrived just after we had had a months-long narrative about US companies slashing token budgets.

现在,我们还没有完全走出这个 DeepSeek 时刻,不过能帮助市场慢慢走出来的一点是,大家会逐渐意识到,K3 确实更便宜,但远没有很多人想象得那么便宜。

Now, we are not out of this DeepSeek moment yet, but among other things that will help drag us out is the recognition that while K3 is, yes, cheaper, it is nowhere near as cheap as some believe.

K3 现在的服务价格大概是 Fable 的三分之一,或者说是 Opus 的一半。

K3 is currently being served at around a third of the price of Fable or half the price of Opus.

这当然还是一笔很有意义的节省,但并不是那种便宜到几乎像白送一样的程度——我觉得很多分析师脑子里想的其实是那种级别。

That's still a meaningful savings, but it's not pennies on the dollar like I believe many analysts have in their heads.

不过,真正存在的情况是,美国公司确实在想办法控制自己的 AI 预算,特别是在越来越多 agentic 的应用场景开始上线之后,这对头部公司的市场表现可能都会带来影响。

Still, what is real is that US companies are looking for ways to manage their AI budgets, especially as more agentic use cases come online with potential implications for the market success of the leading companies.

好,把视角再拉回到之前几轮市场恐慌,还有一种跟营收可持续性有关的担忧,就是所谓营收循环的问题。

Now, zooming back into previous versions of the market freak-out, another one that related to the durability of revenue was the concern around the circularity of revenue.

这个担忧在去年夏末到秋天那段时间明显升温了,因为你开始看到像这样的一些图在到处流传,上面画的是 NVIDIA 和其他公司在整个行业里达成的各种交易关系图。

This concern really picked up between the end of last summer and last fall when you started to see graphics like this one floating around, showing maps of all the deals that NVIDIA and others were making across the industry.

虽然这些交易里很多都只是对 startup labs 和 neoclouds 的小规模投资,但在那个时候,刚好又有报道说 NVIDIA 达成了一项协议,准备向 OpenAI 投资最多一千亿美元。

While many of those deals were small-scale investments in startup labs and neoclouds, at that point, there had been recent reports that NVIDIA had struck a deal to invest up to 100 billion in OpenAI.

有些投资者觉得,这笔潜在投资很可能会直接以芯片收入的形式流回 NVIDIA,从而大幅、而且在他们看来是人为地,抬高公司的利润表现。

Some investors saw this potential investment as likely to flow straight back into NVIDIA as chip revenue, massively and in their minds artificially boosting the bottom line.

顺便说一句,那笔交易最后是在二月份完成的,规模是三百亿美元,另外还有来自 Amazon 的五百亿美元和来自 SoftBank 的三百亿美元。

Now, that deal, by the way, ended up closing at 30 billion in February alongside 50 billion from Amazon and 30 billion from SoftBank.

而对一些人来说,这种循环融资意味着,不只是 OpenAI 和 Anthropic 的收入让人怀疑,大家还担心 NVIDIA、其他 hyperscaler,以及一路往下整个芯片供应链,都是建立在不稳的基础之上的。

And for some, the circular financing means that this isn't just OpenAI and Anthropic's revenue that's suspect, but also a concern that NVIDIA and the other hyperscalers and everyone all the way down the chip supply chain are built on shaky foundations.

因为大家都很喜欢打比方嘛,所以循环融资之所以让人担心的另一个原因,就是历史上循环融资曾经戳破过泡沫。

Now, since people love analogy, the other reason that circular financing has been a concern is the history of circular financing popping bubbles.

有些人就把 dot-com 泡沫的崩盘归因于一种循环融资,或者说 vendor financing,最后失控了。

Some credit the dot-com crash to a type of circular or vendor financing coming unstuck.

像 Cisco 这样的硬件供应商,会通过融资方式把设备卖给那些刚起步的 startup,所以等这些 startup 倒掉的时候,硬件厂商就成了最后接盘的人。

Hardware providers like Cisco would sell equipment to fledgling startups on finance, so when the startups went belly up, the hardware manufacturers were left holding the bag.

结果就是,这种做法在两千年代初实际上被定为非法了,不过这次还是有几个关键区别。

That practice was actually made illegal in the early 2000s as a result, and there are a few key differences this time.

这一次,NVIDIA 实际上并没有参与 vendor financing。

In this case, NVIDIA isn't actually participating in vendor financing.

OpenAI 是用现金在买他们的 GPU。

OpenAI is paying cash for their GPUs.

哪怕这里面有一部分现金,其实是来自 NVIDIA 的投资。

Even if some of that cash is investment money from NVIDIA.

这意味着,如果 OpenAI 把 runway 烧完了,NVIDIA 手里也不会出现债务违约,也不会在资产负债表上炸出一个巨大的窟窿。

What this means is that if OpenAI runs out of runway, NVIDIA doesn't have a debt default on their hands and a massive hole in their balance sheet.

他们只是会持有已经贬值了的 OpenAI 股票。

They just have devalued OpenAI stock.

而且这一次的这些公司,从根本上说也完全不一样。

The companies this time around are also categorically different.

OpenAI 和 Anthropic 不管是在规模、正当性,还是营收体量上,都跟 pets.com 完全不是一回事。

OpenAI and Anthropic are nothing like pets.com in their scale or legitimacy or the scale of their revenue.

AI 市场恐慌的下一类情况,基本上每隔几个月就有可能卷土重来。

The next category of AI market freak-out has a chance to come up renewed every few months.

这种担忧就是,AI 公司拿不出增长得足够快的收入,来证明这么大的投入是合理的。

That concern is AI companies not showing revenue growing fast enough to justify spend.

而自从这波 AI 基础建设开始以来,这一直都是每个季度 hyperscaler 财报里的大主题。

And this has been the big story for hyperscaler earnings every quarter since the AI buildout began.

在今年年初的时候,只要稍微有一点 AI ROI 的迹象,往往就足够让投资者满意了。

At the beginning of the year, some hint of AI ROI tended to be enough for investors.

一月份,Meta 报告了百分之二十四的营收增长,这暗示 AI 正在推动广告收入上升,而这对投资者来说就已经够了。

In January, Meta reported 24% revenue growth suggesting that AI was driving a boost in ad revenue and that was enough for investors.

那天晚上,股价直接飙升了百分之十。

The stock soared by 10% that night.

随着这一年继续往后走,投资者开始对这些 hyperscaler 提出更多要求,不过有些公司还是交得出成绩。

As the year continues, investors are starting to ask a little more from the hyperscalers but some are still able to deliver.

比如说,就在这周三,Google 就展示了,要安抚这类市场情绪,到底得做到什么程度。

As an example, on Wednesday of this week, Google showed just how much it's going to take to appease this particular market.

他们交出了一份远超预期的财报,整体增长达到百分之二十四,cloud 部门同比增长也高得相当惊人,达到百分之八十二。

They delivered a huge earnings beat reporting 24% overall growth and a fairly massive 82% growth for their cloud division year over year.

他们靠提供 AI inference 服务赚得盆满钵满,但股价并没有因此大涨。

They're making money hand over fist serving AI inference but the stock didn't soar.

事实上,投资者盯着看的数字,不是 cloud 部门的增长,而是他们 CAPEX 支出的增长。

In fact, the numbers that the investors were focused on were not the growth of their cloud division but the growth of their CAPEX spend.

Wall Street Journal 是这么写的。

Wrote the Wall Street Journal.

Wall Street 对人工智能投资的容忍度看起来好像没有上限,但 Google 的母公司 Alphabet 在周三找到了那个上限:两千亿美元。

Wall Street's tolerance for artificial intelligence investment might seem to have no limits but Google parent Alphabet found one on Wednesday: $200 billion.

在年初的时候,Google 估计他们的 CAPEX 会在一千八百五十亿美元到这个水平之间,所以两千亿美元其实只增加了百分之八,这基本上都可以归因于供应链里的成本压力,可即便如此,这个两千亿美元的数字,对一些人来说还是成了临界点。

At the beginning of the year, Google estimated that they would spend between $185 billion on CAPEX, making this only an 8% increase that can mostly be chalked up to cost pressure in the supply chain, and yet still that $200 billion number was the breaking point for some.

Zacks Investment Management 的 Brian Mulberry 评论说:“两千亿就是那条不能跨过去的线。”

Brian Mulberry of Zacks Investment Management commented, "The 200 was the do not cross line."

你不可能一下子往外砸这么多现金,却还不跟大家好好解释。

You can't be offloading this much cash and not talking about it.

现在看,这很可能只是一个大整数带来的心理冲击,但它也确实说明了这个问题。

Now likely this is just the psychological shock of a big round number but it also demonstrates the point.

即使增长率达到 82%,投资者还是不确定收入增长能不能继续证明不断攀升的 CAPEX 是合理的。

Even at 82% growth, investors aren't sure that revenue growth will continue to justify the escalating CAPEX.

Google 股价最先有了反应,在隔夜市场下跌了 1.2%。

Google stock was the first to see the result, falling by 1.2% in overnight markets.

所以很明显,市场恐慌里对“收入增长不够快”的担忧,还有一个对应面,就是 CAPEX 长得太快。

So obviously the market freakout concern of revenue growth not growing fast enough has a twin in CAPEX growing too fast.

CAPEX 增得越多,收入要爬的那座山就越高,才能证明它合理。

The more CAPEX grows, the bigger the hill that revenue has to climb to justify it.

对一些投资者来说,担心的是他们根本不明白这一切最后会走向哪里。

To some investors the concern is that they just don't understand where this all ends.

Hyperscalers 现在已经连续三年、每个季度都上调 CAPEX 指引了。

Hyperscalers have now raised CAPEX guidance every quarter for the past three years.

合计 CAPEX 现在看起来明年会超过一万亿美元,而且还有可能继续往上走。

Combined CAPEX is now looking like it will exceed a trillion dollars next year and could head even higher.

当然了,这里挑战的一部分就在于,我们其实是在完全没有先例的水域里前进。

Now of course part of the challenge here is that we are just in totally uncharted waters.

看到公司在不到两年里,营收从 10 亿美元涨到 300 亿美元,这在历史上任何市场、任何时候,都完全没有先例。

Seeing companies go from $1 billion in revenue to $30 billion in revenue in less than two years has absolutely no precedent in the history of any markets anywhere at any time in history.

当然,这也是为什么 2024 年末那套“泡沫论”今年会降温的原因之一——正是 OpenAI 和 Anthropic 这种超高速增长。

Now of course part of the reason that the bubble narrative from late 2024 calmed down this year was exactly this sort of hypergrowth from OpenAI and Anthropic.

而很多投资者当时是在算总共有多少知识工作者席位,把这个数字乘以 20 或 30,然后发现根本没有足够的收入来真正证明这笔巨额 CAPEX 支出合理。

Whereas many investors were counting the number of total knowledge worker seats available, multiplying that by 20 or 30, and not finding enough revenue to really justify this big CAPEX spend.

当巨大的使用场景真正上线,我们开始看到员工每个月花几百美元,甚至几千美元时,这就帮助投资者意识到,这不只是另一类 SaaS,而是某种根本不同的东西。

When gigantic use cases came online and we started to see employees spending hundreds or even thousands of dollars a month, it helped jog investors to realize that this is not just a different category of SaaS and represents something fundamentally different.

不过,对那些喜欢在配置资产时有清晰历史模式可供参考的投资者来说,根本不同也不一定就让人安心。

Still, fundamentally different isn't necessarily comfortable for investors who like having clear patterns in history that they can draw from as they're making their allocations.

而且即使在这种大规模 agent 支出的新背景下,它本身也制造出了新的担忧类别。

And even in that new context of massive agent spend, that in and of itself has created new categories of concern.

AI 市场恐慌里一个反复出现的大主题,就是任何关于 AI 支出上限的暗示。

One of the big recurring AI market freakout themes is any suggestion of limits of AI spend.

当然,去年也有一个类似版本,就是那份臭名昭著的 MIT 报告,声称 95% 的 gen AI 试点都会失败。虽然那项研究在社会科学上简直是彻头彻尾的垃圾,MIT 该为名字和它扯上任何关系而感到羞耻,但大家在乎的只有标题,而那个标题最后传到了华尔街每一张桌子上。

Last year of course we had a version of this in the notorious MIT report that claimed that 95% of gen AI pilots fail and while the study was absolute garbage social science that MIT should be embarrassed to have its name anything associated with, all that mattered was the headline and that headline found its way onto every desk on Wall Street.

今年这个版本稍微没那么蠢,因为它至少是基于现实的,也就是 token CAPEX 这个想法——随着公司开始限制 AI 使用,情况就是这样。

This year's version of that is a little less stupid because at least it's based in reality, which is of course the idea of token CAPEX as companies begin to limit AI use.

不过这一次,标题还是比实际政策夸张得多。

Now once again the headlines are more dramatic than the actual policies.

比如 Uber 已经把用户每月上限设到了 1500 美元,而 Tesla,另一家情况相似的公司,则允许员工每周使用价值 200 美元的 tokens,而且还能申请更高预算。

Uber is capping users, for example, at $1,500 per month and Tesla, another company in a similar boat, is allowing workers to use $200 worth of tokens a week with the ability to request larger budgets.

重要的是,我们几乎还远没触及大多数知识工作者能把这种 token 预算用到接近上限的程度,这意味着即使每家公司都采用类似政策,增长空间依然巨大;但市场还是明白,大家都会想办法控制自己的 token 预算,这就带来了一个新的、隐约逼近的威胁:更便宜的 inference。

Importantly we've barely scratched the surface on the ability for most knowledge workers to use anywhere near that sort of token budgets, meaning that even if every company adopted similar policies there is still an enormous amount of growth to be had, but still the market understands that everyone will be looking for ways to bring their token budgets under control, creating a new looming threat of cheaper inference.

现在还有最后一种值得一提的、反复出现的市场恐慌,虽然今年它其实没那么让人担心,但为了完整起见还是要说一下,那就是 performance plateau,也就是性能平台期,最终又会导致支出下降。

Now one last recurring market freakout that is worth mentioning for completeness, even though it hasn't been much of a concern this year, is the idea of performance plateaus that once again lead to reduced spend.

如果 AI 性能碰到天花板,它能做的事就会受限,公司愿意在它身上花的钱也会受限,总收入增长也会受限,这样 CAPEX 就更难被证明是合理的。

If AI performance hits the wall, that limits what it can do, which limits the amount that companies will spend on it, which limits the total revenue growth, which makes it harder to justify the CAPEX.

这就是 2024 年秋天最大的担忧,当时主流叙事是 pre-training 的 scaling wall。

This was the big fear in the fall of 2024 when the major narrative was the pre-training scaling wall.

那时候还有传言说,Anthropic 和 OpenAI 都在夏天砍掉了旗舰级 pre-training 训练跑,因为他们看到的那点性能提升,根本不值得增加的成本。

There were rumors at the time that both Anthropic and OpenAI had scuttled flagship pre-training runs over the summer as the minor performance boost they were seeing didn't justify the increased cost.

但从今天这个位置回头看,那种说法看起来既老套,又可笑,又疯狂。

Now from where we are today that argument looks quaint, ridiculous, insane.

就好像 Gary Marcus 和 Ed Zitron 是在主动想让你亏钱。

Like Gary Marcus and Ed Zitron were actively trying to lose you money.

几个月后,我们第一次通过 o1 看到了 AI reasoning,才意识到 pre-training 并不是从 AI 里获得更高性能的唯一路径。

A few months later we got a glimpse of AI reasoning for the first time with o1, that pre-training was not the only vector to getting increased performance out of AI.

而在那之后,我们也已经非常清楚地看到,pre-training 根本没有碰到天花板,几乎没人会愿意争辩说 GPT-5 不是和 Opus 3 比起来完全不同的一个怪物。

Now subsequent to that we've also just seen pre-training very clearly not having hit a wall, with pretty much no one being willing to argue that GPT-5 isn't an entirely different beast to, for example, Opus 3.

所以,如果这就是我们在 AI 三年半历史里看到的、投资者最常见的几类恐慌,那现在我们就来看看几个重要的 caveat。

So if these are all the common categories of investor freakouts that we've seen in the three and a half year history of AI, let's now turn to some important caveats.

首先,可能会也可能不会让你感到意外的是,FUD 其实有相当稳定的季节性。

First up, it may or may not surprise you to find that there is a pretty consistent seasonality to the FUD.

这几乎就像投资者想找个理由,在八月少操点心。

It's almost like investors want a reason to not care as much in August.

现在,夏季低迷是市场里一个众所周知的现象,而 AI 热潮似乎又放大了这种效应。

Now summer doldrums are a well-known phenomenon in markets and the AI boom seems to amplify the effect.

这不只是个理论。

This is not just a theory.

有大量研究表明,动能回撤会在夏天发生,尤其是当像半导体这样的动能股在推动这波上涨的时候。

There is a ton of research that momentum breakdowns occur over the summer, particularly when momentum stocks like semiconductors are driving the rally.

而且这次是一次历史性的夏季回调。

And this has been a historic summer breakdown.

上周 Morgan Stanley 指出,这个月动能股已经下跌了 40%,这创下了最差月度纪录,超过了 2021 年初此前最差的 29%。

Last week Morgan Stanley noted that momentum stocks are down 40% this month, making it the worst month on record, beating the previous worst of 29% in early 2021.

当然,这并不是说这一轮回撤全都是市场的一个小怪癖。

Now that's not to dismiss all of this drawdown as a quirk of the market.

Goldman Sachs 在本周早些时候的一份报告里说,对冲基金以前所未有的规模卖出了科技股,但你看到的这些担忧,还是值得放到夏季季节性的背景下去看。

In a note earlier this week Goldman Sachs said that hedge funds have sold tech stocks in record numbers, but it is still worth contextualizing the fears that you see with that summer seasonality.

下面还有一个非常重要的提醒。

Now here's another really important caveat.

我当然会抱怨,有整整一代人看了 The Big Short,觉得 Michael Burry 很酷,然后接下来十年里什么都说是泡沫;但市场总是执着于给一切套上泡沫逻辑,这件事本身,其实是阻止泡沫失控的最大因素之一。

As much as I might bemoan the fact that an entire generation watched The Big Short, thought Michael Burry was cool, and spent the next decade calling everything a bubble, the fact that the market is so determined to have a bubble logic at all times is one of the biggest things preventing a runaway bubble.

每次这个市场有点冲过头,就会有一种压力释放,稍微把它往回拉一点。

Every time this market gets a little over its skis there is a pressure release to moderate it a little.

我们看到的不是 1999 年和 2000 年那种狂热失控的市场,不是隔三差五就有爆发又崩盘的 IPO,带来 100 倍涨幅和巨额亏损。

We're just not seeing the frenetic runaway market we saw in 1999 and 2000 with boom and bust IPOs every other day driving 100x gains and massive losses.

2024 年末那波泡沫讨论之所以消散,是因为大家的末日逻辑变了,而且我们也看到了 Anthropic 和 OpenAI 的数据表现出来了。

The reason why the late 2024 bubble talk dissipated is that agents changed the doomer logic and we saw the numbers show up in Anthropic and OpenAI.

先是合理的担忧,随后又被相反的证据推翻。

Rational concern followed by evidence to the contrary.

现在,说到我认为我们会怎么从这一轮担忧里走出来,我最看好的一个原因,是人们会越来越意识到,像 token 上限这种事,其实远没有“我们最终所需的智能量里,我们现在只用到了极小一部分”这件事重要。

Now when it comes to how I think we drag ourselves out of this round of concern, one of my strongest candidates will be an increasing recognition that things like token caps matter far less than the realization that we're using a vanishingly small portion of the intelligence that will ultimately be in demand.

有多大的空间去扩张,就会给我们多大的空间去扩张。

The room to run will give us room to run.

我也认为,以基础设施扩建实际能达到的速度来看,容量增长速度几乎不可能超过需求增长速度。

I also believe that at the speed at which an infrastructure build-out is even possible, it's almost impossible for capacity to grow faster than demand.

换句话说,建 data centers 真的要花特别特别久的时间,而今年已经宣布的 data centers 里,有一半不是被取消了,就是被推迟了。

In other words, building data centers just takes a really, really long time, and half of data centers that were announced this year have either been canceled or delayed.

看空的人把这说成需求根本没那么大,但这其实完全没有证据;更简单的解释就是,A,拿到许可并建成一个 data center 真的很难,B,到目前为止,data center 建设者在真正回应社区居民关切这件事上,做得惨不忍睹,结果引发了反 data center 活动的高涨。

Bears are calling this a sign that the demand just isn't there, but there is literally zero evidence of this, and the simpler explanation is simply that it's A, really hard to permit and construct a data center, and B, data center builders have done a spectacularly bad job up to this point of actually addressing citizen concerns in their communities, leading to an upswell in anti-data-center activity.

换句话说,这里面本来就有这些内置的减速带,会让事情慢下来,也给每个人更多时间去适应。

In other words, there are just these built-in road bumps that will slow things down and give everyone more time to adjust.

现在,说到这轮关于中国的最新 FUD,我觉得其中一个会逐步化解的方式,就是大家会意识到,如果你没有 inference 去提供服务,那你的 model 再便宜也没什么意义。

Now, when it comes to this latest round of FUD with China, I think one of the ways that it becomes resolved is people realizing that it doesn't really matter if your model is cheap if you don't have the inference to serve it.

Moonshot 在开盘周末就把 compute 完全用光了,而我觉得,把所有中国 AI 公司加在一起,能否支撑哪怕只是一小部分美国公司所服务的用户量,这都极其值得怀疑。

Moonshot was completely tapped out of compute on opening weekend, and I think it is extraordinarily questionable whether all of the Chinese AI companies put together can serve even a tiny fraction of the users that the US companies do.

最后,而且非常重要的一点,看看有多少公司正被这个新机会吸引进来,疯狂涌入——这股力量来自 alternative model architectures 和对更便宜 inference 的追逐。

And lastly, and really importantly, look at the utter explosion of companies flooding into this new opportunity driven by alternative model architectures and the hunt for cheaper inference.

我们几乎每天都能看到 router 被宣布;我们也看到 fine-tuning 和 post-trained model 方向的项目真的在起作用,其中很多还针对特定行业做了垂直化。

We are seeing routers announced every day; we're seeing fine-tuning and post-trained model plays actually working, many of which are verticalized to specific industries.

换句话说,我们看到的是一片繁荣:公司正涌进来解决一个市场机会,而且这还没算上如果美国的 open-source 努力真的出现,可能带来的那些激励。

We're seeing a flourishing, in other words, of companies flooding in to solve a market opportunity, and that's even before we see these potential incentives for US open-source efforts, if those should happen.

投资者 Nick Carter 写道:“美国政府并不欠任何一家 large labs 一个商业模式。如果卖 token 的经济性因为 distillation、廉价克隆、Chinese AI magic 而行不通,美国企业和消费者完全没问题。他们会像其他人一样,受益于 digital cognition 成本的超通缩。hyperscalers 会没事的。只是 OpenAI 和 Anthropic——

Investor Nick Carter wrote, "The US government does not owe either of the large labs a business model. If the economics of selling tokens don't work due to distillation, cheap clones, Chinese AI magic, the American enterprise and consumer will be A-OK. They will benefit from hyper-deflation in the cost of digital cognition just like everyone else. The hyperscalers will be fine. It's just OpenAI and Anthropic

至少按它们现在的形式——不会没事。如果它们愿意适应,就可以开发新的商业模式。token 商人表现不好又怎样?neoclouds 会没事。互联网公司会没事。消费者会拿到更便宜的查询。企业还是会把 AI 纳入流程。AI capex 的超级周期照样会产出 token,不管是 closed weight 还是别的形式。美国公司会消耗这些 token。我的总体 base case 是,

that won't be in their current forms, at least. If they're willing to adapt, they can develop new business models. So what if the token merchants don't do well? The neoclouds will be fine. The internet companies will be fine. The consumers get cheaper queries. The enterprise will still incorporate AI. The AI capex supercycle will still produce tokens, closed weight or not. American firms will consume these tokens. My general base case

在至少未来五年里,每一个 frontier token 都会以溢价被买走,哪怕更便宜的 workload 也在上线。我只是觉得,我们能生产出来的那些溢价的、state-of-the-art token 的数量,仍然会领先于需求,即使同时也会消耗很多非溢价 token。我认为 alternative architectures 和新的 model approaches 的可行性,缓解了

is that every frontier token for at least the next five years gets bought at a premium price, even as cheaper workloads come online. I just think the amount of premium, state-of-the-art tokens that we're going to be able to produce will still be ahead of demand, even with a lot of non-premium tokens being consumed as well. I think the viability of alternative architectures and new model approaches relieves

OpenAI 和 Anthropic 扛起整个市场所承受的更大宏观压力,因为它们把 AI 带来的收入增长分散得更好了。我认为,企业天然的惯性,加上基础设施扩建本身就有的漫长周期,会把适应的时间线拉长,让经济能更好地适应,而不是只看原始 model capability。即便如此,我还是觉得,新的 FUD 会无休无止地接连出现

larger macroeconomic pressure on OpenAI and Anthropic to carry the whole market by better distributing the revenue gains of AI. I think natural enterprise inertia combines with the inherent long horizon of the infrastructure build-out to stretch the adaptive timeline in a way that allows the economy to adapt better than if the only factor was raw model capability. And yet with all of this, I think there will be a never-ending sequence of new FUD

顺带一提,这些 FUD 经常还会和原本就存在的市场季节性撞在一起;但最终,这些相当稳定、很有规律的市场惊慌,反而会降低全面泡沫形成的可能性。总之朋友们,这就是我的看法。这就是我相信的事情,而现在今天的 AI Daily Brief 就到这里。像往常一样,感谢你收听或观看。我们下次再见,peace。

which, by the way, will frequently coincide with pre-existing market seasonality, but also that at the end of the day these fairly consistent and patternistic market freakouts ultimately reduce the likelihood that a full bubble is able to form. Anyways friends, that is my take. That's what I believe, and for now that is going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always. Until next time, peace.
已剔除 2 处广告(点击展开查看)
M1
M10:16广告 · 已剔除

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

To get an ad-free version of the show, go to patreon.com/ai-dailybrief, or you can subscribe to Nōvel podcasts.

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

Now one other quick note, a couple of people have asked me recently if there are better ways that they can support the show.

First of all, you listening is all the support I need.

I appreciate you being here and diving deep with me in this incredibly fascinating and important AI world.

Second, already one of the things that you guys do very frequently, which is incredibly helpful for the show, is just sharing it.

For those who don't know, I recently updated the aidailybrief.ai website to turn every episode into a set of shareable cards.

Some of them are quotes, some of them are themes, some of them are statistics, and each of them can be shared directly to various social media platforms or as an image that you can put anywhere you want.

The most valuable way the show grows is when you share it with new listeners, so that is always appreciated.

Lastly, for whatever reason, Apple and Spotify both very highly reward ratings and reviews, so if you have not at this point left a five-star rating on those platforms, that is helpful as well.

M1
M17:24广告 · 已剔除

Learn more at kpmg.com slash us slash sophisticated, that's kpmg.com slash us slash sophisticated.

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 slash AI Daily Brief.

This episode is brought to you by Retool.

Generating a working app now takes about five 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 slash AI Daily.

Want to accelerate enterprise software development velocity by 5x?

You need Blitzy, the only autonomous software development platform built for enterprise codebases.

Your engineers define the project, a new feature, refactor, or greenfield build.

Blitzy agents first ingest and map your entire codebase, then the platform generates a bespoke agent action plan for your team to review and approve.

Once approved, Blitzy gets to work autonomously generating hundreds of thousands of lines of validated, end-to-end tested code.

More than 80% of the work completed in a single run.

Blitzy is not generating code, it's developing software at the speed of compute.

Your engineers review, refine, and ship.

This is how Fortune 500 companies are compressing multi-month projects into a single sprint, accelerating engineering velocity by 5x.

Experience Blitzy firsthand at blitzy.com.

That's blitzy.com.