人工智能经济有多大？

今天在《AI每日简报》中，我们来聊聊AI经济到底有多大。

Today on the AI Daily Brief, just how big is the AI economy?

在那之前，先看下头条新闻：我们是不是马上就得实名认证才能使用最新的AI模型？

Before that in the headlines, are we about to have to KYC to use the newest AI models?

《AI每日简报》是一档每日更新的播客和视频节目，聚焦AI领域最重要的新闻与讨论。

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

好了，各位。

Alright, friends.

在正式开始之前，先说几个小通知。

Quick announcements before we dive in.

首先，感谢今天节目的赞助商：普华永道、Scrunch、MissionCloud 和 OutSystems。

First of all, thank you to today's sponsors, KPMG, Scrunch, MissionCloud, and Outsystems.

想收听无广告版本的节目？请访问 patreon.com/aideallybrief，或者在苹果播客上订阅。

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

如果你对赞助本节目感兴趣，欢迎发送邮件至 sponsors@aideallybrief.ai。

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

对了，如果你还没看过新版的aideallybrief.ai，建议你去逛一逛。

By the way, if you haven't checked out the new improved aideallybrief.ai, you should check it out.

每期节目不仅有摘要，还被拆成了几个关键数据点，以及十五到二十个可分享的片段，都是按主题整理出的精华内容和经典语录。

Each episode is not only summarized but chunked into the key numbers and a set of between fifteen and twenty shareable chunks that are the key points, key quotes from the show organized by topic.

如果你一直想找办法分享节目里的某个具体部分，就去aideallybrief.ai看看吧。

If you've been looking for a way to share just some specific part of the show, go check it out at aideallybrief.ai.

和往常一样，我们先来看一下MythosFable的最新动态，现在又出现了更多Fable即将重启的迹象，不过这次会加上一些严格的限制。

As per usual, we are starting with our MythosFable watch, where we are getting more signs of a Fable relaunch but with some strict new controls.

AI爆料人M1 Astra在Claude应用里发现了一些新代码片段，透露了Fable重启可能的运作方式。

AI leaker M1 Astra posted some new code strings added to the Claude app giving some hints on how the Fable relaunch might be handled.

首先，Fable的使用似乎会采用积分制，而不是包含在订阅里。

Firstly, it seems that Fable usage will be credit based rather than part of subscriptions.

目前还不清楚Anthropic会不会继续保留试用期，但代码明确显示，Fable的使用最终会单独计费。

It's unclear whether Anthropic will still honor the trial period, but the code strings indicate clearly that Fable usage will be billed separately ultimately.

此外，用户要想使用这个模型，似乎必须向Anthropic提交身份证明文件。

In addition, it appears model access will require users to submit identification documents to Anthropic.

有一段代码提到，你的积分会在身份验证通过后到账。

One code string states your credits will be added once your identity is verified.

对此，不少人感到不满。

And folks are not so happy about this.

海特写道：没有哪个理智的人会为了使用一个限制重重的模型，就把自己的身份信息交给Anthropic。

Hayter wrote, No sensible person is going to give their identity verification to Anthropic just to use a heavily guardrailed model.

虽然我理解海特的立场——毕竟我多年来一直密切关注人们在科技使用中的隐私态度——但我确信，几乎所有人都会把身份信息交给Anthropic，哪怕模型限制再严。

While I understand where Hayter is coming from, having spent a lot of years closely watching people's relationship with privacy when it comes to technology, I am quite sure that basically everyone is going to give their identity verification to Anthropic, even if the model is heavily guardrailed.

事实上，有些人早就认为，只要政府插手，这种情况就不可避免。

Indeed, some believed that this was inevitable as soon as the government intervened.

马克斯·温伯格评论说：我在Fable和Mythos被封禁后四十分钟内就料到会这样。

Max Weinberg commented, I called this within forty minutes of Fable and Mythos getting banned.

这简直就像申请持枪许可证一样。

Seems like the only path forward similar to getting a gun license.

考虑到政府对这些模型的态度，我觉得这个类比还挺贴切的。

Given how the government seems to view these models, I think that's a fair comparison.

顺便说一句，达里奥·阿马代自己也做过这个类比，他说关于Mythos，那些公司都表示：这简直就是个超级武器，用它之前得先拿个持枪证。

By the way, that is a comparison that Dario Amade made himself when he said about Mythos, Companies we gave to it said, this is a super weapon, you should have to own a gun license to use it.

所以说，这沟通水平，真是没得说。

So, you know, just A plus communications all around.

不过现在，这些都还只是基于代码片段的传闻，还没有任何官方确认。

Now, of course, at this stage, this is just rumors based on code snippets so we don't have anything official.

尤其是随着大家期待的Fable可能回归，人们会更加拼命地从各种蛛丝马迹里找线索。

Especially as we get closer to what people anticipate might be the return of Fable, people are going to be looking for any information in the tea leaves.

就像我们昨天聊的，就算Fable真回来了，这场风波的影响也会波及很广，尤其是在华盛顿特区。

Now, as we discussed yesterday, even when we get Fable back, the impact of this whole saga is going to be widespread, not least of which in Washington DC.

在华盛顿，参议员马克·沃纳正准备推出一套针对AI代理的全面新监管框架。

In DC, Senator Mark Warner is preparing to unveil a sweeping new regulation apparatus for AI agents.

那这份法案具体包含哪些内容呢？

So what is in this bill?

首先，法案保障了第三方代理的访问权限。

Well, first of all, the bill protects access for third party agents.

比如，它确保用户可以使用自己的OpenClaw去亚马逊购物，而不是被限制在系统自带的代理里。

For example, it ensures users can send their own OpenClaw to shop on Amazon rather than being locked into the built in agent.

亚马逊和其他平台已经开始尝试各种方式来阻止或限制第三方代理，不过这些措施主要针对的是数据抓取，而不是代理购物行为。

Amazon and other platforms have started experimenting with various methods of blocking or restricting third party agents, although these largely deal with data scraping rather than agentic shopping.

这项法案还确立了‘忠诚义务’的概念，确保代理必须代表用户利益，而不是背后的公司。

The bill also enshrines a concept of a duty of loyalty that ensures agents are acting on behalf of their users rather than the companies that create them.

一位立法助理举了个例子：一个旅行预订代理会优先选择希尔顿，因为和这家酒店集团有未公开的合作关系。

A legislative aide used the example of a travel booking agent preferring to book with Hilton due to an undisclosed partnership with the hotel chain.

这种行为将被新规明确禁止。

This kind of behavior would be barred under the regulations.

沃纳参议员在声明中说：随着代理型AI改变美国人与科技互动的方式，消费者理应在市场上拥有真正的选择权，AI代理必须对其服务的对象负责。

Senator Warner said in a statement: As agentic AI transforms how Americans interact with technology, consumers deserve a real choice in the marketplace, and AI agents must be accountable to the people they serve.

这份讨论草案是迈向建立清晰联邦框架的重要一步，旨在推动创新、保护消费者，并确保美国继续在新兴技术领域引领全球。

This discussion draft is a major step towards building a clear federal framework that promotes innovation, protects consumers, and ensures The United States continues to lead the world in emerging technology.

值得注意的是，这项法案只针对面向消费者的代理，也就是说，它不涉及企业内部的工作流程代理。

Now, it's worth noting that the bill is exclusively concerned with consumer facing agents, meaning it won't deal with internal workflow agents for enterprises.

目前这份法案还只是讨论稿，篇幅也不长，只有25页。

And at the moment, the bill is just a discussion draft and is relatively brief at 25 pages.

它主要要求政府机构根据一些原则制定监管规则，而且大家也普遍认为，今年内这项法案不太可能推进。

It largely instructs government agencies to develop regulations according to certain principles, And there's also an understanding that the bill isn't likely to move forward this year.

一名工作人员承认，这项法案要想进入委员会审议，很可能需要共和党议员共同发起，以及在众议院配套提出类似法案。

A staffer acknowledged that the bill likely needs a Republican co sponsor and a companion bill in the House before it can be put through committee.

但沃纳参议员是华盛顿最有影响力的民主党人之一，我们现在能深入讨论代理行为的具体细节，本身就说明了事情的发展方向。

But Senator Warner is one of the most powerful Democrats in Washington, and and the fact that we are now at the stage where we're getting into the nitty gritty of agent behavior gives you an indication of where things are headed.

现在谈这份法案的具体内容还为时过早，但我对它其实持两种看法。

Now it's too early to get too deep into the substance of this bill, but I'm basically of two minds when it comes to this.

一方面，我们必须密切关注这些监管规定会给代理服务提供方带来多大责任，因为很多初衷良好的提案，最终都成了变相禁止。

On the one hand, it's going to be very important to keep an eye on how much liability these sort of regulations impose on agent providers, as many times we've seen well meaning proposed regulations effectively amount to a backdoor ban.

另一方面，沃纳提出的这种代理中立原则，很可能受到很多人的欢迎，所以也不该轻易否定。

At the same time, these sort of agent neutrality principles that Warner is articulating are probably going to be welcome to lots of folks, and so it's worth not dismissing out of hand.

再来说说政府与人工智能的关系，但这次话题转向了另一个方向：加利福尼亚州谈成了以半价使用Claude的协议。

Now staying in the government's relationship with AI but moving to a very different part of that California has cut a deal for half price Claude.

加利福尼亚州州长加文·纽森宣布与Anthropic达成新协议，将在全州政府范围内推广使用Claude。

Governor Gavin Newsom announced a new agreement with Anthropic to expand the deployment of Claude across the state government.

所有州级部门和地方政府现在都能使用Claude，这是全州首次大规模部署AI工具。

All state departments and local governments will now have access to Claude in the first statewide rollout of an AI tool.

Anthropic还同意提供免费的员工培训和技术支持，价格打五折。

Anthropic has also agreed to provide free workforce training and technical support, all at 50% off.

加州首席信息官兼科技局局长克里斯·吉文表示：很多部门都会转用这个合同，这正是我们的初衷。

California's CIO and Department of Technology director Chris Given said: A lot of departments are going to switch their usage to this contract, and that's very much our intent.

当我们看到大家更倾向于使用某个工具时，我们就希望确保州政府谈到了最优惠的价格。

When we see that folks are going to be using a tool more, we want to make sure that we, as the state, have negotiated the best possible price.

尽管Anthropic与联邦政府关系紧张，纽森办公室表示，这项新合同并非针对华盛顿的回应。

Also, despite Anthropic s ongoing schism with the federal government, Newsom s office said the new contract was not intended as a response to Washington.

纽森也对政府服务中普遍存在的对AI的负面情绪保持警惕，他评论说：AI不该取代政府的人工工作，而应帮助员工更快处理事务、更有效地解决问题，为加州民众带来更好的服务成果。

Newsom was also cautious of negative AI sentiment, especially in government services, commenting, AI should not replace the human work of government, it should help our workers move faster, solve problems more effectively, and deliver better results for Californians.

接下来是另一桩与Anthropic的交易，这次的合作伙伴是亚马逊。

Next up, another deal with Anthropic, and this time it's Amazon.

报道称，Anthropic 已经重新谈判了亚马逊在 130 亿美元投资中获得的优惠条款。

The information reports that Anthropic has renegotiated the sweetheart deal that Amazon locked in as part of their $13,000,000,000 investment.

到目前为止，亚马逊使用 Claude 的费用是按计算时长计费的，相当于批发价。

Until now, Amazon's Claude bill was based on raw computing hours effectively a wholesale rate.

从明年开始的新定价方案中，亚马逊将和其他大型 Anthropic 客户一样，按使用令牌数量付费。

In a new pricing agreement set to begin next year, Amazon will have to pay token based, similar to every other large Anthropic customer.

这次价格调整不仅适用于亚马逊内部使用 Claude，也覆盖了所有基于 Anthropic 模型驱动的产品，比如用于购物的 Alexa。

The pricing adjustment doesn't just apply to internal use of Claude, but also to Amazon products powered by Anthropic's models such as Alexa for shopping.

消息人士称，亚马逊现在正在考虑转向 OpenAI，甚至改用自家研发的 Nova 模型，以节省成本。

Sources said that Amazon is now looking into potential cost savings from switching to OpenAI or even their own in house Nova models.

亚马逊最近对 OpenAI 投资了 500 亿美元，首次获得在其产品中使用 OpenAI 模型的权限，但目前尚无关于该合作定价方式的报道。

Amazon's recent $50,000,000,000 investment in OpenAI allows them to use OpenAI models in their products for the first time, but there's been no reporting on how that arrangement will be priced.

报道还深入挖掘了这两家公司之间日益加剧的矛盾。

The reporting also dug into simmering acrimony between the two companies.

据报道，去年年底，Anthropic 对 Bedrock 平台更新功能太慢感到不满。

Anthropic was reportedly frustrated late last year when new features weren't added to Bedrock fast enough for their liking.

亚马逊方面，报道指出：担心Anthropic的模型未来会变得更贵，一些工程师已经开始主动对其进行精简。

On Amazon's side, the information writes: Fears that Anthropic's models might eventually become more expensive have prompted some engineers to distill them proactively.

消息来源补充说，Anthropic目前仍保有有限的权利，可以使用其模型构建内部使用的小型模型。

The source added that Anthropic does still have some limited rights to use Anthropic models to build their own small models for internal use.

亚马逊的一位发言人否认了这些报道，称：亚马逊与Anthropic的合作关系基于深厚的技术协作，我们仍在持续推动并深化这一合作。

An Amazon spokesperson denied the reporting, commenting: Amazon and Anthropic share a multifaceted partnership grounded in technical collaboration, and we continue to foster that relationship and deepen our work together.

说我们扩展合作会导致成本上升，这是不正确的。

It's incorrect that changes from our expanded collaboration will increase our costs.

Anthropic也声称，他们的服务实际上正在变得更便宜，一位发言人表示：每一代Claude模型完成重要任务的成本都在下降。

Anthropic also claimed that their services are in fact getting cheaper, with a spokesperson for them stating, The cost of getting important work done with Claude falls every generation.

2025年11月，我们大幅下调了Opus的价格，此后价格保持稳定，而模型能力却持续提升，同样的预算每次都能买到更多性能。

In November 2025, we significantly reduced Opus pricing, and that price has held since while the models keep getting more capable, so the same budget buys materially more each cycle.

这些说法，当然都是亚马逊和Anthropic该说的话，但很明显，AI补贴时代的结束，正在彻底改变AI服务的经济格局。

Look, all of those are the things that Amazon and Anthropic are supposed to say, but it's still pretty clear that the end of the AI subsidy era is dramatically changing the economics for AI services.

与此同时，Meta对Codex和Claude的代码使用设定了限制，但原因并非成本问题。

Meta, meanwhile, has placed limits on codex and Claude code, but not because of issues of cost.

根据《The Information》查阅的内部指南，Meta对旗下应用人工智能部门的软件工程师使用主流编码助手设定了严格限制。

According to internal guidelines reviewed by The Information, Meta is placing strict controls on how software engineers in their Applied AI division use the leading coding agents.

应用人工智能团队是Meta新成立的数据标注项目，旨在为前沿人工智能研究收集训练数据。

Applied AI is Meta's recently established data labeling initiative, which seeks to collate training data for their Frontier AI efforts.

一份备忘录要求团队停止在某些任务中使用Codecs和Claude Code，担心模型输出会污染训练数据。

One memo told teams to discontinue the use of codecs and Claude code on certain tasks for fear that model outputs could contaminate training data.

这份文件警告说，这种做法可能引发与合作伙伴公司的严重纠纷。

The document warned that this could lead to serious escalations with partner companies.

今年，Meta一直是Anthropic最大的客户之一，广泛在各个业务环节使用Claude Code。

Now Meta has been one of Anthropic s largest customers this year, pushing Claude code in every corner of their operations.

不过，现在应用人工智能团队必须在不使用AI的情况下解决编程问题，以避免污染训练数据。

However, the applied AI team is now required to solve coding problems without the use of AI to avoid contaminating the training data.

这些限制已广泛覆盖训练流程：员工被禁止使用AI生成用于训练的数据题目，也不允许用AI检查源代码中的漏洞，或根据代码分析生成问题创意。

The restrictions are fairly widespread in training workflows workers have been prohibited from using AI to create programming challenges for use in training data, and they also aren't allowed to use AI to look for bugs in source code or generate ideas for problems based on code analysis.

备忘录指出，这些流程明显属于工程师失去了主导权，我们不希望出现由模型发起的任务。

The memo stated that these workflows fall firmly in the category of the engineering being out of the driver's seat, and we do not want tasks that originated from models.

消息人士称，这些规定是五月推出的，至今仍然有效，表明Meta担心在开发自己的前沿编码模型时，无意中从OpenAI和Anthropic的模型中提炼出知识。

Sources said that these guidelines were introduced in May but remain in place, and they suggest that Meta is worried about inadvertently distilling models from OpenAI and Anthropic as they build their own frontier coding model.

提炼行为违反了两家前沿实验室的服务条款，新政策显示Meta对潜在的法律风险感到担忧。

Distillation violates the terms of service for both frontier labs, and the new policy suggests Meta is concerned about legal exposure.

Anthropic当然一直积极追踪中国实验室的提炼行为。

Anthropic has, of course, been particularly active in documenting distillation efforts from the Chinese labs.

今年二月，他们发布了一份报告，详细记录了提炼攻击；就在本月早些时候，他们还致信国会，指控阿里巴巴大规模使用这种做法。

Back in February, they released a report detailing distillation attacks, and earlier this month, they wrote to Congress complaining that Alibaba had engaged in this practice at a massive scale.

Chubby一语道破：Meta现在正面临每个AI公司迟早都会遇到的问题。

Chubby summed it up: Meta is now facing the exact problem every AI company will soon face.

它希望用自家的内部系统Metacode，取代像Claude Code和Codex这样昂贵的外部编码工具。

It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, Metacode.

但为了打造更出色的编码模型，Meta必须确保不会无意中用竞争对手模型的输出来训练或评估自己的系统。

But to build a better coding model, Meta has to make sure it's not accidentally training or evaluating on outputs from rival models.

这就是提炼的陷阱。

That is the distillation trap.

公司越依赖前沿模型来构建内部AI基础设施，就越难说清楚真正的智能到底来自哪里。

The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

据《金融时报》报道，这还不是Meta最近面临的唯一挑战，报道称谷歌今年早些时候限制了Meta使用Gemini，以应对算力紧张的问题。

Now according to the Financial Times, that is not Meta's only recent challenge, as the Feet reports that Google capped Meta's use of Gemini earlier in the year as a way to deal with a compute crunch.

据消息人士称，谷歌在三月对Meta及其他大型客户实施了使用限额。

According to reports, Google imposed usage limits on Meta and other large customers in March.

消息人士表示，这些限制至今仍在生效，也是Meta停止过度消耗令牌、并鼓励App提升令牌使用效率的原因之一。

Sources said that the restrictions, which remain in place, were part of the reason Meta stopped token maxing and encouraged App to be more token efficient.

其他一些客户也受到了影响，但都没像Meta这么严重，毕竟Meta的令牌需求实在太高了。

Several other clients were reportedly effective, but none to the extent of Meta due to their exceptionally high token demand.

除了用令牌冲排行榜，《金融时报》还提到，Meta用Gemini来自动化部分安全流程，还驱动了一些客服和广告助手聊天机器人。

Aside from burning tokens to top the leaderboards, the Feet reports that Meta uses Gemini to automate some of their safety processes, as well as driving some customer service and advertising help chatbots.

消息人士说，Meta过去用Gemini和Claude，是因为它们比自家的Llama模型表现更好，但最近重点已经转向优先使用Muse Spark——那是Meta今年四月发布的模型。

Sources said that Meta used Gemini and Claude because they were more performant than their in house llama models, but more recently the focus has shifted to prioritize the use of Muse Spark, which was the model that Meta released in April.

关于算力短缺的问题，《金融时报》写道：尽管在芯片、数据中心和电力上投入了数百亿美元，就连最大的科技公司也难以获得足够的算力来应对对先进模型和AI服务的激增需求。

On the topic of the Compute Crunch, the Feet wrote: Despite spending tens of billions of dollars on chips, data centers, and power, even the largest tech companies are struggling to secure enough computing power to support surging demand for advanced models and AI services.

说到这个，AWS 刚刚上调了 GPU 租用的价格。

On that topic, AWS has hiked prices on GPU rentals.

AWS 宣布将 EC2 容量块的价格提高 20%，这会影响那些安排在 NVIDIA GPU 上运行的工作负载。

AWS announced that they would be raising the price for EC2 capacity blocks by 20%, which impacts workloads scheduled to run on NVIDIA GPUs.

这次涨价不会影响使用亚马逊自有 Tranium 芯片的容量块。

The price hike won't affect capacity blocks using Amazon's Tranium chips.

AWS 在 2023 年推出容量块，取代了原先按需租用的 GPU 服务，因为在供应紧张的环境下，那种模式已经难以为继。

Capacity blocks were introduced by AWS in 2023 to replace on demand GPU rentals which were no longer viable in a supply constrained environment.

它们的运作方式类似，但需要提前预订。

They function in a similar way but require reservations in advance.

这次价格调整，又是 AI 推理市场激烈争论中的一个新证据。

Now, the pricing adjustment is another data point in the hotly debated AI inference market.

本月早些时候，你可能还记得，那个广受误解的令牌支出指数连续几周暴跌。

Earlier in the month, you might remember there was that widely misinterpreted token expenditure index crashing for multiple weeks in a row.

我当时就说过，这个指数反映的是令牌路由服务的平均成本，而不是总支出，而且数据来源主要是那些帮客户找更便宜替代方案的公司。不过它确实表明，市场开始转向更便宜的开源令牌了。

As I discussed at the time, the index shows average token costs from token routers, not overall expenditure, and was also sourced from data from the companies whose job was to help their customers find cheaper alternatives, but it did give some indication that the market was starting to turn to cheaper open source tokens.

最近，H100的按需租赁价格大幅下跌，比今年五月的峰值低了40%。

More recently, spot rental prices for H100 fell significantly and are now down 40% from their peak in May.

不过，AWS的这次涨价表明，这个信号可能有点误导性。

The AWS price hike, however, suggests that this could also be a slightly misleading signal.

Semi Analysis指出，他们自己的数据显示，虽然按需GPU租赁价格在下降，但合同价格仍在上涨。

Semi Analysis noted that their own data showed that although spot GPU rentals were falling, contract prices still going up.

他们写道：按需和现货市场主要用于客户跑概念验证、一次性评估、突发性负载或容量溢出。

They wrote: Spot and on demand markets are where buyers run proof of concepts, one off evaluations, burst workloads, and capacity overflow.

这些数据在作为整体数据集的一部分时有一定参考价值，但并不能反映实际生产环境的成本结构。

They can be useful when taken as part of a dataset but are not reflective of where production economics are set.

合同定价反映的是长期稳定的工作负载，这些负载通常对应有计划、可重复、能带来收入的推理或训练需求。

Contract pricing is where sustained workloads show up with the intention of planned, recurring, revenue bearing inference or training demand.

因此，现货价格下跌、合同价格上涨，并不意味着需求减弱。

Falling spot prices alongside rising contract prices are therefore not evidence of weaker demand.

这更可能说明，市场正在从临时性使用转向长期稳定的生产部署。

It is more likely a shift of opportunistic usage towards committed production deployment.

换句话说，他们的结论是：真正的买家正在锁定长期算力，这推高了合同价格。

In other words, they conclude: serious buyers are locking in term capacity and that is pushing contract pricing higher.

最后，今天AI需求推高了内存价格，人们开始寻找‘内存灾难’的替罪羊。

Lastly today, spiking memory prices from AI demand are driving a search for a scapegoat for the Ramageddon.

上周，苹果因内存成本上涨，将多款产品价格上调了高达15%。

Last week saw Apple raise prices on multiple products by as much as 15% due to increased memory costs.

微软迅速跟进，宣布大幅提高Xbox主机的售价。

Microsoft quickly followed suit, announcing a significant price hike for Xbox consoles.

上周的一场会议上，联想表示，内存价格再也不会回到去年的水平，并发布了他们的‘内存灾难生存指南’五步法。

At a conference last week, Lenovo declared that pricing will never return to where it was last year and presented their five step Ramageddon Survival Guide.

现在，有些人把这次价格暴涨当作攻击AI产业的又一个理由，认为高昂的内存成本是在对所有电子产品征税。

Now, some are using the price spike as another reason to hit on the AI industry, viewing high memory costs as a tax on all electronics.

彭博社的乔·魏岑塔尔在最近的一份简报中探讨了这一影响，认为赢得AI竞赛可能需要从非AI经济中抽调真实资源，他称之为‘令人担忧但有可能的路径’。

Bloomberg's Joe Weizenthal pondered the implications in a recent newsletter, suggesting that winning the AI race may require diverting real resources from the non AI economy what he called a worrying but plausible path.

另一些人则把矛头指向内存制造商，指责他们趁机牟取暴利。

Others are looking to the memory producers themselves, accusing them of excessive profiteering.

《华尔街日报》撰文指出，我们正目睹巨额资金从AI提供方——未来可能还包括AI使用者——流向内存芯片制造商。

Writes The Wall Street Journal, We are witnessing an enormous transfer of cash from the providers of AI and perhaps, one day, AI users to the memory chipmakers.

如此大规模的利润转移极为罕见，投资者应当密切关注资金的来源、流向，以及这种趋势能持续多久。

Profit shifts of this scale are rare events, and investors should be paying attention to where the money's coming from, where it's being spent, and how long it will keep flowing.

举个例子，美光在过去三个月里将价格提高了60%以上，过去一年更是翻了四倍，最近还披露其毛利率已达56%，并计划在年底之前达到84%。

Now to give an example, Micron has increased their prices by more than 60% over the past three months and quadrupled them over the past year, and recently reported that they're running at 56% gross margins and targeting 84% margin by the end of the year.

这使得美光的利润率有望位列美国企业第三，仅次于谷歌和英伟达。

That puts them on track to have the third highest profit margin among US companies behind only Google and Nvidia.

就在这样的背景下，《金融时报》报道称，苹果已向特朗普政府申请许可，希望从中国供应商CXMT采购内存，而这家公司目前因涉嫌与中国军方有关联，被列入五角大楼的黑名单。

And it's in this context that the Financial Times reported that Apple has petitioned the Trump administration for clearance to buy memory from Chinese supplier, CXMT, a company that is currently on the Pentagon's blacklist due to alleged ties from the Chinese military.

此外，本周加州还出现了一起集体诉讼，指控三星、SK海力士和美光串通操纵内存市场，抬高消费者价格。这一话题超出了本期节目的讨论范围，至少目前先不深入，但它清楚地表明，内存短缺的形势正变得多么严峻。

Separately, a class action lawsuit was filed in California this week alleging that Samsung, SK Hynix, and Micron conspired to run a memory cartel to inflate consumer memory prices, which is beyond the scope of this show to go deep into, at least for now, but is a good indicator of just how severe the memory shortage is becoming.

今天本可以聊更多头条新闻，但我们已经超时了，就先到这里，接下来进入正题——聊聊AI经济的真实规模。

Now there is even more we could get into today with the headlines, but we are already way over time, so let's close it there and head into the main episode where we are going to talk about the true size of the AI economy.

当前最重要的AI问题，不是谁在用AI，而是谁用得更好。

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

KPMG和德克萨斯大学奥斯汀分校刚刚分析了140万次真实职场中的AI互动，发现了一个令人意外的现象：最有影响力的人并不是更擅长写提示词的人，而是把AI当成一个思维伙伴来合作。

KPMG and the University of Texas at Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会提出问题、引导思考、反复打磨，并不断追求更好的答案。

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

好消息是？

And the good news?

这些行为是可以大规模培养的。

These behaviors are teachable at scale.

如果你正想从单纯使用AI转向真正掌握AI能力，KPMG关于高级AI协作的研究非常值得一看。

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

了解更多，请访问 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated.

网址是 kpmg.com/us/sophisticated。

That's kpmg.com/us/sophisticated.

问你个问题：你上一次真正打开网站查资料，是什么时候？

Quick question: When was the last time you actually visited a website to research something?

如果你和我一样，现在AI已经帮你把这事全干了。

If you're like me, AI pretty much does that work for you now.

这自然给品牌带来一个新问题：如果AI负责发现、研究和决策，你的网站到底为谁而存在？

That of course raises a new question for brands: If AI is doing the discovering, researching, and deciding, who or what is your website really for?

用户行为的这种转变——AI机器人成为你最重要的新访客——正是我的赞助商Scrunch正在全力应对的挑战。

That shift in user behavior, the rise of AI bots becoming your most important new visitors, is what my sponsor Scrunch is taking head on.

Scrunch是一个AI客户体验平台，帮助营销团队了解AI代理是如何使用你们网站的，它们在AI回答中出现在哪里、没出现在哪里，以及是什么阻碍了AI抓取、信任或推荐你的内容。

Scrunch is the AI customer experience platform that helps marketing teams understand how AI agents experience their site, where they show up in AI Answers, where they don't, and what's preventing them from being retrieved, trusted, or recommended.

这不仅仅是曝光度的问题。

And it's not just visibility.

Scrunch能帮你发现那些真正关键的内容缺口、引用缺失和技术障碍，并指导你如何修复，让你的品牌在AI回答中被找到、被选择。

Scrunch shows you the content gaps, citation gaps, and technical blockers that matter, and helps you fix them so your brand is found and chosen in AI Answers.

为我们的听众特别提供：Scrunch免费为你做一次网站审计，揭示AI如何看待你的网站，哪里存在缺口，以及你在AI回答中的表现相比竞争对手如何。

Now for our listeners, Scrunch is providing a free website audit that uncovers how AI sees your site, where there's gaps, and how you're showing up in AI versus the competition.

立即访问 scrunch.com/aidaily，免费检测你的网站。

Run your site through it at scrunch.com/aidaily.

今年，平均一家大型企业会在人工智能上投入1150万美元，但大多数人根本无法证明这笔钱带来了任何回报。

The average enterprise is spending $11,500,000 on AI this year, and most of them can't prove a single dollar came back.

当人工智能真正带来投资回报时，它到底是什么样子的？

What does AI actually look like when it produces ROI?

问问那家把支付处理速度提升了32倍的医疗公司，或者那家把文件研究时间从三个月缩短到十分钟的律师事务所，再或者那个把等待时间减少了99%的客服中心。

Ask the healthcare company that just made their payment processing three twenty times faster, or the law firm whose document research went from three months to ten minutes, or the contact center who reduced wait times by 99%.

这些都是MissionCloud的真实客户，取得了实实在在的成果。

These are real MissionCloud customers with real results.

MissionCloud是CDW旗下的公司，也是AWS的顶级合作伙伴。

MissionCloud is a CDW company and an AWS premier tier partner.

他们是真正以人工智能为先、专注于业务成果的AWS专家，专门打造能推动企业发展的AI解决方案。

They're the AI first outcomes obsessed AWS experts who build AI solutions that drive your business forward.

无论你是满脑子AI想法却不知从何下手，还是已经投入六个月但项目偏离了方向，他们全都见过，也全都解决过。

Whether you're flooded with AI ambitions but no idea where to start, or six months into a deployment that's going sideways, they've seen it and they've fixed it.

别再把预算浪费在那些无法产生实际效果的人工智能项目上了。

Stop burning your budgets on AI that doesn't produce results.

访问 missioncloud.com 开始吧。

Start at missioncloud.com.

本集《AI每日简报》由Outsystems赞助播出，这是一家专为企业打造的领先智能系统平台。

This episode of the AI Daily Brief is brought to you by Outsystems, a leading agentic systems platform built for the enterprise.

全球各地的企业都在使用Outsystems平台构建、协调和管理智能系统，这绝非偶然。

Organizations all over the world are building, orchestrating, and governing agentic systems on the Outsystems platform and with good reason.

Outsystems开放且统一的平台，让团队能够灵活地设计、交付和扩展受控的智能系统。

Outsystems open and unified platform allows teams to architect, deliver, scale governed agentic systems with agility.

无论团队规模大小或技术背景如何，都可以借助Outsystems快速、经济地开发、部署和管理AI应用与智能代理，同时保障可靠性和安全性。

Teams of any size and technical depth can use Outsystems to build, deploy, and manage AI apps and agents quickly and cost effectively without compromising reliability and security.

有了Outsystems，你可以从想法出发，迅速将创意变为现实。

With Outsystems, you can rapidly launch ideas from concept to completion.

作为领先、统一、敏捷且经过企业验证的智能系统平台，Outsystems助你加速增长、降低运营阻力，真正实现AI对企业级价值的驱动。

It's the leading agentic systems platform that is unified, agile, and enterprise proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI.

Outsystems，构建你的智能未来。

Outsystems, build your agentic future.

欢迎回到《AI每日简报》。

Welcome back to the AI Daily Brief.

今天我们要探讨一个关于AI最重要也最引人深思的问题：AI经济到底有多大？

Today we are talking about one of the most interesting and consequential questions around AI, which is just how big is the AI economy?

这个问题可不只是出于好奇或表面的数字而已。

Now this question is much more than a vanity question.

自从2022年底ChatGPT引爆市场，围绕AI的经济活动迅速爆发，人们就开始怀疑：我们是不是正处在一场注定要破裂的泡沫中？

Ever since the chat chippy tee moment in late twenty twenty two and the corresponding explosion of economic activity surrounding AI, there have been questions about whether we were actually in some sort of bubble that inevitably would pop.

但相比其他技术，AI在实用性方面确实表现得相当出色，基本避开了‘是不是泡沫’的质疑。

Now relative to other technologies, frankly, AI has done a pretty good job of outrunning the bubble question at least when it comes to its utility.

换句话说，尽管少数怀疑者一直持保留态度，但大多数人已经公认AI非常实用。

In other words, as much as a handful of skeptics have tried, by and large folks accept that AI is extremely useful.

不过，很多市场历史学家都会提醒：一项技术再有用，也照样可能形成金融泡沫。

But as many market historians will point out, a technology can be extremely useful and still produce a financial bubble.

这两者之间其实并没有必然联系。

These two things don't necessarily have anything to do with one another.

这正是为什么到了2025年，关于泡沫的讨论声浪变得越来越大——因为企业大幅扩大了AI基础设施的建设规模。

This is of course why questions about a bubble got much louder throughout the course of 2025 as companies significantly increased the size of their AI infrastructure build outs.

一开始，超大规模云服务商只是用自由现金流来投资数据中心和其他基础设施，但随着这些现金流逐渐耗尽，他们开始转向其他资金来源，比如借贷和信贷。

Now at first, the hyperscalers were just using their free cash flow to finance data centers and other infrastructure, but as that free cash flow has been used up, they've turned to other sources including debt and credit.

目前，关于泡沫的议论暂时平静了一些，但这种平静从来都持续不了多久。

Now, right now, we're on a little bit of a low ebb when it comes to the bubble talk, but that never lasts for that long.

正因如此，现在反而是个绝佳时机，让我们冷静下来，真正搞清楚当前AI经济到底有多大——换句话说，如今围绕AI产生的经济活动，有多少是基于现实数据，而不仅仅是未来的潜力？

And in that vacuum, it's a really good time to actually try to somewhat dispassionately understand just how big the AI economy is right now, or putting it differently, how much economic activity around AI is justified not only by the future potential but by the numbers today.

这时，Exponential View团队发布了份很棒的报告，名叫《AI经济现状》。

Enter the team at Exponential View who have just released a great report called The State of the AI Economy.

他们梳理了超过一千家AI公司的相关报告。

They went through reports surrounding over 1,000 AI companies.

他们为不同信息来源打了可信度评分——经过审计的财务数据比高管公开言论更有分量，同时他们还去除了重复计算，确保AI支出只算一次。

They gave a confidence score to different sources, meaning that actually audited accounts count for more than public comments by executives, and they made sure to deduplicate so that AI spend was only counted once.

举个例子，一个应用花了100美元，其中60美元给了模型提供商，30美元用于推理托管，这总共只算100美元，而不是190美元。接下来，我们先看大数字，再深入细节。

In other words, dollars 100 in app spend that sends $60 to a model provider and $30 on inference hosting is counted as $100 not $190 So let's talk about the big numbers, and then let's get into the details.

他们指出，与以往的技术平台转型相比，人工智能的需求如今有更明确的收入数据作为支撑。

The top line, they say, is that AI demand is more clearly validated by realized revenue than previous platform shifts.

换句话说，需求真实、规模庞大、增长迅猛。

Or as they put it, demand is real, big, and fast.

事实上，这个行业的发展速度是以往任何一次IT浪潮的三倍。

Indeed, the sector, they say, is growing three times faster than any IT wave before it.

最引人注目的是，过去一年里，人工智能企业累计赚取了1100亿美元，目前年化收入已达1750亿美元。对比一下第三年的收入水平，增长速度确实惊人。

The big headline number is that AI companies have banked $110,000,000,000 over the past twelve months and are at an annualized run rate of $175,000,000,000 Compare that to year three revenue, and the revenue is growing fast.

他们提到，2023年时，人工智能行业还需要180天才能新增10亿美元的累计收入。

They point out that back in 2023, the AI industry needed one hundred and eighty days to add a billion dollars in cumulative revenue.

如今这个速度提升了90倍，每新增10亿美元收入，不到两天就完成了。

It has now gotten 90 times faster at that, needing less than two days to add each new additional billion dollars of revenue.

更重要的是，正如本节目常听的听众所知，对核心产品的需求还带动了超出核心产品本身的经济价值，这就是《Exponential View》所说的‘计算超级周期’。

And what's more, as regular listeners of this show will know, demand for the core product has also launched economic value beyond the core product what The Exponential View dubs a compute supercycle.

2026年的预测显示，全球半导体市场今年的收入将达到1.5万亿美元，几乎是去年7920亿美元的两倍。而对计算能力的需求，也重新激活了他们所说的长期低迷的半导体产业。

2026 projections have the global semiconductor market reaching $1,500,000,000,000 in revenue this year, basically doubling from last year's seven ninety two billion dollars And that demand for compute also has secondary effects, as AI demands reignites what they call a moribund U.

美国

S.

电力行业

Power sector.

从1950年到2008年，美国

Between 1950 and 2008, U.

美国

S.

电力净发电量平均每月增长六太瓦时。

Electricity net generation grew at six terawatt hours per month on average.

但从2008年到2024年，全球金融危机之后，美国的电力净发电量基本停滞，没有新增任何发电量。

Between 2,008 and 2024, however, post global financial crisis, we have effectively been flat, with no new net US electricity generation.

而从2024年到现在，我们看到年增长率达到了历史平均水平的150%，每月新增发电量达到九太瓦时。

Between 2024 and today, we're now seeing annual growth at 150% the historical average, reaching nine terawatt hours per month in annual growth.

不过，我们还是来深入聊聊这些资本支出和当前收入之间的关系。

But let's talk more about the relationship between all this capex spend and current revenue.

他们总结说，眼下这场史上规模最大的科技基建投资正在开始回报。

The way that they sum it up is that the largest buildout in tech history is paying back for now.

今年，超大规模云服务商和新兴云平台的资本支出将达到8480亿美元，自2020年以来累计已达2万亿美元。

Hyperscaler and Neo Cloud CapEx will reach $848,000,000,000 this year, and $2,000,000,000,000 cumulatively since 2020.

尽管大部分资金仍来自企业自有现金，但债务等外部融资渠道的占比正在明显上升。

And while the majority is still coming from balance sheet cash, external funding sources like debt are definitely on the rise.

至于当前的收入状况，报告指出，收入已经能覆盖当期开支，但还无法弥补累计投入的总成本。

Now in terms of the revenue picture today, the report argues that revenues are covering the ongoing expense but not yet the cumulative bill.

从去年第四季度开始，每个季度的收入首次超过了资本支出的折旧部分。

Starting in Q4 of last year, quarterly revenues started to exceed capex depreciation.

更值得注意的是，租金回报率显示，大量基础设施的实际收益远超其长期折旧预期。

And what's more, rental yields suggest that a lot of infrastructure is outperforming its long term depreciation expectations.

这曾经是关于泡沫争议的核心问题之一：有人质疑，如果企业采购的这些GPU一上线就过时了，那它们哪有足够时间收回投资？

This has, at times, been one of the key questions around the bubble, with people arguing that if all these GPUs that companies are buying are out of date immediately, that doesn't give a lot of time for them to make a return on their investment.

但数据表明，较旧的GPU在超过六年折旧期后，依然能持续产生收益，甚至在第七年、第八年，乃至第九年都还能带来可观回报。

However, what the data suggests is that older GPUs are earning yields long beyond their six year depreciation life, seeing meaningful gains all the way into year seven, year eight, and even year nine.

这显然创造了一个更健康的经济环境，有助于逐步偿还这些资本支出。

This obviously creates a much healthier economic scenario in which to pay back all that capex.

不过报告也指出，我们现在还处于非常早期的阶段。

Now as the report points out, we're still really early.

尽管基础设施投入巨大，人工智能的收入仍有很大的增长空间。

Despite all this infrastructure cost, AI revenue still has a ton of room to grow.

举个例子，信息技术行业占美国GDP的9.4%左右。

For one comparison, the IT sector represents around 9.4% of US GDP.

而人工智能的收入，目前只相当于美国GDP的0.42%。

AI revenue, meanwhile, is equivalent to 0.42% of US GDP.

不过这些数字正在快速增长。

These numbers are growing quickly, however.

与2025年第一季度相比，人工智能收入占GDP的比例已经增长了三倍，与2024年第一季度相比更是增长了十倍。

AI revenue relative to GDP has written 3x versus Q1 of twenty twenty five and 10x versus Q1 of twenty twenty four.

尽管我们现在讨论的焦点大多是令牌效率、令牌上限这些话题，但报告指出，人工智能的支出相对于它未来的潜力来说，仍然很小。

And even though so much of our discussion right now is around token efficiency and token caps and things like that, the report points out that AI spending is still relatively small relative to what it might be.

他们举的例子是优步每位工程师消耗1.5千个token，但正如他们指出的，这点消耗在公司的利润表上几乎微不足道。

The example they gave is Uber's 1.5 ks per engineer, which, as they point out, barely dense that company's P and L.

不仅AI的消耗还有很大增长空间，而且它实际上已经在快速增长了。

And not only does AI consumption have room to grow it very clearly is.

报告指出，从聊天转向智能代理正在大幅增加token的使用量，比如一个代理编程任务的token量大约是普通聊天任务的1200倍。

The report points out that the transition from chat to agents is multiplying token use, pointing out that an agentic coding task can have around 1,200 times the tokens of a chat task.

全球每月的token总量现已超过30千万亿，同比增速达到14倍。

Global token volumes are now above 30 quadrillion per month, and are growing 14x year over year.

有趣的是，尽管由于token消耗增长，整体AI支出在上升，但每个单位token的成本却在下降。

Now what's interesting is that even though overall AI spend is going up because of that growth of token consumption, the cost of tokens on a unit basis is going down.

从2024年年中到2026年年中，尽管AI的能力指数从112提升到了158，

Between mid twenty four and mid twenty six, despite the epic capabilities index of AI going from one twelve to one fifty eight, I.

E.

E.

也就是说，AI现在能做的事情多得多，但每百万token的综合价格却从17美元降到了2美元，而每个输出token所处理的输入token数量——也就是平均请求的复杂度——也从12跃升到了36。

Meaning AI can do a lot more now, the blended price per million tokens went from $17 to $2, and the tokens processed per output token, a measure of the intensity of the average request, jumped from 12 to 36.

正如他们所说，价格下降推动了更多使用，让以前不划算的应用变得可行。

As they point out, price declines encourage more use and make previously uneconomical applications viable.

回到资本支出来看，这种新的效率正在提升电力的变现能力。

And bringing it back to CapEx, this new efficiency is increasing electricity monetization.

尽管每个令牌的收入在下降，但每吉瓦电力的变现收入却在上升。

Even though revenue per token is falling, energy monetization per gigawatt is increasing.

自2024年年中以来，每吉瓦装机容量产生的收入基本翻了一番。

Energy monetization has basically doubled since mid-twenty twenty four in terms of the amount of revenue generated per each gigawatt of capacity.

尽管这需要全面调整，但报告认为，基于令牌的定价是AI经济下一阶段的关键，就像数字广告从无法追踪的横幅广告转向按点击付费的可归因广告那样——后者让数字广告年收入从2002年的50亿美元增长到2024年的上千亿美元。

And as much as it requires a broad adjustment, the report argues that token based pricing is key to the next step of the AI economy, comparing it to the moment in digital advertising when we went from untracked banner ads to attributable ad spend with pay per click, which grew annual digital ad revenue from about $5,000,000,000 in 2002 to way over $100,000,000,000 by 2024.

至于价值集中在哪些环节，目前收入仍主要来自芯片，但结构已经开始变化。

Now, in terms of where value is accruing, revenue remains concentrated around chips, but the mix is starting to shift.

由托管服务带来的AI收入占比正在上升。

The portion of overall AI revenue that is from hosting is going up.

基础模型的收入显然也在增长。

Obviously, Foundation Models revenue is going up.

而且第一次，像Cursor这样的公司的应用收入也开始显现了。

And for the first time, app revenue, such as from companies called Cursor, is showing up as well.

事实上，报告认为，价值正向上游向应用和模型层转移。

Indeed, the report argues that value is moving up the stack towards apps and models.

过去一年里，来自应用和模型层的AI收入占比几乎增长了三倍。

The percentage of AI revenue that comes from the app and model layer was up almost 3x over the last year.

价值链条的构成目前仍很不明朗。

And the value stack mix is still very unresolved.

报告提到一点，相信大家都很熟悉：尽管代币成本面临降价压力，各大实验室却在同时向下渗透到基础设施、向上拓展到应用层，我预计这类调整和探索还会持续一段时间。

One of the things the report points out, which will be familiar to all of you guys, is that even as there's pricing pressure around token costs, labs are increasingly pushing both down the stack into infrastructure and up the stack into apps, and I would expect these sorts of shifts and experiments to continue for some time to come.

从商业角度来看，企业正在报告生成式AI带来的影响日益显著。

Now in terms of justifying all this spend from a business perspective, public companies are reporting increased impact of Gen AI.

在财报电话会上提及AI影响的公司比例，已从2023年初的约10%上升到如今的三分之一，即33%，其中更有整整20%的公司给出了具体的数据量化。

The percentage of companies making claims of AI impact on earnings calls has jumped from around 10% in early twenty twenty three to a third at 33% today, with now a full 20% of companies making quantified claims on their earnings calls.

虽然目前这些说法（七成到十成）大多集中在成本节约或效率提升上，但最能体现高AI投入企业优势的指标，还是那些完全没有投入AI的公司与高AI投入公司之间的收入增长差距。

And while most of those claims right now (seven and ten) focus on either cost savings or efficiency, the most dramatic indicator of the comparative success of high AI adopters comes in the comparison of revenue growth between companies with no AI spend and companies with high AI spend.

那些没有投入AI资金的公司，过去三年里的营收增长基本跟美国名义GDP的增速差不多，大概在15%到20%之间。

Companies with no AI spend have grown revenue over the last three years pretty close in line with US nominal GDP between 1520%.

而AI投入强度高的公司（比

Companies with high AI intensity (i.

如

E.

在按营收占比排名的AI支出前25%的公司），在同一时期内的营收增长了超过100%。

Those who are in the top 25 of AI spenders by share of revenue) have seen their revenue grow in that same period by more than 100%.

换句话说，高AI投入企业和零AI投入企业之间的营收增长率差距达到了92%。

In other words, there is a 92% revenue growth differential between high AI spenders and no AI spenders.

他们的结论是：AI的需求比以往任何一次平台变革都更有营收验证。

The conclusion, they write: AI demand is more revenue validated than any prior platform shift.

归根结底，他们表示，投资逻辑取决于价格下降能否带动足够的Token用量，从而让资本支出获得回报。

Ultimately, they say the investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx.

但正如你刚才听到的，许多指标其实比普遍舆论所认为的要乐观得多。

But as you just heard, a lot of the indicators of that are much more positive than the average discourse suggests.

所以朋友们，这就是AI经济的规模。

So friends, that's how big the AI economy is.

年营收已经达到了1750亿美元。

It's on $175,000,000,000 annual run rate.

它的增长速度是以往任何平台转型的三倍。

It's growing three times faster than previous platform shifts.

它正在推动计算超级周期和能源基础设施建设的大幅衍生增长。

It's causing massive secondary growth in the compute super cycle and in the energy buildout.

许多基础设施所需原材料的经济价值，已经超过了原先预期的生命周期。

Many of the raw resources of that infrastructure buildout are seeing economic value past their expected time horizon.

这并不是说市场不会变得过度狂热。

None of this is to say that the market can't get overexuberant.

但我依然认为，关于AI泡沫最深刻的那条推文，是去年十月OpenAI的Rune发的：没人真正做好心理准备，万一这根本不是泡沫呢。

But I continue to think that the most insightful tweet ever about an AI bubble came from OpenAI's Rune back in October of last year, when he tweeted: Not enough people are emotionally prepared for if it's not a bubble.

今天的AI每日简报就到这里。

That's gonna do it for today's AI Daily Brief.

感谢 Exponential View 团队的出色工作。

Great work to the team at Exponential View.

详情请见节目说明中的原始报告链接，一如既往，感谢收听或观看。

Check our show notes for a link to the original report, and thanks for listening or watching as always.

我们下次再会。

Until next time.

再见。

Peace.