美国AI巨头毫无护城河——而中国心知肚明

9月12日，全球最有价值的公司之一的负责人，请求自己的行业放慢脚步。

On September 12, the head of one of the most valuable companies in the world asked his own industry to slow down.

经营Anthropic公司的达里奥·阿莫迪，这家公司制造了名为Claude的AI模型，他写道，监管最强大的AI工具可以大大降低发生严重问题的风险。

Dario Amodi, who runs Anthropic, the company that makes the AI model called Claude, wrote that regulating the most powerful AI tools could greatly reduce the risk that something goes seriously wrong.

大约两周后，他的公司宣布了上市计划，估值超过2万亿美元。

Now a little over two weeks later, his company announced their plan to go public at a price of more than $2,000,000,000,000.

现在，任何想在公开市场出售股票的公司都必须提交一份很长的文件，叫做招股说明书，告诉投资者他们买的是什么，以及可能出什么问题。

Now every company that wants to sell their shares on a public market has to file a very long document called a prospectus, which tells investors what they're buying and what could go wrong.

在Anthropic的招股说明书中，260页里大约有80页是关于风险的。

Now roughly 80 of the 260 pages in Anthropix prospectus were about risks.

在这些风险中，公司警告说，先进的AI可能导致彻底的灾难，甚至可能威胁到人类的生存。

And among those risks, the company warned that advanced AI could lead to absolute catastrophe and could even threaten human survival.

所以让我问你一个问题。

So let me ask you a question.

上一次有公司告诉你它的产品可能威胁到人类的生存，然后还向你要钱，是什么时候？

When was the last time a company told you that its products might threaten the survival of the human race and then ask you for your money.

有趣的是，整个行业都紧跟其后。

Now the interesting part is that the rest of the industry lined up right behind him.

埃隆·马斯克回应说，达里奥说得对。

Elon Musk replied that Dario was right.

山姆·奥特曼，OpenAI的负责人，称赞了阿莫迪的想法，即让外部安全检查员像公司自己的员工一样访问公司的系统。

Sam Altman, who runs OpenAI, praised Amodi's idea of giving outside safety inspectors the same access to a company's system as its own employees have.

他还说OpenAI也会这么做。

And he said that OpenAI would do the same.

然后上周，美国最大的AI和科技公司的负责人齐聚华盛顿，传递了同样的信息。

And then last week, the heads of America's biggest AI and tech companies paraded through Washington carrying the same message.

这项技术太危险了，政府需要介入。

This technology is so dangerous that the government needs to step in.

我觉得这跟保护人类没什么关系。

Now I don't think this has anything to do with protecting humanity.

我认为这完全是为了保护一个没有护城河的业务。

I think it has everything to do with protecting a business that has no moat.

如果华盛顿答应了他们的要求，最终买单的会是你。

And if Washington hands them what they're asking for, you are the one who will end up paying for it.

我们知道这种故事的结局，因为美国以前就干过同样的事。

And we know how that story goes because America has done this exact same thing before.

我们开始吧。

Let's get into it.

这件事跟你有关，哪怕你从来没在聊天机器人里打过一个字，因为现在美国股市全靠这个行业撑着。

Now this matters to you even if you have never typed a single word into a chatbot because the American stock market now rests on this industry.

整个标普500指数大约40%的市值集中在10家公司，其中9家都和AI有关。

Roughly 40% of the value of the whole S and P 500 sits in just 10 companies, nine of which are tied to AI.

在2026年的前三个月，美国经济新增增长中，大约每4美元里就有3美元是由AI行业推动的。

And in the first three months of twenty twenty six, roughly three out of every $4 of new growth in the American economy was driven by the AI industry.

所以我想给你看三件事。

So I wanna show you three things.

第一，什么是护城河，以及没有护城河的企业会怎么样。

First, what a moat is and what happens to a business that doesn't have one.

第二，有证据表明AI公司没有护城河，而且来自中国的新竞争对手让这个问题变得更加严重。

Second, the evidence that the AI companies don't have them and the new competitor from China that makes that problem much worse.

第三，请求政府监管其实只是保护自己企业的一种隐蔽方式，而最终的成本会落到你和我身上。

And third, how a request for government regulation is really just a sneaky way of protecting your business and why that cost will land on you and me.

沃伦·巴菲特让这个术语广为人知。

Now Warren Buffett made this term famous.

城堡的护城河是一圈水，用来阻挡攻击者靠近城墙，而企业的护城河也起到同样的作用。

A castle moat is the ring of water that keeps attackers away from the walls, and a business moat does the same job.

它指的是任何能阻止客户离开你、阻止竞争对手压低你价格的东西。

It is whatever stops your customers from leaving you and stops your rivals from undercutting your price.

而这正是它重要的原因。

And here's why that really matters.

当一家企业获得高额利润时，其他人也会注意到这一点。

When a business makes a big profit, other people notice it too.

他们会开设同样的业务，提供同样的产品，但以更低的价格来争夺这些客户。

They open the same business and offer the same products, but at a lower price to win those customers.

所以第一家生意就得降价来跟上。

So the first business then has to cut its price to keep up.

第二家又降价，这样一直下去，直到大家都没什么钱赚。

The second one cuts again, and this keeps going until nobody is making much money at all.

这就是人们说的，一家的利润是另一家的机会。

Now that's what people mean when they say one company's margin is another company's opportunity.

护城河有几种不同的形式，这里是一些常见的。

Now moats come in a few different shapes, and here are some of the more common ones.

第一，转换成本。

Number one, switching costs.

离开你现在用的生意或服务，要花多少时间、金钱或麻烦？

What would it cost you in time, money, or hassle to leave an existing business or service you are a customer of?

比如说你有一部iPhone。

Let's say you own an iPhone.

想想让你换掉它需要什么。

Think about what it would take to get you to switch.

因为你的照片、信息、应用、密码，还有家庭群聊，全都存在苹果的系统里。

Because your photos, your messages, your apps, your passwords, and the family group chat, they all live inside Apple's system.

而且如果你还有iPad或者MacBook，那手机收到的短信会在笔记本电脑上显示，用iPhone拍的照片也已经同步到iPad上了。

And if you also own an iPad or maybe a MacBook, well, then a text that lands on your phone shows up on your laptop, and a picture you took with that iPhone is already on your iPad.

这种无缝同步确实很方便。

All that synchronicity is convenience.

但竞争对手做的手机和电脑永远不会这样为你服务，因为两家公司都没有理由让对方的产品更好用。

But a phone and a computer made by rival companies will never do that for you because neither company has any reason to make life easier for the other providers.

所以大多数人即使看到竞争对手的手机更便宜，也还是会留在苹果。

So most people stay with Apple even if a rival phone was cheaper.

苹果不需要最低价，因为换系统太麻烦了，大多数顾客宁愿多花钱也不愿折腾。

Apple doesn't need the lowest price because leaving is a hassle that most customers just will pay to avoid.

好的。

Okay.

第二种常见的护城河是网络效应，也就是每增加一个新用户，产品对其他用户来说就更有价值。

A second common mode is the network effect, which is when every new user makes the product itself more valuable to every other user.

那我问你一个问题。

So let me ask you a question.

你听说过多少次有新的信用卡公司席卷全球？

How often have you heard about a new credit card company taking the world by storm?

几乎没听说过，如果有的话。

Rarely, if ever.

商店接受Visa，因为几乎每个顾客都已经有Visa卡。

A store accepts Visa because nearly every customer already carries a Visa card.

顾客持有Visa卡，因为几乎每家商店都已经接受它。

And customers carry Visa cards because nearly every store already accepts it.

新的信用卡公司得同时说服数百万家商店和数百万消费者换用。

A new card company would have to convince millions of stores and millions of shoppers to switch at the same moment.

而且两边都没有理由先行动。

And neither group has a reason to go first.

这真是个大麻烦。

It's a colossal headache.

第三个护城河是成本优势。

A third moat is cost advantage.

想象一下两个铜矿。

Picture two copper mines.

第一个矿的矿石含铜量特别高，我们称之为高品位。

The first sits on rock that is unusually rich in copper, what we call high grade.

第二个矿的矿石含铜量就低多了。

The second sits on rock that holds much less.

要生产一磅铜，第二个矿得挖、运、粉碎更多的矿石，这意味着每卖一磅铜，都要消耗更多的燃料、更多的卡车、更多的工人和更多的电力。

To get one pound of copper, the second mine would have to dig and haul and crush far more rock, which means more fuel, more trucks, more workers, more electricity for every pound it sells.

这就是我们所说的全成本。

This is what we call the all in cost.

从你挖出来的矿石里，实际生产一磅铜的成本是多少？

What does it cost to produce a pound of actual copper from the rock you're digging up?

所以当铜价下跌时，全成本更高的第二个矿会开始亏钱，可能最终停产，而全成本更低的第一个矿仍然在赚钱。

So when the copper price falls, the second mine with the higher all in cost starts losing money and may eventually shut down while the first one with the lower all in cost is still making money.

而且这个优势没人能复制，因为高品位的铜矿不是你想建就能建的。

And nobody can copy that advantage because you can't build high grade copper.

这是大自然几百万年前就决定好的。

Nature decided that millions of years ago.

这里有个有趣的事实。

And here's a fun fact.

大家都在盯着大型科技公司和超大规模数据中心运营商的资本开支，现在每年都快接近一万亿美元了，但大多数人没意识到的是，这些数据中心对金属的需求量巨大。

While everybody is obsessing over the CapEx spends of the major tech companies and hyperscalers now closing in on nearly $1,000,000,000,000 per year, what most people don't realize is that the demand on metal these data centers create is huge.

那笔资本开支里大约有80%都花在了大宗商品上，也就是那些让数据中心真正落地的原材料上。

Roughly 80% of that CapEx spend goes to the commodities, the raw materials that make those data centers a real thing.

包括铜、镍、银、锂、铁矿石，还有更多，这些才让数据中心能运转起来。

That's the copper, that's the nickel, the silver, the lithium, the iron ore, and so much more that make those data centers work.

顺便说一句，如果你对投资这个领域感兴趣，这正是我全职做的事。

Now as a total aside, if you have any interest in investing in that sector, that's what I do full time.

每年一月，我都会在加拿大不列颠哥伦比亚省的温哥华举办一场年度会议，每年大约有11000名投资者到场，来听我请来的100多位主题演讲嘉宾的分享。

And every January, I host an annual conference in Vancouver, British Columbia, and we get about 11,000 investors there every single year who show up to hear from the over 100 keynote speakers that I fly in for the conference.

而且展会上会有超过350家矿业公司，他们会告诉你，他们将供应全世界所需的材料。

And there will be over 350 mining companies on the trade show floor there to tell you they're going to supply the world with the materials the world needs.

所以，加入超过11000名投资者，参加VRIC吧。

So join over 11,000 investors at the VRIC.

详情请访问vricmedia.com。

Check out vricmedia.com for all of those details.

好的。

Okay.

现在回到护城河的重要性。

Now back to the importance of moats.

这里有一个简单的测试，适用于任何企业。

Here's a simple test for any business.

如果明天竞争对手以一半的价格向你的客户提供同样的东西，会有多少客户流失？

If a competitor offered your customers the same thing for half the price tomorrow, how many of them would leave?

如果答案是大多数，那么这家企业就没有竞争护城河，而且这家企业就麻烦了。

And if the answer is most, then that business does not have a competitive moat and that business is in trouble.

那我们就用这个测试来看看那些卖人工智能的公司。

So let's run that test on the companies that are selling artificial intelligence.

现在大多数人想到AI，就是那种聊天窗口，你输入一个问题，然后得到一个答案。

Now most people picture AI as a chat window where you type a question in and you get an answer.

但对于大多数企业或大公司来说，情况有点不同。

But for most enterprise or big business, it's a little bit different.

AI是嵌入到公司自己的软件里的，由像Anthropic或OpenAI这样的公司提供支持。

The AI is built into that company's own software and powered by a company like Anthropic or OpenAI.

但当你使用像Uber这样的应用时，AI在后台悄悄工作，驱动着应用，并根据你的决定和操作进行调整。

But when you use the app like Uber, the AI is working quietly in the background, powering that app and adjusting to your decisions and actions.

AI公司按AI读取和写入的文本量向Uber这样的企业收费。

The AI company bills that business like Uber by the amount of text the AI reads and writes.

这些文本按小块计算，叫做令牌。

And that text is counted in small pieces called tokens.

每个输入的问题和每个输出的答案都会增加令牌费用，就像以前每分钟通话都要计入电话账单一样。

Every question going in and every answer coming out adds tokens to the bill, the way every minute used to add to a phone bill.

现在，关键的部分来了。

Now here's the important part.

像Uber这样的公司的应用，跟所有AI厂商——不管是Anthropic还是OpenAI——交流的方式完全一样。

A company's app like Uber talks to every AI maker, whether Anthropic or OpenAI, exactly the same way.

所以从一个供应商换到另一个很容易，公司也经常这么做。

So switching from one provider to the other is easy, and companies do it all the time.

实际上有一家公司叫OpenRouter，它让企业可以同时接入数百个AI模型，想换就换。

There's actually a company called OpenRouter that lets businesses plug in to hundreds of AI models all at once and move between them whenever they want to.

有一段时间，没人关心切换有多容易，因为没人去算AI到底花了多少钱。

And for a while, nobody cared how easy it was to switch because nobody was counting what AI was costing them.

我来解释一下。

Let me explain.

在AI热潮初期，很多美国公司告诉员工：尽量多用AI，不管需不需要。

During the early AI boom, many American companies told their employees, use as much AI as possible, whether they needed it or not.

这甚至有个外号。

And this even had a nickname.

这被称为“token 最大化”，因为公司消耗的 token 越多，看起来就越现代、越有前瞻性。

It was called token maxing because the more tokens a company burned through, the more modern and forward thinking it looked.

使用 AI 本身就是目标。

Using AI was literally the goal in itself.

但后来账单来了。

But then the bills arrived.

据报道，优步在头三个月就花光了全年的 AI 预算。

Uber reportedly blew through its entire AI budget for the year in just the first three months.

现在，问题的一部分在于 AI 支出很难预测，因为同样的任务每次运行时使用的 token 数量可能不同。

Now part of the problem is that AI spending is very difficult to predict because the same task can use a different number of tokens every time it runs.

你可能也经历过这种情况。

And you've probably experienced a bit of this.

如果你曾经在 AI 平台上输入相同的提示两次，却得到完全不同的结果，那是因为 AI 第二次思考问题的方式不同。

If you've ever put the same prompt in an AI platform two times and gotten a completely different result, That's because the AI thought about the question differently the second time.

这种情况在大规模应用中也会发生，使得 token 的成本非常难以预测。

And this happens at scale as well, making the cost of tokens very difficult to predict.

这确实是个问题。

Now this is an issue.

最近一项调查显示，就在去年春天调查的400家企业中，只有11%能够将AI支出预测准确到10%以内，而超过三分之一的企业误差接近50%。

A recent survey showed that out of 400 businesses surveyed just last spring, only 11% of them were able to forecast their AI spending within 10% of accuracy, while more than a third missed it by nearly 50%.

当像这些大公司一样花费数百亿甚至数千亿美元时，这样的误差就非常严重了。

Now when you're spending tens or hundreds of billions of dollars like these major companies are, that's a really important miss.

于是，公司们开始做每个企业在成本失控时都会做的事情。

So companies began to do what every business does when costs get out of control.

他们把方案推向市场。

They took their offer to market.

他们开始仔细核算账单。

They started counting the bills.

一旦开始核算，他们就开始寻找更便宜的选项。

And once they started counting, they started shopping for cheaper options.

他们发现，大部分工作其实并不需要最先进、最昂贵的模型。

And they learned that they didn't need the most advanced and most expensive model for most of their work.

你不需要每件事都用法拉利。

You don't need the Ferrari for everything.

所以现在，很多公司把昂贵的美国模型留给最困难的任务，其他工作就交给更便宜的选择。

So today, many of them save the expensive American models for their hardest jobs and send everything else somewhere cheaper.

现在，同一家公司OpenRouter追踪了大约12万家同时使用Anthropic和OpenAI的公司的支出。

Now that same company, OpenRouter, tracked the spending of about 120,000 companies that use both Anthropic and OpenAI.

在2026年初，Anthropic占据了这些支出的四分之三，也就是75%。

Now at the start of 2026, Anthropic took about three quarters of that spending, 75%.

但到了9月，又回到了五五开。

But by September, it had swayed back to fifty fifty.

另一个追踪机构，一家叫Ramp的金融公司，观察到它的企业客户在一周之内就在这两家公司之间来回切换，就像打乒乓球一样。

Now a second tracker, it's a finance company called Ramp, watched its business customers switch back and forth between the two companies within a single week, just ping ponging back and forth.

客户们也不避讳原因。

And the customers aren't shy about why.

一家销售软件公司的负责人告诉《华尔街日报》，他把AI工作转向OpenAI是为了省钱，同时也是为了避免对任何单一供应商的依赖。

The head of one sales software company told The Wall Street Journal that he was moving his AI work towards OpenAI to save money, but also to avoid dependence on any one provider.

他说，跟单一模型厂商绑定的蜜月期已经结束了。

He said the honeymoon phase of being tied to one model maker is gone.

另一家叫Retul的软件公司的负责人说得更直白。

The head of another software company called Retul put it even more simply.

我们想找最便宜的模型。

We wanna find the cheapest models.

那这意味着什么呢？

So what does that mean?

我们又回到了价格战的局面。

We're back to a race to the bottom.

还记得那个测试吗？

So remember the test.

一个下午就能走人、不用付任何费用的客户，不是忠诚客户。

A customer who can leave in an afternoon without paying anything is not a loyal customer.

他们只是过客。

They're a visitor.

当你的客户只是过客时，你就只能靠价格竞争。

And when your customers are visitors, you compete on price.

2026年夏天，OpenAI发布了一系列新模型，这些模型的运行成本更低。

In the summer of twenty twenty six, OpenAI released a new family of models built to be cheaper to run.

不久之后，它将这些模型的价格下调了近80%，另一个则下调了近20%。

And shortly afterward, it cut the price of them by nearly 80% and another by nearly 20.

九月底，Anthropic也推出了自己的新系列廉价模型作为回应。

In late September, Anthropic answered with its own new family of cheap options.

而他们最强调的两点就是更低的成本和更高的效率。

And the two things it promoted most were lower cost and better efficiency.

这就是价格战的样子。

Now that's what a race to the bottom looks like.

你降这么多，我就降那么多。

You go this cheap, so I go that cheap.

你来我往，直到没人能赚到钱。

Back and forth until no one's making any money.

而且重要的是要记住，他们在价格上竞争，但这两家公司本来就不赚钱。

And it's important to remember, they're competing on price, but these are two companies who already don't generate any profits.

收入很高，但成本也巨大。

Big revenues, but massive costs.

现在我想说，这个时机真的不能再糟了，因为这两家公司都即将向公众要钱。

Now I wanna say that this timing couldn't really be worse because both of these companies are about to ask the public for money.

目前，Anthropic由创始人、员工和一小部分大型投资者所有。

Right now, Anthropic is owned by its founders, its employees, and a small group of large investors.

上市后，任何有股票账户的人都能买它的股份。

Going public would let anybody with a brokerage account buy a piece of it.

OpenAI也计划采取同样的行动。

And OpenAI is planning the same move.

因此，两家公司都必须说服华尔街，它们拥有持久的业务，价值都超过一万亿美元。

Both have to convince Wall Street, therefore, that they have lasting businesses worth, in each case, more than $1,000,000,000,000.

但想想这样的价格意味着什么。

But think about what a price like that means.

两万亿美元是整个公司的价值，也就是你买下它所有股票要付的价格。

$2,000,000,000,000 is the value of the whole company, the price you'd pay to buy every single one of its shares.

付这个价的投资者并不是在为今年的利润买单，因为Anthropic和OpenAI都没有利润。

And investors paying that are not paying for this year's profits because Anthropic and OpenAI have none.

2025年，Anthropic的收入大约有46亿美元。

In 2025, Anthropic brought in around $4,600,000,000 in revenue.

那是它收到的每一块钱，46亿。

That's every dollar it collected, 4,600,000,000.

但它的运营亏损大约是80亿。

But it posted an operating loss of about 8,000,000,000.

所以投资者现在付的钱，是押注这家公司很多年后能赚到的利润。

So investors are therefore paying today for profits they expect the company to earn many years from now.

这只有在公司能留住客户，并且最终能向他们收更多钱的情况下才说得通。

And that only makes sense if the company can keep its customers and can eventually charge them more.

换句话说，两万亿美元的估值是在赌这家公司有护城河，而你刚刚看到了证据，证明它没有。

In other words, a $2,000,000,000,000 valuation is a bet that the company has a moat, and you've just seen the evidence that it doesn't.

而且这场价格战之所以会变得更加残酷，原因并不来自旧金山。

And the reason this price war is about to get more ruthless is not coming from San Francisco.

你看，到目前为止，这主要是美国公司之间的价格战，尤其是两家公司，Anthropic和OpenAI。

You see, until now, this has been a price war between American companies, and mainly between two, Anthropic and OpenAI.

想象一下，如果第三个竞争者加入这场战斗，在质量上和他们不相上下，但收费却几乎为零。

So imagine a third competitor walks into that fight, matches them on quality, but charges almost nothing.

而这就是今年夏天7月16日发生的事情，一家名为Moonshot的中国公司发布了一个名为Kimi k three的AI模型。

And that's what happened in the summer on July 16 when a Chinese company called Moonshot released an AI model called Kimi k three.

它很快在一个广受关注的排行榜上登顶，这个排行榜用来评估AI模型编写计算机代码的能力，在盲测中击败了Anthropic和OpenAI的最佳模型。

And it quickly took first place on a widely watched leaderboard that ranks AI models at writing computer code, beating the best models from Anthropic and OpenAI in blind tests.

11天后，Moonshot把整个模型放到了互联网上，任何企业都可以免费下载并在自己的电脑上运行，不用付给Moonshot一分钱。

Eleven days later, Moonshot put the entire model on the internet so any business could download it and run it on its computers without paying moonshot a cent.

企业仍然需要支付芯片和电力的费用来运行它，但模型本身是免费发布的。

Now the businesses still have to pay for chips and the electricity to run it, but the model itself was released for free.

所以那些已经表现出会为了更好价格而离开的客户，现在得到了一个几乎和美国最好的模型一样好、却几乎免费的模型，而且很多人正在接受它。

So customers who had already shown they'd leave for a better price were now handed a model nearly as good as the best American ones for almost nothing, and many of them are taking it.

《华尔街日报》最近报道说，很多美国公司越来越多地转向使用像这样的便宜中国模型来处理部分工作。

The Wall Street Journal recently reported that many American companies have increasingly turned to cheap Chinese models like this one for some of their work.

而在OpenRouter上，同样的情况是，美国模型处理的AI流量从一年前的大约70%降到了今年夏天的30%。

And on OpenRouter, that same company, the American models went from handling roughly 70% of the AI traffic a year earlier down to 30% this summer.

这一切都是价格驱动的。

And that was all driven by price.

现在，这对其他所有人，尤其是标普500的每一位投资者来说，都很重要。

Now here's why that matters to everybody else and definitely every investor in the S and P 500.

OpenAI和Anthropic已经承诺在未来几年向大型科技公司支付大约一万亿美元的计算费用。

OpenAI and Anthropic have promised to pay the giant tech companies roughly a trillion dollars for computing power in the coming years.

他们需要这些计算能力来维持他们的商业模式运转。

They need that computing power to keep their business models running.

这两家公司目前都没有盈利。

Neither of those two companies earns a profit.

那他们怎么支付这一万亿美元呢？

So how are they gonna pay a trillion dollars?

他们通过从投资者那里筹集新资金来支付这笔费用。

They pay that by raising new money from investors.

而那些投资者是在为这些公司未来可能成为的更大业务买单。

And those investors are paying for the much bigger businesses they expect these companies to become.

这意味着销售额必须持续快速增长。

That means sales have to keep growing fast.

记住，投资者投资的不是公司现在的样子。

Remember, the investors aren't investing in what the business is today.

他们赌的是公司未来能变成什么样，而这只有在这些企业留住客户、未来能向他们收取更高费用的情况下才能实现。

They're betting on what it could become, and it can only become that if these businesses retain their customers and therefore can charge them more in the future.

现在对AI的需求是真实存在的，任何便宜的国产模型都不会让这种需求消失。

Now demand for AI is super real, and no cheap Chinese model is going to make that disappear.

但它们也不需要这样做。

But they don't need to.

如果更便宜的国产模型不断抢走客户，从而压低价格，那么销售额的增长就会比投资者所期望的要慢。

If cheaper Chinese models keep taking customers and therefore pushing prices down, sales grow more slowly than investors were promised.

增长放缓意味着投资者未来提供的资金会减少，条件也会更苛刻。

And slower growth means investors offer less money in the future and on worse terms.

这样一来，那万亿美元的算力承诺就更难兑现了。

And that makes the trillion dollar promise for computing power harder to pay.

而这万亿美元的承诺是欠微软、亚马逊、谷歌和甲骨文的，他们已经借了数千亿美元，花了大笔钱来兑现这个算力承诺。

And that trillion dollar promise is owed to Microsoft, Amazon, to Google, and Oracle, and they have borrowed and spent hundreds of billions of dollars to make good on that computing power promise.

他们正在建设数据中心。

They're building the data centers.

但他们建这些数据中心、借这些钱，都是基于一个预期，那就是Anthropic和OpenAI会支付那万亿美元的订单。

But they're building those data centers and borrowing that money on the expectation that that trillion dollar order from Anthropic and OpenAI will be paid.

现在，如果这些付款出现疑问，建设就会放缓，资本支出也会减少。

Now, if those payments come into doubt, the building, the CapEx will slow down.

而那些依赖它的行业，从芯片销售到建筑队、发电厂，以及每一个新建的数据中心，都会受到影响。

And the businesses that depend on it from chip sales to construction crews and power plants and every new data center behind it.

这笔支出一直是经济增长的最大推动力，而进行这些投资的公司几乎占了美国股市的一半，而美国人的退休储蓄就放在那里。

That spending has been the biggest driver of economic growth, and the companies doing it make up nearly half of the stock market where Americans keep their retirement savings.

所以现在你知道了，上周那些美国AI和科技公司的CEO们浩浩荡荡走进华盛顿时，真正面临的是什么。

So now you know what was really at stake when that parade of American AI and tech CEOs walked into Washington last week.

但要理解他们想要什么，你得回到大约九十年前，那时美国一个强大的产业正输给一个更便宜的竞争对手。

But to understand what they're asking for, you have to go back about ninety years to the last time a powerful American industry found itself losing to a cheaper competitor.

这些头条新闻往往极其耸人听闻，夸大其词，让你觉得事情会因为错误的原因而迅速发生。

These headlines can be incredibly sensational and hyperbolic and make you think that things are gonna happen really fast for the wrong reasons.

我研究历史是为了寻找先例。

I study history to find precedent.

当我用这个视角来看这些事件时，确实让我感到踏实。

And when I view these events through that lens, I do find it grounds me.

所以让我们回到过去。

So let's go back.

让我们回到铁路时代。

Let's go back to the era of the railroads.

我保证这里有一个重要的教训。

And I promise you there's an important lesson here.

在美国，半个多世纪以来，铁路一直是唯一实用的跨美货物运输方式，这使它们成为华盛顿最大、最富有、人脉最广的公司之一。

For more than half a century in America, the railroads were the only practical way to move goods across America, which made them some of the biggest, richest, and best connected companies in Washington.

然后到了二十世纪三十年代，它们遇到了一个新的、更便宜的竞争对手，那就是卡车。

And then in the nineteen thirties, they faced a new and cheaper competitor, and that was trucks.

因为卡车可以直接在工厂门口装货，然后直接送到商店的地址。

Because a truck could pick up goods right at a factory door and drop them right to a store's address.

它们能去火车到不了的地方，而且收费更低。

They could go places that no train could reach, and they charge less.

铁路公司在价格上赢不了，所以它们推动华盛顿对卡车运输业进行监管。

Now the railroads couldn't win on price, so they pushed Washington to regulate the trucking industry.

当然，它们不会说这是为了保护自己的行业。

Now, of course, they didn't call it protection for their industry.

它们说这些规则会结束它们所谓的恶性竞争。

They said the rules would end what they deemed cutthroat competition.

但更重要的是，这会给道路带来秩序。

But more importantly, this would bring order to the roads.

1935年，华盛顿同意了。

And in 1935, Washington agreed.

从那时起，卡车公司需要政府许可才能运营。

And from then on, a trucking company needed a government permit to operate.

几十年来，新公司经常被拒绝。

And for decades, new companies were routinely turned down.

获准上路的卡车被严格规定能运什么、运到哪里。

The trucks that were allowed on the road were told exactly what they could carry and where they could carry it.

所以很多卡车送完货后，回程却拉不了任何东西。

So many of them would make a delivery somewhere and wouldn't be able to take anything back.

最后只能空车回家。

They'd end up going home empty.

结果，美国人为此多付了近五十年的高额运费。

So Americans ended up paying higher shipping prices for nearly fifty years.

1980年，华盛顿终于放宽了规定，运费立刻下降了大约25%。

Now when Washington finally loosened the rules in 1980, trucking prices immediately fell by about 25%.

现在，这种情况有个名字。

Now there's a name for this.

它叫监管俘获，就是本该被政府管住的公司，最后反而自己来制定规则。

It's called regulatory capture, and it's what happens when companies that a government was supposed to keep in check end up writing the rules instead.

这些规则表面上是对公众的保护，但实际上保护的是那些最初游说政府制定这些规则的公司。

The rules are effectively sold to the public as protection, but they end up protecting the companies who lobbied the government to create them in the first place.

当你在市场上打不过更便宜的竞争对手时，你就让政府替你打败他们。

When you can't beat a cheaper competitor in the market, you ask the government to beat them for you.

而现在，这正是正在发生的事情。

And that's what's happening right now.

美国的人工智能公司以公共安全为名，提出了三种监管要求。

The American AI companies have asked for regulation in three forms, all under the banner of public safety.

第一种是安全测试。

The first is safety testing.

今年七月，Anthropic的CEO呼吁，所有AI模型，无论是美国的、中国的还是其他国家的，在发布前都必须通过政府的安全测试。

In July, Anthropic CEO called for every AI model, American or Chinese or otherwise, to pass government safety tests before it can be released.

听起来挺合理的，但想想实际操作会怎样。

Now that sounds reasonable, but think about what it does in practice.

Anthropic是一家美国公司。

Anthropic is an American company.

它在美国有办公室、律师和员工，能支付测试费用、签署文件，并向监管机构汇报。

It has offices, lawyers, and staff in this country who can pay the tests, sign the paperwork, and answer to the regulators.

而来自海外的免费或更便宜的模型，可能在国内没有任何人能处理这些事。

A free or cheaper model from overseas may not have anybody in the country who can do any of that.

所以，如果美国企业只能用经过测试的模型，那么更便宜的海外模型就会被挡在美国市场之外，而无需任何人写出“禁令”这个词。

So if American businesses can only use tested models, then the cheaper overseas models will be shut out of the American market without anyone ever having to write the word ban.

第二条规则涉及国家安全。

The second rule related to national security.

OpenAI已经要求华盛顿考虑直接禁止中国的AI模型，并将某个实验室描述为受国家控制。

OpenAI has already asked Washington to consider banning Chinese AI models outright, describing one lab as state control.

一旦产品被贴上国家安全威胁的标签，价格就不再重要，因为无论模型多便宜，如果使用它违法，那也没用。

Now once a product is labeled as a national security threat, price no longer matters because it doesn't matter how cheap a model is if it's against the law to use it.

他们推动的第三条规则是针对抄袭问题的。

The third rule they're pushing for is in response to theft.

七月份，白宫指责月之暗面通过抄袭Anthropic的模型来构建Kimi K3。

In July, the White House accused Moonshot of building Kimi k three by copying anthropic's model.

财政部长警告说，作为回应，新的制裁措施正在酝酿中。

And the treasury secretary warned that in response, new sanctions were on the table.

制裁将使美国公司无法与月之暗面做生意，从而切断其模型与美国客户的联系。

Sanctions would make it illegal for American companies to do business with Moonshot, which would cut its model off from American customers.

目前政府尚未公布任何证据，月之暗面也否认了这些指控，但这三点都是值得考虑和应对的合理风险。

Now the government hasn't released any evidence of this claim and Moonshot denies it, but all three are reasonable risk to consider and to address.

但这些规则也恰好保护了那些游说推动它们的公司，并设置了进入壁垒，不仅针对海外竞争对手，也针对资源较少的美国初创企业。

But they also conveniently protect the companies lobbying for them and put up barriers to entry, not just for overseas competitors, but for American startups with fewer resources as well.

重要的是，到九月底，由OpenAI和Anthropic支持的政治团体已经在2026年中期选举上花费了超过5000万美元。

Now it's very important to know that by late September, political groups that were backed by OpenAI and Anthropic had spent more than $50,000,000 on the twenty twenty six midterm elections.

所有这些监管呼吁都建立在公众的恐惧之上。

Now all of this call for regulation rests on public fear.

当人们害怕的时候，规则很容易通过，而没有什么比技术建造者告诉他们这项技术可能会毁灭世界更让人害怕的了。

Rules are easy to pass when people are scared, And nothing scares people quite like the builders of a technology telling them that this technology might end the world.

我想公平地说，有些担忧可能完全是真诚的，强大的AI确实带来真实的风险。

And I wanna be fair here because some of these concerns may be completely sincere, and powerful AI absolutely carries real risks.

但看看同一批公司同时发出的两种信息。

But look at the two messages coming from the same companies at the same time.

对华盛顿和媒体，AI太危险了，政府必须介入。

To Washington and the press, AI is so dangerous that the government must step in.

但对投资者，同样的AI是千载难逢的投资机会，他们应该买一份。

But to investors, that same AI is a once in a generation investment opportunity, and they should buy a piece of it.

一个强大到能毁灭世界的产品，恰好也值两万亿美元。

A product powerful enough to end the world is also conveniently a product worth $2,000,000,000,000.

华盛顿挖的每一条护城河，每一条监管护城河都有代价，而且总是由同一群人买单，那就是客户。

Every moat Washington digs, every regulatory moat has a cost, and it is always paid for by the same people, the customers.

铁路的代价由客户支付了半个世纪。

The railroad's cost was paid for by the customers for half a century.

如果出台规定，把更便宜的中国模型或任何海外模型挡在美国之外，效果也是一样的。

A rule that keeps cheaper Chinese models or any overseas models out of America would work the same way.

美国企业就得按美国的价格来买AI，而他们在欧洲、亚洲或其他地方的竞争对手却可以用更便宜的工具。

American businesses would pay American prices for AI, while their competitors in Europe or Asia or anywhere else could use the cheaper tools.

受保护的公司就能继续维持高价。

The protected companies would keep their prices high.

他们的投资者会继续买入并获利，数据中心建设也会暂时继续下去，但美国经济却要为这项本该提升生产力的技术付出更高的代价。

Their investors would keep buying in and being rewarded, and the data center build out would keep going for now, but the American economy would be paying more for the very technology that is supposed to make it more productive.

护城河能保护城堡，但如果是华盛顿为你挖的护城河，那也意味着里面的人不需要变强，因为他们被保护着。

A moat protects the castle, but a moat that Washington digs for you also means the people inside don't have to get stronger because they're protected.

这里我说的就是AI公司。

And in this case, I'm talking about the AI companies.

铁路公司九十年前就彻底明白了这个道理。

The railroads understood this thoroughly ninety years ago.

最聪明的做法从来不是跟政府对抗，也不是去跟那些自己根本竞争不过的新对手硬拼。

The smartest move was never to fight the government or fight new competitors that they couldn't compete with.

那就是利用政府和他们的资源来为自己谋利。

It was to use the government and their resources to their advantage.

所以下次当AI高管告诉你他们的产品太危险，必须监管时，你得问自己一个非常重要的问题。

So the next time an AI executive tells you that their product is too dangerous to go unregulated, ask yourself a very important question.

这到底对谁有危险？

Who exactly is it dangerous to?

是对我还是对你？

Me or you?

不过说真的，欢迎在评论区告诉我。

But honest question, let me know in the comments.

我是不是漏掉了什么？

What am I missing?

如果你喜欢我的内容，我叫Jay，这里是The JayMartin Show，我每周六都会在这里更新。

If you enjoy my content, my name is Jay and this is The JayMartin Show and I publish here every single Saturday.

如果你喜欢这期内容，帮我个忙。

If you enjoyed this, do me a favor.

点个赞，订阅一下，但最重要的是，把这个视频分享给朋友。

Click like, hit subscribe, but most importantly, share this video with a friend.

你认识的某个人需要看到这个。

Somebody that you know needs to see it.

最后再提醒一下我的年度会议，1月24日和25日在加拿大不列颠哥伦比亚省的温哥华举行。

And one last reminder from my annual conference, January twenty fourth and twenty fifth in Vancouver, British Columbia.

去vricmedia.com查看门票，加入超过11000名投资者和100多位主题演讲者的行列，这些演讲者是我从世界各地请来的，还可以与现场350多家矿业公司中的一家交流，他们会告诉你他们如何准备好为世界提供未来所需。

Check out vricmedia.com for tickets and join over 11,000 investors and over a 100 keynote speakers that I fly in from all over the world and speak with one of over 350 mining companies that will be on the show floor there to tell you how they're ready to supply the world with what they need for the future.

好了。

Alright.

今天就到这里。

That's it for today.

下周见。

See you next week.