安德鲁·费尔德曼谈打造比英伟达大58倍的芯片

对人工智能的需求已经远远超出了所有人的预期和预测。

The demand for AI has outpaced everybody's expectation, everybody's forecast.

所以大家都在拼命追赶。

And so everybody's chasing.

他们争抢芯片、内存，还有数据中心。

They're chasing chips, they're chasing memory, or they're chasing data centers.

今年一月，我们和OpenAI宣布了一项重大合作，这是硅谷历史上最大的交易之一。

OpenAI, we announced in January a huge partnership, one of the biggest deals done in Silicon Valley history.

未来几年，我们为他们提供的硬件设备总额将超过两千亿美元。

We'll be doing more than $20,000,000,000 of hardware for them over the next several years.

我担心，英伟达经常有办法利用其市场优势。

I'm concerned that there are ways frequently for NVIDIA to exercise market strength.

他们完全可以用自己的资产负债表来压制竞争对手。

Are ways for NVIDIA to use their balance sheet to limit competition.

我们有机会让下一代，甚至我们的孩子，不仅不再死于癌症，甚至根本不知道有人曾因癌症去世。

We have a chance for our children or the next generation not only to not die from cancer, but to not know anybody who died from cancer.

安德鲁·费尔德曼，欢迎来到《Sorcery》。

Andrew Feldman, welcome to Sorcery.

非常感谢你邀请我。

Well, thank you so much for having me.

我们现在在巴黎的RAISE活动现场。

We are out here in Paris at RAISE.

还行吧。

Pretty okay.

确实还行。

It's pretty okay.

我还听说你是这儿最受欢迎的人之一。

I also heard you're one of the most popular people here.

这可是我这辈子头一回。

That would be first in my life.

你这么说，我真不好意思。

So it's nice nice of you to say.

不过，一个很大的话题是——我们刚才在镜头外也聊到，硬件现在特别酷。

A big theme, though, is and we were talking about this just off camera is hardware is cool.

半导体现在特别火。

Semiconductors are cool.

现在什么都赶时髦。

Everything is in fashion now.

我们也在潮流里。

We are in fashion.

我们现在供不应求。

We're in demand.

想找到我们可不容易。

We're hard to get.

这变化真不小。

It's a big change.

我觉得是因为AI的需求突然超过了所有人的预期和预测，所以大家都在拼命追赶。

I think what's happened is the demand for AI has sort of outpaced everybody's expectation, everybody's forecast, and so everybody's chasing.

他们在抢芯片，或者抢内存，或者抢数据中心。

They're chasing chips, or they're chasing memory, or they're chasing data centers.

这正是发明新东西、新架构的好机会。

There's an opportunity to invent new things, different architectures.

现在真是思想爆发的年代，特别有意思。

It's really a time of sort of a Cambrian explosion of ideas, and that's really fun.

你去年参加过RAISE活动吧。

So you were at RAISE last year.

我记得，因为你当时摊位是最大的。

I remember this because I think you had the biggest booth here.

周围还围了栏杆。

There was a fence around it.

是的。

Yeah.

而且人特别多。

And a lot of people.

去年和今年最大的区别是什么？

What is the biggest difference between last year and this year?

哦，去年应该是他们第一年参加吧？

Oh, think last year was their first year, right?

我觉得这里的人更多了。

And I think there are more people here.

参会的人数增加了。

There are more participants.

出席的人也更多了。

So are more attendees.

参与的公司也多了。

There are more companies participating.

我觉得大家的参与程度更深了。

I think people are participating in a bigger way.

我觉得昨晚的晚宴上

I think the dinner last night at

秘密地点。

Secret location.

在名叫凡尔赛的一个秘密地点，感觉特别棒。

At a secret location called Versailles was sort of awesome.

我们整天泡在数字世界里，可我们讨论人工智能的时候，却跑去欣赏四百年前一群文盲石匠造出来的东西。

You know, we spend a lot of time in the digital world and the people we talk about AI to go look at something that illiterate Masons made four hundred years ago.

没错。

Right.

亲身体验，感受那种恢弘气势，真是非同寻常。

And to sit in it and to enjoy the grandeur is extraordinary.

这可不是人工智能。

This wasn't AI.

这也不是什么EDA设计或CAD。

This wasn't EDA designer CAD.

那些人就靠笔和纸，跟一群不识字的石匠和其他工匠沟通，却造出了如此不可思议的美。

These were guys with pen and paper communicating with masons and other tradesmen who couldn't read, and they built something unbelievably beautiful.

那真的特别有意思。

And that was really fun.

所以今年你会上台演讲吗？

And so you're going to be on stage this year.

你主要会讲些什么？

What are you mainly going

我和OpenAI的Sachin一起上谈话台。

I'm on talk stage with Sachin from OpenAI.

今年一月，我们宣布了一项重大合作，这是硅谷历史上最大的交易之一。

We announced in January a huge partnership, one of the biggest deals done in Silicon Valley history.

未来几年，我们将为他们提供超过两千亿美金的硬件设备。

We'll be doing more than $20,000,000,000 of hardware for them over the next several years.

我们会聊聊硬件部署、快速AI，以及快速推理在新兴推理AI领域的重要性。

We'll talk a little bit about deploying hardware and fast AI and how important fast inference is the emerging sort of inference AI landscape.

推理现在是个热门话题。

Inference is a hot topic.

人人都喜欢推理。

Everyone loves inference.

没错。

Yeah.

我们是通过训练来打造AI的。

We we we make AI with training.

对吧？

Right?

而我们是通过推理来使用AI的。

And we use AI with inference.

随着模型和我们开发的AI变得越来越有用，大家都想用，而推理正是使用它的关键方式。

And so as the the the models and as the AI we made becomes useful, everybody wants to use it, and inference is the mechanism through which we use it.

现在AI变聪明了，大家想用它，而且一用就要快。

And so now we have smart AI, people want to use it, and when they want to use it, they want to use it and they want to be fast.

这正是我们的切入点——不是稍微快一点，而是快上二十倍。

And that's sort of where we come in, and where the fast is not by a little bit, but by 20x.

所以大家都在用自家的AI，尝试各种新花样，部署GPT、ClogCode或者这些超大模型，整个领域就像突然爆发了一样，大家都在摸索新的工作方式。

And so everybody's using their AI, they're trying new things, they're deploying GPT or ClogCode or one of these supermodels, and it's sort of an explosion of building, of trying new things, of a new way to work.

这挺有意思的。

It's pretty fun.

你们上市都快两个月了，我看到你发过一条帖子，觉得特别有意思。

So we're nearly two months after your IPO, and there was one post that I thought was really cool.

你说自己心里憋着一股气。

You posted that you had a very large chip on your shoulder.

而且这股气，还真是字面意义上的——物理上带着一块芯片。

And it was quite literal and physical.

是的。

Yes.

我

I

希望现在公司上市了，我还能保留那种自嘲的幽默感，保留我平时发社交媒体时的那种语气。

hope that, now that we're running a public company that I wouldn't lose sort of the self deprecating humor or whatever that I enjoy, the sort of tone in my social posts.

所以我们弄了个超大的芯片，就挂在我肩膀上，还专门做了一个固定用的背带。

And so we put a giant chip, we put this sort of on my shoulder like this, and we built a harness for it.

对，那个特别有意思。

Yeah, that was a fun one.

很多人觉得，不管是当CEO还是创业者，日子一定都是顺风顺水，其实完全不是这么回事。

I think that people assume that as a CEO or of any size or an entrepreneur that it's sort of always peaches and cream, and it's just not the case.

背后要付出大量的努力，还有你自己和家人做出的牺牲。

There's an enormous amount of hard work, there's sacrifice that you make and that your family makes.

你陪他们的时间变少了。

They see you less.

这不是一两天的事。

And it's not for a little bit.

也不是说连续几个周末或者两周。

It's not like a weekend or two weeks in a row.

你得拼命干。

You work hard.

这可是持续好几年的事。

It's for years.

所以我觉得，稍微分享一下这些经历，大家应该会愿意听。

And so sharing that a little bit is something I thought would be well received.

上市之后，对你来说最大的变化是什么？

So what's been the biggest difference for you post IPO?

找你帮忙的人多了太多。

The number of people who want something.

真的吗？

Really?

是啊，各种各样的请求一下子暴增了，现在我和我的行政助理、首席了事一起分工处理，光是每天找你提各种要求的邮件就有八十封。

Yeah, has exploded of one form or another, And we get a little better at triaging those between my executive assistant and chief of staff, and just he had 80 emails a day of different people asking you for something.

这根本不算工作。

That's not work.

这不过是形形色色的人想从你这儿得到点什么——想见你、占用你的时间、请你去演讲、请你捐款，或者请你帮忙站台。

That's just this range of people who want something of one form or another, to meet you, your time, for you to present, for you to donate to their cause, for you to woah.

这有点出乎意料。

That was a little unexpected.

你上一家公司，你帮着造就了一百位百万富翁。

Your last company, you helped create a 100 millionaires.

这家公司，你培养的更多了，说不上具体数字，但可能有一千人。

This company, you have, I don't know, countless more, but Maybe a thousand.

一千人左右？

Maybe a thousand?

大概一千人吧。

Maybe thousand.

未来还会有更多千人？

Maybe future thousand?

不会。

No.

在

At

上市的时候，数量更多了。

IPO, was more.

大概有一千人。

It was about a thousand.

这些是投资者和员工，还是

Would that be investors and employees or

都是员工。

It was employees.

真的吗？

Really?

包括现在的和

Current and

以前的。

former.

明白了。

Okay.

哇。

Wow.

作为CEO，你从中学到了什么教训？

And so what lesson do you learn from that as a CEO?

就是，你觉得关键在于

Like, what is I in your

我觉得有几点。

think a couple of things.

你要想做好这份工作，就得真心喜欢创造。

I think to do the job you gotta love to build.

赚钱当然很好，但为在乎的人赚钱，那就特别特别棒。

I think making money is really great and making money for people you care about is really, really great.

当你有机会回报那些信任你的人——那些把职业生涯大半押在你身上的人，比如你的投资人，这真的很棒，他们相信你，但他们自己也分散投资。

And when you get a chance to deliver for people who bet on you, who bet chunks of their career, right, your investors, it's great to deliver for them, they bet on you, but they're diversified.

他们把赌注押在你身上，同时也押了其他二十家公司。

They bet on you and 20 other companies.

当一个人押上五到七年职业生涯，而职业生涯总共三十年，那就是拿出了六分之一的事业来赌你。当你能帮他们实现财务目标时，那种感觉特别棒，我每天都为此感到骄傲。

When someone bets five or seven years of their career, and a career is thirty years, right, they're betting a sixth of their professional career, and when you get to deliver for them and they get to achieve the financial goals that they wanted, that's a great feeling and I'm proud of every day.

这确实是一种非常无私的立场。

It's a very selfless position on that.

有意思的是，现在我们正处在一个AI的超级热潮中，不管你怎么称呼它。

I mean, it's really interesting because today, I mean, in the backdrop, we are in a hyper super cycle, whatever you wanna call it, of AI.

现在冒出来的这些公司，不管是做半导体、模型，还是编程代理，都有很多公司迅速崛起，很快就冲上了十亿美元估值。

And so these companies that are coming up, whether they're in semiconductors or they're in, I don't know, models or, you know, coding agents, there's a lot of companies that are rising up really fast and hitting that billion dollar mark.

有些公司像OpenAI和Anthropic那样有了流动性退出，但其他公司只是靠纸面估值，就已经看到了财富的爆发。

Some are having liquidity events like OpenAI and Anthropic to come, but others are getting that paper markup and they're seeing that kind of wealth creation event.

你该怎么保持一个健康的心态面对这一切？

How do you maintain a healthy mindset around that?

我觉得关键在于你自己。

I think you bring it.

这其实没什么变化，对吧？

It's not a change, right?

我最喜欢和那种人一起工作。

For the type of people I love working with.

他们喜欢动手做东西，哪怕报酬不高、不被看好，也照样做硬核的项目。

They like building stuff, and they like building hard stuff when it paid a little, when it was out of fashion.

那时候硬件没人瞧得上，但他们还是埋头做硬件，因为这就是他们真喜欢做的事。现在硬件又火了，他们照样做，心态很稳——他们热爱的从来不是潮流，而是创造本身。

Hardware was uncool, they were still building hardware because that's what they like to build, and now that it's in fashion, they like to build hardware, and they're sort of even keeled about that, that their passion is the building.

在硅谷，至少我了解的这部分，追钱从来不是致富的路。真正让人快乐的是，做你感兴趣的事，和靠谱又有意思的人一起干。只要你这么做，钱自然会来。但更重要的是，你能做自己喜欢的事，和能让你学到东西的人共事，也能把自己的经验分享出去。

And I think in Silicon Valley, which sort of is what I know, that chasing money is not the path to money, that the path to happiness is working on projects you like with colleagues that are interesting for people with integrity, and if you do that, the money will come, But even more importantly, you'll work on things you like, and you'll work with people you learn a little something from, and you can teach a little something from.

如果你的职业观是这样，那就算遇到低谷，你也能保持平静。

And if that's the way you said about pursuing your career, I think when things are really bad, you're on an even keel.

就算顺风顺水，你照样心平气和，因为你享受的是手头正在做的事。

And when things are really good, you're on an even keel because you you're enjoying what you're doing.

本集由Brex赞助播出，我最推荐的平台。

This episode is brought to you by Brex, my favorite.

你花时间在什么上，就会变成什么样的人。我绝不把时间浪费在本不该存在的工作上。

You become what you spend on, And I refuse to spend my time on work that shouldn't exist.

报销单、追收票据、手动结账，这些事太烦人了。

Expense reports, receipt chasing, and manual closes.

从Vercel、OpenAI、Anthropic、Granola到Deepgram，这些正在打造未来的公司，都做了同一个选择。

The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call.

它们全都用Brex。

They all run on Brex.

Brex是一个智能财务平台，把卡片、支出和银行服务整合成一套系统，还内置了智能财务功能。

Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in.

AI代理能自动处理支出、在消费前就执行预算规则，几分钟内就能完成对账。

AI agents that handle expenses automatically, enforce policy before spend happens, and close your books in minutes.

所以Sorcery用Brex，这样我就能专心做真正重要的事，而不是耗在琐碎事务上。

That's why Sorcery runs on Brex, so I can spend time on building and not busy work.

是时候用上Brex AF了！

It's time to get Brex AF!

了解更多，请访问brex.com/sorcery。

Learn more at brex.com/sorcery.

更多详情请访问 brex.com/sourcery。

That's brex.com/sourcery.

再见。

Bye.

从长远来看，我相信这仅仅是您创业旅程中的一个节点，毕竟您是个创造者。

So looking forward, I'm sure this is also just a mark in your journey because you're a builder.

那么，接下来几个月，就像我们之前说的，这一年里您最期待的是什么？

So what are you most looking forward to over the next couple months, as we said, across this year?

说实话，上市并不是旅程的终点。

Look, getting to an IPO is is not the end of a journey.

它更像是一个阶段性平台。

It's sort of a plateau.

是企业迈入成年的标志。

It's of the arrival at corporate adulthood.

这只是一个阶段的达成，为了能去攀登更高的山峰，我们的机会也变得更大了。

It is the achieving one plateau so that you can climb others, and our opportunity has gotten bigger.

我们的资源更多了。

We have more resources.

我们的知名度也更高了。

We have we're better recognized.

我们能触达更多人，也能用更充足的资源去推进我们的愿景和抱负。

We can reach more people, and we can sort of prosecute our vision and our ambitions with more fuel.

每天让我们兴奋的，就是制造更多芯片、建设更多数据中心，发明推动整个行业前进的技术。

And so that's what we're excited about every day, you know, building more chips, building more data centers, inventing technology that moves the industry forward.

这才是每天激励我们起床、全力以赴的动力。

That's what what drives us and gets us out of bed every day.

我之前听过你和哈里·斯蒂宾斯的那期对话，聊到数据中心和对它们的需求，其实我们不是在追赶未来的需求，而是已经远远落后了。

I think I was listening to a Harry Stebbings episode that you did and talking about data centers and the demand for them, and how we're actually not trying to meet future demand, we're actually just really behind

确实落后了。

We are behind.

落后于当前的实际需求。

On where we are.

那你能说说目前数据中心建设进行到什么阶段了吗？

So could you explain where we are with the data center build out?

没错。

Right.

以前数据中心的建设速度完全取决于房地产的节奏。

So what happened is data centers had historically moved at sort of the speed of real estate.

对吧？

Right?

有人决定盖楼，得等两年，走完审批、浇筑混凝土，楼才刚建起来。

Somebody would decide to build a building, and two years later, after permits, and they poured concrete, the building would go up.

两年前，根本没人关注人工智能。

Two years ago, nobody cared about AI.

那时候我们才刚起步，可人工智能的需求却突然爆发了。

Two years ago, we were at the beginning of this, and the AI explosions happened so quickly.

现在对算力的需求太大了。

There's so much demand for compute.

我们对算力和内存的需求已经远远超出了供给，这些东西到底该往哪儿放？

We've outstripped the demand for compute, for memory, and where do these things go?

它们得放在这些大楼里，但我们建得太慢了，所以全球都在抢建数据中心，现在就是这种情况。

They go to these buildings, and we haven't made them fast enough, and so people are chasing data centers around the world, and that's what's happening.

所以这目前是所有人共同面临的重大瓶颈。

And so it is a major limitation for everybody right now.

我们刚采访了金东宇，聊到了数据中心的新架构。

We just had Tony Kim on, and we were talking about the new architecture of data centers.

你们是怎么和其他公司合作的？无论是通过自建，还是打造合适的软硬件架构？

So how are you working with other companies, whether it's through your build outs, creating the right stack and architecture?

当然。

Sure.

过去二十年，数据中心的结构其实没怎么变，里面的很多基础设施也一直没更新。

So the interesting thing about data centers hadn't changed a lot in twenty years, and a lot of the infrastructure in them hadn't changed.

比如我们用的备用发电机，二十年来都没变过；备用电池、冷却系统和冷板单元，也都是老样子。

I mean, the generators that we use for backup have been unchanged for about twenty years, the type of batteries we use for backup, the chillers and the CDUs.

突然间，对算力的需求暴增，大家都在寻求创新，有人开始用燃料电池，甚至有人直接用喷气发动机轰隆隆地为数据中心供电，涡轮机的设计也在重新思考，整个数据中心领域正掀起一股创新浪潮。

And suddenly there's this sort of intense demand, sort of people are looking to innovate, and they're using fuel cells, and guys are using jet engines like boom to generate power for these data centers, and the turbines are being rethought of and so there's sort of an innovative push in the data center.

对我们来说，虽然我们造出了这块超级大的芯片——比其他任何芯片都大58倍——但它最终还是装进一个类似大学宿舍小冰箱大小的金属机箱里，我们把几个这样的机箱放进标准机架，所以物理空间上的要求，跟部署竞争对手的产品差别并不大。

For us, while we build this sort of super big chip, the chip that's like 58 times larger than any other chip, it a goes server, it goes in a metal enclosure that's about the size of a fridge for a dorm room, a little fridge, and we put a couple of those in a standard rack, and so the physical requirements weren't very different from if you're going to deploy a competitor's product.

托尼真正想说的是，这些数据中心的建筑几十年来都没怎么变过，现在它们变成了制造人工智能的工厂，人人都在试图优化这个工厂的每一个环节。

What Tony is really talking about is sort of these buildings were sort of unimproved for a generation, and now they're part of a factory that makes AI, and people are trying to optimize every part of that factory.

是的。

Yeah.

现在他们甚至打算把数据中心放到太空里。

And now they're also putting them in space.

对。

Yeah.

他们现在在讨论把数据中心放到太空。

Well, they're talking about putting them in space.

我觉得很多技术都是这样，大家会谈论很久，才真正实现。

I think like a lot of technology, you talk about it for a long time before it happens.

你有计划吗？

Do you have plans?

我们在太空应用方面很有优势，因为在太空中最难的问题之一就是让这么多小芯片互相通信。

We are really good for space because one of the hardest problems in space is getting all these little chips to talk to each other.

而因为我们是一整块大芯片，所以根本不存在这个问题。

And because we're a big chip, we don't have that problem.

我觉得短期内不太可能在太空部署数据中心。

I don't think we're in danger in the near term of having a data center in space.

我觉得至少还得五年以上。

I think it's more than five years away.

五年？

Five years?

五年。

Five years.

在我们这个行业，五年可是很长的时间了。

And that's a long time in our world.

就在三年前，根本没人用人工智能。

Mean, three years ago, nobody was using AI.

所以在我们真正把大型数据中心搬到太空之前，地球上还有很多工作要做。

And so I think we've a lot of work in building data centers on Earth before we actually have big production data centers in space.

我很想聊聊芯片设计的变化。

I'd love to talk about how chip design has changed.

协同设计已经成为规划未来三年平台的关键，因为你是在为未来三年做设计。

Co design has become an important part of how do you create that kind of platform for the next three years, because you're building for the next three years.

你是怎么思考这个问题的？又是怎么开展芯片设计的？

So how do you think about that and how do you go about chip design?

过去，人们先造芯片，然后在上面跑软件。

Well, I think historically you made chips and you ran software on them.

芯片和软件之间并没有特别紧密的互动，因为中间隔着一层叫操作系统的软件。

And there wasn't surprisingly a close interaction because there was a layer called an operating system that lived between the chip and the software.

所以英特尔和AMD造芯片，而开发者则为操作系统或者芯片编写软件。

And so Intel and AMD made chips, and people wrote software for the operating system or for the chip.

现在AI规模太大了，速度又这么关键，所以他们开始把软硬件设计一起考虑。

AI has gotten so large, and speed is so important, that what they're doing is they're thinking about the design together.

我们能在软件上做哪些调整，来让硬件发挥更大优势？

What changes could we make in software that would advantage the hardware?

或者在设计硬件的时候，我们能做哪些改动，让软件更容易运行？

Or as we're designing the hardware, what changes could we make that would make the software easier to run?

所以现在软硬件是同步设计的。

And so they're being designed sort of at the same time.

就像任何事情一样，一旦把东西放在一起设计，好处就特别大，这股趋势现在正迅速成型。

And like anything, when you sort of design things together, the advantages are enormous, and so this is something that's really taken shape right now.

我们和OpenAI的合作优势之一，就是能提前看到技术前沿，直接把这些洞察融入我们的设计中。

One of the advantages of our relationship with OpenAI is we get to see exactly where the frontier is going, and we get a chance to roll that into our designs.

谷歌的优势在于，他们的TPU可以和开发Gemini或DeepMind的团队紧密协作，双方能不断互相反馈，这种做法在我们这个行业其实还很新，但威力巨大。

One of the advantages Google has is their TPU can be designed in collaboration with team building Gemini or the team at DeepMind, so they can inform their choices back and forth, and that's an enormously powerful thing that is surprisingly relatively new in our space.

你觉得这个过程中最大的误解或挑战是什么？

What do you think the biggest misconception with that processes and the challenges are?

我觉得大家误以为这很简单，只要把人聚到一个房间里就行了。

Well, I think that the misconception is that it's easy and all you need to do is get in a room.

这其实是个非常难的问题。

It's a very hard problem.

搞软件的人有他们自己的思维方式。

The software guys think one way.

搞硬件的人想问题的方式又有点不一样。

The hardware guys think a slightly different way.

你要是为了让软件更好写，往往就会让硬件变得更难做，对吧？

Anything you do to make it easier to write the software makes it harder to do the hardware, right?

这些权衡真的特别棘手，把两边凑到一块儿，就意味着得在这儿多花点功夫，那儿才能轻松点，结果就是有人的进度得拖，还得加人手。

And these are really hard trade offs, and so bringing them together means these sort of compromises where it will be harder here to make it easier here, and that means somebody's schedule is going to be impacted, somebody's got to add resources.

这种讨论特别难搞。

Those discussions are enormously difficult.

我很好奇，因为我是从外部来看的。

I'm really curious because I come from an outside perspective.

你有内部视角。

You have an inside perspective.

外面传得沸沸扬扬的都是大单子，可我根本不知道 headlines 背后到底是什么情况。

There's a lot of big deals that are being thrown around left and right, and I don't know what's actually under the headline.

所以当 SpaceX 宣布跟谷歌、还有 Reflection 达成数十亿美元的交易时，

So when SpaceX comes out and they say they now have multibillion dollar deals with Google and also with reflection?

这到底意味着什么？

Like, what does that actually mean?

他们到底卖的是什么？

Like, what are they selling them?

很难说清楚。

It's tough to tell.

里面的人也一样摸不着头脑。

Tough to tell inside as well.

真的吗？

Really?

是的。

Yeah.

我觉得有些协议只是名义上宣布了，其实没什么实质内容，不过未来可能会有实际效力。

I I think there have been announced some sort of deals that didn't have teeth, deals that that could have teeth later.

对。

Mhmm.

我觉得X公司当时有闲置的算力，你得想想为什么会有闲置——因为他们的Grock模型根本没怎么被使用。

I I think x had available capacity, and you gotta ask why they had available they had available capacity because their Grock model wasn't used very much.

所以他们那些GPU就一直空着，这可不是个好主意。

So they had these GPUs that were sitting around, and that's a bad idea.

于是他们就把一大块算力整体卖给了或租给了Anthropic。

And so they sold a whole block of them or leased a whole block of them to Anthropic.

结果他们一看，吓了一跳。

And they looked up and said, woah.

这主意还真不错。

That's a pretty good idea.

对吧？

Right?

我们手上有这么多GPU。

We had all these GPUs.

我们的模型没成功，但嘿，我们可以做一个很棒的生意，切入这个GPU极度紧缺的市场，把闲置的资源租出去。

Our model wasn't a success, but, woah, we can have a great business by sort of stepping into what is a constrained market, very hard to get lots of GPUs and lease these.

然后他们就开始想，还能租给谁？还有哪些人需要？

And then they looked around and said, what else can we who else can we lease to?

就这样，这个生意开始了。

And so that's how it started.

至于这些交易的具体细节，我不是很清楚。

Now the specifics of those deals, I'm not super familiar with.

你有没有担心现在这种循环交易的问题？

Are you concerned at all about the circular deals that are going on?

我担心的是，英伟达经常有办法利用它的市场优势。

I'm concerned that there are ways frequently for NVIDIA to exercise market strength.

明白了。

Okay.

对吧？

Right?

NVIDIA 有很多办法利用自己的资产负债表来压制竞争对手。

There are ways for NVIDIA to to use their balance sheet to to to limit competition.

对吧？

Right?

比如他们投资了 Neo Cloud，那 Neo Cloud 就更不可能用非 NVIDIA 的芯片；要是他们投了模型公司，那家公司也会被施压，不敢用别人的芯片。

That that if they invest in in a Neo Cloud, right, the Neo Cloud is less likely to use a non NVIDIA chip If they invest in a model builder, there's pressure not to use other people's chips.

我才更担心这个。

That's what I'm more more worried about.

是的。

Yeah.

在代币圈里，现在也发生着同样的事，一堆免费代币往初创公司手里送。

This is happening in the token world with all the free tokens that are being offered to these startups.

说得太对了。

That's exactly right.

我觉得这些就像是在兜售毒品。

I think these are are drug pushers.

来，小姑娘，尝一点，就一点点。

Here little girl, try a little bit, just a little bit.

我觉得初创公司应该这么做：拿他们的东西，但别上瘾，同时也要找AMD拿一些，来我们这儿看看能不能也搞到点，避免完全依赖单一供应商。

And I think what the start ups should do is, you know, they should take it and then never be dependent, and take some from AMD and come to us and see if we can get you some as well, and avoid dependence.

这种依赖从来都没好下场。

That never ends well.

如果你正在打造下一代AI应用，那你必须了解MongoDB——开发者最爱的数据库平台，专为你的智能代理打造。

If you're building what's next in AI, you need to know MongoDB, the database platform developers love and built for the agents you're running.

MongoDB能实时存储、搜索和分析你的数据，内置Voyage AI的向量搜索和嵌入功能，全部整合在一个系统里。

MongoDB stores, searches, and reasons over your data in real time, with vector search and embeddings from Voyage AI all in the same system.

不用再搭一堆独立的流水线，也不用拼凑十种不同的工具。

No separate pipelines, no stitching together 10 different tools.

这就是为什么财富100强中有75%的企业，以及领先的AI原生初创公司都在使用MongoDB。

It's why 75% of the Fortune 100 and leading AI native startups run on MongoDB.

从第一个用户开始，一路构建并扩展到数十亿个向量。

Build and scale from your first user to billions of vectors.

前往 mongodb.com/ai 了解更多详情。

Go to mongodb.com/ai to learn more.

更多信息请访问 mongodb.com/ai。

That's mongodb.com/ai to learn more.

再见！

Bye!

AssemblyAI 是一个语音AI基础设施平台，数百万开发者都在使用它。

AssemblyAI is a voice AI infrastructure layer millions of developers build on.

他们开发了业内最出色的语音转文字、语音助手和语音理解模型，成为Granolah、Hegen、Ashby和ClickUp等公司的核心基础设施。

They build the industry's best speech to text, voice agent, and speech understanding models that serve as critical infrastructure for companies like Granolah, Hegen, Ashby, and ClickUp.

他们的语音转文字模型在准确性和质量上位居行业首位，而语音理解模型则能超越单纯转录，帮你挖掘洞察、识别说话人，并从语音数据中提取关键信息。

Their speech to text models lead the industry in accuracy and quality and their speech understanding models help you go beyond transcription by uncovering insights, identifying speakers, and highlighting key information from voice data.

现在就去 assemblyai.com/sorcery 注册，即可获得50美元免费额度，立即开始构建语音AI产品。

You can get started today at assemblyai.com/sorcery and get $50 of free credits to start building voice AI products.

就是 assemblyai.com/sourcery。

That's assemblyai.com/sourcery.

这其实正好引出了卡普不久前在CNBC上提到的一个观点。

I mean, this is actually a good transition into something that Karp said not too long ago on CNBC.

他基本上在说，每个公司都应该掌握完整的技术栈。

He was pretty much talking about sovereign AI and how every company should own the full stack.

无论是自己的数据、产品，还是中间的各个环节，包括模型本身，因为你一旦把模型外包出去，别人就可能复刻你的成果，那接下来会不会出现一个全新的Figma？

And so whether it's their data for their products and everything in between, and models too, because you don't want to sell that, then the models will recreate it, and then how about a new Figma?

对于这种向全面自主掌控技术栈的转变，你怎么看主权AI？

How do you feel about sovereign AI in this, like, the shift over to owning the full stack?

我觉得关键在于不要依赖别人。

Well, think the notion of not being dependent.

你的业务不该被英伟达牵着鼻子走。

I mean, you shouldn't be dependent on NVIDIA.

不该只依赖一家供应商。

Shouldn't be dependent on one bottle maker.

这种做法很少能有好结果。

I think that rarely works out well.

对吧？

Right?

作为一个国家或者一家大公司，你更希望拥有多种选择。

What you'd like to be as a nation or as a big company is you'd like to have choices.

你不一定非得掌控整个技术栈，但每一层都该有多个可选方案。

And I don't know if you need to own all the stack, but you like choices at each layer of the stack.

我觉得换种方式理解亚历克斯的话，就是你要想清楚自己的核心优势在哪。

I think a different way to think about what Alex said is, you know, you want to think about where your advantage is, right?

如果你手上有独特数据，千万别轻易放手。

If you're really, you have unique data, be sure you don't give that away, right?

在技术栈的各个层面都要保留选择权，但你的数据带来的价值必须归你所有，不能白白帮别人提升模型。

Be sure you have multiple choices in different parts of the stack, but you get the credit for your data, and it doesn't help make somebody else's model better.

展望未来，从宏观角度看，人工智能的快速发展让我们能创造和实现这么多新事物，你对这些变化带来的外部影响最期待什么？

As we look forward, I guess more on the macro lens, the proliferation of AI and everything that we're able to now create and build and do, what are you excited about on the externalities that come with all of this?

说得对。

Okay.

我觉得，我们这一代人有机会让下一代，甚至我们的孩子，不仅不会死于癌症，甚至根本不知道有人因癌症去世。

I think that we have a chance for our children or the next generation, not only to not die from cancer, but to not know anybody who died from cancer.

我觉得，这个目标在二十五年内是完全可以实现的。

I think that is a real achievable goal in twenty five years.

那该有多了不起啊。

Wouldn't that be something?

说到人工智能能做什么，写更好的代码确实很酷，市场也巨大。

You know, when we think about what AI can do, writing better code is cool, and there's a huge market for that.

但我觉得，它真正能为人类带来的改变，是消灭成年人头号死因。

But I think what it can do to better humanity is rid us of the number one killer of adults.

每当我想到我们做这一切的意义，我就想到这一点。

And I think, you know, that's when I think of what we're all doing this for.

就是这样的成果。

It's an outcome like that.

胰腺癌最近取得了重大突破。

Pancreatic cancer had a huge breakthrough recently.

我觉得现在正是突破的最佳时机，人工智能是攻克那些主要致死疾病的强大工具。

I think the opportunity for breakthroughs right now has never been better, and AI is an extraordinary tool in pursuit of knocking down major, major human killers.

对吧？

Right?

你想啊，要是能把癌症和车祸都解决了，每年能救下多少人。

I mean, if you think if you took out cancer and you took out automobile accidents, you're taking out huge numbers of deaths a year.

你会觉得，这真是做了件了不起的事。

And you say to yourself, well, that's a lot of good that we did.

我觉得自动驾驶特别重要，人类开车实在太差了。

And I think, you know, self driving, humans are terrible drivers.

差劲透了。

Horrible.

开车太差了。

Horrible drivers.

不只是十几岁、十八岁那会儿，或者注意力不集中，也不是父母八十多岁了才这样，根本就是我们平时就不专心。

It's not just when we're 16 or 18 or when we're not paying attention or our parents in their eighties or, you know, it's we don't pay attention.

开车时玩手机，跟老婆聊天，担心工作的事。

We're on the phone, we're talking to our wife, we're worried about work.

这还没算上喝酒或者干其他明显危险的事呢，但我们本来就是差劲的司机，而现在的机器开车比我们强多了。

That's even before people drink or do all sorts of things that are obviously bad, but we're just not good drivers, and machines can drive better than we can today.

十五到四十岁的人，最主要的死因就是车祸。

And so that's the number one killer of of people, whatever, 15 to 40, is car accidents.

要是靠自动驾驶消除了车祸，再把癌症也解决了，那这二十年的技术进步可真够厉害的。

You take that out through self driving, you take out a major killer like cancer, you're going, woah, that's a pretty good twenty year run of technology.

你支持肽类疗法吗？

Are you a peptide fan?

我支持肽类疗法。

I'm a peptide fan.

你也是吗？

You are?

是的。

Yeah.

我不太在意具体是哪种肽，而是觉得我们在理解人体生物学、提升身体表现和延长寿命方面的机会才刚刚开始，未来一定会取得巨大突破，不管是靠肽，还是其他技术，比如下一代GLP-1抑制剂之类的，都会迎来飞跃。

I think not in the specifics of there is a peptide, but rather in that our opportunity to advance our knowledge about the biology of our bodies and how to achieve performance and how to achieve longevity, we're just beginning, and we're going to make some big strides, whether it's with peptides or something else or, you know, the next GLP-one inhibitor or whatever, we're gonna make giant strides.

是啊，看到这么多新药和新药研发公司涌现，真的很令人振奋。

Yeah, it's been really cool to see all the new drug discoveries and new drug discovery companies.

布莱恩·阿姆斯特朗刚成立了一家叫New Limit的公司，就是其中之一。

Brian Armstrong just came out with this company called New Limit, that's one of them.

我觉得他们的目标是彻底消除所有疾病。

I think their goal is to eradicate all diseases.

这想法不错。

That's a good one.

没错，就算这有点狂妄，但十年前根本没人敢这么想。

Right, I mean, even if that's a little hubris, I mean, it wasn't even hubris that was thinkable a decade ago.

是的。

Yeah.

对吧？

Right?

我们的认知现在已经进入了一个领域，说它巨大确实有点惊人，但还不算离谱。

I mean, that we're now in a realm where, wow, that is sort of crazy big, but not insane.

对吧？

Right?

这难道不让人兴奋吗？

I mean, how cool is that?

是的。

Yeah.

很多人总爱谈人工智能的悲观前景，但我觉得我们现在开始更多地关注它实际带来的成果了。

I mean, well, a lot of people like to talk about the doomerism of AI, but we're, I think we're entering a new mix of talking about the actual outcomes with it.

所以，就像你提到的，无论是自动驾驶还是药物研发，这些进展都真的很不错。

And so to your point of what you're talking about, whether it's with AVs or drug discovery, it's really good.

我觉得悲观派的问题在于，他们只看到了账单的一面，对吧？

I think the problem with the doomers is they're only looking at one side of the ledger, right?

要想清醒地看待这件事，你得把账单的两面都看清楚。

I think to look at this with clear eyes, you gotta look at both sides of the ledger.

没错，我们确实会用掉很多电力，AI也确实存在一些真实的风险，这都没错。

You gotta say, look, we're gonna use a lot of power, it's true, and AI has some real risks, it's true.

但另一方面，AI也能带来一些完全不同的改变，比如在教育领域。

And on the other side of the ledger, here's some things it can do really differently, and here's some things it can do that take education, right?

两千年来，我们早就知道该怎么教育孩子，可从来没人真正这么做。

We've known for 2,000 the right way to educate children, and we never do it, ever.

我们明明知道，亚历山大大帝之所以优秀，是因为有亚里士多德这样的顶尖导师一对一教导，每个孩子都应该按自己的学习方式来教，可我们从来不做。

I mean, we knew that the right way to educate Alexander the Great was to have a tutor, and the smartest tutor, Aristotle was, and that you teach each child differently, and you think about their way of learning, and we never do that.

我们只是把孩子塞进教室，用同一个标准教所有人，完全不管他们各自的学习方式有什么不同。

You throw them in a classroom, you teach to some sort of middle level, Each child does not get sort of any different teaching for their different way of learning.

但有了AI，这一切就能做到了。

With AI, we can do that.