Etched：打造更快更便宜的 AI 推理硬件——《像最佳投资者一样思考》第 480 集

几年前我们最后一次公开亮相。

A couple years ago was the last time we were public.

那时候我们刚组建机架团队，还请来了布莱恩·洛勒，他之前负责了英伟达所有HDX和DGX系统的研发，那可是他们收入的八成左右。

At that point, we just started building our rack team, and we brought over Brian Loyler who built all of NVIDIA's HDX and DGX systems, which is, like, 80% of their revenue.

我们就决定，机架和芯片要同步开发。

And we said, we're gonna build the rack at the same time.

在芯片回来之前，我们已经迭代了好几版机架设计。

We actually went through multiple iterations of the rack before the chips even came back.

芯片还没回来，我们就先做了热仿真芯片，模拟出预期的热点分布，用来测试冷板，甚至故意加压把它们炸开。

Before the chips came back, we made thermal chips that had the exact same hot spots as we expected our chips to have so we could build the cold plates, we could over pressurize them and blow them up.

自从芯片真正回来装上冷板后，我们一个漏点都没出现过，因为早就验证过了。

We haven't had a single leak since our chips came back with the cold plates because we already validated them.

我们在台湾有一家工厂，那边有几十号人。

We have a factory in Taiwan, we have a few dozen people out there.

我们在办公室里复刻了一整套测试设备。

We built a clone of a bunch of the test stations in our office.

这层楼有个两兆瓦的数据中心，我们一直进行全天候的开发测试。

We have a two megawatt data center on this floor, and we did twenty four seven development cycles.

大家分白班和夜班，拼命把硬件搭建起来并尽快跑通。

People are doing day shifts and night shifts to actually get the hardware up and running as quickly as possible.

正是这种极致的垂直整合和高度并行的排期，让你能把产品快得多的速度推向市场。

But it's that extreme vertical integration and extreme parallelization of the schedule that lets you get products to market way faster.

回想一下早期团队的建设，两个年轻人要撑起这家公司，需要付出很多。现在有很多非常有才华的年轻创业者，可能是很久以来第一次遇到这种规模和体量的机会。他们肯定都能从你学到的经验中受益，比如如何吸引那些在其他大公司已经拥有成熟职业生涯的顶尖人才加入你。

If you think about the building of the early team and what it required as two young guys building this company, there's lots of very talented young entrepreneurs out there, maybe for the first time of this scope or magnitude in a long time, all of whom probably could benefit from the lessons that you've learned, getting very sophisticated, talented people to come join you even after careers at the other great companies.

如果你要把这个当成一门课来教，比如教年轻人怎么在经验不足、涉世未深的时候搞定顶尖人才，那课程大纲会是怎么样的？

If you were teaching this as a class, like here's how to get lead talent when you're young and inexperienced and naive, what would be the syllabus?

我们在人才策略上采取的是双轨制。

We have a pretty bimodal talent philosophy.

第一步是我们所谓的‘传奇人物’策略。当我们面临极其棘手的技术难题，或者要做一些前人没做过的事情时，我们必须找到全球最顶尖的那个人。

It starts with what we call the legends, which is when we're trying to solve an incredibly hard technical problem and generally do something that hasn't been done before, we need to find the very best person in the world.

通常来说，全球第一、第十和第一百名之间的差距非常大，这直接决定了问题到底能不能解决。

And often, the number one guy in the world versus the number 10 guy versus the number 100 guy, huge difference in whether it's actually possible to solve the problem.

我们建立了一套叫项目制招聘的系统，把各行各业里最难解决的技术问题都梳理出来。

We created the system we call project based recruiting, where we map out all of the hardest technical problems across all industries that anyone has ever had to solve.

我们还会看时效性，也就是那些真正从零到一做出成果的人是谁。

We look at temporality, so who are the people who did the zero to one?

名义上的负责人是谁？

Who is in charge, quote unquote?

实际上干活的人又是谁？

Who actually did the work?

我们尽可能多地接触候选人，然后持续跟进。

We talked to as many people as possible, and then we just track it.

第一次聊完就答应的人确实不多，但聊到第二十次时，愿意加入的人比例会高得让你惊讶。

And you'd be surprised by the amount of people who say yes after the first conversation is pretty low, but the amount of people who say yes after the twentieth conversation is surprisingly high.

一定要锲而不舍地争取。

You really gotta keep at them.

当听到那些真正合适的人说“不”的时候，

When you hear no from somebody who really is

世界上最顶尖的人才，其实意味着你应该等自己多拿几个里程碑再回来。

the best in the world, then that really means, hey, should go ahead and come back when you had a few more milestones proven out.

我觉得最能让人信服的一点是，我们敢大胆承诺，而且一次次真正实现这些目标，这真的特别能激发人的信心。

I think it's one of the most convincing things to see is, hey, we make bold claims, and when you go ahead and hit those again and again and again, that is really belief inspiring.

当我们决定要打造整套机架，而不仅仅是芯片时，我们就想，到底有多少产品是真正大规模出货过的机架级系统，还具备我们想解决的高功率密度？

When we decided we wanted to build a rack and not just a chip, we were looking at this and we're saying, you know, how many products have actually shipped at scale for a rack scale system that actually have the power density that we're trying to solve?

我们就想，如果能施展魔法，世界上最有能力的人该是什么样子？

And we just kind of said, if we were gonna wave a magic wand, what would the best possible person in the world look like?

我们觉得，要是能找到一个从英伟达起步，亲手带过他们所有代际的机架团队，掌握所有相关技术，但依然保持创业者的拼劲，又真正经历过规模化的人，那就是最理想的人选。

And we're like, well, if we could find somebody who, like, started at NVIDIA and built the entire rack team through all their different generations, learned all this different stuff, but is still scrappy, still understands the startup culture, but has seen scale, like, would be the best possible person.

于是我们梳理了英伟达所有与机架级产品相关的团队。

So we mapped all of the different teams that related to all of the different rack scale products at NVIDIA.

最后我们找到了三个人，觉得他们可能符合要求。

And we found three people that we thought could fit the bill.

我们联系了他们所有人，结果有两个已经退休了。

And we talked to all of them, and two of them have just retired.

其中一位本来计划再做完下一代产品就退休。

And one of them was planning to do one more generation for NVIDIA and and then retire.

他叫布莱恩。

His name is Brian.

后来我们慢慢说服他加入我们。

And over time, we convinced him to join.

布莱恩在英伟达创立了HDX和DGX团队，那个团队贡献了英伟达大部分收入，每个季度高达数百亿美元。

Brian started the HDX and DGX team at NVIDIA, which was, you know, a majority of NVIDIA's revenue, tens of billions of dollars a quarter.

另外两位则选择了投资我们。

And the other two guys ended up investing, by the way.

有这样的人在，他一眼就能看出什么是真正的优秀，因为他亲眼见过。

But when you have somebody like that, they just know what good looks like, you know, because they they've seen it.

我们每次和布莱恩讨论，他都会指着我们说：‘这可是我花十亿美元换来的教训。’

And there's so many times where we'd talk to Brian, he'd just point to us and be that's a billion dollar like, a billion dollar lesson I learned.

这可是我花十亿美元换来的教训。

A billion dollar lesson I learned.

嘿，这可省了我们不少精力。

Like, yo, that just saves us cycles.

把像Brian这样的人和Sanford这样的人配在一起。

And you pair someone like Brian with somebody like Sanford.

他们有名字吗？

Do have a name for them?

Brian真是个传奇。

So Brian's a legend.

Sanford又是怎样的人？

What's what's Sanford?

没错。

Yeah.

我们常说，心高气傲的人，能把芯片放进数据中心。

We say chips on shoulders put chips in data centers.

Sanford和Gavin在高中时就是世界机器人锦标赛冠军，而Sanford当时正读大学最后一年。

So Sanford and Gavin in in high school were world robotics champions and Sanford was finishing his senior year of college.

几年前我们联系了他，说嘿，你能来看看我们在做什么吗？

And we called him up a couple years ago and we said, hey, can you come check out what we're doing?

我们需要在平台方面的一些帮助。

We need some help on the platform side.

他过来待了一周，然后我们说，你能在这一周内造出一个冷板吗？

He comes for a week and we say, can you build a cold plate this week?

如果你问任何热工程师，类似这样的人，他们肯定会觉得你简直太天真了。

And if you asked, like, any thermal engineer, anything like that, they they would think you're just, like, totally naive.

对吧？

Right?

我是说，这些事儿得花好几个月才能搞定。

I mean, these things take months to do.

而且，说清楚点，他们确实能行。

And, like, to be clear, they do.

但只要你下定决心，觉得这事儿可行，一周就能取得实实在在的进展。

But you can make real progress in a week if you if you put your mind to it and you think it's possible.

他一周内搞出个装置，真真切切地解决了我们一个关键的电力难题。

And he built a contraption in a week that, like, actually derisked, like, a pretty key power question we had.

把这两者结合起来，他们就干出了惊人的成绩。

And you put those two together, and they've done incredible things.

这两者缺一不可：你需要那种极度执着、不断追问为什么、根本不顾忌风险的人去大胆尝试，同时还需要那些见过大场面、却依然保有初创公司拼劲的人来帮他们落地。

One is not possible without the other because you need the extremely driven people that just keep asking why and don't know where the bodies are buried to, like, take tons of aggressive risks, and then you need the people who've seen scale and still have the startup scrappy mentality to help them along

没错。

the way.

所以这其实是传奇人物加上一点天真，那种回归第一性原理的天赋。

So it's really the legends plus some naivete, law, first principles type talent.

不是说公司里同时有这两种人就够了，关键是他们得真正合作起来。

It's not just that you have both in the company, it's that they're working together.

没错。

That's right.

如果我

If I

想想这个筛选流程，关于你在招聘上的进步，还有为什么这些指标一直在提升，还有什么更值得说的吗？

think about that funnel, anything else more interesting to say about how much better you've gotten at recruiting and why those metrics keep getting better?

其中一个让人惊讶的地方是，你这种反主流的选择，其实本身就筛选了一类人——如果你是个机会主义者，不会去追风口公司，而是会先做足功课，那你就不会来我们这儿。我担心的是，随着我们越来越多地公布产品和细节，如果我们不够谨慎，可能会失去这部分人。

One of the shocking things is, I wish you being such a contrarian bet, kind of self selects, that, like, if you're the kind of person who is somewhat opportunistic, shouldn't go join whatever the hot company is or rather than go ahead and do due diligence, you will not come work here, And it's one of the things I worry about as we announce more and more of the product and its specs, we may lose some of this if we're not very careful.

你得有点疯，才会来我们公司。

You kind of have to be sick in the head to join our company.

你想想看，一个本来在别处做得很好、收入稳定、职业前景清晰的资深工程师，居然愿意说服家人搬去圣何塞，住进这家半导体公司提供的公寓里——而这家公司由两个刚满24岁的年轻人运营，产品还没出来，却要挑战全球最大的公司，还要在有史以来最紧缺的供应链环境中，做出一个他们声称不是提升10%，而是提升十倍的设计。这种人，脑子肯定有点问题。

When you think about it on paper, it's like you, a person who is probably a very accomplished engineer, making a good amount of money, it's liquid, it's predictable somewhere else, you're going to convince your family to move to San Jose and live in this apartment on this housing program for the semiconductor company run by two, what, 24 year olds now, that's pre product that is going against the biggest companies in the world and the most supply constrained environment ever created with a design that they're saying is not going to be, like, 10% better, but it's gonna be 10 x better.

能干出这种事的人，肯定哪里不对劲。

Something must be wrong with you to do that.

我们这儿的人想法不一样，他们不是为了反驳那些不相信的人，而是想证明那些相信他们的人是对的。

People are just wired differently here that they, like, really want to not prove people wrong who don't believe, but prove people right who do believe.

他们真的会把这事当成自己的事。

They just take it personally.

找到这样的人，真的特别有意思。

And, you know, that's really fun to find those people.

而且，坦白说，公司本身的性质就足以筛掉那些不符合要求的人。

And, frankly, just the nature of the company makes it very easy to whittle out the people who aren't like that.

罗布，我们最早聊到的话题之一，就是我开始打听Sohu，也就是咱们第一款产品的名字。

One of the very first things you and I talked about, Rob, was I started asking about Sohu, which is the name of the first product here.

你当时说，细节咱们可以慢慢聊，但你需要知道的是，我们真正专注的是打造一台机器，能够以最高效率、最高质量，大规模地生产这些产品，并不断迭代出新一代的产品。

And you said we can talk about that in great detail, but the thing you should know is that what we're really focused on is building a machine that can, at scale, produce these things and generations of them as efficiently at the highest possible quality level.

所以，我们想打造的其实是一家公司，或者说，这家公司本身就像一台机器，专门负责生产这款产品，以及后续的一系列产品。

So we want to build like the company or the machine that is the company is the thing that will produce this thing and then subsequent things.

我想聊聊公司的几个核心原则或基石。

So I'd like to talk about a few principles or cornerstones of the company.

你之前也提到过其中一些。

You've alluded to some of them.

比如速度。

You've said velocity.

还有垂直整合。

You've said vertical integration.

这些话题现在越来越热门了。

These have become more popular topics.

并行化这块，或许我们也该聊聊。

Parallelization is something maybe that we should talk about.

不过我特别想听听你们团队愿意冒大风险去追求速度的决心。

But I'm especially interested in your guys' willingness to take huge risk to go faster.

能不能分享一个你们最喜欢的故事，讲讲为什么坚持这种理念，以及它让你们做到了哪些其他公司没做到的事。

Maybe tell your favorite story about why this is the philosophy, what it's allowed you to do that maybe other companies haven't done.

这里有很多故事，

There's a number of stories here,

但我最喜欢的是，有一回我们差点就要把芯片送去流片了。

but one of my favorites is there was a time where we were getting close to taping out the chip.

我们突然意识到，等等，有个供应商进度严重滞后。

We realized, wait a minute, one of our vendors is way, way behind schedule.

当时我们面临两个非常糟糕的选择。

And we had two very bad options.

一个办法是继续用现在的供应商，但得把时间表往后推大概一年。

One option, just to keep the current vendor and push our timelines up by on the order of a year.

另一个办法是换掉供应商，从头再来，同样也得把时间往后推一年。

Another option was to push vendors, start over and also get push timelines out by a year.

这两个都不是好选择。

Neither of these was a good option.

所以我们只能去找第三条路。

So we had to go look for option number three.

后来我们发现，他们全都在班加罗尔。

And what that was, was we figured out they're all in Bangalore.

他们以为自己在干活。

They think they're actually going and doing the work.

我们直接派了十几位顶尖工程师跨越半个地球，去班加罗尔待了六个月。

We went out and shipped a dozen of our top engineers across the world to Bangalore for six months.

我也在那儿。

So I was there as well.

我自己在班加罗尔住了四个半月。每天早上，我们都会穿过那些异常繁忙的街道去办公室，总是我们最早到。

I lived in Bangalore for four and a half months personally and every morning, we'd go ahead and walk across the crazy busy Bangalore streets into the office, We'd be the first ones in.

我们出去开发各种各样的工具，比如审查大量即将提交的代码，还构建了一些工具来加快进度，确保能当场做出正确的设计决策。

We'd go out and build a wide variety of tools, both things like, hey, auditing a huge amount of the code that was going in, building a bunch of tools as well to make this go even faster, making sure we're making the right design decisions on the spot right there.

不用来回沟通，当场就能立刻拍板决定。

No to allow back and forth, go ahead and decide immediately.

然后凌晨一点，我们再穿过已经空荡荡的班加罗尔街道走回去，第二天再来一遍。

And then at 1AM, we'd go walk back through the now empty Bangalore streets and do it all again the next day.

我们当时在美国还留了一部分团队。

We still had a bunch of the team in The US.

我们实行的是每天十二小时的交接制度，这样就能实现二十四小时的开发循环。每天早八点和晚八点，大家都会准时上线开会，同步所有数据，然后明确分工，比如我这边一醒来，就得把某某任务搞定。

We ran these twelve hour on each side handoffs, where we had twenty four hour development cycle where at 8AM and 8PM every day, we'd all get on the Zoom, we'd share all the data, and we'd say, when I wake up, like, this must be done.

比如，我们必须得把这款芯片赶出来。

Like, we must get this chip out.

那段时间真的非常紧张。

And it was extremely intense.

与此同时，我们看到其他芯片也处于和我们一样的阶段，用的是同一家供应商，结果却拖了好几年，到现在还没上市。

At the same time, we saw other chips at the same stage as us with that same vendor that ended up taking years, that still aren't out today.

他们到现在甚至都还没完成流片。

They still haven't even taped out today.

这就是把产品推向市场所必需的极端紧迫感。

And it's that level of extreme urgency that's required to bring products to market.

关键点在哪里呢？

What is the key

做这件事？

to doing this?

这已经成了一个套路，主要是因为埃隆·马斯克。像他这样的特殊人才，以及那些试图效仿他的人，都会尝试这样做：找出那个最关键的瓶颈，然后亲自投入大量精力去攻克它。这有点像去班加罗尔那样。

Well, this has become a trope because of Elon mostly that, like, his special skill and others that seek to emulate him would try to do this too, is figure out what the binding constraint is and just flood the zone personally on that thing, which is kind of like going to Bangalore or something.

看起来这是商业运作的核心原则之一，也是任何想要进行这种垂直整合的企业都必须遵循的原则。

It seems like this is a central tenet of the business and of any business that's going to do this kind of vertical integration.

那要把这件事做好的关键是什么？

What's the key to doing that well?

再说一次，关于这个具体的做法，你学到了什么？

Like, again, what have you learned about that specific act?

对我来说，我觉得有两个关键点。

For me, think there are two key tricks to this.

第一，你不可能独自造出芯片。

The first one is that you can't build a chip alone.

这必须是个团队项目，而你最重要的任务就是去找优秀的人才加入你，让他们受到鼓舞，兴奋地去完成这些疯狂的壮举。

It's got to be a team problem, and your most important job is to go get great people to go with you and great people to go ahead and be inspired and excited to go ahead and do crazy things like this.

要求大家说走就走，把生活打乱个半年，甚至像我们之前派出去的那位同事那样待上一年，这确实是个不小的挑战。

It is a huge ask to go say, hey, guys, uproot your lives for six months or in one case, twelve months that we had sent one guy out well ahead.

这滋味确实不好受，但很幸运，我们团队里的人都是出于正确的动机才加入的。

It sucks, but we're lucky to have team members who are in it for the right reasons.

另外一点同样关键，就是决策要快。最糟糕的情况莫过于工厂或供应商在那儿干等着，就等你拍板，结果一直卡着不动。

But think the second big thing too is being able to make decisions very fast, that one of the worst items is when there's a factory or there's a vendor who is waiting for you to go ahead and make some call and has been just stalled.

这种情况经常发生，哪怕只是些很小的事情。

And this happens all the time, even for very small things.

所以，派个人去，把很大一部分责任交给他，让他自己做合理的判断。

So, send folks, delegate a big amount of responsibility to them and say, make a reasonable call.

就算偶尔出错也没关系，我宁愿每次都立刻给出答案，哪怕不是百分百正确，也不愿每次都等一个完美的回复。

Okay, if you're wrong every now and then, but I would much, much rather be right both of the time and give an answer immediately than wait every time for the perfect response.

速度决定成败。

Speed wins.

那为了提速，花钱呢？

What about spending money to go faster?

现在这种边做边学的方式变得特别有意思，随着科技界从软件转向更多硬件，我们以前把大量边做边学的机会外包出去，把生产全丢到海外，自己只在美国当个甩手掌柜。

There's this learn by doing thing which has become so interesting, and as the world has gone away from software and towards more hardware again in the world of technology, that we've outsourced so much of the learn by doing by shipping stuff overseas and effectively just being the ideal guys here in The US.

看来这种情况正在逆转，而你们选择了一种边做边学的方式，主动把自己放进迭代和学习的循环里。

Seems like that obviously is reversing, and you've adopted this way of learning by doing, like you want to be in that iteration learning loop.

没错。

Absolutely.

这其中也包括愿意花钱、愿意承担风险。

And part of that is willingness to spend and take risk with dollars.

是的

Yeah.

你能详细说说这个吗？

Can you talk about that a little bit?

我觉得有句话说得特别好：最大的风险就是不敢冒险。

I think there's a great quote of, like, the biggest risk is not taking risk.

这里的情况也差不多，这个领域每天的营收超过十亿美元，其中很大一部分是靠快速决策带来的。

Very similar here, which is, like, every day, there's over a billion dollars of revenue in this category, and a lot of it's inference.

所以我们每天不发布产品，就是在白白放过大量机会。

So every day we don't ship, we're just leaving tons of opportunity on the table.

因此，如果你能明确看到投资回报，那在花钱上就应该非常果断。

So your willingness to spend money should be extremely high if you can get a very clear ROI out of it.

我们有个概念叫‘预加载’，就是在等一件事完成的时候，既然你知道接下来还要做别的事，有没有办法提前准备，避免拖慢整个进度？

So we have this concept that we call prefetching, which is when you're waiting for one thing to get done, when you know you're going to do other things once you have is there ways that you can paralyze the entire schedule?

比如，我们知道芯片会在某个特定日期回来。

So for example, like, we know our chip is going to come back on a certain date.

我们希望在芯片到位之前，把所有能提前做的工作都做完。

We want it to be that everything possible that could be done without the chip is done before the chip lands.

这需要花一大笔钱。

And this costs a lot of money.

也就是说，我们想在芯片回来之前，先把整个软件栈都搭建好。

This means that, like, we wanna build our entire software stack beforehand.

比如，我们在没有芯片的情况下，就把整套机架运到客户的数据中心，网络、CPU、存储全都配置好，就等着芯片回来就能立刻启动所有数据中心软件。

Like, we shipped racks to customer data centers without our chips in them with all the networking, all the CPUs, all the storage all set up so we could bring all that data center software up before the chips came back.

我们动用了700多块FPGA，把完整的芯片设计映射到FPGA集群上，提前用全套推理系统跑了一遍十几个模型。

It meant that we took over 700 FPGAs and put the entire full reticle chip on an FPGA cluster and ran a dozen different models with our full inference stack on them before the chips came back.

我们还专门做了一个模拟芯片热特性的热模型，根据它提前设计好了散热板。

It means that we built a thermal chip to mock the thermal profile of our chip and build cold plates based on that before the chip came back.

这意味着整条生产线都已准备就绪。

It means we had the entire production line ready.

电路板我们也反复迭代了好几个版本。

It means we did many revs of the circuit board.

这意味着，芯片一回来，整个产品就能立刻投入使用。

It means the entire product was ready to go before the chips came back.

这就是我们能获得的优势。

And this is what it gives you.

还有一家非常知名的AI芯片公司，从拿到芯片到能在机架上运行推理，花了整整十个月。

There is another very famous AI chip company that took ten months to go from getting their silicon back to having them running inference in a rack.

这件事他们还公开向投资者宣布了，当时引起了很大轰动。

And this was, like, publicly announced to their investors and it was a really big deal.

而我们只用了四十天就完成了。

We were able to do it in forty days.

这是因为芯片回来的时候，所有事情都已经准备好了，没什么好折腾的了。

And it's because by the time the chip came back, everything was boring.

软件早就写好了。

The software was already written.

机架也早就摆好了。

The rack was already there.

生产线早就准备好了。

The production line was already set up.

我们就是不停地干，赶紧把所有东西都凑齐。

We were just go, go, go, go, go, get everything together.

你不可能每次都面面俱到。

You don't always catch everything.

边做边调整，然后直接开干。

You make some tweaks on the fly, and then off you go.

在那个情况下，这也是关键因素之一。

In that particular case too, that was a big part of it.

而且，轮班制度也起了很大作用。

Also, like, shift, I think, made a big difference too.

没错。

Totally.

我们真的安排了白班和夜班轮流上。

We went out and literally had a day shift and a night shift.

有些团队成员早上十点来，干到半夜十二点才走。

There were team members who would come in at 10AM and leave at around midnight.

她则是半夜十二点来，早上十点才下班。

She would come in at midnight and leave at 10AM.

为了赶那四十天的工期，大家真是连轴转，一刻不停。

You're running around the clock to get to that forty days.

对，公司一半以上的人就住在办公室附近，所以搞这种排班确实方便多了。

Yeah, I mean over half the company lives next to the office, so it makes it easier to do that type of thing.

给他们付钱让他们这么做，对吧？

You pay them to do that, right?

额外付钱的话，你们现在还这么操作吗？

Pay them extra, do you still do that?

市场这只无形的手确实很管用。

The invisible hand does wonders.

对我来说也挺有效的。

I mean, works for me too.

我们俩都在那儿。

We're both there.

我想退后一步，聊聊这里更广泛的生态系统。

I'd love to take one big step back and talk a bit about just the broader ecosystem here.

供应端的短缺、全球系统和相关供应链的风险暴露，现在天天都是《华尔街日报》的头版新闻。

The amount of shortages on the supply side, the exposure of risks in the global system and the supply chain around this stuff has become like everyday Wall Street Journal front page news.

比如人们关注、投资、热衷的那些内存股。

Like the stocks that people are watching and investing in and excited about, about the memory stocks.

五年前，这些还都是乏味的普通商品，毫无亮点。

These were boring commodity nothing burgers five years ago.

而现在，它们成了全球关注的焦点。

And now they're the center of global attention.

如果你因为一直参与其中，来评估一下让这种产品成为可能的全球互联供应链，随便聊聊吧。

If you just assess because you've been building in it, the global connected supply chain that's required to make stuff like this possible, just riff on it.

什么最让你感到担忧？

Like what scares you?

哪些方面做得不错？

What's working well?

哪些地方需要改进？

What needs to change?

你希望通过打造这个东西，最终改变什么？

What do you hope you change by virtue of how you build this thing?

你对当前这个局面怎么看？

What's your assessment of this story right now?

我觉得供应链故事里最被低估的一点是，这些产品你买回来之后，根本不可能跟供应商就再没联系了。

I think that one of the most undervalued pieces of a supply chain story is almost none of these things are you buy them and they don't talk to the vendor again.

你必须得持续协作。

You have to go collaborate.

我认为，能和台积电或者内存供应商成功合作，关键就在这里。

That is the most important part, being successful, I think, in ships with TSMC or with memory vendors.

你需要建立这种合作关系。

You need that partnership.

我觉得对于台积电来说，很多人并不明白它到底为什么这么有价值。

I think that for TSMC in particular, people don't understand why it is so valuable.

人们只看到它的技术，而它的技术确实是全球最好的。

People look at the tech and the tech is the best in the world.

但对我来说，真正的价值全在服务上。

But for me, the real value is all in the service.

台积电的客户服务，比我见过的任何其他行业任何公司都要好得多。

TSMC customer service is way, way better than I have seen at any other company in any other industry.

比如你发现，通过某个调整能提升良率，你可以直接提建议，然后他们愿意自己出钱做实验，看看能不能真的提高良率。

It's the kind of thing where if you say, hey, you can improve your yield by making this change, You can go make them a recommendation and then we'll go run an experiment on their own dime in our case to see if they could actually get the higher yield.

当我们发现建议是对的、实验成功后，他们就会把整个生产线都改过来。

And when we found that we were right and the experiment worked, they moved over the rest of the line.

这种事在大多数行业根本不可能发生。

And that kind of thing doesn't happen in most industries.

要是我去钢铁厂说，嘿，你们得改一下钢材配方，他们大概会回我：滚蛋。

If I go to, like, the steelworks plant and say, hey, I want you to change the composition of the steel, and they'll say, screw you.

不是台积电。

Not TSMC.

这就是他们能成为第一、最终胜出的原因。

It is why they are the number one and why they're going to win.

其中非常关键的一点是电力供应和通电速度。

One of the things that matters a ton is power availability and time to power.

问题是，你想要的电力越多，供应就越紧张。

And the problem is the more power you want, the more shortage there is.

这其实跟芯片集群的情况很像，比如为什么Colossus要每小时收12美元来运行黑洞？

It's actually very similar to chip clusters, which is like, why is Colossus charging $12 an hour for black holes?

因为他们是唯一能一次性提供两万个这种资源的地方。

It's because they're the only place you can buy 20,000 of them at once.

对吧？

Right?

为什么500兆瓦的数据中心这么难找？

Why is the 500 megawatt data center so hard to find?

原因完全一样。

It's the exact same reason.

我们得想想，怎么才能让每兆瓦发出更多的电。

You know, one of the things we need to think about is, like, how do we get way more juice out of each megawatt?

大家都在整个系统层面寻找解决方案，不只是提升PUE，还在开发全新的硬件，力求用每兆瓦算出更多的任务。

People are looking throughout the entire stack, whether it's just improving the PUE, but also entirely new hardware to get the most tokens per megawatt to solve this problem.

但说到底，盖新大楼太难了。

But fundamentally, building new buildings is hard.

从100兆瓦升级到1吉瓦，再从1吉瓦到10吉瓦、100吉瓦，其实更容易些，而我们现在已经在这些时间线上逼近极限了。

It's much easier to go from a 100 megawatt to a gigawatt from a gigawatt to ten and ten to a 100, and we are pushing the limits of what's possible on these timelines.

所以很多人既在努力扩大数据中心的规模，也在拼命提升单个数据中心的容量。

So there's a lot of people trying to scale in their data centers as much as trying to scale them out.

这种系统有意思的地方在于，它替代了什么。

One of the interesting things about a system like this is what it replaces.

确实如此。

Yeah.

如果我拿这个机架跟Blackwell、Rubens或者接下来可能出现的其他设备相比，我该怎么理解它们之间的区别？

If I think about a rack like this versus, I don't know, a set of Blackwells or something or Rubens or whatever is coming next, how should I conceptualize that?

这不仅仅是功耗的问题。

It's not just Watts.

还涉及到物理空间。

It's also physical space.

Rebors在最近的报告里提到过，这确实是个问题——根本没地方安放这些系统。

The Rebors talked about this in their recent reviews called that literally this is a problem, that there's literally no space to put the systems.

我该怎么理解这个机架所代表的或替代的其他计算单元？

How should I conceptualize what this represents or replaces in terms of other units of compute?

客户在部署模型时通常是这么想的：当我建数据中心或者搭建集群时，不是抽象地觉得‘我喜欢这些芯片，功耗和占地也不错’之类的。

Here's how customers think about deploying models generally, which is when I'm building a data center or I'm building a cluster, it's not like in the abstract of like, oh, I like these chips and like, yeah, this is the power of footprint and so forth.

而是我有一个真实的生产负载要处理，为了让我的产品有用，必须达到某个特定的处理速度。

It's like, I have a real production workload I'm trying to serve and for my product to be useful, there's a certain speed I need to serve for that.

有些产品要求速度非常快，有些则慢一些。

And for certain products, it's really fast and certain products, it's really slow.

不管速度是多少，这就是我的目标速度。

Whatever the speed is, this is my speed.

关键问题是，在给定的功耗下，我能同时服务多少用户，还能保证这个速度？

The question is, in a given amount of power, how many users can I serve while guaranteeing that speed?

换句话说，也就是交互性，或者说吞吐量，我的实际处理能力是多少？

So another way to put it is ISO, what's called interactivity, what is my throughput?

我们才刚刚进入人工智能基础设施爆发的初期阶段，那时候大家只关心速度。

We are just finishing kind of the early innings of the AI infrastructure boom where people really just cared about speed.

你知道，以前的GPU根本达不到其他类型芯片的速度，比如那些SRAM芯片，每秒能处理成千上万个token，这才催生了大量让人兴奋的新应用场景。

You know, GPUs were not able to reach a lot of the speeds of other types of chips like all these SRAM chips, thousands of tokens per second, and that enabled tons of new use cases that got people very excited.

现在出现了一波全新的AI芯片，我们就是其中之一，都能达到这些速度。

There's an entirely new wave of AI chips, us being one of them, that are all going to be able to hit these speeds.

那接下来的问题就是，既然都能达到这个速度，你能同时服务多少用户？

The question then is, if you're hitting these speeds, what is the number of users you can serve at the same time?

反过来讲，如果我有一个100兆瓦的数据中心，能同时运行多少个软件代理？

And by proxy, if I have a 100 megawatt data center, how many software agents can I run at the same time?

所以大家在评估的时候，我们的硬件通常能在保持同等交互速度的前提下，把并发量提升一个数量级。

So when people are doing that evaluation, our hardware is going to generally be able to get you an order of magnitude more concurrency at a given level of interactivity.

这直接体现在每瓦特能处理的 token 数、每美元能处理的 token 数上，这些都是大家真正大规模部署那些巨型混合专家模型时最关心的指标。

So that directly translates into tokens per watt, tokens per dollar, all the things people care about when they're actually serving these giant mixture of expert models at scale.

现在有个很出名的交互性曲线。

There's these now famous interactivity curves.

没多少人会公开这些数据，但你

Not many people publish them, but you

能看到Blackwell的曲线。

can see a Blackwell curve.

还能看到A和B的曲线，虽然比Blackwell稍差一点，但这家公司依然值八千亿美元。

You can see an A and B curve, which is a little bit worse than Blackwell's, and it's still an $800,000,000,000 company.

所以，如果你想想把这条曲线往外推，不只是稍微推远一点，而是大幅推远，这会带来什么影响？其中最让你兴奋、觉得能实现的事情有哪些？

So if you think about what then the impacts are of shifting that curve, not just a little bit further out, but much further out, what are the things that most excite you about what this will enable?

我想出去看看

I I want to go out and

解决一些最棘手的问题，而且我想用短得多的时间就把它们搞定。

solve some of the hardest problems, and I want to go solve these in much less time.

小时候有些事情，我根本没想过自己能亲眼看到。

There were things growing up that I was not sure I'd be able to live to see.

比如免疫动力学猜想，当年在大学时我们聊过，我当时觉得这辈子都看不到它被证明，结果现在居然被AI模型做出来了，虽然花了挺长时间。

For example, the immunodysics conjecture, one of the things that we talked about in college, and I was not sure I'd see that proven in my life, And this was done by an AI model and it was done over a long period of time.

但如果你能把同一个模型的速度提升十倍，这些突破就能大大加速。

But if you're able to then run the same model 10 times faster, you can go shrink the time to go have these breakthroughs.

数学里还有大量类似的问题，我担心有些可能得花一千年才能证明。

And there's a huge number of other problems in math like this as well that I worry it will take a thousand years to go prove a thing like this.

你要么用更聪明的模型，要么用同样智能但运行快得多的模型。

You'd either have a much smarter model or a model of the same intelligence running much faster.

这样时间就能压缩，我就能亲眼看到结果了。

You can then shrink that and I can see it.

看到这些突破一个个实现，真的太让人激动了。

It's so cool seeing these breakthroughs get made.

我真的特别期待看到更多这样的成果出现。

I am so, so excited to see much more of this happen.

我觉得大家往往把任务和应用放在太短的时间框架里去考量。

I think too often people think about tasks and applications and stuff in these very short time horizons.

聊天速度提升50%确实不错，但算不上什么颠覆性的改变。

Doing a chat and it's like 50% faster is nice, but it's not like game changing.

随着智能体运行的时间跨度越来越长，模型能力越来越强，你会看到那些原本需要数月算力的庞大工程变得可行。

As these agents go longer and longer time horizon and the models get more and more capable, you're going to see gigantic bodies of work that would take months of compute.

我们是从实际耗时这个角度来考虑的。

And we think about this in wall clock time.

如果你去问一家实验室的预训练研究员，他们会告诉你，实际耗时往往是最关键的指标之一。所谓实际耗时，就是指从你开始运行模型到最终跑完并拿到结果所花费的时间。

Like if you talk to a pre training researcher at a lab, they'll tell you that wall clock time often is one of the most important things that matters, and what wall clock time means is the time from, you know, starting your run to finishing it to actually get data back.

如果你能把原本需要六个月才能跑完的任务缩短到两个月，你就能进行更多的迭代尝试，大家也会去调整模型架构，从而进一步压缩实际耗时。

If you can shrink this time from a six month run to a two month experiment, you're gonna be able to do many more iterations and people will make changes on the model architectures to actually improve the wall clock time.

我们在思考应用场景时也是类似的逻辑，而长周期任务中实际耗时的显著缩短，正是超低延迟解码最让人兴奋的地方。

Very similar here in terms of how we think about the use cases, which is the exciting part about super low latency decode is wall clock time on long horizon tasks becomes much shorter.

原来需要一年才能完成的计算任务，现在一个月就能搞定；原来一个月的任务，现在三天就能完成；原来三天的任务，现在七个小时就能做完，以此类推。

So a year long compute build would now take a month and that month long compute build will now take three days and that three day compute build will now take seven hours and so forth and so forth.

这一点真的很难真正体会到，因为模型现在才刚有能力做这些事。

So that's the thing that I think is really hard to internalize because the models are just getting capable enough to do this stuff.

几个月前，Cursor 发布他们让一群编程代理从零开始用一周时间造出一个完整浏览器的时候，我觉得特别震撼。

I thought it was really cool months ago when Cursor published that they had a bunch of coding agents build an entire browser from scratch in a week.

简直不可思议。

Totally nuts.

这种事很快就会在不到一小时内完成。

And that will soon happen in under an hour.

未来会有大量类似的事情发生，多个强大的并行代理一起协作完成同一个任务。

And there's going to be many of those types of things that are going to happen with these massive parallel agents all working on a given task.

这些系统的终极限制是什么？

What are the ultimate limitations of these systems?

这只是一个物理层面的问题吗？

Is this just like a physics question?

比如，理论上我们能快多少倍、便宜多少倍？

Like, how many times faster, cheaper can we get theoretically?

还有很多

There's a lot

空间可以挖掘，就像人们常说的那样。

of room at the bottom, as they say.

如果你想想芯片之间的延迟，在英伟达的产品上，从一个芯片传到另一个芯片大概需要四千纳秒。

If you think about chip to chip latencies, on an NVIDIA product, you're looking at four thousand nanoseconds to go from one chip to another.

我们能做得比那好得多。

We'll be able to do much better than that.

数学上的极限是什么？

What's the mathematical limit?

是光速。

Is speed of light.

你只需要极短的时间，比如两三个纳秒就能完成。

You can do it in just a handful, like two, three nanoseconds.

今天它们是四千零四个。

And they're 4,004 today.

底层还有很大的提升空间。

There is a lot of room at the bottom.

像能效这种事，我们确实可以通过降低电压省下大量能耗，但还能再降，能降得低得多。

The same thing with things like power efficiency that, sure, we're able to go and save huge amount by bringing the voltage down by so much, but you can go lower, you can go much, much lower.

这非常有挑战性，但当我想到二三十年后的未来，就觉得这几乎是必然的。

It's very challenging, but when I think about twenty, thirty years in the future, then I think it's inevitable.

还有规模经济和集群扩展的问题，很长一段时间里，八块芯片就是最大的扩展规模，甚至还有NVL72，把数量推到七十二块，但其实还能大得多。

And also for economies of scale, for cluster scale up, for a long time, eight chips was the biggest scale up domain and it even had the NVL 72, bringing it to, well, 72, but you can be way, way bigger.

比如你看一座晶圆厂，一栋单体建筑就值四百亿美元，里面却只有寥寥几条生产线。

You look at like a fab, for example, you have a $40,000,000,000 single monolithic building with only a handful of lines running through it.

未来也可能出现类似的超级集群，投资四百亿甚至一千亿美元，建一个巨型‘模型工厂’，只为一两个模型服务，却支撑海量用户，实现同样的规模效应——同一个模型，服务无数人。

You could have the same kind of thing for some futuristic mega cluster, $40,000,000,000, $100,000,000,000 as a giant mega token factory serving one or a handful of models for a massive number of users to get that same economies of scale thing, same model, massive number of people.

你提到内核工程是你第一份工作，这曾经是个没人听过的领域，现在却成了人人都在谈论的关键，因为要从硬件底层榨出更多性能。

You mentioned kernels engineering and that being your first job, That has emerged as a thing that nobody had ever heard of in their lives to now something that you hear about all the time, the importance of it, to eke more performance out of the bare raw metal.

那什么时候AI能完全独立搞定这些事？

When will that just be something that AI does entirely as well?

现在人类还是最厉害的内核工程师吗？

Are humans still the best kernel engineers?

他们是不是在用AI系统来辅助工作？

Are they doing it with the assistance of AI systems?

人类在设计这些东西的过程中，还能参与多久？

Like, how far down will humans still be in the loop of designing these things?

那人类什么时候才会彻底退出这个环节？

Like, when will that go away?

现在都是人机协作的混合模式。

Today, all very hybrid.

最顶尖的内核代码，还是靠人类和AI合作写出来的。

And the best kernels are still written by human AI collaborations.

任何AI模型都是由一些基础组件构成的，比如矩阵运算、卷积、芯片间的通信操作等等。

Also, any AI model is built up of these fundamental primitives like matmoles, like convolutions, like chip to chip operations collectives.

让这些操作重叠起来，并且跑得飞快，这非常关键。

And making these overlap and making these really fast matters enormously.

内核工程师的核心工作就是去搞清楚，我能在哪里实现重叠。

And it's the kernel of the matter's job to go ahead and figure out, where can I overlap?

我该怎么分配内存？

How do I allocate memory?

我该怎么确保万一出现重传之类的问题，不会让整个流水线卡住？

How do I verify that if there's some issue, like a retransmit, doesn't stall the whole pipeline?

这些工作确实很有挑战性，但每次优化都能让模型性能提升百分之三到四，而且这样的优化机会非常多。

These things are very challenging, but they can go and make your model performance be, say, 3%, 4% better per optimization, and you can do so many.

我们在规划软件架构时，总是希望走在前面，预判未来的发展趋势。

And we thought about our software stack, we wanted to go skate it where the puck is going to be.

三年前的时候，开发软件基本上只有两条路可选。

And three years ago, there were kind of two ways you could build software.

其中一条就是重金投入图编译器。

One of them was to invest heavily in graph compilers.

这些东西性能不算高，但开箱就能用。

These things are not very performant, but they work out of the box.

它们不需要人工去逐个调优内核，但我们偏偏选了另一条路。

They don't require a human to go come in and tweak all the kernels, but we went the opposite direction.

我们从内核入手来编程，所以很长一段时间里，它们根本没法开箱即用。

We are kernels first programming and that means that it, for a long time, did not work out of the box.

但如果你是内核专家，就能榨出极致的性能。

But if you were a kernel's expert, you could get incredibly, incredibly high performance.

关键是，现在代码模型越来越强，越来越多的内核生成工作已经能交给它们了。

And the thing about this is that now, as the coding models get better and better, they're doing more and more of the kernel generation task.

等模型再聪明一点，迟早能全包了。

And when the models keep getting smarter, it'll eventually do all of it.

到时候，它的表现会超越人类。

It will become superhuman.

所以我们决定，跟着世界的发展方向走。

So we're going to build where the world is going.

即便到了今天，我们在设计性能分析工具或者调试栈的时候，

And even today, we think about our profiling tools or debugging stack.

更多考虑的是模型会怎么使用这些工具，而不是

We think about it from how will the model use these tools more than we think about how

人类会怎么使用它们。

will humans use these tools.

我们内部有时会做一些实验，让 Codecs 仅凭我们的文档，完全靠自己从零开始把 GPT OSS 跑起来，结果它好像一晚上就搞定了。

We sometimes run experiments internally and we had Codecs actually get GPT OSS running from scratch just based off of our docs completely by itself and it did it I think overnight.

我们花了很多心思在选游戏上，意思就是要把精力投在最有把握的项目上。

We think about game selection a lot, and what we mean by that is making sure we're investing our energy in the right bets.

因为不管选什么方向，都需要付出巨大的努力。

Because regardless of what you choose to work on, it will take tremendous effort.

我们一开始就明确决定，不做一个通用的图编译器，不支持随便什么 PyTorch 代码，不支持随便什么 CUDA 代码，也不支持随便什么 ONNX 模型。

And one of the things that we started with was, you know, the decision explicitly not to build an arbitrary graph compiler, not to support arbitrary PyTorch, not to support arbitrary CUDA, not to support arbitrary ONNX graphs.

相反，我们设想了一个未来，真正重要的模型不会超过一百个，而且从底层数学角度看，它们都长得非常像。我们要基于物理原理构建基础组件，尽可能把这些模型跑得快一点，再快一点，同时让最顶尖的客户能直接操作硬件，想怎么玩就怎么玩。

But instead, we envisioned a world where there was going to be under a 100 models that actually mattered, and they were all going to look very similar from the underlying mathematical perspective, and that we were going to build primitives using physics that were going to accelerate these as much as humanly possible, and we were going to allow the most sophisticated customers to have direct access to the hardware and do whatever they want.

这让我们省下了大量时间，根本不用去开发编译器，还让我们获得了更好的性能。

And that has saved us a tremendous amount of time not having to build a compiler and that has allowed us to actually get much more performance.

有趣的是，刚开始的时候，很多人觉得这想法不靠谱，只有高频交易领域的人认真对待我们。

And funnily enough, when we started, a lot of people dismissed this idea and the only people that took us seriously were in high frequency trading.

他们也都讨厌编译器。

They all hate compilers too.

他们全都是自己写内核代码。

They all write their own kernels.

已经有几十个来自高频交易领域的人加入我们团队，因为他们认同我们的理念。

And we've had dozens of people from high frequency trading join the team because they saw this philosophy too.

垂直整合的边界在哪里？

What are the limits to vertical integration?

你们怎么决定该在哪儿划线？

Like, how do you know where to draw the line?

我先问这个问题，是想聊聊更广阔的市场。

And I'm starting with this question to talk a bit about the broader market.

整个市场的状况让我特别感兴趣，因为绝大多数AI芯片都只被少数几家客户买走。

The circumstances of the broader market are really interesting to me, where the vast majority of chips, of AI chips, get bought by a very small set of customers.

这些客户中的很多，其实自己也在研发AI芯片。

Many of those customers are themselves trying to design their own AI chips.

OpenAI发布了Talapeno。

OpenAI announced TALAPENO.

这局面挺有意思，世界上最有价值的东西，居然全靠那么几家芯片厂商和买家来流通。

It seems like this very funny circumstance for like the most valuable thing in the world all kind of flows through a couple chip makers, a couple chip buyers.

他们好像都在琢磨着去干对方的活儿。

They all seem to be thinking about doing each other's job.

然后你再看，这些芯片最终进了数据中心，又冒出一堆云服务商和推理提供商，构成了栈的另一层。

And then you've got the circumstance where like, okay, then these things go into data center and you've got neo clouds and inference providers in this other part of the stack.

还有模型的构建者和提供者。

You've got model builders and providers.

我甚至能想象这样一个世界：你要是硬件做得最牛，干脆连模型也设计了，数据中心也自己建了。

Like, I could imagine a world where, because you have the best hardware, you design models and you build data centers.

你会超出当前的业务范围。

You leak outside of your current vertical.

那么，你是怎么考虑该在哪儿划清业务边界的呢？

So, like, how do you think about where to draw the lines for the business?

我们有一句话：生产就是产品。

We have a saying that production is the product.

归根结底，我们清楚推理服务将成为全球最大的市场。

Ultimately, matters here is we know inference is going to be the biggest market in the world.

谁生产的token最多，谁就会成为世界上最有价值的公司。

Whoever produces the most tokens is going to be the most valuable company in the world.

所以我们所有的决策都围绕一个目标：如何让尽可能多的token产能上线。

So all the decisions we make is how do we get the most token capacity online as possible?

其中一部分是打造一款性能卓越的产品，拥有更高的吞吐量、更低的延迟，这样每颗芯片都能支撑更多的token运行。

And part of that is building a really good product that has way more throughput, that can run at way better latencies and so forth so we can per chip we make get way more tokens online.

另一部分是，除非绝对必要，否则我们不会涉足栈中的其他环节，只为实现大规模运营。

Another part of it is like not doing parts of the stack unless we absolutely have to to get to giant scale.

所以我们有些部分是必须做的，为了实现大规模部署，比如自己做机架而不是只做芯片，比如采用自研模型而不是用JDM模式。

So there are parts that we decided to do because it was absolutely required to get the scale, like building the rack instead of just building the chips, like doing a Centimeters model instead of doing a JDM model.

但有些事情现在对我们来说其实是干扰项。

But there are parts of it that are kind of noise to us right now.

比如，我们目前根本不会自己建数据中心。

Like, we're not going and building our own data centers today.

这其实并不能帮我们提升在线算力。

That doesn't actually help us get more capacity online.

一般来说，我们的客户自己会解决供电问题，还会移动他们的集群，就是为了尽快部署我们的芯片，因为它们的吞吐量实在太高了。

In general, our customers are actually making power and moving their clusters around to get our chips online because they're such high throughput.

如果其他环节成了瓶颈，我们当然会去整合，但现实是，我们只专注一件事：尽可能多地把算力上线。

If there was a world where other things are constrained, we would totally go and integrate with them, but the reality is we're just purely focused on getting as many tokens online as possible.

我觉得这还是回到了规模效应的问题。

I think this comes down to economies of scale again.

栈里的某些环节确实能带来巨大的规模优势，但其他部分则不然。

A certain part to the stack and there are huge economies of scale and others there aren't.

比如在设计模型这方面，规模效应非常明显。

For example, on designing models, huge economies of scale there.

芯片制造也是同样的情况。

For chip fabrication, same story.

但比如说，你想想在机架里头做某个小金属零件，就没这么明显的规模优势。

But for example, if you think about building some small metal part inside of that rack, there's not that same effect.

我们认为自然的边界在底部的芯片和顶部的模型层，中间的整个环节我们都来补上。

We think the natural boundaries are on the chip side, on the bottom, and the model layer at the top, and we'll fill the whole gap between.

几周前，有个为前沿公司运行下一代AI芯片的人，想挖我们的一位架构师。

A few weeks ago, there was a guy who was running a next generation AI chip for one of the frontier companies, and he's trying to recruit one of our architects.

结果他反过来被对方挖角的人给反向挖了，那人都没挖成，自己先被撬了。

And, actually kinda did an UNO reverse card and started recruiting the person trying to recruit our guy.

结果一周之内，我们就把他招来了。

Then within a week, we hired him.

当时我正散步，正好赶上我们最后敲定录用通知。

And I was going on a walk as we were kind of finalizing the offer.

我当时就说，你可是负责这么重要的项目啊。

I was like, you're leading this super important project.

那你为什么决定要来呢？

Why are you deciding to join?

他的回答特别有意思，意思是，对我们公司来说，这个产品能不能成功根本不是生死攸关的事。

And his answer is super interesting, which was, it fundamentally is not existential for my company for this product to win.

对于谷歌的TPU来说，收入主要来自搜索业务。

For Google with TPUs, the revenue comes from search

就算TPU失败了，谷歌也不会垮。

Google won't fail if TPUs fail.

没错。

That's right.

Meta就算MTIA失败了，也不会出问题。

Meta won't fail if MTIA fails.

微软就算Maya失败了，也不会有事。

Microsoft won't fail if Maya fails.

OpenAI也不会因为Jalapeno失败而倒闭。

And OpenAI won't fail if Jalapeno fails.

归根结底，这是我们的产品。

Ultimately, this is our product.

世界上最好的芯片由一家只造芯片的公司打造，这完全在情理之中。

It is completely unsurprising that the best chip in the world is built by a company that only builds that chip.

那就是英伟达。

It's NVIDIA.

对吧？

Right?

但对我们来说，尽可能多地上线算力资源是关乎生死存亡的大事。

And for us, it is completely existential for us to get as much token capacity online as possible.

这样才能吸引到人才，并获得供应商和客户的高度支持，因为他们也和我们一样重视这件事。

It recruits a set of talent and recruits a support from suppliers and from customers that view it with the level of intensity that we do.

你看看那些原始的算力数据。

Look at the raw FLOP stats.

你可以拿实验室或者超大规模厂商造的任何芯片来比。

You compare any of the chips built by the labs or by the hyperscalers.

比如FB8乘FB8的FLOP密度，就比Blackwell B300低。

The FLOP density for say FB8 times FB8 is lower than the Blackwell B300.

这很合理，因为他们不用去冒那个险。

And that makes sense because they don't have to go take the risk.

他们只要造出个差不多的产品，不用交那笔英伟达的溢价税就行。

They just have to go build a similar enough product and not pay the NVIDIA tax.

你们的财务团队并不是因为犯了大错而亏钱。

Your finance team isn't losing money on big mistakes.

而是通过成千上万个没人注意的小决定慢慢流失的。

It's leaking through a thousand tiny decisions nobody's watching.

Ramp能在花钱之前就把好关。

Ramp puts guardrails on spending before it happens.

实时限额，自动规则，再也不用到处救火。

Real time limits, automatic rules, zero firefighting.

去 ramp.com/invest 试试看。

Try it at ramp.com/invest.

随着业务增长，Vanta 也能同步扩展，自动化合规流程，并为你提供一个关于安全和风险的统一事实来源。

As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk.

更多信息请访问 vanta.com/invest。

Learn more at vanta.com/invest.

每家投资公司都是独一无二的，通用 AI 并不了解你的具体流程。

Every investment firm is unique and generic AI doesn't understand your process.

Rogo 可以做到。

Rogo does.

这是一个专为华尔街打造的 AI 平台，能对接你的数据，理解你的业务流程，并生成实际可用的结果。

It's an AI platform built specifically for Wall Street connected to your data, understanding your process, and producing real outputs.

去 rogo.ai/invest 看看他们吧。

Check them out at rogo.ai/invest.

从 OpenAI 到 Cursor 再到 Perplexity，这些顶尖的 AI 和软件公司都借助 WorkOS 一夜之间就能满足企业级要求，而不是耗时数月。

The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months.

访问 workos.com，跳过枯燥的底层基建，专心做好你的产品。

Visit workos.com to skip the unglamorous infrastructure work and focus on your product.

Ridgeline 正在重新定义资产管理技术，它不只是软件供应商，更是真正的合作伙伴。

Ridgeline is redefining asset management technology as a true partner, not just a software vendor.

他们帮助多家公司实现了五倍增长，推动更快发展、更智能的运营，以及更强的竞争力。

They've helped firms five x in scale, enabling faster growth, smarter operations, and a competitive edge.

访问 ridgelineapps.com，看看他们能为你的公司带来什么突破。

Visit ridgelineapps.com to see what they can unlock for your firm.

当我想到你们在打造这个解决方案时，整个过程其实是在逐一攻克一系列极其困难的挑战。

As I think about you guys building the solution, the process of doing so is solving a sequence of really hard challenges.

其中最难突破的一关是什么？

What has been the single episode that was the hardest to overcome?

我们当时在设计芯片。

We were designing the chip.

我们搭建了一个庞大的 FPGA 集群，用来验证整颗芯片是否能正常运行。

We built this massive, massive FPGA cluster to go out and verify that full chip worked as is.

FPGA 是数字器件。

And FPGAs are digital entities.

你可以测试数字逻辑，但没法测试模拟逻辑。

You can go test digital logic, but not analog logic.

芯片回来后，我们发现出现了数据保持和自适应错误的问题。

And it turns out that when the chip came back, we began to go see issues in our retention of adaptation of incorrect results.

我们意识到，时钟域交叉处的反压逻辑出了问题，这会导致芯片输出错误结果，而且这个问题极其难解。

And we realized, wait a minute, there's a problem with a back pressuring logic across a clock domain crossing is failing and this is going to cause the ship to produce wrong results and it is very, very hard to solve.

我们发现，解决这个问题的唯一办法，就是把芯片上的两个时钟信号对齐到五十皮秒以内——也就是五十万亿分之一秒，还得每秒对二十亿次，每次都达到这种精度。

We realized there was one and only one way to solve it, as we had to go line up two clock signals on our chip to within fifty picoseconds, that is literally fifty trillionths of a second, and we had to go get these signals aligned to this super small granularity and do it on every chip 2,000,000,000 times a second.

很多人说这根本不可能做到。

A lot of people said this was impossible.

当时甚至有人选择了离开。

We had people quit.

是的。

Yeah.

当时很多人都觉得这根本没法解决，结果还真让我们碰上了。

That people literally were like, this problem is unsolvable, and that's the luck, guys.

遇到这种问题，第一步就是先假设它能解决。

Well, when you have a problem like that, step one is, okay, let's assume the problem is solvable.

那它到底是怎么被解决的呢？

How would it be solved?

我们意识到，必须找到办法把时钟相位调整一皮秒、十皮秒。

Well, we realized what we have to be able to do is find a way to go ahead and move our clock phase by a picosecond, ten picoseconds.

我们有个想法：要是这两个时钟本来就稍微错开一点呢？

And we had an idea, what if we had these two clocks, they had them just a little bit apart from each other.

后来我们发现，只要先测出相位差，再利用漂移机制，等恰到好处的时机让它们对齐，就能非常稳定地把相位锁定在精确位置。

We figured out that, hey, if you go out and figure out the phase, if we then go out and use the drifting mechanism go ahead and wait for just the right amount of time to get that always lined up, we can do this extremely reliably and then lock the phases exactly where they have to be.

这样就能确保这种问题再也不会发生。

We can guarantee this never happens.

大家都没想到这真能成，而且效果还出乎意料的好——但我们就是做到了。

People were, I think, blown away that this worked, and it worked as well as it actually did, but we made it work.

这花了多长时间？

How long did that take?

实际上大概用了两周。

It was actually about two weeks.

那可是整整两周的黑暗期。

It was dark two weeks.

那两周特别吓人，但正是在这种时候，你最需要投入精力。当你觉得一切都无望的时候，恰恰最难坚持，但你越早解决这个问题，就能越早回到开发和扩大生产规模上。

It was a very scary two weeks, but it was the kind of thing where when that kind of thing happens, that is the most important time to go ahead and be investing effort, that is the hardest time to go do it when you feel like things are hopeless, but like the sooner you solve that problem, the sooner you can get back to building and scaling up production to mass volumes.

我觉得我们的很多经历都像加文说的那样：先假设这是可能的。

I think a lot of our story is like, as Gavin says, assume it is possible.

假设制造出算力高得多的芯片是可能的。

Like, assume it is possible to have a chip with way more flops on it.

假设实现芯片之间更低延迟的系统是可能的。

Assume it is possible to have a system with way lower latency between chips.

假设建立一个带宽高得多的共享内存池是可能的，那该怎么做到呢？

Assume it is possible to create a shared memory pool that can run at way higher bandwidths, like, how would one do it?

我们做实验的时候，经常要做几十个，结果全都不成功。

A lot of the time when we do experiments, we will do dozens of experiments and all of them will fail.

但只要有一个能成就行了。

But we only need one to work.

我记得有好几次，Gavin带头搞芯片启动，做了大约三十块板子的实验，最后只有三块成功了。

There's multiple times I mean, Gavin, I think, was leading the charge and our chip bring up with, I think, 30 different board experiments, and three of them worked.

这三块板子的价值，简直无可估量。

And all three of them are worth their weight in gold.

有人常来找我说，Gavin，你的实验几乎都没成功。

This is one of the things that people come to me and say, Gavin, almost none of your experiments works.

我只需要运气好那么一次就够了。

I only got to get lucky once.

我问的这些故事里，有一个关于公司历史上艰难时刻的，就是融资这件事。

So one idea for one of these stories that I'm asking about, you know, difficult moments in the company's history is around the ability to raise capital to fund the thing.

你刚开始的时候知道需要资金，但没想到最后需要筹集和花掉这么多钱，才能把解决方案做出来。

I think when you started it, you knew you'd need capital, but you did not know you'd need the quantum of capital that you've ultimately raised and are spending to build the solution.

而且你之前从来没融过资。

And you hadn't raised money before.

这些都是全新的事情，对吧？

These are all new things, right?

有过好几次，真的特别艰难，我就在现场，亲眼看着你差点儿融不到钱——要是没筹到那些钱，公司根本撑不下去。

There was moments where it was really, really difficult because I was there, I saw it moments where it was extremely difficult to raise the money that you did, without which the company would not exist.

公司早就垮了。

It would have died.

就像很多伟大的故事一样，也经历过好多次濒临死亡的时刻。

And like many great stories, there are many near death moments.

但在这个新领域，钱的问题特别不一样，这可不是软件行业。

But money specifically in this new world, this isn't software.

你不是只要一点点钱就行的。

You don't just need a little bit of money.

你能不能讲讲，最早期融资最难的地方在哪？那时候你还没做出能让人眼前一亮、彻底折服的产品。

Maybe you could tell the story about the true hardest part about raising money early on, before you had something that you could show people and be so proud of in performance that you could show them blow their socks off.

那时候你们只是在谈论一个想法。

It was just you guys talking about an idea.

但说说早期融资有多难吧，那真不是一般的辛苦。

But talk about the early difficulties raising money because it was pretty hardcore.

我们确实经历过一些非常紧张的时刻。

We've had some intense moments.

这让我想起大概2024年初，我们还没拿到A轮投资的时候。

It reminds me of probably early twenty twenty four before we raised our series a.

那时候我们已经完成了足够的架构设计，确认了芯片架构是可行的。

And we were at this point where we had done enough of the architecture, done enough of the design that we knew that the chip architecture was sound.

接下来就得真刀真枪地把它做出来。

We had to go build it.

但还有大量工作等着我们。

There was a lot more to do.

我们已经准备进入所谓的物理设计阶段了。

We were ready to go into what's called the physical design stage.

我们需要跟一家物理设计供应商签协议，这至少得花40.5亿美元。

We needed to sign an agreement with a physical design vendor, which it will cost you at least $4,050,000,000 dollars.

后来我们意识到，随着模型越来越大，像那些巨型MoE模型陆续出现，我们不仅要造芯片，还得搭建整个集群——包括主板、互连结构、冷板，还得搞定所有网络方案等等。这花费远超我们账上那1500万美元。

And then we had this realization as the models were getting bigger and bigger and you're seeing these giant MOE models come out, that we were going to need to build the entire cluster, not just the chip, but we were gonna need to build boards, we're gonna need to build interconnects, we need to build cold plates, we're gonna need to figure out all of the networking and everything, and that this was gonna cost a lot more than the $15,000,000 we had in the bank.

你们当时肯定觉得挺吓人的吧。

And you're like, man, that was scary.

就是那种时刻，心里想：天哪，这根本负担不起。我当时甚至开始琢磨，回哈佛读书有多难。

That like, you're sitting in that moment, and you think, holy crap, we can't afford this, and I began looking at it like, how hard is it to go back to Harvard?

2023年底的时候，我们整理了一份备忘录。

At the end of twenty three, we put together this like memo.

我们在这上面花了大概一百个小时，因为心里没底，不知道投资人会不会相信我们要融的那笔钱。

We spent like a hundred hours on this because we're like, we have no idea how people are gonna believe us when we ask for the amount of money we're about to ask for.

这份备忘录大概有30页。

And it was like 30 pages.

内容非常技术化，也很深入，涵盖了我们需要构建的所有不同模块、必须达成的里程碑、市场的演变趋势、新的应用场景、每个token的成本，以及各种模型测算。

It was like extremely technical and in-depth of all the different things we needed to build and like all the milestones we needed to hit and how the market was gonna evolve and all the new use cases and the cost per token and all this modeling.

然后我们去找投资人，硅谷所有主要的投资人都直接拒绝了。

And then we went and talked to investors and every major investor in the Valley passed immediately.

他们就说：两个刚毕业的哈佛学生，连芯片都没流过片，连测试芯片都没有，推理能力都不确定，这市场到底有多大谁说得准。

They were just like, okay, two kids that just finished Harvard, haven't taped out a chip, no test chip, inference, who knows if this is gonna be a big market.

所有东西都只集中在训练上。

Everything's gonna be training.

模型还在胡说八道呢。

The model's still hallucinating.

这整个可能就是个泡沫。

This could all be a bubble.

那时候，半导体领域A轮最大的融资也就四到五千万美元左右。

You know, at the time, the biggest semiconductor fundraisers for a series a was around like forty, fifty million dollars.

我们坐在那儿，一笔一笔算着账。

We were looking at this and we were just like tallying the bill.

我们觉得，接下来十二个月，我们可能会花掉一亿美元。

And we're like, we think we're gonna spend a $100,000,000 in the next twelve months.

我们真的想做成这件事。

We really wanna do this.

如果我们真想做到规模化，实现我们说的性能目标，那这事儿需要巨额资金投入。

Like, if we want to actually get to scale and, like, actually get the performance we're talking about, like, this is going to be extremely capital intensive.

我们到底该怎么把这事办成？

How the hell are we going to pull this off?

我觉得我们一开始能推进这个项目，是因为我们问了自己一个问题：怎么才能用最少的钱来做这件事？

I think that one of our key ways we got started with this process, we thought to ourselves, what is the cheapest possible way we could do this?

我们决定，干脆自己省到极致，除了吃泡面什么都不买，然后把所有钱都砸在流片上，这样大概只需要三千万美元就够了——这数字低得离谱。后来我们还真找了个债主，愿意贷款给我们，帮我们跨过这个门槛。

Decided, well, made almost nothing, and if I ate nothing but ramen, then we would go ahead and spend basically just the money for the mass, left hand amputer tape out, and that would be that, and if so, could go and do it on $30,000,000, an insanely low number, And that we actually went out and got a debt provider, they wanted to go ahead and lend us the money we needed to go across this threshold.

从那以后，我们开始一个接一个地推动新的目标，再做一件事，再做一件事，再做一件事。

From there, I think it was first to go catalyze a series of other, hey, maybe we can go do one more thing, one more thing, one more thing.

对。

Yeah.

所以这时候我们意识到，如果真想把这家公司做起来，那就必须全力以赴，绝不能敷衍了事。

So we're at this moment where we're like, if we really want to build this company, because like, we're not going to half ass it.

我们不可能去搞什么测试芯片，花上几年时间，眼睁睁看着整个AI市场爆发，而我们却还在原地踏步。

We're not going to go do a test chip and spend years on it and like, let the entire AI market boom while we could be building the product.

既然要做，就要做到底。

Like, if we're going do it, we're going go all the way.

我们得想办法凑到一亿美元。

We're gonna need to find a way to get a $100,000,000.

我记得当时我和Gavin深夜坐在库比蒂诺的办公室里，面面相觑，心想：这地方能不能再砍掉五十万？

I mean, I remember Gavin and I were like sitting down in the office in Cupertino late at night, just like looking at each other and we're like, could we cut 500 k here?

这里能砍掉十万吗？

Could we cut a 100 k here?

我们能说服大家多久不领薪水？

How long could we convince everyone not to take a salary?

我们当时心想，这账根本算不平啊。

And we're like, holy fuck, the math is not going to close.

我们真的得解决这个问题。

We we really need to solve this.

那段时间大概有几周，我们完全进入了生存模式，给所有可能认识投资人的人打电话，说我们需要一亿美元来做这件事。

And there was a period of a few weeks where you kind of just go into survival mode, and you call every person that could possibly know an investor, and you're like, we need a 100,000,000 to do this.

如果我们做成这件事，我们认为这可能会成为历史上最重要的公司之一。

If we do this, we think this could be one of the most important companies of all time.

你认识谁愿意下重注吗？

Do you know somebody that wants to take an aggressive bet?

谁愿意像我们一样，真心相信我们？

That wants to, like, believe in us?

所有的信息都在这儿了。

Like, here's all the information.

我们完全透明，没有任何隐瞒。

Like, we're an open book.

这就是我们的团队。

Here's the team.

都是些非常优秀的人。

It's great people.

我们一直拼命在干。

We've been working super hard.

这些事我们都是破纪录完成的，但还有上百件事等着做。

We've done these things in record time, but we have these, you know, 100 things to go.

你真想做这件事吗？

Like, do you want to do this?

一旦开始滚雪球，这儿来一百万，那儿来两百万，你就会想，好，这个月不用为钱发愁了。

And like, the snowball starts and you get a million here and 2,000,000 here and you're like, okay, we're not going to run out of money this month.

来了五百万的支票，一千万的支票。

You get a $5,000,000 check, $10,000,000 check.

你就会想，嗯，也许能买那些FPGA了。

You're like, okay, maybe I can buy those FPGAs.

后来雪球真的滚起来了，说真的，我们特别幸运，最后把所有事情都凑齐了。

And, you know, the snowball happened where, you know, we were very lucky that we ended up putting it all together.

我们开了一次董事会。

We had a board meeting.