NVIDIA首席执行官黄仁勋

大卫，我得说，我真希望每期节目都能有英伟达的完整制作团队。这次录制时不用操心开关摄像机，也不用担心录制过程中出什么差错，感觉真好。

I will say, David, I would love to have NVIDIA's full production team every episode. It was nice not having to worry about turning the cameras on and off and making sure that nothing bad happened myself while we were recording this.

是啊，光是设备就够呛。我是说，那些从摄像机里取出来的硬盘。

Yeah. Just the gear. I mean, the drives that came out of the camera.

好吧，下期开始家庭工作室要用红龙摄像机了。

Alright. Red cameras for the home studio starting next episode.

没错，太棒了。

Yeah. Great.

好，我们开始吧。

Alright. Let's do it.

是你吗？是你吗？是你吗？现在谁掌握了真相？是你吗？

Is it you? Is it you? Is it you? Who got the truth now? Is it you?

是你吗？是你吗？让我坐下。直说吧。又一个故事即将上演。

Is it you? Is it you? Sit me down. Say it straight. Another story on the way.

谁掌握了真相？

Who got the truth?

欢迎收听本期《Acquired》，这是一档讲述伟大科技公司及其背后故事与成功法则的播客。我是本·吉尔伯特。

Welcome to this episode of Acquired, the podcast about great technology companies and the stories and playbooks behind them. I'm Ben Gilbert.

我是大卫·罗森塔尔。

I'm David Rosenthal.

我们是本期主持人。听众朋友们，先说重点——这期节目对我和David来说简直酷毙了。没错。在花费近两年时间、投入约500小时研究英伟达后，我们专程飞赴其总部与黄仁勋本人进行了对话。这位英伟达创始人兼CEO所执掌的公司，正是推动当前这场AI革命的核心力量。

And we are your hosts. Listeners, just so we don't bury the lead, this episode was insanely cool for David and I. Yeah. After researching NVIDIA for something like five hundred hours over the last two years, we flew down to NVIDIA headquarters to sit down with Jensen himself. And Jensen, of course, is the founder and CEO of NVIDIA, the company powering this whole AI explosion.

录制本期节目时，英伟达市值高达1.1万亿美元，是全球市值第六大企业。此刻正是公司发展的关键转折点，市场预期高得惊人——可以说直冲云霄。在我们研究过的所有企业中，他们拥有最令人叹服的战略地位，与竞争对手的差距之大前所未见。

At the time of recording, NVIDIA is worth $1,100,000,000,000 and is the sixth most valuable company in the entire world. And right now is a crucible moment for the company. Expectations are set high. I mean, sky high. They have about the most impressive strategic position and lead against their competitors of any company that we've ever studied.

但所有人都想知道：英伟达的疯狂增长能否持续？AI会成为下一个万亿美元级的技术浪潮吗？我们对此有多大把握？如果答案是肯定的，随着市场格局逐渐成型，英伟达真能保持这种近乎荒谬的统治地位吗？

But here's the question that everyone is wondering. Will NVIDIA's insane prosperity continue for years to come? Is AI gonna be the next trillion dollar technology wave? How sure are we of that? And if so, can NVIDIA actually maintain their ridiculous dominance as this market comes to take shape?

黄仁勋带我们回顾了从图形处理到数据中心再到AI的转型历程，讲述了他们如何多次绝处逢生。他为创业者提供了大量建议，并在节目尾声动情地分享了创始人的心路历程。

So Jensen takes us down memory lane with stories of how they went from graphics to the data center to AI, how they survived multiple near death experiences. He also has plenty of advice for founders, and he shared an emotional side to the founder journey toward the end of the episode.

确实。尽管自认为做足功课，但这次访谈让我对公司和黄仁勋作为创始人兼领导者都有了全新认知——事实证明我们原先的了解远远不够。

Yeah. I got new perspective on the company and on him as a founder and a leader just from doing this despite you know, we thought we knew everything before we came in advance, and, it turned out we didn't.

看来主角确实懂得更多。

Turns out the protagonist actually knows more.

没错。

Yes.

好了听众朋友们，快来加入Slack讨论组。那里正热烈探讨关于这家AI公司及其生态系统的方方面面，还有我们近期其他节目内容。访问acquired.fm/slack即可加入，期待与你相遇。

Alright. Well, listeners, join the Slack. There is incredible discussion of everything about this company AI, the whole ecosystem, and a bunch of other episodes that we've done recently going on in there right now. So that is acquired.fm/slack. We would love to see you.

免责声明：本节目不构成投资建议。我和David可能持有讨论企业的相关投资，节目内容仅作信息交流与娱乐用途。现在有请黄仁勋。黄先生，欢迎来到《Acquired》，我们想从故事时间开始——让我们把时钟拨回到1997年。

And without further ado, this show is not investment advice. David and I may have investments in the companies we discuss, and this show is for informational and entertainment purposes only. On to Jensen. So, Jensen, this is Acquired, so we wanna start with story time. We wanna wind the clock all the way back to I believe it was 1997.

当时你们正准备发布Riva 128——计算机史上最大规模的图形芯片之一，首个实现完全3D加速的图形管线。但与此同时，公司资金链即将断裂。

You're getting ready to ship the Riva one twenty eight, which is one of the largest graphics chips ever created in the history of computing. It is the first fully three d accelerated graphics pipeline for a computer. Yeah. And you guys have You're running out of money. Of cash left.

于是你决定完全在模拟环境中进行测试，根本不接收实体样机，直接用公司剩余资金盲订了量产。所以你这是在Riva 128上孤注一掷。结果产品面世后，DirectX的32种混合模式它只支持其中8种。

And so you decide to do the entire testing in simulation rather than ever receiving a physical prototype, you commission the production run sight unseen with the rest of the company's money. So you're betting it all right here on the Riva one twenty eight. Yeah. It comes back. And of the 32 DirectX blend modes, it supports eight of them.

你既要说服市场购买它，还得劝开发者只使用这8种混合模式。请详细说说当时——

And you have to convince the market to buy it, and you gotta convince developers not to use anything but those eight blend modes. Walk us through what that

另外24种根本不重要。

The other 24 weren't that important.

等等，第一个问题：这原本就是计划的一部分吗？你们什么时候意识到——

Okay, so wait, wait, first question. Was that the plan all along? Like when did you realize that

我们...我本该更早察觉的，但发现时为时已晚。我们本应实现全部32种模式。但既然木已成舟，我们只能尽力而为。那真是个非常时期。

we I should have even gone for realized I didn't learn about it until it was too late. We should have implemented all 32. Yeah. But we built what we built and so we had to make the best of it. That was really an extraordinary time.

记得Vivo 01/2020是MV3版本吗？MV1和MV2基于正向纹理映射，不用三角形而是曲线，并对曲线进行细分。由于渲染的是更高级的对象，我们本质上避开了Z缓冲。当时以为这是绝佳的渲染方案，结果证明完全错了。所以Riva 128成了我们公司的重置键。

Remember Vivo 01/2020 was MV3. MV1 and MV2 were based on forward texture mapping, no triangles, but curves and it tessellated the curves. And because we were rendering higher level objects, we essentially avoided using Z buffers. And we thought that that was going to be a good rendering approach and turns out to have been completely the wrong answer. And so what Revo Run '28 was, was a reset of our company.

要知道1993年公司创立时，我们是史上首家面向消费者的3D图形公司。当时Windows还是纯软件渲染系统，我们专注将PC改造成加速PC。但等到发现走错路时，微软已推出DirectX——这与英伟达架构根本不相容。尽管我们是行业先驱，此时已冒出30个竞争对手。

Now remember at the time that we started the company in 1993, we were the only consumer three d graphics company ever created. And we were focused on transforming the PC into an accelerated PC because at the time Windows was really a software rendered system. And so anyways, Riva 01/1928 was a reset of our company because by the time that we realized we had gone down the wrong road, Microsoft had already rolled out DirectX. It was fundamentally incompatible with NVIDIA's architecture. 30 competitors have already shown up, even though we were the first company at the time that we were founded.

世界已天翻地覆。关于公司战略的抉择...坦白说我们犯过很多错误。但在那个关键节点，我们做出了一连串极其正确的决策。1997年可能是英伟达最辉煌的时刻，因为当时我们已被逼到绝境。

So the world was a completely different place. The question about what to do as a company strategy at that point, I would have said that we made a whole bunch of wrong decisions. But on that day that mattered, we made a sequence of extraordinarily good decisions. And that time, 1997, was probably NVIDIA's best moment. And the reason for that was our backs were up against the wall.

时间所剩无几，资金即将耗尽，许多员工开始失去希望。我们该怎么办？首先我们认清现实：DirectX已成既定标准。

We were running out of time. We're running out of money. And for a lot of employees, running out of hope. And the question is, what do we do? Well, the first thing that we did was we decided that, look, DirectX is now here.

我们决定不再对抗，转而思考如何为其打造最完美的硬件。Riva 128成为世界上首个完全硬件加速的3D渲染管线——从坐标变换、投影到帧缓冲，每个环节都由硬件加速。我们还实现了纹理缓存。

We're not gonna fight it. Let's go figure out a way to build the best thing in the world for it. Rivo one twenty eight is the world's first fully accelerated hardware accelerated pipeline for rendering three d. And so the transform, the projection, every single element all the way down to the frame buffer was completely hardware accelerated. We implemented a texture cache.

我们将总线限制、帧缓冲限制推至当时物理条件允许的最大值。我们制造了人们前所未想的最大芯片，采用了最快的存储器。本质上，如果我们造出那枚芯片，世上便无更快之物。同时，我们设定的成本点远高于预估竞争对手所能承受的最高价位。

We took the bus limit, the frame buffer limit to as big as physics could afford at the time. We made the biggest chip that anybody had ever imagined building. We used the fastest memories. Basically, if we built that chip, there could be nothing that could be faster. And we also chose a cost point that is substantially higher than the highest price that we think that any of our competitors would be willing to go.

若我们设计得当，加速所有环节，完整实现DirectX功能，并尽可能扩大规模，那么显然无人能造出更快之物。

If we built it right, we accelerated everything, we implement everything in DirectX that we knew of, and we build it as larger as we possibly could, then obviously nobody can build something faster than that.

如今在某种程度上，英伟达仍在延续这种做法。你们当时是消费产品公司对吧？最终买单的是终端消费者。

Today, in a way, you kind of do that here at NVIDIA too. You were a consumer products company back then, right? It was end consumers who were gonna have to pay the money to buy them.

没错。但我们观察到当时PC产业尚未成熟，市场存在渴求——人人都追逐更快的设备。若当年性能能达到市售产品的十倍，我们确信庞大的发烧友市场会趋之若鹜。事实证明PC产业确实存在愿为顶尖技术买单的巨大发烧友市场。

That's right. But we observed that there was a segment of the market where people were because at the time the PC industry was still coming up and it wasn't good enough. Everybody was clamoring for the next fastest thing. And so if your performance was 10 times higher this year than what was available, there's a whole large market of enthusiasts who we believe would have gone after it. And we were absolutely right that the PC industry had a substantially large enthusiast market that would buy the best of everything.

这一现象至今仍在特定领域成立。比如3D图形技术永远不够完美——当年我们选中这项技术正是因其永无止境的改进空间。我们称之为可持续技术机遇，正因它永远不够好，技术才能持续进步。

To this day, kind of remains true. And for certain segments of the market where the technology is never good enough, three d graphics, when we chose the right technology, three d graphics is never good enough. And we call it back then, three d gives us sustainable technology opportunity because it's never good enough. And so your technology can keep getting better. We chose that.

我们还决定采用名为仿真的技术。有家叫IQOS的公司，在我联系当日恰逢倒闭——他们已无客户。我说愿买下全部库存且无需承诺。选择仿真器的原因在于：若按常规流程，芯片流片后开始软件开发，待发现所有bug再重新流片...

We also made the decision to use this technology called emulation. There was a company called IQOS. And on the day that I called them, they were just shutting the company down because they had no customers. And I said, hey, look, I'll buy what you have in inventory and no promises are necessary. The reason why we needed that emulator is because if you figure out how much money that we have, if we taped out a chip and we got it back from the fab and we started working on our software, By the time that we found all the bugs, because we did the software, then we taped out the chip again.

公司早就破产了。

Well, we would have been out of business already.

而你们的竞争对手早已追赶上来。

And I And your competitors would have caught up.

更不用说我们早就...

Well, not to mention we would have been out

破产了。谁

of business. Who

在乎吗？没错。所以如果你反正都要倒闭，那个计划显然不是真正的计划。公司通常遵循的流程——设计芯片、编写软件、修复漏洞、流片新芯片，如此循环——这种方法根本行不通。

cares? Exactly. So if you're gonna be out of business anyways, that plan obviously wasn't the plan. The plan that companies normally go through, which is build the chip, write the software, fix the bugs, tape out the new chip, so on and so forth. That method wasn't gonna work.

于是问题来了：如果我们只有六个月时间且只能流片一次，那你显然会流出一颗完美芯片。我记得和团队领导讨论时他们说：‘黄仁勋，你怎么保证它能完美？’我回答：‘我知道它会完美，因为如果不完美，我们就破产了。所以我们必须让它完美，我们只有一次机会。’

And so the question is, if we only had six months and you get the tape out just one time, then obviously you're gonna tape out a perfect chip. So I remember having conversation with our leaders and they said, but Jensen, how do you know it's gonna be perfect? And I said, I know it's gonna be perfect because if it's not, we'll be out of business. And so let's make it perfect. We get one shot.

我们通过购买这台仿真器实现了芯片的虚拟原型。德怀特和软件团队编写了整个软件栈，在仿真器上运行程序，然后整组人就坐在实验室里等着画面渲染出来。

We essentially virtually prototype the chip by buying this emulator. And Dwight and the software team wrote our software, the entire stack and ran it on this emulator and just sat in the lab waiting for windows to paint.

当时每帧要六十秒？哦，可能更久。

It was like sixty seconds per frame or Oh, something like easily.

我记得是每小时一帧的样子。我们就这么干坐着看渲染进度。到了决定流片那天，我认定芯片已经完美无缺，所有能测试的都提前测试过了，于是告诉大家：‘就这样了，我们要流片了，它会是完美的。’如果你确信流片的芯片完美无缺，接下来该做什么？

I actually think that it was an hour per frame, something like that. And so we just sit there and watch a paint. And so on the day that we decided to tape out, I assumed that the chip was perfect and everything that we could have tested, we tested in advance and told everybody this is it. We're gonna tape out the chip, it's gonna be perfect. Well, if you're gonna tape out a chip and you know it's perfect, then what else would you do?

这才是关键问题。如果你按下回车键流片时确信芯片完美无缺，接下来该做什么？答案当然是直接量产。

That's actually the good question. If you knew that you hit enter, you taped out a chip and you knew it was gonna be perfect, then what else would you do? Well, the answer obviously go to production.

还有营销轰炸。

And marketing blitz.

对，没错。

Yeah. Yeah.

以及开发者关系维护。

And developer relations.

因为既然有了完美芯片，我们就坚信不疑。

Off because Kick you got a perfect chip. And so we got in our head that we have a perfect chip.

这其中有多少是你的主意，又有多少是来自你的联合创始人、公司其他成员或董事会？是不是所有人都觉得你疯了？

How much of this was you and how much of this was, like, your cofounders, the rest of the company, the board? Was everybody telling you you were crazy?

不，所有人都清楚我们毫无胜算。不这么做才叫疯狂，因为——

No, everybody was clear we had no shot. Not doing it would be crazy because otherwise

你可能会

you might

反正横竖都要破产。所以除此之外的任何选择都是疯狂的。这看起来是相当合乎逻辑的决定。说实话，现在听我这么描述，你们大概会觉得确实挺合理。但关键是，它成功了。

as You're well go gonna be out of business anyways. So anything aside from that is crazy. So it seemed like a fairly logical thing. And quite frankly, right now, just I'm describing it, everybody's you're probably thinking, yeah, it's pretty sensible. Well, it worked.

没错。所以我们直接跳过了测试环节投入量产。

Yeah. And so we take that out and went directly to production.

那么给创业者的启示就是：当你对像Revo 128或CUDA这样的项目有坚定信念时，就该押上整个公司。这招对你屡试不爽。所以你的经验教训似乎是——没错，继续全押筹码，因为至今每次都能成功。你怎么看待这个策略？

So is the lesson for founders out there when you have conviction on something like the Revo one twenty eight or, CUDA, go bet the company on it. And this keeps working for you. So it seems like your lesson learned from this is, yes, keep pushing all the chips in because so far it's worked every time. How do you think about that?

不，不是的。我全押筹码时，是确信它能成功。注意，我们假设流片的是完美芯片。之所以能流片完美芯片，是因为我们在流片前模拟了整个芯片。

No. No. When you push your chips in, I know it's gonna work. Notice, we assume that we taped out a perfect chip. The reason why we taped out a perfect chip is because we emulated the whole chip before we taped it out.

我们开发了完整的软件栈，对所有驱动和软件进行了质量测试。跑遍了手头所有游戏，测试了每个VGA应用程序。所以当你全押筹码时，本质上是在说：我要把未来所有风险事项提前解决。

We developed the entire software stack. We ran QA on all the drivers and all the software. We ran all the games we had. We ran every VGA application we had. And so when you push your chips in, what you're really doing is when you bet the farm, you're saying, I'm gonna take everything in the future, all the risky things and I pull it in advance.

这可能才是真正的经验。直到今天，我们能预取的所有未来事项，能模拟的所有可能性，都会提前处理。

And that is probably the lesson. And to this day, everything that we can prefetch, everything in the future that we can simulate today, we prefetch it.

我们经常讨论这点，刚在Costco主题节目里就谈到——全押筹码的前提是你确信它能成功。

And we talk about this a lot. We were just talking about this on our Costco episode. You wanna push your chips in when you know it's gonna work.

所以每次我们看到你做出一个‘是的’。打赌公司已经模拟了这个动作。你知道吗？

So every time we see you make a Yeah. Bet the company move you've already simulated it. You know?

是的。是的。是的。

Yeah. Yeah. Yeah.

你觉得CUDA的情况也是如此吗？

Do you feel like that was the case with CUDA?

是的。事实上，在CUDA之前，有CG。对吧。所以我们已经在探索这个概念：如何在我们芯片之上创建一个可以用更高级语言和表达方式描述的抽象层。我们如何利用GPU进行CT重建、图像处理等任务。

Yeah. In fact, before there was CUDA, there was CG. Right. And so we were already playing with the concept of how do we create an abstraction layer above our chip that is expressible in a higher level language and higher level expression. How can we use our GPU for things like CT reconstruction, image processing.

我们已经沿着这条路走下去了。所以有一些积极的反馈和一些直觉上的积极反馈让我们认为通用计算是可能的。如果你看看可编程着色器的流水线，它就是一个处理器，高度并行，大规模多线程，是世界上唯一能做到这一点的处理器。因此，可编程着色器的许多特性表明，CUDA有很大的成功机会。

We were already down that path. And so there were some positive feedback and some intuitive positive feedback that we think that general purpose computing could be possible. And if you just looked at the pipeline of a programmable shader, it is a processor and is highly parallel and it is massively threaded and it is the only processor in the world that does that. And so there were a lot of characteristics about programmable shading that would suggest that CUDA has a great opportunity to succeed.

如果有一个庞大的机器学习从业者市场，他们最终会出现并想要进行所有这些伟大的科学计算和加速计算，那么这是真的。但在你开始投入现在大约一万个人年的时候，是的，在构建那个平台时，是的，你有没有觉得，‘哦，天哪。我们可能在机器学习需求之前就进行了投资，因为我们比全世界意识到这一点早了大约十年？’

And that is true if there was a large market of machine learning practitioners who would eventually show up and wanna do all this great scientific computing and accelerated computing. But at the time when you were starting to invest what is now something like ten thousand person years Yeah. In building that platform Yeah. Did you ever feel like, oh, man. We might have invested ahead of the demand for machine learning since we're like a decade before the whole world is realizing it?

我想说是也不是。你知道，当我们看到深度学习，当我们看到AlexNet并意识到它在计算机视觉中的惊人效果时，我们有足够的智慧，如果你愿意这么说的话，回到第一原则并问，是什么让这个东西如此成功？当一个新的软件技术、一个新的算法出现，并突然超越了三十年的计算机视觉工作时，你必须退一步问自己，但为什么？从根本上说，它是可扩展的吗？如果它是可扩展的，它还能解决什么问题？

I guess yes and no. You know, when we saw deep learning, when we saw AlexNet and realized its incredible effectiveness in computer vision, we had the good sense, if you will, to go back to first principles and ask, what is it about this thing that made it so successful? When a new software technology, a new algorithm comes along and somehow leapfrogs thirty years of computer vision work, you have to take a step back and ask yourself, but why? And fundamentally, is it scalable? And if it's scalable, what other problems can it solve?

我们有几个观察结果。第一个观察当然是，如果你有大量的示例数据，你可以教会这个函数进行预测。嗯，我们基本上发现了一个通用函数逼近器，因为维度可以尽可能高。因为每一层都是一次训练一层，没有理由不能构建非常非常深的神经网络。好吧。

And there were several observations that we made. The first observation, of course, is that if you have a whole lot of example data, you could teach this function to make predictions. Well, what we've basically done is discovered a universal function approximator because the dimensionality could be as high as you want it to be. Because each layer is trained one layer at a time, there's no reason why you can't make very, very deep neural networks. Okay.

所以现在你只需要推理一下。对吧？好吧。所以我回到十二年前。你可以想象我脑子里正在进行的推理，我们发现了通用函数逼近器。

So now you just reason your way through. Right? Okay. So now I go back to twelve years ago. You could just imagine the reasoning I'm going through in my head that we've discovered a universal function approximator.

事实上，我们可能还发现了另外几项技术，一个通用计算机，你可以而且你已经可以

In fact, we might've discovered with a couple more technologies, a universal computer that you And you've can

关注ImageNet竞赛。是的。你正在引导这个话题。

paying attention to the ImageNet competition. Yeah. You're leading up to this.

没错。当时我们已经在研究计算机视觉，试图让CUDA成为优秀的视觉系统。但多数计算机视觉算法并不适配CUDA架构。所以我们当时正苦苦思索解决方案。

Yeah. Yeah. And the reason for that is because we were already working on computer vision at the time, and we were trying to get CUDA to be a good computer vision system. Or most of the algorithms that were created for computer vision aren't a good fit for CUDA. And so we're sitting there trying to figure it out.

突然AlexNet横空出世。这简直令人着迷——它的效果如此惊艳，迫使你退后一步思考：为什么它能成功？当你理清思路后就会意识到：在通用函数逼近器的世界里，哪些问题...

All of a sudden AlexNet shows up. And so that was incredibly intriguing. It's so effective that it makes you take a step back and ask yourself, why is that happening? So by the time that you reason your way through this, you go, well, what are the kind of problems in a world where a universal function approximator

能够...

can Well,

我们知道多数算法始于科学原理。你想理解因果关系，并据此创建可扩展的模拟算法。但对许多问题而言，我们并不关心因果，只在乎可预测性。

we know that most of our algorithms start from principled sciences. Okay. You wanna understand the causality and from the causality, you create a simulation algorithm that allows us to scale. Well, for a lot of problems, we kind of don't care about the causality. We just care about the predictability of it.

比如我真的需要知道你为什么选这款牙膏吗？因果不重要，我只需要准确预测你的选择。购买热狗的人会搭配番茄酱和芥末——根本原因重要吗？无关紧要。

Like, do I really care for what reason you prefer this toothpaste over that? I don't really care the causality. I just wanna know that this is the one you would have predicted. Do I really care that the fundamental cause of somebody who buys a hotdog, buys ketchup and mustard? It doesn't really matter.

重要的是可预测性。这适用于电影推荐、音乐推荐，甚至天气预报。我们懂得热力学原理，理解太阳辐射...

It only matters that I can predict it. It applies to predicting movies, predicting music. It applies to predicting, quite frankly, weather. We understand thermodynamics. We understand radiation from the sun.

了解云层效应，掌握海洋影响，通晓所有这些变量——最终只想知道今天该不该穿毛衣，不是吗？

We understand cloud effects. We understand oceanic effects. We understand all these different things. We just want to know whether we should wear sweat or not. Isn't that right?

没错。世上多数问题都不需要追究因果，我们只需模拟系统并预测结果。

Yep. And so causality for a lot of problems in the world doesn't matter. We just wanna emulate the system and predict the outcome.

这能创造惊人利润。若能预测社交媒体下个爆款内容，那将是...

And it can be an incredibly lucrative market. If you can predict what the next best performing feed item to serve into a social media feed, turns out that's a

极具价值的事物。

hugely valuable thing.

说到这个，我非常喜欢你举的例子，比如牙膏、番茄酱、音乐、电影。

With that. I love the examples you pulled, you know. Toothpaste, ketchup, music, movies.

当你意识到这一点时，你会突然明白——等等，一个通用的函数逼近器，一个机器学习系统，一种能从示例中学习的工具，可能带来巨大的机遇，因为它的应用范围极其广泛。从我们刚谈到的商业领域一直到科学领域。于是你意识到，这可能会影响世界上很大一部分产业。几乎世界上所有的软件最终都可能以这种方式编写。

And when you realize this, you realize, hang hang on a second. A universal functional approximator, a machine learning system, you know, something that learns from examples could have tremendous opportunities because it's just the number of applications is quite enormous. And everything from obviously we just started talking about commerce all the way to science. And so you realize that maybe this could affect a very large part of the world's industries. Almost every piece of software in the world would eventually be programmed this way.

如果是这样的话，那么计算机和芯片的构建方式实际上可以彻底改变。剩下的问题只是，你有没有勇气把筹码押在这上面？

And if that's the case, then how you build a computer and how you build a chip, in fact, can be completely changed. And realizing that the rest of it is just comes with, you know, do you have the courage to put your chips behind it?

这就是我们现在的处境，也是NVIDIA现在的处境。嗯。但我很好奇，在AlexNet之后的几年里，当时我和Ben刚进入科技行业和风投行业。

So that's where we are today, and that's where NVIDIA is today. Mhmm. But I'm curious in the know, there's a couple years after AlexNet. Mhmm. And this is when Ben and I were getting into the technology industry and the venture industry ourselves.

我2012年加入微软。

I started at Microsoft in 2012.

嗯。

Yeah.

是的，就在AlexNet之后，但在机器学习甚至主流工程界开始讨论之前。

Yes. So right after Alex Knapp, but before anyone was talking about machine learning and even the mainstream engineering community.

有那么几年，对世界上很多人来说，这些看起来像是科学实验。是的。硅谷的科技公司，尤其是社交媒体公司，如谷歌、Facebook、Netflix等，正从中获得巨大的经济价值。显然，这后来催生了许多事情，包括几年后的OpenAI。但在那几年里，当你看到硅谷释放出如此巨大的经济价值时，你当时是什么感觉？

There were those couple of gears there where to a lot of the rest of the world, these looked like science projects. Yeah. The technology companies here in Silicon Valley, particularly the social media companies, they were just realizing huge economic value out of this, the Googles, Facebooks, the Netflixs, etcetera. And obviously that led to lots of things, including OpenAI a couple years later. But during those couple years, when you saw just that huge economic value unlock here in Silicon Valley, How are you feeling during those times?

首先想到的当然是如何改变我们的计算架构。其次是在哪里能找到最早的应用可能性？如果我们建造这样的计算机，人们会用它来做什么？幸运的是，与全球大学和研究人员的合作是我们公司的天性，因为我们已经在开发CUDA，而CUDA的早期使用者正是研究人员，因为我们让超级计算民主化了。CUDA不仅仅用于AI。

The first thought was, of course, reasoning about how we we should change our computing stack. The second thought is where can we find earliest possibilities of use? If we were to go build this computer, what would people use it to do? And we were fortunate that working with the world's universities and researchers was innate in our company because we were already working on CUDA and CUDA's early adopters were researchers because we democratize supercomputing. CUDA is not just used as you know, for AI.

CUDA几乎应用于所有科学领域，从分子动力学到成像、CTB构建、地震处理、天气模拟、量子化学，应用不胜枚举。因此CUDA在科研中的运用极为广泛。当意识到深度学习可能极具潜力时，我们很自然地联系全球每位AI研究者——包括Yan LeCun、Andrew Ng和Jeff Hinton——询问如何助力他们的研究。

CUDA is used for almost all fields of science. Everything from molecular dynamics to imaging, CTB construction, to seismic processing, to weather simulations, quantum chemistry, the list goes on. So the number of applications of CUDA in research was very high. And so when the time came and we realized that deep learning could be really interesting, it was natural for us to go back to the researchers and find every single AI researcher on the planet and say, how can we help you advance your work? And that included Yan LeCun and Andrew Ng and Jeff Hinton.

就这样我结识了这些学者。我常参加各类AI会议，正是在那里初次见到Iliya Suskabur。当时的核心问题是：我们能构建怎样的系统和软件栈来推动研究进展？尽管当时GAN这类技术看起来像玩具——记得初次见到Goodfellow时，GAN生成的只是32x32像素的模糊猫图——但我们坚信其发展潜力。

And that's how I met all these people. And I used to go to all the AI conferences and that's where I met Iliya Suskabur there for the first time. And so it was really about at that point, what are the systems that we can build and the software stacks we can build to help you be more successful to advance the research? Because at the time it looked like a toy, but we had confidence that even GAN, the first time I met Goodfellow, the GAN was like 32 by 32. And it was just a blurry image of a cat, but how far can it go?

我们深信两点：其一，深度学习具有可扩展性，因其分层训练机制允许数据集和模型的持续扩大；其二，规模越大性能越优。是的，这很符合逻辑。

And so we believed in it. We believed that one, you could scale deep learning because obviously it's trained layer by layer and you could make the datasets larger and you could make the models larger. And we believe that if you made that larger and larger, it would get better and better. Yep. Kind of sensible.

我认为与研究者们的讨论互动构成了我们所需的正向反馈系统。所有的突破都源于那个研究交流的闭环——我不断往返于研究前线，那里才是一切发生的源头。

And I think the discussions and the engagements with the researchers was the exact positive feedback system that we needed. I would go back to research. It was that's where it all happened.

OpenAI成立是在20...呃...

When OpenAI was founded in '20 Yeah.

15年？

'15?

对。如今回想那确实是关键转折点，但当时即便科技圈内人也多持观望态度——"这到底是什么？"

Yeah. I mean, that was such an important moment that's obvious today now. But at the time, I I think most people, even people in tech, were like, what is this? Yeah. Yeah.

当时你有参与吗？毕竟你与Iliya等研究者关系密切，某种程度上是将人才从谷歌、Facebook引流出来——直白地说——但重组研究社区并开放资源确实意义重大。你参与其中了吗？

Yeah. Were were you involved in it at all? Like, you know, because you were so connected to the researchers, to Ilia, taking that talent out of Google and Facebook to be blunt, but reseeding the research community Yeah. And opening it up was such an important moment. Were you involved in it at all?

虽未参与创立，但认识其中多位核心成员：Elon、Peter Beale、Ilya等。如今我们有些优秀员工正是创始成员。当时我就知道他们需要正在研发的超级计算机。

I wasn't involved in the founding of it, but I knew a lot of the people there. And Elon, of course, I knew. And Peter Beale was there, and Ilya was there. And we have we have some great employees today that were there in the beginning. And I knew that they needed this amazing computer that we were building.

我们正在开发初代DGX系统——如今Hopper重70磅含3.5万零件承载1万安电流，但当年交付OpenAI的首台DGX仅是内部试用版。那天令人难忘。早期成功关键在于助力研究者突破边界：我清楚当前技术尚不成熟，但确信几次迭代后必将惊艳世界。

And we're building the first version of the DGX, which, you know, today when you see a hopper, it's 70 pounds, 35,000 parts, 10,000 amps. But DGX, the first version that we built was used internally and I delivered the first one to OpenAI. That was a fun day. But most of our success was aligned around in the beginning just about helping the researchers get to the next level. I knew it wasn't very useful in its current state, but I also believe that in a few clicks, it could be really remarkable.

这种信念体系源于与众多杰出研究者的互动，也源于亲眼见证逐步的进展。最初，论文每三个月才发表一篇，而如今每天都有新论文问世。我通过持续关注arXiv上的论文，怀着对深度学习进展的兴趣，尽己所能阅读这些文献。你能实时目睹进步的发生，呈指数级增长的实时演进。

And that belief system came from the interactions with all these amazing researchers and it came from just seeing the incremental progress. At first, the papers were coming out every three months, and then papers today are coming out every day. So you could just monitor the archive papers, I took an interest in learning about the progress of deep learning, and and to the best of my ability, read these papers. And you could just see the progress happening, you know, in real time, exponentially in real time.

据我们采访的一些研究者表示，业界似乎没人预料到单纯扩大模型规模竟能让语言模型变得如此实用。他们原以为必须进行某种算法变革。但当参数突破100亿大关，尤其是超过1000亿后，模型突然变得精准度更高、实用性更强、拟真度更佳。你首次见到真正的大型语言模型时是否感到震惊？还记得当时的感受吗？

It even seems like within the industry, from some researchers we spoke with, it seemed like no one predicted how useful language models would become when you just increase the size of the models. They thought, oh, there has to be some algorithmic change that needs to happen. But once you cross that 10,000,000,000 parameter mark and certainly once you cross the 100,000,000,000, they just magically got much more accurate, much more useful, much more lifelike. Were you shocked by that the first time you saw a truly large language model? And do you remember that feeling?

我对语言模型的初印象是其设计之精妙——通过遮蔽词语来预测下一个词。这是自监督学习的巅峰之作。我们拥有海量文本数据，既然已知正确答案，就让模型来猜测。

Well, my first feeling about the language model was how clever it was to just mask out words and make it predict the next word. It's self supervised learning at its best. We have all this text. You know, I know what the answer is. I'll just make you guess it.

因此BERT给我的第一感觉就是惊艳。随后的问题是如何扩展它？对任何新事物，首先要发现其趣味性，然后直观理解其原理，最后从第一性原理出发推演发展路径。我们当时就预见到BERT必将走向更大规模。

And so my first impression of BERT was really how clever it was. And now the question is how can you scale that? The first observation on almost everything is interesting, and then try to understand intuitively why it works. And then the next step, of course, is from first principles, how would you extrapolate that? And so obviously we knew that BERT was gonna be a lot larger.

这些语言模型的特性之一在于信息编码，对吧？它们本质是信息压缩器。在人类语言和文本中，本就蕴含着大量逻辑推理的编码痕迹，我们本就习惯用文字描述推理过程。

Now, one of the things about these language models is it's encoding information. Isn't that right? It's compressing information. And so within the world's languages and texts, there's a fair amount of reasoning that's encoded in it. We describe a lot of reasoning things.

所以如果说多步推理能力可以通过阅读习得，我毫不意外。人类本就通过阅读获取常识和推理能力，机器学习模型为何不能从中习得部分推理能力？而基于推理能力，自然可能涌现出更多新能力。沉浸式学习与推理能力的发展本就是相辅相成的。

And so if you were to say that few step reasoning is somehow learnable from just reading things, I wouldn't be surprised. A lot of us, we get our common sense and we get our reasoning ability by reading. And so why wouldn't a machine learning model also learn some of the reasoning capabilities from that? And from reasoning capabilities, you could have emergent capabilities. Immersion abilities are consistent with intuitively from reasoning.

某些方面或许可预测，但这丝毫不减其神奇。合乎逻辑的事实反而更显非凡。

And so some of it could be predictable, but still, it's still amazing. The fact that it's sensible doesn't make it any less amazing.

确实。

Right.

我能在脑海中完整构建自动驾驶汽车的整个计算机系统及所有模块。看到它始终保持在车道内行驶，这种成就感让我欣喜若狂。

I could visualize literally the entire computer and and all the modules in a self driving car. And the fact that it's still keeping lanes makes me insanely happy. And so

这让我想起大学第一次操作系统课，当我终于通过这门课将编程语言与电气工程课程的知识贯通时，那种顿悟感——从底层到顶层理解了冯·诺依曼计算机的完整运作机制，至今想来仍是奇迹。

I even remember that from my first operating systems class college when I finally figured out all the way from programming language to the electrical engineering classes bridged in the middle by that OS class. I'm like, oh, I think I understand how the Von Neumann computer works soup to nuts, and it's still a miracle.

对。对。对。对。完全正确。

Yeah. Yeah. Yeah. Yeah. Exactly.

对。对。当你把所有东西拼在一起时，这仍然是个奇迹。没错。

Yeah. Yeah. When you put it all together, it's still a miracle. Yeah.

好的，听众朋友们。现在是个绝佳时机向大家介绍我们节目非常期待的新伙伴——WorkOS。

Okay, listeners. Now is a great time to tell you about a new friend of the show we are very excited about, WorkOS.

是的。WorkOS是企业级就绪平台，被OpenAI、Cursor、Perplexity、Vercel、Plaid等数百家成功企业所采用。

Yes. WorkOS is the enterprise ready platform used by OpenAI, Cursor, Perplexity, Vercel, Plaid, and literally hundreds of other winning companies.

那么这些公司都用WorkOS做什么呢？想象你是一家快速发展的初创公司。你已实现产品市场匹配，并开始收到大型企业客户的主动询盘。非常激动人心。但随后他们发来了安全调查问卷。

So what are all these companies using WorkOS for? Imagine you're a fast growing startup. You've got product market fit, and you're getting inbound interest from big enterprise customers. Very exciting. But then they send you their security questionnaire.

没错。这份问卷长达47页，里面的要求读起来像字母汤。你们支持SAML 2.0吗？能对接我们的Okta吗？有SCIM用户配置功能吗（s-c-i-m）？

Yep. And it's, like, 47 pages long with requirements that kinda sound like alphabet soup. Do you support saml two dot o? Can you integrate with our Okta? Do you have SCIM provisioning, s c I m?

那RBAC呢（r-b-a-c）？而你心里想着：我连这些缩写是什么意思都不知道，更别说怎么实现了。

What about RBAC, r b a c? And you're thinking, I have no idea what these acronyms even mean, let alone how to implement them.

关键在于：这些不是锦上添花的功能，而是交易拦路虎。没有单点登录（SSO）、没有SCIM、没有RBAC、没有审计日志，你根本不可能达成企业级交易，就这么简单。

So here's the thing. These are not nice to haves. These are deal blockers. Without SSO, without SCIM, without RBAC, without audit logs, you simply cannot close enterprise deals, period.

但这些功能都不会让你的核心产品变得更好。用我们在Acquired节目里最爱的比喻来说：它们不会让啤酒更好喝。如果你正在开发设计工具，花六个月构建SAML认证并不会让你的设计工具更强大。

But none of these features make your core product better. They don't make your beer taste better, to use our favorite analogy here on Acquired. So if you're building like a design tool, spending six months building SAML authentication doesn't make your design tool more powerful.

这时WorkOS就派上用场了。他们为企业功能打造了'Stripe式'解决方案。WorkOS将企业认证需求转化为即插即用的API，尽可能消除不必要的复杂性。

So this is where WorkOS comes in. They've built Stripe for enterprise features. WorkOS turns enterprise authentication requirements into drop in APIs, abstracting away as much unnecessary complexity as possible.

因此，你的团队无需花费数月研读SAML规范，几分钟内就能实现企业级单点登录。WorkOS负责用户配置、权限管理、审计日志等所有企业IT要求的必备功能。

So instead of your team spending months reading SAML specs, you can implement enterprise SSO in minutes. WorkOS handles user provisioning, permissions, audit logs, all the checkbox items that enterprise IT requires.

所以无论你是初创公司想赢得首个企业客户，还是已具规模正全球扩张，WorkOS都是最快实现企业级准备就绪的途径。

So whether you are a seed stage company trying to land your first enterprise customer or already big and expanding globally, WorkOS is the fastest path to becoming enterprise ready.

只需访问workos.com或直接联系他们的Slack支持。那里有真正的工程师快速解答问题。联系时记得提是本和大卫推荐你的。

Just visit workos.com or just message their Slack support. They have real engineers in there who answer questions fast. And when you get in touch, just tell them Ben and David sent you.

我们有几个问题想请教。有些是关于英伟达文化的，但其他问题普遍适用于公司建设。首先我们听说你有40多位直接下属，这种组织架构与传统公司截然不同。你认为英伟达有什么特别之处，能让你拥有这么多直接汇报者，而不必费心呵护或关注高管的职业发展？你似乎对此不以为然。

We have some questions we wanna ask you. Some are cultural about NVIDIA, but, others are generalizable to company building broadly. And the first one that we wanted to ask is, we've heard that you have 40 plus direct reports and that this org chart works a lot differently than a traditional company org chart. Do you think there's something special about NVIDIA that makes you able to have so many direct reports, not worry about coddling or focusing on career growth of your executives? And you're like, no.

你来这里就是要全力以赴做出世界上最了不起的工作。现在就去干吧。

You're just here to do your freaking best work and the most important thing in the world. Now go.

嗯。

Mhmm.

第一，这个情况属实吗？第二，英伟达是否有特殊机制使之成为可能？

A, is that correct? And b, is there something special about NVIDIA that enables that?

我不认为这是英伟达的特例。我认为是我们有勇气建立这样的体系。英伟达不是按军队模式构建的，没有那种将军上校的等级制度，我们根本不是那样设置的。

I don't think it's something special with NVIDIA. I think that we had the courage to build a system like this. NVIDIA is not built like a military. It's not built like a like the armed forces where you have, you know, generals and colonels. You know, we just we're not set up like that.

我们没有建立自上而下的命令控制与信息分发系统。公司更像一个计算堆栈：最底层是架构，之上是芯片，再往上是软件。每层都有不同模块，每个模块都由人管理。所以对我来说，公司架构就像台计算机——计算堆栈里的人们管理着系统不同部分，汇报关系与你在堆栈中的位置无关。

We're not set up in a command and control and information distribution system from the top down. We're really built much more like a computing stack. And a computing stack, the lowest layer is our architecture and then there's our chip and then there's our software. On top of it, there are all these different modules and each one of these layers of modules are people. And so the architecture of the company to me is a computer with a computing stack with people managing different parts of the system and who reports to whom Your title is not related to anywhere you are in the stack.

关键在于谁最擅长运作那个功能层的模块，谁就是负责人，那个人就是指挥飞行员。这是我们的一个特质。你们...

It just happens to be who is the best at running that module on that function on that layer. It is in charge, and that person is the pilot in command. And so that's one characteristic. Have you

一直以来都是这样看待公司的吗？从最初期就是如此。

always thought about the company this way? Even from the earliest days.

是的，基本上是这样。因为你的组织架构应该是构建产品这台机器的设计蓝图，对吧？

Yeah. Pretty much. Yeah. And the reason for that is because your organization should be the architecture of the machinery of building the product. Right?

这就是公司的本质。然而每家公司的架构看起来都一模一样，却生产着不同的东西。这怎么可能合理？你明白我的意思吗？是的。

That's what a company is. And yet everybody's company looks exactly the same, but they all build different things. How does that make any sense? Do you see what I'm saying? Yeah.

制作炸鸡、翻转汉堡与烹饪中式炒饭的工艺流程截然不同。那么为什么生产机器和流程会完全相同？在我看来，大多数公司的组织架构图都大同小异，这根本不合理——这边是一个业务部门，那边是另一个业务部门，全都号称自主运营。这些安排对我而言毫无意义。

How you make fried chicken versus how you flip burgers versus how you make Chinese fried rice is different. And so why would the machinery, why would the process be exactly the same? And so it's not sensible to me that if you look at the org charts of most companies, it all kind of looks like this. And then you have one group that's for a business and you have another for another business, you have another for another business, and they're all kind of supposedly autonomous. And so none of that stuff makes any sense to me.

关键取决于我们要构建什么，以及什么样的公司架构最适合实现它。这是首要原则。在信息系统和促进协作方面，我们搭建了类似神经网络的结构。公司里有句口号叫'使命就是老板'，所以我们会明确界定使命究竟是什么。

It just depends on what is it that we're trying to build and what is the architecture of the company that best suits to go build it. So that's number one. In terms of information system and how do you enable collaboration, we kind of wired up like a neural network. And the way that we say is that there's a phrase in the company called mission is the boss. So we figure out what is the mission of what is the mission?

然后我们集结最优秀的技能、团队和资源来实现这个使命。这种方式会打破传统组织边界，看似不合常理，但确实有点像神经网络运作模式，明白吗？

And we go wire up the best skills and the best teams and the best resources to achieve that mission. And it cuts across the entire organization in a way that doesn't make any sense, but it looks like a little bit like a neural network. You know?

你所说的使命，是指像英伟达'打造Hopper芯片'这样的具体目标？而不是更宽泛的'加速计算'领域？

When you say mission, you mean mission like NVIDIA's mission is Build Hopper. Yeah. Okay. So it's not like further accelerated computing. No.

就像我们现在要交付DGX云服务这样具体。

It's like we're shipping DGX Cloud.

打造Hopper芯片，或是为Hopper构建系统，或是开发适配Hopper的CUDA，或是为Hopper的CUDA开发CUDNN——这些就是具体使命，明白吗？

Build Hopper or somebody else's, build a system for Hopper. Somebody has, build CUDA for Hopper. Somebody's job is build CUDNN for CUDA for Hopper. Somebody's job is the mission. Right?

所以你看，每个人的使命就是完成某个具体任务。

Is is so, you know, your mission is to do something.

与传统结构相比，这种模式有哪些权衡取舍？

What are the trade offs associated with that versus the traditional structure?

弊端在于领导者的压力相当大。原因在于，在命令与控制体系中，上级比下属拥有更多权力。而他们权力更大的原因在于更接近信息源。在我们公司，信息会快速传播给许多不同层级的员工，通常是在团队层面共享。

The downside is the pressure on the leaders is fairly high. And the reason for that is because in a command and control system, the person who you reports to has more power than you. And the reason why they have more power than you is because they're closer to the source of information than you are. In our company, the information is disseminated fairly quickly to a lot of different people. And it's usually at a team level.

比如刚才的机器人技术会议上，我们讨论某些事项并做出决策时，会议室里有应届毕业生、三位副总裁和两名高管团队成员。当我们共同推理并做出决定的那一刻——

For example, just now I was in our robotics meeting and we're talking about certain things and we're making some decisions. And there are new college grads in a room. There's three vice presidents in a room. There's two E staffs in a room. And at the moment that we decided together, we reasoned through some stuff, we made a decision.

所有人都是同步获知信息的。因此没有人比其他人拥有更多权力。明白吗？应届毕业生和高管团队是同时知晓决策结果的。所以高管团队、我的下属管理层以及我本人，你们获得职位的资格是基于解决问题的能力以及帮助他人成功的能力。

Everybody heard it exactly the same time. So nobody has more power than anybody else. Does that make sense? The new college grad learned at exactly the same time as the E staff. And so the executive staff and the leaders that work for me and myself, you earn the right to have your job based on your ability to reason through problems and helping other people succeed.

而不是因为掌握了某些特权信息——比如只有我知道答案是3.7。明白吗？所有人都知道。

And it's not because you have some privileged information that I knew the answer was 3.7 and only I knew. Know? Right. Everybody knew.

我们最近制作NVIDIA第三期节目时——就是刚发布的那期——特别在过去两年里做了个思想实验。嗯。考虑到你们所涉及的技术水平和整体难度，你们的产品发布周期令人印象深刻。我们当时就在想，你能想象苹果每年发布两款iPhone吗？

When we did our most recent episode of NVIDIA part three that we we just released, we sort of did this thought exercise, especially over the last couple years. Mhmm. Your product shipping cycle has been very impressive, especially given the level of technology that you are working with and the difficulty of this all. We sort of said, like, could you imagine Apple shipping two iPhones a year?

我们这么说只是为了举例说明。

And we said that for illustrative purposes.

只是为了举例，并非针对苹果或其他公司。

For illustrative purposes. Not to pick on Apple or whatever.

一家科技公司每年发布两次旗舰产品或同款旗舰产品两次。

A tech company shipping two flagship products or their flagship product twice per year.

对。或者一年举办两次全球开发者大会之类的。

Yeah. Or, you know, two WWDCs a year.

是啊。你似乎有些什么

Yeah. There seems to be something You

不能 好吧，你确实很难想象这种事会在这里发生。有没有其他公司，无论是现在的还是历史上的，是你所仰望、钦佩，或许从中汲取了一些灵感的？

can't Well, you can't really imagine that whereas that happens here. Are there other companies, either current or historically, that you look up to, admire, maybe took some of this inspiration from?

过去三十年里，我读过不少商业书籍。就像阅读任何东西一样，首先你要享受它，对吧？享受它，从中获得启发，但不是照搬。这些书的目的不在于此。它们的意义在于分享经验。

In the last thirty years, I've read my fair share of business books. And as in everything you read, you're to, first of all, enjoy it, right? Enjoy it, be inspired by it, but not to adopt it. That's not the whole point of these books. The whole point of these books is to share their experiences.

你应该问自己：在我的世界里这意味着什么？在我所处的环境中意味着什么？对我正在经历的事情有何启示？对我想要实现的目标有何关联？对英伟达在我们公司当前发展阶段和能力范围内又意味着什么？

And you're supposed to ask, what does it mean to me in my world? And what does it mean to me in the context of what I'm going through? What does this mean to me in the environment that I'm in? And what does this mean to me and what I'm trying to achieve? And what does this mean to NVIDIA in the age of our company and the capability of our company?

所以你需要自问：这对你意味着什么？在吸收了这些多元化的知识后，我们应该制定自己的策略。这就是我的处事方式——从所有人身上汲取灵感，学习经验。教育本就是免费的资源。

And so you're supposed to ask yourself, what does it mean to you? And then from that point, being informed by all these different things that we're learning, we're supposed to come up with our own strategies. What I just described is kind of how I go about everything. You're supposed to be inspired and learn from everybody else. And the education is free.

当有人谈论新产品时，你应该去倾听而非忽视。可能是竞争对手，也可能是相邻行业。

When somebody talks about a new product, you're supposed to go listen to it. You're not supposed to ignore it. You're supposed to go learn from it. And it could be a competitor. It could be adjacent industry.

甚至可能与我们毫无关联。我们对外界动态了解得越多越好。但最终还是要回归本质：这对我们意味着什么？

It could be nothing to do with us. The more we're we learn from what's happening out in the world, the better. But then you're supposed to come back and ask yourself, you know, what does this mean to us?

没错。但你不能只是模仿。

Yeah. You don't just wanna imitate, though.

正是如此。

That's right.

对，对。我特别欣赏这种'学习但不模仿，广采众长'的理念。说到英伟达今天的成就，数据中心堪称不可思议的第三要素——这绝非显而易见的发展路径。

Yeah. Yeah. I love this tee up of learning but not imitating and learning from a wide array of sources. There's this sort of unbelievable third element, I think, to what NVIDIA has become today, and that's the data center. It's certainly not obvious.

我无法从AlexNet以及你与研究社区和社交媒体反馈的互动中推断出什么。归根结底，是供应商和公司在做决定。

I can't reason from AlexNet and your engagement with the research community and social media feedback All in, vendors deciding and the company deciding.

全力以赴，

All in,

我们将在数据中心开启一段为期五年的全力以赴之旅。

all We're gonna go on a five year all in journey on the data center.

是啊，是啊。

Yeah, yeah.

这是怎么发生的？

How did that happen?

是的。我们进军数据中心的旅程，我想说大约十七年前就开始了。总有人问我，公司未来可能面临哪些挑战？我一直认为，NVIDIA的技术需要插入电脑，而电脑必须放在你旁边，因为它要连接显示器——这终将限制我们的发展机会。毕竟能装GPU的台式机数量有限，我们能驱动的CRT和LCD显示器也有限。

Yeah. Our journey to the data center happened, I would say almost seventeen years ago. I'm always being asked, what are the challenges that the company could see someday? And I've always felt that the fact that NVIDIA's technology is plugged into a computer and that computer has to sit next to you because it has to be connected to a monitor, that will limit our opportunity someday because there are only so many desktop PCs that plug a GPU into. And there's only so many CRTs and the time LCDs that we could possibly drive.

所以问题是：如果我们的计算机不必连接显示设备，岂不是妙极了？这种分离让我们能在别处进行计算。有天我们的工程师演示给我看——它真正实现了抓取帧缓冲、编码成视频并流式传输到接收设备，将计算与显示分离。

So the question is, wouldn't it be amazing if our computer doesn't have to be connected to the viewing device? That the separation of it made it possible for us to compute somewhere else. And one of our engineers came and showed it to me one day, and it was really capturing the frame buffer and coding it into video and streaming it to a receiver device, separating computing from the viewing.

从很多方面说，这就是云游戏。云游戏。比如，

Many ways, that's cloud gaming. Cloud gaming. Like,

十八年前。没错。实际上

eighteen years ago. Yeah. In fact

实际上，那就是我们启动GFN的时候。我们清楚GFN将是场持久战，因为你要解决包括光速在内的各种难题。

in fact, was when we started GFN. We knew that GFN was going to be a journey that would take a long time because you're fighting all kinds of problems, including the speed of light.

放眼望去，延迟无处不在。

Latency everywhere you look.

没错。

That's right.

听众们，GFN，GeForce NOW。

For listeners, GFN, GeForce NOW.

GeForce NOW。GeForce NOW。我们一直在努力

GeForce NOW. GeForce NOW. And we've been working

开发GeForce，那是你们首个云端产品。

on GeForce That's your first cloud product.

没错。看，GeForce NOW是英伟达首款数据中心产品。我们的第二款数据中心产品是远程图形技术，将我们的GPU部署在全球企业数据中心，这进而引领我们开发出第三款产品，它结合了CUDA和我们的GPU，演变成超级计算机，随后不断迭代升级。其重要性在于——若能分离英伟达计算发生地与成果享用地的界限，市场机遇将呈爆发式增长。

That's right. And look GeForce NOW was NVIDIA's first data center product. And our second data center product was remote graphics, putting our GPUs in the world's enterprise data centers, which then led us to our third product, which combined CUDA plus our GPU, which became a supercomputer, which then worked towards more and more and more. And the reason why it's so important is because the disconnection between where NVIDIA's computing is done versus where it's enjoyed, if you can separate that, your market opportunity explodes.

是啊。确实。

Yeah. Yeah.

这完全正确。因此我们不再受限于桌边台式机的物理限制，也不受限于每人一块GPU的约束。位置变得无关紧要——这才是真正卓越的洞见。

And it was completely true. And so we're no longer limited by the physical constraints of the desktop PC sitting by your desk. And we're not limited by one GPU per person. And so it doesn't matter where it is anymore. And so that was really the great observation.

这是个很好的提醒。对我而言，英伟达的数据中心业务已与AI发展画上等号。嗯。但这种对等并不成立。有趣的是，你们之所以能在AI和数据中心领域蓄势待发，正是因为此前通过三个以上产品积累了构建数据中心计算机的经验。

It's a good reminder. The data center segment of NVIDIA's business to me has become synonymous with how is AI going. Mhmm. And that's a false equivalence. And it's interesting that you were only this ready to sort of explode in AI and the data center because you had three plus previous products where you learned how to build data center computers.

正是如此。尽管那些市场不像AI这样引发翻天覆地的技术变革。但这就是你们的学习路径。

Exactly. Even though those markets weren't these, like, gigantic world changing technology shifts the way that AI is. Yeah. That's how you learned.

没错。你想为未来的机遇铺路。不能等到机会摆在眼前才伸手去抓。所以必须提前预见。

Yeah. That's right. You wanna pave the way to future opportunities. You can't wait until the opportunity is sitting in front of you for you to reach out for it. And so you have to anticipate.

作为CEO，我们的工作是预见转角处的机会，预判机遇将来会出现在哪里。即使我不完全确定具体是什么或何时出现，也要让公司处于临近位置。就像站在果树附近，等苹果掉下来时就能飞身接住。你们懂我的意思吧？

Our job as CEO is to look around corners and anticipate where will opportunities be someday. And even if I'm not exactly sure what and when, how do I position the company to be near it? To be just standing kind of near under the tree And we can do a diving catch when the apple falls. You guys know what I'm saying?

是啊。但你必须...

Yeah. But you've got to

要足够近才能飞身接住。

be close enough to do the diving catch.

没错。回溯到2015年的OpenAI。如果当时没有在数据中心打下基础，现在就不可能为OpenAI提供支持。

Yeah. Rewind to 2015 and OpenAI. If you hadn't been laying this groundwork in the data center, you wouldn't be powering OpenAI right now.

但认为计算将主要在显示设备之外完成，绝大多数计算将远离计算机本身——这个洞见很准确。实际上，云计算和当今所有计算都关乎这种分离。通过将计算放在数据中心，我们能克服延迟问题，毕竟你无法超越光速。端到端的光速延迟只有120毫秒左右，并不算长。

But the idea that computing will be mostly done away from the viewing device, that the vast majority computing will be done away from the computer itself, That insight was good. In fact, cloud computing, everything about today's computing is about separation of that. And by putting it in a data center, we can overcome this latency problem, meaning you're not gonna overcome speed of light. Speed of light end to end is only a hundred and twenty milliseconds or something like that. It's not that long.

从数据中心到地球任何角落。

From a data center to an Anywhere on the planet.

对。

Yeah.

是的。所以我们能看到。真正实现全球覆盖。

Yeah. And so we can see. And to end literally across the planet.

没错。如果能解决这个问题——具体数字我忘了，大概是70到100毫秒——其实并不长。关键是，如果消除其他所有障碍，光速延迟完全不是问题。这样就能建造超大规模数据中心，实现惊人突破。

Yeah. Right. So if you could solve that problem approximately something like that, I forget the number, but it's seventy milliseconds, a hundred milliseconds, but it's not that long. And so my point is, if you could remove the obstacles everywhere else, then speed of light should be perfectly fine. And you could build data centers as large like, and you could do amazing things.

而我们用作电脑或电视的这个小巧设备，无论何种电脑，它们都能瞬间变得无比强大。这一洞见在十五年前确实颇具前瞻性。

And this little tiny device that we use as a computer or your TV as a computer, whatever computer, they all they can all instantly become amazing. And so that insight, you know, fifteen years ago was a good one.

说到光速，InfiniBand。是啊，就像戴维一直在极力推荐这个方向。嗯哼。

So speaking of the speed of light, InfiniBand. Yeah. Like, Dave David's, like, begging me to go here. Uh-huh.

我也经历过同样的时期。

I've had the same time.

你完全预见到InfiniBand会比其他任何人意识到的更早发挥巨大作用。收购Mellanox时，我认为只有你独具慧眼，看出这是训练大语言模型的必备条件，并且你以超强魄力完成了收购。为何你能看到别人看不到的？

You totally saw that InfiniBand would be way more useful, way sooner than anyone else realized. Yeah. Acquiring Mellanox, I think you uniquely saw that this was required to train large language models, and you were super aggressive in acquiring that company. Why did you see that when no one else saw that?

这有几个原因。首先，若想成为数据中心企业，仅靠制造处理芯片是行不通的。数据中心与台式机或手机的区别不在于处理器——数据中心里的台式机使用的CPU和GPU几乎相同，对吧？

Well, there were several reasons for that. First, if you wanna be a data center company, build building the processing chip isn't the way to do it. A data center is distinguished from a desktop computer versus a cell phone, not by the processor in it. A desktop computer in a data center uses the same CPUs, uses the same GPUs apparently, right? Very close.

因此关键不在芯片，不在于处理芯片本身，而在于其网络架构。是基础设施的布局，计算力的分布式部署，安全机制的构建，网络系统的设计等等。这些特性与Mellanox而非英伟达相关联。

And so it's not the chip. It's not the processing chip that describes it, but it's the networking of it. It's the infrastructure of it. It's the how the computing is distributed, how security is provided, how networking is so on and so forth. And so those characteristics are associated with Mellanox, not Nvidia.

当我最终认定英伟达要构建未来计算机——这些未来计算机将以数据中心形态存在时，当我们决心转型为数据中心导向型企业时，就必须进军网络领域。这是其一。其二，我观察到云计算始于超大规模架构，即利用标准化组件服务海量用户，在一台计算机上实现多用户虚拟化。而AI则关乎分布式计算——单个训练任务需要协调数百万处理器协同工作。

And so the day that I concluded that really Nvidia wants to build computers of the future and computers of the future are gonna be data centers embodied in data centers. And we wanna be data center oriented company, we really need to get into networking. And so that was one. The second thing is observation that whereas cloud computing started in the hyperscale, which is about taking commodity components, a lot of users and virtualizing many users on top of one computer. AI is really about distributed computing where one job, one training job is orchestrated across millions of processors.

这几乎与超大规模架构背道而驰。用现成的标准化以太网设计超大规模计算机，对Hadoop、搜索查询等应用完全适用。

And so it's the inverse of hyperscale almost. And the way that you design a hyperscale computer with off the shelf commodity ethernet, which is just fine for Hadoop, it's just fine for search queries, it's just fine for all of those things.

但当你需要跨节点分片模型时就...

But not when you're sharding a model across

当需要跨节点分片模型时就不适用了。对吧？这一观察表明，理想的网络类型并非传统以太网。超级计算领域的网络解决方案才是最佳选择。正是这两个理念的结合让我确信Mellanox是最佳收购目标——他们是全球领先的高性能网络公司。

Not multiple when you're sharding a model across. Right? And so that observation says that the type of networking you want to do is not exactly ethernet. And the way that we do networking for supercomputing is really quite ideal. And so the combination of those two ideas convinced me that Mellanox is absolutely the right company because they're the world's leading high performance networking company.

我们已经在高性能计算的多个领域与他们合作过。而且我真的很喜欢那里的人。以色列团队是世界一流的。我们现在那里有大约3200名员工。这是我做过的最好的战略决策之一。

We worked with them in so many different areas in high performance computing already. Plus I really liked the people. Israel team is world class. We have some 3,200 people there now. And it was one of the best strategic decisions I've ever made.

当我们进行调研时，特别是NVIDIA系列第三部分，我们与许多人交谈过，很多人都告诉我们，Mellanox的收购是有史以来科技公司最好（如果不是最好的话）的收购之一。我

When we were researching, particularly part three of our NVIDIA series, we talked to a lot of people and many people told us the Mellanox acquisition is one of, if not the best of all time by any technology company. I

也这么认为。是的。而且这与我们平时的工作完全不同。这让所有人都感到惊讶。

think so too. Yeah. And it's so disconnected from the work that we normally do. It was surprising to everybody.

但换个角度想，你当时就站在行动发生的地方附近。是的。所以你可以很快意识到，一旦苹果（指机会）可以购买，比如，哦，LLM（大语言模型）即将爆发。我会需要那个。每个人都会需要那个。

But frame this way, you were you were standing near where the action was. Yeah. So you could figure out as soon as that Apple sort of becomes available to purchase, like, oh, LLMs are about to blow up. I'm gonna need that. Everyone's gonna need that.

我想我在其他人之前就知道了这一点。

I think I know that before anyone else does.

是的。你希望自己靠近机会。你不必做到完美。你希望自己站在树下。即使你没有在苹果落地前接住它，只要你是第一个捡起它的人就行。

Yeah. You wanna position yourself near opportunities. You don't have to be that perfect. You wanna position yourself near the tree. And even if you don't catch the apple before it hits the ground, long as you're the first one to pick it up.

你希望自己靠近机会。所以我很多工作就是让公司靠近机会，并让公司具备在每一步变现的能力，这样我们才能持续发展。

You wanna position yourself close to the opportunities. And so that's kind of a lot of my work is positioning the company near opportunities and having company having the skills to to monetize each one of the steps along the way so that we can be sustainable.

你刚才说的让我想起了巴菲特和芒格的一句名言：大致正确比精确错误要好。

What you just said reminds me of a great aphorism from Buffett and Munger, is it's better to be approximately right than exactly wrong.

是的。就是这样。是的。这句话很好。

Yeah. There you go. Yeah. That's a good one.

这是值得遵循的好建议。

Good one to live by.

没错。好了，听众朋友们。现在正是感谢我们Acquired新合作伙伴Sentry的好时机。Sentry拼写为s-e-n-t-r-y，就像站岗的哨兵。

Yeah. Alright, listeners. This is a great time to thank a new partner of ours here at Acquired, Sentry. That's s e n t r y, like someone standing guard.

是的。Sentry帮助开发者调试错误和延迟问题，几乎涵盖所有软件问题，并在用户发怒前修复它们。正如其官网所言，它被超过400万软件开发者认为'还不错'。

Yes. Sentry helps developers debug errors and latency issues, pretty much any software problem, and fix them before users get mad. As their homepage puts it, it's considered, quote unquote, not bad by over 4,000,000 software developers.

今天我们要讨论Sentry如何与Acquired生态中的另一家公司Anthropic合作。Anthropic原本使用较旧的基础设施监控系统，但鉴于其庞大的规模和复杂性，他们转而采用Sentry来更快发现和解决问题。

So today, we're talking about the way that Sentry works with another company in the acquired universe, Anthropic. Anthropic used to have some older infrastructure monitoring that was in place, but at their massive scale and complexity, they instead adopted Sentry to help them find and fix issues faster.

没错。崩溃在AI领域可能造成巨大问题。如果你正在运行像模型训练这样的大型计算任务，一个节点故障就可能影响数百或数千台服务器。Sentry帮助他们检测故障硬件，从而在引发连锁问题前快速剔除。Sentry让他们能在几小时内而非数天调试重大问题，尽快恢复训练任务。

Yep. Crashes can be a massive problem in AI. If you're running a huge compute job like training a model and one node fails, it can affect hundreds or thousands of servers. Sentry helped them detect bad hardware so they could quickly reject it before causing a cascading problem. Sentry enabled them to debug massive issues in hours instead of days so they could get back to their training runs.

如今Anthropic依赖Sentry实时追踪异常、分配错误并分析故障，覆盖其研究团队使用的所有主要语言，包括Python、Rust和C++。据Anthropic团队表示，Sentry为开发者提供了调试问题所需全部信息的统一平台。

And today, Anthropic relies on Sentry to track exceptions, assign errors, and analyze failures in real time across all the primary languages used by Anthropics research teams, including Python, Rust, and c plus plus According to the Anthropic team, Sentry gives our developers one place where they have all the information they need to debug an issue.

Sentry还有个有趣的新动态：本月推出了名为SEER的AI调试器。这个AI代理能利用Sentry的问题上下文和代码库，不仅猜测问题根源，还能针对具体应用提供可直接合并的修复方案。

And one other fun update in the world of Sentry is that as of this month, Sentry now has an AI debugger called SEER. SEER is an AI agent that taps into all the issue context from Sentry and your code base to not just guess, but root cause gnarly issues and propose merge ready fixes specific to your application.

我们非常兴奋能与Sentry合作。他们的客户名单令人印象深刻，不仅有Anthropic，还包括Cursor、Vercel、Linear等。如果你想像13万家组织那样快速修复故障代码——从独立开发者到世界顶级公司——可以访问sentry.io/acquired了解更多，他们为所有Acquired听众提供两个月免费试用。记得告诉他们是Ben和David推荐来的。

We are pumped to be working with Sentry. They've got an incredible customer list, including not only Anthropic, but Cursor, Vercel, Linear, and more. If you wanna fix broken code like the over 130,000 organizations using Sentry from indie hobbyists to some of the biggest companies in the world to find and fix broken code fast. You can check out sentry.i0/acquired to learn more, and they are offering two free months to all Acquired listeners. That's Sentry, s e n try,.i0/acquired, and just tell them that Ben and David sent you.

我想暂时抛开NVIDIA话题，既然我们有很多创业者听众，请教些关于公司建设的建议。第一个问题是：初创公司早期最大的竞争对手其实是'做不出人们想要的东西'，这样的公司很可能会倒闭。

I wanna move away from NVIDIA, if you're okay with it, and ask you some questions since we have a lot of founders that listen to this show, sort of advice for company building. The first one is when you're starting a startup in the earliest days, your biggest competitor is, you don't make anything people want. Like, company is likely to die

不仅仅是消费需求问题。

Not just consumption.

因为人们实际上并不像你那样在乎

Just because people don't actually care as much as you do about what

没错。

That's right.

是的。在后期阶段，企业必须深思熟虑竞争策略。我很好奇，对于那些已经实现产品市场匹配、正在成长中且处于有趣增长市场的公司，您会给出什么建议？他们应该在哪里寻找竞争，又该如何应对？

Yeah. In the later days, you actually have to be very thoughtful about competitive strategy. And I'm curious, what would be your advice to companies that, you know, have product market fit, that are starting to grow, they're in interesting growing markets? Where should they look for competition, and how should they handle it?

关于竞争有各种思考方式。我们更倾向于定位在服务尚未显现的需求领域。

Well, there are all kinds of ways to think about competition. We prefer to position ourselves in a way that serves a need that usually hasn't emerged.

我记得您或英伟达其他人用过‘零美元市场’这个说法

I've heard you or others in NVIDIA, I think, use the phrase $0

市场。

market.

对，完全正确。这是我们表达‘市场尚未存在，但我们相信终将形成’的方式。通常当你这样定位时，所有人都会困惑你为何在此布局。

Yeah. That's exactly right. Yeah. It's our way of saying there's no market yet, but we believe there will be one. Usually when you're positioned there, everybody's trying to figure out why are you here.

明白吗？比如我们最初进入汽车领域时，是因为相信未来汽车将主要由软件定义。若真如此，就需要极其强大的计算能力。当我们这样定位时，有位CTO曾对我说：‘汽车可受不了蓝屏死机’。

Right? Because when we first got into automotive, because we believe that in the future, the car is going to be largely software. And if it's gonna be largely software, a really incredible computer is necessary. And so so when we positioned ourselves there, most people I I still remember one one of the one of the CTOs told me, you know what? Cars cannot tolerate the blue screen of death.

我回应道：‘没人能接受蓝屏，但这改变不了未来每辆车都将成为软件定义汽车的事实’。十五年后证明我们基本正确。我们常选择‘非消费领域’布局，当市场真正形成时，往往鲜有同类竞争者。我们在PC游戏领域早期入场，如今英伟达已是该领域巨头。

And I said, I don't think anybody can tolerate that, but that doesn't change the fact that someday every car will be a software defined car. I think, you know, fifteen years later, we're largely right. And so oftentimes there's non consumption, and we like to navigate our company there. And by doing that, by the time that the market emerges, it's very it's very likely there aren't that many competitors shaped that way. And so we were early in PC gaming, and today, NVIDIA is very large in PC gaming.

我们重新构想了设计工作站的可能形态，如今全球几乎所有工作站都采用英伟达技术。我们重塑了超级计算的实现方式及其受益群体——使其民主化。如今英伟达在加速计算领域举足轻重。我们还重新设想了软件开发范式。

We reimagined what a what a design workstation would be like. And today, just about every workstation on the planet uses NVIDIA's technology. We reimagine how supercomputing ought to be done and who should benefit from supercomputing, that we would democratize it. And look today, NVIDIA's and accelerated computing quite large. We reimagine how software would be done.

如今这被称为机器学习；对计算方式的革新，我们称之为AI。我们以超前十年的眼光构想这些变革，因此在‘零美元市场’耕耘约十年之久。现在，我正全力投入Omniverse的研发。

And today it's called machine learning and how computing would be done, we call it AI. And so we reimagine these kinds of things, try to do that about a decade in advance. And so we spent about a decade in $0 markets. Mhmm. And today, I spent a lot of time on Omniverse.

而Omniverse就是一个典型的零美元业务的例子。而且现在大概有

And Omniverse is a, you know, classic example of a $0 business. And There's like

40个客户了。

40 customers now.

亚马逊、宝马。是的。不。这很酷。

Amazon, BMW. Yeah. No. It's cool.

这很酷。

It's cool.

假设你确实获得了这十年的领先优势，但后来其他人也搞明白了，有人紧追不舍。创业者可以做哪些结构性的事情来保持领先？你可以一直保持高强度工作，说我们会比他们更努力、更聪明。这在某种程度上是有效的，但这些只是战术。从战略上，你能做些什么来确保维持这种领先优势？

So let's say you do get this great ten year lead, but then other people figure it out, you got people nipping at your heels. What are some structural things that someone who's building a business can do to sort of stay ahead? And you can just keep your pedal to the metal and say, we're gonna outwork them, and we're gonna be smarter. And, like, that works to some extent, but those are tactics. What strategically can you do to sort of make sure that you can maintain that lead?

通常，如果你开创了市场，最终会拥有人们所说的护城河。因为如果你的产品做得好，并且围绕你形成了一个完整的生态系统来服务终端市场，你实际上就创建了一个平台。有时是基于产品的平台，有时是基于服务的平台，有时是基于技术的平台。

Oftentimes, if you created the market, you ended up having what people describe as moats. Because if you build your product right and it's enabled an entire ecosystem around you to help serve that end market, you've essentially created a platform. Sometimes it's a product based platform. Sometimes it's a service based platform. Sometimes it's a technology based platform.

但如果你进入得早，并且有意识地帮助生态系统与你一起成功，最终你会拥有这个由网络组成的网络，所有这些开发者和客户都围绕着你建立起来。是的。而这个网络本质上就是你的护城河。所以，我不喜欢从护城河的角度来思考这个问题。原因是这样会让你专注于在城堡周围建造东西。

But if you were early there and you you were mindful about helping the ecosystem succeed with you, you ended up having this network of networks and all these developers and all these customers who are who are built around you. Yep. And that network is essentially your moat. And so, you know, I I don't love thinking about it in the context of a moat. And the reason for that is because you're now focused on building stuff around your castle.

我更喜欢从构建网络的角度来思考问题，这个网络是为了让其他人也能享受终端市场的成功。你不是唯一一家从中受益的公司，而是与包括我在内的许多人一起享受这种成功。

I tend to like thinking about things in the context of building a network, and that network is about enabling other people to enjoy the success of the final market. You know, that you're not the only company that enjoys it, but you're enjoying it with a whole bunch of other people, including me.

我很高兴你提到这一点，因为我想问你。至少在我看来，听起来你也这么认为

I'm so glad you brought this up because I wanted to ask you. In my mind at least and sounds like in yours

嗯哼。

too Mhmm.

英伟达绝对是一家平台公司，嗯。世界上真正有影响力的平台公司屈指可数。我认为公平地说，在最初几年起步时，你们还是一家技术公司而非平台公司。我能想到的所有试图一开始就做平台的公司都失败了。必须先从技术做起。

NVIDIA is absolutely a platform company Mhmm. Of which there are very few meaningful platform companies in the world. I think it's also fair to say that when you started for the first few years, you were a technology company and not a platform company. Every example I can think of of a company that tried to start as a platform company fails. You gotta start as a technology first.

你们是什么时候考虑转型为平台的？比如最早的显卡只是技术产品，当时还没有CUDA，也没有平台概念。

When did you think about making that transition to being a platform? Like your first graphics cards were technology. There was no CUDA. There was no platform.

是的，你的观察没错。但在公司内部，我们始终自视为平台公司。因为从创立第一天起，我们就拥有名为UDA的架构——也就是CUDA里的UDA。

Yeah. What you observed is not wrong. However, inside our company, we were always a platform company. And the reason for that is because from the very first day of our company, we had this architecture called UDA. It's the UDA of CUDA.

CUDA是计算统一...嗯...设备架构？

CUDA is Compute Unified Yeah. Device architecture?

没错。关键在于，尽管Riva 128最初只有计算机图形功能，但我们设计的架构适用于各类加速器。开发者可以基于这个架构编程。实际上，英伟达最初的商业战略是成为PC内的游戏主机，而游戏主机需要开发者支持——这就是为什么我们很早就聘请了开发者关系专员。

That's right. And the reason for that is because what we've done, what we essentially did in the beginning, even though Riva one twenty eight only had computer graphics, the architecture described accelerators of all kinds. And we would take that architecture and developers would program to it. In fact, NVIDIA's first strategy, business strategy, was we were going to be a game console inside the PC. And a game console needs developers, which is the reason why NVIDIA, a long time ago, one of our first employees was a developer relations person.

正因如此，我们认识所有游戏开发者和3D开发者，我们了解...

And so it's the reason why we knew all the game developers and all the three d developers, and we knew

认识所有人。所以最初的商业计划是类似于...

we knew everyone. So was the original business plan to like

有点像要打造DirectX那样。

Sort of like to build DirectX.

对，在PC领域与任天堂、世嘉竞争。

Yeah. Compete with Nintendo and Sega as, like, with PCs.

英伟达最初的架构叫DirectNV，就是Direct NVIDIA。而DirectX是个让操作系统能直接...的API

Original NVIDIA architecture was called DirectNV. Direct NVIDIA. Yeah. And DirectX was an API that made it possible for operating system to directly

向网络致辞。

Address the network.

对。硬件。没错。但DirectX在

Yeah. Hardware. Yeah. But DirectX didn't exist when

你们创立英伟达时还不存在。对吧？而这正是导致你们

you started NVIDIA. Right? And that's what And made your strategy wrong for the

第一家公司战略失误的原因。'93年是的。那时我们有了DirectX

first company. '93 Yeah. We had DirectX

是的。

Yeah.

然后在1995年，你知道的，DirectX正式发布了。

And which in 1995 became, you know well, DirectX came out.

所以这是个重要的教训。

So this is an important lesson.

我们始终是一家以开发者为导向的公司。

You We were always always a developer oriented company.

最初的尝试是让开发者基于DirectNV开发，他们就会为我们的芯片开发，然后我们就能拥有一个平台。对。正是如此。实际情况是微软早已建立了所有这些开发者关系。所以确实如此。

The initial attempt was we will get the developers to build on DirectNV, and then they'll build for our chips, and then we'll have a platform. And Yeah. Exactly. What played out is Microsoft already had all these developer relationships. So Exactly.

你们以艰难的方式学到了这个教训，就像

You learned the lesson the hard way of like Yeah.

哎呀，我们只能那样做了。那时候他们就说，哦，这可以做成一个开发者平台。我们就收下了，谢谢啊。

Yikes. We just gotta That's what did back in the day. They're like, oh, that could be a developer platform. We'll take that. Thank you.

你知道吗？不。但他们有很多优势。他们的做法截然不同，而且他们很多事情都做对了。我们却犯了很多错误。

You know? No. But they had a lot. They did it very differently, and and they did a lot of things right. We did a lot of things wrong.

不过话虽这么说

But but having said

对抗九十年代的微软。我是说，那简直

against Microsoft in the nineties. I mean, that's

是啊，就像现在面对英伟达一样。没错。

Yeah. It's like having NVIDIA today. Yeah.

不，情况大不相同，但我理解你的意思。我们当时根本无力与他们竞争。看看现在，当CUDA出现时，虽然有OpenGL和DirectX，但可以说还有另一个扩展，那就是CUDA。这个CUDA扩展让原本为运行DirectX和OpenGL而付费的芯片，为CUDA建立了用户基础。

No. It's a lot different, but I appreciate that. But we were nowhere near competing with them. If you look now, when CUDA came along and there was OpenGL, there was DirectX, but there's still another extension, if you will, and that extension is CUDA. And that CUDA extension allows a chip that got paid for running DirectX and OpenGL to create an installed base for CUDA.

所以这就是

And so that's the

结局

end

的这就是

of And is this the

你当时如此强硬的策略。根据我们的研究，正是你坚持每块英伟达芯片都必须运行CUDA的强硬态度起了关键作用。

strategy you were so militant. And I think from our research, it really was you being militant that every NVIDIA chip will run CUDA.

是的。如果你是一个计算平台，所有东西都必须兼容。我们是地球上唯一一个所有加速器在架构上都相互兼容的加速器厂商。从未有过这样的先例。全球实际部署的CUDA GPU数量高达数亿。

Yeah. If you're a computing platform, everything's gotta be compatible. We are the only accelerator on the planet where every single accelerator is architecturally compatible with the others. None that has ever existed. There are literally a couple of 100,000,000.

对吧？如今全球有2.5亿到3亿活跃的CUDA GPU在使用中，它们全都架构兼容。如果MV30、MV35、MV39和MV40都各不相同，那还怎么构建计算平台？但三十年来，它们始终保持完全兼容。

Right? 250,000,000, 300,000,000 installed base of active CUDA GPUs being used in the world today, and they're all architecturally compatible. How would you have a computing platform if MV 30 and MV 35 and MV 39 and MV 40, they're all different. Right? At thirty years, it's all completely compatible.

所以这是我们公司唯一不可妥协的原则。其他一切都可以商量。

And so that's the only unnegotiable rule in our company. Everything else is negotiable.

我想，Kudo可以说是UDA的重生，但现在看来，UDA的历史可以一直追溯到你们所有芯片的源头。

I mean, and I guess Kudo was a rebirth of UDA, but understanding this now, UDA going all the way back, it really is all the way back to all the chips you've ever heard.

没错。实际上，UDA的基因延续到了我们现在的每一款芯片。哇哦。

Yeah. Yeah. Yeah. In fact, UDA goes all the way back to all of our chips today. Wow.

郑重声明，我没有给在听的任何创始CEO提供帮助。说实话，当你问这个问题时——我会传授什么经验？我不知道。成功企业和成功CEO的特质已经被描述得很充分了，这样的案例有很多。

For the record, I didn't help any of the founding CEOs that are listening. I gotta tell you, you know, while you were asking that question, what lessons would I impart? I don't know. I mean, the characteristics of successful companies and successful CEOs, I think are fairly well described. There are a whole bunch of them.

我只是觉得创立成功企业难如登天。当我看到这些杰出企业被建立起来时，心中只有敬佩，因为我深知这有多艰难。虽然大家采取的措施大同小异，但确实有些明智的做法。

I just think starting successful companies are insanely hard. It's just insanely hard. And when I see these amazing companies getting built, I have nothing but admiration and respect because I just know that it's insanely hard. And I think that everybody did many similar things. There are some good, smart things that people do.

也有些愚蠢的决策。你可能做对所有事仍然失败，也可能犯很多错误——我就犯过不少——却依然成功。所以显然这不是绝对的。我认为技能是可以在过程中学习的。

There are some dumb things that you can do. You could do all the right, things and still fail. You could do a whole bunch of dumb things, and I did many of them, and still succeed. So obviously, that's not exactly right. You know, just I think skills are are the things that you can learn along the way.

但在关键时刻，必须要有特定条件同时满足。我确实认为市场必须...成为助力成功的因素之一。当然这还不够，毕竟很多人还是失败了。

But at an important moment, certain circumstances have to come together. And and I do think that that the market has to, you know, be one of the agents Yeah. To help you succeed. It's not enough, obviously, because a lot of people still fail.

你还记得英伟达历史上某些时刻吗？就是你们做了一堆错误决定，却因为运气和能力的总和而侥幸过关。有没有这样的瞬间...

Do you remember any moments in NVIDIA's history where you're like, oh, we made a bunch of wrong decisions, but somehow we got saved because, it takes the sum of all the luck and all the skill Yeah. In order to succeed. Do you remember any moments where you

其实我认为你们从Rebo 128起步是完全正确的。正如我提到的，Rebo 128时期我们做了许多至今仍堪称明智的决策——我们设计芯片的方式至今未变，因为天啊，当时根本没人这么做过。我们当时穷尽所有手段是出于绝望，因为我们别无选择。但你知道吗？那恰恰是做事应有的方式。

actually thought that you started with Rebo one hundred twenty eight was spot on. Rebo one hundred twenty eight, as I mentioned, number of smart decisions we made, which are smart to this day, how we design chips is exactly the same to this day, because gosh, nobody's ever done it back then. And we pulled every trick in the book in a desperation because we had no other choice. Well, guess what? That's the way things ought to be done.

现在所有人都这么做了对吧？大家这样做是因为——能一次完成的事何必做两次？能一次流片成功的芯片为什么要做七次？对吧？

And now everybody does it that way. Right? Everybody does it because why should you do things twice if you can do it once? Why tape out a chip seven times if you could tape it out one time? Right?

因此最有效率、最具成本效益、最具竞争力的就是速度技术。速度即性能，上市时间即性能，这些道理都适用。

And so the most efficient, the most cost effective, the most competitive speed is technology. Right? Speed is performance. Time to market is performance. All of those things apply.

既然能一次完成，何必重复劳动？Rebo 128在产品规格制定、市场需求判断、市场空白评估等方面都做出了极其出色的决策。老天，我们当时那些决定简直好得不可思议——毕竟我们当时被逼到了绝境。

So why do things twice if you could do it once? And so REBA one hundred twenty eight made a lot of great decisions in how we spec products, how we think about market needs and lack of, and how do we judge markets and all of this. Man, we made some amazing amazingly good decisions. Yeah. We were, you know, back against the wall.

我们只剩最后一次机会了。但是

We only had one more shot to do it. But

当你全力以赴并见识到自己真正的能力后，下次为什么还要保留余力呢？我们应该永远保持全力出击的状态

Once you pull out all the stops and you see what you're capable of, why would you put stops in next time? Like, let's keep stops out all the time

说得对。

That's right.

每次都该如此。

Every time.

没错。

That's right.

不过从运气因素来看，是否可以说1997年正是消费者开始真正重视游戏3D图形性能的转折点？

Is it fair to say though, maybe on the luck side of the equation, thinking back to 1997, that that was the moment where consumers tipped to really, really valuing three d graphical performance in games?

哦对了。比如说运气。我们来聊聊运气。如果卡马克决定使用加速技术——要知道《毁灭战士》完全是软件渲染的。而NVIDIA的理念是，尽管通用计算非常了不起，能推动软件和IT等领域发展，但我们认为有些应用如果不加速就无法实现或成本会很高。

Oh yeah. So for example, luck. Let's talk about luck. If Carmack decided to use acceleration because remember, Doom was completely software rendered. And the NVIDIA philosophy was that although general purpose computing is a fabulous thing and it's gonna enable software and IT and everything, we felt that there were there were applications that wouldn't be possible or would be costly if it wasn't accelerated.

它应该被加速。3D图形是其中之一，但不是唯一。恰好它是第一个也是极其重要的一个。我还记得最初几次见到约翰时，他非常坚持使用CPU，认为软件渲染器已经足够好。说实话，当时即使有加速器，《毁灭战士》的性能表现都很难达到那种水平。

It should be accelerated. And three d graphics was one of them, but it wasn't the only one. And it just happens to be the first one and a really great one. And I still remember the first times we met John, he was quite emphatic about using CPUs and the software render was really good. I mean, frankly, if you look at Doom, the performance of Doom was really hard to achieve even with accelerators at the time.

如果不做过滤，不需要双线性过滤的话，它的表现其实相当不错。

If you didn't filter, if you didn't have to do bilinear filtering, it did a pretty good job.

但《毁灭战士》的问题在于你需要卡马克来编程。

The problem with Doom though was you needed Carmack to program it.

没错。必须由卡马克来编程。那确实是天才级的代码。不过即便如此，软件渲染器的表现已经非常出色了。

Yeah. You needed Carmack to program it. Exactly. It was it was a genius piece of code. And but nonetheless, software renderers did a really good job.

但如果他没有决定转向OpenGL并为《雷神之锤》做加速，说实话，我们现在的成就会由哪个杀手级应用带来呢？正是卡马克和斯威尼，通过《雷神之锤》和《虚幻》，创造了前两个面向消费者的3D杀手级应用。我欠他们太多了。

And but and and if he hadn't decided to go to OpenGL and accelerate accelerate for Quake, frankly, you know, what would be the killer app that put us here? Right. And so Carmack and Sweeney, both between Unreal and Quake, created the first two killer applications for for consumer three d. Yeah. And so I I owe owe them a great deal.

我想快速回到刚才的话题。你说不知道创始人能从这些故事中学到什么。但我认为，看看当今所有大型科技公司——可能除了谷歌——它们起步时都像你这样，专注于服务开发者，计划为开发者构建平台和工具。无一例外。

I wanna come back real quick too. You know, you said you told these stories and you're like, well, I don't know what founders can take from that. I I actually do think, you know, if you look at all the big tech companies today, perhaps with the exception of Google, they did all start, and understanding this now about you, by addressing developers, planning to build a platform and tools for developers. You know, all of them.

苹果不是，亚马逊也不是。

Apple, not Amazon.

不过AWS确实是这么起家的。所以我觉得这其实印证了你的观点：这绝对不意味着能保证成功。

I guess with AWS, that's how AWS started. So I think that actually is a lesson to your point of like, that won't guarantee success by any means.

没错。

Right.

但那样的话，如果苹果掉下来，你就会在树旁晃悠。

But that'll get you hanging around the tree if the apple falls.

没错。尽管我们有很多好主意，但不可能囊括世界上所有的好点子。而拥有开发者的好处就是你能见识到大量优秀的创意。

Yeah. As many good ideas as we have, you don't have all the world's good ideas. And and the benefit of having developers is you get to see a lot of good ideas.

是的。随着我们逐渐接近尾声，我们花了大量时间回顾过去，而我想稍微展望下未来。我知道身处AI前沿的你们肯定经常思考这个问题。我们正进入一个时代，当人们使用软件时，软件带来的生产力能极大放大他们创造的影响力和价值，长远来看这对人类必定是极好的。

Yep. Yeah. Well, as we we start to drift toward the end here, we spend a lot of time on the past, and I wanna think about the future a little bit. I'm sure you spend a lot of time on this being on the cutting edge of AI. You know, we're moving into an era where the productivity that software can accomplish when a person is using software can massively amplify the impact and the value that they're creating, which has to be amazing for humanity in the long run.

短期来看，在我们摸索适应期间难免会有波折。随着AI越来越强大，越来越擅长提升生产力，你认为针对由此产生的岗位替代问题，有哪些解决方案？

In the short term, it's gonna be inevitably bumpy as we sort of figure out what that means. What do you think some of the solutions are as AI gets more and more powerful and better at accelerating productivity for all the displaced jobs that are gonna come from it?

首先必须确保AI安全。AI安全有几个关键领域：机器人技术和自动驾驶涉及完整的安全体系，我们致力于功能安全、主动安全等各类安全标准——何时需要人工介入？何时可以无人参与？如何逐步减少人类干预但保持宏观监督？在信息安全方面，偏见、虚假信息及尊重创作者权益等问题都亟待重视。

Well, first of all, we have to keep AI safe. And there's a couple of different areas of AI safety that's really important. Obviously, robotics and self driving car, there's a whole field of AI safety and we've dedicated ourselves to functional safety and active safety in all kinds of different areas of safety. When to apply human in the loop, when is it okay for human not to be in the loop, how do you get to a point increasingly human doesn't have to be in the loop, but human largely in the loop? In the case of information safety, obviously bias, false information and appreciating the rights of artists and creators, that whole area deserves a lot of attention.

你已看到我们的实践：我们未全网爬取数据，而是与盖蒂图片社和Shutterstock合作，以商业公平方式开发生成式AI。对于大语言模型和未来更具自主性的AI，只要条件允许（我认为会持续很久），解决方案就是保持人类监督。应避免AI在数字环境中自主进化，需通过数据收集、训练、测试和验证的完整流程才能部署更新。

You've seen some of the work that we've done. Instead of scraping the internet, partnered with Getty and Shutterstock to create commercially a fair way of applying artificial intelligence generative AI. In the area of large language models and the future of increasingly greater agency AI, clearly the answer is for as long as it's sensible, and I think it's gonna be sensible for a long time, is human in the loop. The ability for an AI to self learn and improve and change out in the wild in a digital form should be avoided. We should collect data, we should carry the data, we should train the model, we should test the model, validate the model before we release it on the wild again.

必须保持人类监督。许多行业已示范如何构建对人类有益的安全系统，比如飞机自动驾驶采用双驾驶员制，结合空管系统和多重冗余设计，这些安全哲学同样适用于自动驾驶等领域。我认为现有的大量AI安全模式都值得借鉴应用。

Human is in the loop. Yep. There are a lot of different industries that have already demonstrated how to build systems that are safe and good for humanity. Obviously the way autopilot works for a plane and two pilot system and then air traffic control and redundancy and diversity, and all of the basic philosophies of designing safe systems apply as well in self driving cars and so on and so forth. So I think there's a lot of models of creating safe AI, and I think we need to apply them.

关于自动化，我认为AI更可能创造而非消灭岗位——当然'近期'的定义有待商榷。生产力提升首先带来行业繁荣，企业会因拓展新领域而增聘人手。问题的关键在于：如果公司认为提高生产力就要裁员，那只能说明他们缺乏新创意。

With respect to automation, my feeling is that, and we'll see, but it is more likely that AI is gonna create more jobs. And in the near term, the question is what's the definition of near term? And the reason for that the first thing that happens with productivity is prosperity. Prosperity, the companies get more successful, they hire more people because they want to expand into more areas. And so the question is, if you think about a company and say, okay, if we improve the productivity, need fewer people.

实际上，当企业因生产力提升而盈利时，通常会招聘更多人才开拓新领域。只要人类在医药研发、交通运输、零售娱乐、科技创新等领域还有未实现的创意，行业繁荣就会催生更多就业机会。回望历史，当今产业规模远超千年前，正是因为人类拥有无穷创意。

Well, that's because the company has no more ideas, but that's not true If for most you become more productive and the company becomes more profitable, usually they hire more people to expand into new areas. And so long as we believe that there are more areas to expand into, that there are more ideas in drugs, this drug discovery, there are more ideas in transportation, there are more ideas in retail, there are more ideas in entertainment, that there's more ideas in technology. So long as we believe that there are more ideas, the prosperity of the industry, which comes from improved productivity, results in hiring more people, more ideas. Now, you go back in history, we can fairly say that today's industry is larger than the industry, the world's industries a thousand years ago. And the reason for that is because obviously humans have a lot of ideas.

我相信生产力提升将孕育更多促进繁荣的创意。虽然整体就业增长不保证个体不被淘汰——这显而易见——但我的判断是AI更可能创造净增岗位。

And I think that there's plenty of ideas yet for prosperity and plenty of ideas that can be begat from productivity improvements. But my sense is that it's likely to generate jobs. Now, obviously, net generation of jobs doesn't guarantee that any one human doesn't get fired. Okay. I mean, that's obviously true.

更有可能的是，某人会因其他人使用AI而失去工作，而非直接被AI取代。因此我认为，首要之事是每个人都应学习使用AI以提升自身生产力。每家公司也应借助AI提高生产力，从而创造更多繁荣、雇佣更多员工。我猜测就业率反而会上升，我们将创造更多工作岗位。

And it's more likely that someone will lose a job to someone else, some other human that uses an AI, and not likely to an AI, but to some other human that uses an AI. And so I think the first thing that everybody should do is learn how to use AI so that they can augment their own productivity. And every company should augment their own productivity to be more productive so that they can have more prosperity, hire more people. And so I think jobs will change. My guess is that we'll actually have higher employment, we'll create more jobs.

我认为各行业生产力将提升。当前面临劳动力短缺的产业很可能会借助AI重振旗鼓，恢复增长与繁荣。我的视角略有不同，但确实认为就业将受影响。我鼓励大家都去学习AI技术。

I think industries will be more productive. Many of the industries that are currently suffering from lack labor, workforce is likely to use AI to get themselves off their feet and get back to growth and prosperity. So I see it a little bit differently, but I do think that jobs will be affected. And I I'd encourage everybody just to learn AI.

这很贴切。我们在Acquired节目中经常讨论一个概念，称之为'莫里茨对摩尔定律的推论'，灵感来自红杉资本迈克·莫里茨...

This is appropriate. There's a version of something we talk about a lot on Acquired. We call it the Moritz corollary to Moore's law after Mike Moritz from

嗯哼。

Uh-huh.

是啊，我知道。对。

Yeah. I know. Yeah.

红杉是我们公司的首轮投资者。没错。

Sequoia was the first investor in our company. Yeah.

当然。这背后有个精彩故事：当迈克与道格接替唐·瓦伦丁时，他查看红杉的回报率，盯着包含思科的第三或第四期基金（可能是第四期），心想'我们怎么可能超越这个纪录？'

Of course. Yeah. Yeah. The great story behind it is that when Mike was taking over for Don Valentine with Doug, he was sitting and looking at Sequoia's returns and he was looking at fund three or four, I think it was four maybe that had Cisco in it. He was like, how are we ever gonna top that?

唐的成就我们永远无法企及。后来他意识到，随着计算成本降低并渗透到更广泛的经济领域，我们能触及的市场规模也会扩大。而AI，你的论点本质上是...

Can't, Don's gonna have us beat, we're never gonna beat that. And he thought about it and he realized that, well, as compute gets cheaper and it can access more areas of the economy because it gets cheaper and can get adopted more widely, well, then the markets that we can address should get bigger. Yeah. And AI, your argument is basically

正是。AI将...

Exactly. AI will

产生同样效果

do the same