英伟达第二部分：机器学习公司（2006-2022）

热身时听的Swedish House Mafia《Greyhound》还在我脑海里回响。

Still got Swedish House Mafia Greyhound in my head from the pump up.

不错，不错。有趣的是

Nice. Nice. It is funny how

所有GPU公司都这样，我为了准备这次节目看了很多英伟达和AMD的主题演讲，大家都搞得像科技霓虹灯秀一样。简直就像是加密货币兴起前的氛围。欢迎收听《Acquired》第十季第六期，这是一档讲述伟大科技公司及其背后故事与策略的播客。我是本·吉尔伯特，西雅图Pioneer Square Labs联合创始人兼董事总经理，同时也是我们风投基金PSL Ventures的负责人。

all, like, GPU companies like, I was watching a bunch of NVIDIA keynotes and AMD keynotes to get ready for this, and everyone is so, like, techno, neon, lighting. Like, it's like crypto before crypto. Welcome to season 10 episode six of Acquired, the podcast about great technology companies and the stories and playbooks behind them. I'm Ben Gilbert, and I am the cofounder and managing director of Seattle based Pioneer Square Labs and our venture fund, PSL Ventures.

我是大卫·罗森塔尔，旧金山的天使投资人。

And I'm David Rosenthal, and I am an angel investor based in San Francisco.

我们就是本期主持人。大卫，小时候我常盯着后院篝火发呆，思考那些跳动的火焰究竟是随机舞动，还是说——如果我知晓世间所有变量：空气状况、木材的物理构造、环境中的每个参数——其实它的轨迹是可预测的。虽然当时可能不知道这个术语，但本质上是在思考火焰是否具有可建模性。现在我们当然知道答案是肯定的，但要实现这种预测所需的数据和算力极其庞大。

And we are your hosts. When I was a kid, David, I used to stare into backyard bonfires and wonder if that fire flickering was doing so in a random way, or if I knew about every input in the world, all the air, exactly the physical construction of the wood, all the variables in the environment, if it was actually predictable. And I don't think I knew the term at the time, but modelable. If I could know what the flame could look like if I knew all those inputs. And we now know, of course, it is indeed predictable, but the data and compute required to actually know that is extremely difficult.

而这正是英伟达如今在做的事。

But that is what NVIDIA is doing today.

本，这个开场太棒了。我一直在想：本到底要引出什么话题呢？

Ben, I love that intro. That's great. I was thinking, like, where is Ben going with this?

当我看着黄仁勋分享NVIDIA的Omniverse愿景时，我突然意识到NVIDIA确实已经构建了完整的基石——包括硬件、供开发者使用的软件、面向用户的应用程序，以及通过他们难以置信的高效GPU架构来模拟物理世界一切的服务。各位听众，这些基石已不仅服务于游戏玩家。它们使得创建数字孪生来重现现实世界成为可能，比如预测机翼气流、模拟细胞互动以快速发现新药（无需触碰培养皿），甚至精确建模预测气候变化的影响。这里蕴含的内容太丰富了，尤其是NVIDIA如何从生产普通显卡发展到如今横跨游戏、企业数据中心、科学计算等多个行业的全栈布局，现在甚至为制造商提供现成的自动驾驶汽车架构。以他们的运营规模，这些技术进步简直超出人类想象。

And this was occurring to me as I was watching Jensen sharing the Omniverse vision for NVIDIA and realizing NVIDIA has really built all the building blocks, the hardware, the software for developers to use that hardware, all the user facing software now and services to simulate everything in our physical world with their unbelievably efficient and powerful GPU architecture. And these building blocks, listeners, aren't just for gamers anymore. They are making it possible to recreate the real world in a digital twin to do things like predict airflow over a wing or simulate cell interaction to quickly discover new drugs without ever once touching a petri dish or even model and predict how climate change will play out precisely. And there is so much to unpack here, especially in how NVIDIA went from making commodity graphics cards to now owning the whole stack in industries from gaming to enterprise data centers to scientific computing, and now even basically off the shelf self driving car architecture for manufacturers. And at the scale that they're operating at, these improvements that they're making are literally unfathomable to the human mind.

举个例子，如今训练一个单一的语音识别机器学习模型——仅仅一个模型——所需的数学运算（加减乘除）总量实际上超过了地球上所有沙粒的数量。

And just to illustrate, if you are training one single speech recognition machine learning model these days, one, just one model, The number of math operations, like adds or multiplies to accomplish it is actually greater than the number of grains of sand on the earth.

我完全知道你引用的研究出处，因为我读过同样的内容，当时我的反应是：这简直太离谱了。

I know exactly what part of the research you got that from because I read the same thing, and I was like, you gotta be freaking kidding me.

是不是很疯狂？在我们共同查阅的所有研究中，我想不出比这更能说明他们成就背后所需的数据量和计算规模有多么不可思议——而这一切竟然发生在单张显卡上，更显得这一切有多么难以想象的微小。

Isn't that nuts? I mean, there's just nothing better in all of the research that you and I both did, I don't think, to better illustrate just the unbelievable scale of data and compute required to accomplish the stuff that they're accomplishing and how unfathomably small all of this is, the fact that that happens on one graphics card.

没错。太棒了。

Yep. So great.

好了听众们，现在正是感谢我们最喜爱的公司之一Anthropic的最佳时机，他们最新的突破性模型Claude Sonnet 4.5已成为Acquired节目工作流程的核心部分。

Alright, listeners. Now is a great time to thank one of our favorite companies that has become a core part of our workflow for Acquired, Anthropic, and their latest breakthrough model, Claude Sonnet 4.5.

是的。在研究这些标志性公司时，我们不断提出诸如'他们处理这种情况的方式有何独特之处'、'这个策略有多新颖'、'之前有其他公司尝试过吗'等问题。这类问题及给出深思熟虑答案的能力，正是当今企业在构建AI时所需要的——而Claude确实能推理并回答这些问题。

Yes. As we research these iconic companies, we're constantly asking questions like, what was unique about the way they approach this situation, Or how novel was that strategy? And had any other companies tried it before? These kind of questions and the ability to produce thoughtful answers to them are exactly what today's enterprises need when building with AI. And Claude can actually reason through and answer them.

Claude Sonnet 4.5绝非普通模型。它是全球最优秀的编程模型，也是构建复杂智能体最强大的工具。Shopify和Netflix的工程师称其为强力思维伙伴，并表示它正彻底改变开发效率。Canva在其部分产品中使用Claude，认为这是重大突破。各企业都对Sonnet 4.5赞不绝口。

Claude Sonnet 4.5 isn't just another model. It's the best coding model in the world and the most capable for building complex agents. Engineers at Shopify and Netflix call it their powerful thinking partner and tell us that it is transforming their development velocity. And Canva, which uses Claude for some of its products, calls it a big leap forward. Companies are loving Sonnet 4.5.

一个显而易见的事实是：擅长编程的模型天生就擅长所有分析型任务。让Claude擅长代码重构的特性，同样使其能高效处理数千份监管文件或复杂财务分析。通过Anthropic API，Claude能无缝对接企业现有工作流，新增的记忆与上下文管理功能更让智能体长时间运行仍能保持关键信息不丢失。

And one thing that's become clear is that making a model great at coding also makes it great at any analytical task right out of the box. So the same thing that makes Claude great at refactoring code bases also makes it great at, say, combing through thousands of regulatory documents or doing complex financial analysis. Claude integrates seamlessly with enterprises existing workflows through Anthropics API and now has new memory and context management features that let agents run longer without losing critical information.

无论您是在扩展工程团队，还是构建下一代智能应用，Claude都能与您共同应对复杂挑战，而非简单代劳。它确实是您真正的智能思维伙伴。

So whether you're scaling an engineering team or building the next generation of intelligent applications, Claude thinks through complexity with you, not just for you. It is truly your intelligent thought partner.

立即访问claude.ai/acquired免费试用Claude，并享受Claude Pro三个月五折优惠。若需咨询企业版服务，只需告知是Ben和David推荐即可。

So head on over to claude.ai/acquired to try Claude for free and get 50% off Claude Pro for three months. And if you wanna get in touch about their enterprise offerings, just tell them that Ben and David sent you.

听完本期节目后，欢迎加入Slack社区acquired.fm/slack与我们交流。好了David，闲话少说，开始吧。听众们请注意，本节目不构成投资建议。David与我可能持有相关证券仓位，请务必自行研究。

And after you finish this episode, come join the slack, acquired.fm/slack, and talk about it with us. Alright, David. Without further ado, take us in. And as always, listeners, this is not investment advice. David and I may hold positions in securities discussed, and, please do your own research.

很好。我正想确认你这次说了免责声明，因为我们要详细讨论多年来投资NVIDIA股票的经历。这段旅程可谓惊心动魄——上次讲到我们勇敢的主人公黄仁勋和NVIDIA在『GPU公司』篇章尾声时，2004-2006年间处境相当艰难。他们不仅一次，而是两次绝处逢生。

That's good. I was gonna make sure that you said that this time because we're gonna talk a lot about investing in investors in NVIDIA stock over the years. It has been a wild wild journey. So last we left our plucky heroes, Jensen Huang and NVIDIA in the end of our NVIDIA, the GPU company years, ending kinda roughly, you know, 02/2004, 02/2005, 02/2006. They had cheated death not once, but twice.

第一次是在刚起步时竞争超饱和的显卡市场。后来他们刚跳出火坑，又落入英特尔全力围剿的熔炉——就像当年所有PCI芯片都要屈服于英特尔主板那样面临商品化危机。但他们英勇反击，与微软结盟，最终实现了GPU可编程化。

The first time in the super overcrowded graphics card market when they were first getting started. And then once they sort of, you know, jumped out of that frying pan into the fire of Intel now gunning for them, coming to commoditize them like all the other, you know, PCI chips that plugged into the Intel motherboard back in the day. And they bravely fend them off. They team up with Microsoft. They make the GPU programmable.

这太不可思议了。他们推出了搭载可编程着色器的GeForce 3显卡。他们为Xbox提供动力。他们与微软共同创造了CG编程语言。这就是我们现在看到的成就。

This is amazing. They come out with programmable shaders with the GeForce three. They power the Xbox. They create the CG programming language with Microsoft. And so here we are.

现在是2004年2月、2005年2月，这是一家相当令人印象深刻的公司。科技泡沫破裂后，这家上市公司的股票一路飙升。他们已经征服了显卡市场。当然，还有ATI的存在，这个我们稍后还会提到。但我觉得公司在头十年里建立了三个非常重要的优势。

It's now 02/2004, 02/2005, and this is a pretty impressive company. Public company stock is high flying after the tech bubble crash. They've conquered the graphics card market. Of course, there's ATI out there as well, which will come up again. But there's three pretty important things that I think the company built in the first ten years.

首先，我们上次讨论了很多关于他们芯片的六个月发布周期。我们谈到了这点，但没具体说他们发布产品的速度。我其实列了个小清单：1999年推出首款GeForce显卡GeForce 256，2000年发布GeForce 2。

So one, we talked about this a lot last time, the six month ship cycles for their chips. We talked about that, but we didn't actually say the rate at which they ship these things. I actually wrote down like a little list. So in the 1999, they shipped the first GeForce card, GeForce two fifty six. In the 2000, GeForce two.

2000年推出GeForce 2 Ultra，2001年发布具有可编程着色器的重磅产品GeForce 3，六个月后又推出GeForce 3 Ti 500。

In the 2000, GeForce two Ultra. 2001, G Force three, that's the big one with the programmable shaders. Then six months later, the G Force three t I 500.

我是说，正常周期应该是两年，竞争对手大概是十八个月——这些对手现在基本都被远远甩在后面了。

I mean, the normal cycle, I think we said, was two years, maybe eighteen months for most of their competitors who just got largely left in the dust.

我在想，确实，现在竞争对手都消失了，但考虑下英特尔。英特尔多久才发布一次新产品？更别提根本性的新架构了。从286到386再到奔腾，最后到...不知道，奔腾5之类的。

Well, I was just thinking, you know, yeah, the competitors are gone at this point, but I'm thinking about Intel. How often did Intel ship new products, let alone fundamentally new architecture? Architecture. You know, there was the two eighty six and then the three eighty six and the Pentium, and it got up to Pentium, I don't know, five, whatever.

老兄，我感觉英特尔的产品周期差不多和汽车换代一样漫长。

Dude, I feel like the Intel product cycle is approximately the same as a new body style of cars.

是的，正是如此。

Yes. Exactly.

每隔五六年，似乎就会出现一次重大的新架构变革。

Every five, six years, there seems to be a meaningful new architecture change.

英特尔是摩尔定律的推动者，对吧？这些家伙以极快的速度推出新架构，并一直延续至今。第二，我们上次遗漏了一个超级重要的事情，它成为了NVIDIA今天一切成就的重要基石——他们为自己的显卡编写了驱动程序。我们要特别感谢一位名叫杰里米的听众，他在Slack上联系我们并提供了大量信息，包括Asianometry YouTube频道的内容。

And Intel is the driver of Moore's Law. Right? Like, these guys ship and bring out new architectures at warp speed, and they've continued that through to today. Two, one thing that we missed last time that is super important and becomes a big foundation of everything NVIDIA becomes today that we're gonna talk about, they wrote their own drivers for their graphics cards. And we owed a big thank you for this and many other things to a great listener, very kind listener named Jeremy who reached out to us in Slack and pointed us to a whole bunch of stuff including, the Asianometry YouTube channel.

太棒了！这周我大概看了25个Asianometry的视频。

So good. I've probably watched like 25 Asianometry videos this week.

确实非常棒！向他们致以崇高的敬意。当时所有其他显卡公司和大多数外设公司都让下游合作伙伴来评估驱动程序的适用性，而NVIDIA是第一个站出来说'不，我们要掌控这一切'的公司。我们要确保使用NVIDIA显卡的消费者在任何系统上都能获得良好体验。

So so good. Huge shout out to them. But all the other graphics cards companies at the time and most peripheral companies, they let the further downstream partners rate the drivers for what they were doing. Nvidia was the first one that said, no no no, we wanna control this. We wanna make sure consumers who use Nvidia cards have a good experience on whatever systems they're on.

这意味着：一方面他们能保证质量，另一方面公司内部开始培养出一批精通底层软件开发的芯片工程师。很少有其他芯片公司具备这样的能力。

And that meant, a, that they could ensure quality, but, b, they start to build up in the company this, like, base of really nitty gritty low level software developers in this chip company. And there are not a lot of other chip companies that have capabilities like this.

没错。他们这样做实际上承担了更高的固定成本——为不同操作系统、不同OEM厂商、不同主板编写兼容驱动程序需要雇佣大量人员，成本非常高昂。但他们持有类似苹果的世界观：我们要尽可能掌控用户体验，确保使用我们产品的用户获得最佳体验。

No. And what they're doing here is taking on a bigger fixed cost base. I mean, it's very expensive to employ all the people who are writing the drivers for all the different operating systems, all the different OEMs, all the different boards that it has to be compatible with. But they viewed it as it's kind of an Apple esque view of the world. We want the control or as much control as we can get over making sure that people using our products have a great user experience.

所以他们愿意承受这种短期支出的痛苦，以换取长期的收益

So they were sort of willing to take the short term pain of that expense for the long term benefit of that

提升用户对其产品的体验。那些高端游戏玩家追求极致体验，他们会不惜花费304美元、500美元购买英伟达顶级显卡，装进自己组装的电脑里，他们希望一切都能完美运行。

improved user experience with their products. That their users, high end gamers that want the best experience, you know, they're gonna go out. They're gonna spend the time $3.04, $500 on an Nvidia top of the line graphics card. They're gonna drop it into the PC that they built. You know, they want it to work.

我记得以前折腾驱动程序的痛苦经历，这真的非常重要。当然，他们还有第三个优势——可编程着色器技术（ATI后来也效仿了），这些都是创新成果。此时此刻，所有这些都服务于游戏市场。

I remember messing around with drivers back in the day and things not working, like, this is super important. So all this is focused on, of course, they have the third advantage in the company is programmable shaders, you know, which ATI copies as well. But, like, they innovated, like, they've, you know, done all this. So all of this at this time, it's all in service of the gaming market.

大卫，这里要埋个伏笔——当你提到可编程着色器开发者时，'英伟达开发者'这个概念此前根本不存在。以前人们只是开发能在操作系统上运行的软件，计算任务可能会分流到显卡。但那时并没有专门为GPU开发的语言和库。这是他们首次与开发者建立直接联系，开始能说：'针对我们硬件开发，你会获得独特优势'。

And one seed to plant here, David, when you say the programmable shaders developers, the notion of a NVIDIA developer did not exist until this moment. It was people who wrote software that would run on the operating system. And then from there, maybe it would get that compute load would get offloaded to whatever the graphics card was. But it wasn't like you were developing for the GPU, for the graphics card with a language and a library that was specific to that card. So for the very first time now, they start to build a real direct relationship with developers so that they can actually start saying, look, if you develop for our specific hardware, there are advantages for you.

准确说是针对我们的游戏显卡。我们讨论的这些开发者都是游戏开发者，所有技术最终都服务于游戏市场。别忘了他们是上市公司，还与微软达成了重要合作。

And really, our specific gaming cards. Like, everything we're talking about, these developers, they're game developers. All of this stuff, it's all in service to the gaming market. So, you know, again, they're a public company. They have this great deal with Microsoft.

他们联合推出CG语言，为Xbox提供支持。华尔街对他们青睐有加，公司市值从科技泡沫破裂后的不足10亿美元，到2004-2005年间已增长至50-60亿美元。股价持续飙升。

They bring out CG together. They're powering the Xbox. Wall Street loves them. They go from sub a billion dollar market cap company after the tech crash up to 5 to $6,000,000,000 kind of by 02/2004, 02/2005. Stock keeps going on a tear.

到2007年年中，公司市值已接近200亿美元。这太棒了——这就是纯粹的游戏产业故事。他们在视频游戏这个庞大且持续增长的市场中，建立了巨大的技术优势和开发者生态系统。

By mid two thousand seven, the stock reaches just under $20,000,000,000 market cap. You know, this is great, and this is all the stories. Like, this is pure play gaming. These guys have built such a great advantage and a developer ecosystem in a large and growing market clearly, which is video games.

单就这一点而言，这已经是值得追逐的浪潮了。我是说，看看现在的游戏市场规模？1800亿美元左右吧？当我们与参与开创这个行业的特里普·霍金斯或诺兰·布什内尔交谈时，这个市场还几乎为零。而现在英伟达正站在一个惊人的转折点上。

Which on its own, that would be a great wave to surf. I mean, I think what's the gaming market today? A 180,000,000,000 or something? And when we talked to Trip Hawkins who sort of, like, helped invent it or Nolan Bushnell, you know, was zero then. And so NVIDIA is sort of, like, on a wave that's at an amazing inflection point.

他们完全可以乘着游戏产业的东风，成为举足轻重的

They can totally just ride this gaming thing and be an important

企业。但动力正在消退。我是说，作为创始人怎么可能不满足——岂止是满足，应该说是远超预期吧？你会觉得：没错，我已成为这个巨大市场里的领军企业，这股浪潮短期内看不到尽头。要知道，99.9%的创始人——这个群体本身就以雄心勃勃著称——都会对此心满意足

company. Running out of steam. I mean, like, how could you not be not just satisfied, but, like, more than satisfied with this as a founder? You're like, yes, I am the leading company in this major market, this huge wave that I don't see ending anytime soon. You know, 99.9% of founders who are themselves as a class, like, you know, very ambitious are gonna be satisfied

。但黄仁勋不会。

with that. But not Jensen.

但黄仁勋不会。当这一切发生时，他开始思考：下一篇章是什么？我虽然统治着这个市场，但还想继续成长。我不愿让英伟达只做游戏公司。

But not Jensen. So while all this is happening, he starts thinking about, well, what's the next chapter? You know, I'm dominating this market. I wanna keep growing. I don't want NVIDIA to be just a gaming company.

上回我们讲到那个近乎杜撰的小故事：斯坦福研究员给黄仁勋发邮件说，多亏你们的产品，我儿子让我去本地Fry's电子超市买现成的GeForce显卡，我把它们装进办公室电脑后——他好像是量子化学研究员——运行模型的速度比实验室超算快十倍。所以谢谢你，

So we ended last time with the little, you know, almost a surely apocryphal story of a Stanford researcher, you know, sends the email to Jensen and he's like, you know, thanks to you. My son told me to go buy off the shelf, you know, GeForce cards at the local Fry's Electronics and I stuffed them into my PC at work and, you know, I ran my models on on this. He's a I think it was a quantum chemistry researcher supposedly. It was 10 times faster than the supercomputer I was using in in the lab. And so thank you.

让我能在有生之年完成毕生研究。

I can get my life's work done in my lifetime.

黄仁勋特别喜欢这句话，每次GTC大会他都会提到。

And Jensen loves that quote. It comes out at every GTC.

这个故事如果让心存疑虑的听众听到，可能会引发两个疑问。首先是实际问题：我们刚说这里一切都与游戏有关，但这里却有个研究人员——用GeForce显卡做化学建模的科研人员。他到底用什么语言在编程？

So that story, if you're a skeptical listener, might beg two questions. First is a practical one. You know, we just said everything's about gaming here. And here's like a researcher, like a scientific researcher doing, you know, chemistry modeling using GeForce cards for that. What's he writing this in?

结果发现是可编程着色器对吧？他们把本为图形设计的CG语言生搬硬套过来，把所有研究内容都转换成图形术语——即便他们解决的并非图形问题，然后用CG语言编写。说真的，这可不是胆小者能做的事。

Well, it turns out Programmable shaders, right? Yeah. They were shoehorning CG, which was built for graphics. They were translating everything that they were doing into graphical terms, even if it was not a graphical problem they were trying to solve, and writing it in CG. This is not for the faint of heart, so to speak.

没错，所以一切都带着隐喻色彩。这位量子化学研究员本质上是在对硬件说：听着，把我给你的数据想象成三角形，而我要做的数据转换就像给三角形打点光效。

Right. So everything is sort of metaphorical. He's a quantum chemistry researcher, and he's basically telling the hardware, okay. So imagine this data that I'm giving you is actually a triangle. And imagine that this way that I wanna transform the data is actually like applying a little bit of lighting to the triangle.

然后我要你输出你认为正确的像素颜色，我再把它转译回量子化学需要的结果。你明白为什么这种操作不够理想了吧。

And then I want you to output something that you think is the right color pixel, and then I will translate it back into the result that I need for my quantum chemistry. Like, you can see why that's suboptimal.

是的。他认为这是个有趣的市场，希望英伟达能服务这个领域。但要真正做好这件事，可是个浩大工程——公司花了十多年才走到今天这步。

Yep. So he thinks this is an interesting market. He wants NVIDIA to serve it. If you really wanna do that right, it is a massive undertaking. It was ten plus years to get to the company to this point.

要知道，CG语言只是开发生态中的一小部分。要让开发者能用GPU做通用计算，就像我们讨论的这样，需要构建完整的软件栈。当年他们和微软合作开发CG，就像现在苹果搭建完整开发框架的差别——或者说，比微软当年为Windows开发.net框架的差距更大，懂我意思吗？

You know, what CG was is like a small sliver of the stack of what you would need to build for developers to use GPUs in a general purpose way like we're talking about. You know, it's kinda like, they worked with Microsoft to make CG. It's like the difference between working on CG and like Microsoft building the whole dot net framework for developing on Windows, you know. Or today, even better, Apple. Right?

比如苹果提供给iOS和Mac开发者的一切

Like, everything Apple gives to iOS and Mac developers

对。

Right.

在Mac上开发。

To develop on Mac.

对。这个类比虽然不完美，但就像是苹果不再只是说'用Objective-C为我们的平台写代码吧，祝你好运'，而是会说'好吧，你需要UI框架吗？那AppKit和Cocoa Touch怎么样？还有ARKit、StoreKit、HomeKit这些SDK和框架呢？'本质上你需要整个抽象层叠加在编程语言之上，才能让为特定领域开发软件变得真正可行——尤其是那些你知道会借助该硬件大受欢迎的领域。

Right. Yeah. The analogy is not perfect, but it's like instead of just Apple saying, okay, objective c is the way that you write code for our platforms. Good luck. They're like, okay.

所以当黄仁勋决定带领公司投身这一事业时，他接下的担子很重。我们之前聊过他们一直在自研驱动程序，这说明他们拥有大量底层——我不是说低水平，而是指基础设施层面、接近硬件的系统级编程人才。这让他们能从这里起步，但依然任重道远。于是作为英伟达的 discerning investor，你现在要问的第二个问题就是：黄仁勋，你让公司承担如此宏大的使命，

Well, will you need UI framework? So how about AppKit and Cocoa Touch? And how about all these other SDKs and frameworks like ARKit and like StoreKit and like HomeKit? It's basically you need the whole sort of abstraction stack on top of the programming language to actually make it very accessible to write software for domains and disciplines that you know are gonna be really popular using that hardware. Exactly.

商业依据是什么？向我证明市场存在。我的意思是，要是唐·瓦伦丁在场，他听完大概会直接说：给我看市场数据。

So when Jensen commits himself and the company to pursuing this, he's biting off a lot. Now we talked about they've been writing their own drivers, so they have actually a lot of very low level I don't mean low level like bad. I mean low level like infrastructure, like close, very difficult systems oriented programming talent within the company. So that kinda enables them to start here, but like still this is big. So then the second question, if you're a discerning investor, particularly in Nvidia, that you wanna ask at this point in time is like, okay, Jensen, you're committing the company to a big undertaking.

这背后的商业逻辑是什么？让我看到市场潜力。换作唐·瓦伦丁坐在这里听黄仁勋演讲，他肯定会说：证明市场存在。

What's the business case for that? Show me the market. I mean, Don Valentine at this point would be sitting there listening to Jensen and being like, show me the market.

不仅如此，它还要展示市场的潜力，更要预测市场何时会到来，以及我们需要投入多少时间、资金和资源，才能在市场真正形成时开发出对其有用的产品。因为虽然CUDA的开发始于2006年2月，但在NVIDIA内部，这个平台在六年多的时间里都未能达到实用阶段。

And not only is it show me the market, but it's how long will the market take to get here, And it's how long is it gonna take us and how many dollars and resources it gonna take us to actually get to something that's useful for that market when it materializes. Because while CUDA development began in 02/2006, that was not a useful usable platform for six plus years at NVIDIA.

没错。这更接近于微软或苹果的开发环境规模，而非NVIDIA之前所做的——仅仅是提供一些API并与微软合作，让你能为他们的设备编程。

Yep. This is closer to on the order of the Microsoft development environment or the Apple development environment than what NVIDIA was doing before, which was like, hey, we made some APIs and worked with Microsoft so that you can program for my thing.

对。我要快速跳到未来，来说明这项工程的疯狂程度。我在LinkedIn上搜索了目前在NVIDIA工作且头衔中包含'CUDA'的员工。有1,100名员工专门负责CUDA平台。

Right. I'm gonna flash way forward just to illustrate the insane undertaking of this. I searched LinkedIn for people who work at NVIDIA today and have the word CUDA in their title. There are 1,100 employees dedicated specifically to the CUDA platform.

我惊讶的不是11,000人。好吧。那么，这个市场的定位在哪里？是的，本。

I'm surprised it's not 11,000. Yeah. Okay. So, like, where's the market for this? Yes, Ben.

你提出了第三个问题：这需要什么条件才能实现，以及市场何时能在时间和成本等方面准备好。但即便抛开这些不谈，首要问题是——这个市场存在吗？目前的答案很可能是否定的。

You asked the, you know, the third question which is, okay, the intersection of what does this take to do this and when is the market gonna get there in time and cost and all that. But even just put that aside, is there a market for this? Is the first order question. And the answer to that is probably no at this point in time.

他们的目标是科学计算领域。对吧？那些在特定科学领域的研究人员，现在需要超级计算机或访问超级计算机来运行他们认为需要数周或数月的计算。如果能更便宜或更快地完成，岂不更好？是不是这类

And what they're aiming at is scientific computing. Right? It's researchers who are in science specific domains who right now need supercomputers or access to a supercomputer to run some calculation that they think is gonna take weeks or months. And wouldn't it be nice if they could do it cheaper or faster? Is that kind

他们瞄准的市场？是的。他们在进攻克雷（Cray）超级计算机那样的市场。你知道的，伟大的公司。对吧？

of the market they're looking at? Yeah. They're attacking, like, the Cray market, like Cray supercomputers, like that kind of stuff. You know, great company. Right?

但是，今天的英伟达已经不是那样了。对吧。他们曾经主导着市场。你知道吗？是的。

But, like, that's no NVIDIA today. Right. And they were dominating the market. You know? Yeah.

这是科研计算领域。你知道，药物研发。他们可能在想很多这类工作，哦，也许我们可以进入更专业的领域，比如好莱坞、建筑和其他专业图形领域。是的。是的。

It's scientific research computing. You know, it's drug discovery. It's probably a lot of this work they're thinking, oh, maybe we can get into more professional, like Hollywood and architecture and other professional graphics domains. Yeah. Yeah.

是的。当然。但是，你知道，把这些都加起来，可能也就几十亿美元的市场规模，也许整个市场加起来。嗯。对于任何理性的人来说，都不足以证明投入时间和成本去开发这个市场的合理性。

Yeah. Sure. But, you know, you sum all that stuff up and, like, maybe you get to a couple billion dollar market, maybe, like total market. Mhmm. And not enough to justify the time and the cost of what you're gonna have to build out to go after this to any rational person.

所以，你看，我们来了。黄仁勋和英伟达，他们正在做这件事。他全心投入。他已经深信不疑。2006年2月、2007年2月、2008年2月，他们投入大量资源开发即将成为CUDA的技术，我们稍后会讲到。

So, you know, here we come. Jensen and Nvidia, like, they are doing this. He is committed. He's drunk the Kool Aid. 02/2006, 02/2007, 02/2008, they're pouring a lot of resources into building what will become CUDA that we'll get to in a second.

其实，在这个时间点它已经是CUDA了。

Like, it already is CUDA at this point in time.

我认为黄仁勋的动机有两方面。一方面是他对这个市场非常着迷。他喜欢开发硬件来加速特定计算用例的想法，觉得这很奇妙，而且他希望通过计算机为人类做更多事情。但另一方面，这肯定也是商业模式的觉醒——在过去十三四年里，他在各个领域都被商品化，而他现在看到了实现持久差异化的路径，就像，哇。

And I think Jensen's psychology here is sort of twofold. One is he is enamored with this market. He loves the idea that they can develop hardware to accelerate specific use cases in computing that he finds sort of fanciful, and and he likes the idea of making it more possible to do more things for humanity with computers. But the other part of it is certainly a business model realization where he has spent the last, gosh, at this point, thirteen, fourteen years being commoditized in all these different ways. And I think he sees a path here to durable differentiation where he's like, woah.

掌控平台。你知道，这有点像苹果的模式，通过不仅拥有软件差异化的硬件，还与使用这些定制软件的开发者建立关系。这样我就能打造一个在行业内真正有影响力的公司。

To own the platform. You know, it's kind of the Apple thing again, to own the platform and to build hardware that's differentiated by not only software, but relationships with developers that use that custom software. Like, then I can build a really sort of like a a company that can throw its weight around in the industry.

百分之百。詹森，我不确定他当时是否用了这个说法，因为可能会招致批评，但他或许确实这么说过。我觉得他并不在意。他后来肯定用过——他的思路是，这不仅仅是‘我们建好了，他们就会来’（当时流行的说法），他用的短语是‘如果你不建造，他们就无法到来’。

A 100%. Jensen, I don't know if he used it at the time because he probably would have gotten pilloried, but maybe he did. I don't think he cared. He certainly, has used it since the way he thought about this was, it wasn't just like if we build it, they will come, which is what was going on. The phrase he uses is if you don't build it, they can't come.

所以这甚至不是‘是的，我相当确信只要我们建好，他们就会来’那么简单。这比那更进一步。意思是，‘如果我们不建造，他们连来的可能性都没有’。我不知道他们是否会来，但如果我们不建造，他们就绝对来不了。因此华尔街在2006年2月、2007年2月、2008年2月大多选择忽视这一点。

So it's not even like, yeah, I'm pretty sure if we build it, they will come. It's one step removed from that. It's like, well, if we don't build it, they can't even possibly come. I don't know if they will come, but they can't come if we don't build it. So Wall Street is mostly willing to ignore this in 02/2006, 02/2007, 02/2008.

公司当时仍保持着强劲增长。在金融危机前夕，其市值一路飙升。但后来，你上次提到的那个人——我记得交易可能在2006年宣布，2007年2月完成——AMD收购了ATI。没错。ATI曾是个非常强劲的竞争对手。

The company's still growing really nicely. There this great market cap run leading up to right before the financial crisis. But then, you know, who you mentioned last time, I think it gets announced in 2006 maybe and closes in 02/2007, AMD acquires ATI. Yep. And ATI was a very legit competitor.

这是英伟达整个发展历程中唯一长期存在的真正竞争对手。但现在AMD收购了它，我记得收购价大约是67亿美元之类。

It's the only standing legit competitor to NVIDIA through its whole life. But now AMD acquired it, I think they acquired it for, what, $67,000,000,000, something like that.

差不多这个数。

Something like that.

所以这是笔巨资，之后他们还投入了大量资源。他们收购可不只是为了获取人才。他们态度很明确：不，不，不。

So it was a lot of money, and then they put a lot of resources. Like, they weren't just acquiring this to, you know, get some talent. Like, they're like, no. No. No.

这将是我们重要的产品线。我们会全力支持这个项目。

This is gonna be a big product line for us. We're putting a lot of weight behind this.

我们对AMD的研究不如对NVIDIA那么深入，但AMD的Radeon系列——曾经是ATI的Radeon产品线——正是人们将AMD视为一家主要为游戏场景生产GPU的公司的方式。

We haven't done the research into AMD the way we have into NVIDIA, but the AMD Radeon line, which used to be the ATI Radeon line, that is how you think about AMD as a company is that they make these GPUs mostly for the gaming use case.

没错。在被收购前，我记得高中毕业刚上大学时组装的第一台电脑里就装了Radeon显卡。当时我可能属于少数派，Nvidia市场份额更大，但不知为何那个阶段我就是更喜欢ATI。他们确实很靠谱。

Yep. Before the acquisition, I think the first PC I built in, like, end of high school, beginning of college, I think I had a Radeon card in it. I think I was probably in the minority. I think Nvidia was bigger, but for whatever reason I liked ATI at that point in time. So like, they were legit.

而现在Nvidia全力押注新赛道时，AMD仍坚守游戏市场——正如我们所说这是个蓬勃发展的领域。你的竞争对手现在手握全部资源，AMD则全力追击。2008年中期Nvidia财报暴雷，这其实很自然。

Well, so here's Nvidia now focusing on this whole other thing. And you're still in the gaming market, which like we said is like massive rising tide. Your competitor now has all these resources and AMD that's fully dedicated to going after it. Mid two thousand eight, NVIDIA whiffs on earnings. Like, this is natural.

他们分心了。当然会这样。于是股价遭遇重创。

They took their eye off the ball. Of course, they did. And, the stock gets just hammered.

因为CUDA赋能的所有领域都尚未形成收入来源，他们完全忽视了游戏业务。

Because in anything that CUDA empowers is not yet a revenue driver, and they've totally taken their eye up off of gaming.

是的。我们说过其市值高峰约200亿美元，之后暴跌80%。这不只是金融危机的影响。现在回想金融危机都觉得有点遥远，当时人们看到道指单日跌5%就惊慌失措——

Yes. So, you know, we said the high was around a $20,000,000,000 market cap. It drops 80%, eight zero. This isn't just the financial crisis. It's almost quaint, I think, you know, for me thinking back on the financial crisis now and, like, people freaking out the Dow, you know, during the S and P dropping 5% in a day.

放现在也就是个普通周四罢了，懂吗？

Like, oh, that's a Thursday these days. You know?

我们录制节目的那天正好是星期四。

It is literally the Thursday that we are recording.

没错。一家科技公司的股票暴跌80%，即便是在金融危机期间，这不仅仅是进了惩罚区，简直是被彻底抛弃了。

Yes. For a company stock to drop 80%, a technology company stock, even during the financial crisis, they're not just in the penalty box. They're, like, getting kicked to the curb.

对。他们完了吗？现在的头条新闻都在问：英伟达的辉煌是否就此终结？

Right. Are they done? The headlines at this point are, is NVIDIA's run over?

如果你是此时大多数CEO，你可能会打电话给高盛，或者艾伦公司，或是弗兰克·夸特罗，然后开始兜售这家公司，因为还能怎么翻身呢？

If you're most CEOs at this point in time, you're probably calling up Goldman or, you know, Allen and Company or Frank Quatro and then you're shopping this thing because how are you gonna recover?

但黄仁勋不是这样。

But not Jensen.

显然黄仁勋没有这么做。相反，他继续开发CUDA并不断完善它。这里只是做个背景铺垫——虽然我们在Acquired节目里经常为各种事情兴奋，但我认为CUDA是过去十年、二十年甚至更长时间里最伟大的商业案例之一。

But not Jensen, obviously. So instead, he goes and builds CUDA and continues to build CUDA. And, this is, you know, just a set context. Like, get excited about a lot of stuff on Acquired. But I think CUDA is, like, one of the greatest business stories of the last ten years, twenty years, more.

我不知道。那你怎么看？

I don't know. What do you think then?

我是说，这绝对是我们报道过最大胆的赌注之一，但可编程着色器也是如此，英伟达最初尝试打造更高效的四边形聚焦架构也是如此。

I mean, I'd say it's one of the boldest bets we've ever covered, but so we're programmable shaders and so was NVIDIA's original attempt to make a more efficient quadrilateral focused Yeah.

图形领域。那些都是大赌注。但这次赌注的规模完全不同。这是我们《Acquired》节目不常涉及的那种赌注。

Graphics. Those were big bets. I think this is this is a bet on another scale, though. This is a bet that we don't cover that often on Acquired.

那些赌注相对于公司当时的规模确实很大，但这次赌注...

Those were big bets relative to the company's size at the time, but this bet is

就像是iPhone级别的豪赌。没错，这就是iPhone量级的赌注。

like an iPhone sized bet. That's exactly what this is. It's an iPhone sized bet.

这是当公司已成为价值数十亿美元企业时下的赌注。

It is a bet the company when you are already a several billion dollar company.

没错。如果他们成功且这个市场成型，这将创造一家划时代的公司。那么CUDA是什么？它是英伟达的计算统一设备架构。

Yes. An attempt to create something that if they are successful and this market materializes, this will be a generational company. Yep. So what is CUDA? It is NVIDIA's compute unified device architecture.

正如我们本期节目多次提到的，它是一个完整的——我强调完整——开发框架，用于在GPU上执行任何你需要的计算。

It is as we've referred to, you know, thus far throughout the episode, a full, and I mean full, development framework for doing any kind of computation that you would want on GPUs.

是的。特别有趣的是，我听Jensen既把它称为编程语言，也称作计算平台。它兼具这些身份——本质上是一个API。

Yeah. And in particular, it's interesting because I've heard Jensen reference it as a programming language. I've heard him reference it as a computing platform. It is all of these things. It's an API.

它是C或C++的扩展，因此某种程度上也算一种语言。但关键在于，它构建了完整的框架和库体系，支持超高层次的应用程序开发，为数百个行业提供高度抽象的开发层，最终通过CUDA与GPU及其他底层硬件通信。

It is an extension of c or c plus plus, so there's a way that it's sort of a language. But importantly, it's got all these frameworks and libraries that live on top of it, and it enables super high level application development, you know, really high abstraction layer development for hundreds of industries at this point to communicate down to CUDA, which communicates down to the GPU and everything else that they have done at this point.

这正是其精妙之处。记得我们发布第一部分内容的同一天——就是几周前那期关于英伟达的节目——Ben Thompson在Strathecari上对Jensen做了精彩访谈。那次访谈中Jensen对CUDA的定义和重要性的阐述，

This is what's so brilliant. So right after we released right, the same day that we released part one. Yep. The first NVIDIA episode we did a couple weeks ago, Ben Thompson had this amazing interview with Jensen on Strathecari. And Jensen in this interview, think, puts what CUDA is and and how important it is.

我认为比其他任何地方都更透彻。这是Jensen对Ben说的原话：'我们持续推动CUDA生态系统发展已十五年之久，通过GPU、加速库、系统和应用的全栈优化，同时不断扩展平台覆盖领域。我们从卓越的芯片出发，但针对每个科学领域、行业和应用，我们都构建了完整的技术栈。'

I think better than I've seen anywhere else. So this is Jensen speaking to Ben. We've been advancing CUDA in the ecosystem for fifteen years and counting. We optimize across the full stack iterating between GPU, acceleration libraries, systems, and applications continuously all while expanding the reach of our platform by adding new application domains that we accelerate. We start with amazing chips, but for each field of science, industry, and application, we create a full stack.

'我们拥有超过150个SDK，服务领域涵盖游戏设计、生命与地球科学、量子计算、人工智能、网络安全、5G和机器人技术。'接着他谈到实现这一切的艰辛——这正是我们试图强调的重点——他说：'你必须意识到这是全新的编程范式，作为处理器公司或计算平台公司所需的一切都必须从零构建。所以我们组建了编译器团队，

We have over a 150 SDKs that serve industries from gaming and design to life and earth sciences, quantum computing, AI, cybersecurity, five g, and robotics. And And then he talks about what it took to make this. This is like the point we give or try to like hammer home here. He says, you have to internalize that this is a brand new programming model and everything that's associated with being a program processor company or a computing platform company had to be created. So we had to create a compiler team.

必须考虑SDK开发、库构建，需要向开发者布道我们的架构优势，甚至要协助他们推广这一愿景，从而激发平台软件需求——诸如此类的工作永无止境。'

We had to think about SDKs. We had to think about libraries. We had to reach out to developers and evangelize our architecture and help people realize the benefits of it. And we even had to help them market this vision so that there would be demand for their software that they write on our platform and on and on and on.

这太疯狂了，但也令人惊叹。当他说这是全新的编程范式时——我记得他用了'范式'这个词——这完全符合事实。因为此前大多数编程语言和计算平台主要考虑串行执行，而CUDA则宣告：'听着，我们的GPU架构将依靠海量核心并行运算，这是彻底的并行编程与架构革命。'

It's crazy. It's amazing. And when he says that it's a whole new programming, I think he says maybe paradigm or way of programming, It is literally true because most programming languages up to this point and most computing platforms primarily contemplated serial execution of programs. And what CUDA did was it said, you know what? The way that our GPUs work and the way that they're gonna work going forward is tons and tons of cores all executing things at the same time, parallel programming, parallel architecture.

如今，他们最新的消费级显卡上集成了超过10,000个核心。这种并行程度简直疯狂，甚至可以说‘令人尴尬’，而CUDA从设计之初就是为并行执行而生的。

Today, there's over 10,000 cores on their most recent consumer graphics card. So insanely, or dare I say embarrassingly parallel, and CUDA is designed for parallel execution from the very beginning.

这就是行业里常说的‘令人尴尬的并行’这个流行语。

That's the, like, catchphrase in the industry of embarrassingly parallel.

实际上这是个技术术语。我不明白为什么说它‘尴尬’。它本质上是指软件具有极高的可并行性，意味着所有需要运行的计算都是独立的，不需要依赖前一个结果就能开始执行。就好比说，如果你按顺序执行这些指令而不想办法并行处理，那才叫尴尬。

And it's actually kind of a technical term. I don't know why it's embarrassing. It's basically the notion that this software is so parallelizable, which means that all of the computations that need to be run are independent. They don't depend on a previous result in order to start executing. It's sort of like it would be embarrassing for you to execute these instructions in order instead of finding a way to do it parallel.

并不是说并行本身令人尴尬。尴尬的是如果你还在用CPU按老办法串行处理这些任务。

It's not that it's parallel that's embarrassing. It's embarrassing if you were to do it the old way on CPUs serially.

我认为这就是隐含的意思。

I think that's the implication.

明白了。

Got it.

明白了。这太明显了，简直是令人尴尬的并行。

Got it. This is so obvious that it's embarrassingly parallel.

好的，现在说得通了。接下来是最精彩的部分——我们要花几分钟谈谈这有多天才。我们刚才描述的一切，这整个工程，简直就像建造埃及金字塔一样宏伟。

Okay. Now it makes sense. Now here's the coup de grace. We're gonna spend a few minutes talking about how brilliant this was. Everything we just described, this whole undertaking, like, it's like building the Pyramids Of Egypt or something here.

它完全免费。NVIDIA至今——虽然这可能会改变，我们会在节目最后讨论——从未对CUDA收取过一分钱。任何人都可以下载、学习、使用它，等等。所有这些成果都建立在NVIDIA所做工作的基础上。但是，本，这个'但是'是什么？

It is entirely free. NVIDIA, to this day, now this may be changing, we'll talk about this at the end of the episode, has never charged a dollar for CUDA. But anyone can download it, learn it, use it, you know, blah blah blah. You know, all of this work stand on the shoulders of everything NVIDIA has done. But, Ben, what is the but?

它是闭源的，且专属于NVIDIA的硬件。

It is closed source and proprietary exclusively to NVIDIA's hardware.

没错。你做任何这类工作，都只能在NVIDIA芯片上部署。这甚至不是NVIDIA在服务条款里规定你不能在AMD芯片或其他平台上部署这么简单。

That's right. You do any of this work, you cannot deploy it on anything but NVIDIA chips. And that's not even just like, oh, NVIDIA put in there, like, terms of service that you can't deploy this on, you know, AMD chips or or whatever.

字面意义上就是无法运行。

Literally doesn't work.

不行。这是全栈式的。就像你开发一个iOS应用然后试图在Windows上部署，根本行不通。它与硬件深度集成。

Nope. It's full stack. It's like if you were to develop an iOS app and then try and deploy it on Windows, like, it wouldn't work. It is integrated with the hardware.

所以OpenCL目前是主要竞争对手，他们确实允许OpenCL应用程序在其芯片上运行，但CUDA的任何内容都无法在其他平台运行。

So OpenCL is sort of the main competitor at this point, and they do actually let OpenCL applications run on their chips, but nothing in CUDA is available to run elsewhere.

这太棒了。好的。现在你可以看到这就像苹果一样，就是苹果的商业模式。苹果将他们构建的这个惊人平台生态系统全部开放给开发者，然后通过销售硬件获得非常非常高的毛利率。这就是为什么黄仁勋如此出色。

It's so great. Okay. So now you can see this is just like Apple, and it's the Apple business model. Apple gives away all of this amazing platform ecosystem that they built to developers and then they make money by selling their hardware for very very healthy gross margins. But this is why Jensen is so brilliant.

因为早在2006年他们开始这段旅程时，甚至更早之前，那时还没有iOS，没有iPhone。看起来这并不像是个好模式。事实上，大多数人认为这是个愚蠢的模式，认为苹果输了，Mac又蠢又小众，而Windows和Intel才是开放生态的赢家。

Because back when they started down this journey in thousand six, even before that when they started and then all through it, there was no iOS. There was no iPhone. Like, it wasn't obvious that this was a great model. In fact, most people thought this was a dumb model that, like, Apple lost and the Mac was stupid and niche and, like, Windows and Intel is what won the open ecosystem.

不过，Windows和Intel确实有专有的开发环境和全套开发工具。

Well, but Windows and Intel did have proprietary development environments and, you know, full stack dev tools.

哦，是的。这里有很多微妙之处。他们本身并不算开源，但可以在任何硬件上运行。

Oh, yeah. There's a lot of nuance here. It's not like they were, like, open source per se, but it could run on any hardware.

嗯，除了实际上并不能。它只能运行在Intel、IBM、微软联盟的硬件上。不能在Power PC上运行，也不能在苹果的任何产品上运行。

Well, except that it couldn't. It could only run on the Intel, IBM, Microsoft alliance world. It wasn't running on Power PCs. It wasn't running on anything Apple made.

确实如此。

That's true.

有趣的是，在某些方面，NVIDIA像苹果；而在另一些方面，他们又像微软、Intel、IBM联盟，只不过是完全整合在一起，而不是三家独立公司。

It's funny. In some ways, NVIDIA is like Apple. In other ways, they're like the Microsoft, Intel, IBM alliance except fully integrated with each other instead of being three separate companies.

是的，这么说可能挺贴切的。某种程度上介于两者之间，这里存在细微差别。还记得克莱·克里斯坦森在iPhone早期抨击苹果时说的那些话吗？

Yeah. That's maybe a good way to put it. It is sort of somewhere in between. There is nuance here. Remember when Clay Christensen was bashing on Apple in the early days of the iPhone being like

对，哦，是的。

Yeah. Oh, yeah.

开放系统会赢，安卓会赢，苹果要完蛋了。封闭系统从来行不通，必须模块化。

Open's gonna win. Android's gonna win. Apple is doomed. You know, closed never works. You gotta be modular.

不能搞整合。克莱很了不起，是最伟大的战略家之一，但我觉得这恰恰代表了当时所有人的想法——大家都认为苹果模式糟透了。

You can't be integrated. And like, you know, Clay was amazing and one of the greatest strategic but I think that's just representative to me of, like, everybody thought that, like, the Apple model sucked.

没错。除非你达到规模效应，否则确实糟糕。而当时几乎没人相信英伟达能达到支撑这种投资所需的规模，或者存在能让他们实现这种规模的市场。

Yeah. I mean, it sucks unless you're at scale. And at the time, there was very little to believe that Nvidia was going to have the scale required to justify this investment or that there was a market to let them achieve the scale to justify this.

问题就在这里。即便你说：好吧黄仁勋，我相信你，也认同如果能实现的话这是个好模式。但当时就算你是唐·瓦伦丁或其他投资人，可能也会质疑——毕竟他们可能还持有股票——你要实施这套打法所需的市场规模在哪里？

That's the thing. Even if you were to say, okay, Jensen, I believe you, and I agree with you that this is a good model if you can pull it off. At the time, you could be Don Valentine or whoever looking around, and maybe Don was still looking around because they probably still held the stock being like, where's the market that's gonna enable the scale you need to run this playbook?

好了听众朋友们，现在该聊聊我们另一家喜爱的公司Statsig了。自上次报道后他们有个激动人心的消息：完成了C轮融资，估值达到11亿美元。

Alright, listeners. It's time to talk about another one of our favorite companies, Statsig. Since you last heard from us about Statsig, they have a very exciting update. They raised their series c, valuing them at 1,100,000,000.

是啊，重大里程碑。祝贺团队。时机也很有趣，因为实验领域现在确实非常火热。

Yeah. Huge milestone. Congrats to the team. And timing is interesting because the experimentation space is, really heating up.

没错。那么为什么投资者对STAT SEG的估值超过十亿美元？因为实验已成为全球顶尖产品团队产品栈中的关键部分。

Yes. So why do investors value STAT SEG at over a billion dollars? It's because experimentation has become a critical part of the product stack for the world's best product teams.

是的。这一趋势始于Web 2.0时代的公司，如Facebook、Netflix和Airbnb。这些公司面临一个问题：如何在员工规模扩大到数千人的同时，保持快速、去中心化的产品和工程文化？实验系统就是这个答案的重要组成部分。

Yep. This trend started with web 2 dot o companies like Facebook and Netflix and Airbnb. Those companies faced a problem. How do you maintain a fast, decentralized product and engineering culture while also scaling up to thousands of employees? Experimentation systems were a huge part of that answer.

这些系统让这些公司的每个人都能访问一套全球产品指标，从页面浏览量到观看时长再到性能表现。每当团队发布新功能或产品时，他们都可以衡量该功能对这些指标的影响。

These systems gave everyone at those companies access to a global set of product metrics, from page views to watch time to performance. And then every time a team released a new feature or product, they could measure the impact of that feature on those metrics.

因此，Facebook可以设定一个公司范围内的目标，比如增加应用内停留时间，然后让各个团队自行想办法实现。将这种做法扩展到数千名工程师和产品经理身上，砰，你就获得了指数级增长。难怪实验现在被视为必不可少的基础设施。

So Facebook could set a company wide goal like increasing time in app and let individual teams go and figure out how to achieve it. Multiply this across thousands of engineers and PMs, and boom, you get exponential growth. It's no wonder that experimentation is now seen as essential infrastructure.

没错。如今最优秀的产品团队，如Notion、OpenAI、Rippling和Figma，同样依赖实验。但他们不再自行构建，而是直接使用Statsig。而且他们不仅用Statsig做实验。过去几年里，Statsig已经添加了快速产品团队所需的所有工具，如功能开关、产品分析、会话回放等等。

Yep. Today's best product teams like Notion, OpenAI, Rippling, and Figma are equally reliant on experimentation. But instead of building it in house, they just use Statsig. And they don't just use Statsig for experimentation. Over the last few years, Statsig has added all the tools that fast product teams need, like feature flags, product analytics, session replays, and more.

所以，如果你想帮助团队的工程师和产品经理更快地构建产品并做出更明智的决策，请访问statsig.com/acquired，或点击节目说明中的链接。他们提供非常慷慨的免费套餐、5万美元的初创企业计划，以及适合大公司的实惠企业合同。只要告诉他们是本和大卫推荐你的。好了，你要带我们去02/1112吗？

So if you would like to help your team's engineers and PMs figure out how to build faster and make smarter decisions, go to statsig.com/acquired, or click the link in the show notes. They have a super generous free tier, a $50,000 startup program, and affordable enterprise contracts for large companies. Just tell them that Ben and David sent you. Alright. So are you gonna take us to 02/1112?

还是说我们要重新回到这里？

Or are we hopping back in here?

要是世界能像小说那样运作，真的是一条直线就好了。但从来都不是直线。我们终会到达彼岸，正是这一点拯救了英伟达，让整个计划得以运转。不过他们在这期间确实遭遇了些波折。股价暴跌，那是2008年，我只是纯粹在个人推测。

If only the world were a a works like fiction and it were actually like a truly straight line. It's never a straight line. We will get there and that is what saves NVIDIA and makes this whole thing work. But they have some misadventures in between. So stocks getting hammered, it's 2,008 and, I'm just completely speculating on my own.

但他们当时处境艰难。他们坚持继续投资CUDA，致力于让GPU通用计算成为现实。我确实好奇他们是否想过：'我们总得做点什么来安抚股东吧。得证明我们在努力商业化。'

But they're in the penalty box. They're committed to continuing to invest in CUDA and making general purpose computing on GPU a thing, I do wonder if they felt like, well, we gotta do something to appease shareholders here. You know, we gotta show that we're trying to be commercial here.

嗯。

Mhmm.

时间来到2008年2月。你知道2008年2月的科技界在发生什么吗？是移动革命。就在那个二月，英伟达推出了Tegra芯片和平台。

So it's 02/2008. What's going on in 02/2008, you know, in the tech world? It's mobile. So in 02/2008, they launched the Tegra chip and platform within NVIDIA.

这可能不是拯救公司的关键。

This may not be what saved the company.

这确实不是公司的救命稻草。这更像是马戏团小丑车式的尝试——或许我对英伟达太苛刻了。但Tegra到底是什么？可能有人还记得这个名字。

This is not what saved the company. This is more a clown car style. Oh, that's maybe that's too rough on NVIDIA. But what was Tegra? People might recognize that name.

这是一款面向智能手机的全系统级芯片，直接与高通、三星竞争。它是一款基于ARM架构的处理器，包含系统级芯片所需的所有组件，用于驱动安卓设备。这简直是天马行空的转型，因为除了手机图形处理可能涉及英伟达的核心技术外，其他方面完全背离了他们的专长领域。但说真的，如果有什么场景最适合集成显卡，那非智能手机莫属。

It was a full on system on a chip for smartphones competing directly with Qualcomm, with Samsung. Like, it was a processor, like an ARM based CPU plus all the other stuff you would need for a system on a chip to power Android headsets. I mean, this is like a wild departure for it leverages none of NVIDIA's core skill sets except maybe graphics being part of smartphones. But, like, come on. If there's ever a use case for integrated graphics, it's smartphones.

没错。

Right.

没错。低功耗，小体积。

Right. Low power, smaller footprint.

是的，完全正确。知道吗？这是整个研究中最让我着迷的部分。你知道首款搭载Tegra芯片的产品是什么吗？

Yep. Totally. Do you know this is one of my favorite parts about the whole research. Do you know what the first product was that shipped using a Tegra chip?

不知道。是

No. It was

微软Zune HD媒体播放器。这个答案本身就说明了一切。不过Tegra系统至今仍以某种形式存在，它曾为初代特斯拉Model S的触控屏提供算力。在自动驾驶技术出现之前，它就是Model S车载娱乐系统和触控屏的处理器。

the Microsoft Zoom HD media player. That just tells you pretty much everything you need to know. It did though, the Tegra system, it is still around sort of to this day. It powered the original Tesla Model s touchscreen. So like before any of the autopilot autonomous driving stuff, they were the processor powering just the infotainment, the touchscreen infotainment in the Model s.

而我

And I

我认为这实际上有助于英伟达进入汽车市场。至今为止，Tegra平台仍是任天堂Switch的主要处理器。

think that actually starts to help Nvidia get into the automotive market. The Tegra platform still to this day is the main processor of the Nintendo Switch.

哦，他们是把它重新用于这个用途了吗？

Oh, they repurposed it for that?

是的，就是用于这个。而且我想他们还有自己的NVIDIA Shield专有游戏设备产品线，不过我不确定是否有人购买那些产品。

Yeah. For that. And they I think they still have their NVIDIA shield proprietary gaming device stuff that I don't know that anybody buys those.

哦，这就说得通了，因为他们基本上自PS3之后就退出了所有游戏主机市场。是的。所以有趣的是他们这个蓬勃发展的游戏部门却不支持任何主机——除了任天堂Switch。我总在想，他们为什么要接手Switch业务？因为他们基本上已经完成了这部分工作。

Oh, this makes so much sense because they basically have walked away from every console since the PlayStation three. Yep. And so it's interesting that they have this thriving gaming division that doesn't power any of the consoles except the Nintendo Switch. And I always sort of wondered, like, why did they take on the Switch business? Because they kinda already had it done.

这不是为了显卡业务，而是为了给Tegra芯片找个去处。

It's not for the graphics cards. It was as somewhere to put the Tegra stuff.

真有意思。顺便说个趣事，这些GPU公司在向移动端转型方面都不太成功。有个有趣的命名故事——你知道ATI Radeon后来变成AMD Radeon桌面系列后发生了什么吗？他们尝试做移动GPU，结果不太理想，最终剥离业务并把所有知识产权卖给了另一家公司。

Fascinating. Quick aside, it's funny how these GPU companies have not been good at transitioning to mobile. There's like a funny naming thing, but do you know what happened to so there's the ATI Radeon, which became the AMD Radeon desktop series. They tried to make mobile GPUs. It didn't go great, and they ended up spinning that out and selling all that IP to another company.

你知道是哪家公司吗？

Do you know the company?

噢，我不知道。是苹果吗？

Oh, I do not. Was it Apple?

是高通。今天讨论的是高通的移动GPU部门，高通在移动领域很擅长，所以这对它来说是自然而然的选择。你知道那系列移动GPU处理器叫什么吗？不知道。它叫Adreno，a r d e n o处理器。

It is Qualcomm. And it today is Qualcomm's mobile GPU division, and Qualcomm's good at mobile and so it's a natural home for it. Do you know what that line of mobile GPU processors is called? No. It is the Ardeno, a r d e n o processors.

那你知道为什么它叫Adreno或Ardeno吗？

And do you know why it's called the Ardeno or Ardeno?

不知道。听起来很耳熟，但确实不知道。

No. That sounds super familiar, but no.

这些字母是从Radeon重新排列而来的。

The letters are rearranged from Radeon.

太棒了。是啊，真不错。

That's great. Yeah. That's great.

所以你是说英伟达在移动图形领域的努力没有取得太大成功？

So you're saying NVIDIA's mobile graphics efforts didn't quite pan out?

不，我们在索尼那期节目里没怎么讨论这个，但在我看来，整个安卓价值链生态系统中没有任何利润可图，而且谷歌是故意维持这种状态的。

No. We didn't talk about this as much in the Sony episode, but my impression of the whole Android value chain ecosystem is that there's no profits to be made anywhere and Google keeps it that way on purpose.

讽刺的是，他们现在通过Play商店赚了很多钱。

Ironically, they make a lot of money now on the Play Store.

没错，就是Play商店和广告。

Yeah. The Play Store and ads.

对。我认为他们主要的变现方式是不必向他人支付搜索流量获取费用。

Right. I do think the primary way that they monetize it is not having to pay other people to acquire the search traffic.

是的。但我是说对合作伙伴而言。比如，如果你在安卓生态中从芯片到硬件全链条生产——哦对了，除非你是规模巨头，否则这些产品设计出来就是以极低价销售的，这里根本没有利润空间。

Right. But I mean for, like, partners. Like, if you are making Oh, yeah. Everything from chips all the way up through hardware in the Android ecosystem, I don't think you're making like, maybe if you were the scale player, but, like, these things are designed to sell for dirt cheap as in products. Like, there's no margin to be had here.

嗯嗯。另外，在继续之前，你刚才提到了AMD移动显卡芯片的题外话。我看到你的题外话了，我要再加一个我们必须包含的题外话——NZ的那些家伙告诉过我们这个。

Yep. Yep. Also, before we continue, you just did the sidebar on the AMD mobile graphics chip. I see your sidebar. I'm gonna raise you one more sidebar that we have to include that you know because the NZS guys told us about this.

所以当英伟达进军移动领域时，他们在2011年收购了一家名为iSera的英国移动基带公司。你应该明白我要说什么了。

So when Nvidia is going after mobile, they buy a mobile baseband company called iSera, a British company called iSera in 2011. You know where I'm going with this.

哦，是的。我确实这么认为。这太棒了。这是个值得回头再研究的好种子项目。

Oh, yes. I do. This is so good. It's a good seed plant to come back to later.

你知道，因为他们正在投资移动领域，Integra将会成为热门话题等等。然后几年后当他们基本关闭整个业务时，他们关闭了从iSera收购的项目，并裁掉了所有员工。那些在英伟达收购时赚了大钱的iSera创始人们，他们离开后创立了一家叫Graphcore的公司，我们会在节目最后稍微聊到它。这可能是主要的看空英伟达的论点之一。

You know, because they're investing in mobile, Integra is gonna be a thing blah blah blah. And, then a few years later when they end up pretty much shutting down the whole thing, they they shut down what they bought from iSera, they lay everyone off. The iSera founders who made a lot of money when Nvidia bought them, They go off and they found a company called Graphcore that, we're gonna talk about a little bit at the end of the episode. It's, you know, maybe one of the primary sort

看空英伟达的论点之一。

of NVIDIA bear cases.

看空英伟达的论点，所谓的'英伟达杀手'们。他们现在已经筹集了约7亿美元风险投资，并涉足了一些移动业务。

NVIDIA bear cases, NVIDIA killers out there. They've now raised about 700,000,000 in venture capital and pick up some, mobile.

某种程度上，这有点像贝索斯和jet.com的关系。是的。如果Jet当初成功了的话。我觉得这就是Graphcore与英伟达的类比关系。

In some ways, it's kind of like Bezos and jet.com. Yes. If Jet had been successful. I think that's sort of the graph core to Nvidia analogy.

是的。不过，我的意思是，现在下结论还为时过早，是否真有人能在与英伟达的竞争中取得重大成功。尽管我认为现在的市场可能讽刺性地足够大，足够容纳...是的。英伟达可以成为巨头，同时也能存在其他许多大型公司。

Yes. Well, I mean, jury's still out if, anybody's gonna be really successful in competing with NVIDIA. Although, I think the market now is probably ironically big enough that Large. Yeah. NVIDIA can be the whale and there can be plenty of big other companies too.

总之，好吧。回到故事主线。英伟达在2000年代末到2010年代初这段时间里磕磕绊绊地前行。有些年份增长率可能是10%，有些年份则基本持平。这家公司完全进入了平台期。

So anyway, okay. Back to the story. So NVIDIA's bumping along through all of this in the early late two thousands, early twenty tens. You know, some years, growth is, like, 10%, maybe it's flat in others. Like, this company has completely gone sideways.

2011年2月，他们又一次在财报上失手。股价再次腰斩50%。老套的情节。我本不想说这话。甚至不确定能不能这样评价Jensen。

In 02/2011, they whiff on earnings again. Stock goes through another 50% drawdown. It's cliche. I don't I was gonna say it. I don't even know if you can say it about Jensen.

看吧，我们又来了。公司再次陷入困境。换作别人早就放弃了，但他们显然没有。结果呢？基本上算是发生了奇迹。

Like, here we are. The company is screwed again. Like, everybody else would have given up, but obviously not them. So what happens? Basically, a miracle happens.

除了奇迹，我想不出其他方式来描述这件事。所以这可能并不是Jensen战略案例的好教材，因为它需要奇迹才能成功。

I don't know that there's any other way that you can describe this except like a miracle. So maybe this is actually not a great strategy case study of Jensen because it required a miracle.

嗯，Jensen会说这是有意为之，他们确实掌握了市场时机，战略正确，投资获得了回报，而且他们一直都在这么做。

Well, Jensen would say it was intentional, that they did know the market timing and that the strategy was right and the investment was paying off and that they were doing this the whole time.

嗯哼。当然。当然啦，Jensen。

Uh-huh. Sure. Sure, Jensen.

实际上，即使在Ben Thompson的采访中，我记得他说Ben基本上列出了——这些难以置信的事情怎么会在恰好的时机发生？而他的回应是：哦，没错。这都是我们计划好的。一切尽在掌握。

In fact, even in the Ben Thompson interview, I think he said Ben basically lays out, like, how did all these implausible things happen at exactly the right time? And and his response is, oh, yes. We planned it all. It was so intentional.

Jensen根本没有预见到AlexNet会出现，因为当时没人能预见AlexNet。所以在2009年2月，一位普林斯顿计算机科学教授——也是普林斯顿本科校友，就像鄙人一样。哇哦。好地方。名叫李飞飞。

Jensen did not plan AlexNet or see it coming because nobody saw AlexNet coming. So in 02/2009, a Princeton computer science professor and also undergrad alum of Princeton, just like yours truly. Woo. Wonderful place. Named Fei Fei Li.

他们的专长是人工智能和计算机科学。她开始着手一个名为ImageNet的图像分类项目。这个灵感其实源自普林斯顿大学八十年代的一个古老项目WordNet，那是对词语进行分类。而这个ImageNet是对图像进行分类。她的想法是创建一个包含数百万张标注图片的数据库，这些图片都附有正确标签，比如这是狗、这是草莓之类的。

Their specialty is artificial intelligence and computer science. Starts working on an image classifying project that she calls ImageNet. Now the inspiration for this was actually a way old project from, I think, the eighties at Princeton called WordNet that was, like, classifying words. This is classifying image ImageNet. And her idea is to create a database of millions of labeled images, like images that they have a correct label applied to them, like this is a dog or this is a strawberry or something like that.

有了这个数据库，人工智能图像识别算法就可以对照数据库进行测试，看看它们的表现如何。比如，我们看着一张图片说那是草莓，但不告诉算法答案，让算法自己判断它认为是草莓还是狗什么的。她和合作者们开始做这个项目，非常酷。

And that with that database, then artificial intelligence image recognition algorithms could run against that database and see how they do. So like, oh, look at this image of, you know, you and I were looking at be like, that's a strawberry. But you don't give the answer to the algorithm and the algorithm figures out if it thinks it's a strawberry or a dog or whatever. So she and her collaborators start working on this. It's super cool.

他们建立了这个数据库，使用亚马逊的Mechanical Turk众包平台来完成。然后其中一人——不确定是李飞飞还是其他人——想到：既然有了这个数据库，我们希望人们使用它，那不如搞个竞赛吧。

They build the database. They use Mechanical Turk, Amazon Mechanical Turk to build it. And then one of them, I'm not exactly sure who, if it was Fei Fei or somebody else, has the idea of, like, we know we've got this database. We want people to use it. Well, let's make a competition.

这在计算机科学学术界很常见：举办算法竞赛。他们决定每年举办一次，任何团队都可以提交针对ImageNet数据库的算法进行比拼，看谁的误差率最低、正确识别的图片比例最高。这个竞赛大获成功，为她在AI研究界赢得了巨大声誉。

This is, a very standard thing in computer science academia of, like, let's have a competition, an algorithm competition. So we'll do this annually, and anyone, any team can submit their algorithms against the ImageNet database, and they'll compete. Like, who can get the lowest error rate, like, most number of images percentage of the images correct. And, this is great. So it brings her great renown, becomes popular in the AI research community.

第二年她就被斯坦福大学挖走了。我想这没什么问题，因为我也去过那里。她还在...

She gets poached away by Stanford the next year. I guess that's okay because I went there too, so that's fine. Did she? She's still

也在那里吗？

there too?

我知道，我忍不住要提这个。她就是让我觉得特别投缘的那种人，你懂吗？

I know. I I couldn't resist. I couldn't resist. I'm just she's like a kindred spirit to me. Do you know?

我知道你清楚，但我敢说大多数听众并不了解她现在在斯坦福大学担任的捐赠讲席教授职位。

I know you do know, but I bet most listeners do not know what her endowed tenure chair is at Stanford today.

我知道。她是红杉讲席教授。

I do. She is the Sequoia chair.

没错。斯坦福大学计算机科学系红杉资本讲席教授。太酷了。她为何能成为红杉资本讲席教授？这一切又与英伟达有何关联？

Yes. The Sequoia capital professor of computer science at Stanford. So cool. Why does she become the Sequoia capital chair? And what does all this have to do with NVIDIA?

2012年的比赛中，多伦多大学的一个团队提交的算法赢得了比赛。而且不是险胜，而是以巨大优势获胜。他们衡量标准是：在数据库所有图像中，你的错误识别率是多少？这个算法以超过10%的优势胜出。

Well, in the twenty twelve competition, a team from the University of Toronto submits an algorithm that wins the competition. And it doesn't just win it by, like, a little bit. It wins it by a lot. So the way they measure this is the 100% of the images in the database, what percentage of them did you get wrong? So it wins it by over 10%.

我记得下一届比赛的错误率大概是15%左右。

I think it had a 15% error rate or something in the next.

之前最好的成绩也就25%左右。是的。这就像有人首次突破四分钟一英里。实际上某种程度上，这比四分钟一英里更令人惊叹，因为他们不是靠蛮力达到的，而是尝试了完全不同的方法。

Like, all the best previous ones have been, like, 25 something percent. Yes. This is like someone breaking the four minute mile. Actually, in some ways, it's more impressive than the four minute mile thing because they just didn't brute force their way all the way there. They, like, tried a completely different approach.

没错。然后突然之间就证明，我们能达到的准确度远超所有人的想象。

Yes. And then boom, showed that we could get way more accurate than anyone else ever thought.

那么这种方法是怎样的呢？他们组建了一个团队，由Alex Krusevskiy担任主要领导者。当时他是一名博士生，与Ilya Sutzkever和Jeff Hinton合作。Jeff Hinton是Alex的博士生导师。他们将其命名为AlexNet。

So what was that approach? Well, they called the team, which was composed of Alex Krusevskiy, was the primary, lead of the team. He was a PhD student in collaboration with Ilya Sutzkever and Jeff Hinton. Jeff Hinton was the PhD advisor of of Alex. They call it AlexNet.

它是什么？这是一种卷积神经网络，属于人工智能中被称为深度学习的分支。虽然深度学习对这个应用场景来说是新的，但Ben你并不完全正确。这个概念已经存在很久很久了，深度学习神经网络并非新想法。这些算法其实已经存在了几十年，只是它们对计算资源的需求极其庞大。

What is it? It is a convolutional neural network which is a branch of artificial intelligence called deep learning. Now deep learning is new for this use case, but Ben is you weren't exactly right. It had been around for a long time, a very long time and deep learning neural networks, this was not a new idea. The algorithms had existed for many decades I think but they were really really really computationally intensive.

训练深度神经网络模型需要海量计算资源——规模堪比地球上所有沙粒的数量级。在传统计算机架构下，要让这些模型在实际应用中发挥作用是完全不可能的。

They required to train the models to do a a deep neural network. You need a lot of compute, like, on the order of, you know, like grains of sand that exist on Earth. It was completely impossible with a traditional computer architecture that you could make these work in any practical applications.

人们也在预测，随着摩尔定律的发展，我们何时才能实现这个目标？当时看来仍遥不可及，因为不仅需要摩尔定律持续生效，还需要NVIDIA提出的大规模并行架构方案——这种架构带来的性能飞跃不仅源于单位面积内晶体管数量的增加，更关键的是实现了程序并行运算。

And people were forecasting too, like, when with Moore's Law, when will we be able to do this? And it still seemed like the far future because not only did Moore's law need to happen, but you also needed the NVIDIA approach of massively parallelizable architecture where suddenly you could get all these incredible performance gains, not just because you're putting, you know, more transistors in a given space, but because you're able to run programs in parallel now.

没错。AlexNet正是将这些古老理念在GPU上实现了。具体来说，他是在NVIDIA GPU上通过CUDA实现的。这一刻的重要性怎么强调都不为过——不仅对NVIDIA，对整个计算机科学、技术领域、商业世界、乃至我们每天盯着手机屏幕的生活而言，这都是人工智能的大爆炸时刻，而NVIDIA和CUDA正处于核心位置。

Yes. So Alex Nat took these old ideas and implemented them on GPUs. And to be very specific, he implemented them in CUDA on NVIDIA GPUs. We cannot overstate the importance of this moment, not just for NVIDIA, but for, like, computer science, for technology, for business, for the world, for us staring at the screens of our phones all day every day. This was the big bang moment for artificial intelligence, and NVIDIA and CUDA were right there.

有意思的是，2012到2013年间还有个类似案例：NVIDIA长期考虑过将其架构定位为通用计算，甚至想过是否要把GPU改名为GPGPU（通用图形处理器）。不过最终他们决定保留原名，只开发了CUDA。

Yep. It's funny. There's another example within the next couple of years, 2012, 2013, where NVIDIA had been thinking about this notion of general purpose computing for their architecture for a long time. In fact, they even thought about should we relaunch our GPUs as GP GPUs, general purpose graphics processing units. And, of course, they decided not to do that, but just built CUDA.

这其实是个暗号：'我们为这个技术寻找市场多年无果，干脆就说它能用于任何场景吧'。

Which is code word for, like, we've been searching for years for a market for this thing. We can't find a market, we'll just say, you can do use it for anything.

没错。深度学习当时引起了很大轰动，主要源自AlexNet竞赛。2013年，NVIDIA的研究科学家Brian Catanzaro与斯坦福大学的其他研究人员（包括吴恩达）发表了一篇论文，他们采用了Google Brain团队提出的无监督学习方法——该团队此前发布了相关研究成果，使用了上千个节点。这正处早期神经网络热潮时期，大家都在尝试各种酷炫技术。而这个团队仅用三个节点就实现了相同效果。

Right. And so deep learning's generating a lot of buzz, you know, a lot from this AlexNet competition. And so in 2013, Brian Catanzaro, who's a research scientist at NVIDIA, published a paper with some other researchers at Stanford, which included Andrew Ng, where they were able to take this unsupervised learning approach that had been done inside the Google Brain team, where they had sort of the Google Brain team had sort of published their work on this, and it had a thousand nodes. And, you know, this is a big part of the sort of early neural network hype cycle of people trying cool stuff. And this team was able to do it with just three nodes.

完全不同的模型架构，高度并行化，在极短时间内以高性能计算（HPC）方式完成海量运算。这最终成为了cuDNN的核心——这个深度神经网络库直接集成在CUDA中，让全球非硬件/软件工程师的数据科学家能轻松在NVIDIA硬件上编写高性能深度神经网络。AlexNet加上Brian和吴恩达的论文，彻底打破了人们认为不可逾越的界限，为后续团队提供了更高效、更低能耗的实现方案。

So totally different models, super parallelized, lots of compute for a super short period of time in a really high performance computing way or HPC as it would sort of become known. And this ends up being the very core of what becomes cooDNN, which is the library for deep neural networks that's actually baked into CUDA that makes it easy for data scientists and research scientists everywhere who aren't hardware engineers or software engineers to just pretty easily write high performance deep neural networks on NVIDIA hardware. So this AlexNet thing plus then Brian and Andrew Ng's paper, it just collapses all these sort of previously thought to be impossible lines to cross and just makes it way easier and way more performant and way less energy intensive for other teams to do it in the future.

是的，特别是针对深度学习。现在大家都知道这非常重要——既然能训练计算机自主识别图像，那么让它自主观察、自动驾驶、下国际象棋、下围棋、甚至让你用最新iPhone拍出惊艳照片（即使参数没调好）也就不足为奇了。

Yep. And specifically to do deep learning. So I think at this point, like, everybody knows that this is pretty important, but it's not that much of a leap to say if you can train a computer to recognize images on its own, that you can then train a computer to see on its own, to drive a car on its own, to play chess, to play Go, to make your photos look really awesome when you take them on the latest iPhone, even if you don't have everything right.

最终还能让你描述场景，然后由transformer模型生成令人难以置信的非人工创作画作。

To eventually let you describe a scene and then have a transformer model paint that scene for you in a way that is unbelievable that a human didn't make it.

更重要的是，对于黄仁勋和NVIDIA瞄准的市场，同源AI技术能预测你接下来可能想看什么内容，以及哪种广告对你特别有效。刚才提到的这些人物——相信很多人都认得——基本都被谷歌挖走了，比如李飞飞就去了谷歌。

Yep. And then most importantly, for the market that Jensen and Nvidia are looking for, you can use the same branch of AI to predict what type of content you might like to see next show up in your feed of content and what type of ad might work really, really, really well on you. So basically, all of these people we were just talking about, I bet a lot of you recognize their names. They get scooped up by Google. Fei Fei Li goes to Google.

Brian去了百度，现在又回到NVIDIA从事应用AI研究。

Brian went to Baidu, and he's back at NVIDIA now doing applied AI.

Brian去了百度，Jeff Hinton去了Facebook。即便抛开自动驾驶这些你不看好的领域，单是这项技术开启的数字广告市场，就是个价值数万亿美元的超级市场。

Brian went to Baidu. Jeff Hinton goes to Facebook. So, you know, all the other markets, like, even throw out say you don't believe in self driving cars, you don't think it's gonna happen or any of this other stuff. Like, it doesn't matter. Like, the market of advertising of digital advertising that this enables is a freaking multi trillion dollar market.

有趣的是，这感觉像是找到了杀手级应用场景，但其实这只是最简单的用例，最显而易见的那个。

And it's funny because, like, that feels like, oh, that's the killer use case. But that's just the easiest use case. That's the most, like

是的。

Yes.

那些模型不需要特别出色，因为它们不需要生成独特输出，只是协助提高效率——这些都有明确标注的数据集。但快进十年后，我们现在有了这些疯狂的Transformer模型，参数动辄数亿甚至数十亿。曾经认为只有人类能做的事，现在机器也能做了，而且发展速度前所未有。没错。

Obvious well labeled dataset that these models don't have to be amazingly good because they're not generating unique output. They're just assisting in making something more efficient. But then, like, flash forward ten more years, and now we're in these crazy transformer models with I don't know if it's hundreds of millions or billions of parameters. Things that we thought only humans could do are now being done by machines, and it's like it's happening faster than ever. Yep.

所以大卫，就像你说的，神经网络和深度学习在广告领域催生了大摇钱树。但这只是最浅层的应用，对吧。

So I think to your point, David, it's like, oh, there was this big cash cow enabled by, you know, neural networks and deep learning in advertising. Sure. But that was just the easy stuff. Right.

但这是必经之路。这个市场最终促成了规模化建设和技术发展。本·汤普森采访黄仁勋时也说过——这是本的原话：

But that was necessary, though. This was finally the market that enabled the building of scale and the building of technology to do this. And Yes. In the Ben Thompson Jensen interview, Ben actually says this when he's sort of realizing this talking to Jensen. He says, this is Ben talking.

在零边际成本的互联网世界，内容爆炸式增长，价值会流向那些帮你筛选内容的人（他讲的是聚合理论）。然后他说：黄仁勋，我听懂你的意思了——虽然价值流向内容导航者，但总得有人制造芯片和软件来支撑这个体系。就像当年Windows是用户界面层，英特尔是Wintel垄断的另一半。

The way value accrues on the Internet in a world of zero marginal costs where there's just an explosion and abundance of content, that value accrues to those who help you navigate the content. And he's talking about aggregation theory. Duh. Then And he says, what I'm hearing from you, Jensen, is that, yes, the value accrues to people that help you navigate that content, but someone has to make the chips and the software so that they can do that effectively. And it's like it sort of used to be with Windows was the consumer facing layer and Intel was the other piece of the Wintel monopoly.

现在谷歌、Facebook等消费端公司都依赖英伟达。这位置可太棒了——事实也确实如此。

This is Google and Facebook and a whole host of other companies on the consumer side and they're all dependent on NVIDIA. And And that sounds like a pretty good place to be. And indeed, it was a pretty good place to be.

真是个绝佳的去处。

Amazing place to be.

天啊。问题是市场多年来都没意识到这一点。我是说，我也没意识到，你可能也没意识到。我们这群在科技行业做风险投资的人本应该是最早察觉的。

Oh my gosh. The thing is like the market did not realize this for years. And I mean, I didn't realize this and I you probably didn't realize this. We were the class of people working in tech as venture capitalists that should have.

哦，你知道马克·安德森的那句话吗？哦，不知道。哦，这太棒了。好吧。那是几年前的事了。

Oh, do you know the Marc Andreessen quote? Oh, no. Oh, this is awesome. Okay. So it's a couple years later.

所以现在情况越来越明显了，但回到2016年。马克·安德森接受采访时说，我们投资了很多将深度学习应用于各个领域的公司，每家公司实际上都是在英伟达的平台上构建的。就像九十年代人们都在Windows上开发，或者两千年代末都在iPhone上开发一样。他还说，为了好玩，我们公司内部有个游戏：如果我们是对冲基金，会投资哪些上市公司。

So it's, like, getting more obvious, but it's 2016. And Marc Andreessen gave an interview. He said, we've been investing in a lot of companies applying deep learning to many areas, and every single one effectively comes in building on NVIDIA's platforms. It's like when people were all building on Windows in the nineties or all building on the iPhone in the late two thousands. And then he says, for fun, our firm has an internal game of what public companies we'd invest in if we were a hedge fund.

我们会把所有钱都投给英伟达。

We'd put in all of our money to NVIDIA.

这就像Paradigm那样，对吧？他们把某个基金的所有资金都集中起来，在比特币3000美元左右时全仓买入。我们当时都该这么做的。说真的，英伟达股价在2012、13、14、15年时每股从未超过5美元。而今天我们录制时，股价大概是220美元左右。过去一年的最高价远超300美元。

This is like, it was Paradigm, right, that called all of their capital in one of their funds and put it into Bitcoin when it was, like, $3,000 a coin or something like that. We all should have been doing this. So literally, Nvidia stock in 20, like, recent like, is now known 2012, 1314, '15, it doesn't trade above, $5 a share. And NVIDIA today as we record this is I think about $2.20 a share. The high in the past year has been well over 300.

如果你当时意识到发生了什么...而且说实话，在那几年里，要看清形势并不难。哇，这影响太巨大了。

Like, if you realized what was going on and and and again, in a lot of those years, it was not that hard to realize what was going on. Wow. Like, it was huge.

有趣的是。2017年和2018年加密货币的情况我们稍后会讲到。但在2018年，股价曾暴涨至每股65美元左右。甚至直到2019年初，你还能以这个价格买到。我发过推特，我们会在YouTube版本里把走势图放上屏幕。

It's funny. So there was even and we'll get to what happened in 2017 and 2018 with crypto in a little bit. But there was a massive stock run up to, like, $65 a share in 2018. And even as late as, I think, the very beginning of 2019, you could have gotten it. I tweeted this, and we'll put the graph on the screen in the YouTube version here.

在2019年那次暴跌中，你本可以以每股34美元的价格入手。如果你放大那张走势图——就是下一条推文里的——你会发现事后看来那次小崩盘根本不值一提。在股价疯狂飙升至3.5美元或历史新高的过程中，你甚至不会注意到它。

You could have gotten it in that crash for $34 a share. In 2019. If you zoom out on that graph, which is the next tweet here, that you can see that, like, in retrospect, that little crash just looks like nothing. You don't even pay attention to it in the crazy run up that they had to $3.50 or whatever their their all time high was.

是啊，太疯狂了。还有更离谱的：直到2016年——要知道AlexNet诞生于2012年——英伟达才重回2007年作为纯游戏公司时200亿美元的市值巅峰。

Yeah. It's wild. And a few more wild things about this. It's not until 2016 and again, AlexNet happens in 2012. It's not until 2016 that NVIDIA gets back to the $20,000,000,000 market cap peak that they were in 2007 when they were just a gaming company.

这几乎花了十年时间。

That's almost ten years.

我确实没从你描述的角度思考过，但突破发生在2010到2012年。很多人本有机会的，尤其是因为黄仁勋当时在公开场合反复强调——他在财报电话会议上都在谈论这个。

I really hadn't thought about it the way that you're describing it, but the breakthrough happened in 2010, 2011, 2012. Lots of people had the opportunity, especially because freaking Jensen's talking about it on stage. He's talking about it on earnings calls at this point.

他根本没打算保密。

He's not keeping this a secret.

没错。他几乎是在向全世界宣告这就是未来，但人们仍然持怀疑态度。没人抢购股票。我们眼睁睁看着奇迹通过他们的硬件和软件发生，就连那些仔细研究黄仁勋言论、紧密跟踪行业的半导体分析师，当时也觉得他在台上鼓吹'未来属于神经网络，我们要全力投入'时像个疯子。

No. He's, like, trying to tell us all that this is the future, and people are still skeptical. Everyone's not rushing to buy the stock. We're watching this freaking magic happen using their hardware, using their software on top of it. And, like, even semiconductor analysts who are, like, students of listening to Jensen talk and following the space very closely, sort of think he sounds like a crazy person when he's up there espousing that the future is neural networks, and we're gonna go all in.

我们并没有调整业务方向，但从他在财报电话会议上对此事与游戏业务的关注度对比来看，大家都觉得他是不是疯了。不过我认为人们

And we're not pivoting the business, but from the amount of attention that he's giving in earnings calls to this versus the gaming. I mean, everyone's just like, are you off your rocker? Well, I think people

他们早已失去信任和兴趣，你看，在CUDA技术推出这么多年后，他们早期甚至不知道AlexNet会出现。Jensen觉得GPU平台能实现CPU架构做不到的事，他坚信会有突破，但确实没预料到会是这样的突破。所以多年来他一直在说‘我们正在搭建这个平台’

had just lost trust and interest, you know, after, like, there were so many years of, like, they were so early with CUDA and early to know, again, they didn't even know that this like, they didn't know AlexNet was gonna happen. Right. Jensen felt like the GPU platform could enable things that the CPU paradigm could not, and he really, like, had this faith that something would happen. But, like, he didn't know this was gonna happen. And so for years, he was just saying that, like, we're building it.

‘他们终会到来’，明白吗？

They will come. You know?

具体来说是这样的：GPU加速了图形处理负载，所以我们把图形负载从CPU上剥离了。CPU很棒，它是处理各种灵活任务的主力，但图形处理需要独立环境，需要那些炫酷的风扇和强力散热。

And And to be more specific, it was that well, look. The GPU has accelerated the graphics workload. So we've taken the graphics workload off of the CPU. The CPU is great. It's your primary workhorse for all sorts of flexible stuff, but we know graphics needs to happen in its own separate environment and have all these fancy fans on it and get super cooled.

它还需要矩阵变换运算，核心就是矩阵乘法。当时人们开始意识到——就像那位传闻中的教授说的——既然矩阵变换能为他所用，或许这种矩阵运算对其他领域也很有用。确实，它在科学计算领域派上了用场。

And it needs these matrix transforms. The math that needs to be done is matrix multiplication. And there was starting to be this belief that like, oh, well, because the, you know, professor the apocryphal professor told me that he was able to use these program, the matrix transforms to work for him. You know, maybe this matrix math is really useful for other stuff. And sure, it was for scientific computing.

说实话，深度学习依赖大规模并行矩阵运算这个需求，简直是为NVIDIA量身定制的。他们看着自己的GPU说：‘我们正好有你们需要的东西’。

And then honestly, like, it fell so hard into NVIDIA's lap that the thing that made deep learning work was massively parallelized matrix math. And they're like, NVIDIA is just, like, staring down at their GPUs, like, I think we have exactly what you are looking for.

没错。Brian Catasaro在同一个采访里谈到这事时说：‘深度学习恰好是所有需要高吞吐量计算的应用中最重要的那个。’这简直是本世纪最轻描淡写的说法了。

Yes. There's a that same interview with Brian Catasaro. He says about when all this happened. He says, the deep learning happened to be the most important of all applications that need high throughput computation. Understatement of the century.

当英伟达意识到这一点时，几乎是瞬间就抓住了机遇。整个公司都紧紧抓住了这个机会。黄仁勋有很多值得称赞的地方。他不仅描绘了未来愿景，而且始终密切关注行业动向。公司也对任何正在发生的事情保持着高度警觉。

And so once NVIDIA saw that, it was basically instant. The whole company just latched onto it. There's so many things to laud Jensen for. Know, he was painting a vision for the future, but he was paying very close attention. And the company was paying very close attention to anything that was happening.

当他们发现这一趋势时，完全没有错失良机。

And then when they saw that this was happening, they were not asleep at the switch.

确实百分之百同意。有趣的是，从某些角度看这像是历史的偶然，但从另一些角度看又显得如此必然——图形处理本身就是个典型的并行计算问题，因为屏幕上每个像素都是独立的。我是说，他们并没有为每个像素配备独立核心。

Yeah. 100%. It's interesting thinking about the fact that in some ways, feels like an accident of history. In some ways, it feels so intentional that graphics is an embarrassingly parallel problem because every pixel on a screen is unique. I mean, they don't have a core to drive every pixel on the screen.

最新英伟达显卡也只有一万个核心，这听起来很疯狂对吧？但屏幕上的像素数量远多于此。所以它们并不是每个时钟周期都在同时处理所有像素。但巧合的是，神经网络同样可以完全并行计算，其中每个运算都独立于其他需要进行的计算。

There's only 10,000 cores on the most recent NVIDIA graphics cards, but there's not which is crazy. Right? But there's way more pixels on a screen. So, you know, they're not all doing every single pixel at the same time, every clock iteration. But it worked out so well that neural networks also can be done entirely in parallel like that, where every single computation that is done is independent of all the other computations that need to be done.

因此这些计算也能在这套超级并行核心系统上完成。当你把所有这些东西都归结为数学时，不禁会想：在寻找'还能用并行矩阵乘法解决哪些问题'的世界里，这两个庞大应用居然使用了同类型数学方法，这很有趣。可能还存在更多应用，甚至可能有更大的市场。

So they also can be done on this super parallel set of cores. It's just you gotta wonder, like, when you kinda reduce all this stuff to just math, it is interesting that these are two very large applications of the same type of math in the search space of the world of what other problems can we solve with parallel matrix multiplication, there may be more. There may even be bigger markets out there.

完全同意。我认为这很可能就是黄仁勋现在为英伟达描绘的愿景的重要组成部分——我们稍后会谈到——这仅仅是个开始。机器人、自动驾驶汽车、元宇宙，这些都将到来。有趣的是，我们刚才还在开玩笑说2016、2017年暴涨之前没人预见到这点。

Total well, I think they probably will be A big part of Jensen's vision that he paints for NVIDIA now, which we'll get to in a sec, is this is just the beginning. You know, there's robotics, there's autonomous vehicles, there's the Omniverse. It's all coming. It's funny. We just joked about how like nobody saw this before the run up in 2016, 2017.

那些年里，马克·安德森是否通过个人账户赚到钱不得而知——真想问问他。但到了2018年，另一个典型的可并行计算问题当然是加密货币挖矿。因此很多人在2016、2017年购买消费级英伟达显卡搭建矿机。而当2018年加密货币寒冬来临，ICO狂潮结束，矿机需求骤降。这对英伟达影响巨大，导致其收入实际下滑。

There were all these years where like, Marc Andreessen knew, you know, whether he made money in his personal account or not, you know, love to ask him. But then in 2018, another class of problems that are embarrassingly paralyzable is of course cryptocurrency mining. And so a lot of people were going out and buying consumer Nvidia, you know, graphics cards and using them to set up crypto mining rigs in 2016, 2017. And then when the crypto winter hit in 2018 and the end of the ICO craze and all that, the mining rig demand fell off. And this had become so big for NVIDIA that their revenue actually declined.

对，没错。这里有几件有趣的事情。我们来从技术角度聊聊原因。加密货币挖矿的运作方式本质上就是不断猜测和验证。

Right. Yeah. So couple interesting things here. Let's talk about technically why. So the way crypto mining works is effectively guess and check.

你实际上是在暴力破解一种加密方案。挖矿时，你试图发现某个难以找到的答案。如果猜错了，就递增数值继续猜测。虽然这是极度简化的说法，技术上并不完全准确，但这样理解是对的。好比解数学题时，你可能需要尝试数百万次才能找到正确答案，第一次就猜中的概率微乎其微，从概率学角度看这种事几乎不会发生。

You're effectively brute forcing an encryption scheme. And when you're mining, you know, you're trying to discover the answer to something that is hard to discover. So you're guessing if that's not the right thing, you're incrementing, you're guessing again. And that's a vast oversimplification and not technically exactly right, but that's the right way to think about it. And if you were gonna guess and check at a math problem and you had to do that on the order of a few million times in order to discover the right answer, you could very unlikely discover the right answer on the first time, but, you know, that probabilistically is only gonna happen to you once if ever.

这些芯片的厉害之处在于：第一，它们拥有海量核心。这类问题具有高度并行性——你不仅能用一个核心猜，还能同时用一万个核心并行猜测，接着再启用另外一万个。第二，它涉及矩阵运算。于是在游戏和神经网络之外，这十年间又出现了第三种应用场景，恰好都是这类芯片的专长领域。

And so well, the cool thing about these chips is that, a, they have a crap ton of cores. So the problem like this is massively parallelizable because instead of guessing and checking with one thing, you can guess and check with 10,000 at the same time, and then 10,000 more, and then 10,000 more. And the other thing is, it is matrix math. So yet again, there's this third application beyond gaming, beyond neural networks. There's now this third application in the same decade for the two things that these chips are uniquely good at.

有趣的是，你可以专门为加密挖矿或AI打造硬件，英伟达及其竞争对手确实都这么做了。但通用型GPU碰巧在这两个领域都表现得异常出色。

And so it's interesting that, like, you could build hardware that's better for crypto mining or better for AI, and both of those things have been built by NVIDIA and their competitors now. But the sort of, like, general purpose GPU happened to be pretty darn good at both of those things.

至少比CPU强了不知道多少个数量级。

Well, at least way way way better than a CPU.

就像英伟达的初创竞争对手Cerebras今天说的：GPU做这类运算比CPU强上千倍，但比起理想状态仍然差上千倍。对于某些AI运算，其实存在更优化的解决方案。

Yeah. As some of NVIDIA's startup competitors put it today, and Cerebras is the one that I'm thinking of, they sort of say, well, the GPU is a thousand times better or, you know, much much better than a CPU for doing this kind of stuff. But it's like a thousand times worse than it should be. There exist much more optimal solutions for, you know, doing some of this this AI stuff.

有意思。这确实引出一个问题：在这些应用场景中，到底多好才算足够好？

Interesting. Really begs the question of, like, how good is good enough in these use cases.

没错。现在快进来看，NVIDIA和其他所有新兴公司参与的游戏，本质上仍是加速计算的竞赛，但现在的焦点是如何将工作负载从CPU转移到GPU上进行加速。这很有趣。

Right. And now, I mean, to flash way forward, the game that NVIDIA and everyone else, all these upstarts are playing is really it's still the accelerated computing game, but now it's how do you accelerate workloads off the GPU instead of off the CPU. Interesting.

回到加密货币寒冬。NVIDIA股价再次遭受重创，跌幅达50%。这种事似乎每五年就要发生一次。

Well, back to crypto winter. Nvidia stock gets hammered again. It goes through another 50% drawdown. This is just like every five years this has gotta happen.

这很耐人寻味，因为归根结底，这完全超出了他们的控制范围。人们购买这些芯片的用途并非其设计初衷，他们根本不知道客户买来做什么。所以他们甚至无法获得准确的市场渠道信息——我们是在卖给矿工，还是卖给游戏玩家？

Which is fascinating because at the end of the day, it was a thing completely outside their control. Like, people were buying these chips for a use case that they didn't build the chips for. They had really no idea what people were buying them for. So it's not like they could even get really good market channel intelligence on are we selling to crypto miners or are we selling to, you know, people that are gonna use these for gaming.

他们只是把货卖给百思买，然后消费者去百思买购买。

They're selling to Best Buy, and then people go buy them in Best Buy.

对。有些人会批量采购，比如真正要建矿场的数据中心。但更多人只是用家用设备在地下室挖矿，所以他们无法掌握完整信息。而当币价暴跌，挖矿变得无利可图时，这种你既没有主动争取、又难以追踪真实需求的市场就会迅速萎缩。

Right. And some people are buying them wholesale, like, if you're actually starting a data center to mine. But a lot of people are just doing this in their basement with consumer hardware, so they don't have perfect information on this. And then, of course, the price crashing makes it either unprofitable or less profitable to be a minor. And so then your demand dries up for this thing that you, a, didn't ask for and, b, had poor visibility into knowing if people were buying in the first place.

因此在这个阶段，管理层在华尔街眼中显得极其无能，因为他们完全无法理解自己业务的实际情况。

So the management team just looks terrible to the street at this point because they had just no ability to understand what was going on in their business.

而且我认为当时华尔街仍对深度学习持怀疑态度——'这什么玩意儿？黄仁勋在搞什么？'所以随便找个理由就抛售股票。

And I think a lot of street was still was still had this hangover of skepticism about is this deep learning thing? Like, what? Jensen? Okay. And so, you know, it's kinda any excuse to sell off.

虽然短暂，但很快就达到了50%的深度，因为GPU在深度学习中的应用场景，尤其是企业级应用场景，一下子就爆发了。这非常有趣。如果你看NVIDIA的财报，他们用几种不同的方式报告财务数据，其中一种方式是将业务划分为几个不同板块：游戏消费板块和数据中心板块。而数据中心板块，他们在数据中心做什么？全都是AI。

It took, but anyway, that was short lived to the 50% depth because, with the use case and specifically the enterprise use case for GPUs for deep learning, like, it just takes off. And so this is really interesting. If you look at NVIDIA's, they report financials a couple different ways, but one of the ways they break it out is few different segments is the gaming consumer segment and then their data center segment. And it's like data center, like what are they in the data center? Well, all the Data centers AI.

没错。我们讨论的所有内容都是在数据中心完成的。比如，谷歌不会去买一堆NVIDIA的GPU然后连到他们软件工程师的笔记本电脑上。

Right. All of the stuff we're talking about, it's all done in the data center. Like, Google isn't going and buying, you know, a bunch of NVIDIA GPUs and hooking them up to the laptops of their software engineers. Like

Stadia还存在吗？我记得那是用于云游戏之类的，好像还有...

Is Stadia still a thing? Like, I think that's used for cloud gaming and some like, there there

但我的重点是，这一切都发生在数据中心。

are But if it's all happening in the data center is my point.

对，对。我想说的是，每次我看到数据中心收入时，在我脑海中，我几乎把它等同于他们的机器学习板块。

Right. Right. I guess what I'm saying my argument is every time I see data center revenue, I in my mind, I sort of make it synonymous with this is their ML segment.

是的，是的。我就是这个意思。我同意。

Yes. Yes. That's what I'm saying. I I agree. Yeah.

现在说到数据中心，这又非常有趣，因为他们过去卖这些显卡是打包上架，消费者购买。他们也为科学计算市场做过一些特殊显卡之类的。但这个数据中心的机会，天啊，你知道卖给数据中心的设备能卖多少钱吗？让RTX 3090看起来像零花钱。

Now the data center, this is really interesting again because they used to sell these cards that would get packaged, put on a shelf, a consumer would buy them. Yeah. They made some specialty cards for the scientific computing market and stuff like that. But this data center opportunity, like, man, do you know the prices that you can sell gear to data centers for? Like, it makes the RTX thirty ninety look like a pittance.

RTX 3090是他们最昂贵的高端显卡，消费者能买到的，价格曾是3000美元。现在大概是2000美元。但如果你要买...最新的不是A100。

And the RTX thirty ninety, which is their most expensive high end graphics card that you can buy as a consumer, was $3,000. Now it's like $2,000. But if you're buying I don't know. What's the latest? It's not the a 100.

是H100。

It's the h 100.

A100刚发布不久，他们就宣布了H100。

So the a 100, they just announced the h 100.

那要多少钱？一张卡就要20或30美元吗？

And that's what? Like, 20 or $30 in order to just get one card?

是的。而且人们正在大量购买这些产品。

Yeah. And people are buying a lot of these things.

是啊，太疯狂了，简直难以置信。

Yeah. It's crazy. It's crazy.

有意思。我发推文说过这事，我有点说错了，但就像所有事情一样，这里面有细微差别。特斯拉已经宣布要自研硬件，他们肯定是为了车载推理设备做的，比如特斯拉的全自动驾驶计算机。

It's funny. I tweeted about this. I was sort of wrong, but then like everything, there's nuance. You know, Tesla has announced making their own hardware. They're certainly doing it for the on the car, the inference stuff, like the full self driving computer on Teslas.

他们现在自己生产那些芯片。特斯拉Dojo是他们宣布的训练中心，他们还宣布要为此自研硅芯片。实际上他们还没做到，所以目前仍在用英伟达芯片进行训练。他们现有的计算集群仍在运行，我算了一下，假设了一些价格。

They now make those chips themselves. The Tesla Dojo, which is the training center that they announced, they announced they were also gonna make their own silicon for that. They actually haven't done it yet. So they're still using NVIDIA chips for their training. The current compute cluster that they have, that they're still using, I wanna say I did the math and, like, assumed some pricing.

我认为他们向英伟达支付了5千万到1亿美元，用于购买所有这些计算资源。哇。

I think they spent between 50 and a $100,000,000 that they paid NVIDIA for all of the compute in that Wow.

集群。一个客户。

Cluster. One customer.

这是一个客户在单一应用场景下的采购规模。

That's one customer for one use case at that one customer.

疯狂。我是说，这体现在他们的财报里。现在节目进行到接近当下时间点，用最新数据最能说明问题。我直接快进展示数据中心业务现状：两年前这个部门收入约30亿美元，只有游戏业务收入的一半。

Crazy. I mean, you see this show up in their earnings. So we're at the part of the episode where we're close enough to today that it's best illustrated by the today numbers. So I'll I'll just flash forward to what the data center segment looks like now. So two years ago, they had about 3,000,000,000 of revenue, and it was only about half of their gaming revenue segment.

游戏业务，从2006年到AlexNet出现，再到2020年这十几年间，始终是王者。年收入近60亿美元。数据中心业务30亿美元，但多年增长停滞。而过去两年竟疯狂增长三倍。数据中心业务翻了三倍。

So gaming, you know, through all this, through 2006 to AlexNet, all the way, you know, another decade forward to 2020, gaming is still king. It generates almost 6,000,000,000 in revenue. The data center revenue segment was 3,000,000,000, but had been pretty flat for a couple years. So then insanely over the last two years, it three x. The data center segment three x.

现在年收入超过105亿美元，规模已与游戏业务相当。太疯狂了。2010年代中期趋势就很明显，但当企业真正入场，大量采购硬件部署在数据中心——无论是超大规模云服务商谷歌、微软、亚马逊，还是企业自建私有云或所谓的本地数据中心——爆发式增长就来了。

It is now doing over 10 and a half billion a year in revenue, and it's basically the same size as the gaming segment. It's nuts. It's amazing how it was, like, sort of obvious in the mid twenty tens. But when the enterprises really showed up and said, we're buying all this hardware and putting it in our data centers. And then whether that's the hyperscalers, the, like, cloud folks, Google, Microsoft, Amazon putting it in their data centers, or whether it's companies doing it in their own private clouds or whatever they wanna call it these days, on prem data centers.

现在每个人都在数据中心使用机器学习硬件。

Everyone is now using machine learning hardware in the data center.

没错。而且英伟达以非常、非常、非常健康的毛利率出售这些产品，达到了苹果级别的毛利率。

Yep. And NVIDIA is selling it for very, very, very healthy gross margins, Apple level gross margins.

是的，正是如此。

Yes. Exactly.

说到数据中心，有几件事值得一提。其一，这非常英伟达风格。2018年，他们确实修改了消费级显卡GeForce的用户协议条款，禁止将其用于数据中心。

So speaking of the data center, couple things. One, in this is so NVIDIA. In 2018, they actually do change the terms of the user agreements of their consumer cards of GeForce cards that you cannot put them in data centers anymore.

他们大概觉得，确实需要开始进行一些市场细分了。而且我们知道企业客户有更强的支付意愿，这很值得。我是说，你买这些疯狂的数据中心显卡，它们的晶体管数量翻倍。实际上，它们甚至没有视频输出接口。也就是说，你无法将这些数据中心GPU当作普通显卡使用。

They're like, course, we really do need to start segmenting a little bit here. And, we know that the enterprises have much more willingness to pay, and it it is worth it. I mean, you buy these crazy data center cards, they have, like, twice as many transistors. And, actually, they don't even have video outputs. Like, you can't use the data center GPUs.

比如A100就没有视频输出。所以它们实际上不能作为显卡使用。

Like, the a 100 does not have video out. So they actually can't be used as graphic cards.

哦，是的。Linus Tech Tips有个很酷的视频讲这个，他们不知怎么搞到了一块A100，然后跑了一些基准测试，但实际上无法用它来运行游戏。

Oh, yeah. There was a there's a cool, Linus Tech Tips video about this where they get a hold of an a 100 somehow, and then they run some benchmarks on it, but they can't actually, like, drive a game on it.

哦，真有趣。

Oh, fascinating.

是啊，太好玩了。

Yeah. So fun.

数据中心设备虽然性能超强，但显然不适合用来运行游戏，因为你没法把它连接到电视或显示器上。但有趣的是，他们现在反其道而行之，对不想花三万美元购买专业设备、试图在家搭建小型数据中心设备的人说：不，你没法用机架安装

Data center stuff is, like, super high horsepower, but, of course, like, useless to run a game on because you can't pipe it to a TV or a monitor. But then it's interesting that they're sort of artificially doing it the other way around and saying, for those of you who don't wanna spend $30,000 on this and are trying to, like, make your own little rig at home, your own little data center rig at No, you cannot rack

这些东西。别想着去Fries买一堆T Force显卡。讽刺的是整个行业最初就是这么起步的。总之在2020年，他们收购了以色列数据中心计算公司Mellanox——这家公司主要专注于数据中心内部的网络计算——没错，花了约70亿美元，将其整合进他们拓展数据中心的宏图里。

these things. Don't think about going to fries and buying a bunch of T Forces. Ironic because that's how the whole thing started. But anyway, in 2020, they acquire an Israeli data center compute company called Mellanox that I believe focuses on like, networking compute within the data center Yep. For about seven billion, integrate that into, you know, their ambitions in building out the data center.

理解Mellanox带来的价值在于：现在他们能在数据中心内部硬件之间实现超高带宽、超低延迟的互联。此时他们已经拥有NVLink技术——就像苹果的专有互连技术（AMD称之为无限架构）——这是一种芯片间超高带宽连接方案。Mellanox让他们能在数据中心部署这些极高带宽的交换机，使得所有搭载NVIDIA硬件的机箱能实现极速互联。

And the way to think about what Mellanox enables them to do is now they're able to have super high bandwidth, super low latency connectivity in the data center between their hardware. So at this point, they've got NVLink, which is their it's like the what does Apple call it? A proprietary interconnect, or I think AMD calls it the infinity fabric. It's the, like, super high bandwidth chip to chip connection. So think about what Mellanox lets them do is it lets them have these extremely high bandwidth switches in the data center to then let all of these different boxes with NVIDIA hardware and then communicate super fast to each other.

这太棒了。因为像特斯拉这样的客户——正如我之前举例的——他们不是购买显卡的企业用户，而是向NVIDIA采购整体解决方案，购买的是装满各种设备的大型机箱。

That's awesome. Because, of course, these data centers that's the other thing about, you know, customers like that Tesla example I gave. They're not buying cards, the enterprise because they're buying solutions from NVIDIA. They're buying big boxes with lots of stuff in them.

你说解决方案，我听到的是毛利率。

You say solutions, I hear gross margin.

这句话说得太好了。我们应该把它裱起来挂在

That's such a great quote. We should, like, frame that and put it on

墙上，

the wall of the,

收购来的博物馆里。

the acquired museum.

确实，收购Mellanox不仅让我们现在拥有了超高速连接技术，更重要的是这为NVIDIA引入了他们现在常说的计算三足鼎立中的第三条腿——你原本有CPU，它很棒，是主力军，是通用计算机；然后有GPU，特别是他们强化过的通用GPU（GPGPU）。

It is true that acquiring Mellanox not only, like, enables this now we have the super high connectivity thing, but this is what leads to this introduction of this third leg of the stool of computing for NVIDIA that they talk about now, which is you had your CPU. It's great. It's your workhorse. You know, it's your general purpose computer. Then there's the GPU, which is really a GP GPU that they've really beefed up.

他们为企业级数据中心专门加入了张量核心，能极速执行机器学习专用的4x4x4矩阵乘法，还集成了各种非游戏场景的数据中心专用AI模块。现在他们说：有了CPU和GPU，现在还有DPU（数据处理单元）。这个脱胎于Mellanox技术的DPU，正是实现数据中心内高效通信和数据转换的关键。现在你对黑盒子的认知要从机架上的单个设备，升级为把整个数据中心看作黑盒子。

And they've really like, for the enterprise, for these data centers, they've put tensor cores in it to do the machine learning specific four by four by four matrix multiplication super fast and do that really well. And they've put all this other non gaming data center specific AI modules onto these chips and and this hardware. And now what they're saying is you've got your CPU, you've got your GPU, now there's a DPU. And this data processing unit that's, like, kinda born out of the Mellanox stuff is the way that you really efficiently communicate and transform data within data centers. So the unit of how you think about like, the black box just went from a box on a rack to now you can kinda think about your data center as the black box.

你可以在极高抽象层进行编程，NVIDIA会协助处理数据中心的资源调度。现在正是感谢节目好朋友ServiceNow的好时机。我们曾向听众介绍过ServiceNow传奇的创业故事，以及他们如何成为过去十年表现最佳的企业之一。但仍有听众询问ServiceNow具体业务，今天我们就来解答这个问题。

And you can write at a really high abstraction layer, and then NVIDIA will help handle how things move around the data center. Now is a great time to thank good friend of the show, ServiceNow. We have talked to listeners about ServiceNow's amazing origin story and how they've been one of the best performing companies the last decade, but we've gotten some questions from listeners about what ServiceNow actually does. So today, we are gonna answer that question.

首先，最近媒体常用一个说法：ServiceNow是企业级的'AI操作系统'。具体来说，22年前ServiceNow成立时只专注于自动化——将实体文书转化为软件工作流，最初服务于企业内部的IT部门。后来他们在这个平台上不断扩展，处理更强大复杂的任务。

Well, to start, a phrase that has been used often here recently in the press is that ServiceNow is the, quote, unquote, AI operating system for the enterprise. But to make that more concrete, ServiceNow started twenty two years ago focused simply on automation. They turned physical paperwork into software workflows initially for the IT department within enterprises. That was it. And over time, they built on this platform going to more powerful and complex tasks.

他们从仅服务于IT部门扩展到人力资源、财务、客户服务、现场运营等其他部门。在过去二十年的过程中，ServiceNow已经完成了连接企业各个角落并实现自动化所需的所有繁琐基础工作。

They were expanding from serving just IT to other departments like HR, finance, customer service, field operations, and more. And in the process over the last two decades, ServiceNow has laid all the tedious groundwork necessary to connect every corner of the enterprise and enable automation to happen.

所以当人工智能出现时，从定义上讲，AI本质上就是高度复杂的任务自动化。而谁已经构建了支持这种自动化的平台和企业连接架构？正是ServiceNow。因此回答'ServiceNow如今做什么'这个问题时，他们声称'连接并赋能每个部门'绝非虚言。

So when AI arrived well, AI kinda just by definition is massively sophisticated task automation. And who had already built the platform and the connective tissue with enterprises to enable that automation? ServiceNow. So to answer the question, what does ServiceNow do today? We mean it when they say they connect and power every department.

IT和HR部门用它管理全公司的人员、设备和软件许可证；客户服务部门用它检测支付失败并内部路由到正确的团队或流程；供应链组织用它进行产能规划，整合其他部门的数据和计划确保协同一致。不再需要在不同应用间反复切换重复录入数据。最近ServiceNow还推出了AI助手，任何岗位的员工都能创建AI代理处理繁琐事务，让人力专注于宏观工作。

IT and HR use it to manage people, devices, software licenses across the company. Customer service uses ServiceNow for things like detecting payment failures and routing to the right team or process internally to solve it. Or the supply chain org uses it for capacity planning, integrating with data and plans from other departments to ensure that everybody's on the same page. No more swivel chairing between apps to enter the same data multiple times in different places. And just recently, ServiceNow launched AI agents so that anyone working in any job can spin up an AI agent to handle the tedious stuff, freeing up humans for bigger picture work.

ServiceNow去年入选《财富》全球最受赞赏公司榜单和《快公司》最佳创新者工作场所，正是源于这一愿景。若您希望在业务各环节利用ServiceNow的规模与速度优势，请访问servicenow.com/acquired，只需提及是本和大卫推荐即可。

ServiceNow was named to Fortune's world's most admired companies list last year and Fast Company's best workplace for innovators last year, and it's because of this vision. If you wanna take advantage of the scale and speed of ServiceNow in every corner of your business, go to servicenow.com/acquired and just tell them that Ben and David sent you.

感谢ServiceNow。

Thanks, ServiceNow.