本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
你喜欢我的雄鹿队T恤吗？
You like my Bucks t shirt?
我超爱你的雄鹿队T恤。
I love your Bucks t shirt.
我第一次去是在两周前？当时我去Benchmark开会，那里的怀旧氛围简直难以置信。
I went for the first time, what, two weeks ago? When I was down for meeting at Benchmark, and the nostalgia in there was just unbelievable.
真不敢相信你之前没去过。我知道Jensen是Denny's的常客，但我觉得如果我们邀请他，他会愿意在雄鹿队见面的。
I can't believe you hadn't been before. I know Jensen is a Denny's guy, but I feel like he would meet us at Bucks if we asked him.
至少我们该想办法弄些NVIDIA的纪念品挂到雄鹿队的墙上。完全合适。就这么定了，我们干吧。
Or at the very least, we should figure out some NVIDIA memorabilia to get on the wall at Bucks. Totally. Fit right in. Alright. Let's do it.
干吧。真相在谁手里？是你吗？是你吗？是你吗？
Let's do it. Who got the truth? Is it you? Is it you? Is it you?
现在真相在谁手里？是你吗？是你吗？是你吗？让我坐下说。
Who got the truth now? Is it you? Is it you? Is it you? Sit me down.
直说吧。另一个
Say it straight. Another
欢迎收听《收购》第十三季第三集，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。
Welcome to season thirteen episode three of Acquired, the podcast about great technology companies and the stories and playbooks behind them. I'm Ben Gilbert.
我是大卫·罗森塔尔。
I'm David Rosenthal.
我们是您的主持人。今天我们要讲述一个本以为已经完结的故事——英伟达。但过去十八个月的发展如此疯狂，听众们，这值得单独做一整期节目。所以今天是我们关于英伟达的第三集，讲述AI革命的故事，我们如何走到今天，以及为什么这一切现在发生——从最底层的原子与硅开始。有件疯狂的事我查了文字记录来确认：
And we are your hosts. Today, we tell a story that we thought we had already finished, NVIDIA. But the last eighteen months have been so insane listeners that it warranted an entire episode on its own. So today is a part three for us with NVIDIA telling the story of the AI revolution, how we got here, and why it's happening now starting all the way down at the level of atoms and silicon. So here's something crazy that I did a transcript search on to see if it was true.
在我们2022年4月的节目中，从未提到过'生成式'这个词。变化就是如此之快。难以置信，太疯狂了。而全球AI浪潮的时机巧合得令人咋舌，且非常有利。
In our April 2022 episodes, we never once said the word generative. That is how fast things have changed. Unbelievable. Totally crazy. And the timing of all of this AI stuff in the world is unbelievably coincidental and, very favorable.
让我们回到十八个月前。整个2022年，我们目睹金融市场从公开股票到早期初创公司再到房地产，都因利率快速上升而暴跌。加密货币和Web3泡沫破裂，银行倒闭。似乎整个科技经济——可能连带很多其他领域——正步入漫长寒冬。包括英伟达。
So recall back to eighteen months ago. Throughout 2022, we all watched financial markets from public equities to early stage startups to real estate just fall off a cliff due to rapid rise in interest rates. The crypto and web three bubble burst, banks fail. It seemed like the whole tech economy and potentially a lot with it was heading into a long winter. Including NVIDIA.
包括英伟达，他们曾因误判需求而进行了大规模库存减记。
Including NVIDIA, who had that massive inventory write off for what they thought was over ordering.
是啊。哇，变化真大。
Yep. Wow. How things have changed.
没错。但到了2022年，就在一切看似最黯淡的时刻，一项突破性技术经过实验室多年研究终于实用化——基于革命性Transformer机器学习机制的大语言模型（LLM）横空出世。先是OpenAI的ChatGPT成为史上最快达到1亿活跃用户的应用，随后微软、谷歌等几乎所有公司迅速跟进。2022年11月，AI无疑迎来了它的'网景时刻'，甚至可能是'iPhone时刻'——时间会给出答案。
Yeah. But by the 2022, right when everything looked the absolute bleakest, a breakthrough technology finally became useful after years in research labs, large language models or LLMs built on the innovative transformer machine learning mechanism burst onto the scene. First, with OpenAI's ChatGPT, which became the fastest app in history to a 100,000,000 active users, and then quickly followed by Microsoft, Google, and seemingly every other company. In November 2022, AI definitely had its Netscape moment, and time will tell, but it may have even been its iPhone moment.
嗯，这确实是Jensen坚信的观点。
Well, that is definitely what Jensen believes.
是的。今天我们将深入探讨这项突破如何诞生、背后的关键人物，以及为什么这一切都建立在英伟达的硬件和软件基础之上。若想及时获取新剧集通知，请注册acquired.fm/email。您还将获得两项独家福利：一是下期节目线索，二是已发布剧集的后续更新内容。
Yep. Well, today, we will explore exactly how this breakthrough came to be, the individuals behind it, and, of course, why the entire thing has happened on top of NVIDIA's hardware and software. If you wanna make sure you know every time there's a new episode, go sign up at acquired.fm/email. You'll also get access to two things that we aren't putting anywhere else. One, a clue as to what the next episode will be, and two, follow ups from previous episodes from things that we learned after release.
收听后可到acquired.fm/slack与我们讨论本期内容。想了解更多David和我的动态？欢迎关注我们的访谈节目a c q two。接下来几期我们将对话引领当今AI潮流的CEO们，还有与Doug DeMiro的精彩访谈——虽然本想和他聊的不止保时捷，但毕竟在Doug车库里只待了十一小时左右。
You can come talk about this episode with us after listening at acquired.fm/slack. If you want more of David and I, check out our interview show, a c q two. Our next few episodes are about AI with CEOs leading the way in this world we are talking about today. And a great interview with Doug DeMiro, where, we wanted to talk about a lot more than just Porsche with him. But, you know, we only had eleven hours or whatever we had in Doug's garage.
所以关于汽车行业的讨论，以及Doug的创业历程等丰富内容，我们都留在了a c q two节目里。快去收听吧！最后公告：很多人询问——我们也收到大量邮件——那些帽子何时补货？现在它们回来了。
So a lot of the, car industry chat and learning about Doug and his journey and his business, we saved for a c q two. So go check it out. One final announcement. Many of you have been wondering, and we've been getting a lot of emails, when will those hats be back in stock? Well, they're back.
限时供应！您可在acquired.fm/store购买ACQ刺绣帽。趁它们再次被永久封存进迪士尼宝库前，赶快下单吧。
For a limited time, you can get an ACQ embroidered hat at acquired.fm/store. Go put your order in before they, go back into the Disney vault forever.
这太棒了。我终于能给珍妮买一个属于她自己的，这样她就不会再偷我的了。
This is great. I can finally get Jenny one of her own so she stops stealing mine.
是的。那么，闲话少说，本节目不构成投资建议。大卫和我可能在讨论的公司中有投资，本节目仅供信息参考和娱乐目的。大卫，历史和事实。
Yes. Well, without further ado, this show is not investment advice. David and I may have investments in the companies we discuss, and this show is for informational and entertainment purposes only. David, history and facts.
哦，天啊。一方面，我们只有十八个月的时间可以讨论。
Oh, man. Well, on the one hand, we only have eighteen months to talk about.
但我知道你不会从十八个月前开始讲起。
Except that I know you're not gonna start eighteen months ago.
另一方面，我们有几十年的基础研究需要涵盖。所以当我开始研究时，我首先想到的是2022年4月的旧节目。我听完第二期结尾时，天啊，我都忘了这事。我想Jensen可能希望我们都忘了这事——在2021年英伟达的某张财报幻灯片上。
On the other hand, we have decades and decades of foundational research to cover. So when I was starting my research, I went to the natural first place, which was our old episodes from April 2022. And I was listening to them, and I got to the end of the second one. And, man, I had forgotten about this. I think Jensen maybe wishes we all had forgotten about this in one of NVIDIA's earnings slides in 2021.
他们展示了总可寻址市场，声称有1万亿美元的TAM。计算方式是：他们服务的客户所在行业总价值100万亿美元，而他们只需占据其中的1%。幻灯片上还有些相当推测性的内容，比如自动驾驶汽车和Omniverse，我记得机器人技术占了很大比重。
They put up their total addressable market, and they said they had a $1,000,000,000,000 TAM. And the way that they calculated this was that they were gonna serve customers who provided a $100,000,000,000,000 worth of industry, and they were gonna capture just 1% of it. And there were some stuff on the slide that was fairly speculative, you know, like autonomous vehicles and the Omniverse, and I think robotics were a big part of it.
他们的论点基本是：汽车加工厂加所有这些加起来有100万亿美元，我们只需拿下1%，因为他们的计算能力肯定能占到1%。我不是说这不对，但这种市场分析方式确实很粗糙。
And the argument is basically like, well, cars plus factories plus all these things added together is a 100,000,000,000,000, and we can just take 1% of that because surely their compute will amount to 1% of that, which I'm not arguing is wrong, but it is a very blunt way to analyze that market.
是啊。通常这不是创业的正确打开方式。懂吗？就那种‘我们只要拿下这个大市场的1%就够了’之类的空话。
Yeah. It's usually not the right way to, think about starting a startup. You know? Oh, if we can just get 1% of this big market, blah blah blah.
这是我所能想到的最自上而下的市场评估方式了。
It's the topiest down way I can think of to size a market.
所以Ben你在英伟达专题第二集的结尾犀利地指出了这点，你说要证明英伟达当前估值合理，实际上得相信所有这些——自动驾驶汽车、机器人技术等等——都会很快实现。
So you, Ben, rightly so called this out at the end of NVIDIA part two, and you're like, you know, I think to justify where NVIDIA is trading at the moment, you kinda actually gotta believe that all of this is gonna happen and happen soon. Autonomous cars, robotics, everything.
没错。关键是我觉得他们当时要达到那种估值，唯一的途径就是为现实世界中所有这些硬件提供算力支持。
Yeah. Importantly, I felt like the way for them to become worth what they were worth at that time literally had to be to power all of this hardware in the physical world.
对。我简直不敢相信自己说过这话，因为那既非本意也缺乏依据。我当时只是为了帮你唱反调而胡乱找理由。我们那期节目大部分时间都在讨论英伟达驱动的机器学习如何催生了极具价值的应用——社交媒体信息流推荐系统。正是这些推荐让Facebook和Google成长为互联网巨头，而英伟达支撑着整个体系。所以我当时随口说了句‘或许...’
Yep. I kinda can't believe that I said this because it was unintentional and uninformed, but I was kinda grasping at straws trying to play devil's advocate for you. And we just spent most of that whole episode talking about how machine learning powered by NVIDIA ended up having this incredibly valuable use case, which was powering social media feed recommenders. And that Facebook and Google had grown bigger than anyone ever imagined on the Internet with those feed recommendations, and NVIDIA was powering all of it. And so I just sorta idly proposed, well, maybe.
但要是其实不需要相信那些也能认为英伟达值万亿美元呢？万一互联网、软件和数字世界会持续扩张，而英伟达能为新的基础层提供算力呢？有可能吗？记得我们当时都表示‘嗯...说不准’。
But what if you don't actually need to believe any of that to still think that NVIDIA could be worth a trillion dollars? What if maybe, just maybe, the Internet and software and the digital world are gonna keep growing, and there will be a new foundational layer that NVIDIA can power? Is that possible? And I think we were both like, yeah. I don't know.
这期节目就到这里吧。
Let's end the episode.
好的。当然。我们没当回事，就觉得，行吧。特殊情况处理。
Yeah. Sure. We shrugged it off, we were like, alright. Carve outs.
但疯狂的是，至少在这个时间框架内，黄仁勋那张万亿市场规模幻灯片上的大多数预测都未实现。但那个疯狂的问题可能真的成真了。从英伟达的收入和利润来看，确实如此。简直不可思议。
But the crazy thing is that, of course, at least in this time frame, most things on Jensen's trillion dollar TAM slide have not come to pass. But that crazy question just might have come to pass. And from NVIDIA's revenue and earnings standpoint, definitely has. It's just wild.
好吧。那我们是怎么走到这一步的？
Alright. So how did we get here?
让我们倒带讲讲这个故事。2012年，人工智能迎来了大爆炸时刻——或者用当时更谦虚的说法叫机器学习。那就是AlexNet。我们上期节目详细讨论过，这是多伦多大学三位研究者向ImageNet计算机科学竞赛提交的算法。
Let's rewind and tell the story. So back in 2012, there was the big bang moment of artificial intelligence, or as it was more humbly referred to back then, machine learning. And that was AlexNet. We talked a lot about this on the last episode. It was three researchers from the University of Toronto who submitted the AlexNet algorithm to the ImageNet computer science competition.
ImageNet竞赛需要参赛者分析1400万张人工标注的图片，比如标注图片是草莓、猫、狗之类的。
Now, ImageNet was a competition where you would look at a set of 14,000,000 images that had been hand labeled with what the pictures were of, like of a strawberry or a cat or a dog or whatever.
大卫，你之前告诉我这是当时亚马逊土耳其机器人（Mechanical Turk）有史以来最大规模的应用——用来标注ImageNet数据集。
And, David, you were telling me it's the largest ever use of Mechanical Turk up to that point was to label the ImageNet dataset.
是啊，太疯狂了。要知道在这次竞赛和AlexNet出现之前，没有任何机器学习算法能准确标注图片。所以成千上万的人在土耳其机器人上每小时赚几美元来标注这些图片。
Yeah. It's wild. I mean, until this competition and until AlexNet, there was no machine learning algorithm that could accurately label images. So thousands of people on Mechanical Turk got paid however much, $2 an hour to label these images.
是的。如果我没记错我们那期节目的话，基本上情况就是AlexNet团队的表现远超以往任何人。完全是质的飞跃。我记得错误率从25%的图像误标率骤降到仅15%。这相对于之前缓慢的渐进式进步简直是巨大的跨越。
Yeah. And if I'm remembering from our episode, basically, happened is the AlexNet team did way better than anybody else had ever done. The complete step change better. I think the error rate went from mislabeling images 25% of the time to suddenly only mislabeling them 15% of the time. And that was like a huge leap over the tiny incremental progress that had been made along the way.
你说得完全正确。他们的方法彻底改变了互联网、谷歌、Facebook的命运，当然还有英伟达——他们实际上使用了古老的算法。这个计算机科学和人工智能分支叫神经网络，特别是卷积神经网络，早在六十年代就存在，但训练过程计算量极大。所以没人认为实际训练和使用这些网络具有可行性，至少在我们有生之年不行。而多伦多这群人做的，就是跑到当地百思买或加拿大类似商店。
You are spot on. And the way that they did it and what completely changed the fortunes of the Internet, of Google, of Facebook, and certainly of NVIDIA was they actually used old algorithms. A branch of computer science and artificial intelligence called neural networks, specifically convolutional neural networks, which had been around since the sixties, but they were really computationally intensive to train. And so nobody thought it would be practical to actually train and use these things, at least not anytime soon or in our lifetimes. And what these guys from Toronto did is they went out probably to their local Best Buy or equivalent in Canada.
他们买了两块GeForce GTX 580显卡，当时最顶级的型号。然后用CUDA——英伟达的GPU软件开发平台——编写了他们的卷积神经网络算法。天啊，他们就用价值一千美元的消费级硬件训练出了这个模型。
They bought two GeForce GTX five eighties, which were the top of the line cards at the time. And they wrote their algorithm, their convolutional neural network in CUDA, in NVIDIA's software development platform for GPUs. And by god, they trained this thing on, like, a thousand dollars worth of consumer grade hardware.
本质上，其他人多年来尝试的算法都不具备显卡那种大规模并行能力。所以如果你能真正发挥显卡的全部算力，或许就能运行某些独特的新算法，用超级计算机实验室所需时间和成本的一小部分完成训练。
And, basically, the algorithm that other people had been trying over the years just wasn't massively parallel the way that a graphics card sort of enables. So if you actually can consume the full compute of a graphics card, then perhaps you could run some unique novel algorithm and do it on, you know, a fraction of the time and expense that it would take in these supercomputer laboratories.
没错。之前所有人都试图在CPU上运行这些模型。CPU很棒，但一次只能执行一条指令。而GPU能同时执行数百甚至数千条指令。所以GPU、英伟达显卡、加速计算——黄仁勋和公司喜欢这么称呼——你可以把它想象成巨大的阿基米德杠杆。
Yeah. Everybody before was trying to run these things on CPUs. CPUs are awesome, but they only execute one instruction at a time. GPUs, on the other hand, execute hundreds or thousands of instructions at a time. So GPUs, NVIDIA graphics cards, accelerated computing, what Jensen and the company likes to call this, you can really think of it like a giant Archimedes lever.
无论摩尔定律如何发展，芯片上晶体管数量如何增加，只要你的算法能并行运行（虽然并非所有问题都适用，但很多可以），你就能将摩尔定律的效果放大数百倍、数千倍，如今甚至是数万倍，以远超常规方式的速度执行运算。
Whatever advances are happening in Moore's Law and the number of transistors on a chip, if you have an algorithm that can run-in parallel, which is not all problem spaces, but many can, then you can basically lever up Moore's Law by hundreds of times or thousands of times or today, tens of thousands of times and execute something a lot faster than you otherwise could.
最有趣的是最初存在一个天然并行的市场——图形处理，屏幕上每个像素都不依赖于相邻像素，完全可以独立计算后输出。所以屏幕上数万乃至如今数十万像素都能并行处理。当时英伟达没意识到的是，AI、加密货币等基于线性代数和矩阵运算的领域，都将成为加速计算的新前沿——把任务从CPU转移到GPU和其他并行处理器上，这恰恰是他们为图形处理开发的技术的新应用场景。
And it's so interesting that there was this first market called graphics that was obviously parallel, where every pixel on a screen is not sequentially dependent on the pixel next to it. It literally can be computed independently and output to the screen. So you have however many tens of thousands or now hundreds of thousands of pixels on a screen that can all actually be done in parallel. And little did NVIDIA realize, of course, that AI and crypto and all this other linear algebra, matrix math based things that turned into accelerated computing, pulling things off the CPU and putting them on GPU and other parallel processors, was an entire new frontier of other applications that could use the very same technology they had pioneered for graphics.
是的，这些内容相当实用。Alex Knapp的这一刻和多伦多的三位研究者开创了新局面。Jensen称之为，而且他说得完全正确，这是AI的大爆炸时刻。
Yeah. It was pretty useful stuff. And this Alex Knapp moment and these three researchers from Toronto kicked off. You know, Jensen calls it, and he's absolutely right. The big bang moment for AI.
David，上次我们完整讲述这个故事时，提到了多伦多的这个团队。但我们没有跟进这三人团队后来做了什么。
So, David, the last time we told this story in full, we talked about this team from Toronto. We did not follow what this team of three went on to do afterwards.
没错。基本上我们说的是，这些人工作的自然结果就是——哦，实际上你可以用这个技术来推送社交媒体信息流中的下一条内容，比如Instagram或YouTube的信息流之类。这释放了数十亿的价值。这些人和该领域的所有其他研究者，全都被谷歌和脸书网罗了。嗯，确实如此。
Yeah. So, basically, what we said was it turned out that a natural consequence of what these guys were doing was, oh, actually, you can use this to surface the next post in a social media feed on, like, an Instagram feed or the YouTube feed or something like that. And that unlocked billions and billions of value. And those guys and everybody else working in the field, they all got scooped up by Google and Facebook. Well, that's true.
于是，谷歌和脸书开始大量购买英伟达的GPU。但事实证明这个故事还有另一个被我们完全忽略的篇章，它始于你提出的问题，Ben。这些人是谁？是的。组成AlexNet团队的三个人分别是博士生Alex Krusevskiy，他的导师是传奇计算机科学教授Jeff Hinton。
And then as a consequence of that, Google and Facebook started buying a lot of NVIDIA GPUs. But turns out there's also another chapter to that story that we completely skipped over, and it starts with the question you asked, Ben. Who are these people? Yes. So the three people who made up the AlexNet team were, of course, Alex Krusevskiy, who was a PhD student, under his faculty adviser, the legendary computer science professor, Jeff Hinton.
关于Jeff Hinton有个惊人的冷知识。你知道他的曾曾祖父母是谁吗？
I have an amazing piece of trivia about Jeff Hinton. Do you know who his great, great grandparents were?
不知道，完全没概念。
No. I have no idea.
他是George和Mary Boole的曾曾孙。就是布尔代数和布尔逻辑的那个布尔。
He is the great, great grandson of George and Mary Boole. You know, like, Boolean algebra and Boolean logic?
这家伙天生就是做计算机科学研究的料。天啊。
This guy was born to be a computer science researcher. Oh my god.
对吧？这些都是计算和计算机科学的基础内容。
Right? Foundational stuff for computation and computer science.
我也不知道有人叫布尔，但这就是它的来源。太搞笑了。
I also didn't know there were people named bool, but that's where that came from. That's hilarious.
是啊。你知道那些与、或、异或、或非运算符吗？这些都源自乔治和玛丽。太疯狂了。所以他是系里的导师。
Yeah. You know, the and, or, x, or, nor operators. That comes from George and Mary. Wild. So he's the faculty adviser.
团队里还有第三个人，Alex在实验室的博士生同学Ilya Sutskever。如果你知道我们接下来要讲什么，你现在可能已经在座位上激动得跳起来了。Ilya是OpenAI的联合创始人兼现任首席科学家。没错。AlexNet之后，Alex、Jeff和Ilya做了件非常自然的事。
And then there was a third person on the team, Alex's fellow PhD student in this lab, when Ilya Sutskever. And if you know where we're going with this, you are probably jumping up and down right now in your seat. Ilya is the cofounder and current chief scientist of OpenAI. Yes. So after AlexNet, Alex, Jeff, and Ilya do the very natural thing.
他们开了家公司。我不知道他们在公司具体做什么，但成立公司是顺理成章的事。
They start a company. I don't know what they were doing in the company, but, it made sense to start one.
不管他们做了什么，公司很快就会被收购。
And whatever they did, it was gonna get acquired real fast.
六个月内被谷歌收入麾下。他们就这样被谷歌招致麾下，加入了一群被谷歌在该领域实质垄断的学者和研究人员行列。具体有三位：格雷格·科拉多、杰夫·迪恩和著名斯坦福教授吴恩达。这三人刚在谷歌内部组建了谷歌大脑团队，旨在全面推动由AlexNet开启的AI研究工作。
By Google within six months. So they get scooped up by Google. They join a bunch of other academics and researchers that Google has been monopolizing, really, in the field. Three specifically, Greg Corrado, Jeff Dean, and Andrew Ng, the famous Stanford professor. The three of them had just formed the Google Brain team within Google to turbocharge all of this AI work that has been unleashed by AlexNet.
当然，最终目的是为谷歌创造巨额利润。
And, of course, to turn it into huge amounts of profit for Google.
事实证明，通过Facebook或谷歌在互联网上精准投放个性化广告
Turns out individually serving advertising that's perfectly targeted on the Internet through Facebook or Google
或者YouTube。
Or YouTube.
是门利润极其丰厚的生意，而且消耗了大量英伟达GPU。
Is an enormously profitable business and one that consumes a whole lot of NVIDIA GPUs.
没错。大约一年后，谷歌还收购了著名的DeepMind。几乎同时，Facebook挖走了计算机科学教授杨立昆——他也是该领域的传奇人物。这两家公司基本形成了对顶尖AI研究人才的双头垄断。不过当时，没人会误以为这些公司和研究者正在开发真正人类级别或接近人类水平的智能。
Yes. So about a year later, Google also acquires DeepMind famously. And then right around the same time, Facebook scoops up computer science professor, Jan Lecun, who also is a legend in the field. And the two of them basically establish a duopoly on leading AI researchers. Now at this point, nobody is mistaking what these companies and these people are doing for true human level intelligence or anything close to it.
这种AI只擅长特定任务，比如我们讨论过的社交媒体信息流推荐。谷歌大脑团队与杰夫、亚历克斯、伊利亚合作的重要项目之一就是重构YouTube算法。正是这次改造，让YouTube从谷歌收购的亏损怪胎变成了今日的绝对霸主。我们做YouTube专题节目时（2013、2014年左右），距离这个转型期并不久。
This is AI that is very good at narrow tasks. Like, we talked about social media feed recommendations. So the Google Brain team and Jeff and Alex and Ilya, one of the big projects they work on is redoing the YouTube algorithm. And this is when YouTube goes from, like, money losing, you know, crazy thing that Google acquired to the just absolute juggernaut that it is today. I mean, back then in, like, 2013, 2014, we did our YouTube episode not that long after.
YouTube视频的大部分观看量来自其他网页的嵌入播放。这是当他们将其整合到社交媒体网站时发生的。他们开始推送信息流，启动自动播放功能。所有这些都源自人工智能研究。
The majority of views of YouTube videos were embeds on other web pages. This is when they build it into a social media site. They start the feed. They start autoplay. All this stuff is coming out of AI research.
谷歌收购DeepMind后的一些著名成果。DeepMind开发了一系列算法来节省冷却成本。而Facebook，当然，他们可能是这一代最后的赢家，因为他们利用了所有这些成果，加上杨立昆（Jan Lecun）的研究和大量研究人员的招聘。这距离他们收购Instagram才过去几年。天啊，我们真该重新回顾那段历史，因为Instagram无论如何都是一笔绝佳的收购。
Some of the other stuff that happens at Google, famously after they acquired DeepMind. DeepMind built a bunch of algorithms to save on cooling costs. And Facebook, of course, they probably had the last laugh in this generation because they're using all this work, and Jan Lecun is doing his thing and hiring lots of researchers there. This is just a couple years after they acquired Instagram. Man, we need to, like, go back and redo that episode because Instagram would have been a great acquisition anyway.
但正是信息流中由AI驱动的推荐系统，让Instagram成为Facebook价值1000亿美元的资产。
But it was AI powered recommendations in the feed that made that into a $102,105 $100,000,000,000 asset for Facebook.
我认为你并没有夸大其词。这确实是Instagram现在对Meta的实际价值。顺便说一句，我在Instagram广告上买了很多东西，因为他们的定向投放确实有效。
And I don't think you're exaggerating. I think that is literally what Instagram is worth to Meta now. By the way, I have bought a lot of things on Instagram ads so that the targeting works.
绝对有效。Astro Teller（当时至今仍负责Google X）在《纽约时报》的文章中有句精彩的话：'我不认为这甚至包括DeepMind的贡献。仅Google Brain团队为谷歌带来的利润增长，就足以覆盖Google X的所有开支。'
It absolutely does. There's this amazing quote from Astro Teller who ran Google x at the time and still does in a New York Times piece where he says that the gains from Google Brain during this period I don't think this even includes DeepMind. Just the gains from the Google Brain team alone in terms of profits to Google more than funded everything they were doing in Google X.
Google X有产出过任何盈利项目吗？
Which has there ever been anything profitable out of Google X?
Google Brain。
Google Brain.
是的，我是说没错。
Yeah. I mean yeah.
我们就此打住。这让我们回到2015年，当时硅谷的一些人开始意识到，谷歌和Facebook在人工智能领域的双头垄断实际上是个非常、非常严重的问题。大多数人对此毫无察觉。这两个人的预见性确实很强。
We'll leave it at that. So this takes us to 2015, when a few people in Silicon Valley start to realize that this Google, Facebook, AI duopoly is actually a really, really big problem. And most people had no idea about this. This is really visionary of these two people.
而且不仅仅是对其他大型科技公司来说是个问题——你可以提出这样的论点。问题在于，比如Siri表现糟糕。当时所有拥有大量消费者接触点的公司，其人工智能技术都相当落后。但担忧的原因远不止于此。
And not just a problem for, like, the other big tech companies, because you could make the argument. It's a problem because, like, Siri's terrible. All the other companies that have lots of consumer touch points have pretty bad AI at the time. But the concern is for a much greater reason.
我认为这里存在三个层面的担忧。第一显然是其他科技公司。其次是初创企业面临的问题。这对初创企业非常不利。当人工智能成为这代技术的主要价值驱动力时，你如何与谷歌和Facebook竞争？
I think there are three levels of concern here. One, obviously, is the other tech companies. Then there's the problem of startups. This is terrible for startups. How are you gonna compete with Google and Facebook when this is the primary value driver of this generation of technology?
我是说，确实可以从另一个角度看待Snap和Musically的遭遇——它们不得不卖给字节跳动变成TikTok，最终归入中国公司旗下。也许是商业决策问题，也许是执行不力或其他原因阻碍了这些平台实现独立规模。Snap现在是上市公司，但它远不如Facebook。或许是因为它们无法获得与Facebook和谷歌同等水平的人工智能研究资源。
I mean, there really is another lens to view what happened with Snap, what happened with Musically, and having to sell themselves to ByteDance and becoming TikTok and going to the Chinese. Maybe it was business decisions. Maybe it was execution or whatever that prevented those platforms from getting to independent scale. Snap's a public company now, but, like, it's no Facebook. Maybe it was that they didn't have access to the same AI researches that Facebook and Google had.
这倒是个有趣的问题。虽然结论可能多跳了几步，但不失为一个好玩的假想敌来思考。
That feels like an interesting question. It's probably a couple steps too far in the conclusion, but still sort of a fun straw man to think about.
一个有趣的假想敌。但无论如何，这绝对是个问题。第三层问题在于——这对整个世界都很糟糕，这么多人才都被锁在谷歌和Facebook里。
A fun straw man. Nonetheless, this is definitely a problem. The third layer of the problem is just, like, this sucks for the world that all these people are locked up in Google and Facebook.
现在或许是个合适的时机说明一下。OpenAI的创立动机源于希望在大科技公司之前率先发现AGI（人工通用智能）。DeepMind也是如此。当时这注定是一条曲折迂回的道路，因为无论是当时还是现在，都没人知道通往AGI的最佳路径。但OpenAI创立时的核心理念是：无论谁先发现AGI，都将迅速变得无比庞大和强大，掌握巨大控制权——而这一切最好在开放的环境中进行。
This is probably a good time to mention. This founding of OpenAI was motivated by the desire to find AGI or artificial general intelligence first before the big tech companies did. And DeepMind was the same thing. It was gonna be this winding and circuitous path at the time since really nobody knew then or knows now the best path to get to AGI. But the big idea at OpenAI's founding was whoever figures out and finds AGI first will be so big and so powerful so quickly, they'll have an immense amount of control, and that is best in the, open.
于是这两个人
So these two people
对此深感忧虑的两人在2015年召集了一场至关重要的晚宴——地点偏偏选在了
who are quite concerned about this convene a very fateful dinner in 2015 at, of all places
是瑰丽酒店吗？
Is it the Rosewood?
沙丘路瑰丽酒店。当然了。如果选在丹尼斯餐厅或伍德赛德的巴克斯这类地方会更有意思。
The Rosewood Hotel on Sand Hill Road. Naturally. It would have been way better if it were a Denny's or Bucks and Woodside or something like that.
但这确实表明了OpenAI的起源。它与英伟达这类公司从底层草根奋斗起家的模式截然不同。明白吗？这是来自权力高层和既有资本的声音在说：不，我们需要凭空创造出一个新事物。
But it does actually just show, like, where the seeds of OpenAI come from. It is very different than this sort of organic scrappy way that the Nvidias of the world got started. You know? This is powers on high and existing money saying, no. We need to will something into existence.
没错。
Yep.
所以，那两个神秘人物自然是埃隆·马斯克和时任Y Combinator总裁的山姆·奥特曼。他们共进晚餐时，几乎邀请了谷歌和脸书所有顶尖的AI研究员。他们直接发问：要怎样才能让你们离开，打破这两家公司的垄断？几乎所有人的回答都是：没可能。
So, of course, those two shadowy figures are Elon Musk and Sam Altman, who at the time was president of Y Combinator. So they get this dinner together, and they invite basically all of the top AI researchers at Google and Facebook. And they're like, yo, what is it gonna take for you to leave and to break this duopoly? And the answer from almost all of them is nothing. You can't.
我们为什么要离开？在这里我们快乐得像蛤蜊一样。
Why would we ever leave? We're happy as clams here.
我们能招到想要的人才，组建了顶尖团队，还有源源不断的资金支持。
We've gotten to hire the people that we want. We've built these great teams. There's a money spigot pointed at our face.
没错。我们不仅拿着天文数字的薪水，还能直接与业界最优秀的AI研究者共事。如果留在学术界——比如华盛顿大学（计算机科学顶尖学府）或多伦多大学（这些人的母校）——你仍处于分散的市场。但进入谷歌或脸书，你就与整个领域最优秀的人为伍。
Right. Not only are we getting paid just ungodly amounts of money, but we get to work directly with the best AI researchers in the field. If we were still at academic institutions, you know, say you're at the University of Washington, amazing academic institution for computer science, one of the top in the world, or the University of Toronto where these guys came from, you're still at a fragmented market. If you go to Google or you go to Facebook, you're with everybody. Yep.
所以几乎所有人都拒绝了，只有一人对埃隆和山姆的提议动了心。引用凯特·梅茨当时在《连线》杂志的精彩报道（我们会在资料来源中附链接）：'问题在于，能解决这些AI难题的人才大多已被谷歌和脸书收入麾下，晚宴上没人确信能说服这些思想家加入初创企业——即便有马斯克和奥特曼背书。但至少有一位关键人物对跳槽持开放态度。'报道引述这位关键人物的话：'我知道存在风险，但尝试这件事会非常有趣。'这位关键人物就是伊利亚·苏茨克沃。
So the answer is no from basically everybody, except there's one person who's intrigued by Elon and Sam's pitch. And to quote an amazing Wired article from the time by Kate Metz that we will link to in our sources, Quote, the trouble was so many of the people most qualified to solve all these AI problems were already working for Google and Facebook, and no one at the dinner was quite sure that these thinkers could be lured to a new startup even if Musk and Altman were behind it. But one key player was at least open to the idea of jumping ship, and then they have a quote from that key player. I felt there were risks involved, but I also felt it would be a very interesting thing to try. And that key player was Ilya Sutskever.
没错。晚餐会后，伊利亚离开谷歌，成为埃隆和山姆支持的新独立非营利AI研究实验室OpenAI的联合创始人兼首席科学家。
Yep. So after the dinner, Ilya leaves Google and signs up to become, as we said, cofounder and chief scientist of a new independent AI nonprofit research lab backed by Elon and Sam, OpenAI.
好了听众朋友们，现在正是感谢我们最爱的合作伙伴——已成为Acquired和Anthropic工作流程核心的Claude Sonnet 4.5最新突破性模型——的最佳时机。
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。当我们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？这种策略有多新颖？之前有其他公司尝试过吗？这类问题以及能够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确实能够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不仅仅是又一个模型。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的工程师称其为强大的思考伙伴，并表示它正在改变他们的开发速度。Canva在其部分产品中使用Claude，称其为重大飞跃。企业们对Sonnet 4.5爱不释手。
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.
有一点已经变得很清楚：让一个模型擅长编程，也会让它天生擅长任何分析性任务。因此，使Claude擅长重构代码库的特性，同样让它擅长梳理数千份监管文件或进行复杂的财务分析。Claude通过Anthropic的API与企业现有工作流无缝集成，并新增了记忆和上下文管理功能，使智能体能够长时间运行而不丢失关键信息。
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.
因此，无论你是在扩展工程团队，还是构建下一代智能应用，Claude都会与你共同思考复杂问题，而不仅仅是替你思考。它确实是你的智能思考伙伴。
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.
现在访问claude.ai/acquired即可免费试用Claude，并享受Claude Pro三个月五折优惠。如果你想了解他们的企业服务，只要告诉他们是本和大卫推荐你的。
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.
好的。大卫，OpenAI成立了。那是2015年。八年后我们有了ChatGPT。从那时到现在看似是条超级直线路径。
Okay. So, David, OpenAI is formed. It's 2015. Here we are eight years later, and we have ChatGPT. Super linear path from there to here.
对吧？结果发现并非如此。
Right? Turns out, no.
正如我们刚才谈到的，当前阶段的AI在特定应用场景下表现优异，但与如今的GPT-4相比还相去甚远。当时它的能力相当有限，主要原因之一是可实际用于训练模型的数据量非常有限。以AlexNet为例，1400万张图片在互联网的宏大图景中不过是沧海一粟。
So as we were talking about a little bit, AI at this point in time, super good for narrow use cases, looks nothing like GPT four today. The capabilities that it had were pretty limited. And one of the big reasons was that the amount of data that you could practically train these models on was pretty limited. So the AlexNet example, you're talking about 14,000,000 images. In the grand scheme of the Internet, 14,000,000 images is a drop in the bucket.
这既是硬件也是软件的限制。在软件层面，我们当时根本没有算法能狂妄到考虑用整个互联网的数据训练一个基础模型——这种想法根本不存在。
And this was both a hardware and a software constraint. On the software side, we just didn't actually have the algorithms to sort of suppose that we could be so bold to train one single foundational model on the whole Internet. Like, it wasn't a thing.
没错，那在当时简直是个疯狂的想法。
Yeah. That was a crazy idea.
确实。人们虽然对语言模型的概念感到兴奋，但实际上我们并不清楚如何从算法上实现它。2015年，当时还在OpenAI的安德烈·卡帕西（后来领导特斯拉AI部门，如今又回到OpenAI）发表了那篇开创性的博客《神经网络不合理的有效性》。大卫，虽然这期节目我们不会深入探讨，但要注意循环神经网络与2012年论文中的卷积神经网络是有所不同的。
Right. People were excited about the concept of language models, but we actually didn't know how we could algorithmically get it done. So in 2015, Andre Karpathy, who was then at OpenAI and went on to lead AI for Tesla and is actually now back at OpenAI, writes this seminal blog post called the unreasonable effectiveness of neural networks. And, David, I don't think we're gonna go into it on this episode, but note that recurrent neural networks are a little bit of a different thing than convolutional neural networks, which was the 2012 paper.
技术前沿已经进化了。
The state of the art had evolved.
是的。大约在同一时期，2016年稍晚时候NVIDIA官方频道发布了一段1分45秒的短视频，里面出现了两位人物：年轻的伊利亚·苏茨克沃，以及安德烈·卡帕西。以下是安德烈在视频中的原话：'语言模型是我最看好的算法之一。'
Yes. And right around that same time, there is also a video that hits YouTube, right, a little bit later in 2016 that is actually on NVIDIA's channel and has two people in this very short one minute and forty five second video. One is a young Ilya Sitzkever, and two is Andre Karpathy. And here is a quote from Andre from that YouTube video. One algorithm I'm excited about is a language model.
这个理念是：你可以将海量数据输入网络，它会自动识别词语在句子中的排列规律。比如用人们在互联网上的对话数据，本质上可以训练出聊天机器人，但这种方式能让计算机理解语言运作机制和人际交互模式。终有一天，我们将能像人类交谈那样与计算机对话。哇。
The idea that you can take a large amount of data and you feed it into the network, and it figures out the pattern in how words follow each other in sentences. So for example, you could take a large amount of data on how people talk to each other on the Internet. You can train basically a chatbot, but you can do it in a way that the computer learns how language works and how people interact. Eventually, we'll use that to talk to computers just like we talk to each other. Wow.
这是2015年？那时距离Transformer模型问世还有两年，Karpathy正在OpenAI工作。他不仅提出了聊天机器人的构想或倡导这一理念，说明相关讨论早已存在。甚至在Transformer这种实现方法出现前，他就隐约意识到——这里有个关键点——模型能发现句子中词语的排列规律。
This is 2015? This is two years before the transformer while Karpathy is at OpenAI. He both comes up with the idea or espouses the idea of a chatbot, so that sort of had already been discussed. But even before we had the transformer, the method to actually pull this off, he sort of had the idea that and there's an important part here. It figures out the pattern in how words follow each other in sentences.
所以核心观点是：语言的基本结构和对知识的理解方式，其实就蕴含在训练数据本身中，而非依赖人工标注。
So there's this idea that the very structure of language and the way to interpret knowledge is actually embedded in the training data itself rather than requiring labeling.
这太酷了！今年春季GTC大会上，Jensen和Ilya进行了炉边对话。整场对话精彩绝伦，你真该去看看。其中就谈到这个问题。
This is so cool. So at Spring GTC this year, Jensen did a fireside chat with Ilya. And it's amazing. You should go watch the whole thing. But in it, this question comes up.
Jensen故意扮演反方角色说：嘿，有人认为GPT-3、GPT-4、ChatGPT这些大语言模型只是在概率预测句子中的下一个词，它们并不真正具备知识。而Ilya给出了绝妙回应。
Jensen kinda poses as a straw man. Like, hey. Some people say that GPT three, four, chat GPT, everything going on, all these LLMs, they're just probabilistically predicting the next word in a sentence. They don't actually have knowledge. And Ilya has this amazing response to that.
他说：好吧，想象一本侦探小说。在结尾处，侦探把所有人召集到房间宣布：我现在要说出凶手的名字，那个人就是____。LLM越准确地预测出这个空缺词——
He says, okay. Well, consider a detective novel. Yes. At the end of the novel, the detective gathers everyone together in a room and says, I am now going to tell you all the name of the person who committed the crime, and that person's name is blank. The more accurately an LLM predicts that next word, I.
也就是凶手姓名——就越证明它不仅理解小说内容，更掌握了人类层面的通用知识与智能。因为要猜中凶手，需要调动全部人生经验和世界认知。而当今的LLM（GPT-3、GPT-4、LAMA、Bard等）确实能推断出凶手。
E, the name of the criminal, ipso facto, the greater its understanding, not only of the novel, but of all general human level knowledge and intelligence. Because you need all of your experience in the world and as a human to be able to guess who the criminal is. And the LLMs that are out there today, GPT three, GPT four, LAMA, Bard, these others, they can guess who the criminal is.
噢没错！这个关于「理解vs预测」的议题要重点标记，正是当下的热点。David，现在是不是该快进到2017年讨论Transformer论文了？
Oh, yeah. Put a pin in that. Understanding versus predicting. It's a hot topic du jour. So, David, is now a good time to fast forward two years to 2017 to the transformer paper?
当然。本，给我们讲讲Transformer吧。
Absolutely. Ben, tell us about the transformer.
好的。2017年谷歌的Transformer论文问世，标题叫《注意力机制就是全部所需》。
Okay. So Google 2017 transformer paper. Paper comes out. It's called attention is all you need.
这是来自谷歌大脑团队的成果，对吧？
And it's from the Google brain team. Right?
是的。
Yes.
就是Ilya刚离开的那个团队。
That Ilya just left.
刚离开。两年前去创办OpenAI了。先铺垫一下背景：机器学习在自然语言处理领域长期用于自动纠错或外语翻译等功能。但2017年谷歌这篇论文提出了一种全新模型，彻底改变了这些领域并开辟了新方向。当时的情况是这样的。
Just left. Two years before to start OpenAI. So machine learning on natural language, just to set the table here, had long been used for things like autocorrect or foreign language translation. But in 2017, Google came out with this paper and discovered a new model that would change everything for these fields and unlock another one. So here is the scenario.
假设你要把英语句子翻译成法语。按顺序逐字翻译看似可行，但任何有过海外旅行经历的人都知道，不同语言的词序经常需要调整。所以这种翻译方式很糟糕。比如西班牙语中'美国'这个词，按逐字翻译从一开始就会出错。
You're translating a sentence from English to French. You could imagine that a way to do this would be one word at a time in order. But for anyone who's ever traveled abroad and tried to do this, you know that words are sometimes rearranged in different languages. So that's a terrible way to do it. You know, United States in Spanish is, failure on the very first word in that example.
那么引入注意力机制这个概念，它是这篇研究论文的关键部分。这个注意力机制，作为Transformer论文中相当神奇的组成部分，字面意思就是它听起来的那样。它是让模型在不同时刻关注输入文本不同区域的一种方式。在考虑翻译中选择下一个词时，你可以查看大量上下文。因此，对于你即将用法语输出的每个单词，你可以浏览整个输入单词集，以确定在决定下一步操作时应重点考虑哪些词。
So enter this concept of attention, which is a key part of this research paper. So this attention, this fairly magical component of the transformer paper, it literally is what it sounds like. It is a way for the model to attend to different areas of the input text at different times. You can look at a large amount of context while considering what word to pick next in your translation. So for every single word that you're about to output in French, you can look over the entire set of inputted words to figure out what words you should weight heavily in your decision for what to do next.
这就是为什么如果你拟人化地看待AI和机器学习，它们之前的适用性如此有限——就像是一个注意力持续时间非常非常短的人类。
This is why AI and machine learning was so narrowly applicable before if you anthropomorphize it and you think of it like a human, it was like a human with a very, very short attention span.
没错。现在到了神奇的部分。虽然它确实会查看整个输入文本来考虑下一个词应该是什么，但这并不意味着它完全抛弃了位置的概念。它使用了一种称为位置编码的技术，因此不会完全忘记单词的位置。所以它有这个很酷的特性：既能权衡与你特定单词相关的重要部分，又能理解位置信息。
Yes. Now here's the magical part. While it does look at the whole input text to consider what the next word should be, it doesn't mean that it throws away the notion of position entirely. It uses a technique called positional encoding, so it doesn't forget the position of the words altogether. So it's got this cool thing where it weights the important part relevant to your particular word, and it still understands position.
还记得我说过，注意力机制每次选择输出单词时都会查看整个输入。
So remember I said the attention mechanism looks over the entire input every time it's picking what word to output.
这听起来计算量非常大。
That sounds very computationally hard.
是的。用计算机科学的术语来说，这意味着注意力机制的时间复杂度是O(n²)。
Yes. In computer science terms, this means that the attention mechanism is o of n squared.
噢，这让我回想起了大学计算机入门课时的恐惧感。
Oh, that's given me the heebie teebies back to my intro CS classes in college.
哦，等我们看完这集就知道了。剧情会更深入。所以，显然是的。传统上，你会说这非常非常低效，实际上意味着你的上下文窗口（即令牌限制，即提示长度）越大，计算成本就会以二次方增长。所以输入翻倍意味着计算输出的成本翻四倍，输入翻三倍则意味着成本翻九倍。
Oh, just wait till we get through this episode. It gets deeper. So, obviously, yes. Traditionally, you'd say this is very, very inefficient, and it actually means that the larger your context window, aka token limit, aka prompt length gets, the more computationally expensive it gets on a quadratic basis. So doubling your input means quadrupling the cost to compute an output, or tripling your input means nine times
成本。这变得非常棘手。
the cost. It gets real gnarly.
没错。成本会迅速飙升。但GPU来拯救我们了。对我们来说好消息是，这些Transformer比较可以并行完成。所以尽管有很多比较要做，如果你有带大量核心的大GPU芯片，你可以同时完成所有比较。
Yeah. It gets real expensive real fast. But GPUs to the rescue. The amazing news for us here is that these transformer comparisons can be done in parallel. So even though there are lots of them to do, if you have big GPU chips with tons of cores, you can do them all at exactly the same time.
而之前实现这一点的技术，比如循环神经网络或LSTM（长短期记忆网络，一种循环神经网络）等等，这些都需要在开始下一步之前知道每一步的输出，在你选择下一个词之前。换句话说，它们是顺序的，因为它们依赖于前一个词。现在有了Transformer，即使你输入的文本长达一千个词，只要那个大GPU有足够多的核心，它也能在可测量的时间内完成，就像只有十个词一样快。所以这里的重大创新是，你现在可以并行训练基于序列的模型。
And previous technologies to accomplish this, like recurrent neural networks or LSTMs, long short term memory networks, which is a type of recurrent neural network, etcetera. Those required knowing the output of each step before beginning the next one, before you picked the next word. So in other words, they were sequential since they dependent on the previous word. Now with transformers, even if your string of text that you're inputting is a thousand words long, it can happen just as quickly in humid measurable time as if it were 10 words long, supposing that there were enough cores in that big GPU. So the big innovation here is you could now train sequence based models in a parallel way.
以前根本无法训练这种规模的模型，更不用说经济高效地训练了。
You couldn't train models of this size at all before, let alone cost effectively.
是的。这太重要了，可能对所有听众来说，开始听起来非常像我们今天生活的世界。
Yeah. This is huge, and probably for all listeners out there starting to sound very familiar to the world that we live in today.
没错。我刚才有点偷换概念，从翻译转到了使用上下文窗口和令牌长度这样的词。你大概能看出这是要往哪里发展。
Yeah. I sort of did a slight of hand there morphing translation to using words like context window and token length. You can kinda see where this is going.
是的。这篇关于Transformer的论文在2017年发表，意义重大。但不知为何，外界有段时间并未完全意识到这一点。当然，谷歌很清楚它的重要性。
Yep. So this transformer paper comes out in 2017. The significance is huge. But for whatever reason, there's a window of time where the rest of the world doesn't quite realize it. So Google obviously knows how important this is.
大约有一年时间，谷歌的AI研究（尽管Ilya已离职且OpenAI已成立）再次领先于该领域的其他所有人。那时谷歌在Gmail推出了智能撰写功能，还展示了让AI机器人代用户拨打本地商户电话的技术——还记得他们在IO大会上演示的那个功能吗？
And there's like a year where Google's AI work, even though Ilya has left and OpenAI is a thing now, accelerates again beyond anybody else in the field. So this is when Google comes out with Smart Compose in Gmail, and they do that thing where they have an AI bot that'll call local businesses for you. Remember that demo from IO that they did?
那个功能最终上线了吗？
Did that ever ship?
我不确定。也许上线了。毕竟这可是谷歌，技术实力摆在那里。
I don't know. Maybe it did. Maybe. I mean, this is Google here. Like, the capabilities are there.
产品表现也印证了这点。那时他们开始大力投资Waymo。但归根结底，真正的应用还是回归到搜索广告和YouTube视频推荐——这段时间他们简直势不可挡。而OpenAI等机构当时还未采用Transformer架构。
The product sends about as much. This is when they really start investing in Waymo. But, again, where it really manifests is just back to serving ads in search and recommending YouTube videos. Like, they're just crushing it in this period of time. OpenAI and everyone else, though, they haven't adopted transformers yet.
他们某种程度上还停留在过去，仍在进行那些偏研究性质的计算机视觉项目。比如那时他们开发了玩Dota2（《刀塔2》）游戏的AI。
They're kinda stuck in the past, and they're still doing these really research y computer vision projects. So, like, this is when they build a bot to play Dota two, defense of the agents two, the video game.
那些成果确实令人惊叹。那个AI仅通过计算机视觉——即分析屏幕截图进行推理——就击败了世界顶级Dota玩家。这非常困难，因为Dota2不是能一览全局的游戏，AI必须基于单名玩家的视角信息智能构建整个战局。这绝对是当时最前沿的研究。
And super impressive stuff. Like, they beat the best Dota players in the world at Dota by literally just consuming computer vision, like consuming screenshots and inferring from there. And that's a really hard problem because DOTA two is not a game where you get to see the whole board at once. So it has to do a lot of, like, really intelligent construction of the rest of the game based on just a single player's worth of input. So it's unbelievably cutting edge research.
在过去一代人看来，它基本上就是个更快的马。
For the past generation. It's a faster horse, basically.
也许吧。是的，我是说，他们还在做一些事情，比如'宇宙'项目——那是个用三维建模世界来训练自动驾驶汽车的系统。现在你基本听不到相关消息了，但他们当时构建了整个体系。我记得是借用《侠盗猎车手》作为虚拟环境，然后利用GTA世界进行汽车的计算机视觉训练。
Maybe. Yeah. I mean, they were also doing stuff like, universe, which was the three d modeled world to train self driving cars. You don't really hear anything about that anymore, but they built this whole thing. I think it was using Grand Theft Auto as the environment, and then it was doing computer vision training for cars using the GTA world.
我是说，那确实很疯狂，但有点东一榔头西一棒子。
I mean, it was crazy stuff, but it was kinda scattershot.
没错，确实很分散。我想说的是，当时仍局限在特定应用场景里。那个时间节点上他们根本没做任何接近GPT的事情。而谷歌那时候已经转向其他方向了。
Yeah. It was scattershot. And I guess what I'm saying is it was still in this narrow use case world. They weren't doing anything approaching GBT at this point in time. Meanwhile, Google had kinda moved on.
是的。现在我想为OpenAI及当时该领域的其他所有人辩护一点，他们并非闭目塞听。要实现Transformer所赋予的能力——本，你稍后会谈到这点——需要巨大的计算资源成本。GPU、NVIDIA和Transformer使之成为可能。但要处理你所提及的那种规模的大模型，其花费对非营利机构乃至除谷歌外的任何机构来说都难以为继。
Yep. Now one thing I do wanna say in defense of OpenAI and everybody else in the field at the time, they didn't just have their heads in the sand. To do what transformers enabled you to do, which, Ben, you're gonna talk about in a sec, cost a lot in computing power. GPUs and NVIDIA and the transformer made it possible. But to work with the size of models you're talking about, you're talking about spending an amount of money that's certainly for a nonprofit, and anybody really except Google was untenable.
没错。有意思的是，大卫，你直接跳到了昂贵的大模型这个话题。我们之前所做的不过是讨论如何将一句话翻译成另一句话。Transformer的应用并不必然要求你消耗整个互联网资源去构建一个基础模型。
Right. It's funny, David. You made this leap to expensive and large models. All we were doing before was merely talking about translating one sentence to another. The application of a transformer does not necessarily require you to go and consume the whole Internet and create a foundational model.
但让我们聊聊这个。正如我们现在所知，Transformer非常适用于另一类任务。对于给定的输入句子，除了翻译成目标语言外，它们还能作为下一个词预测器，判断序列中后续该出现什么词。你甚至可以通过在文本语料库上进行预训练，帮助模型理解该如何预测下一个词。稍微回溯一下，让我们回到Transformer之前最先进的循环神经网络。
But let's talk about this. Transformers lend themselves quite well as we now know to a different type of task. So for a given input sentence, instead of translating to a target language, they can also be used as next word predictors to figure out what word should come next to the sequence. You could even do this idea of pre training with some corpus of text to help the model understand how it should go about predicting that next word. So backing up a little bit, let's go back to the recurrent neural networks, the state of the art before transformers.
除了它们是串行而非并行的问题外，它们还存在另一个问题——上下文窗口非常短。虽然可以做下一个词预测，但由于它只能记住前几个词的内容，到段落结尾时就会忘记开头的内容，无法同时保留所有信息，所以这种预测作用有限。
Well, they had this problem in addition to the fact that they were sequential rather than parallel. They also had a very short context window. So you could do a next word predictor, but it wasn't that useful because it didn't know what you were saying more than a few words ago. By the time you'd get to the end of the paragraph, it would forget what was happening at the beginning. It couldn't sort of hold on to all that information at the same time.
因此，这种基于Transformer预训练的下一个词预测模型开始展现出强大能力：它能消化海量文本，然后基于超长上下文完成下一个词预测。没错，我们正在接近大语言模型的概念。这里我们先快进做个演示（稍后会回到主线故事）——在OpenAI首个生成式预训练Transformer模型GPT-1中，它采用无监督预训练，意味着模型学习的是未经标注的原始语料库。
So this idea of a next word predictor that was pre trained with a transformer could really start to do something pretty powerful, which is consume large amounts of text, and then complete the next word based on a huge amount of context. Yep. We're starting to come up to this idea of a large language model. And we're gonna flash forward here just for a moment to do some illustration, and then we'll come back to the story. In GPT one, the first OpenAI model, this generative pretrained transformer model, GPT, it used unsupervised pretraining, which basically meant that as it was consuming this corpus of language, it was unlabeled data.
这个模型仅通过阅读就能推断语言结构和含义，这对机器学习而言是革命性的。传统观点认为必须用高度结构化的数据训练模型，否则如何理解数据意义？但这种新范式表明：数据本身就能揭示其含义。就像儿童认知世界的过程，父母偶尔会纠正说'不对，这是红色'，但绝大多数时候他们都在通过观察进行自我学习。
The model was inferring the structure and meaning of language merely by reading it, which is a very new concept in machine learning. The canonical wisdom is that you needed extremely structured data to train your model on, because how else are you going to learn what the data actually means? This was a new thing. You can learn what the data means from the data itself. It's like how a child consumes the world where only occasionally does their parents say, no.
不对，不对，你搞错了。那其实是红色。但大多数时候，他们只是通过观察世界来自学。
No. No. You have that wrong. That's actually the color red. But most of the time, they're just self teaching by observing the world.
作为两岁孩子的家长可以证实这点。
As a parent of a two year old can confirm.
在无监督预训练阶段之后，还会进行有监督微调。无监督预训练通过海量文本学习通用语言规律，然后针对特定任务在标注数据集上微调，使模型真正具备实用价值。
And then a second thing happens after this unsupervised pre training step where you then have supervised fine tuning. The unsupervised pre training used a large corpus of text to learn the sort of general language, and then it was fine tuned on labeled datasets for specific tasks that you sort of really want the model to be actually useful for.
为了让人们理解为什么说这种超大数据量训练成本高得离谱：GPT-1参数量约1.2亿，GPT-2达到15亿，GPT-3暴涨至1750亿，而尚未官宣的GPT-4据传训练参数量高达1.7万亿。这已经与当年的AlexNet不可同日而语了。
So to give people a sense of why we're saying that the idea of training on very, very, very large amounts of data here is crazy expensive, GPT one had roughly a 120,000,000 parameters that it was trained on. GPT two had 1,500,000,000. GPT three had a 175,000,000,000. And GPT four, OpenAI hasn't announced, but it's rumored that it has about 1,700,000,000,000 parameters that it was trained on. This is a long way from AlexNet here.
它的扩展速度堪比英伟达的市值。有个有趣的发现：参数越多，预测下一个词的准确率就越高。这些模型在参数低于100亿时基本表现糟糕，甚至可能低于1000亿参数时也是如此，它们只会产生幻觉或胡言乱语。
It's scaling like NVIDIA's market cap. There is this interesting discovery, basically, that the more parameters you have, the more correctly you can predict the next word. These models were basically bad sub 10,000,000,000 parameters. I mean, maybe even sub a 100,000,000,000 parameters. They would just hallucinate or they would be nonsensical.
有趣的是，当你观察那些10亿参数级别的模型时，你会觉得它们永远不可能产出任何有用的东西。但仅仅通过增加训练数据和参数，性能就会大幅提升。基于Transformer的模型因并行计算展现出奇妙的扩展性，所以当你投入海量数据进行训练时——
It's funny when you look at some of the, like, 1,000,000,000 parameter models. You're like, there is no chance that turns into anything useful ever. But by merely adding more training data and more parameters, it just gets way way better. There's this weirdly emergent property where transformer based models scale really well due to the parallelism. So as you throw huge amounts of data at training them
你还可以投入大量英伟达GPU来处理这些数据。
You can also throw huge amounts of NVIDIA GPUs at processing that.
没错。而模型的输出会出乎意料地神奇变好。虽然我反复强调这点，但这确实令人惊讶——我们完全不改变结构，只是提供更多数据，让模型长时间运行并大幅增加参数规模。
Exactly. And the output sort of unexpectedly gets magically better. I mean, I know I keep saying that, but it is like, wait. So we don't change anything about the structure. We just give it way more data and let it run these models for a long time and make the parameters of the model way bigger.
而且没有研究者预料到它们能如此出色地理解世界，但随着模型规模不断增大，这种能力就自然而然地出现了。
And, like, no researchers expected them to reason about the world as well as they do, but it just kinda happened as they were exploring larger and larger models.
为OpenAI辩护一下，他们早就知道这点。但训练这些模型需要购买或租赁云端GPU的费用高得惊人。即便是谷歌，在这个时间节点也开始自研TPU芯片——虽然他们仍在从英伟达采购大量硬件，但已开始布局自主供应。
So in defense of OpenAI, they knew all this. But the amount of money that you would have to spend to buy GPUs or to rent GPUs in the cloud to train these models is prohibitively expensive. And, you know, even Google at this point in time, this is when they start building their own chips, TPUs. Because, you know, they're still buying tons of hardware from NVIDIA, but they're also starting to source their own here.
是的。关键是他们此时正准备向公众发布TensorFlow。他们建立了一个开发框架，心想：如果开发者使用我们的软件，那或许也该运行在我们专为该软件优化的硬件上。因此他们确实有套令人信服的软硬件协同发展逻辑。
Yep. And importantly, they've, at this point, are getting ready to release TensorFlow to the public. So they have a framework where people can develop for stuff, and they're like, look. If people are developing using our software, then maybe it should run on our hardware that's optimized to work with that software. So they actually do have this very plausible story around why their hardware, why their software framework.
当他们将其开源时，这确实是个令人意外的举动，因为人们都震惊了。你知道吗？为什么谷歌会免费送出这么重要的东西？但这提前了三四年，是个极具先见之明的举措，确实让很多人开始大规模使用谷歌的架构计算。
It was kind of a surprising move when they open sourced it because people were like, gasp. You know? Why is Google giving away the farm for free here? But this was three, four years early and a very prescient move to really get a lot of people using Google architecture compute at scale.
没错。全都在谷歌云上。是的。这样一来，整个OpenAI的闹剧似乎什么都没实现，而全球的AI资源比以往任何时候都更被锁定回谷歌手中。2018年，埃隆对此极度沮丧，基本上大发雷霆后退出并拆分了OpenAI。
Yep. All within Google Cloud. Yep. So with this, it starts to look like maybe this whole OpenAI boondoggle didn't actually accomplish anything, and the world's AI resources are more than ever just locked back into Google. So in 2018, Elon gets super frustrated by all this, basically throws a hissy fit and quits and pieces out of OpenAI.
围绕这件事有很多戏剧性情节，我们暂时不细说。他可能向团队发出了最后通牒，要么接管并掌控一切，要么离开。谁知道呢？毕竟是埃隆。但无论发生了什么，这成为OpenAI团队的重要催化剂，真正是一个历史转折的关键时刻。
There's a lot of drama around this that we're not gonna cover now. He may or may not have given an ultimatum to the rest of the team that he would either take over and run things or leave. Who knows? It's Elon. But whatever happened, this turns out to be a major catalyst for the rest of the OpenAI team and truly a history turning on a knife point moment.
这对埃隆来说可能也是个超级糟糕的决定。不过，这又是另一天的故事了。
It was also a probably super bad decision by Elon. But, again, story for another day.
在信号旗的文章中对此有很好的解释，我们会在资料来源中附上链接。作者提到，那年秋天，OpenAI的一些人更加清楚地意识到，成为尖端AI公司的成本将会上升。谷歌大脑的Transformer开辟了一个AI可以无限改进的新领域，但这意味着需要投入无尽的数据进行训练，成本高昂。OpenAI做出了重大决策，转向这些Transformer模型。2019年3月11日，OpenAI宣布成立营利性实体，以便筹集足够资金支付追求最雄心勃勃AI模型所需的计算能力。
So there's this great explanation of what happened in the semaphore piece that, we'll link to in our sources. The author says, that fall, it became even more apparent to some people at OpenAI that the costs of becoming a cutting edge AI company were going to go up. Google Brain's transformer had blown open a new frontier where AI could improve endlessly, but that meant feeding endless data to train it, a costly endeavor. OpenAI made a big decision to pivot toward these transformer models. On 03/11/2019, OpenAI announced it was creating a for profit entity so it could raise enough money to pay for all the compute power necessary to pursue the most ambitious AI models.
公司当时写道：‘我们希望增强筹集资金的能力，同时仍服务于我们的使命，而我们已知的现有法律结构都无法达到这种平衡。’OpenAI表示将限制投资者的利润，任何超额部分将归原始非营利组织所有。不到六个月后，OpenAI接受了微软10亿美元的投资。
We want to increase our ability to raise capital while still serving our mission, and no preexisting legal structure that we know of strikes the right balance, the company wrote at the time. OpenAI said it was capping profits for investors with any excess going back to the original nonprofit. Less than six months later, OpenAI took a $1,000,000,000 investment from Microsoft.
是的。我相信这大部分，如果不是全部，是由于萨姆·奥特曼的影响和接管。所以，一方面，你可以持怀疑态度说，好吧，萨姆，你把非营利组织变成了今天价值300亿美元的实体。另一方面，了解这段历史后，这几乎是他们唯一的选择。
Yeah. And I believe this is mostly, if not all, due to Sam Altman's influence and taking over here. So, you know, on the one hand, you can look at this sort of skeptically and say, okay, Sam. You took your nonprofit and you converted it into an entity worth $30,000,000,000 today. On the other hand, knowing this history now, this was kinda the only path they had.
他们不得不筹集资金以获取计算资源来与谷歌竞争。而萨姆则出去与微软达成了这些具有里程碑意义的交易。
They had to raise money to get the computing resources to compete with Google. And Sam goes out and does these landmark deals with Microsoft.
确实。太不可思议了。他们当时的观点是，这么做本质上是因为这将极其昂贵。我们仍秉持着让通用人工智能造福全人类的使命，但实现这一目标的成本将高得离谱，因此我们需要转型为营利性企业并持续运营，最终通过商业收入来资助我们的研究以实现这一使命。
Yeah. Truly amazing. And their opinion at the time of why they're doing this is, basically, this is gonna be super expensive. We still have the same mission to ensure that artificial general intelligence benefits all of humanity, But it's gonna be ludicrously expensive to get there, and so we need to basically be a for profit enterprise and a going concern and have a business that funds our research eventually to pursue that mission.
没错。2019年，他们完成了向营利性公司的转型。微软如你所说投资了10亿美元，成为OpenAI的独家云服务提供商——这对英伟达而言将变得极为重要，稍后会详细讨论。2020年6月，GPT-3面世。
Yep. So 2019, they do the conversion to a for profit company. Microsoft invests a billion dollars, as you say, and becomes the exclusive cloud provider for OpenAI, which is going to become highly relevant here for NVIDIA. More on that in a minute. June 2020, g p t three comes out.
2020年9月，微软获得了底层模型的独家商业使用权用于其产品。2021年GitHub Copilot发布，微软又向OpenAI投资了20亿美元。最终这一切在2022年11月30日达到高潮，用黄仁勋的话说就是'震惊全球的AI'——OpenAI推出了ChatGPT。
In September 2020, Microsoft licenses exclusive commercial use of the underlying model for Microsoft products. 2021, GitHub Copilot comes out. Microsoft invests another $2,000,000,000 in OpenAI. And then, of course, this all leads to 11/30/2022, in Jensen's words, the AI heard around the world. OpenAI comes out with ChatGPT.
正如你所说，本，这是史上最快达到1亿用户的产品。2023年1月，微软再次向OpenAI投资100亿美元，宣布将GPT整合到所有产品中。5月GPT-4发布，这基本上把我们带到了今天。关于OpenAI和微软的细节我们以后需要另做专题讨论，但今天的关键点是：第一，得益于这一切，生成式AI作为面向用户的产品展现出巨大机遇。
As you said, Ben, the fastest product in history to reach a 100,000,000 users. In January 2023, this year, Microsoft invests another $10,000,000,000 in OpenAI, announces they're integrating GPT into all of their products. And then in May, GPT four comes out, and that basically catches us up to today. We eventually need to go do a whole another episode about all the details here of OpenAI and Microsoft. But for today, the salient points are, one, thanks to all this, generative AI as a user facing product emerges as this enormous opportunity.
第二，为实现这一点，需要海量的GPU算力——这显然利好英伟达。但同样重要的是第三点：现在很明显，企业获取和提供这种算力的主要方式将通过云服务。这三者的结合，基本上成为了英伟达千载难逢的绝佳机遇。
Two, to facilitate that happening, you needed enormous amounts of GPU compute, obviously benefiting NVIDIA. But just as important, three, it becomes obvious now that the predominant way that companies are gonna access and provide that compute is through the cloud. And the combination of those three things turns out to be basically the single greatest moment that could ever happen for NVIDIA.
没错。你把这些都铺垫好了。到现在为止我在想：所以这算是OpenAI和微软专题？这和英伟达有什么关系？天啊，这里有个关于英伟达的精彩故事要讲。
Yes. So you're teeing all of this up. And so far, I'm thinking, so this is like the OpenAI and Microsoft episode? Like, what does this have to do with NVIDIA? And, god, there's a great NVIDIA story here to be told.
那么让我们来谈谈英伟达方面的情况。好了，听众朋友们。现在该聊聊我们另一家喜爱的公司Statsig了。自从我们上次提及Statsig后，他们迎来了一个激动人心的消息——完成了C轮融资，估值达到11亿美元。
So let's get to the NVIDIA side of it. 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.
没错。这是个重大里程碑，祝贺团队。时机也很有趣，因为实验平台领域正在真正升温。
Yeah. Huge milestone. Congrats to the team. And timing is interesting because the experimentation space is, really heating up.
是的。那么投资者为何将Statsig估值超过十亿美元？因为实验已成为全球顶尖产品团队技术栈中的关键组成部分。
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.
对。这一趋势始于Web 2.0时代的公司，如Facebook、Netflix和Airbnb。这些公司面临一个难题：如何在保持快速、去中心化的产品工程文化的同时，将团队规模扩展到数千人？实验系统就是这个答案的重要组成部分。
Yep. This trend started with web two 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.
这些系统让公司每位成员都能获取全局产品指标，从页面浏览到观看时长再到性能表现。每当团队发布新功能时，他们都能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'增加应用使用时长'这样的公司级目标，让各团队自行探索实现路径。当数千名工程师和产品经理同时推进时，就会产生指数级增长。难怪实验系统如今被视为核心基础设施。
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.
确实。当今顶尖产品团队如Notion、OpenAI、Rippling和Figma同样依赖实验系统。但他们不再自建，而是直接使用Statsig。Statsig不仅提供实验功能——过去几年还整合了快速迭代团队所需的所有工具，包括功能开关、产品分析、会话回放等。
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.
如果你想帮助团队的工程师和产品经理找到加速开发和做出更明智决策的方法，请访问statsig.com/acquired，或点击节目说明中的链接。他们提供极其慷慨的免费套餐、5万美元的初创企业计划，以及适合大企业的实惠企业合同。只需告诉他们是本和大卫推荐你的。好了，来说英伟达。
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. Okay. So NVIDIA.
好的。我们刚才在本期节目前半部分阐述了三点：第一，生成式AI已成为可能并开始获得关注；第二，它需要难以置信的海量GPU算力进行训练；第三，企业使用这些算力的主要方式似乎将通过云端。这三者的结合，我认为是我们节目迄今为止最完美的案例，印证了那句老话——对英伟达而言，幸运就是准备遇到机遇的时刻。
Okay. So we just said these three things that we've painted the picture of on the first part of the episode here that, a, generative AI is, like, possible, a thing, and it's now getting traction. B, it requires an unbelievably massive amount of GPU compute to train. And three, it looks like the predominant way that companies are going to use that compute is gonna be in the cloud. The combination of these three things is, I think, the most perfect example we've ever covered on this show of the old saying about luck being what happens when preparation meets opportunity for NVIDIA here.
显然，当前机遇是生成式AI。而在准备方面，英伟达过去五年确实在疯狂努力构建面向数据中心的新计算平台——一个GPU加速的计算平台，旨在取代英特尔主导的旧x86 CPU架构。多年来，他们确实取得了一些进展，数据中心业务也在增长，但人们总在质疑：
So, obviously, the opportunity is generative AI. But the preparation front, NVIDIA has literally just spent the past five years working insanely hard to build a new computing platform for the data center. A GPU accelerated computing platform to, in their minds, replace the old CPU led Intel dominated x 86 architecture in the data center. And for many years, I mean, they were getting some traction, right, and the data center segment was growing for NVIDIA. But people were like, okay.
你们希望这成为现实，但凭什么认为它会实现？
You want this to happen, but, like, why is it gonna happen?
没错。总有些零散工作负载——比如加密货币这样的疯狂事物，或是学术实验室里AI研究者将其用作超级计算机——会用到你们酷炫的GPU加速。但长久以来，英伟达的数据中心业务始终面临一个根本问题：企业为何要将软件栈的大规模部分迁移到GPU上？
Right. There's these little workloads here and there that will toss you, Jensen, that we think can be accelerated by your cool GPUs. And then, you know, crazy things like crypto happened, and there was, like, AI researchers in academic labs that are using it as, you know, supercomputers. But for the longest time, the data center segment of NVIDIA, it just wasn't clear that organizations had enormous parts of their software stack that they were gonna shift to GPUs. Like, why?
驱动因素是什么？现在我们知道了可能的答案——那就是AI。
What's driving this? And now we know what could be driving it, and that is AI.
不仅是可能——看看他们最新季度财报，这根本就是正在发生的现实。
Not only could be, but if you look at their most recent quarter, absolutely freaking is.
好的，那么现在问题来了，为什么是它在驱动？大卫，你愿意听我简单讲讲计算机体系结构的计算机科学课吗？
Okay. So now it begs the question, why is it driving it? And David, are you open to me giving a little computer science lecture on computer architecture?
哦，请讲吧。
Oh, please do.
好的，我得拿出我最像教授的架势来。
Alright. I need to do my best professor impression here.
老兄，我大学时超爱计算机科学，那些课是我最喜欢的。
Dude, I loved computer science in college. They were my favorite classes.
我得说，做这些节目时，台积电真的让我重温了在计算机科学课堂上那种‘哦，原来是这样运作’的兴奋感。这真的很有趣。让我们退一步想想经典的计算机体系结构——冯·诺依曼架构。如今大多数计算机和CPU都基于这种架构，它能将程序存储在计算机内存中并运行。你明白为什么这会成为主流架构。
I will say doing these episodes, this TSMC, it really does bring back the thrill of being in a CS lecture and being like, oh, that's how that works. Like, it's just really fun. So let's take a step back and consider the classic computer architecture, the Von Neumann architecture. Now the Von Neumann architecture is what most computers, most CPUs are based on today, where they can store a program in the computer's memory and run that program. You can imagine why this is the dominant architecture.
否则我们就需要为每项任务专门设计一台计算机。关键要知道计算机内存能存储两种内容：程序使用的数据，以及程序本身的指令——那些代码行。在我们即将描述的例子中，这一切都极度简化了，因为我不想涉及缓存、内存速度、内存位置等复杂问题。就保持简单吧。
Otherwise, we'd need a computer that is specialized for every single task. The key thing to know is that the memory of the computer can store two different things. The data that the program uses, and the instructions of the program itself, the literal lines of code. And in this example we're about to paint, all of this is wildly simplified because I don't wanna get into caching and speeds of memory and, you know, where memory is located, not located. So let's just keep it simple.
在冯·诺依曼架构中，处理器执行这个用汇编语言编写的程序——汇编语言会编译成处理器能理解的字节码。所以它是用指令集架构（ISA）编写的，比如来自ARM的指令集。
So the processor in the Von Neumann architecture executes this program written in assembly language, which is the language that compiles down to the byte code that the processor itself can speak. So it's written in an instruction set architecture, an ISA from ARM, for example.
或者更早的英特尔。
Or Intel before that.
是的。程序的每一行都非常简单。我们将以这个例子为例，使用一些汇编语言伪代码将数字二和三相加得到五。
Yes. And each line of the program is very simplistic. So we're gonna consider this example where I'm gonna use some assembly language pseudo code to add the numbers two and three to equal five.
本，你准备在《Acquired》节目上现场编程吗？
Ben, are you about to program live on Acquired?
嗯，这是伪汇编语言代码。第一行是从内存中加载数字二。我们要从内存中取出它，并将其加载到处理器的一个寄存器中。现在数字二就实实在在地存放在CPU上，随时可以进行操作。这是第一行代码。
Well, it's pseudo assembly language code. So the first line is we're gonna load the number two from memory. We're gonna fetch it out of memory, and we're gonna load it into a register on the processor. So now we've got the number two actually sitting right there on our CPU ready to do something with. That's line of code number one.
第二，我们将以完全相同的方式将数字三加载到第二个寄存器中。这样我们就有两个CPU寄存器存放着不同的数字。第三行，我们将执行加法操作，在CPU上将两个寄存器的值相加，并将结果存储到第三个寄存器或其中一个寄存器中。这是一个更复杂的指令，因为需要实际执行算术运算。但这些正是CPU非常擅长的事情。
Two, we're gonna load the number three in exactly the same fashion into a second register. So we've got two CPU registers with two different numbers. The third line, we're gonna perform an add operation, which performs the arithmetic to add the two registers together on the CPU and store the value in some either third register or into one of those registers. So that's a more complex instruction since it's arithmetic that we actually have to perform. But these are the things that CPUs are very good at.
我们对从内存中获取的数据进行数学运算。然后示例中的第四行也是最后一行代码是，我们将把刚刚计算出的、暂时存放在CPU寄存器中的数字五，写回到内存中的某个地址。所以这四行代码就是：加载、加载、加法、存储。
We're doing math operations on data fetched from memory. And then the fourth and final line of code in our example is we are going to take that five that has just been computed and is currently held temporarily in a register on the CPU, and we're gonna write that back to an address in memory. So the four lines of code are load, load, add, store.
这一切听起来很熟悉。
This all sounds familiar to me.
所以你可以看到这四个步骤中的每一步都只能同时执行一个操作。每一个步骤都占用CPU的一个时钟周期。如果你听说过千兆赫兹，那就是每秒的时钟周期数。一台1千兆赫兹的电脑在一秒钟内可以运行我们刚才编写的简单程序2.5亿次。但你会发现这里有个问题。
So you can see each of those four steps is capable of performing one and only one operation at a time. And each of these happens with one cycle of the CPU. So if you've heard of gigahertz, that's the number of cycles per second. So a one gigahertz computer could handle the simple program that we just wrote 250,000,000 times in a single second. But you can see something going on here.
我们四个时钟周期中有三个被用于向内存加载和存储数据。这就是所谓的冯·诺依曼瓶颈。它是人工智能的核心限制之一，至少历史上一直如此。每个步骤必须按顺序执行，且一次只能执行一个。所以在这个简单例子中，给这台电脑增加更多内存实际上没有帮助。
Three of our four clock cycles are taken up by loading and storing data to memory. Now, this is known as the Von Neumann bottleneck. And it is one of the central constraints of AI, or at least it has been historically. Each step must happen in order and only one at a time. So in this simple example, it actually would not be helpful for us to add a bunch more memory to this computer.
我无法利用这些内存做任何事情。提高时钟速度也只能带来有限的改善。即使我将时钟速度提高一倍，程序执行速度也只能快两倍。如果我在进行某些AI工作时需要百万倍的速度提升，仅靠提高时钟速度是无法实现的。这根本做不到。
I can't do anything with it. It's also only incrementally helpful to increase the clock speed. If I double the clock speed, I can only execute the program twice as fast. If I need like a million x speed up for some AI work with what I'm doing, I'm not gonna get it there with just a faster clock speed. That's not gonna do it.
当然，提高内存读写速度会有所帮助。但我在这方面受到物理定律的限制。通过导线传输数据的速度是有上限的。最具讽刺意味的是，这个瓶颈实际上会随着时间的推移而恶化而非改善。因为CPU速度越来越快，内存容量越来越大，但架构仍然受限。
And it would, of course, be helpful to increase the speed at which I can read and write to memory, memory. But I'm kinda bound by the laws of physics there. There's only so fast that I can transmit data over a wire. Now the great irony of all of this is that the bottleneck actually gets worse over time, not better. Because the CPUs get faster and the memory size increases, but the architecture is still limited.
这个被称为总线的单一通道很麻烦，我无法充分享受性能提升，因为所有数据都要挤过这个通道，而且每个时钟周期只能使用一次。所以真正的突破当然是制造非冯·诺依曼架构的计算机，让程序能够并行执行，大幅增加处理器或核心数量。这正是英伟达在硬件方面所做的，而所有AI研究人员则在软件层面找到了利用方法。但有趣的是，大卫，现在我们解决了这个问题后，限制因素不再是时钟速度或核心数量。对于这些超大规模语言模型来说，真正让我们担心的是芯片上的内存容量。
So this one pesky single channel known as a bus, I don't actually get to enjoy the performance gains nearly as much as I should because I'm jamming everything through that one channel, and that only gets to sort of be used one time per every clock cycle. So the magical unlock, of course, is to make a computer that is not a von Neumann architecture, to make programs executable in parallel and massively increase the number of processors or cores. And that is exactly what NVIDIA did on the hardware side, and all these AI researchers figured out how to leverage on the software side. But interestingly, now that we've done that, David, the constraint is not the clock speed or the number of cores anymore. For these absolutely enormous language models, it's actually the amount of on chip memory that concerns us.
我以为你要说的是这个。这就是为什么数据中心和英伟达的工作如此重要。
I thought you were going. And this is why the data center and what NVIDIA has been doing is so important.
是的。我们在亚洲计量学YouTube频道（我们在台积电那期节目中也提到过）上有个很棒的视频会讲到这点。但现在的限制实际上在于芯片上有多少高性能内存可用。这些模型需要同时全部加载到内存中，它们会占用数百GB的空间。虽然内存容量在增长——我是说我们会一直进步——但像H100这样的芯片内存大约是80GB。
Yes. There's this amazing video that we'll link to on the Asianometry YouTube channel that we link to also on the TSMC episode. But the constraint today is actually in how much high performance memory is available on the chip. These models need to be in memory all at the same time, and they take up hundreds of gigabytes. So while memory has scaled up, I mean, we're gonna get flashing all the way forward, the h one hundreds on chip RAM is like 80 gigabytes.
内存容量的增长速度远不及模型规模的实际扩张速度。训练AI所需的内存容量简直大得离谱。因此，必须将多个芯片、多台芯片服务器以及多机架的芯片服务器联网组成一个所谓的'计算机'（这里我特意加了引号），才能真正训练这些模型。值得注意的是，由于我们之前讨论过的极紫外光刻技术（EUV，在台积电那期节目中提到过）的特殊限制，我们无法制造更大的内存芯片。
The memory hasn't scaled up nearly as fast as the models have actually scaled in size. The memory requirements for training AI are just obscene. So it becomes imperative to network multiple chips and multiple servers of chips and multiple racks of servers of chips together into one single computer. And I'm putting computer in air quotes there, in order to actually train these models. It's also worth noting, we can't make the memory chips any bigger due to a quirk of the extreme ultraviolet photolithography that we talked about, the EUV on the TSMC episode.
芯片已经达到了光罩的最大尺寸。这是物理和波长的限制。在没有新的、尚未商业化的发明之前，确实无法蚀刻出更大的芯片。所以最终意味着你需要大量内存，这些内存必须非常靠近处理器，全部并行运行并以最快速度传输数据。再次声明，这是极度简化的说法，但你应该能理解为什么这一切变得如此重要。
Chips are already the full size of the reticle. It's a physics and wavelength constraint. You really can't etch chips larger without some new invention that we don't have commercially viable yet. So what it ends up meaning is you need huge amounts of memory, very close to the processors, all running in parallel with the fastest possible data transfer. And again, this is a vast oversimplification, but you kinda get the idea of why all of this becomes so important.
好的，回到数据中心话题。英伟达正在做的事情，我认为其他公司都没有做到，这也解释了为什么对他们而言，让这个全新的生成式AI世界——被黄仁勋称为'新计算时代'——在数据中心运行如此重要。过去五年英伟达做了三件事：第一，也是最重要的，与你刚才说的内容相关，本——他们在2020年完成了一笔史上最成功的收购，当时没人意识到其价值。
Okay. So back to the data center. And here's what NVIDIA is doing that I don't think anybody else out there is doing, and why it's so important for them that all of this new generative AI world, this new computing era, as Jensen dubs it, runs in the data center. So NVIDIA has done three things over the last five years. One, and probably most importantly, related to what you're talking about, Ben, they made one of the best acquisitions of all time back in 2020, and nobody had any idea.
他们收购了以色列一家名为Mellanox的小型特色网络公司。
They bought a quirky little networking company out of Israel called Mellanox.
其实规模不小。他们为此支付了70亿美元。
Well, it wasn't little. They paid $7,000,000,000 for
没错。而且它已经是上市公司了，对吧？
Yeah. And it was already a public company. Right?
是的。
It was. Yep.
是的。但它确实很独特。Mellanox是什么？Mellanox的主要产品是一种叫InfiniBand的技术，我们在与Crusoe的Chase Lockmiller合作的ACQ两期节目中讨论了很多。
Yep. But it was definitely quirky. Now what was Mellanox? Mellanox's primary product was something called InfiniBand, which we talked about a lot with Chase Lockmiller on our ACQ two episode with him from Crusoe.
实际上，InfiniBand是一个开源标准，由一个联盟管理。当时有很多参与者，但传统观点认为，InfiniBand速度更快、带宽更高，是在数据中心传输数据更高效的方式。归根结底，以太网是最低共同标准，所以无论如何大家都得实现以太网。因此大多数公司退出了市场，Mellanox成了几乎唯一的InfiniBand规格提供商。
And actually, InfiniBand was an open source standard or managed by a consortium. There were a bunch of players in it, but the traditional wisdom was, well, InfiniBand is way faster, way higher bandwidth, a much more efficient way to transfer data around a data center. At the end of the day, Ethernet is the lowest common denominator, and so everyone had to implement Ethernet anyway. And so most companies actually exited the market, and Mellanox was kind of the only InfiniBand spec provider left.
对。你刚才说，等等，InfiniBand是什么？它是与以太网竞争的标准，是数据中心机架间传输数据的一种方式。
Yeah. So you said, wait. What is InfiniBand? It is a competing standard to Ethernet. It is a way to move data between racks in a data center.
而在2020年，人们都觉得以太网够用了。数据中心机架间为什么需要比以太网更高的带宽？有什么应用需要每秒3200千兆比特的带宽？结果发现，如果你试图将数百甚至更多GPU作为一个计算集群来训练巨型AI模型，确实需要它们之间有超高速的数据互连。
And back in 2020, everybody was like, Ethernet's fine. Why do you need more bandwidth than Ethernet between racks in a data center? What could ever require 3,200 gigabits a second of bandwidth running down a wire in a data center? Well, it turns out if you're trying to address hundreds, maybe more than hundreds of GPUs as one single compute cluster to train a massive AI model, yeah, you want really fast data interconnects between them.
没错。人们原以为，哦，超级计算机用于学术目的可能需要。但企业市场在共享云计算数据中心需要的是以太网，这就够了。大多数工作负载会在单个机架上完成，最多扩展到该机架的多台计算机。但肯定不需要将多个机架联网。
Right. People thought, oh, sure, for supercomputers for these academic purposes. But what the enterprise market needs in my shared cloud computing data center is Ethernet, and that's fine. And most workloads are gonna happen right there on one rack, and maybe maybe maybe things will expand to multiple computers on that rack. But certainly, they won't need to network multiple racks together.
这时英伟达介入，黄仁勋说：嘿，傻瓜们。数据中心就是计算机。听好了，整个数据中心需要变成一台计算机。当你开始这么想时，就会意识到：天啊，我们真的要通过机架间的线缆传输海量数据，如何把它们视为片上内存或尽可能接近片上内存——尽管那些设备在三英尺外的机箱里？
And NVIDIA steps in, and you got Jensen saying, hey, dummies. The data center is the computer. Listen to me when I tell you the whole data center needs to be one computer. And when you start thinking that way, you start thinking, jeez, we're really gonna be cramming huge amounts of data through wires that are going between these like, how can we sort of think about them as if it's all sort of on chip memory or as close as we can make it to on chip memory even though that's in a box located three feet away?
是的。这就是过去五年英伟达宏伟数据中心计划的第一部分。第二部分是2022年9月，英伟达出人意料地宣布了一款新芯片——不仅是新芯片，而是一个全新品类：Grace CPU处理器。英伟达开始做CPU了。
Yep. So that's piece number one of NVIDIA's grand data center plan over the last five years. Piece number two is in September 2022, NVIDIA makes a quite surprising announcement of a new chip. Not just a new chip, an entirely new class of chips that they are making called the Grace CPU processor. NVIDIA is making a CPU.
这简直就像异端邪说。
This is, like, heretical.
但是，Jensen，我以为所有计算都将被加速。我们为什么要在这些ARM CPU上浪费时间？
But, Jensen, I thought all computing was gonna be accelerated. What are we doing here on these ARM CPUs?
没错。这些Grace CPU不是用来装进笔记本电脑的。它们是专为数据中心解决方案设计的CPU组件，从底层架构开始就旨在与这些庞大的GPU集群协同工作。
Yeah. These Grace CPUs are not for putting in your laptop. They are for being the CPU component of your entire data center solution that is specifically from the ground up design to orchestrate with these massive GPU clusters.
这是一场持续三十年的芭蕾舞终章。还记得显卡曾是英特尔主板上PCIe插槽的附庸吗？后来我们快进到未来，NVIDIA制造的这些GPU成为数据中心里漂亮的独立机箱，或是放在图形程序员身旁的小型工作站——你可以直接对GPU编程。当然它们还需要搭配CPU，所以当时用的是AMD或英特尔，或是授权某些CPU。而现在他们说：知道吗？
This is the endgame of a ballet that has been in motion for thirty years. Remember when the graphics card was subservient to the PCIe slot in Intel's motherboard? And then eventually, you know, we fast forward to the future, NVIDIA makes these GPUs that are these beautiful standalone boxes in your data center, or perhaps these little workstations that sit next to you while you're doing graphics programming, you're directly programming your GPU. And then, of course, they need some CPU to put in that, so they're using AMD or Intel or they're licensing some CPU. And now they're saying, you know what?
我们索性连CPU也自己做了。于是现在我们有了一个机箱，里面是NVIDIA的完整解决方案：RGPU、自研CPU、NVLink互联、InfiniBand组网——欢迎来到新时代。
We're actually just gonna do the CPU too. So now we make a box, and it's a fully integrated NVIDIA solution with RGPUs, our CPUs, our NVLink between them, our InfiniBand to network it to other boxes, and, you know, welcome to the show.
在讨论这一切意味着什么之前——我猜你马上要说到这个——先谈谈战略布局的第三支柱。剧透警告：你说解决方案，我听到的是毛利率。第三部分就是GPU。在NVIDIA当前Hopper架构的数据中心GPU之前，公司只有单一GPU架构。
One more piece to talk about the third leg of the stool there, strategy before we get to what it all means that I think you're about to go to. Spoiler alert. You say solution, I hear gross margin. The third part of it is the GPUs. Up until NVIDIA's current GPU generation, the hopper generation of GPUs for the data center, there was only one GPU architecture at NVIDIA.
同样的架构，同样的台积电晶圆产出的芯片，有些流向消费级游戏显卡，有些则成为数据中心的百核GPU。而从2022年9月起，他们将两条产品线拆分为不同架构：以计算机科学家格蕾丝·霍珀（我想这位是美国海军少将）命名的Hopper架构。
And that same architecture and those same chips from the same wafers made at TSMC, Some of them went to consumer gaming graphics cards, and some of those dies went to a 100 GPUs in the data center. It was all the same architecture. Starting in September 2022, they broke out the two business lines into different architectures. So there's the Hopper architecture named after great computer scientist Grace Hopper. I think rear admiral in the US Navy, Grace Hopper.
明白吗？Grace CPU、Hopper、GPU、Grace Hopper，就是那些h100系列。那是给数据中心用的。而在消费级市场，他们以Ada Lovelace命名推出了全新架构Lovelace，也就是RTX 40系列。
Get it? Grace CPU, Hopper, GPU, Grace Hopper, the h one hundreds. That was for the data centers. And then on the consumer side, they start a whole new architecture called Lovelace after Ada Lovelace. And that is the RTX 40 XX.
所以你现在买顶级RTX 40游戏显卡，它的架构已经和驱动ChatGPT的h100系列完全不同了。它采用独立架构，这非常重要——因为Hopper架构开始使用所谓'晶圆基板芯片'(COWOS)技术。
So you buy, you know, top of the line RTX 40, what have you, gaming card right now. That is no longer the same architecture as the h one hundreds that are powering Chat two PT. It's got its own architecture. This is a really big deal because what they do with the hopper architecture is they start using what's called chip on wafer on substrate, c o w o s.
一提到CoOS，硬核半导体极客们就会开始滔滔不绝讨论这个话题。
Co OS. When you start talking to the real semi nerds, that's when they start busting out the Co OS conversation.
这部分会让我们的部分听众特别兴奋。本质上，这回归到内存对GPU和AI工作负载至关重要的概念。通过垂直堆叠方式，可以在GPU芯片上集成更多内存。这是台积电最前沿的技术，而英伟达将芯片架构分化为游戏级（不配备最新COWAS技术）和...
This is when a certain segment of our listeners are gonna get really excited. So, essentially, what this is, back to this whole concept of memory being so important for GPUs and for AI workloads. This is a way to stack more memory on the GPU chips themselves, essentially by going vertical in how you build the chips. This is the absolute bleeding edge technology that is coming out of TSMC. And by NVIDIA bifurcating their chip architectures into a gaming segment that does not have this latest COWAS technology.
这使得他们能垄断台积电大量产能专供h100系列生产COWAS芯片，从而让这些芯片比市面其他GPU拥有大得多的内存。
This allows them to monopolize, like, a huge amount of TSMC's capacity to make the COWAS chips specifically for these h one hundreds, which allows them to have way more memory than other GPUs on the market.
没错。这就解释了为什么现在芯片总是不够用——本质上是台积电产能问题。你提到的两个密切相关的组件：CoAOS晶圆基板芯片和高带宽内存。半导体分析有篇精彩文章指出，2.5D芯片本质上就是通过CoAOS组装技术让内存尽可能贴近处理器。
Yes. So this gets to the point of why can't they seem to make enough chips right now? Well, it's literally a TSMC capacity problem. So there's these two components that are extremely related that you're talking about, the CoAOS chip on wafer on substrate and the high bandwidth memory. So there's this great post from semi analysis where the author points out a 2.5 d chip, which is basically how you assemble this CoAOS stuff to get the memory really close to the processor.
当然，所谓2.5D其实是3D的，但3D另有含义（比2.5D更立体），所以业界创造了这个折中说法。台积电的2.5D芯片封装技术，就是把多个活性硅晶片（比如逻辑芯片和高带宽内存堆栈）集成到单一硅基板上。虽然还有更复杂的技术细节，但关键是CoAOS目前是GPU和AI加速器芯片封装最主流的技术，也是高带宽内存协同封装的主要方法。
And of course, 2.5 d, it is literally three d, but three d means something else. It's even more three d, so they came up with this 2.5 d denominator. Anyway, the 2.5 d chip packaging technology from TSMC is where you take multiple active silicon dies, like the logic chips and the stack of high bandwidth memory, and they stack them on one piece of silicon. And there's more complexity here, but the important thing is CoAOS is the most popular technology for GPUs and AI accelerators for packaging these chips. And it's the primary method to co package high bandwidth memory.
再次强调，请记住，当前最关键的是尽可能获取靠近CPU的高带宽内存，紧邻逻辑单元，以实现交易推理的最佳性能。目前CoWoS约占台积电容量的10%至15%，许多设施专为生产这类芯片定制。因此当英伟达需要预订更多产能时，他们很可能已经预定了台积电总产能10%至15%中的很大部分。台积电需要建造更多晶圆厂，英伟达才能获得更多支持CoWoS的产能。
Again, remember, think back to the thing that's most important right now is get as much high bandwidth memory as you can closest to the CPU next to the logic to get the most performance for trading in inference. So Coas represents right now about 10 to 15% of TSMC's capacities, and many of the facilities are custom built for exactly these types of chips that they're producing. So when NVIDIA needs to reserve more capacity, there's a pretty good chance that they've already reserved some large part of the 10 to 15% of TSMC's total footprint. And TSMC needs to, like, go make more fabs in order for NVIDIA to have access to more COAS capable capacity.
是的。众所周知，台积电完成这些需要数年时间。
Yeah. Which as we know, it takes years for TSMC to do this.
没错。现在还有更多实验性探索。比如必须提到的是，确实存在内存计算的实验。随着我们逐渐远离冯·诺依曼架构，在开放接受新计算架构的当下，有人正在探索：何不直接在数据存储的内存中进行处理，而非通过铜线传输数据到CPU这种高损耗、高成本且耗能的方式？
Yep. There are more experimental things that are happening. Like, I would be remiss not to mention, there are actually experiments of doing compute in memory. Like, as we shift away from von Neumann and sort of all bets are off now that we're open to new computing architectures, there are people exploring. Well, what if we just process the data where it is in memory instead of doing the very lossy, expensive, energy intensive thing of moving data over the copper wire to get it to the CPU?
这其中存在各种权衡取舍，但深入当前学术计算机科学领域非常有趣，他们正在真正重新思考：计算机究竟是什么？
All sorts of trade offs in there, but it is very fun to sort of dive into the academic computer science world right now where they really are rethinking, like, what is a computer?
英伟达构建的这三样东西——专用Hopper数据中心GPU架构、Grace CPU平台、基于Mellanox的网络堆栈，现已形成生成式AI数据中心的完整解决方案。本，我说的解决方案是指...
So these three things that NVIDIA has been building, the dedicated Hopper data center GPU architecture, the Grace CPU platform, the Mellanox powered networking stack, They now have a full suite solution for generative AI data centers. And Ben, when I say solution
我听到的是利润率。但明确一点：作为英伟达，你不需要提供什么解决方案就能获得高利润。价格由供需关系决定，他们目前正在全力增加供应。相信我，出于各种原因，英伟达希望所有需要H100的人都能获得H100。
I hear margins. But let's be clear. You don't need to offer some sort of solution to get high margins if you're NVIDIA. Price is set where supply meets demand, and they're adding as much supply as they possibly can right now. Like, believe me, for all sorts of reasons, NVIDIA wants everyone who wants h one hundreds to have h one hundreds.
但目前价格就像——我给你一张空白支票，英伟达，金额随你填。所以他们现在的利润率疯狂高，纯粹因为这些产品的需求远超供应。
But for now, the price is kind of like a I'll write you a blank check, and NVIDIA, you write whatever you want on the check. So their margins are crazy right now just literally because there's way more demand than supply for these things.
好的。那么我们来拆解一下他们实际在卖什么。就像你说的，本，当然你可以直接去买H100，很多人确实这么做。所以，我不关心Grace CPU。
Yes. Okay. So let's break down what they're actually selling. So like you were saying, Ben, of course, you can and lots of people do just go buy h 100. So, like, I don't care about the Grace CPU.
我也不在乎Mellanox这些东西。我自己运营数据中心，非常擅长这个。
I don't care about this Mellanox stuff. I'm running my own data center. I'm really good at it.
最可能这么做的正是超大规模运营商，或者如英伟达所称的CSP——云服务提供商。
And the people who are most likely to do this are the hyperscalers or as NVIDIA refers to them, the CSPs, the cloud service providers.
这就是AWS、Azure、Google，还有Facebook内部使用的方案。
This is AWS. This is Azure. This is Google. This is Facebook for their internal use.
英伟达，别给我你们组装的DGX服务器。直接把芯片给我，我会按自己的方式集成。
Like, NVIDIA, don't give me one of these DGX servers that you assemble. Just give me the chip, and I will integrate it the way that I wanna integrate it.
我是世界级的数据中心架构师和运营商。我不要你们的解决方案，只要你们的芯片。所以他们大量销售这些。当然英伟达也在扶持生态系统中的新兴云提供商，比如我们的朋友Crusoe，还有CoreWeave和Lambda Labs（如果你听说过的话）。
I am a world class data center architect and operator. I don't want your solution. I just want your chips. So they sell a lot of those. Now NVIDIA, of course, has also been seeding new cloud providers out there in the ecosystem, like our friends at Crusoe, also CoreWeave and Lambda Labs, if you've heard of them.
这些都是英伟达密切合作的新型GPU专用云服务商。所以他们向所有这些云提供商销售H100以及之前的A100。
These are all new GPU dedicated clouds that NVIDIA is working closely with. So they're selling h one hundreds and a one hundreds before that to all these cloud providers.
假设你是《财富》500强中任意一家非科技公司。天啊，你绝对不想错过生成式AI这班船，而且你还有自己的数据中心？那么，NVIDIA为你准备了DGX解决方案。
But let's say you are an arbitrary company in the Fortune 500 that is not a technology company. And my god, do you not wanna miss the boat on generative AI, and you've got a data center of your own? Well, NVIDIA has a DGX for you.
没错。他们确实有。这是一套完整的基于GPU的超级计算机解决方案，装在箱子里直接接入你的数据中心就能使用。市面上没有其他类似产品。
Yes. They do. Full GPU based supercomputer solution in a box that you can just plug right into your data center, and it just works. There's nothing else on the market like this.
而且它全部运行CUDA。整个开发者生态系统都使用同样的语言，开发者们完全清楚如何为这个系统编写软件。
And it all runs CUDA. It is all speaking the exact language of the entire ecosystem of developers that know exactly how to write software for this thing.
这意味着你现有的AI或其他领域的开发人员，他们正在开发的所有项目都能直接迁移到你这台崭新的AI超级计算机上运行，因为全都基于CUDA。
Which means that whatever developers you already had who were working on AI or anything else, everything they were working on is just gonna come right over and run within your brand new shiny AI supercomputer because it all runs CUDA.
太棒了。
Amazing.
稍后再详谈CUDA。但正如我们所说，你说解决方案，我听到的是毛利率。NVIDIA这些DGX系统每台售价约15万到30万美元。这很疯狂。现在有了Hopper、Grace和Mellanox这三条新产品线，这些系统正变得更加集成化、专有化和高性能化。
More on CUDA in a minute. But as we said, you say solution, I hear gross margin. NVIDIA sells these DGX systems for, like, a 150 to $300,000 a box. That's wild. And now with all these three new legs, the Stool Hopper, Grace, and Mellanox, these systems are just getting way more integrated, way more proprietary, and way better.
如果你想购买顶配的DGX H100系统，单台起价50万美元。而DGX GH200 SuperPOD——就是黄仁勋最近展示的那面'AI墙'，那个装满AI的巨型房间——价格就更惊人了。
So if you wanna buy a new top of the line DGX h 100 system, the price starts at $500,000 for one box. And if you wanna buy the DGX g h 200 SuperPOD, this is the AI wall that Jensen recently unveiled, the huge, like, room full of AI.
它大概有20个机架那么宽。想象一下数据中心里整整一排的规模。
And it's, like, 20 racks wide. Imagine an entire row in a data center.
没错。这是256个Grace Hopper DGX机架全部连接在一起形成的整面墙。他们宣称这是首个即购即用的AI数据中心，可以直接购买并用于训练万亿参数的GPT-4级别模型。至于价格嘛——当然是‘请联系我们’。
Yes. This is 256 Grace Hopper DGX racks all connected together in one wall. They're billing this as the first turnkey AI data center that you can just buy and can train a trillion parameter GPT four class model. The pricing on that is call us. Of course, it is.
但我估计要几亿美元。应该不到十亿，但轻松上亿是肯定的。
But I'm imagining, like, hundreds of millions of dollars. Like, I doubt it's a billion, but hundreds of millions easily.
太疯狂了。咱们聊聊H100吧。我手头有张产品卡，他们造的这个东西简直离谱。它是2022年9月发布的，作为A100的继任者。
Wild. Well, let's talk about the h 100. I've got baseball card right here on this, insane thing that they've built. So they launched it in September 2022. It's the successor to the a 100.
单块H100 GPU售价4万美元。所以你刚才说的天价就是这么来的。
One GPU, one h 100 costs $40,000. So that's how you get to that price point you're talking about.
这就是他们卖给亚马逊、谷歌和脸书的价格。
That's what they're selling to Amazon and Google and Facebook.
对。你提到50万美元的价位——那其实是把84万美元的H100装进盒子，配上灰色CPU，再系个漂亮蝴蝶结的打包价。
Right. And you mentioned that $500,000 price point. The $500,000 is the $840,000 h one hundreds in a box with the gray CPU and, you know, the nice bow around it.
没错。我们来算一下这笔账。8乘以4万，就是32万美元。这基本上意味着英伟达通过销售这套解决方案能额外获得18万美元的利润。它搭载的是ARM处理器。
Yep. Which do the math on that. So eight times 40,000, that's $320,000. So that's essentially an extra $180,000 of margin that NVIDIA NVIDIA is getting out of selling the solution. It's an ARM CPU.
对他们来说制造这个几乎零成本。
It doesn't cost them anything to make that.
这些4万美元的H100芯片本身就有利润空间。每次他们捆绑销售更多组件时，整套系统的利润就更高。这就是捆绑销售的经济学原理——当你打包更多产品并提供更多客户价值时，自然有权获得更多利润。不过这里主要是为了说明定价机制。
And these $40,000 h one hundreds have margin of their own. So, like, every time they bundle more, there's more margin in the fully assembled. I mean, that's literally bundle economics. You are entitled to margin when you bundle more things together and provide more value for customers. But just to, like, illustrate the way that this pricing works.
选择H100的原因是它比A100快30倍——要知道A100只是两年半前的老产品——在AI训练方面快9倍。H100是专为训练语言模型设计的，比如全自动驾驶视频处理。它极易扩展，拥有18,500个CUDA核心。还记得我们之前讨论冯·诺依曼架构的例子吗？
So the reason you want an h 100 is they're 30 times faster than an a 100, which mind you is only like two and a half years older, is nine times faster for AI training. The h 100 is literally purpose built for training LMs, like the full self driving video stuff. It's super easy to scale up. It's got 18 and a half thousand CUDA cores. Remember when we were talking about the Von Neumann example earlier?
那个例子中单个计算核心只能处理四条汇编指令。而现在这个被他们称为GPU的H100，拥有18,500个能运行CUDA软件的核心，640个专为矩阵乘法优化的张量核心，以及80个流式多处理器。我们现在讨论的是什么级别的硬件？
Like, that is one computing core that is able to handle, you know, those four assembly language instructions. This one h 100, which they're calling a GPU, has 18 and a half thousand cores that are capable of running CUDA software. It's got 640 tensor cores, which are highly specialized for matrix multiplication. They have 80 streaming multiprocessors. So what are we up to here?
这玩意儿接近有两万个独立核心。它的能耗明显高于A100。关键点在于：英伟达每推出新一代产品都大幅提高功耗要求。他们既在探索物理极限，又受制于物理定律——有些性能突破只能通过暴增能耗来实现。
Close to 20,000 unique cores on this thing. It's got meaningfully higher energy usage than the a 100. I mean, a big takeaway here is that NVIDIA is massively increasing the power requirement every time they come out with a next generation. They're both figuring out how to push the edge of physics, but they're also constrained by physics. Some of this stuff is only possible with way more energy.
单块H100重量就达70磅（约32公斤）。
This thing weighs 70 pounds. This is one h 100.
Jensen在每次主题演讲中都对此大做文章，比如，'哦，我举不动它'。
Jensen makes a big deal about this every keynote that he gives them. Like, oh, I can't lift it.
它由3.5万个部件组成，拥有2500亿个晶体管，需要机器人来组装。不仅组装需要实体机器人，设计它还需要人工智能。他们现在实际上在用AI设计芯片本身。我的意思是，他们彻底重新定义了计算机的概念。
It's got a quarter trillion transistors across 35,000 parts. It requires robots to assemble it. Not only does it require physical robots to assemble it, it requires AI to design it. They're actually using AI to design the chips themselves now. I mean, they have completely reinvented the notion of what a computer is.
完全正确。这都是Jensen向客户推销的一部分。是的，我们的解决方案非常昂贵。但他喜欢用那句口头禅——买得越多，省得越多。
Totally. And this is all part of Jensen's pitch here to customers. Yes. Our solutions are very expensive. However, he uses the line that he loves, the more you buy, the more you save.
前提是你能搞到一些的话。
If you could get your hands on some.
没错。他的意思是，比如说，假设你是麦当劳，想开发一个生成式AI系统让顾客点餐之类的。如果你试图在现有数据中心基础设施上构建和运行这个系统，长期来看，计算资源会耗费你更多时间和成本，远不如直接购买我的超级计算单元。
Right. But what he means by that is, like, okay. Say you're McDonald's, and you're trying to build a generative AI so that, I don't know, customers can order or something. You're using it in your business. If you were gonna try and build and run that in your existing data center infrastructure, it would take so much time and cost you so much more over the long run-in compute than if you just went and bought my super pod here.
你可以即插即用，一个月内就能上线运行。
You can plug and play and have it up and running in a month.
是的。由于这都是加速计算，在上面运行的任务，你根本不可能用其他方式完成，或者会消耗更多能源、时间和成本。购买并在这里运行你的工作负载，或者从任何云服务提供商那里租用并在这里运行，确实有其合理性——因为结果产生得更快、更便宜，或者根本就是唯一可行的方案。
Yep. And by the fact that this is all accelerated computing, the things you're doing on it, you literally wouldn't be able to do otherwise or might take you a lot more energy, a lot more time, a lot more cost. There is a very valid story to buying and running your workloads here or renting from any of the cloud service providers and running your workloads here is more performant because the results just happen much faster, much cheaper, or at all.
是的，你提到了能量。就像，这也是Jensen的观点。他说，没错。这些事情需要消耗大量能量，但替代方案消耗的能量更多。
Yep. You mentioned energy here. Like, this is also Jensen's argument. He's like, yes. These things take a ton of energy, but the alternative takes even more energy.
所以如果我们假设这些事情会发生，实际上我们是在节省能量。不过这里有个前提，就是这些事情只能在这些类型的机器上实现。所以他推动了整个事情，但他的观点是有道理的。
So we are actually saving energy if you assume this stuff is going to happen. Now there's a bit of caveat here in that it can't happen except on these types of machines. So he enabled this whole thing, but he has a point.
哦，不过我完全认同。我是说，确实存在一个非常现实的案例——你只需要训练模型一次，然后就可以反复进行推理。我认为用压缩来类比模型训练非常贴切。大语言模型正在将整个互联网的文本压缩成小得多的模型权重。这样做的好处不仅是以小体积存储了大量实用性，还能在每次需要模型给出答案时，以相对低廉的计算成本进行推理。
Oh, I totally buy it, though. I mean, I think there's a very real case around, look, you only have to train a model once, and then you can do inference on it over and over and over again. I mean, the analogy I think makes a lot of sense for model training is to think about it as a form of compression. LLMs are turning the entire Internet of text into a much smaller set of model weights. This has the benefit of storing a huge amount of usefulness in a small footprint, but also enabling a very inexpensive amount of compute, again, relatively speaking, in the inference step for every time that you need to prompt that model for an answer.
当然，这样做的代价是一旦你将所有训练数据编码进模型，重新训练的成本会非常高。所以你最好第一次就做对，或者找到后期微调的方法——这也是许多机器学习研究者正在攻克的课题。但我总认为一个合理的类比是压缩一个超多图层的Photoshop文件。任何处理过3GB大小PS文件的人都知道，这种文件根本不可能直接发给客户。
Of course, the trade off you're making there is once you encode all of the training data into the model, it is very expensive to redo it. So you better do it right the first time or figure out little ways to modify it later, which a lot of ML researchers are working on. But I always think a reasonable comparison here is to compress a zillion layer Photoshop file. For anybody that's ever dealt with, oh, I've got a three gigabyte Photoshop file. Well, that's not a thing you're gonna send to a client.
你会把它压缩成JPEG格式再发送。在很多方面，这个JPEG作为包含PS文件所有图层的压缩副本反而更实用。但代价是你永远无法从这个压缩后的小JPEG还原出原始文件。所以这个类比就像：你让大家不必每次都制作完整的PSD文件，因为在绝大多数情况下使用JPEG就足够了。
You're gonna compress it into a JPEG, and you're gonna send that. And the JPEG is in many ways more useful as a compressed facsimile of the original layers comprising the Photoshop file. But the trade off is you can never get from that compressed little JPEG back to the original thing. So I think the analogy here is like, you're saving everyone from needing to make the full PSD every time because you can just use the JPEG the vast vast majority of the time.
希望我们现在已经相对清晰地描绘了这两方面：一是使生成式AI成为可能的技术进步，二是英伟达为何能超越显而易见的原因占据优势地位——特别是由于这些工作负载以数据中心为核心，而他们过去五年一直在从根本上重构数据中心架构。是的，除此之外，英伟达最近还宣布了其云战略中又一个令人惊叹的举措。就像我们说的，如今如果你是一家AI初创公司想要使用H100或A100芯片，很可能会选择云服务——无论是超大规模云供应商还是Crusoe、CoreWeaver、Lambda Labs等专业GPU云平台，通过租赁GPU来实现。Ben，你在这方面做过研究。
So hopefully, we've now painted a relatively coherent picture of both the advances that made the generative AI opportunity possible, that it has truly become a real opportunity, And why NVIDIA, even above the obvious reasons, was just so well positioned here, particularly because of the data center centric nature of these workloads and that they had been working so hard for the past five years to fundamentally rearchitect the data center. Yep. So on top of all this, NVIDIA recently announced yet another pretty incredible piece of their cloud strategy here. So today, like we've been saying, if you want to use h one hundreds and a one hundreds, say you're an AI startup, the way you're probably gonna do that is you're gonna go to a cloud, either a hyperscaler or a dedicated GPU cloud like Crusoe or CoreWeaver or Lambda Labs or the like, and you're gonna rent your GPUs. And Ben, you did some research on this.
那么，这大概要花多少钱？
So, like, what does that cost?
哦，我今天刚看了公有云上的定价页面。我想我看的是Azure和AWS。你可以以每小时约30美元的价格获得一台DGX服务器，配备8块A100显卡。或者你可以去AWS租用p5.48xlarge实例，那是8块H100显卡，我认为是HGX服务器，每小时约100美元。所以大概是三倍的价格。
Oh, I just looked at the pricing pages on public clouds today. I think Azure and AWS were where I looked. You can get access to a DGX server that's 8 a one hundreds for about $30 an hour. Or you can go over to AWS and get a p five dot 48 x large instance, which is eight h one hundreds, which I believe is an HGX server for about a $100 an hour. So about three times as much.
再说一次，当我说你可以获得访问权限时，我并不是真的指你能获得。我的意思是，那是标价。对吧。
And again, when I say you can get access, I don't actually mean you can get access. I mean, that's the price. Right.
如果你真能获得访问权限，那就是你要支付的费用。没错。好的。这只是获取GPU的成本。但如果你购买我们刚才讨论的所有东西，假设你是麦当劳或UPS之类的企业，你会说：'我真的很欣赏黄仁勋，我要买你推销的产品'。
If you could get access, that's what you would pay for it. Correct. Okay. That's just getting the GPUs. But if you buy everything we were talking about a minute ago, say you're McDonald's or UPS or whoever, And you're like, you know, I really like Jensen, I buy what you're selling.
我想要这个完整的集成方案。我想要一个即插即用的AI超级计算机整机。但我完全依赖云端，不再运营自己的数据中心。现在NVIDIA推出了DGX Cloud。
I want this whole integrated package. I want an AI supercomputer in a box that I can plug into my wall and have it run. But I'm all in on the cloud. I don't run my own data centers anymore. NVIDIA has now introduced DGX Cloud.
是的。当然，你可以从亚马逊、微软、谷歌或甲骨文那里租用这些实例。但是，就像...
Yeah. And, of course, you could rent these instances from Amazon, Microsoft, Google, Oracle. But, like
你得不到那种完全集成的解决方案。
You're not getting that full integrated solution.
没错。而且你获得的集成方式取决于云服务提供商想如何通过他们的专有服务来实现集成。说实话，你的团队里可能没有合适的人手来以准裸机的方式处理这些事务。即使不在你自己的数据中心而是从云端租用，根据你的员工情况，你可能实际上只需要通过网页浏览器，使用一个友好简单的网页界面，从可信来源加载一些模型，轻松与你的数据配对，点击运行即可，完全不必担心管理亚马逊或微软等云应用的复杂性，或是更接近硬件层的复杂操作。
Right. And you're getting some integration the way that the cloud service provider wants to create the integration using their proprietary services. And to be honest, you might not have the right people on staff to be able to deal with this stuff in a pseudo bare metal way. Even if it's not in your data center and you're renting it from the cloud, you might actually need, based on your workforce, to just use a web browser and just use a real nice, easy web interface to load some models in from a trusted source that you can easily pair with your data and just click run and not have to worry about any of the complexity of managing a cloud application that's in Amazon or Microsoft or something a little bit scarier and closer to the metal.
是的。NVIDIA推出了DGX Cloud，这是一个虚拟化的DGX系统，目前通过其他云平台（如Azure、Oracle和Google）向您提供。
Yep. So NVIDIA has introduced DGX Cloud, which is a virtualized DGX system that is provided to you right now via other clouds, so Azure and Oracle and Google.
没错。这些设备就存放在其他云服务提供商的数据中心里。
Right. The boxes are sitting in the data centers of these other CSPs.
对。它们部署在其他云服务商那里。但对客户来说，感觉就像是在租用自己专属的设备。
Right. They're sitting in the other cloud service providers. But as a customer, it looks like you have your own box that you're renting.
你通过NVIDIA登录DGX Cloud网站，全是直观的所见即所得操作。平台集成了Hugging Face，可以直接部署上面的模型。你可以上传数据，所有操作都非常直观——用'所见即所得'来形容最贴切。
You log in to the DGX Cloud website through NVIDIA, and it's all nice WYSIWYG stuff. There's an integration with Hugging Face where you can easily deploy models right off of Hugging Face. You can upload your data. Like, everything is just really WYSIWYG. It's probably the way to describe it.
这太不可思议了。NVIDIA居然通过其他云平台推出了自己的云服务。
This is unbelievable. NVIDIA launched their own cloud service through other clouds.
NVIDIA确实拥有六个数据中心，但我认为他们并没有用这些来支撑DGX Cloud。
And NVIDIA does have, I think, six data centers, but that I don't believe is what they're actually using to back DGX Cloud.
不是的。DGX Cloud起售价每月3.7万美元，提供的是基于A100的系统（不是H100）。这对NVIDIA及其合作伙伴来说利润高得离谱。有听众帮我们估算过，现在搭建同等配置的A100 DGX系统成本约12万美元——注意这还是上一代产品。
No. So starting price for DGX Cloud is $37,000 a month, which will get you an a 100 based system, not an h 100 based system. So the margins on this are insane for NVIDIA and their partners. A listener helped us out and estimated that the cost to actually build an equivalent a 100 DGX system would be today something like a 120 k. Remember, this is the previous generation.
这不是普通的H100设备。你可以以每月3.7万美元的价格租用它。因此，英伟达及其云合作伙伴在这项设备上的资本支出仅需三个月就能回本。对英伟达而言更长远重要的是，企业购买这些设备后，英伟达现在与这些公司建立了直接销售关系，不一定需要通过Azure、谷歌或AWS等云服务商中介，尽管计算资源仍托管在他们的云平台上。
This is not a h one hundreds. And you can rent it for 37 k a month. So that's three month payback on the CapEx for this stuff for NVIDIA and their cloud partners together. And even more for NVIDIA, more important longer term, for enterprises that buy this, NVIDIA now has a direct sales relationship with those companies, not necessarily intermediated by sales through Azure or Google or AWS, even though the compute is sitting in their clouds.
这一点至关重要，因为首席财务官Colette Kress在上次财报电话会议中提到，数据中心业务部门约一半收入来自云服务提供商(CSPs)。其次是消费互联网公司，再之后才是企业客户。这里有几个耐人寻味的点：首先，天啊，这部分收入竟然集中在5到8家云服务提供商身上。
Which is crucially important because at this point, the CFO Colette Kress said on their last earnings call that about half of the revenue from the data center business unit is CSPs. And then I believe after that is the consumer Internet companies, and after that is enterprises. So there's a few interesting things in there. One of which is, oh my god. Their revenue for this is concentrated among, like, five to eight companies with these CSPs.
其次，他们并不完全掌握客户关系。通过CUDA平台，他们拥有开发者关系。英伟达目前拥有空前强大的开发者生态系统。但就实际客户而言，半数收入仍需通过云提供商中介。另一个有趣的现象是，即便在AI爆发的今天，数据中心第二大客户群体仍是消费互联网公司。
Two, they don't necessarily own the customer relationship. They own the developer relationship through CUDA. You know, they've got this unbelievable ecosystem right now of NVIDIA developers that's stronger than ever. But in terms of the actual customer, half of their revenue is intermediated by cloud providers. The second interesting thing about this is even today, in this AI explosion, the second biggest segment of data centers is still the consumer Internet companies.
这些公司仍在使用机器学习技术优化社交媒体算法和精准广告投放——这部分业务规模甚至超过了直接从英伟达采购的企业客户。因此DGX云服务战略本质上是将部分云服务商收入转化为直接客户收入。
It's still all that stuff we were talking about before of the uses of machine learning to figure out what should show up in your social media algorithms and match ads to you, that's actually bigger than all of the direct enterprises who are buying from NVIDIA. So the DGX Cloud Play is a way to sort of shift some of that CSP revenue into direct relationship revenue.
时间来到2023年5月，英伟达公布了2024财年第一季度财报（他们采用特殊的1月财年截止日，所以2024Q1实际是2023Q1）。季度营收环比增长19%至72亿美元，这非常亮眼——要知道2022年他们经历了加密货币崩盘和资产减记的艰难时期。
So all of this brings us to 2023. In May, NVIDIA reported their q one fiscal twenty four earnings. NVIDIA's on this weird January fiscal year end thing. So q one twenty four is essentially q one twenty three. But anyway, in which revenue was up 19% quarter over quarter to 7,200,000,000.0, which is great because remember they had a terrible 2022 with the write offs and crypto falling off a cliff and all that.
确实。在2023年3月那次Strathecari专访中，黄仁勋明确表示'去年无疑是令人失望的一年'。要知道同年正是ChatGPT发布的年份，这家公司经历的过山车式发展实在令人惊叹。
Yes. It's amazing that in that Strathecari interview when was that? In March 2023, Jensen said last year was unquestionably a disappointing year. This is the year ChatGPT was released. It is wild, the roller coaster this company
这一切都发生在如此短暂的时间框架内。
has been on. The time frame is so compressed here.
当然，这其中部分原因是以太坊转向权益证明机制，英伟达的加密货币业务就此终结——我敢肯定他们其实对此欣喜若狂。但另一个原因是他们向台积电下了大量产能预订单，后来发现用不上，只好进行资产减值。从会计角度看，这就像去年财务上的一大笔亏损，一个巨大的污点。但现在，天啊，他们简直庆幸当初锁定了所有产能。
And part of that, of course, is Ethereum moving to proof of stake, the end of the crypto thing for Nvidia, which I'm sure they're actually thrilled about. But part of it was they also put in a ton of preorders for capacity with TSMC that then they thought they weren't gonna need, so they had to write down. So from an accounting perspective, it looks like a big loss, like a really big blemish on their finances last year. But now, oh my god, are they glad that they reserved all that capacity.
没错。这些产能实际上将变得极具价值。说到这个，你看他们第一季度财报就很亮眼，环比增长19%，但随后他们扔出了重磅炸弹——由于数据中心对生成式AI算力的需求空前激增，英伟达预计第二季度营收将达到110亿美元，较第一季度环比再增53%，同比增幅达65%。股价直接疯了。
Yep. It's actually going to be quite valuable. So speaking of, you know, this q one earnings is, like, great, up 19% quarter over quarter, but then they dropped the bombshell due to unprecedented demand for generative AI compute in data centers. NVIDIA forecasts q two revenue of $11,000,000,000, which would be up another 53% quarter over quarter over q one and sixty five percent year over year. The stock goes nuts.
盘后交易暴涨25%。是的。这家公司现在市值万亿美元——至少这次财报让他们跻身万亿俱乐部。但想想看，一家原本估值约8000亿美元的公司，财报后直接飙升25%。
25% in after hours trading. Yep. This is a trillion dollar company, or at least this made them a trillion dollar company. But, like, a company that was previously valued at around $800,000,000,000 popped 25% after earnings.
其实比这更疯狂。去年四月我们做节目时，英伟达还是全球市值第八大公司，市值约6600亿美元。虽然从高点略有回落，但大体是这个量级。后来一度暴跌至3000亿美元以下。结果短短几个月内，现在又重回万亿之上。
Well, and it's even crazier than that. Back when we did our episodes last April, NVIDIA was the eighth largest company in the world by market cap, had about a 660,000,000,000 market cap. That was down slightly off the highs, but that was kind of the order of magnitude back then. It crashed down below 300,000,000,000. And then within a matter of months, it's now back up over a trillion.
简直疯狂。这一切在上周我们录制本期节目时达到高潮——英伟达发布2024财年第二季度财报。通常我们不会在Acquired节目里单独讨论某次财报，毕竟在历史长河中谁在乎呢？但这次是历史性事件。我认为这是上市公司有史以来最惊人的财报之一，甚至没有之一。
Just wild. And then all of this culminates last week at the time of this recording when NVIDIA reports q two fiscal twenty four earnings. And this earnings release we usually don't talk about, like, individual earnings releases on acquired because, like, in the long arc of time, who cares? This was a historic event. I think this was one of, if not the most incredible earnings release by any scaled public company ever.
展开剩余字幕（还有 363 条）
说真的，无论未来如何，上周都注定载入史册。
Seriously, no matter what happens going forward, last week was a historic moment.
最让我震惊的是，单是他们数据中心业务板块当季就实现100亿美元营收。较前一季度直接翻倍。三个月内，该业务营收从40亿左右飙升至100亿——这可是实打实交付产品给客户产生的营收，不是预订单。
The thing that blows my mind the most is that their data center segment alone did $10,000,000,000 in the quarter. That's more than doubling off of the previous quarter. In three months, they grew from 4 ish billion to 10,000,000,000 of revenue in that segment. And revenue only happens when they deliver products to customers. This isn't preorders.
这不是点击量。这不是那些挥挥手的花哨玩意。这是我们实际交付给客户的产品，他们本季度比上季度多支付了60亿美元。
This isn't clicks. This isn't wave your hands around stuff. This is we delivered stuff to customers, and they paid us an additional $6,000,000,000 this quarter than they did last quarter.
以下是完整数据：本季度公司总收入135亿美元，环比增长88%，同比增长超过100%。正如你所说，本，数据中心部门收入103亿美元。也就是说，135亿中有103亿来自这个五年前对公司而言几乎不存在的业务部门。环比增长141%，同比增长171%。
So here are the full numbers. For the quarter, total company revenue of 13,500,000,000.0, up 88% from the previous quarter and over a 100% from a year ago. And then, Ben, like you said, in the data center segment, revenue of 10,300,000,000.0. So 10.3 out of 13.5 for a segment that basically didn't exist five years ago for the company. That's up a 141% from q one and a 171% from a year ago.
这可是100亿美元啊。如此规模的增长率前所未见。没有。市场也从未见过。
This is $10,000,000,000. That kind of growth at this scale, never seen anything like it. No. Neither has the market.
没错。
That's right.
这是我第一次注意到这点。黄仁勋在第一季度财报中就提到过，所以并非首次。但他再次提到了万亿美元市场规模。这次没用幻灯片展示，而是直接口头提及。
And so this this is the first time I noticed it. Jensen had talked about this in q one earnings, so it wasn't the first time. But he brings back the trillion dollar TAM. Not in a slide, I think, this time. He just talks about it.
不。但这次他用了一种我认为更好的阐述方式。这次不一样了。
No. But in a new way that I think is a better way to slice it. This time, it's different.
听着，我们会花些时间讨论对此的看法，但这次确实大不相同。这次他将英伟达的万亿美元机遇定位在数据中心领域。他是这么说的：全球数据中心目前拥有价值1万亿美元的硬件资产。
You know, look, we'll spend a while here now talking about what we think about this, but this is very different. This time, he frames NVIDIA's trillion dollar opportunity as the data center. And this is what he says. There is $1,000,000,000,000 worth of hard assets sitting in data centers around the world right now.
每年增长2500亿美元。
Growing at 250,000,000,000 a year.
数据中心每年用于更新和增加资本支出的投入是2500亿美元。而NVIDIA无疑拥有最全面、最完善、最连贯的平台，将成为未来这些数据中心处理大量计算工作负载的形态。这与那种‘我们要在这个100万亿美元的行业中分得1%’的说法截然不同。
Annual spend on data centers to update and add to that CapEx is $250,000,000,000 a year. And NVIDIA has certainly the most cohesive, fulsome, and coherent platform to be the future of what those data centers are gonna look like for a large amount of compute workloads. This is a very different story than, like, oh, we're gonna get 1% of this $100,000,000,000,000 of industry out there.
现在你必须相信的是，这些AI工作负载及其创建的应用程序正在创造真正的用户价值。而且有相当有力的证据。比如ChatGPT让OpenAI现在的年收入传闻已超过10亿美元，甚至可能达到几十亿美元，并且仍在显著增长。
And the thing you have to believe now because whenever someone paints a picture, you say, okay. What do I have to believe? The thing you have to believe is there is real user value being created by these AI workloads and the applications that they are creating. And there's pretty good evidence. I mean, ChatGPT made it so OpenAI is rumored to be doing over a billion dollar run rate now, maybe multiple single digit billions, and still growing meaningfully.
这就是最耀眼的例子。可以说，它是整个繁荣期的‘网景导航器’。但尤其是这些财富五百强企业的赌注在于，每个人的私有应用程序和无数其他公共界面都将出现类似GPT的体验。Jensen将其描述为未来每个应用都会有一个GPT前端，这将是一种更自然的与计算机交互的方式。
And so that is like the shining example. Again, that's the Netscape Navigator here of this whole boom. But the bet, especially with all these Fortune five hundreds, is that there are going to be GPT like experiences in everyone's private applications, in a zillion other public interfaces. I mean, Jensen frames it as in the future, every application will have a GPT front end. It will be a way that you decide that you wanna interact with computers that is more natural.
我认为他的意思并不是说取代点击按钮，而是说每个人都能某种程度上成为程序员，而编程语言就是英语。所以当人们问为什么大家都在花这么多钱时，答案在于那些有购买力的世界高管们——他们上季度刚给NVIDIA开出了100亿美元的支票——从目前看到的数据深信这项技术将足以改变世界，值得他们下如此大的赌注。而我们尚不清楚的是，这是否属实？
And I don't think he means, like, versus clicking buttons. I think he means everyone can kinda become a programmer, but the programming language is English. And so when you're sort of like, well, why is everyone spending all of this money? It is that the world's executives with the purchasing power to go write a $10,000,000,000 check last quarter to NVIDIA for all this stuff, wholeheartedly believes from the data they've seen so far that this technology is gonna change the world enough for them to make these huge bets. And the thing that we don't know yet is, is that true?
类似GPT的体验是否会成为未来长期的主流？目前已有相当好的证据表明人们喜欢这类技术，它在改变人们的生活方式、日常工作、职业发展、教育等方面非常有用。但这就是你必须相信的前提。现在是个好时机感谢我们的节目好友ServiceNow。我们曾向听众讲述过ServiceNow惊人的起源故事，以及他们如何成为过去十年表现最佳的公司之一，但我们也收到听众关于ServiceNow实际业务的疑问。
Is the GPT like experiences going to be an enduring thing for the far future or not? There's pretty good evidence so far that people like this stuff and that it's quite useful in transforming the way that, you know, everyone lives their lives and goes about day to day and does their jobs and goes through school and, you know, on and on and on. But that is the thing you have to believe. 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.
所以今天，我们就来回答这个问题。
So today, we are gonna answer that question.
首先，最近媒体频繁引用的一句话是：ServiceNow是企业的‘AI操作系统’。但具体来说，ServiceNow二十二年前创立时仅专注于自动化，最初将企业IT部门的纸质流程转化为软件工作流。仅此而已。随着时间的推移，他们在这个平台上构建了更强大复杂的任务处理能力。
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.
他们的服务范围从IT扩展到人力资源、财务、客户服务、现场运营等部门。在过去二十年里，ServiceNow完成了连接企业各个角落、实现自动化所需的繁琐基础工作。
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.
当AI时代来临，本质上AI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企业级平台和连接架构？正是ServiceNow。要回答‘ServiceNow如今做什么’这个问题，他们所谓‘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用它管理全公司的人员、设备和软件许可；客户服务部门用它检测支付失败并内部路由到对应团队处理；供应链部门用它进行产能规划，整合其他部门数据确保协同一致。不再需要跨系统重复录入数据。最近ServiceNow还推出了AI助手，任何岗位员工都能创建AI代理处理琐事，让人专注更高价值的工作。
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，只需告知是Ben和David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。好的，David，分析时间。在讨论其他内容前，我们必须先谈谈CUDA。本期节目已讨论了很多硬件，但自第二部分后我们就没再提及英伟达拼图中的这块关键部分。
Thanks, ServiceNow. Okay. So, David, analysis. We gotta talk about CUDA before we even start analyzing anything else here. Talked about a lot of hardware so far on this episode, but there's this huge piece of the NVIDIA puzzle that we haven't talked about since part two.
众所周知，CUDA是黄仁勋与Ian Buck等英伟达团队成员2006年启动的项目，实质是对科学计算的押注——让人们能将显卡用于图形处理之外的领域，为此需要强大的软件工具支持。这也是黄仁勋眼中‘或许能与开发者建立专属关系’的萌芽，不同于依附微软或英特尔开发生态，英伟达可以培育自己的开发者体系——这对公司至关重要。如今CUDA已成为我们讨论的所有AI应用的基础架构，正如黄仁勋演讲中常提及的‘CUDA平台’‘CUDA语言’。
And CUDA, as folks know, was the initiative started in 2006 by Jensen and Ian Buck and a bunch of other folks on the NVIDIA team to really make a bet on scientific computing, that people could use graphics cards for more than just graphics, and they would need great software tools to help them do that. It also was the glimmer in Jensen's eye of, oh, maybe I can build my own relationship with developers. And, you know, there can be this notion not of a Microsoft or an Intel developer who happens to be able to, you know, have a standard interface to my chip, but I can have my own developer ecosystem, which has been huge for the company. So CUDA has become the foundation that everything that we've talked about, all the AI applications are written on top of today. So, you know, you hear Jensen in these keynotes reference CUDA the platform, CUDA the language.
我花了一些时间试图弄清楚，比如当我观看开发者会议，甚至学习一些CUDA程序时，如何准确描述它的特性？
And I spent some time trying to figure out, like, when I was watching developer sessions and, like, literally learning some CUDA programs, what is the right way to characterize it?
那么今天该如何正确描述它呢？因为它已经发生了很大演变。
And what is the right way to characterize it today? Because it has evolved a lot.
是的。如今的CUDA是一个自底向上的完整体系：包含编译器、运行时环境、以及调试器和性能分析器等开发工具。它本身就是一门编程语言——CUDA C++，拥有行业专用库，兼容自2006年2月以来发布的所有显卡。这对开发者至关重要，因为你的CUDA代码可以在任何设备上运行。
Yes. So today, CUDA is starting from the bottom and going up, a compiler, a runtime, a set of development tools like a debugger and a profiler. It is its own programming language, CUDA c plus plus It has industry specific libraries. It works on every card that they ship and have shipped since 02/2006, which is a really important thing to know. And if you're a CUDA developer, your stuff works on everything.
只要是NVIDIA产品，都采用这套统一接口。它具备多层优化抽象和现成库，通过调用这些代码库可以简化开发流程，避免重复造轮子。你可以选择用C++编写并依赖其编译器在NVIDIA硬件上高效运行，或者直接用CUDA C++原生语言自行实现。关键在于它具备惊人的灵活性，拥有强大的技术支持和庞大的开发者社区作为后盾。
Anything NVIDIA, all this unified interface. It has many layers of abstractions and existing libraries that are optimized. So these libraries of code that you can call to keep your development work short and simple instead of reinventing the wheel. So, you know, there are things that you can decide that you wanna write in c plus plus and just rely on their compiler to make it run well on NVIDIA hardware for you, or you can write stuff in their native language and try to implement things yourself in CUDA c plus plus The answer is it's incredibly flexible. It is very well supported, and there's this huge community of people that are developing with you and building stuff for you to build on top of.
观察CUDA开发者数量增长曲线：2006年2月发布后，历时四年达到10万用户；到2016年（发布13年后）突破100万；仅两年后就翻倍至200万。也就是说，前1300万用户用了13年，而第二个百万只用了两年。
If you look at the number of CUDA developers over time, it was released in 02/2006. It took four years to get the first 100,000 people. Then by twenty sixteen, thirteen years in, they got to a million developers. Then just two years later, they got to 2,000,000. So thirteen years to add their first 13,000,000, then two years to add their second.
2022年开发者数量突破300万，而到2023年5月就达到了400万注册开发者。这为NVIDIA构筑了巨大的竞争壁垒。不过有趣的是，当我们与NVIDIA人员交流时，他们并不将其视为对抗竞品的护城河。
2022, they hit 3,000,000 developers. And then just one year later, in May 2023, CUDA has 4,000,000 registered developers. So at this point, there's a huge moat for NVIDIA. And I think when you talk to folks there, and frankly, when we did talk to folks there, they don't describe it this way. They don't think about it like, well, CUDA is our moat versus competitors.
他们的理念更像是：我们预见了加速计算的未来，认为更多工作负载应该通过并行化提升效率。为此我们投入上千名全职软件工程师，不遗余力地打造编程语言、编译器、基础框架等全套体系，只为让开发者能最大限度基于我们的平台创新——这才是构建开发者生态的正确方式。
It's more like, well, look, we envisioned a world of accelerated computing in the future, and we thought there are way more workloads that should be paralyzed and made more efficient that we want people to run on our hardware. And we need to make it as easy as possible for them to do that. And we're going to go to great lengths and have one, two thousand people that work at our company. They're gonna be full time software engineers building this programming language and compiler and foundation and framework and everything on top of it to let the maximum number of people build on our stuff. That is how you build a developer ecosystem.
虽然是不同的语言，但关键在于他们对这种语言赋予他们在公司中的权力怀有极大的敬畏。
It's different language, but the bottom line is they have a huge reverence for the power that it gives them at the company.
这是我们上期节目提到过的话题，但在做这期节目时我有了更清晰的认识。英伟达自认为——我也相信他们确实是——一家平台公司，特别是在本周公布惊人财报、本季度发生的一切以及股价表现之后。最近流行的一种观点你肯定经常听到：哦，这种剧情我们见过。思科当年就是这样。从更长时间尺度看，英特尔也经历过。
This is something we touched on on our last episode, but has really crystallized for me in doing this one. NVIDIA thinks of themselves as, and I believe is, a platform company, especially this week after the blowout earnings and everything that happened this quarter and the stock and whatnot. Sort of a popular take out there that you've been seeing a lot is, oh, we've seen this movie before. This happened with Cisco. You could say over a longer timescale, this happened with Intel.
没错。这些硬件供应商、半导体公司，风光时风光无限，大家都想增加资本支出。但当它们不再热门时，就真的无人问津了。但我不认为这完全适用于英伟达。他们确实生产半导体和数据中心设备。
Yeah. These hardware providers, these semiconductor companies, they're hot when they're hot, and people wanna, you know, spend CapEx. And then when they're not hot, they're not hot. But I don't think that's quite the right way to characterize NVIDIA. They do make semiconductors, and they do make data center gear.
但本质上，他们是一家平台公司。更恰当的类比是微软。他们打造操作系统，构建编程环境，还开发许多
But, really, they are a platform company. The right analogy for NVIDIA also is Microsoft. They make the operating system. They make the programming environment. They make many of
应用程序。是的。思科没有开发者生态，英特尔也从未建立过开发者生态。微软有开发者，英特尔有微软依赖，但英特尔自身没有开发者。
the applications. Right. Cisco doesn't really have developers. Intel never had developers. Microsoft had developers, and Intel had Microsoft, but Intel didn't have developers.
英伟达拥有开发者生态。他们构建了一种非冯·诺依曼架构的新体系，逆五十年技术潮流而行，让每个GPU都配备流处理器单元。可想而知，这需要全新的编程语言、编译器等一系列工具来适配这种计算模式。这就是CUDA，而且它确实大获成功。
NVIDIA has developers. I mean, they've built a new architecture that is not a von Neumann computer. They've bucked fifty years of progress, and instead, every GPU has a stream processor unit. And as you'd imagine, you need a whole new type of programming language and compiler and everything to deal with this new computing model. And that's CUDA, and it freaking works.
还有无数人以此为生进行开发。
And there's all these people that develop their livelihood in it.
你跟Jensen或公司其他人交谈时，他们会告诉你，我们是一家基础计算机科学公司。我们不只是在这里卖硬件。
You talk to Jensen, and you talk to other people at the company, and they will tell you, we are a foundational computer science company. We're not just slinging hardware here.
没错。我觉得这很有意思。他们无疑是一家平台公司，同时也是一家系统公司。实际上他们在销售大型主机。
Yeah. I mean, it's interesting. They're a platform company for sure. They're also a systems company. They're effectively selling mainframes.
我的意思是，这和当年的IBM没什么不同。他们试图向你推销一个价值1亿美元的数据中心整体解决方案，完全集成，开箱即用。
I mean, it's not that different than IBM way back when. They're trying to sell you a, you know, a $100,000,000 wall that goes in your data center, and it's all fully integrated, and it all just works.
是的。或许IBM确实是个很好的类比，就像老派IBM那样。他们制造底层技术，生产硬件，研发芯片。
Yeah. And maybe IBM actually is a really good analogy, like old school IBM here. Make the underlying technology. They make the hardware. They make the silicon.
他们为芯片开发操作系统，为客户定制解决方案。他们包办一切，然后以整体方案的形式出售。
They make the operating system for the silicon. They make the solutions for customers. They make everything, and they sell it as a solution.
好的。那么在开始分析前还有几点需要补充。我想强调一个重要事项——让我们看看时间线。因为我直到录制前两小时才发现这个。
Yep. Okay. So couple other things to catch us up here as we're starting analysis. One big point I wanna make is, let's look at a timeline. Because I didn't discover this until, like, two hours before we started recording.
2019年3月，英伟达宣布以70亿美元现金收购Mellanox。当时英特尔也在考虑收购，但被英伟达横刀夺爱。可以说当时没人真正理解英伟达的意图及其重要性。但问题是为什么？原来英伟达预见到新兴模型需要跨多服务器、多机柜运行，因此极其重视机器间的带宽问题。
In March 2019, NVIDIA announced they were acquiring Mellanox for $7,000,000,000 in cash. And I think Intel was considering the purchase, and then NVIDIA came in and kinda blew out of the water. And it is fair to say nobody really understood what NVIDIA was going to do there and why it was so important. But the question is why? Well, NVIDIA knew that these new models coming out would need to run across multiple servers, multiple racks, and they put a huge level of importance on the bandwidth between the machines.
当然，他们是怎么知道的呢？2019年8月，英伟达发布了当时最大的基于Transformer的语言模型Megatron。83亿参数，在512块GPU上训练了9天，按当时零售价计算训练成本约50万美元，这在当时是模型训练的一笔巨款——想想看，那不过是四年前的事？
And, of course, how did they know that? Well, in August 2019, NVIDIA released what was at the time the largest transformer based language model called Megatron. 8,300,000,000 parameters trained on 512 GPUs for nine days, which at the time at retail would have cost something like half a million dollars to train, which at the time was a huge amount of money to spend on model training, which is, what, only four years ago?
但如今看来，那已经显得过时了。
But Today, that's quaint.
英伟达这么做是因为他们在公司内部进行了大量研究，并与所有从事AI研究的公司合作。他们当时就意识到：'没错，这项技术会成功，而且需要最快的网络支持。'我认为这就是为什么其他人都没看出Mellanox技术的价值所在。确实如此。
NVIDIA did that because they do a huge amount of research at the company, and they work with every other company doing AI research. And they were like, oh, yes. This stuff is gonna work, and this stuff is gonna require the fastest networking available. And I think that has to do with why no one else saw how valuable the Mellanox technology could be. Yep.
关于英伟达目前的业务，我还想谈谈'数据中心即计算机'这个概念。黄仁勋去年接受本·汤普森采访时精彩地阐述了他们如何全栈构建系统——他们的理想是客户能拥有并运行DGX SuperPOD。他说：'我们全栈构建系统，但以解耦方式进入市场，融入行业的计算架构。'我想他是在表达：客户需要以多种不同方式使用我们的技术。
Another thing that I wanna talk about for NVIDIA's business today is this notion of the data center is the computer. And Jensen did a great interview with Ben Thompson last year where he talks about the idea that they build their systems full stack. Like, their dream is that you own and operate a DGX SuperPOD. And he says, we build our systems full stack, but we go to market in a disaggregated way, integrating into the compute fabric of the industry. So I think that's his sort of way of saying, look, customers need to use us in a bunch of different ways.
因此我们需要保持灵活性。但我们希望打造的每个组件都能实现这样的效果：如果全部组装起来，将带来非凡体验；而如果客户只想部分使用、通过云服务使用，或是云提供商想采用我们的技术，我们也会提供相应解决方案。这本质上是以系统思维打造产品——全栈构建系统，但以解耦方式推向市场。
So we need to be flexible on that. But we wanna build each of our components such that if you do assemble them all together, it's this unbelievable experience, and we'll figure out how to provide the right experience to you if you only wanna use them in piecemeal ways or you wanna use us in the cloud or the cloud providers wanna use us. Again, it's build the product as a system. Build the system full stack, but go to market in a disaggregated way.
我记得在那次采访中，本敏锐地捕捉到这点并追问：'等等，你们是在自建云服务吗？'黄仁勋当时的回应是：'也许吧，我们拭目以待。'果然，后来他们以这种'或许会试试看'的态度推出了DGX Cloud。
And I think if I remember right in that interview, Ben picked up on this and was like, wait. Are you building your own cloud? And Jensen was like, well, maybe. We'll see. And, of course, then they launched DGX Cloud in a well, maybe we'll see sort of way.
没错。可以预见未来可能会出现更多英伟达全资运营的数据中心。说到这些，我们必须谈谈利润率数字——上季度他们的毛利率达到70%，并预测下季度将升至72%。要知道在CUDA之前的时代，当他们还是标准化显卡制造商时，毛利率只有24%。
Yeah. You could imagine there are more NVIDIA data centers likely on the way that are, fully owned and operated. Speaking of all of this, we gotta talk to some numbers on margin. This last quarter, they had a gross margin of 70%, and they forecasted for next quarter to have a gross margin of 72%. I mean, if you go back pre CUDA when they were a commoditized graphics card manufacturer, it was 24%.
他们的毛利率已从24%提升至70%。除了几个季度因特殊一次性事件影响外，基本上每个季度都在线性增长，这得益于他们不断加深的行业护城河和差异化优势。当前的高点显然是暂时的，源于全球企业甚至部分政府（如英国和中东某些国家）的芯片短缺——简直像开空白支票一样，只求能获得英伟达硬件。
So they've gone 24 to 70 on gross margin. And with the exception of a few quarters along the way for these strange one time events, that's basically been a linear climb quarter over quarter as they've deepened their moat and as they've deepened their differentiation in the industry. We're definitely at a place right now that I think is temporary due to the supply shortage of the world's enterprises and in some cases, even governments. You look at The UK or some of the Middle Eastern countries, like blank check. I just need access to NVIDIA hardware.
这种局面终将结束，但我认为65%以上的超高毛利率不会大幅下滑。是的，我是说...
That's gonna go away, but I don't think this very high, you know, 65% plus margin is gonna erode too much. Yes. I mean,
我认为有两点：首先，我完全认同我们刚才讨论的——英伟达不仅是硬件公司，不仅是芯片公司，更是平台公司，其业务具有高度差异化。如果你想训练GPT或同类模型...
I think two things here. One, I really do believe what we were talking about a minute ago that NVIDIA is not just a hardware company. They're not just a chips company. They are a platform company, and there is a lot of differentiation baked into what they do. If you wanna train GPT or a GPT class model
只有一个选择。你只能...
There's one option. You're doing
在英伟达平台上进行。没错，虽然市面上有许多低于GPT级别的模型可供选择，特别是推理市场比训练市场更开放，可以在其他平台上操作。但英伟达是最佳选择，不仅因其硬件优势，也不仅因其数据中心解决方案。
it on NVIDIA. There's one option. And, yes, we should talk about there's lots of less than GPT class stuff out there that you can do, and especially inference is more of a wide open market versus training that you can do on other platforms. But they're the best, and they're not just the best because of their hardware. They're not just the best because of their data center solutions.
也不仅因CUDA技术。而是所有这些因素的综合。另一个能说明其领先程度的例证是——我们还没讨论中国市场呢。
They're not just the best because of CUDA. They're the best because of all of those. So the other sort of illustrative thing for me that shows how wide their lead is, we haven't talked about China yet.
H800芯片的国度。
The land of a eight hundreds.
是的。那么情况如何？去年，中国大陆的销售额占英伟达总收入的25%，其中很大一部分是销售给中国的超大规模云计算服务提供商，如百度、阿里巴巴、腾讯等。
Yes. So what's going on? Last year, China was 25% or sales to Mainland China was 25% of NVIDIA's revenue. And a lot of that is they were selling to the hyperscalers to the cloud providers in China, Baidu, Alibaba, Tencent, others.
顺便说一下，百度可能拥有全球最大的模型。他们的GPT竞品参数超过万亿，实际上可能比GPT-4还要庞大。
And by the way, Baidu has potentially the largest model of anyone. Their GPT competitor is over a trillion parameters and may actually be larger than GPT four.
哇，我之前不知道。确实很疯狂。然后在九月份，拜登政府宣布了相当全面的监管措施，禁止销售先进计算基础设施。
Wow. I didn't know that. Yep. It's wild. So then, I believe also in September, the Biden administration announced pretty sweeping regulations and bans on sales of advanced computing infrastructure.
大卫，那是出口管制，别说是禁令。虽然...没错，这是个微妙的界限。政府出台的措施几乎等同于禁令了。
David, they're export controls. Don't say bans. I mean, yes. That's a fine line. This is pretty close to bans, what the administration introduced.
因此，英伟达不能再向中国客户销售顶级的H100或A100芯片。所以他们推出了符合性能管制要求的阉割版型号——A800和H800。
As part of that, NVIDIA can no longer sell their top of the line h one hundreds or a one hundreds to anybody in China. So they created a nerfed SKU, essentially, that meets the regulations, the performance regulations, the a 800 and h eight hundreds.
我认为他们基本上就是降低了NVLink的数据传输速度。就像买了顶级A100，但没有所需的高速数据连接，这基本上导致无法训练大模型。
Which I think they basically just crank down the NVLink's data transfer speeds. So it's like buying a top of the line a 100, but not with as fast of data connections as you need, which basically makes it so you can't train large models.
没错。或者说你无法像使用最新硬件那样高效或快速地训练模型。最说明问题的是，这些芯片和机器在中国依然热销。即便性能受限，它们仍是中国能获得的最佳硬件平台——我认为这在全球任何地方都成立。
Right. Or you can't train them as well or as fast as you could with the latest stuff. The incredibly telling thing is that those chips and those machines are still selling like hotcakes in China. They're still the best hardware and platform that you can get in China, even a crippled version. And I think that's true anywhere in the world.
最近它们的需求甚至出现了更急剧的增长，因为许多中国公司正在解读形势，认为出口管制可能会变得更加严格，所以趁还能买到时赶紧入手这些A800芯片。
And there's been a even a more recent spike of them because a lot of Chinese companies are reading the tea leaves and saying, oh, export controls might get even more severe, so I should get them while I still can, these a eight hundreds.
是的。我想不出比这更能说明他们领先优势有多大的例子了。
Yep. So, I mean, I can't think of a better illustration of just how wide their lead is.
没错，说得很到位。稍微提一下英伟达的其他业务——虽然这期节目主要讲数据中心板块。
Yeah. That's a great point. Talking about the rest of NVIDIA just for a moment. I mean, this episode is about the data center segment.
哦，你是说他们还做游戏显卡？
But Oh, you mean they still make gaming cards too?
值得探讨的是，Omniverse开始展现出巨大潜力。半年前他们的发布会上，已有700家企业注册成为客户。这之所以有趣，在于它可能是两个不同领域的交汇点：拥有光线追踪的革命性3D图形技术（演示效果令人震撼），以及人工智能。由于这两类工作负载都高度可并行化，他们在这两个市场早有布局。
It is worth talking about this idea that Omniverse is starting to look really interesting. As of their conference six months ago, they had 700 enterprises who had signed up as customers. And the reason this is interesting is it could be where their two different worlds collide. Three d graphics with ray tracing, which is new and amazing and the demos are mind blowing, and AI. They have been playing in both of these markets since the workloads are both massively parallelizable.
这原本就是他们进入AI市场的原因。如果回顾我们第一期节目，英伟达最初的使命是将图形技术打造成叙事媒介。后来他们发现：天啊，我们的硬件同样擅长其他需要并行计算的任务。但奇妙的是，Omniverse预示的未来可能是需要同时调用顶尖图形能力和AI能力的应用场景。考虑到英伟达在图形软硬件和AI软硬件领域都是头号供应商，这种叠加优势简直——
That is the sort of original reason for them to be in the AI market. If you recall back to way back our part one episode, the original mission of NVIDIA was to make graphics a storytelling medium. And then their mission has expanded as they've realized, my god, our hardware is really good at other stuff that needs to be parallelized too. But fascinatingly, with Omniverse, the future could actually look like applications where you need both amazing graphical capability and AI capability for the same application. And, I mean, for all the other amazing uniqueness about NVIDIA that we've been talking about and how well positioned they are, adding this on top where they're the number one provider for graphics hardware and software and AI hardware and software.
哦对了，现在正出现一个需要同时运用这两种能力的巨大应用场景。如果成真，他们绝对会大获全胜。
Oh, and by the way, there's this huge application emerging where you actually do need both. They're just gonna knock it out of the park if that comes true.
最近有个主题演讲展示了一个超酷的演示。可能是在SIGGRAPH大会上，NVIDIA创建了一个完全光线追踪的游戏环境。看起来就像3A级游戏大作，视觉效果惊人。基本上与现实难以区分，但你要仔细观察才能发现这不是真实的，也不是在和一个真人对话。
There was a super cool demo at a recent keynote. It might have been at SIGGRAPH where NVIDIA created a game environment, you know, fully ray traced game environment. It looks like a triple a game. You know, it looks amazing. You know, basically, distinguishable from reality, but, like, you really gotta look hard to tell that this isn't real, and this isn't a real human you're talking to.
所以你在和一个不可操作角色（NPC）对话，它正在给你发布任务。他们展示的这个演示看起来太震撼了。然后他们透露，那个角色对你说的话都不是预设脚本，全部是由AI动态生成的。
So there's a nonplayable character that you're talking to, an NPC, who's giving you, like, a mission. And they show this demo. It looks amazing. Then they're like, the script, the words that that character was saying to you were not scripted. That was all generated with AI dynamically.
你会觉得这太疯狂了。想想看，你玩电子游戏时角色都是预设好的。但在你描述的这个世界里，可以有由生成式AI控制的虚拟角色，它们没有预设脚本，拥有自己的智能，并推动故事发展。
So you're like, holy crap. You know, you think about you play a video game, the characters are scripted. But in this world that you're talking about, you can have generative AI controlled avatars that are unscripted that have their own intelligences, and that drives the story.
完全同意。或者想象一架飞机不仅在进行风洞模拟，还能利用实时天气数据模拟数百万小时的飞行时间，并用AI预测未来天气。这样你就能通过AI生成的图形模拟，预知飞机可能遇到的所有现实情况。我是说，这类应用未来会越来越多。
Totally. Or, you know, an airplane that's in a simulation of not just a wind tunnel, but simulating millions of hours of flying time using real time weather that's actually going on in the world and using AI to project the weather in the future. So you can sort of know the real world potential things that your aircraft could encounter all in a generated graphical AI simulation. I mean, there's gonna be a lot more of this stuff to come.
没错，完全正确。
Yep. Totally.
关于NVIDIA还有件事我们上期没谈到，他们员工效率很高。公司有2.6万名员工，听起来很多，但相比之下市值仅两倍的微软有22万人。按市值计算，微软每美元对应的员工数是NVIDIA的五倍。虽然这个比较有点滑稽，毕竟NVIDIA市值暴涨是最近的事。
Another thing to know about NVIDIA that we really didn't talk about on the last episode, they're pretty employee efficient. They have 26,000 employees. And that sounds like a big number, but for comparison, Microsoft, whose market cap is only twice as big, has 220,000. So that is five x the number of employees per dollar of market cap going on over at Microsoft. And this is a little bit farcical since, you know, NVIDIA only recently has had such a massive market cap.
但NVIDIA正在打造的平台规模，已经达到微软量级了。
But the scale of the platform that NVIDIA is building is on the order of magnitude of Microsoft scale.
没错。他们每位员工对应的市值高达4600万美元。太疯狂了，难以置信。
Right. They have $46,000,000 of market cap per employee. Wild. Crazy.
我认为这反映了他们的企业文化，因为我们接触过那里的一些人。那确实是一种非常独特的文化。虽然是一家大型科技公司，但你从没听说过英伟达有其他大公司那种常见的荒唐事。据我所知——可能我错了——英伟达没有那种居家办公或强制返岗的政策。
Which I think translates into the culture there as we've gotten to know some folks there. It really is a very unique kinda culture. Like, it is a big tech scale company, but you never hear about the same kinda silly big tech stuff that you hear at other companies at NVIDIA. As far as I know, I could be wrong on this. There is no, like, you know, oh, work from home or return to the office policy at NVIDIA.
完全没有。就是单纯地完成工作，没人强迫你必须来办公室。而且他们的产品迭代周期还加快了。
It's like, no. It's just like, you do the job, and, you know, nobody's forcing anybody to come into the office here. And, like, they've accelerated their ship cycles.
嗯，我还感觉那里有点像'要么做毕生事业，要么别来'的氛围。据说黄仁勋直接管理40个下属，他的办公室基本就是个空会议室，因为他总在四处奔波。他总在打电话，和这个人那个人交谈。直接管理40人时，你根本没空操心谁的职业抱负。
Well, I also get the sense that it's a little bit of a do your life's work or don't be here situation. Like Jensen is rumored to have 40 direct reports, and his office is basically just an empty conference room because he's just bouncing around so much. And he's on his phone, and he's talking to this person and that person. And, like, you can't manage 40 people directly if you're worrying about someone's career ambitions.
是的。他亲口说过：'我有40个直接下属，他们都是各自领域的世界顶尖专家。这就是他们毕生的事业。'
Yep. He's talked about this. He's like, I have 40 direct reports. They are the best in the world at what they do. This is their life's work.
我从不过问他们的职业抱负。对刚毕业的新人我们会指导，但如果你是资深员工，在这里工作了二十年，已经是世界顶尖专家，我们就追求极致效率。
I don't talk to them about their career ambitions. Like, I don't need to. Like, you know, yeah, for, like, recent college grads, we do mentoring. But if you're a senior employee, you've been here for twenty years. You're the best in the world of what you do, and we're hyper efficient.
我每周七天都是早上5点开始工作，你们也一样。
And I start my day at 5AM seven days a week, and you do too.
这太疯狂了。
It's crazy.
是啊。我最近在听黄仁勋的一次访谈时，他说的那段话简直绝了。采访快结束时，记者问他：'黄仁勋，你和英伟达做了这么多惊人的事情，你是怎么放松的？' 黄仁勋的回答是——这可是原话——
Yeah. There's actually this amazing quote from Jensen that I heard on an interview with him that I was listening to. Towards the end of the conversation, the interviewer asked him, you know, Jensen, you and NVIDIA do these just amazing things. What do you do to relax? And Jensen's answer is, I'm really this is a quote, direct quote.
我随时都在放松。我喜欢在工作中放松，因为工作对我而言就是放松。解决问题让我放松，取得成就让我放松。而且他是百分之百认真的。
I relax all the time. I enjoy relaxing at work because work is relaxing for me. Solving problems is relaxing for me. Achieving something is relaxing for me. And he's a 100% serious.
简直是一万分认真。
Like, a thousand percent serious.
黄仁勋多大年纪了？
How old is Jensen?
老兄已经60岁了。
The dude is 60 years old.
感觉他的同龄人要么已经决定退休享清福，要么就是在经营公司的同时享受生活。我觉得还有一批人是这样的。但这些对他来说毫无吸引力。我隐约感觉他还能再干三十年，而且他把公司架构设计成符合这个长远规划的样子。我不认为公司里有任何接班人在准备接手。
It kinda feels like all of his peers have either decided to retire and relax or are, you know, relaxing while running their companies. I think there's another crop of people that are doing that. And that is just not at all interesting to him or what he's doing. And I kinda get the sense like he's got another thirty years in him, and he's architected the company in such a way that that's the plan. I don't think there's anyone else there where they're like getting ready for that person to take over.
我认为公司是Jensen对未来思想、意志、动力和信念的延伸，事情就是这样发展的。
I think the company is a extension of Jensen's thoughts and will and drive and belief about the future, and that's kind of what happens.
我不知道是否存在Jensen和Lori Huang基金会。但如果有，他也不会把时间花在这上面。他没有购买体育特许经营权，也没有购买超级游艇。或者即便买了，他也绝口不提，而是直接在游艇上办公。
I don't know if there is or isn't a Jensen and Lori Huang foundation. But if there is, he's not spending his time on it. He's not buying sports franchises. He's not buying mega yachts. Or if he is, he isn't talking about them, and he's working from them.
是啊，他也没收购社交媒体平台和报纸。
Yeah. He's not buying social media platforms and newspapers.
没错，完全同意。
Yeah. Totally.
我的意思是，这很能说明问题——当你观看他们的主题演讲时，台上只有Jensen和客户演示。不像苹果发布会那样，Tim Cook会叫上其他苹果员工。这就是Jensen的个人秀。
I mean, it is quite telling that when you watch one of their keynotes, it's Jensen on stage, and it's some customer demos. But it's not like the Apple keynotes where Tim Cook's calling up another Apple employee. It's the Jensen show.
确实。没人会指责Tim Cook工作不努力。但你去看看那些主题演讲，Tim做完开场白就交接给其他人，然后是一连串高管轮流发言。
Right. Nobody would accuse Tim Cook of not working hard, I don't think. But you go to those keynotes, and it's like, Tim does the welcome and then the handoff. And, you know, a parade of other executives talk about stuff.
早上好。早上好。
Good morning. Good morning.
蒂姆·苹果。我爱这名字。喜欢蒂姆·苹果。我们得找机会请蒂姆上节目。那一定会很棒。
Tim Apple. I love it. Love Tim Apple. We gotta have Tim on the show sometime. That would be amazing.
是啊，给他发短信。
Yeah. Text him.
给他发短信。
Text him.
好的，听众朋友们。现在正是感谢我们Acquired新合作伙伴Sentry的好时机。拼写是s-e-n-t-r-y，就像站岗的哨兵。没错。
Alright, listeners. This is a great time to thank a new partner of ours here at Acquired, Sentry. That's s e n t r y, like someone standing guard. Yes.
Sentry帮助开发者调试错误和延迟问题，几乎能解决任何软件故障，并在用户发怒前修复它们。正如其官网所言，它被超过4,000,000名软件开发者认为'还算不赖'。
Sentry helps developers debug errors and latency issues, pretty much any software problem, and fix them before users get mad. As their homepage puts it, it's considered, quote unquote, not bad by over 4,000,000 software developers.
今天我们要讨论Sentry如何与被收购宇宙中的另一家公司Anthropic合作。Anthropic原本使用较旧的基础设施监控系统，但在其庞大规模和复杂性面前，他们转而采用Sentry来更快发现和解决问题。
So today, we're talking about the way that Sentry works with another company in the acquired universe, Anthropic. Anthropic used to have some older infrastructure monitoring that was in place, but at their massive scale and complexity, they instead adopted Sentry to help them find and fix issues faster.
没错。在AI领域，崩溃可能造成巨大问题。当你运行像模型训练这样的大型计算任务时，一个节点故障可能影响数百甚至数千台服务器。Sentry帮助他们检测故障硬件，从而在引发连锁问题前快速剔除。Sentry让他们能在几小时而非数天内调试重大故障，尽快恢复训练任务。
Yep. Crashes can be a massive problem in AI. You're If running a huge compute job like training a model and one node fails, it can affect hundreds or thousands of servers. Sentry helped them detect bad hardware so they could quickly reject it before causing a cascading problem. Sentry enabled them to debug massive issues in hours instead of days so they could get back to their training runs.
如今，Anthropic依赖Sentry实时追踪异常、分配错误并分析故障，覆盖其研究团队使用的所有主要编程语言，包括Python、Rust和C++。据Anthropic团队称，Sentry为开发人员提供了调试问题所需的所有信息。
And today, Anthropic relies on Sentry to track exceptions, assign errors, and analyze failures in real time across all the primary languages used by Anthropic's Anthropics research teams, including Python, Rust, and c plus plus According to the Anthropic team, Sentry gives our developers one place where they have all the information they need to debug an issue.
Sentry领域的另一个有趣更新是，本月起Sentry推出了名为SEER的AI调试器。SEER是一个AI代理，它能利用Sentry的所有问题上下文和代码库，不仅猜测问题根源，还能针对应用程序提出可直接合并的修复方案。
And one other fun update in the world of Sentry is that as of this month, Sentry now has an AI debugger called SEER. SEER is an AI agent that taps into all the issue context from Sentry and your code base to not just guess, but root cause gnarly issues and propose merge ready fixes specific to your application.
我们非常兴奋能与Sentry合作。他们拥有令人印象深刻的客户名单，不仅包括Anthropic，还有Cursor、Vercel、Linear等。如果你想像全球超过13万家组织（从独立开发者到世界顶级公司）一样快速修复代码问题，可以访问sentry.io/acquired了解更多信息。Sentry为所有Acquired听众提供两个月的免费试用，只需告诉他们Ben和David推荐了你。
We are pumped to be working with Sentry. They've got an incredible customer list, including not only Anthropic, but Cursor, Vercel, Linear, and more. If you wanna fix broken code like the over 130,000 organizations using Sentry from indie hobbyists to some of the biggest companies in the world to find and fix broken code fast. You can check out sentry.i0/acquired to learn more, and they are offering two free months to all Acquired listeners. That's Sentry, s e n t r y, dot I o slash Acquired, and just tell them that Ben and David sent you.
好了。说到核心竞争力？让我们聊聊这个。对于新听众来说，这个环节我们会探讨是什么让公司能够实现持续的差异化回报，或者说，如何比最接近的竞争对手更盈利并保持这种优势。
Alright. Power? Let's talk power. Alright. So for listeners who are new to the show, this is the section where we talk about what it is about the company that enables them to achieve persistent differential returns, or in other words, to be more profitable than their closest competitor and do so sustainably.
英伟达之所以引人入胜，是因为他们虽然有一个直接竞争对手，但最有趣的竞争形式并非如此。去中介化才是。表面上，英伟达与AMD竞争，但AMD并没有像英伟达那样从台积电保留全部产能，至少在高性能GPU的2.5D封装工艺上没有。AMD也没有CUDA的开发者生态系统。
And NVIDIA is fascinating because they sort of have a direct competitor, but that's not the most interesting form of competition for them. Disintermediation is. Sure. Ostensibly, there's NVIDIA versus AMD, but, like, AMD doesn't have all this capacity reserved from TSMC, at least not for the 2.5 d packaging process for the high end GPUs. AMD doesn't have the developer ecosystem from CUDA.
AMD是最接近的直接竞争对手，但真正的挑战来自亚马逊开发Trainium和Inferentia，微软可能通过与AMD合作自建芯片的传闻，谷歌的TPU，以及Facebook通过PyTorch在开发者社区的影响力来扩展底层技术。英伟达面临许多竞争压力，但都不是直接的。
They're the closest direct comp, but it's Amazon building Trainium and Inferentia. It's if Microsoft decides to go and build their own ship as they're rumored to with AMD. It's Google and the TPU. Facebook developing PyTorch and then leveraging their foothold with PyTorch with the developer community to figure out how to extend underneath of PyTorch. There's a lot of competitive vectors coming at NVIDIA, but not directly.
更不用说那些现在也成为他们直接竞争对手的数据中心硬件供应商了。是的，英特尔等等。
Not to mention all the data center hardware providers that are their direct competitors now too. Yep. Intel, etcetera, on down the line.
是的。综上所述，他们拥有诸多优势。我们逐一讨论时，不妨先全部列出，再决定哪些值得探讨。关于反定位策略，我个人认为目前并不存在这种情况。英伟达的所有业务领域，没有哪家公司会主动选择不跟进——毕竟现在谁不想成为英伟达呢？
Yep. Now all that said, they've got a lot of powers. So as we move through these one by one, I think let's just say them all, and we can decide if there's something to talk about here. Counter positioning is the one where I actually don't think there's anything here. I don't think there's anything that NVIDIA does where there's another company that's actively choosing not to do that because any company would wanna be NVIDIA right now.
我原本赞同你的观点，但事实上当前数据中心领域存在明显的反定位策略。几年前黄仁勋和英伟达就公开宣称要重构数据中心架构，而所有现存的数据中心硬件和计算供应商都有充分动机不跟进这种变革。
I would have agreed with you, but I actually think there is strong counter positioning in the data center world right now. NVIDIA and Jensen put a flag in the ground several years ago where they said, we are going to rearchitect the data center. And all the existing data center hardware and compute providers had strong incentives not to do that.
但就目前而言，你认为其他数据中心硬件供应商具体在哪些方面没有跟进？
But, like, right now, what do you think other data center hardware providers what are they not doing?
确实。你说得对。他们现在也都在尝试将GPU引入数据中心。
Yeah. Fair point. They're all trying to put GPUs in the data center too.
所有人都会亦步亦趋地追赶英伟达，只是落后几年而已。这就是当前市场现状。
Everyone's just gonna chase exactly what NVIDIA is doing years behind them. That's the market right now.
没错。好吧，有道理。
Yep. Okay. Fair enough.
核心问题是：英伟达能否在关键领域持续保持领先？我认为当前对这家公司的全面分析就在于——在大规模客户和主流市场关注的关键维度上，他们能否持续领先那些只会跟风模仿的竞争者？毕竟其利润空间如此诱人，谁都不想拱手相让。好，第二个优势：规模经济。这简直就是CUDA的完美写照。
And the question is, will NVIDIA be able to stay ahead in ways that matter? That I think is the entire analysis on the company right now is in what ways that matter to customers at large scale and large markets will they be able to sustainably be ahead of people that are just chasing them and trying to copy what they're doing because the margin profile is so fat and juicy that people don't wanna pay it. Yep. So the second one, scale economies. This has CUDA written all over it.
当你的规模足以分摊这些成本时，你可以进行大规模的固定成本投资。当你有400万开发者想在平台上开发时，任何投入都能被合理化——如今领英上有1600名英伟达员工的职位名称包含CUDA。实际人数肯定更多，有些人可能只标注了软件工程师之类的头衔。数千人的投入却几乎不从软件直接获利，顶多获得象征性收入，但这些成本被整个开发者生态分摊了。
You can make massive fixed cost investments when you have the scale to amortize that cost across. And when you have 4,000,000 developers who wanna develop on your platform, you can justify whatever it is, 1,600 people who actively on LinkedIn at NVIDIA today have the word CUDA in their job title. I mean, I'm sure it's actually even more than that who just aren't you know, they're saying software or something like that. But thousands of people of an investment that they don't make any money on software. They may they make a de minimis amount on software, but that is amortized across the entire developer base.
我认为这很值得
I think it's worth
关于这点我想再补充些，我们上期节目也讨论过。在我看来，这就像苹果iOS与安卓的对比。苹果有成千上万的开发者专注iOS开发，安卓同样拥有遍布广泛生态的庞大开发者群体。
saying a bit more here on this too, which we also talked about in our last episode. To me, the dynamics here are a lot like Apple and iOS Yes. Versus Android. Apple has thousands and thousands and thousands of developers working on iOS. Android also has thousands and thousands of developers working on it across a widespread ecosystem.
但苹果是高度控制的软硬一体模式，安卓则不然。作为用户，你可能收到最新系统更新，也可能收不到。
But at Apple, it's all tightly controlled, it's coupled with hardware. At Android, it's not. And, like, as a user, maybe you'll get the latest operating system update. Maybe you won't.
这个比喻非常精准——英伟达就是AI领域的苹果。PyTorch则类似安卓，因为它是开源的，有多家公司参与维护。OpenCL在图形领域相当于安卓，但表现糟糕且落后。AMD为自家硬件推出的ROCm是CUDA的竞品，但太新缺乏普及。
I think this is exactly the right framing here, that NVIDIA is the apple of AI. And PyTorch is sort of Android because it's open source, and it's got a bunch of different companies that care about it. OpenCL is the Android as it pertains to graphics, but it's pretty bad and pretty far behind. RockM is the CUDA competitor made by AMD for their hardware. But again, new, not a lot of adoption.
他们正在推进，但选择开源是因为意识到无法直接对抗英伟达，需要差异化策略。没错，他们完全在复刻苹果的战术手册。
They're working on it, but they've open sourced that because they realized they can't go directly head to head with NVIDIA. They need some different strategy. But, yes, they are absolutely running the Apple playbook here.
确实。当前局势对英伟达比iOS对安卓更有利，因为英伟达十六年来从几十人到如今数千工程师持续完善CUDA。而开源生态中的安卓对标方案才刚起步——iOS与安卓的时间差仅一两年，但英伟达至少领先十年，可能接近十五年。
Yep. And I think in the current state of things, it's even more favorable to NVIDIA than iOS versus Android because NVIDIA has had first dozens and then hundreds and now thousands of engineers working on CUDA for sixteen years. Meanwhile, the Android equivalent out there in the open source ecosystem has only just been getting going. You know, if you think about the delta of the timeline between iOS and Android, it was a year and a half, two years. There's a probably at least ten, probably closer to fifteen year lead than NVIDIA has.
于是我们和一些人讨论过这个问题，我们想知道开源生态系统发生了什么变化？有没有类似Android的替代品？就连我们交谈中最乐观的人也认为，现在Facebook确实将PyTorch转移到一个基金会并脱离Facebook，这意味着其他公司可以投入几十名工程师来开发它。你觉得这很棒吧？
And so we talked to a few people about this, and we're like, oh, what's going on in the open source ecosystem? Is there an Android equivalent? And even the most bullish people we talked to were like, oh, yeah. You know, now that Facebook has really moved PyTorch into a foundation and outside of Facebook, that means that other companies can now contribute, you know, couple dozen engineers to work on it. And you're like, cool?
所以AMD将投入几十名，甚至可能100名工程师来开发PyTorch。谷歌、Facebook和其他公司也会这样做。而NVIDIA有数千名工程师在十年前就开始研究CUDA了。
So AMD is gonna contribute a couple dozen, maybe a 100 engineers to work on PyTorch. And so will Google, and so will Facebook, and so will everybody else. NVIDIA has thousands of engineers working on CUDA ten years ahead.
大卫，我给你发了这张图表，显示自2006年2月启动以来，我估计每年从事CUDA开发的员工数量。如果你计算曲线下的面积并求积分，大约有1万人年的投入进入了CUDA。祝你好运吧。
I sent you this graph, David, of my estimated number of employees working on CUDA per year since inception in 02/2006. And then if you look at the area under the curve and just take the integral, it's approximately 10 thousand person years that have gone into CUDA. Like, good luck.
再次强调，开源是非常强大的力量。市场激励机制绝对支持这种情况发生。
Now, again, open source is a very powerful thing. The market incentives are absolutely there for this to happen.
没错。有趣的是，每道护城河只有在城堡足够小时才有效。如果终点线的奖品变得足够大，你就需要更宽的护城河，并且需要想办法更严密地防守城堡。我混用了太多比喻，但你应该明白我的意思。
Right. That is the interesting point is every moat only works if the castle is sufficiently small. If the prize at the end of the finish line becomes sufficiently large, you're gonna need a bigger moat, And you need to figure out a, you know, how to defend the castle harder. I'm mixing so many metaphors here, but you get the idea.
是的。我很喜欢这个比喻。
Yeah. I love it.
当可寻址市场是1000亿美元时，这道护城河完全没问题。但如果面对的是万亿美元的市场机会呢？可能就不够了。基本上这意味着利润率会下降，竞争会随着时间的推移变得更加激烈。
This was a perfectly fine moat when the addressable market was a $100,000,000,000. Is it at a trillion dollar market opportunity? Probably not. Basically, it means margins come down and competition gets more fierce over time.
我认为NVIDIA完全理解这一点，因为正如我之前提到的，部分原因与COVID有关。但我们在第一部分就讨论过，NVIDIA如何通过将显卡发布周期缩短至六个月来挽救公司，而当时竞争对手的周期是一到两年。这种情况持续了数年，之后他们又恢复为年度发布周期。每年都有GTC大会。自COVID以来，NVIDIA重新加速回到了六个月的发布周期。
And I think NVIDIA totally gets this because part of this, as I was alluding to, is COVID related. But we talked way back in part one about how NVIDIA ended up to save the company moving to a six month shipping cycle for their graphics cards when their competitors were on a one to two year shipping cycle. That persisted for several years, and then they relaxed back to a annual shipping cycle. There were annual GTCs. Since COVID, NVIDIA has reaccelerated to a six month shipping cycle.
自COVID以来，他们大多数年份每年举办两次GTC大会，对于他们所涉及的技术复杂度而言简直疯狂。想象一下苹果一年办两次WWDC。没错，这就是NVIDIA正在发生的事。
They've been doing two GTCs a year, most years since COVID, which is insane for the level of technology complexity that they're doing. Yep. Imagine Apple doing two WWDC's a year. Yeah. That's what's happening in NVIDIA.
太疯狂了。
It's crazy.
一方面，这是文化使然；另一方面，这相当于承认：我们现在必须全力以赴才能甩开竞争对手。
So on the one hand, that's a culture thing. On the other hand, that is an acknowledgment of, like, we need to be pedal to the floor right now to outrun competition.
我们已经建立了一些结构性防御措施来保护业务，但为了保持领先，仍需以史上最快的速度奔跑，因为我们参与的是一场极具吸引力的竞赛。
We've built some structural ways to defend the business, but we need to continue running as fast as we've ever run to stay ahead because it's such an attractive race that we're in.
没错。好了，这就是规模经济。现在我们来谈谈转换成本。
Yep. Alright. So that's scale economies. Let's move to switching costs now.
迄今为止，所有重要成果——尤其是LLM模型训练——都建立在NVIDIA之上。仅这一点就积累了海量代码和组织惯性。因此，即使从软件角度考虑，要转换平台也很困难。但2023年已有企业——包括超大规模厂商和拥有自建数据中心的财富500强——正在做出将影响未来至少五年的数据中心采购与部署决策，因为这类架构重构并不频繁发生。
So far, everything of consequence, especially model training, especially on LLMs, has been built on NVIDIA. And that alone is just a big pile of code and a big amount of organizational momentum. So switching away from that, even from the software perspective, is gonna be hard. But there are companies today in 2023, both at the hyperscalers and Fortune 500 companies that own their own data centers, making data center purchase and rollout decisions that will last at least the next five years. Because these data center rearchitectures don't happen very often.
因此你最好相信，英伟达正竭尽全力趁领先优势时尽可能多地出货，以锁定未来数年的数据中心架构。
And so you better believe that NVIDIA is trying as hard as they can to ship as much product as they can while they have the lead in order to lock in that data center architecture for the next n years.
是的。我们为这期节目采访了许多人，但最有趣的对话之一是与我们最欣赏的公开市场投资者——NCS Capital团队进行的。
Yeah. We talked to many people in preparation for this episode, but one of the most interesting conversations was with some of our favorite public market investors out there, the NCS Capital guys.
这期节目的许多洞见都是从他们那里借鉴来的。
Who I stole many insights from for this episode.
他们实在太棒了。显然，我长期关注英伟达在这个领域的发展。他们指出，数据中心收入和资本支出是人类已知最具粘性的收入类型。光是数据中心采购和架构标准化决策涉及的组织转换成本——天啊，这话说起来都拗口——财富500强这类企业最多十年才会变更一次架构。
Oh, they're just so great. And, obviously, I've been following NVIDIA in this space for a long time. They made the point that data center revenue and data center CapEx is some of the stickiest revenue that is known to humankind. Just the organizational switching costs involved in data center procurement and data center architecture standardization decisions. God, that's a mouthful even to say Fortune 500 companies and the like is like, they're not changing that more than once a decade at most.
所以即便我们正处在这个由生成式AI热潮带来的泡沫时刻，尚未完全了解所有应用场景。英伟达正利用这股热潮抢占市场份额。我看到网上有人说他们喜欢这种供应紧张的状态，但我认为并非如此。我觉得他们正在千方百计寻求产能，以把握当前机遇。
So even if we're sort of in this bubbly moment around the excitement of generative necessarily know the full set of applications. NVIDIA is leveraging this excitement to go get some lock in. I've seen some people on the Internet being like, they love how supply constrained they are. I don't think so. I think they're looking for capacity in every way they can get it to exploit this opportunity while it exists.
我完全同意。虽然我们没和英伟达CFO科莱特聊过这个，但强烈怀疑如果我是他们，会很乐意牺牲部分当前毛利率来提升销量。
I completely agree with that. Yeah. I think, you know, again, we didn't talk to Colette, NVIDIA's CFO, about this. But I strongly suspect if I were them, I would be happy to trade some of this gross margin right now for increased throughput on sales.
没错。但台积电只有一家，能生产所谓2.5D架构的晶圆厂也有限。所以
Yep. But there's only one TSMC, and there's only so many fabs that they have that can do the what do they call it? The 2.5 d architecture. So
我们应该讨论垄断资源吗？
Should we talk cornered resource?
是的。这可能是教科书级的垄断资源案例。英伟达在台积电拥有大量产能，而其竞争对手都无法触及。他们确实有点幸运地获得了这种垄断资源——当初他们预订那些晶圆产能是为了其他用途，部分用于加密货币挖矿，但AMD没有这样的资源。
Yeah. This is probably the textbook cornered resource. NVIDIA has access to a huge amount of capacity at TSMC that none of their competitors can get their hands on. I mean, they did luck into this cornered resource a little bit. They reserved all that wafer supply for a different purpose, partially crypto mining, but AMD doesn't have it.
需要说明的是，AMD在台积电确实拥有大量产能用于其他产品，比如数据中心CPU，他们在这方面确实做得很好。但英伟达最终在台积电的CoAOS产能上独占了一条宽阔的赛道，他们必须趁此机会尽可能充分利用。
AMD does have a ton of capacity, it's worth saying, at TSMC for their other products, data center CPUs, which they've actually been doing very well in. But NVIDIA did end up with this wide open lane all to themselves on CoAOS capacity at TSMC, and they gotta make the most of that for as long as they have it.
没错。不过我想补充的是，正如我们在台积电专题中讨论过的，这并非大宗商品。虽然台积电是代工厂，但它与大宗商品截然相反，尤其是在最高端的先进制程领域。
Yep. And I guess to say a little more though, it's not like this is not a commodity as we talked about on our TSMC episode. Although TSMC is a contract manufacturer, it is the opposite of a commodity, especially at the highest end leading edge.
这就像外星人带来的发明，地球上没几个人真正掌握这项技术。
It's like a invention delivered by aliens that very few humans know how to actually do.
确实。
Yes.
值得承认的是，大语言模型训练领域基本是双雄争霸。虽然我们一直在强调英伟达，但谷歌TPU也有量产。只不过你只能通过谷歌云获取。我不确定是否必须使用TensorFlow框架（这个框架相对于PyTorch正在失去热度），但使用TPU肯定不像使用英伟达硬件那样成为行业标准。
It is worth acknowledging. It's kind of a two horse race for LLM training. I know we've been harping on NVIDIA, but Google TPUs are also manufactured at volume. You can just only get them through Google Cloud. And I think I don't know if you have to use the TensorFlow framework, which has been waning in popularity relative to PyTorch, but it's certainly not an industry standard to use TPUs the way that it is to use NVIDIA's hardware.
我怀疑谷歌内部大量TPU算力正用于Bard项目，以及谷歌搜索中的各种功能。据我所知，他们已经在搜索中整合了大量生成式AI能力。
I suspect a lot of the volume of the, TPUs is being used internally by Google for Bard, for doing stuff in Google Search. Like, I know they've added a lot of the generative AI capability to search.
没错。完全同意。关于这点有两个观察。仅从商业和市场讨论范畴来看，这是谷歌战略冲突造成的主要后果。显然，理想做法应该像英伟达那样——客户希望通过云服务购买算力。
Yep. Totally. Two points on this. Just sticking to the scope of this business and market discussion, This is a major casualty of a strategy conflict at Google. Obviously, the way you wanna do this is the way NVIDIA is doing this of, like, your customers wanna buy through the cloud.
你希望入驻所有云平台。但显然谷歌不可能同时进驻AWS、Azure、Oracle和所有新兴云服务商，他们只会存在于GCP。
You wanna be in every cloud. But, obviously, Google is not gonna be in AWS and Azure and Oracle and all the new cloud providers. They're only gonna be in GCP.
也许吧，大卫。
Maybe, David.
不过我想说的是，从更宏观的视角来看，这对谷歌是合理的，因为他们的主营业务是自有产品。
But I was gonna say, though, through the expanded lens, though, I think this makes sense for Google because their primary business is their own products.
确实。他们运营着人类历史上最赚钱的业务之一。所以任何能巩固优势、延长领先地位的事，他们都应该去做。
Right. And they run among the most profitable businesses the world has ever seen. So anything they can do to further advantage and extend that runway, they probably should do.
贯穿所有这些变化，有个事实始终未变：上一代AI通过机器学习赋能社交媒体和互联网应用，创造了人类已知最强劲的现金流喷泉——这一切至今依然适用，对谷歌也是如此。
Nothing has changed through all of this with respect to the fact that what the previous generation of AI enabled with machine learning with regard to social media and Internet applications being the most profitable cash flow geysers known to man, none of that has changed. That is still true in this current world and still true for Google.
是的。我最后强调的一点是网络效应。他们拥有大量开发者和客户群体，能够分摊技术投资成本，并实现多方共赢。想想看，有人在CUDA基础上开发库，你可以直接复用他人构建的模块来编写代码。用CUDA能写出惊艳的程序，代码量却不多，因为它调用了大量现有资源。
Yep. And the last one that I had highlighted is network economies. They have a large number of developers out there and a large number of customers that they can amortize these technology investments across and who all benefit from each other. I mean, remember, there are people building libraries on top of CUDA, and you can use the building blocks that other people built to build your code. You can write amazing CUDA programs that just don't have that many lines of code because it's calling other preexisting stuff.
英伟达在2006年做出了当时代价高昂的战略决策——确保每块出厂GPU都完整支持CUDA。如今回头看堪称神来之笔，现在有五亿块支持CUDA的GPU可供开发者使用，这形成了极强的吸引力。我把这归入网络效应范畴。
And NVIDIA made a decision in 2006 that at the time was very costly, like big investment decision. But it looks genius in hindsight to make sure that every GPU that went out the door was fully CUDA capable. And today, there are 500,000,000 CUDA capable GPUs for developers to target. It's just very attractive. I'm putting this in network economies.
我认为这更像是规模经济而非网络效应。但你可以想象2012年2月时，很多人对英伟达的做法不以为然：为什么我们要压缩软件体积，却要让CUDA占用大量空间？为什么要在硬件上做出这些妥协？人们真的会用CUDA吗？如今看来这个决策简直天才。
I think it's probably more a scale economy than a network economy, but you could imagine a lot of people ho humming around NVIDIA in 02/2012 saying, why do I have to make it so that my software fits on this tiny little footprint and we can include CUDA taking up a huge amount of space on this thing and make all these trade offs in our hardware so that we can why are people gonna use CUDA? And today, it just looks so genius.
没错，我们在节目中多次讨论过这点，包括与汉密尔顿、赫尔默和陈毅本人的对话。对于英伟达这样的平台型企业，其特殊影响力正是规模经济与网络效应的结合体，这也是你想表达的观点。
Yeah. I mean, we've talked about this many times on the show, including with Hamilton, Helmer, and Chen Yi themselves. But for platform companies like NVIDIA clearly is, there is this special brand of power that is a combination of scale economies and network economies, and this is what you're getting at.
确实，他们的品牌影响力毋庸置疑。
Yep. They do have branding power for sure.
是的，我觉得值得深入探讨下这个话题。
Yeah. I actually think it's worth talking about this a little bit.
这就是'买IBM永远不会被开除'的现代版——在AI时代，英伟达就是当代的IBM。
This is the nobody gets fired for buying IBM. I mean, NVIDIA is the modern IBM in the AI era.
是的。听着，我还没有足够的信心能拍桌子打包票。但考虑到这家公司的创立背景、存在时间，以及他们在另一个完全不同的图形业务领域也拥有市场领先产品——既面向消费者又涵盖专业图形领域——我认为这确实为他们的品牌增添了些许力量，特别是当麦当劳的首席信息官和高管团队在做采购决策时。毕竟，人人都知道英伟达。
Yep. Look, I don't feel confident enough to, like, pound the table on this. But given the nature of how the company started, how long they've been around, and the fact that they also have the market leading product in a totally different business in graphics, you know, which is both consumers, but also professional graphics. I think that probably does lend some brand power to them, especially when the CIO and the c suite at McDonald's is making a buying decision here. Like, everybody knows NVIDIA.
你是说他们将消费者品牌的影响力延伸到了企业领域。
You're saying that they carried their consumer brand into their enterprise posture.
虽然在技术栈的底层力量上相差甚远，但我不认为这对他们有害。他们一直被视为技术领导者，几十年来全世界都知道他们能实现的技术堪称神奇。
This is way, way, way down the stack in power, but I don't think it's hurt them. They've always been known as a technology leader, and the whole world has known for decades at this point that the stuff that they can enable is magical.
没错。这里还有个‘强者愈强’的现象，我敢打赌上个季度的营收成果，已经远远超过了他们从消费者端获得的任何品牌收益。事实就是——看啊，所有人都在买英伟达，我不买才是傻子。
Yeah. There's a big strength leads to strength thing here too, where I bet the revenue results from last quarter massively dwarf any brand benefit that they ever got from the consumer side. I think it's just the fact that, like, hey, look, everyone else is buying NVIDIA. I'd be an idiot not to.
短期内没人会因为采购英伟达而被炒鱿鱼。确实。
Nobody is getting fired for buying NVIDIA anytime soon. Yep.
对。或者对他们形成重大依赖，又或者选定那个开发平台。道理很简单：如果你在业务上创新，就不该在底层平台上冒险。你只想成为价值链中唯一的风险点。
Right. Or taking a big dependency on them or targeting that development platform. It's just the like, if you're innovating in your business, you don't wanna take risk on the platform you're building on top of. You wanna be the only risk in the value chain.
好吧。那么最后一点，就是处理能力了。
Alright. Then the last one, right, is process power.
是的。尽管我确信你可以提出一些关于他们处理能力的论点，但这可能是最弱的一个。只是因为其他所有优势都显得更有价值。
Yeah. And this is probably the weakest one, even though I'm sure you could make some argument that they have process power. It's just that all the other powers are so much more valuable.
这总是很难区分。没错。我认为这里的论点就像是英伟达的文化和他们过去明显拥有的六个月发布周期，后来一度失去，现在又恢复了。我不确定。但我觉得这里可以提出一个论点。
It's always so tricky to tease out. Yep. You know, I think the argument here would just be like NVIDIA's culture and their six month shipping cycle that clearly they had in the past, and they didn't have for a while, and now they have again. I don't know. I think you can make an argument here.
这可行吗？让我们做个思维实验。他们的任何竞争对手，在任何领域，真的能转向六个月发布周期吗？那会非常困难。是的。
Is it feasible? Let's do a thought exercise. Could any of their competitors, really in any domain, move to a six month ship cycle? That'd be really hard. Yeah.
你想，像苹果这样规模的公司能一年办两次WWDC吗？显然不能。
You know, could a Apple sized company do two WWDC's a year? Like, no.
问题在于，这真的重要吗？现在有这么多人正在使用H100。事实上，除非你在训练GPT-4模型，大多数工作负载都可以在H100上运行。我只是不确定这一点是否真的那么重要，或者比其他因素更重要。我给你举个例子。
The question is, does that actually matter? There are so many people that are using a one hundreds right now. And in fact, most workloads can be run on a one hundreds unless you're doing model training of GPT four. I just don't know that it actually matters that much or as much as other factors. And I'll give you an example.
AMD确实在他们最新的一款GPU上采用了3D封装技术。这是一种更先进的方式，无需硅中介层就能实现真正的铜对铜直接连接。我稍微深入了一些细节。但基本上，这比H102.5D用来确保内存极度接近计算单元的技术更复杂。这重要吗？
AMD does have three d packaging on one of their latest GPUs. It's a more sophisticated way of doing real copper to real copper direct connection without a silicon interposer. I'm getting into a little bit of the details. But basically, it's more sophisticated than the process that the h 102.5 d is using to make sure that memory is extremely close to compute. And does that matter?
并不重要。重要的是我们一直在讨论的其他所有因素，没有人会因为这个小改进就做出购买决定。
Not really. What matters is everything else that we've been talking about, and nobody's gonna make a purchase decision on this thing because it's, you know, a little bit of a better mousetrap.
是的。深入思考后，我认为品牌确实是NVIDIA当前的重要优势。
Yeah. Thinking about this more, I think actually brand is a really important power for NVIDIA right now.
没错。而且强者愈强，你能理解他们为何要把握此刻机遇。嗯，剧本环节？
Yeah. And in a strength leads to strength way, so you can see why they're trying to sort of seize this moment. Yep. Playbook?
好的，我们进入剧本环节吧。
Alright. Let's move on to Playbook.
我想指出一点，黄仁勋不断将此刻称为AI的iPhone时刻。通常理解是他指代一种与计算机交互的新主流方式。但还有另一种解读——大卫，当我说一家通过软件实现差异化、继而拓展至服务的硬件公司时，这听起来耳熟吗？
So one thing that I wanna point out is Jensen keeps referring to this as the iPhone moment for AI. And when he says it, the common understanding is that he means a new mainstream method for interacting with computers. But there's another way to interpret it. Does this sound familiar, David, when I say a hardware company differentiated by software that then expanded into services?
是的，确实如此。
Yes. Yes. It does.
用AI的iPhone时刻自比NVIDIA为苹果，这颇具戏谑意味。因为两者的相似度惊人：NVIDIA提供垂直整合的软硬件堆栈，开发者使用其工具进行开发，他们出货量最大因此开发者有强烈动机瞄准该市场，高端个人买家对价格最不敏感且最认可优质体验。
It's quite tongue in cheek to be referring to the iPhone moment of AI when referring to oneself NVIDIA as the Apple. Because I really think that the parallels are uncanny that they have this vertically integrated hardware and software stack provided by NVIDIA. You use their tools to develop for it. They've shipped the most units, so developers have a big incentive to target that market. It's the best individual buyers to target because they're the least cost sensitive, and they appreciate you building the best experiences for them.
这就像是iPhone，但在许多方面更胜一筹，因为目标客户是企业而非消费者。唯一区别在于苹果市值长期滞后于其用户价值体现，而NVIDIA目前完全处于超预期状态。
I mean, it's the iPhone, but in many ways, it's better because the target is a b to b target instead of consumers. Yep. The only way in which it's different is Apple has always had a market cap that sort of lagged its proven value to users, whereas NVIDIA right now is, exactly over their skis.
好吧，我们把这个留到最后的牛市和熊市环节再讨论。
Well, let's save that for bull and bear at the end.
很好。第二点是他们已从硬件公司转型为真正的系统公司。虽然英伟达的芯片通常领先，但芯片间的比较其实无关紧要。这不是竞争的主战场。关键在于多GPU和多机架GPU如何与所有硬件、网络和软件协同工作，形成一个整体系统。
Great. The second one is that they've moved on from becoming a hardware company to truly being a systems company. While NVIDIA's chips are typically ahead, it really doesn't matter on a chip to chip comparison. That is not the playing field. It is all about how well multiple GPUs and multiple racks of GPUs work together as one system with all the hardware and networking and software that enables that.
他们彻底改变了竞争方向，我认为很多公司都可以从中学习。我的第三点来自Jensen在同一个Strathecari采访中的话：'通过做别人做不到的事来打造伟大公司，而不是通过与人争夺谁都能做的事来建立企业。'我觉得这句话非常精辟，体现在许多有趣的方面。
They have just entirely changed the vector of competition, which I think lots of companies can learn from. And my third one here is this quote that Jensen had again from the same Strathecari interview, which is, you build a great company by doing things that other people can't do. You don't build a company by fighting other people to do things that everyone can do. And I think it's so salient. It comes out in all these interesting ways.
其中一个例子是英伟达从未投入资源开发CPU，直到现在出现了差异化方式和真正需要自研CPU的理由。顺便说，他们的做法其实差异化不大——是在现成的ARM架构上加入自己的秘制配方，并非像苹果那样从零打造M3芯片。
One of which is NVIDIA never dedicated resources to building a CPU until there was a differentiated way and a real reason for them to build their own CPU, which is now. And the way that they're doing it, by the way, is not terribly differentiated. It's an off the shelf ARM architecture that they're putting some of their own secret sauce on, but it's not like they're doing Apple style m three creation of a chip from scratch.
这不是他们的明星产品。
It's not the hero product.
没错。英伟达在很多方面都运用了这个原则，我们上期节目也讨论过。如果他们认为某个机会利润率低，就不会追逐。更优雅的说法是：'我们不想参与谁都能做的竞争，只想做唯有我们能做的事。'
Right. There are many ways that NVIDIA sort of applies this where I think we talked about in the last episode. If they think it's gonna be a low margin opportunity, they don't go after it. But the nicer way to say that is, we don't wanna compete for things that anybody can do. We wanna do things that only we can do.
哦对了，当我们做这些事时，会充分实现它们的价值。
Oh, and by the way, we will fully realize the value of those things when we do them.
是的。我认为英伟达在这方面或许有一个相关的策略主题，就是在时机成熟时出击。我猜测过去十到十五年间，黄仁勋和公司内部的大部分竞争动力和动机，实际上是与英特尔对抗。正如我们在前几期节目中多次谈到的，英特尔曾试图扼杀他们。我们采访过一位将英特尔比作乡村俱乐部、英伟达比作战斗俱乐部的人。
Yep. And I think there's maybe a related playbook theme here for NVIDIA of strike when the timing is right. I suspect that a lot of the inner competitive drive and motivation for Jensen and the company over the past ten, fifteen years here has been to really fight against Intel. Intel tried to kill them, as we talked about many times in the previous episodes. We talked to somebody who framed it as Intel was the country club, and NVIDIA is the fight club.
当年，英特尔这个乡村俱乐部不想让英伟达加入。英特尔控制着主板，控制着最重要的CPU芯片。英特尔最终会把所有其他芯片集成并商品化到主板上。如果他们做不到这点，就会尝试自己制造这些芯片。
And back in the days, the Intel country club didn't wanna let NVIDIA in. Intel controlled the motherboard. Intel controlled the most important chip was the CPU. Intel would integrate and commoditize all other chips into the motherboard eventually. And if they couldn't do that, well, then they'd try and make the chips themselves.
他们对英伟达使出了所有这些手段，而英伟达勉强存活下来。后来在数据中心领域，英特尔长期掌控着数据中心。PCI Express作为数据中心互联标准存在了太久，英伟达不得不屈居其中。我敢说他们对此深恶痛绝。但十年前他们并没有立即转身说：猜猜怎么着？
And they tried to run all these playbooks on NVIDIA, and NVIDIA just barely survived. And then in the data center, Intel controlled the data center for so long. PCI Express, you know, that was the interconnect in the data center for so long, and NVIDIA had to live in there. And I'm sure they hated every single minute of it. But they didn't turn around ten years ago and just be like, guess what?
我们也要做CPU了。他们一直等到时机成熟。
We're making a CPU too. They waited until the time was right.
这很疯狂。他们过去必须接入别人的服务器。后来开始制造能接入他人机架、整排设备和架构的服务器。接着又开始打造自己的整排机柜。而现在，他们即将开始运营装满服务器的整栋自建数据中心。
It is crazy. They used to have to plug into other people's servers. And then they started making servers that plugged into other people's racks and rows and architectures. And then they started making their own entire rows and walls. And at some point here, they're gonna start running their own buildings full of servers too.
他们会宣布：我们不需要接入任何外部设备。
And they're gonna say, we don't have to plug in anything.
没错。但我认为对很多其他领导者来说，要像他们这样保持耐心会很难。
Yep. But I think for a lot of other leaders, it would have been hard to have the patience that they've had.
确实。只有那些提前十年布局行业、极具创造力并实现真正突破性创新，且对巨大市场判断极其准确的人，才能做到他们正在做的事。没错。除非你同时满足这三个条件，否则这些都与你无关。
Totally. You only get to do the stuff they're doing if you invested ten years ahead of the industry, were wildly inventive and innovative in creating these, like, true breakthrough innovations, and we're really, really right about huge markets. Yep. None of this stuff applies unless you're doing those three things.
是的。如果刚才你说的那些都不成立，财富500强的CIO们根本不会做出采购决策。
Yeah. Fortune 500 CIOs aren't making buying decisions if none of what you just said isn't true.
对。所以在总结看涨和看跌观点前，我想和你探讨个有趣的话题。回想我们那期关于AWS的节目，我们讨论了很多AWS如何牢牢锁定客户。他们的数据库优势简直坚不可摧。
Right. So there's this interesting conversation I wanted to have with you ahead of winding it up with the bull and bear case. So think back to our AWS episode. We talked a lot about how AWS is just locked in. The databases are a ridiculously durable advantage.
一旦你的数据被迁移到特定云平台——经常是通过装满硬盘的半挂卡车运输
Once your data has been shipped to a particular cloud, often literally in semi trucks full of hard drives
Snowball设备。没错。确实很难
Snowball. Yeah. Is hard to
迁移出去。这里有个耐人寻味的问题：谷歌、微软、亚马逊这些赢下云计算1.0时代的公司，能否凭借这个立足点继续赢得云AI时代？一方面你会觉得当然可以，因为我希望训练AI模型时数据就在旁边。
move off of it. There's this sort of interesting question of will winning Cloud one point o for all these Google, Microsoft, Amazon, will that toehold actually enable them to win in the Cloud AI era? On the one hand, you'd think, yes. Absolutely. Because I wanna train my AI models right next to where my data is.
把数据转移到其他地方进行训练的成本实在太高了。
It's really expensive to move my data somewhere else to do that.
举例来说，微软是OpenAI的独家云基础设施提供商，据我们所知，OpenAI完全运行在NVIDIA的基础设施上，但他们通过微软采购所有设备。
Case in point, Microsoft is the exclusive cloud infrastructure provider for OpenAI, which runs, as far as we know, solely on NVIDIA infrastructure, but they buy it all through Microsoft.
没错。另一方面，客户所要求的体验是完整的NVIDIA堆栈体验，而不是这种——哦，你们找到了最低成本的商品销售方式来提供类似我想要的体验。有时云提供商不得不提供A100或H100给我，因为我的代码过于复杂，无法为任何他们提供的第一方加速计算设备重新架构，那些设备对他们来说更便宜。我不知道。我只是觉得在过去五年左右的时间里，我第一次对这些现有云提供商的护城河稍微侧目，并认为，也许真的存在与他们竞争的途径，云领域并非已成定局。
Right. On the other hand, the experience that customers are demanding is the full stack NVIDIA experience, not this, oh, you found the cheapest possible cost of goods sold way to offer me something that's like the experience that I want. And sometimes the cloud providers have to offer me an a 100 or an h 100 because my code is way too complicated to ever rearchitect for whatever accelerated computing devices they're offering me that's first party and cheaper for them. I don't know. I just think for the first time in the last five years or so, I've sort of cocked my head a little bit at the moat of these existing cloud providers and said, Maybe there really is a vector to compete with them, and cloud is not a settled frontier.
是的。不过这里有点贬义。云是数据中心的委婉说法，对吧？超大规模运营商和公有云远不止数据中心这么简单。
Yeah. Well, this is pejorative here. Cloud is a euphemism for data centers. Right? There's so much more to the hyperscalers and public clouds than just data centers.
对吧？
Right?
但从物理层面来说，它们就是数据中心。
But physically, they're data centers.
没错。比喻上，Equinix和AWS之间有一英里的距离。是的。但它们都是数据中心。而且根据Jensen的说法，数据中心正在发生根本性的转变。
Yeah. There is a mile of distance metaphorically between, like, an Equinix and AWS. Yep. But they're data centers. And there is a fundamental shift, at least according to Jensen, a fundamental shift that is happening in data centers.
所以我认为这确实会造成云市场必须应对的一些变动局面。
So I think that probably does create some shifting sands that the cloud market is gonna have to navigate.
没错。我敢打赌，你在云计算1.0时代的布局将极大决定你在AI云时代的位置。因为说到底，如果客户需要英伟达的产品，云服务商就有十足动力确保你的应用能在他们云端完美运行。
Yep. I bet the way it plays out is that where you landed in cloud one point o strongly dictates where you will land in this AI cloud era. Because at the end of the day, if customers are demanding NVIDIA stuff, then the cloud providers have every incentive in the world to make it so that you can run your applications great in their cloud.
但还不止如此。Crusoe存在，CoreWeave存在，Lambda Labs也存在。这些是资金雄厚的初创公司，拥有数十亿美元投资，许多聪明人都认为这里蕴藏着云服务级别的重大机遇。
But also, like, there's more to this too. Crusoe exists. CoreWeave exists. Lambda Labs exists. These are well funded startups with billions of dollars that a lot of smart people think there's a major cloud sized opportunity for.
确实。这在几年前是不可能发生的。
Yep. That would not have happened a few years ago.
千真万确。好吧，我们来分析看涨和看跌的观点，给这个问题做个了结。
Super true. Alright. Let's do the bull case and bear case and bring this one home.
哦天哪。我们一直在尽可能拖延这个话题。现在这才是问题的关键所在。
Oh, boy. We've been trying to delay this as long as possible. This is the crux of the question right now.
是的。部分问题在于，如果GPU真的成为年产值1000亿美元的市场，他们现有的护城河还够宽吗？目前数据中心GPU市场规模约300亿美元，明年可能增长到500亿。如果真像大家预期的那样发展，这里面的利润空间太大，大公司不可能不重金投入。Meta就砸了数百亿美元搞元宇宙。
Yeah. I mean, part of it is, is their existing moat big enough if GPUs actually become a $100,000,000,000 a year market? I mean, right now, GPUs in the data center are like a $30,000,000,000 a year market going to like a $50,000,000,000 next year. And, like, if this actually goes the way that everyone seems to think it's gonna go, there's just too many margin dollars out there for these big companies to not invest heavily. Meta threw tens of billions of dollars making the metaverse.
苹果据传为头显投入了150亿美元。亚马逊在设备上砸了数百亿，按说这笔投资糟透了。Echo音箱有什么回报可言？
I mean, Apple's put $15,000,000,000 into rumored into their headset. Amazon's put tens of billions of dollars into devices, which by all means was a terrible investment. How is Echo paying anything back?
天啊，完全跑题了。我太失望了。我家全用Echo生态，但它越来越蠢。在这个AI能力飞速进步的世界里，我的Echo怎么反而变笨了？
Oh, man. Total sidebar. I'm so disappointed. I have standardized my house on the Echo ecosystem, and it keeps getting dumber. How in this world of incredibly accelerating AI capabilities are my echoes getting dumber?
呃，他们需要在Trainium和Inferentia芯片上再加把劲。
Well, they need to trainium and Inferentia a little bit harder.
老天。好吧。吐槽完毕。
Jesus. Okay. Rant over.
是啊。我是说，千万别低估科技巨头们为潜在巨大回报砸下数百亿的能力。这些都是利润离谱的垄断企业——除了亚马逊其实没那么赚钱。
Yeah. I mean, never doubt Big Tech's ability to throw tens of billions of dollars into something if the payoff could be big enough. These are ludicrously profitable monopolies, except for Amazon's not that profitable.
AWS是赚钱的。
AWS is.
对。但谷歌、脸书、苹果这些公司，迟早会结束这场胆小鬼博弈。有些公司会押上全部筹码说：我们也有聪明工程师，我们也能搞定这事。
Yeah. But Google, Facebook, Apple, at some point here, there's a game of chicken that ends. And some of these companies go all in and say, yeah, we have smart engineers too. Like, we're gonna figure this out.
没错。但也别低估科技巨头对自以为能实现的事情——尤其是重大战略转型——的执行无能。
Yeah. But also never underestimate the inability of big tech to execute on stuff that it thinks it can, especially with major strategy shifts.
好的。好的。行吧。那我们正式开始。熊市情景。
Yeah. Yeah. Alright. So let's actually do this. Bear case.
我们先从熊市情景开始。
Let's start with the bear case.
所以你刚才阐述的，我认为是第一个熊市情景，即整个科技生态系统中其他所有人现在都达成一致并有动机说，我要从英伟达的蛋糕中分一杯羹。而这些公司拥有难以计量的资源。
So you just illustrated, I think, bear case number one, which is literally everybody else in the technology ecosystem is now aligned and incentivized to say, I wanna take a piece of NVIDIA's pie. And these companies have untold resources.
没错。更准确地说，让我们看看PyTorch。既然现在所有或许多开发者都在使用PyTorch，这确实能让PyTorch聚合客户，从而给他们提供去中介化的机会。也许吧。但需要在底层编写大量新代码并交付大量硬件。
Yep. And to put a finer point on that, let's look at PyTorch for a minute. Now that all the developers or lots of developers are using PyTorch, it does enable PyTorch to aggregate customers, which gives them the opportunity to disintermediate. Maybe. You gotta write a lot of new stuff underneath and ship a lot of hardware.
我是说，云服务提供商在这方面已经采取了一些措施。它最初是由Meta开发的。虽然是开源的，但如果它实际上由Meta掌控，其他公司很难投入其中。所以现在PyTorch已经转移到一个由多家公司共同贡献的基金会。而且，把PyTorch和英伟达相提并论完全是错误的类比。
I mean, the cloud service providers have taken some steps here. It was originally developed by Meta. And while it's open source, it's still hard for all these companies to invest in it if it's really sort of owned and controlled by Meta. So now PyTorch has been moved out into a foundation that a lot of companies are contributing to. And, it is a absolute false equivalence to be like PyTorch versus NVIDIA.
但用本·汤普森的聚合理论术语来说，如果你聚合了客户，就有机会获取更多利润空间、实现去中介化、引导注意力流向。PyTorch就有这样的机会。这似乎是许多云服务提供商试图竞争的维度，他们会说：看，如果你为PyTorch开发，在我们的平台上也能运行得很好。是的，确实如此。
But in real Ben Thompson aggregation theory parlance, if you aggregate the customers, you have the opportunity then to take more margin, to disintermediate, to direct where that attention is going. And PyTorch has that opportunity. That feels like the vector that a lot of these CSPs will try and compete on and say, look, if you're building for PyTorch, it runs really well on our thing too. Yep. For sure.
毫无疑问这种情况会发生。好吧。所以这算是第一个熊市情景下的第二个熊市情景。
No doubt that that's gonna happen. Alright. So that's bear case number two kinda as part of bear case number one.
下一个问题是，市场实际规模并不像市值反映的那么大。我认为未来12到18个月内很可能会出现投资者信心危机，届时我们会发现GPT或许没有想象中那么实用，人们可能并不需要这种扑面而来的聊天功能。这种信心危机，这种小型泡沫破裂会逐渐影响到美国企业的首席信息官和首席执行官们，使得在董事会中推动这项我们今年已达成共识、而我正试图提议改变的重大基础采购和全面预算重构变得更加困难。就像加密货币热潮退去那样，某些公司的支出将会因此放缓。
The next one is, like, literally, the market isn't as big as the market cap reflects. I think there's a pretty reasonable chance that there's some falter in the next twelve to eighteen months where there's a crisis of confidence among investors where at some point, something will come out where we all observe, oh, maybe GPTs aren't as useful as we thought. Maybe people don't want chat in their faces. And that crisis of confidence, that mini bubble burst will trickle out to America's CIOs and CEOs, make it harder to advocate in the boardroom to make this big fundamental purchase and rearchitecture of our whole budget from this year that we agreed on that I'm trying to propose us changing. There's a crypto like element to a excitement bubble bursting that will, for some companies, slow their spend.
问题在于这种情况何时会发生——因为这并非是否会发生的问题，而是何时会发生的问题。鉴于当前围绕AI的所有炒作，我很难相信它会比大众预期的更有用，并且会毫无回调地线性发展下去，让所有人的热情只增不减。它最终可能远超预期，但总会在某个时间点出现低谷。关键在于英伟达如何在这场信心危机中应对。
And the question is sort of like, when that happens because it's not an if, it's a win. I have a hard time believing that given all the hype around everything right now, AI will be even more useful than everyone believes, and it will continue in a linear fashion where without any drawdowns, everyone's excitement only gets bigger from here. It may end up being way more useful than anyone thought, but there, at some point, will be some valley or trough. And it's sort of about how does NVIDIA fare during that crisis of confidence.
有趣的是，我们为这期节目采访了很多人，包括一些顶尖的AI研究者、实践者，以及正在从事相关工作的公司创始人和高管团队。几乎所有人都对这个问题的回答如出一辙：'没错，现在确实存在过度炒作。'
It's funny. You know, again, we talked to a lot of people for this episode, including a set of some of the foremost AI researchers and practitioners out there and founders and c suites of companies that are doing all this. And pretty much to a tee, they all said the same thing when we asked them about this question. They all said, yeah. This is overhyped right now.
当然，这显而易见。但以十年为尺度来看，你还没见识到真正的变革。我们相信即将到来的颠覆性改变，其程度远超你现在的想象。
Of course. Obviously. But on a 10 time scale, you haven't seen anything yet. The transformative change that we believe is coming, you can't even imagine.
关于过度炒作最耐人寻味的是，它已经切实体现在营收上。所有购买这些算力资源的人都抱有某种信念。对英伟达而言，由于这种信念以营收形式呈现，它就是真实的。因此他们只需确保能平稳渡过这个阶段，直到客户实际获得的价值与全球CIO们当前的前瞻性投资相匹配。
The most interesting thing about the overhype is that it's actually showing up in revenue. It's everyone who is buying access to all this compute believes something. And for NVIDIA, because it's showing up in the form of revenue, the belief is real. Then so they just need to make sure that they smooth the gap to customers actually realizing as much value as the CIOs of the world are currently investing ahead of.
没错。我认为现在值得讨论的子话题是：生成式AI真的配得上所有赞誉吗？
Yep. So I think the sub point to that that's worth the discussion right now is like, okay. Generative AI. Yeah. Is it all it's cracked up to be?
大卫，我大概有一个月没问你这个问题了。但一个月前你还拍着桌子坚称：'我根本不需要用ChatGPT，从没觉得它有用，总是胡言乱语，我根本想不起来用它。'
Well, David, I haven't asked you about this in, like, a month or so. But a month ago, you were pounding the table insisting to me, like, I have no need for I've never used ChatGPT. I can't find it to be useful. It's hallucinating all the time. I never think to use it.
这不在我的工作流程中。话说，你现在进展到哪一步了？
It's not a part of my workflow. Like, where are you at?
基本上还在原地踏步，包括强迫自己为了这期节目大量试用它。但与此同时，随着我们采访更多人，我意识到大卫·罗森塔尔的用例在这里根本不重要。首先，作为一家企业，我们是高度专业化、独特的小众存在，节目内容的精确性和我们投入的思考深度才是重中之重。
Still basically there, including forcing myself to try to use it a bunch in preparation for this episode. But also, as we talk to more people, I think I've realized that, like, David Rosenthal's use case doesn't really matter here at all. Right. A, because as a business, we are such a hyper specialized, unique little unicorn thing where accuracy and the depth of the work and thought that we ourselves put into episodes is the paramount thing.
而且我们没有同事。我们业务的许多方面都很奇怪，比如从不需要为会议准备简报。
Well, and we have no coworkers. There's so many things about our business that is weird. Like, we never have to prepare a brief for a meeting.
没错。所有这些外部准备的材料都是出于热爱。而我们内部根本不需要准备任何文件。
Right. All this stuff. Anything external that we prepare is a labor of love for us. And there is nothing we prepare internal.
我认识有人用ChatGPT来制定OKR。我问什么是OKR？他们说真希望自己生活也能这样，所以才让ChatGPT代劳。
I know people who use ChatGPT to set their OKRs. And I'm like, okay. What's an OKR? And they're like, I wish my life were like that too. That's why I have ChatGPT do it.
老实说，通过实践交流和阅读，我认为它在编写代码方面确实有极具说服力的应用场景。无论你是哪个级别的开发者，从零基础到精英程序员，都能通过GitHub Copilot获得远超以往的效率。这难道没有价值吗？当然有价值。
Right. Honestly, like, I think through doing this and talking to some folks and reading, I think there's a very compelling use case for it for writing code right now. No matter what level of software developer you are, from zero all the way up through elite software developer, you can get a lot more leverage out of this thing in GitHub Copilot. So is that valuable? Like, for sure, that's valuable.
没错。大语言模型在编写和辅助编写代码方面表现出色，我完全相信这个应用场景。
Yeah. The LLMs are unbelievably good at writing and helping you write code. I'm a huge believer in that use case.
是的。然后我认为，你知道，还有一些更具推测性的内容，但你现在其实可以隐约看到，比如我最近提到的NVIDIA那个游戏演示，哦，你在和一个非脚本化的非玩家角色对话。我们最近和Runway CEO克里斯·瓦伦苏埃拉做了两期ACQ节目，他们的技术被用在《瞬息全宇宙》里。他说那只是冰山一角。
Yep. And then I think, you know, there's the slightly more speculative stuff, but you can actually sort of see it now of, like, that gaming demo that I mentioned recently from NVIDIA of, oh, you're talking to a non playable character that wasn't scripted. We did an ACQ two episode recently with Chris Valenzuela from the CEO of Runway. That was used in everything everywhere all at once. And he said, that's just the tip of the iceberg.
就像，目前在生成式AI领域已经可以实现的事情，正在现实中发生的那些
Like, the stuff that you can do that is happening that's out there today with generative AI in these domains is
令人震惊。没错。我想你说的是人们可能会基于自身经历持悲观态度。每次尝试使用生成式AI应用时，它都无法融入你的工作流程，你觉得它没什么用。
astounding. Yeah. I think what you're saying is one could be a bear on your own experience. Every time you try to use a generative AI application, it doesn't fit into your workflow. You don't find it useful.
用户粘性不足。但另一方面，实际上AI将成为众多细分领域的总和。有游戏市场、写作市场、创意写作市场，
You're not sticky. But on the other hand, actually, what AI will be is a sum of a whole bunch of niches. There's a video game market. There's a writing market. There's a creative writing market.
有软件开发市场、营销文案市场。你知道，这类细分领域数不胜数，而你恰好可能不属于前几个受益领域。
There's a software developer market. There's a marketing copy market. You know, there's a million of these things, and you just may happen to not fall into one of the first few niches of it.
没错。至少对我个人而言，我最初抱有强烈的怀疑态度，因为时机实在太完美了。懂吗？就像那些UBC们刚向所有人宣扬加密货币是未来之类的话，
Yep. I think for me, at least, again, just speaking personally too, I had a very strong element of skepticism initially because the timing was just too perfect. You know? It was like Yes. All UBCs out there.
结果利率就涨到了5%，你的世界瞬间崩塌。
You just told everybody about how crypto is the future and whatever you're talking about, and then interest rates went to, you know, 5% and your world fell off a cliff.
唉。那些在外面筹集资金的人，动不动就说未来是AI的时代。
Ugh. The number of people who were, like, out raising a fund, and they're like, the future is AI.
是啊，没错。
Yeah. Right.
现在绝对是投资的最佳时机。而且
This is the best time ever to be investing. And
所以我当时很大一部分想法就是：拜托，各位。
so there was a large part of me that I was just like, come on, guys.
没错，太完美了。你说得对。
Yeah. It's too perfect. You're right.
确实太完美了。但最近这几个月英伟达的表现说明——抛开那些不谈——财富500强企业正在采用这些技术，CIO们正在采用这些技术，英伟达正在赚取真金白银。
It's too perfect. But this most recent couple months in this quarter for NVIDIA has shown that put all that aside. Fortune five hundreds are adopting this stuff. CIOs are adopting this stuff. NVIDIA is selling real dollars.
同时也了解到训练这些模型所需的成本，以及知识效用的阶梯式增长——从十亿参数到百亿参数，再到两千亿乃至万亿参数模型。没错，这里面确实正在发生些什么。
And learning also about what it takes to train these models and the step scale function of knowledge and utility going from a billion parameters to 10,000,000,000 parameters to 200 to a trillion parameter models. Yeah. Like, something's going on there for sure.
这让我想到下一个看跌观点：模型性能将足够成熟，届时所有训练完成，我们将转向推理阶段。而大部分计算负载将集中在推理环节，英伟达在该领域的差异化优势较弱。我对此持怀疑态度有多重原因，尽管这是主流论调。一个重要原因是——Transformer架构绝非技术演进的终点。
So this leads me to my next bear case, which is the models will get good enough, and then they'll all be trained, and then we'll shift to inference. And most of the compute load will be on inference where NVIDIA is less differentiated. There's a bunch of reasons I don't believe that. That is a popular narrative, though. One of the big reasons I don't believe that is the transformer is not the end of the road.
大卫，在我们的大量研究中可以清晰看到，目前已有许多超越Transformer架构的技术处于研究阶段。用户体验只会更加神奇高效。因此还存在第二个看跌观点：当前我们采用暴力穷举式方案处理这些任务，相关收益全归英伟达所有，因为只有他们能提供这种'暴力方案'。但随着时间的推移，比如观察谷歌的Chinchilla或Llama2，它们用比GPT-4更少的参数实现了相当的质量（当然评判标准见仁见智），这些高质量模型都证明了参数精简的可能性。所以潜在看跌论点在于：未来模型会更智能且降低算力需求。
In a bunch of the research that we did, David, it's very clear that there are things beyond the transformer that are in the research phase right now. And the experiences are only gonna get more magical and only gonna get more efficient. So there's sort of a second bear case there, which is right now, we threw a brute force kitchen sink at trading these things, and all of that revenue accrued to NVIDIA because they're the ones that make the kitchen sinks. And over time, like, you look at Google's Chinchilla or Llama two, they actually use less parameters than GPT four and have equivalent quality or, you know, many other people can be the judge of that, but were high quality models with less parameters. So there is this potential bear case around future models will be more clever and not require as much compute.
需要指出的是，即便在今天，绝大多数AI工作负载仍与LLM不同（至少直到最近才改变）。LLM堪称人类历史上需要海量算力的任务巅峰。问题在于：这种趋势会持续吗？许多近期出现的惊艳AI体验（如扩散模型和整个图像生成领域）都通过更低成本的模型训练实现——这期节目我们讨论较少正是因为它们算力需求较低。事实上，许多任务根本不需要全网训练数据和万亿参数就能完成。
It's worth saying that even today, the vast majority of AI workloads don't look like LLMs, at least until very recently. LLMs are like the current maxima in human history of jobs to be done that require a ton of compute. And I guess the question is, will that continue? I mean, many other magical recent AI experiences have happened with far less expensive model training, like diffusion models and the entire genre of generative AI on images, which we really haven't talked about a lot on this episode because they're less compute intensive. But many tasks don't require an entire Internet of training data and a trillion parameters to pull off.
确实，这个观点很有道理。我认为工作负载向推理转移也有其合理性，这个趋势确实正在发生，我同意你的看法。
Yep. That makes sense to me. And I think there also is some merit to workloads are shifting to inference. That is happening. I agree with you.
我不认为训练环节会消失。但回想谷歌时代，直到最近训练才是资金投入的重点，也是所有人的关注核心。随着应用规模扩大，推理环节——即模型训练完成后为获取输出所需的计算——自然会占据更大比重。
I don't think training is going anywhere. But until recently, you know, thinking back to the Google days, training was what everybody was spending money on. That's what everybody was focused on. As usage scales with this stuff, then inference. And inference, of course, being the compute that has to happen to get outputs out of the models after they're already trained, that becomes a bigger part of the pie.
正如你所说，支撑推理的基础设施和生态系统确实比训练环节更缺乏差异化。好的，这些是主要看跌观点。可能还存在关于中国的看跌因素，这个确实值得关注，因为将对许多企业造成困扰。
And as you say, the infrastructure and ecosystems around doing that is less differentiated than training. Yep. Okay. Those are the bear cases. There's probably also a bear case around China, which is a legitimate one because that's gonna be a problem for lots of people.
在可预见的未来，他们都无法实质性地开拓这个巨大市场。
A large market that they won't be able to address for the foreseeable future in a meaningful way.
那么整体上会发生什么呢？显然，中国正在加速发展本土生态系统和竞争对手，这将形成一个封闭的市场。那么会有什么结果呢？究竟会发生什么？
And just what's gonna happen generally. Like, obviously, China is racing to develop their own homegrown ecosystems and competitors, and, like, that's gonna be a closed off market. So what's gonna come out of there? What's gonna happen?
没错，这绝对也是一个因素。我最后要说的是看跌观点，但结果证明并非如此。对大多数公司来说，如果他们以如此高的市盈率交易，并且刚刚经历了收入和营业利润的巨大实际增长，当这种系统冲击消失时，公司会在业务放缓时遭受无法弥补的损害。股票薪酬是个问题。
Yep. That's definitely one too. My last one is a bear case, but it ends up not being a bear case. For most companies, I would say that if they were trading at this very high multiple and they just experienced this tremendous real growth in revenue and operating profit, that that sort of spike to the system when it goes away will irreparably harm the company when things slow down. Stock compensation's an issue.
员工士气是个问题。客户认知是个问题。但这是英伟达。
Employee morale is an issue. Customer perception's an issue. But this is NVIDIA.
是啊，这没什么新鲜的。
Yeah. This is nothing new.
他们经历过多少次在长期低迷后，凭借令人震撼的创新浴火重生——当这种情况发生时，他们可能是最有能力应对的公司，或者说具备最佳应对特质的公司。
The number of times that they've risen from the ashes after, you know, years long terrible sentiment with something mind blowingly innovative, they're probably the best positioned company or the company with the best disposition to handle that when it happens.
哦，我觉得这个说法很精彩。你在语言表达上提升了训练模型的水准。
Oh, I thought that's a great turn of phrase there. You, upped your training model on the language there.
你真该看看参数量有多大。
You should see the number of parameters.
我很喜欢。
I love it.
好的，我来列举几个主要观点。第一，Jensen在加速计算方面说得没错。目前大多数工作负载并未加速，它们仍依赖于CPU。
Alright. Just to list the bulk cases. One, Jensen is right about accelerated computing. The majority of workloads right now are not accelerated. They're bound to CPUs.
这些工作负载是可以加速的，这将使当前仅5%到10%的加速工作负载比例在未来提升至50%以上。并行计算量会大幅增加，而NVIDIA将从中获益最多。
They could be accelerated, and that shifts from some crazy low number, like five or 10% of workloads being accelerated today to 50 plus percent in the future. And there's way more compute happening in parallel, and that mostly accrues to NVIDIA.
哦，我想补充一个细节。表面上很多人会觉得这种说法很夸张，但在生成式AI领域和我们本期讨论的内容中，这个观点其实很有道理。Jensen和NVIDIA并不是说传统计算会消失或萎缩，而是说AI计算将叠加到所有领域，其所需计算量将远超通用计算。
Oh, I have one nuance I wanna add to that. On the surface, I think a lot of people look at that, and they're like, yeah, come on. But I think there actually is a lot of merit to that argument in the generative AI world and everything we've talked about on this episode. I don't think Jensen and Nvidia are saying that traditional compute is going away or getting get smaller. I think what he's saying is that AI compute will be added onto everything, and the amount of compute required for doing that will dwarf what's happening in general purpose compute.
所以并不是说人们会停止使用SharePoint服务器，或者产品会停用现有接口。而是生成式AI将融入所有这些领域，催生出新用例——这些用例仍会使用传统的CPU通用计算，但实现智能功能所需的工作负载规模将会大得多。
So, like, it's not that people are gonna stop running SharePoint servers or that whatever products you use are gonna stop using their whatever interfaces that they use. It's that generative AI will be added to all of those things, and the use cases will pop up, which will also use traditional general purpose CPU based computing. But the amount of workloads that go into making those things magical is just gonna be so much bigger.
没错。另外关于软件开发有个普遍现象：编写可并行化代码非常困难，除非有现成框架。即便是多线程编程——比如计算机科学课上遇到的竞态条件或信号量问题——这些都是最难调试的。很多本可加速的领域进展缓慢，开发难度高是主要原因之一。
Yep. Also, just a general statement on software development. Writing parallelizable code is really hard unless you have a framework to do it for you. Even writing code with multiple threads, like if anybody remembers a CS college in class where they had a race condition or they needed to write a semaphore, these are the hardest things to debug. And I would argue that a lot of things that could happen in an accelerated way aren't just because it's harder to develop for.
如果我们未来生活在NVIDIA重构计算机概念的世界里——从冯·诺依曼架构转向他们开发的流处理器架构，并且拥有让应用开发和迁移变得同样简单的全栈方案。特别是当所有硬件都已部署在数据中心后，只要操作足够简便，确实会有大量工作负载值得加速。
And so if we live in some future where NVIDIA has reinvented the notion of a computer to shift away from von Neumann architecture into this stream processor architecture that they've developed, and they have the full stack to make it just as easy to write applications and move existing applications. Especially once all the hardware has been bought and paid for and sitting in data centers, there's probably a lot of workloads that actually do make sense to accelerate if it's easy enough to do so.
是的，这很棒。但你的观点是，市面上存在大量潜在的、可加速的计算资源尚未被加速利用。
Yeah. That's great. But so your point is that there's a lot of latent accelerated addressable computing out there that just hasn't been accelerated yet.
没错。就像这个工作负载并不那么昂贵，我也不打算花钱请工程师重构系统。所以维持现状就好。
Right. It's like, this workload's not that expensive, and I'm not gonna pay an engineer to go rearchitect the system. So it's fine how it is.
你觉得怎么样？
How about that?
我认为这种情况很普遍。看涨理由一：Jensen关于加速计算的观点是正确的。看涨理由二：Jensen关于生成式AI的观点是正确的。结合加速计算，这将使数据中心支出大规模转向英伟达硬件。正如我们提到的，传闻OpenAI通过ChatGPT实现了超10亿美元的经常性收入。
I think there's a lot of that. So bull case one, Jensen is right about accelerated computing. Bull case two, Jensen is right about generative AI. I mean, combined with accelerated computing, this will massively shift spend in the data center to NVIDIA's hardware. And as we've mentioned, OpenAI is rumored to be doing over a billion dollars in recurring revenue on ChatGPT.
所以我估计大概有30亿美元——这是我听过最可信的估算。这可能是明年的预测。但他们不是唯一玩家，比如谷歌的Bard——我在准备这期节目时发现它非常实用——虽然没直接变现，但通过这种方式留住了我这个谷歌搜索用户。即便现在也创造了大量实际经济价值。
So I think there's let's call it 3,000,000,000 because that's the most sort of credible estimate that I've heard. And maybe that was a forecast for next year. But, like, they're not the only one. I mean, Google with Bard, which I found tremendously useful actually preparing for this episode, is not directly monetizing that, but they're sort of retaining me as a Google search customer by doing it. There is a lot of real economic value even today.
虽然远不及估值中隐含的预期，但最保守的情况是英伟达必须事事顺利。而看涨迹象表明事情正在朝有利方向发展。第三点看涨理由：英伟达行动极其迅速。无论出现什么技术发展，很难想象他们不会找到方法占据优势。
Not nearly the amount that sort of baked into the valuation, but I suppose the bare case of this is that everything has to go right for NVIDIA. But the bull cases, indications are things are going right for NVIDIA. Yep. Third, bull cases, NVIDIA just moves so fast. Whatever the developments are, it's hard to believe that they're not gonna find a way positioned to capture it.
这就是企业文化使然。第四点是你早前提到的：数据中心现有万亿级设备规模，每年还有2500亿美元用于更新扩容，英伟达可能夺取可观份额。他们现在年收入是多少？大概300亿美元？
That's just a cultural thing. Four is the point that you brought up earlier that there's a trillion dollars installed in data centers, 250,000,000,000 more being spent every year to refresh and expand capacity, and that NVIDIA could take a meaningful share of that. I think today, what's their annual revenue at? Like, 30,000,000,000 or something?
嗯，如果你按当前季度年化计算的话，大概是50.5美元左右。对。
Well, if you run rate this current quarter, then it's at, like, $50.50. Plus. Yeah.
所以目前来看，他们约占数据中心当前支出的20%。你可以想象这个比例会更高。
So right now, that puts them at, like, 20% of the current data center spend. You could imagine that being much higher.
好的等等。这包括游戏收入。大约是40因为数据中心收入40是10。所以年化40。
Okay. Wait. That includes the gaming revenue. It's about 40 because the data center revenue is 40 is 10. So 40 annualized.
明白了。所以18%。对。哇。但你可以想象这个比例会逐渐上升。
Alright. So 18%. Yeah. Whoo. But you could imagine that creeping up.
再说一次，如果加速计算和生成式AI的预期成真，他们会扩大那个250的数字，并从中获取更大份额。
Again, if the accelerated computing and generative AI belief comes true, like, they'll expand that two fifty number, and they'll take a greater percent of it.
是的。
Yep.
验证这个计算有个有趣的方法，就是看生态系统中其他公司在财报中披露的数据。台积电上季度财报显示，AI硬件目前仅占其收入的6%。但所有迹象表明，他们预计AI收入未来五年将以每年50%的速度增长。哇。所以我们试图从客户工作负载的角度来看，问：这在那里有用吗？
An interesting way to do a sort of a check on this math is to look at what other people in the ecosystem are reporting in their numbers. TSMC, in their last earnings, said that AI hardware currently only represents 6% of their revenue. But all indications over there is that they expect AI revenue to grow 50% per year for the next five years. Wow. So we're trying to come at it from the customer workload side and say, is it useful there?
但如果你从另一个角度思考，英伟达的供应商们在预测什么？他们必须言行一致，投资建设新的晶圆厂来实现这一目标，包括封装和芯片制造的所有环节。所以台积电一旦出错代价高昂。没错，这又是一个看涨的理由。
But if you come at it from this other side of what do NVIDIA suppliers forecasting? And they have to put their money where their mouth is building these new wafer fabs to be able to facilitate that, and packaging and all the other things that go into chips. So it's expensive for TSMC to be wrong. Yep. So that's another bull case.
在我离开前还有最后一个想法要分享。
The last one that I have before leaving you with one final thought.
你是说你还有一件事要说？
Are you saying you have one more thing?
是的。那就是英伟达不是英特尔。我认为这是你帮我意识到的最大启示。
Yes. Is that NVIDIA isn't Intel. And I think that's the biggest realization that you helped me have.
它也不是思科。
And it's not Cisco.
没错。我们上期节目做的类比是错误的。他们是微软。他们控制着整个软件栈，同时能与开发者和客户生态系统建立关系。而且，我的意思是，可能比微软更胜一筹，因为他们还制造所有硬件。
Yeah. The comparison we were making in the last episode was wrong. They are Microsoft. They control the whole software stack, and they simultaneously can have relationships with the developer and customer ecosystems. And, I mean, it may even be better than Microsoft because they make all the hardware too.
是啊，或许像老牌IBM。
Yeah. Maybe old school IBM.
没错。想象一下如果IBM在当今规模的计算机市场运营。那时候计算机市场还非常小。
Right. Imagine if IBM operated in a computing market of today's magnitude. Computing was tiny little market back then.
对。我是说，确实如此。PC浪潮才真正颠覆了IBM的统治地位——用现在的话说就是个人电脑、边缘计算、设备端计算。IBM当时主导了企业级主机计算周期。如果你相信黄仁勋过去五年所说的一切以及他如何引领公司，我们正在回归一个由现代版数据中心主导的主机计算周期。
Right. I mean, was like that. I mean, it took the PC wave to disrupt IBM, which was a personal computer in today's parlance, edge computing, you know, device based computing. IBM dominated the b two b mainframe cycle of computing. And again, if you believe everything Jensen is saying and how he steered the company for the last five years, we are going back into a centralized data center modern version of a mainframe dominated computing cycle.
是的。
Yep.
我推测大量推理计算会在边缘端完成。想想我们口袋里那些尚未充分利用的强大算力吧。未来手机会处理大量机器学习任务，遇到复杂问题再调用云端模型。
I suspect a lot of inference will get done on the edge. You think about the insane amount of compute that's walking around in our pockets that is not fully leveraged right now. There's gonna be a lot of machine learning done on phones that are gonna, like, call up to cloud based models for the hard stuff.
毫无疑问。但我认为训练环节短期内不会发生在边缘端。
No doubt. I don't think training is happening at the edge anytime soon, though.
确实。我完全同意。就像我们讨论台积电那期节目，David，我想最后留给你一个思考：要与英伟达竞争需要什么？因为台积电那集给我的最大启示是：哇，这需要相信政府会投入数十亿美元并招募所有人才。
No. I certainly agree with that. Alright. Well, just like our TSMC episode, I wanted to end and leave you with a thought, David, of what it would take to compete with NVIDIA. Because my big takeaway from the TSMC episode was like, wow, that's a lot of things you have to believe about a government putting billions of dollars in and hiring all this talent.
我不禁思考：对标英伟达需要什么？假设你能设计出同等水平的GPU芯片——AMD、谷歌和亚马逊正在这样做。你还需要建立芯片间互联技术（比如NVLink），这领域玩家极少。当然还得和富士康这样的硬件组装商合作，把这些芯片集成到DGX这样的服务器里。
And I was like, what's the equivalent for NVIDIA? So here's what you would need to do to compete. Let's say you could design GPU chips that are just as good, which arguably AMD, Google, and Amazon are doing. You'd, of course, then need to build up the chip to chip networking capabilities like NVLink that very few have. And you'd, of course, need to build relationships with hardware assemblers like Foxconn to actually build these chips into servers like the DGX.
即便你完成了所有那些，你还需要打造能与Mellanox匹敌的服务器间和机架间网络能力——Mellanox曾是InfiniBand市场的霸主，现已被英伟达全资控股，这种技术几乎无人能及。就算你做到了这些，你还得说服所有客户购买你的产品，这意味着你的产品必须比英伟达更优、更便宜或两者兼具，而非仅仅持平。
And even if you did all that, you'd need to create server to server and rack to rack networking capabilities as good as Mellanox, who was the best on the market with InfiniBand that NVIDIA now fully owns and controls, which basically nobody has. And even if you did all that, you'd need to go convince all the customers to buy your thing, which means it would need to be either better or cheaper or both, not just equal to NVIDIA.
而且必须远超这个品牌才行，短期内没人会因采购英伟达而被解雇。这就是行业标准——要想说服CIO，你必须在这方面比英伟达强十倍。
And by a wide margin too to this brand, you're not gonna get fired for buying NVIDIA anytime soon. Like, this is the canonical. You gotta be 10 x better than NVIDIA on this stuff if you're gonna convince a CIO.
没错。即便获得了客户需求，你还需要与台积电签约，获得其最新尖端晶圆厂的制造能力来实现2.5D COAS光刻和封装技术——当然这些产能早已被抢占。所以祝你好运吧。就算你搞定了制造，还需要开发出媲美或超越CUDA的软件。这显然需要上万人工年的投入，不仅耗资数百亿美元，更要耗费大量时间。
Yep. And even if you got the customer demand, you'd need to contract with TSMC to get the manufacturing capability of their newest cutting edge fabs to do this 2.5 d COAS lithography and packaging, which there, of course, isn't anymore of. So, you know, good luck getting that. And even if you figured out how to do that, you'd need to build software that is as good or better than CUDA. And, of course, that's gonna take ten thousand person years, which would, of course, cost you not only billions and billions of dollars, but all that actual time.
即便你完成所有投资并协调好一切，当然还需要说服开发者转而使用你的产品而非CUDA。而英伟达也不会坐以待毙，因此你必须以破纪录的速度完成所有这些，不仅要追赶他们，还要超越他们在此期间新开发的所有能力。所以核心结论是：正面竞争几乎不可能。未来若有人能取代英伟达在AI和加速计算领域的地位，要么来自他们未察觉的侧翼突袭，要么就是AI领域最终不需要加速计算——后者可能性微乎其微。
And even if you made all these investments and lined all of this up, you'd, of course, need to go and convince the developers to actually start using your thing instead of CUDA. Well, NVIDIA also wouldn't be standing still, so you'd have to do all of this in record time to catch up to them and surpass whatever additional capabilities they developed since you started this effort. So I think the bottom line here is nearly impossible to compete with them head on. And if anybody's gonna unseat NVIDIA in the future of AI and accelerated computing, it's either gonna be from some unknown flank attack that they don't see, or the future will turn out to just not be accelerated computing in AI, which seems very unlikely.
确实。这么说来，我们只能得出和马克·安德森相同的结论了。我们讨论的是哪一年来着？
Yeah. Well, when you put it that way, I think the conclusion that we can come to is that Marc Andreessen was right. In what years was this that we're talking about on
大概是2015年左右吧。
It was like 2015 or something.
对，2015或2016年。
Yeah. Like 2015, 2016.
他们本应该把16z筹集到的每一分钱都投入到英伟达的股票每日市价中。
They should have put every dollar of every fund that a 16 z raised into NVIDIA's market price of the stock every single day.
没错。因为他们当时看到所有这些初创公司都在做深度学习、机器学习，早期的AI。而这些都建立在英伟达的基础上。他们本应该对所有这些公司说不，然后把钱全投给英伟达。马克又一次说对了。
Yeah. Because they were seeing all of these startups doing deep learning, machine learning at the time, early AI. And they were all building on NVIDIA. And they should've just said, no, thank you to all of them and put it all in NVIDIA. Mark is right once again.
强者愈强。就是这样。
Strength leads to strength. There you go.
就是这样。听众朋友们，我承认这期节目概括了很多细节，特别是对技术背景的听众，也包括金融领域的听众。我们的目标是让这期节目成为更具持久性的《英伟达三部曲》宏观视角特辑，而不是讨论他们上季度表现如何，以及对接下来三个季度的影响。希望这期内容能比单纯的英伟达时事评论更有生命力。非常感谢你们和我们一起踏上这段旅程。
There it is. Well, listeners, I acknowledge that this episode generalized a lot of the details, especially for technical listeners out there, but also for the finance folks who are listening. Our goal was to make this more of a lasting NVIDIA part three big picture episode than sort of how did they do last quarter, and what are the implications on that of the next three quarters. So hopefully, this holds up a little bit longer than just some current NVIDIA commentary. But thank you so much for going on the journey with us.
是的。正如我们在节目中提到的，我们要向许多热心帮助过我们的人致谢，包括那些本可以把时间花在更重要事情上的人。我们真的非常非常感激。
Yep. We also, as we've alluded to throughout the show, we owe a bunch of thank yous to lots of people who were so kind to help us out, including people who have way better things to do at their time. So we're very, very grateful.
比如，首先要感谢英伟达的伊恩·巴克，他负责数据中心业务，也是当年发明CUDA的原始团队成员之一。非常感谢他为准备这期节目接受我们的采访。
I mean, one, Ian Buck from NVIDIA, who leads the data center effort and is one of the original team members that invented CUDA way back when. Really grateful to him for speaking with us to prep for this.
绝对要感谢。还要特别鸣谢节目听众朋友ABC数据的杰里米，他主动为我们准备了四份PDF？完全自发地。简直是为我们疯狂整理了大量技术细节的私人博客文章。
Absolutely. Also, big shout out to friend and listener of the show, Jeremy from ABC Data, who prepared Four PDFs for us? Completely unprompted. Like, an insane write up for us about a lot of the technical detail behind this. Private blog posts.
是的。私人博客文章。我们收购的社区简直是最棒的。你们总是让我们惊艳。非常感谢。
Yeah. Private blog posts. So our acquired community is just the best. Like, you guys continue to blow us away. So thank you.
Hugging Face的首席技术官Julian，AI two的Orinizioni，OctoML的Luis，当然还有我们在NZS Capital的朋友们。感谢大家协助这项研究。
Julian, the CTO of Hugging Face, Orinizioni from AI two, Luis from OctoML, and of course, our friends at NZS Capital. Thank you all for helping us research this.
确实。好吧。拆分事项。我们换个话题。拆分事项。
Indeed. Alright. Carve outs. Let's shift gears. Carve outs.
你有什么发现？
What you got?
我和妻子最近沉迷看《双面女间谍》。
My wife and I have been on an alias binge.
哦，哇。是詹妮弗·加纳那部吗？
Oh, wow. Yeah. Jennifer Garner?
没错。当年播出时我完全没看过。它就像是完美的两千年代初垃圾食品——当你晚上躺在沙发上还有一小时空闲时的最佳选择。
Yes. I never saw it when it came out. It is like the perfect early two thousands junk food when you have one more hour at the end of the day, and you're just laying on the couch.
所以我一天结束时总抽不出多余时间。我有个两岁的孩子。但我真的很感激。因为十六年后她上大学时，我会把这件事记在清单上。
Then I never have one more hour at the end of the day. I have a two year old. But I really appreciate it. For, sixteen years from now when she goes to college, I'll keep that on my list.
哦，你在玩游戏啊。
Oh, you play games.
啊，确实如此。但那是研究。我只是在测试最新的图形技术。
Oh, that's true. But that's research. I'm just checking out the latest graphics technology.
所以我对《双面女间谍》的评论是它有点做作。他们经常重复自己。我是说，观察电视从那时到现在变化多大很诡异，因为如今他们制作非常相似的剧集，但手法含蓄得多。基调更阴暗。留给观众更多想象空间。
So my review of alias is it's a little bit campy. They repeat themselves pretty often. Mean, it's weird to observe how much TV has changed between now and then because they make very similar shows today, but they're just much more subtle. They're much darker. They leave much more sort of to the imagination.
而在21世纪初，一切都那么直白露骨，还要重复三遍。我很高兴这部剧没有笑声背景音，但绝对值得一看。有时你得想象它有不同配乐，因为每集都有类似《黑客帝国》风格的歌曲。
And in the early two thousands, everything was just so, like, explicit and on the nose and restated three times. I'm just glad the show doesn't have a laugh track, but it's well worth the watch. Sometimes you have to imagine it has a different soundtrack because every episode has like a matrix type song to it.
没错。这就像是电视剧版的《黑客帝国》。
Yes. That's right. This is like the TV version of The Matrix.
对吧？是的。但它很棒。不知道怎么说，我们看得很开心。
Right? Yes. But it's great. I don't know. We're having a lot of fun watching it.
我的车，与我现阶段的生活相关，也是我最近才想起并发现的事物。我们刚和Woah一起完整看完了第一部电影，第一部迪士尼电影。一个重要的里程碑。她简直爱死它了。我觉得我们选了一部很棒的作品。
My car, related for my stage of life, also something I missed and discovered recently. We just watched our first full movie, full Disney movie with our Woah. A major milestone. And she freaking loved it. And I think we picked a great one.
《海洋奇缘》，之前我和珍妮都没看过。后来稍微了解了一下，你知道皮克斯近几年有多可惜地走下坡路吗？真是令人失望。我是说，他们还是皮克斯，但已经不再是那个皮克斯了。
Moana, which neither Jenny or I had seen before. And in reading just a little bit about it afterwards, you know how super sadly Pixar kinda fell off in recent years? Like, such a bummer. I mean, they're still Pixar, but like, they're not Pixar.
它不再像从前那样每次都能保证成功了。
It's not the guaranteed hit every time that it used to be.
没错。《海洋奇缘》和《魔发奇缘》等作品属于皮克斯被收购后迪士尼动画回归本真的那一代作品，就像艾斯纳时代的迪士尼动画，各方面都表现出色。我们超爱这部电影。是和没有孩子的兄嫂一起看的，他们30岁住在旧金山，也超喜欢。
Yep. So Moana came out in this kinda generation with Tangled and some of the other stuff out of actual Disney animation after the Pixar acquisition that are just like, these are returned to form, Eisner era, Disney animated, just like fires on all cylinders. And we loved it. We watched with our brother and sister-in-law who don't have kids and are 30 living in San Francisco. They loved it.
我们的女儿超爱它。强烈推荐《海洋奇缘》，无论你处于人生的哪个阶段。
Our daughter loved it. Highly recommend Moana no matter what life phase you're in.
好的。太棒了。我这就把它加入我的观影清单。
Alright. Great. Adding it to my list.
而且还有巨石强森配音。这还有什么可挑剔的？就是这样。
And it's got the rock. How can you complain? There you go.
各位听众，如果你想在我们发布新一期节目时收到通知，确保不会错过，或者想在下期节目中获得小提示玩猜谜游戏，又或者想跟进我们从上期听众反馈中学到的内容——嘿，这里有条信息想分享给大家。我们将独家通过邮件发送这些内容，请访问acquire.fm/email获取。
Well, listeners, if you want to be notified every time we drop a new episode and you wanna make sure you don't miss it, and you want little hints to play a guessing game at our next episode, or you want follow ups from our previous episode in case we learn from listeners, hey, here's a little piece of information that we wanted to pass along. We will exclusively be dropping those in the email, acquire dot f m slash email.
太有趣了。我猜你接下来要聊我们的Slack。看着Slack里大家讨论这期节目的线索特别有意思。我们写了小预告，我当时还想，大家肯定能马上猜出来。
It was so fun. I think you're about to talk about our Slack. It was so fun watching people in Slack talk about the hints for this episode. We wrote the little teaser, and I was like, oh, everybody's gonna know exactly what this is.
结果没人猜中。我太震惊了。
No one got it. I was shocked.
是啊，最后确实有人猜到了，但花了好几天时间。
Yeah. Eventually, somebody did, but it took a couple days.
对了，我们有顶帽子，你应该买一顶。这可不是什么高利润商品，我们只是很兴奋看到更多人戴着ACQ的标识到处走。
Yeah. We have a hat. You should buy it. And this is not a thing that we make a lot of margin on. We just are excited about more people sporting the ACQ around.
所以快来参与这个行动吧。给你的朋友们也看看。
So participate in the movement. Show it to your friends.
虽然这不是我们的超级播客周边，但你知道的...
It's not our super pod, but, you know
是啊。
Yeah.
这个播客是超级播客。
The pod is the super pod.
成为Acquired LP。你可以，靠近厨房帮我们每季挑选一期节目，我们大约每隔一个月会进行一次Zoom通话。访问Acquired.fm/lp。在任何播客播放器中查看ACQ two获取更多Acquired内容，并加入Slack讨论，acquired.fm/slack。听众们，我们下次见。
Become an Acquired LP. You can, come closer to the kitchen and help us pick an episode once a season, and we'll do a Zoom call every other month or so. Acquired.fm/lp. Check out ACQ two for more Acquired content in any podcast player, and come talk about this in the Slack, acquired.fm/slack. Listeners, we'll see you next time.
我们下次见。
I'll see you next time.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。