Huang Renxun's latest 10,000-word interview: AGI is coming soon, AI will completely change productivity.

wallstreetcn · 19:52

机器学习的飞轮才是最重要的。仅仅拥有强大的GPU并不能保证一家公司在AI领域取得成功。

10月4日，英伟达首席执行官黄仁勋做客访谈节目Bg2 Pod，与主持人Brad Gerstne和Clark Tang进行了一场广泛的对话。

他们主要讨论了如何将智能扩展到AGI、英伟达的竞争优势、推理与训练的重要性、AI领域未来的市场动态、AI对各个行业的影响、Elon的孟菲斯超级集群与X.ai、OpenAI等话题。

黄仁勋强调了AI技术的迅速演变，尤其是通向通用人工智能（AGI）道路上的突破。他表示，AGI助理即将以某种形式出现，并且会随着时间的推移变得更加完善。

黄仁勋还分享了英伟达在计算革命中的领导地位，指出通过降低计算成本和创新硬件架构，英伟达在推动机器学习和AI应用中占据了显著优势。他特别提到英伟达的“护城河”，即其十年积累的软硬件生态系统，使得竞争者很难通过单一的芯片改进超越。

此外，黄仁勋赞扬了xAI和马斯克团队在短短19天内完成十万个GPU的孟菲斯超级集群建设，称这是“前所未有”的成就。这一集群无疑是全球最快的超级计算机之一，将在AI推理和训练任务中发挥重要作用。

谈及AI对生产力的影响，黄仁勋乐观地表示，AI将极大提升企业的效率，带来更多的增长机会，并不会导致大规模失业。同时，他也呼吁业界加强对AI安全性的关注，确保技术的开发和使用有益于社会。

全文要点总结如下：

（AGI助理）很快就会以某种形式出现……一开始它会非常有用，但并不完美。然后随着时间的推移，它会变得越来越完美。
我们在10年内将计算的边际成本降低了100000倍。我们整个堆栈都在增长，整个堆栈都在创新。
人们认为设计出更好的芯片的原因在于它有更多的触发器、更多的位和比特……但机器学习不仅仅是软件，它关乎整个数据管道。
机器学习的飞轮才是最重要的。你必须考虑如何让这个飞轮更快。
仅仅拥有强大的GPU并不能保证一家公司在AI领域取得成功。
马斯克对大型系统的工程和建设以及资源调配的理解是独一无二的... 十万个GPU作为一个集群... 19天内完成。
AI不会改变每一项工作，但会对人们的工作方式产生巨大影响。公司使用AI提高生产力时，通常表现为更好的收益或增长。

AGI和AI助理的演变

Brad Gerstner：

今年的主题是将智能扩展到AGI。两年前我们做这件事的时候，我们是在AI时代做的，那是在ChatGPT的两个月前，考虑到所有这些变化，这真是令人难以置信。所以我觉得我们可以用一个思想实验和一个预测来开始。

如果我把AGI通俗地想象成我口袋里的私人助理，如果我认为AGI就是那个口语助理，我已经习惯了。它知道我的一切。它对我有完美的记忆，可以和我交流。他们可以帮我预订酒店，或者帮我预约医生。看看当今世界的变化速度，你认为我们什么时候才能拥有个人助理？

黄仁勋：

很快就会以某种形式出现。而且随着时间的推移，这个助理会越来越好。这就是我们所知道的美妙技术。所以我认为一开始它会非常有用，但并不完美。然后随着时间的推移，它会变得越来越完美。就像所有的技术一样。

Brad Gerstner：

当我们看看变化的速度时，我认为马斯克说过，真正重要的唯一事情就是变化的速度。我们确实感觉到变化的速度已经急剧加快，这是我们在这些问题上见过的最快变化速度，因为我们已经在AI领域摸爬滚打了十年，甚至更长。这是你职业生涯中见过的最快变化速度吗？

黄仁勋：

这是因为我们重新发明了计算。很多事情之所以发生，是因为我们在10年内将计算的边际成本降低了100000倍。摩尔定律应该是100倍左右。我们通过多种方式实现了这一点。首先，我们引入了加速计算，将CPU上效率较低的工作放在GPU上。我们通过发明新的数值精度来实现这一点。我们通过新的架构来实现这一点，发明了张量核心，以系统的方式构建MV Link，以及非常非常快的内存、以及使用MV Link进行扩展并在整个堆栈上工作。基本上，我描述的关于英伟达如何做事的一切都导致了超摩尔定律的创新速度。

现在真正让人惊奇的是，从此之后，我们从人工编程转向了机器学习。机器学习的神奇之处在于，机器学习可以学得非常快。事实证明如此。因此，当我们重新制定分配计算的方式时，我们做了很多，各种并行性。张量并行性，各种管道并行性。我们擅长在此基础上发明新算法和新训练方法，所有这些技术，所有这些发明都是相互叠加的结果。

回顾过去，如果你看看摩尔定律是如何运作的，软件是静态的。它是预编译的，就像放入商店的收缩筏一样。它是静态的，下面的硬件以摩尔定律的速度增长。现在，我们整个堆栈都在增长，整个堆栈都在创新。所以我认为，现在我们突然看到了扩展。

这当然是非凡的。但我们过去谈论的是预训练模型和在那个层面上的扩展，以及我们如何将模型大小翻倍，因此相应地翻倍，数据大小也翻倍。结果，所需的计算能力每年增加四倍。这是一件大事。但现在我们看到了后训练的扩展，我们看到了推理的扩展。所以人们过去认为预训练很难，推理很容易。现在一切都很难。这很有道理，但认为人类的所有思维都是一次性的想法有点荒谬。所以，必须有一个快速思考、慢速思考、推理、反思、迭代和模拟等概念。现在它正在出现。

英伟达的竞争护城河

Clark Tang：

我认为，关于英伟达最容易被误解的事情之一就是真正的英伟达模式有多深。我认为有一种观念，如果有人发明了一种更好的芯片，他们就赢了。但事实是，你花了十年时间构建从GPU到CPU再到网络的完整堆栈，尤其是支持应用程序的软件和库。在英伟达上运行。所以我认为你谈到了这一点，但当你想到英伟达今天的护城河时，，您认为今天的视频模式比三四年前大还是小？

黄仁勋：

好吧，我很感谢你认识到计算是如何变化的。事实上，人们认为（而且很多人现在仍然这样认为）设计出更好的芯片的原因在于它有更多的触发器、更多的位和比特。您明白我的意思吗？您会看到他们的主题演讲幻灯片。它有所有这些触发器、条形图和诸如此类的东西。这些都很好。我的意思是，看，马力确实很重要。是的。所以这些事情从根本上来说很重要。

然而，不幸的是，这都是想法。这都是在软件是运行在Windows上的某个应用程序并且软件是静态的意义上的想法，对吗？这意味着改进系统的最佳方法就是制造越来越快的船。但我们意识到机器学习不是人类编程。机器学习不仅仅是软件，这关乎整个数据管道。事实上，机器学习的飞轮才是最重要的。那么你如何看待我启用这个飞轮？一方面，让数据科学家和研究人员在这个飞轮中高效工作，而这个飞轮从一开始就开始了。很多人甚至没有意识到需要AI来管理数据，来教导AI。而AI本身就相当复杂。

Brad Gerstner：

AI本身在改进吗？它也在加速吗？再说一次，当我们考虑竞争优势时，是的，没错。这是所有这些的组合。

黄仁勋：

正是如此，正是因为有了更聪明的AI来管理数据，才会导致这种情况。我们现在甚至有了合成数据生成和各种不同的数据管理方式，向其呈现数据。所以在你接受培训之前，你就已经涉及了大量的数据处理。所以人们会想，哦，Pytorch，这是世界的开始，也是世界的终结。这非常重要。

但别忘了，在Pytorch之前和之后，飞轮的意义在于你必须思考的方式，我该如何思考整个飞轮，如何设计一个计算系统，一个计算架构，帮助你利用这个飞轮，让它尽可能高效。这不是一个应用程序训练的大小。这说得通吗？这只是一步。好吧。飞轮上的每一步都很难。所以你应该做的第一件事，不是思考如何让Excel更快，如何让doom更快，那是过去的事情，不是吗？现在你必须考虑如何让这个飞轮更快？这个飞轮有很多不同的步骤，机器学习并不容易，你们都知道。

OpenAI 或 X 或 Gemini 团队所做的事情都不容易，他们深思熟虑地思考着我们。我的意思是，他们做的事情并不容易。所以我们决定，你看，这才是你应该考虑的。这是整个过程，你想加速其中的每一部分。你要尊重多尔斯定律，多尔斯定律表明，如果这是30%的时间，而我将其加速了三倍，那么我并没有真正加速整个过程。这有道理吗？你真的想创建一个系统来加速每一步，因为只有做整个事情，你才能真正实质性地改善周期时间和飞轮，也就是学习率，最终才是导致指数级增长的原因。

所以，我想说的是，我们对公司真正做的事情的看法会体现在产品上。注意，我一直在谈论这个飞轮，整个网站。是的，没错。我们加速了一切。

现在，主要关注点是视频。很多人都专注于物理AI和视频处理。想象一下前端。每秒有TB的数据进入系统。举个例子，一个管道会接收所有这些数据。首先要准备好训练。是的，这样整个过程就可以加速。

Clark Tang：

今天人们只考虑文本模型。是的，但未来是，这个视频模型，使用一些文本模型，比如o1，在我们到达那里之前真正处理大量数据。

黄仁勋：

是的。所以语言模型会涉及到一切。但我们，这个行业花费了巨大的技术和精力来训练语言模型，训练这些大型语言模型。现在，我们在每一步都使用大型语言形式。这非常了不起。

Brad Gerstner：

我听到你说的是，在组合系统中，是的，优势会随着时间的推移而增长。所以我听到你说，我们今天的优势比三到四年前更大，因为我们正在改进每个组件。这就是组合，当你想到，例如，作为一个商业案例研究，英特尔，相对于你现在的位置，它拥有主导模式，在堆栈中占据主导地位。也许，再简单归纳一下，将你的竞争优势与他们在其周期高峰期所拥有的竞争优势进行比较。

黄仁勋：

英特尔之所以与众不同，是因为他们可能是第一家在制造工艺工程和制造方面表现出色的公司。制造上面说的就是制造芯片。设计芯片，在x86架构中构建芯片，制造越来越快的x86芯片，这是他们的才华所在，他们将其与制造融合在一起。

我们公司有点不同，我们认识到这一点，事实上，并行处理并不要求每个晶体管都表现出色，串行处理要求每个晶体管都表现出色。并行处理需要大量的晶体管才能更具成本效益。我宁愿晶体管多10倍，速度慢20%。然后晶体管少10倍，速度快20%。这有道理吗？他们想要相反的结果。因此，单线程性能、单线程处理和并行处理非常不同。因此，我们观察到，事实上，我们的世界并不是越走越好。我们想要做得非常好，尽可能好，但我们的世界真的在不断进步。

并行计算、并行处理很难，因为每个算法都需要一种不同的重构方式和重新构建算法的架构。人们没有意识到的是，你可以有3个不同的 CPU。它们都有自己的C编译器。你可以将软件编译到那个轴上。

这在加速计算中是不可能的。提出架构的公司必须提出自己的 Open GL。所以我们革命性地进行了深度学习，因为我们的领域特定库叫做cuDNN（深度神经网络库），一个领域特定库称为光学。我们有一个领域特定库，称为 cuQuantum。

Brad Gerstner：

对于位于下方的行业特定算法，您知道，每个人都关注的 Pytorch 层。就像我经常听到的那样。

黄仁勋：

如果我们不发明它，上面的任何应用程序都无法工作。你们明白我的意思吗？所以是英伟达真正擅长的算法。在底层架构之上的科学之间的传播，这就是我们真正擅长的。

英伟达正在构建一个完整的AI计算平台，包括硬件、软件和生态系统

Clark Tang：

现在所有的注意力都集中在推理上。但我记得，两年前，我和Brad共进晚餐时问过你一个问题，你认为你的护城河在推理方面会像在训练方面一样强大吗？

黄仁勋：

我不确定我是否说过它会更强。

Clark Tang：

你刚才提到了很多这些元素，两者之间的可组合性，或者，我们不知道整体组合。对于客户来说，能够在两者之间保持灵活性非常重要。但是，既然我们现在已经处于推理时代，你能不能谈谈？

黄仁勋：

推理训练就是在那种规模上进行推理。我的意思是，你是对的。所以如果你训练得当，那么很有可能你会推理得当，如果你在没有任何考虑的情况下在这个架构上构建它，它将在这个架构上运行。好吧，你仍然可以去优化它以适应其他架构，但至少，因为它已经在英伟达上构建了架构，它将在英伟达上运行。

现在，另一个方面当然只是一种资本投资方面，也就是当你训练新模型时，你希望用你最好的新设备进行训练。这会留下你昨天使用的设备，而这些设备非常适合推理。所以在新的基础设施背后有一系列免费的设备可以兼容。因此，我们非常严格地确保我们始终兼容，这样我们留下的一切都将继续保持卓越。

现在，我们还投入了大量精力不断重新发明新算法，以便当时间到来时，Hopper架构比他们购买时好两倍、三倍、四倍，这样，这个基础设施就会继续真正有效。因此，我们所做的所有工作，改进新算法、新框架。请注意，它有助于我们拥有的每一个安装基础。Hopper更适合它，Ampere更适合它，甚至Volta也更适合它。

我想Sam Altman刚刚告诉我，他们最近刚刚停用了OpenAI的Volta基础设施。所以我认为，我们留下了这条安装基础的痕迹，就像所有计算安装基础都很重要一样。英伟达涉足每一个云端，包括本地和边缘。

VILA的视觉语言模型在云端创建，无需修改，在机器人边缘也能完美运行。它们都具有良好的兼容性。因此，我认为架构兼容性对于大型设备非常重要，对于 iPhone 和其他设备也是如此。我认为安装基础对于推理非常重要。

黄仁勋：

但真正让我受益匪浅的是，我们正在努力在新的架构中训练这些大型语言模型。我们能够思考，在某一天时机成熟时，如何创建在推理方面表现出色的架构。因此，我们一直在思考推理模型的迭代模型，以及如何为此创建非常互动的推理体验，对吧，您的个人代理。您不想说完话后就离开并思考一段时间。您想很快与您互动。那么我们该如何创建这样的东西呢？

MVLink这样我们就可以采用这些非常适合训练的系统。但是当你完成它时，推理性能会非常出色。所以你想优化这个到第一个Token的时间。而到第一个Token的时间实际上非常难做到，因为到第一个Token的时间需要大量的带宽。但如果你的上下文也很丰富，那么你需要大量FLOPS。因此，你需要无限量的带宽，同时无限量的FLOPS才能实现几毫秒的响应时间。所以这个架构真的很难实现。我们为此发明了伟大的Blackwell MVLink。。

Brad Gerstner：

本周早些时候和 Andy Jassy（亚马逊总裁兼CEO）一起吃饭，Andy说，我们有 Tranium、Inferentia 即将到来。我认为大多数人再次将这些视为英伟达的问题。但接下来，他说英伟达是我们的重要合作伙伴，并且将继续是我们的重要合作伙伴。就我所见，未来世界将依靠英伟达。

所以当你想到正在构建的定制ASIC时，它们将用于目标应用。也许是Meta的推理加速器，也许是亚马逊的训练，或者谷歌的TPU。然后你想想你今天面临的供应短缺，这些因素会改变这种动态吗？或者它们会补充他们从你那里购买的系统？

黄仁勋：

我们只是在做不同的事情。是的，我们试图完成不同的事情。现在英伟达正在尝试为这个新世界、这个机器学习世界、这个生成式AI世界、这个代理式AI世界构建一个计算平台。我们试图创造，在计算领域，如此深刻的一点是，经过60年的发展，我们重新发明了整个计算堆栈。从编程到机器学习，从CPU到GPU，从软件到AI，应用程序从软件到AI。从软件工具到AI。因此，计算堆栈和技术堆栈的每个方面都发生了变化。

我们想做的是创建一个随处可用的计算平台。这确实是我们所做工作的复杂性，我们所做工作的复杂性在于，如果你仔细想想我们所做的事情，就会发现我们正在构建整个AI基础设施，我们把它看作一台计算机。我之前说过，数据中心现在是计算的单位。对我来说，当我想到计算机时，我考虑的不是芯片。我在考虑这个东西。这是我的思维模型，里面的所有软件、所有编排、所有机器，都是我的使命。这就是我的计算机。

我们每年都试图建造一台新的。是的，这太疯狂了。以前从来没有人这样做过。我们每年都试图建造一台全新的。每年，我们都提供两到三倍的性能。因此，每年，我们都会将成本降低两到三倍。每年，我们都会将能源效率提高两到三倍。因此，我们要求客户不要一次性购买所有东西，每年只买一点，对吗？好的。这样做的原因是，我们希望它们的成本在未来保持平均水平。现在，所有东西在架构上都是兼容的，所以以我们现在的速度单独构建这些东西是非常困难的。

现在，双重困难的部分是，我们接受所有这些，而不是将其作为基础设施或服务出售，我们不同意所有这些。我们将它集成到GCP、AWS、Azure、X 中。所以每个人的集成都不同。我们必须将我们所有的架构库、所有算法和所有框架集成到他们的框架中。我们将我们的安全系统集成到他们的系统中，我们将我们的网络集成到他们的系统中，对吗？然后我们基本上进行10次集成，现在我们每年都这样做。这就是奇迹。

Brad Gerstner：

我们，我的意思是，你每年都试图这样做，这太疯狂了。那么是什么驱使你每年都这样做呢？

黄仁勋：

是的，那就是当你系统地分解它时。你们分解得越多，每个人分解得越多，他们就越惊讶。是的。今天整个电子生态系统如何能够致力于与我们合作，最终构建出一个集成到所有这些不同生态系统中的计算机立方体，并且协调如此无缝。因此，显然我们向后传播的是API、方法、业务流程和设计规则，而向前传播的是方法、架构和API。

Brad Gerstner：

他们本来就是这样的。

黄仁勋：

几十年来，他们一直在努力。是的，而且随着我们的发展也在不断发展。但是，这些 AP 必须整合在一起。

Clark Tang：

有人只需调用 OpenAI API，它就可以工作。就是这样。

黄仁勋：

对。是的，有点疯狂。这是一个整体。这就是我们发明的，这个庞大的计算基础设施，整个星球都在与我们合作。它融入了任何地方。你可以通过戴尔销售它，也可以通过惠普销售它。它托管在云端。它无处不在，无处不在。人们现在在机器人系统中使用它，机器人和人类机器人，它们在自动驾驶汽车中。它们在架构上都是兼容的。相当疯狂。

Brad Gerstner：

这太疯狂了。

黄仁勋：

我不想让你留下我没有回答问题的印象。事实上，我回答了。当我们真正分层基础时，我的意思是思考的方式。我们只是在做一些不同的事情。是的，作为一家公司，我们希望了解情况，我对公司和生态系统周围的一切都非常了解，对吗？

我知道所有人都在做其他事情，他们在做什么。有时这对我们是不利的，有时不是。我非常清楚这一点，但这并没有改变公司的目标。是的，公司的唯一目标是构建一个可以无处不在的平台架构。这就是我们的目标。

我们不会试图从任何人那里夺取任何份额。英伟达是做市商，而不是份额获取者。如果你看看我们公司没有展示的幻灯片，你会发现这家公司没有一天谈论市场份额，内部也没有。我们谈论的都是我们如何创造下一个东西？

我们可以在这个飞轮中解决的下一个问题是什么？我们如何才能更好地为人们服务？我们如何将过去需要大约一年时间的飞轮缩短到大约一个月？是的。那光速是多少？不是吗？

所以我们在考虑所有这些不同的事情，但有一件事我们不会，我们不会，我们对所有事情都了如指掌，但我们确信我们的使命是非常独特的。唯一的问题是这个使命是否必要。这有意义吗？所有公司，所有伟大的公司都应该把这个作为核心。这是关于你在做什么？

当然。唯一的问题是，它有必要吗？它有价值吗？对。它有影响力吗？它对人们有帮助吗？我确信你是一名开发人员，你是一家生成式AI初创公司，你即将决定如何成为一家公司。

你不必做出的一个选择是我支持哪一款 A6？如果你只支持CUDA，你可以去任何地方。你以后总是可以改变主意的。但我们是通往AI世界的入口，不是吗？

一旦你决定加入我们的平台，其他的决定你都可以推迟。你以后可以随时构建自己的基础。我们并不反对。我们不会因此而生气。当我与所有GCP、GCP Azure 合作时，我们会提前几年向他们展示我们的路线图。

他们没有向我们展示他们的基本路线图，而且这从未冒犯过我们。这有道理吗？我们创造，我们处于一个。如果你有一个唯一的目标，你的目标有意义，你的使命对你和其他人来说都是珍贵的，那么你可以透明。请注意，我的路线图在 GTC 上是透明的。我的路线图对我们在 Azure、AWS 和其他公司的朋友来说更深入。我们做任何这些事情都没有问题，即使他们正在构建自己的资产。

Brad Gerstner：

我认为，当人们观察业务时，您最近表示对Blackwell的需求非常疯狂。您说，工作中最困难的部分之一就是在世界缺乏您能够生产和提供的计算能力的情况下，用情感工具对人们说“不”。但批评者说了这些。稍等片刻。他们说这就像2000年的思科一样，我们过度建设光纤。这将是繁荣与萧条的交替。我想到23年初我们吃饭的时候。23年1月的那顿晚餐上，英伟达的预测是，2023年的收入将达到260亿美元。你做到了600亿美元。

黄仁勋：

就让事实真相大白吧。这是世界历史上最大的预测失败。对。我们至少可以承认。

GPU在AI计算中扮演着越来越重要的角色

Brad Gerstner：

那是，那是我们在11月22日非常兴奋，因为我们有来自Inflection的Mustafa这样的人，而没有来自Character的人来我们办公室谈论投资他们的公司。他们说，好吧，如果你不能投资我们的公司，那就买英伟达，因为世界上每个人都在试图获得英伟达芯片来构建这些将改变世界的应用程序。当然，寒武纪时刻发生在ChatGPT上。尽管如此，这25位分析师仍然专注于加密货币赢家，以至于他们无法想象世界上正在发生的事情。所以最终规模更大了。用非常通俗的英语来说，对Blackwell的需求是疯狂的，而且只要你能，只要你能预见，它就会一直这样下去。当然，未来是未知的，不可知的。但为什么批评者会错得这么离谱，认为这不会像思科在2000年那样过度建设。

黄仁勋：

思考未来的最佳方式是从第一原则出发，对吗？好的，那么，对于问题，我们正在做的事情的第一原则是什么？第一，我们在做什么？我们正在做的第一件事就是重新发明计算，不是吗？我们刚才说过，未来计算的方式将是高度机器学习的。是的，高度机器学习的。好的，几乎我们所做的一切，几乎每一个应用程序，Word、Excel、Powerpoint、Photoshop、Premier，AutoCAD，你最喜欢的应用程序都是手工设计的。我向你保证，未来它将高度机器学习。对吧？所以所有这些工具都会如此，最重要的是，你会有机器，代理来帮助你使用它们。好的。所以现在我们知道这是事实。对吧？我们已经重新发明了计算。我们不会回头了。整个计算技术堆栈都在被重新发明。好的。既然我们已经做到了这一点，我们说过软件会有所不同。软件能写的东西会有所不同。我们使用软件的方式也会有所不同。所以现在让我们承认这一点。所以这些就是我现在的基本事实。是的。

现在的问题是会发生什么？让我们回顾一下过去的家庭计算。过去的计算机投入了1万亿美元。我们看看，只要打开门，看看数据中心，看看它。这些计算机是你想要的未来吗？答案是否定的。你那里有所有这些CPU。我们知道它能做什么，不能做什么。我们只知道，我们有1万亿美元需要现代化的数据中心。所以现在，在我们说话的时候，如果我们要在未来四五年内对这些旧东西进行现代化改造。这不算不合理。

所以我们有一个趋势，你正在与那些必须对其进行现代化改造的人进行交谈。是的，他们正在GPU上对其进行现代化改造。就是这样。

我们再做一次测试。你有500亿美元的资本支出。你喜欢花选项A，选项B，为未来建立资本支出，对吗？

或者像过去一样建立资本支出，现在你已经拥有了过去的资本支出，对吗？是的，对。它就在那里。反正也没有好转多少。莫尔斯定律基本结束了。那为什么要重建它呢？

我们只拿出500亿美元，投入生成式AI，对吗？所以现在你的公司变得更好了。对吗？现在你会投入这500亿美元中的多少？好吧，我会投入500亿美元的100%，因为我已经有了四年的基础设施，这是过去的。

所以现在你只是，我只是从某人从第一原理思考的角度来推理，这就是他们正在做的事情。聪明的人在做聪明的事情。现在，第二部分是这样的。那么我们就有价值一万亿美元的产能。去吧，比尔。

价值数万亿美元的基础设施。大概是1500亿美元。好的。所以我们有1万亿美元的基础设施需要在未来四五年内建设。好吧，我们观察到的第二件事是软件的编写方式不同，但软件的使用方式也不同。

未来，我们会有代理。我们公司会有数字员工。在你的收件箱里，你会看到这些低矮的面孔上的小点。未来，事情意味着AIS的低矮图标。对吧？我会把这些发给他们。

我不再用C++编程电脑了，我要用提示来编程AI。对吧？现在，这和我今天早上跟我聊天没什么不同。

我来这里之前写了很多电子邮件。我当然在提示我的团队。我会描述背景，描述我所知道的基本限制，描述他们的任务。我会留下足够的空间，我会给出足够的方向，让他们明白我需要什么。我想尽可能清楚地说明结果应该是什么，但我留下了足够的模糊空间，一点创意空间，这样他们就可以给我惊喜。

对吧？这和我今天提示AI没什么不同。是的，这正是我提出AI的方式。因此，我们将要现代化的基础设施之上，将会有一个新的基础设施。这个新的基础设施将是操作这些数字人的AI工厂，它们将全天候运行。

我们将为世界各地的所有公司提供这些设备。我们将在工厂中拥有它们，我们将在自主系统中拥有它们。对吗？所以有一整层计算结构。这一整层我称之为AI工厂，世界必须制造，但今天根本不存在。

所以问题是，这有多大。目前还不知道。可能有几万亿美元。我知道现在的情况，但当我们坐在这里建造时，美妙之处在于，这个新数据中心的现代化架构和AI工厂的架构是一样的。这是一件好事。

Brad Gerstner：

您能否说清楚，您有一万亿的旧东西。您必须进行现代化。您至少有一万亿的新AI工作负载即将到来。是的，您今年的收入将达到1250亿美元。曾经有人告诉过你，这家公司的市值永远不会超过10亿美元。当你今天坐在这里时，有什么理由吗？对，如果你在数万亿的Tam中只有1250亿美元，那么你将来的收入将不会是现在的2倍或3倍。你的收入没有增长有什么原因吗？没有。

黄仁勋：

正如你所知，并不是所有事情都如此，公司只受鱼塘大小的限制，金鱼池只能这么大。所以问题是，我们的鱼塘是什么？我们的池塘是什么？这需要很多想象力，这就是为什么做市商在不创建新鱼塘的情况下考虑未来的原因。回顾过去并试图夺取市场份额很难弄清楚这一点。对。份额获取者只能这么大。当然。做市商可以非常大。当然。

所以，我认为我们公司拥有的好运是，从公司成立之初，我们就必须创造市场，才能在其中畅游。当时人们没有意识到这一点，但现在人们已经意识到了，但我们处于创造3D游戏PC市场的起点。我们基本上发明了这个市场，以及所有的生态系统和显卡生态系统，我们发明了这一切。因此，需要发明一个新的市场，以便以后为其服务，这对我们来说是一件非常舒服的事情。

黄仁勋：我为OpenAI的成功感到高兴

Brad Gerstner：

众所周知，OpenAI 本周以 1500 亿美元的估值筹集了 65 亿美元。我们都参与了。

黄仁勋：

是的，真的为他们感到高兴，真的很高兴他们走到了一起。是的，他们做了一件伟大的事情，团队也做得很好。

Brad Gerstner：

据报道，他们今年的收入或营业收入将达到50亿美元左右，明年可能会达到100亿美元。如果你看看今天的业务，它的收入大约是谷歌首次公开募股时的两倍。他们有2.5亿，是的，每周平均用户数为2.5亿，我们估计这是谷歌首次公开募股时的两倍。如果你看看这家公司的市盈率，如果你相信明年会有100亿美元，那么它大约是预期收入的15倍，也就是谷歌和Meta在首次公开募股时的市盈率。想象一家22个月前收入为零、每周平均用户数为零的公司。

跟我们谈谈OpenAI作为合作伙伴对你的重要性，以及OpenAI作为推动公众对AI的认识和使用的力量。

黄仁勋：

好吧，这是我们这个时代最重要的公司之一，一家追求AGI愿景的纯AI公司。不管它的定义是什么。我几乎不认为定义是什么完全重要，我也不认为时机很重要。我知道的一件事是，AI将随着时间的推移拥有能力路线图。而这个能力路线图将非常壮观、奇特。在此过程中，早在它达到任何人对AGI的定义之前，我们就会充分利用它。

你所要做的就是，现在，在我们说话的时候，去和数字生物学家、气候技术研究人员、材料研究人员、物理科学家、天体物理学家、量子化学家交谈。你可以去问视频游戏设计师、制造工程师、机器人专家。选你最喜欢的。无论你想选择哪个行业，你都要深入研究，和重要的人交谈，问他们，AI是否彻底改变了你的工作方式。你收集这些数据点，然后回头问问自己，你想有多怀疑。因为他们不是在谈论AI的概念优势。他们现在谈论的是将来使用AI。现在，农业技术、材料技术、气候技术，你选择你的技术，你选择你的科学领域。它们正在进步。AI正在帮助他们推进他们的工作。

现在，正如我们所说，每个行业、每个公司、每个高度、每所大学。难以置信。对吧？绝对是这样。我会以某种方式改变商业。我们知道这一点。我的意思是，我们知道它是如此切实。

今天。它正在发生。它正在发生。所以，我认为ChatGPT 的觉醒引发了它，这完全令人难以置信。我喜欢他们的速度和他们推动这一领域发展的独特目标，这真的很重要。

Brad Gerstner：

他们建立了经济引擎，可以为下一个模型前沿提供资金。我认为硅谷正在形成一种共识，即整个模型层、商品化的 Llama 使许多人能够以非常低廉的价格建立模型。所以早期我们有很多模型公司。这些，特征、语调和凝聚力都列在清单上。

很多人质疑这些公司是否能够在经济引擎上建立逃逸速度，从而继续资助下一代。我自己的感觉是，这就是你看到整合的原因。OpenAI 显然达到了速度。他们可以资助自己的未来。我不清楚其他许多公司是否能做到。这是对模型层现状的公平评估吗？我们将像在许多其他市场一样，将这种整合到能够负担得起的市场领导者身上，他们拥有经济引擎和应用程序，可以让他们继续投资。

仅仅拥有强大的GPU并不能保证一家公司在AI领域取得成功

黄仁勋：

首先，模型和AI之间存在根本区别。是的。模型是必不可少的要素。对。对于AI来说，它是必要但不充分的。对。所以，AI是一种能力，但用于什么，对吗？那么它的应用是什么？对吗？软件驾驶汽车的AI与人类机器人的AI有关，但并不相同，后者与聊天机器人的AI有关，但并不相同。

所以你必须了解分类法。是的，堆栈的分类法。在堆栈的每一层，都会有机会，但不是堆栈的每一层都为每个人提供无限的机会。

现在，我刚刚说了一句话，你所做的就是用GPU替换模型这个词。事实上，这是我们公司 32 年前的一个伟大观察，即 GPU、图形芯片或GPU与加速计算之间存在根本区别。加速计算与我们在 AI 基础设施方面所做的工作不同。它们是相关的，但并不完全相同。它们是相互叠加的。它们并不完全相同。而且这些抽象层中的每一个都需要完全不同的技能。

真正擅长构建GPU的人不知道如何成为一家加速计算公司。我可以举例说明，有很多人制造 GPU。我不知道哪一个是后来的，我们发明了 GPU，但你知道我们不是，我们不是今天唯一一家制造GPU的公司，对吗？到处都有 GPU，但它们不是加速计算公司。有很多人这样做。他们的加速器可以进行应用程序加速，但这与加速计算公司不同。例如，一个非常专业的AI应用程序，对吧，这可能是一件非常成功的事情，对吗？

Brad Gerstner：

这就是 MTIA（Mata自研的下一代AI加速芯片）。

黄仁勋：

对。但它可能不是那种带来影响力和能力的公司。所以你必须决定你想成为什么样的人。所有这些不同领域可能都有机会。但就像建立公司一样，你必须注意生态系统的变化以及随着时间的推移哪些东西会被商品化，认识到什么是功能，什么是产品，对，什么是公司。好的。我刚刚讲过，好吧，你可以用很多不同的方式来思考这个问题。

xAI和孟菲斯超级计算机集群已经到了“20万到30万个GPU集群的时代”

Brad Gerstner：

当然，有一家新进入者有钱、有智慧、有野心。那就是 xAI。是的，对。而且，有报道称你和 Larry Ellison（甲骨文创始人）和马斯克共进晚餐。他们说服你放弃 100000个H100芯片。他们去了孟菲斯，在几个月内就建立了一个大型连贯超级集群。

黄仁勋：

三点，不要划等号，好吗？是的，我和他们共进晚餐。

Brad Gerstner：

你认为他们有能力建立这个超级集群吗？有传言说他们想要另外十万个 H200，对吧，来扩大这个超级集群的规模。首先，跟我们谈谈 X 和他们的野心以及他们取得的成就，但同时，我们已经到了20万到30万个GPU集群的时代了吗？

黄仁勋：

答案是肯定的。然后首先，承认成就。从概念的那一刻到数据中心准备好让英伟达在那里安装我们的设备，再到我们启动它、连接好它并进行第一次训练的那一刻，这一切都值得。

黄仁勋：

好的。所以第一部分就是在这么短的时间内建造一个巨大的工厂，水冷、通电、获得许可，我的意思是，这就像超人一样。是的，据我所知，世界上只有一个人能做到这一点。我的意思是，马斯克对大型系统的工程和建设以及资源调配的理解是独一无二的。是的，这真是令人难以置信。当然，他的工程团队也很出色。我的意思是，软件团队很棒，网络团队很棒，基础设施团队很棒。马斯克对此深有体会。

从我们决定与工程团队、网络团队或基础设施计算团队、软件团队一起开始规划的那一刻起，所有的准备工作都提前了。然后所有的基础设施、所有的物流、当天运来的技术和设备数量、视频基础设施和计算基础设施，以及培训所需的所有技术，19 天都悬而未决，你想要什么吗？做了。

退一步想想，你知道 19 天是多少天吗？19 天是几周吗？对吧？如果你亲眼看看，技术的数量是令人难以置信的。所有的布线和网络，英伟达设备的网络与超大规模数据中心的网络非常不同。好的，一个节点需要多少根电线。计算机的背面全是电线，而将这一大堆技术和所有软件集成在一起，真是不可思议。

所以我认为马斯克和X团队所做的，我非常感激他承认我们与他一起进行的工程工作以及规划工作等等。但他们取得的成就是独一无二的，以前从未有过。只是从这个角度来看。十万个 GPU，作为一个集群，这很容易成为地球上最快的超级计算机。你建造的超级计算机通常需要三年的规划时间。然后他们交付设备，需要一年的时间才能让它们全部运转起来。是的，我们说的是 19 天。

Clark Tang：

英伟达的功劳是什么？

黄仁勋：

一切都已经正常运转了。是的，当然，还有一大堆 X 算法、X 框架、X 堆栈等等。我们说我们有大量逆向集成要做，但规划非常出色。只是预先规划。

大规模分布式计算是未来AI发展的重要方向

Brad Gerstner：

一端是正确的。马斯克是一端。是的，你，但你回答这个问题时一开始就说，是的，这里有 20到 30万个GPU集群。是的，对。这能扩展到 50万个吗？能扩展到 100 万个吗？你的产品需求是否取决于它扩展到 200 万个？

黄仁勋：

最后一部分是否定的。我的感觉是分布式训练必须有效。我的感觉是分布式计算将被发明。某种形式的联邦学习和分布式计算，异步分布式计算将被发现。

我对此非常热衷和乐观，当然，要意识到的是，缩放定律过去是关于预训练的。现在我们已经转向多模态，我们已经转向合成数据生成，后训练现在已经扩展得令人难以置信。合成数据生成、奖励系统、基于强化学习，然后现在推理缩放已经达到了顶峰。一个模型在回答你的答案之前已经进行了令人难以置信的 10000 次内部推理。

这可能并非不合理。它可能已经完成了树搜索。它可能已经在此基础上进行了强化学习。它可能，它可能已经进行了一些模拟，肯定做了很多反思，可能查找了一些数据，查看了一些信息，不是吗？所以他的背景可能相当大。我的意思是，这种类型的智能是。好吧，这就是我们所做的。这就是我们所做的。不是吗？因此，对于能力，这种扩展，我刚刚进行了计算，并将其与模型大小和计算大小每年 4 倍进行复合。

另一方面，需求在使用方面持续增长。我们认为我们需要数百万个GPU吗？毫无疑问。是的，现在这是肯定的。所以问题是，我们如何从数据中心的角度来构建它？这在很大程度上与数据中心是一次几千兆瓦还是一次 250 兆瓦有关。我的感觉是，你会同时得到两者。

Clark Tang：

我认为分析师总是关注当前的架构赌注，但我认为这次谈话中最大的收获之一是，你正在考虑整个生态系统和未来很多年。所以，因为英伟达只是在扩大或扩大规模，是为了满足未来的需求。这并不是说，你只能依赖一个拥有 50 万甚至一百万个GPU集群的世界。当分布式训练出现时，你就会编写软件来实现它。

黄仁勋：

我们七年前就开发了 Megatron。是的，这些大型训练任务的扩展会发生。所以我们发明了Megatron，所以，所有正在进行的模型并行性，所有分布式训练的突破和所有批处理以及所有这些东西都是因为我们做了早期的工作，现在我们正在为下一代做早期的工作。

AI改变了工作方式

Brad Gerstner：

那么我们来谈谈草莓和o1。我认为他们以o1命名很酷。这意味着招募世界上最优秀、最聪明的人，并将他们带到美国。我知道我们都对此充满热情。所以我喜欢这个想法，建立一个思考的模型，将我们带到下一个扩展智能的水平，对吧，这是对这样一个事实的致敬：正是这些通过移民来到美国的人，成就了我们，让我们成为现在的样子，将他们的集体智慧带到了美国。

黄仁勋：

当然。还有外星智慧。

Brad Gerstner：

当然。这是由我们的朋友 Noam Brown 带头的。推理时间推理作为扩展智能的全新载体有多重要，与仅仅构建更大的模型是分开的。

黄仁勋：

这是一件大事。这是一件大事。我认为，很多智能不能先验地完成。对。很多计算，甚至很多计算都不能重新排序。我的意思是，无序执行可以优先完成，很多事情只能在运行时完成。

所以，无论你是从计算机科学的角度还是从智能的角度来思考，太多的事情都需要背景。环境，对吧。还有质量，你正在寻找的答案类型。有时，一个快速的答案就足够了。这取决于答案的后果，影响。这取决于答案的使用性质。所以，有些答案，请花一个晚上，有些答案需要一周的时间。

是的。对吧？所以我完全可以想象我向我的AI发送一个提示，告诉它，考虑一晚。考虑一夜。不要马上告诉我。我希望你考虑一整晚，然后明天再告诉我。你对我最好的答案和理由是什么。所以，我认为从产品的角度来看，现在的质量，智能的细分。会有一次性的版本。当然。还有一些需要五分钟。

对吧？还有人类。所以如果你愿意的话，我们将成为一个庞大的员工群体。他们中有些是AI中的数字人，有些是生物人，我希望有些甚至是超级机器人。

Brad Gerstner：

我认为，从商业角度来看，这是一个被严重误解的事情。你刚刚描述了一家公司，它的产出量相当于一家拥有 15 万人的公司，但你只用 5 万人就做到了。没错。现在，你并没有说我要解雇所有员工。不。你仍在增加组织中的员工数量，但该组织的产出量将大幅增加。

黄仁勋：

这，这经常被误解。AI不是我。AI不会改变每一项工作。AI将对人们的工作方式产生巨大影响。让我们承认这一点。AI有潜力带来令人难以置信的好处。它也有潜力造成伤害。我们必须建立安全的AI。是的，让我们打好这个基础。是的。好的。

黄仁勋：

人们忽视的部分是，当公司使用AI提高生产力时，它很可能会表现为更好的收益或更好的增长，或两者兼而有之。当这种情况发生时，首席执行官的下一封电子邮件很可能不是裁员。

Brad Gerstner：

当然是公告，因为你在成长。

黄仁勋：

原因是我们有更多的想法，我们可以探索，我们需要人们帮助我们在自动化之前仔细考虑。所以自动化部分，AI可以帮助我们做到。显然，它也会帮助我们思考，但仍然需要我们去弄清楚我想解决什么问题。我们可以解决的问题有上万亿。那么，公司需要解决什么问题，选择这些想法，并找出自动化和扩展的方法。因此，随着我们变得更有生产力，我们将雇佣更多的人。人们忘记了这一点，如果你回到过去，显然我们今天的想法比 200 年前更多。这就是为什么 GDP 更大、就业人数更多的原因。尽管我们在底层疯狂地自动化。

Brad Gerstner：

这是这个时期的一个非常重要的点，我们正在进入一个几乎所有人类生产力、几乎所有人类繁荣都是自动化的副产品。过去 200 年的技术。我的意思是，你可以看看亚当·斯密和谢姆彼得的创造性破坏，你可以看看过去 200 年来人均 GDP 增长图表，现在它正在加速。

是的，这让我想到了这个问题。如果你看看 90 年代，我们美国的生产力增长率大约是每年 2.5% 到 3%，好吗？然后在 2010 年，它放缓到大约 1.8%。然后过去 10 年是生产力增长最慢的十年。所以这就是我们固定数量的劳动力和资本或产出量，实际上是有记录以来最慢的。

很多人都在争论这个原因。但如果世界真的像您所描述的那样，我们要利用和制造智能，那么我们是不是正处于人类生产力急剧扩张的边缘呢？

黄仁勋：

这是我们的希望。这是我们的希望。当然，我们生活在这个世界上，所以我们有直接的证据。

我们有直接的证据，要么是孤立的案例，要么是个别研究人员，他们能够利用AI以难以想象的超大规模探索科学。这就是生产力。百分百衡量生产力，或者我们正在以如此高的速度设计出如此令人难以置信的芯片。我们正在构建的芯片复杂性和计算机复杂性正在呈指数级增长，而公司的员工基础并不是衡量生产力的标准，对吧。

我们开发的软件越来越好，因为我们使用AI和超级计算机来帮助我们。员工人数几乎呈线性增长。生产力的另一个体现。

所以，我可以深入研究，我可以抽样检查很多不同的行业。我可以亲自检查。是的，你说得对。商业。没错。

所以我可以，当然，你不能，我们不能，我们可能会过度拟合。但它的艺术性当然是概括我们所观察到的是什么，以及这是否会在其他行业中体现出来。

毫无疑问，AI是世界上已知的最有价值的商品。现在我们要大规模生产它。我们，我们，我们所有人都必须擅长，如果你被这些AI包围，它们做得非常好，比你好得多，会发生什么。当我回想起来，这就是我的生活。我有 60 个直接下属。

他们在自己的领域是世界一流的，而且他们做得比我好。比我好很多。我与他们互动毫无困难，我也能毫不费力地设计他们。我也能毫不费力地编程他们。所以我认为人们要学习的是，他们都将成为 CEO。

他们都将成为 AI 代理的 CEO。他们有能力拥有创造力，嗯，一些知识，以及如何推理，如何分解问题，这样你就可以对这些 AI 进行编程，以帮助你实现像我一样的目标。这就是经营公司。

AI安全需要多方共同努力

Brad Gerstner：

现在。你提到了一些东西，那就是不协调，安全的AI。你提到了中东正在发生的悲剧。我们拥有很多自主权，而且很多AI正在世界各地使用。那么让我们来谈谈坏人、安全AI、与华盛顿的协调。你今天感觉如何？我们走在正确的道路上吗？我们有足够的协调水平吗？我认为马克·扎克伯格曾说过，我们打败坏AI的方法是让好AI变得更好。你如何描述你对我们如何确保这对人类产生积极的净收益的看法，而不是让我们生活在这个反乌托邦的世界里。

黄仁勋：

关于安全的讨论确实很重要，也很好。是的，抽象的观点，将AI视为一个巨大的神经元网络的概念性观点，并不是那么好，对吧。好的。原因是，众所周知，AI和大型语言模型是相关的，而不是一回事。我认为有很多正在做的事情非常好。第一，开源模型，以便整个研究社区、每个行业和每个公司都可以参与AI，是的，并学习如何利用这种能力进行应用。非常好。

第二，人们低估了致力于发明AI以保证AI安全的技术数量。是的，AI可以整理数据、携带信息、进行训练，创建AI是为了协调AI，生成合成数据以扩展AI的知识，使其减少幻觉。所有被创建用于矢量化或图形化或任何其他用于通知 AI、保护 AI 以监控其他 AI 的 AI 系统，这些 AI 系统创建的安全 AI 正在受到赞扬，对吗？

Brad Gerstner：

那么我们已经建立了。

黄仁勋：

那。我们正在建立这一切。是的，在整个行业中，方法论、红队、流程、模型卡、评估系统、基准测试系统，所有这些，所有这些正在以令人难以置信的速度构建的线束。我想知道，庆祝。你们明白吗？是的，你知道。

Brad Gerstner：

而且，没有，没有，没有政府法规说你必须这样做。是的，今天在这个领域中构建这些AI的参与者正在认真对待这些关键问题，并围绕最佳实践进行协调。没错。

黄仁勋：

所以这还没有得到充分重视，也没有得到充分理解。是的。需要有人，需要，每个人都需要开始谈论AI，这是一个AI系统，是一个工程系统，是经过精心设计的，从第一原则构建的，经过充分测试的，等等。记住，AI是一种可以应用的能力。我不认为有必要对重要技术进行监管，但也不要过度监管，以至于有些监管要针对大多数应用进行。所有已经监管技术应用的不同生态系统现在都必须监管现在融入AI的技术应用。

所以，我认为，不要误解，不要忽视世界上为AI而必须启动的大量法规。不要只依赖一个宇宙银河系。AI委员会可能能够做到这一点，因为所有这些不同的机构的成立都是有原因的。所有这些不同的监管机构的成立都是有原因的。回到最初的原则，我会。

开源与不开源的对立是错误的

Brad Gerstner：

你们推出了一个非常重要、非常庞大、非常强大的开源模型。

黄仁勋：

Nemotron。

Brad Gerstner：

是的，很明显，Meta为开源做出了重大贡献。我发现当我阅读 Twitter 时，有很多关于开放与封闭的讨论。您如何看待开源，您自己的开源模型，能否跟上前沿？这是第一个问题。第二个问题是，您知道，拥有开源模型和闭源模型，它们为商业运营提供动力，这是您对未来的看法吗？这两件事是否为安全创造了健康的张力？

黄仁勋：

开源与闭源与安全有关，但不仅仅是安全。例如，拥有闭源模型绝对没有错，它们是维持创新所必需的经济模型的引擎。好吧，我完全赞成这一点。我认为封闭与开放的对立是错误的。

因为开放是许多行业得以激活的必要条件，现在，如果我们没有开源，所有这些不同的科学领域如何能够激活，激活AI。因为他们必须开发自己的特定领域的AI，他们必须使用开源模型开发自己的AI，创建特定领域的AI。它们是相关的，不是，再说一遍，不一样。仅仅因为你有一个开源模型并不意味着你就有AI。所以你必须有那个开源模型来创建AI。所以金融服务、医疗保健、交通运输，这些行业、科学领域的列表现在已经因为开源而得以实现。

Brad Gerstner：

难以置信。你看到对你的开源模型的需求很大吗？

黄仁勋：

我们的开源模型？首先。Llama 下载。显然，是的，马克和他们所做的工作令人难以置信。超乎想象。是的。它完全激活并吸引了每一个行业、每一个科学领域。

好的，当然。我们做 Nemotron的原因是为了生成合成数据。直观地说，一个AI会以某种方式坐在那里循环并生成数据来学习自己。这听起来很脆弱。你能绕这个无限循环多少次，这个循环值得怀疑。然而，我脑海中的画面有点像你找了一个超级聪明的人，把他关进一个软垫房间，关上门大约一个月，出来的可能不是一个更聪明的人。所以，所以，但你可以让两三个人坐在一起，我们有不同的AI，我们有不同的知识分布，我们可以来回进行质量保证。我们三个人都可以变得更聪明。

因此，你可以让 AI 模型交换、互动、来回传递、讨论强化学习、合成数据生成等，这种想法在直觉上是有意义的，可以提出建议并有意义。因此，我们的模型Nemotron 350B 是是世界上最好的奖励系统模型。因此，这是最好的批评。

有趣。这是一个非常棒的模型，可以增强其他人的模型。因此，无论其他人的模型有多好，我都会建议使用Nemotron 340B 来增强和改进它。我们已经看到 Llama 变得更好，使所有其他模型都变得更好。

Brad Gerstner：

作为在 2016 年交付 DGX1 的人，这真的是一段不可思议的旅程。您的旅程既不可思议又令人难以置信。就像只是在早期幸存下来一样非常了不起。您在 2016 年交付了第一台 DGX1。我们在 2022 年迎来了寒武纪时刻。

所以我要问您一个我经常想得到答案的问题，那就是，在 60 个直接下属的领导下，您能维持现在的工作多久？您无处不在。您正在推动这场革命。您玩得开心吗？还有什么其他您更愿意做的事情吗？

黄仁勋：

这是关于过去一个半小时的问题。答案是“我很享受”。很棒的时光。我无法想象我更愿意做的其他事情。让我们看看。我认为，我认为不应该给人留下我们的工作总是充满乐趣的印象。我的工作并不总是充满乐趣，我也不指望它总是充满乐趣。我曾经期望它总是充满乐趣吗？我认为它总是很重要。

是的，我不会太严肃地对待自己。我非常认真地对待工作。我非常认真地对待我们的责任。我非常认真地对待我们的贡献和我们的时刻。

这总是充满乐趣吗？不。但我一直都喜欢它吗？是的。就像所有的事情一样，无论是家庭、朋友还是孩子。它总是充满乐趣吗？不。我们总是喜欢它吗？绝对。

所以我认为，我，我能做多久？真正的问题是，我能保持相关性多久？这才是最重要的，这个问题的答案只能是我将如何继续学习？今天我更加乐观。我这么说不仅仅是因为我们今天的主题。我对我说相关性和继续学习的能力更加乐观，因为AI。我每天都在用它，我不知道，但我相信你们都在用。我几乎每天都在用它。

我没有一个研究不涉及AI。是的，没有一个问题，即使我知道答案，我也会用AI再三核对。是的，令人惊讶的是，我接下来问的两三个问题，揭示了一些我不知道的东西。你选择你的主题。你选择你的主题。我认为AI是一个导师。

AI 是助手，AI 是合作伙伴，可以与我一起集思广益，检查我的工作，伙计们，这完全是革命性的。我是一名信息工作者。我输出的是信息。所以我认为他们对社会的贡献非常了不起。所以我认为，如果是这样的话，如果我能保持这种相关性，继续做出贡献，我知道这项工作足够重要，是的，我想继续追求它，我的生活质量令人难以置信。所以我会。

Brad Gerstner：

说我无法想象你和我在这个领域已经工作了几十年，我无法想象错过这一刻。这是我们职业生涯中最重要的时刻。我们非常感谢这种合作关系。

黄仁勋：

怀念未来十年。

Brad Gerstner：

思想伙伴关系。是的，你让事情变得更聪明。谢谢你。我认为你作为领导层的一部分真的很重要，对吧，这将乐观而安全地引领这一切向前发展。所以谢谢你。

黄仁勋：

和你们在一起。真的很开心。真的。谢谢。

The flywheel of machine learning is the most important. Just having a powerful GPU does not guarantee a company's success in the field of AI.

On October 4th, NVIDIA CEO Huang Renxun was a guest on the interview program Bg2 Pod, and had an extensive conversation with hosts Brad Gerstne and Clark Tang.

They mainly discussed how to expand intelligence to AGI, NVIDIA's competitive advantages, the importance of reasoning and training, the market dynamics of the future in the AI field, the impact of AI on various industries, and topics such as Elon's Memphis supercluster and X.ai, OpenAI, among others.

Huang Renxun emphasized the rapid evolution of AI technology, especially breakthroughs on the path to General Artificial Intelligence (AGI). He mentioned that AGI assistants will soon appear in some form and will become more sophisticated over time.

Huang Renxun also shared NVIDIA's leadership position in the computing revolution, pointing out that by reducing computing costs and innovating hardware architecture, NVIDIA has a significant advantage in driving machine learning and AI applications. He particularly mentioned NVIDIA's "moat", the decade-long accumulation of a software and hardware ecosystem, making it difficult for competitors to surpass with a single chip improvement.

Furthermore, Huang Renxun praised xAI and the Musk team for completing the construction of the Memphis supercluster with one hundred thousand GPUs in just 19 days, calling it an "unprecedented" achievement. This cluster is undoubtedly one of the fastest supercomputers globally, and will play a crucial role in AI reasoning and training tasks.

Regarding the impact of AI on productivity, Huang Renxun optimistically mentioned that AI will greatly enhance enterprise efficiency, bring more growth opportunities, and not lead to mass unemployment. At the same time, he called for the industry to strengthen its focus on AI security to ensure that the development and use of technology are beneficial to society.

The summary of the key points of the full text is as follows:

Siasun Robot&Automation assistants will soon appear in some form... At first, it will be very useful, but not perfect. Then over time, it will become more and more perfect.
We have reduced the calculated marginal costs by 100,000 times in 10 years. Our entire stack is growing, the entire stack is innovating.
People think that the reason for designing better chips is that it has more triggers, more bits, and bytes... but machine learning is not just about software, it's about the entire data pipeline.
The flywheel of machine learning is the most important. You have to think about how to make this flywheel faster.
Merely having a powerful GPU does not guarantee a company's success in the AI field.
Musk's understanding of large-scale system engineering, construction, and resource allocation is unique... 100,000 GPUs as a cluster... completed in 19 days.
AI will not change every job, but will have a huge impact on how people work. When companies use AI to increase productivity, it often manifests as better profits or growth.

Evolution of AGI and AI assistants

Brad Gerstner:

This year's theme is to expand intelligence to AGI. When we did this two years ago, we were in the AI era, which was two months before ChatGPT, considering all these changes, it's truly incredible. So I think we can start with a thought experiment and a prediction.

If I popularly imagine AGI as my personal assistant in my pocket, if I think of AGI as that conversational assistant I'm used to. It knows everything about me. It has perfect memory of me and can communicate with me. They can help me book hotels or make doctor appointments. Looking at the speed of change in the world today, when do you think we will have a personal assistant?

Huang Renxun:

It will soon appear in some form. And over time, this assistant will get better and better. This is the wonderful technology we know. So I think at first it will be very useful but not perfect. Then over time, it will become more and more perfect. Just like all technologies.

Brad Gerstner:

When we look at the speed of change, I think Musk once said, the only thing that really matters is the speed of change. We do feel that the speed of change has accelerated dramatically, this is the fastest rate of change we have seen on these issues because we have been in the AI field for ten years, or even longer. Is this the fastest rate of change you have seen in your career?

Huang Renxun:

This is because we have reinvented computing. Many things have happened because we have reduced the marginal cost of computing by 100,000 times in 10 years. Moore's Law should be around 100 times. We have achieved this in multiple ways. Firstly, we introduced accelerated computing, moving the less efficient work on CPUs to GPUs. We achieved this through inventing new numerical precisions. We achieved this through new architectures, inventing tensor cores, building the MV Link system, very fast memory, extending with MV Link, and working across the entire stack. Essentially, everything I described about how NVIDIA operates has led to an innovation pace beyond Moore's Law.

What is truly amazing is that from then on, we transitioned from manual programming to machine learning. The magic of machine learning is that it can learn very quickly. And it has been proven. So, when we redefined the way computations are allocated, we did a lot, all kinds of parallelism. Tensor parallelism, various pipeline parallelisms. We excel in inventing new algorithms and training methods on this basis, and all these technologies, all these inventions are the cumulative result.

Looking back, if you look at how Moore's Law worked, software was static. It was precompiled, like a shrink-wrapped raft put in a store. It was static, while the hardware underneath grew at the pace of Moore's Law. Now, our entire stack is growing, the whole stack is innovating. So, I think now we suddenly see expansion.

This is certainly extraordinary. But what we used to talk about was pre-trained models and the scalability at that level, and how we doubled the model size, hence doubled the data size accordingly. As a result, the required computing power increases fourfold each year. This is a big deal. But now we see the post-training expansion, we see the inference expansion. So people used to think pre-training was hard, inference was easy. Now everything is hard. It makes sense, but thinking that all human thinking is one-shot ideas is a bit absurd. So, there must be concepts of fast thinking, slow thinking, inference, reflection, iteration, and simulation. Now it is emerging.

NVIDIA's competitive moat

Clark Tang:

I think one of the most easily misunderstood things about NVIDIA is how deep the true NVIDIA model is. I think there is a notion that if someone invents a better chip, they win. But the fact is, you spend a decade building a complete stack from GPUs to CPUs to the network, particularly the software and libraries supporting applications. Running on NVIDIA. So, when you talk about this, when you think about NVIDIA's moat today, do you think today's video model is larger or smaller than three or four years ago?

Huang Renxun:

Okay, I really appreciate your recognition of how calculations are changing. In fact, people think (and many still do) that the reason for designing better chips is because they have more triggers, more bits, and bytes. Do you understand what I mean? You will see their keynote presentation slides. It has all these triggers, bar graphs, and things like that. All of these are good. What I mean is, look, horsepower is indeed important. Yes. So these things are fundamentally important.

However, unfortunately, these are all ideas. These are ideas in the sense that software is running some application on Windows and software is static, right? This means that the best way to improve the system is to create faster ships. But we realize that machine learning is not human coding. Machine learning is not just software, it's about the entire data pipeline. In fact, the flywheel of machine learning is the most important. So how do you view me enabling this flywheel? On one hand, enabling data scientists and researchers to work efficiently in this flywheel, which started from the beginning. Many people didn't even realize the need for AI to manage data, to teach AI. And AI itself is quite complex.

Brad Gerstner:

Is AI itself improving? Is it also accelerating? Again, when we consider competitive advantage, yes, that's right. It's a combination of all these.

Huang Renxun:

It is precisely because of smarter AI managing data that this situation arises. We now even have synthetic data generation and various different ways of presenting data to it. So before you undergo training, you have already been involved in a lot of data processing. So people would think, oh, Pytorch, this is the beginning of the world and also the end of the world. This is very important.

But don't forget, before and after Pytorch, the significance of the flywheel is in how you must think, how do I think about the whole flywheel, how to design a computing system, a computing architecture, to help you leverage this flywheel, make it as efficient as possible. It's not about the size of the application training. Does that make sense? It's just one step. Alright. Every step on the flywheel is challenging. So the first thing you should do is not think about how to make Excel faster, how to make Doom faster, that's in the past, right? Now you have to think about how to make this flywheel faster? This flywheel has many different steps, machine learning is not easy, you all know.

The things that the OpenAI or X or Gemini team is doing are not easy, they are carefully considering our needs. I mean, what they are doing is not easy. So we decided, you look, this is what you should consider. This is the whole process, you want to accelerate every part of it. You must respect Moore's Law, Moore's Law shows that if it's 30% of the time, and I speed it up three times, then I haven't really accelerated the entire process. Does that make sense? You really want to create a system to speed up every step, because only by doing the whole thing, can you really substantially improve the cycle time and the flywheel, which is the learning rate, ultimately leading to exponential growth.

So, what I want to say is that our view of what the company is really doing will be reflected in the products. Note, I have been talking about this flywheel, the entire website. Yes, that's right. We are accelerating everything.

Now, the main focus is on video. Many people are focused on physical AI and video processing. Imagine the front end. TB of data enters the system every second. For example, a pipeline will receive all this data. First, training needs to be prepared. Yes, this way the whole process can be accelerated.

Clark Tang:

Today, people only think about text models. Yes, but in the future, it's the video model, using some text models, like o1, to really process large amounts of data before we get there.

Huang Renxun:

Yes. So the language model will involve everything. But us, this industry has spent significant technology and effort to train the language models, training these large language models. Now, we are using large language forms at every step. This is amazing.

Brad Gerstner:

I hear you saying, in composite systems, yes, the advantage will grow over time. So I hear you say, our advantage today is greater than three to four years ago because we are improving every component. This is a combination, when you think of, for example, as a business case study, intel, relative to where you are now, it has the dominant mode, dominant position in the stack. Perhaps, to put it simply, compare your competitive advantage to theirs during their peak in the cycle.

Huang Renxun:

What sets Intel apart is the possibility that they may be the first company to excel in manufacturing process engineering and manufacturing. When it comes to manufacturing, it's about making chips. Designing chips, building chips in the x86 architecture, making faster x86 chips, that's where their talent lies, they integrate it with manufacturing.

Our company is a bit different, we recognize this fact, in fact, parallel processing does not require every transistor to perform well, while serial processing requires every transistor to perform well. Parallel processing requires a large number of transistors to be cost-effective. I would rather have ten times more transistors, 20% slower. Then ten times fewer transistors, 20% faster. Does this make sense? They want the opposite result. Therefore, single-thread performance, single-thread processing, and parallel processing are very different. Therefore, we observe that, in fact, our world is not getting better and better. We want to do very well, as best as possible, but our world is truly advancing constantly.

Parallel computing, parallel processing is difficult because each algorithm requires a different way of restructuring and re-architecting the algorithm. What people don't realize is that you can have three different CPUs. They each have their own C compilers. You can compile software to that axis.

This is not possible in accelerating calculations. Companies proposing architectures must come up with their own Open GL. So we have revolutionized deep learning because our domain-specific library is called cuDNN (Deep Neural Network Library), one domain-specific library is called optical. We have a domain-specific library called cuQuantum.

Brad Gerstner:

For industry-specific algorithms below, you know, everyone is focusing on Pytorch layers. Like I often hear.

Huang Renxun:

If we don't invent it, none of the applications above can work. Do you understand what I mean? So, algorithms are really what NVIDIA excels at. The dissemination of science on top of the underlying architecture is what we excel at.

NVIDIA is building a complete AI computing platform, including hardware, software, and ecosystem.

Clark Tang:

Now, all the attention is focused on reasoning. But I remember, two years ago, when Brad and I had dinner together, I asked you a question. Do you think your moat in reasoning will be as strong as in training?

Huang Renxun:

I'm not sure if I said it would be stronger.

Clark Tang:

You just mentioned a lot of these elements, the combinability between the two, or we don't know the overall combination. For customers, maintaining flexibility between the two is very important. But now that we are in the era of reasoning, can you talk about it?

Huang Renxun:

Reasoning training is to reason on that scale. What I mean is, you are correct. So, if you train properly, there is a good chance you will reason properly, it will run on this framework if you build it on this framework without any consideration. Well, you can still optimize it to fit other architectures, but at least, because it was built on nvidia architecture, it will run on nvidia.

Now, on the other hand, of course, it is also a capital investment aspect, that is, when you train new models, you want to train with your best new devices. This will leave out the devices you used yesterday, which are very suitable for inference. So behind the new infrastructure, there is a range of free devices that are compatible. Therefore, we rigorously ensure that we always remain compatible, so everything we leave behind will continue to excel.

Now, we are also putting a lot of effort into constantly reinventing new algorithms so that when the time comes, the Hopper architecture is two, three, four times better than when they were purchased, so that this infrastructure will continue to be truly effective. Therefore, all the work we do, improving new algorithms, new frameworks. Note that it helps us have every installed base. Hopper is better for it, Ampere is better for it, and even Volta is better for it.

I just heard from Sam Altman that they have recently stopped using OpenAI's Volta infrastructure. So I think, we leave traces of this installed base, just as all computing installed bases are important. Nvidia enters every cloud, including local and edge.

The VILA visual language model is created in the cloud, no modifications needed, and it can run perfectly on siasun robot&automation edges. They are both highly compatible. Therefore, I think architectural compatibility is very important for large devices, as well as for iPhones and other devices. I believe the installed base is crucial for inference.

Huang Renxun:

But what really benefits me greatly is that we are working on training these large language models in new architectures. We are able to think about how to create architectures that perform exceptionally well in reasoning when the time is right. Therefore, we have been thinking about the iterative model of inference models and how to create a very interactive inference experience for this, right, your personal agent. You don't want to leave and think for a while after finishing speaking. You want to interact with you quickly. So how do we create something like this?

This way, we can use these systems that are very suitable for training. However, when you complete it, the inference performance will be excellent. So you want to optimize this to the time of the first Token. Achieving the time of the first Token is actually very difficult because it requires a lot of bandwidth. But if your context is also rich, then you need a lot of FLOPS. Therefore, you need an unlimited amount of bandwidth and an unlimited amount of FLOPS to achieve a response time of a few milliseconds. So this architecture is really difficult to achieve. We have invented the great Blackwell MVLink for this.

Brad Gerstner:

Earlier this week, I had dinner with Andy Jassy (Amazon's President and CEO), and Andy said, we have Tranium, Inferentia coming soon. I think most people once again view these as Nvidia's issues. But then he said Nvidia is our important partner and will continue to be our important partner. As I see it, the future world will depend on Nvidia.

So when you think about the custom ASICs being built, they will be used for target applications. Maybe Meta's inference accelerators, maybe Amazon's training, or Google's TPU. And then you think about the supply shortages you are facing today, will these factors change this dynamic? Or will they complement the systems they buy from you?

Huang Renxun:

We are just doing different things. Yes, we are trying to accomplish different things. Nvidia is now trying to build a computing platform for this new world, this machine learning world, this generative AI world, this agentive AI world. We are trying to create, in the field of computing, to such a profound degree that after 60 years of development, we have reinvented the entire computing stack. From programming to machine learning, from CPU to GPU, from software to AI, applications from software to AI. From software tools to AI. Therefore, every aspect of the computing stack and technology stack has changed.

What we want to do is create a ubiquitous computing platform. This is indeed the complexity of our work, the complexity of our work is that if you think carefully about what we are doing, you will find that we are building the entire AI infrastructure, we see it as a computer. I have said before, the data center is now the unit of computation. For me, when I think of a computer, I am not thinking about chips. I am thinking about this thing. This is my mental model, with all the software, all the orchestration, all the machines inside, it is my mission. This is my computer.

Every year we try to build a new one. Yes, that's crazy. No one has ever done this before. Every year, we try to build a completely new one. Every year, we provide two to three times the performance. Therefore, every year, we reduce costs by two to three times. Every year, we improve energy efficiency by two to three times. Therefore, we ask customers not to buy everything at once, just buy a little each year, right? Ok. The reason for this is that we want their costs to remain average in the future. Now, everything is compatible in architecture, so it is very difficult to build these things separately at our current pace.

The double challenge right now is that we accept all of these, instead of selling them as infrastructure or services, we disagree with all of them. We integrate it into GCP, AWS, Azure, and X. So everyone's integration is different. We have to integrate all of our architecture libraries, all algorithms, and all frameworks into their frameworks. We integrate our security system into their system, we integrate our network into their system, right? Then we basically do 10 integrations, and now we do this every year. That's a miracle.

Brad Gerstner:

We, I mean, you try to do this every year, it's crazy. So what drives you to do this every year?

Huang Renxun:

Yes, that's when you systematically break it down. The more you break it down, the more surprised everyone is. Yeah. How the entire electronic ecosystem can be dedicated to working with us today, ultimately building a computer cube integrated into all these different ecosystems, and coordinating so seamlessly. So obviously what we propagate backward is APIs, methods, business processes, and design rules, while what we propagate forward is methods, architecture, and APIs.

Brad Gerstner:

That's how they were.

Huang Renxun:

For decades, they have been working hard. Yes, and they are also continuously developing along with our progress. However, these APs must be integrated together.

Clark Tang:

Some people just need to call the OpenAI API, and it works. That's it.

Huang Renxun:

Yes. Yes, it's a bit crazy. This is a whole. This is what we invented, this immense computing infrastructure, the entire planet is collaborating with us. It integrates anywhere. You can sell it through Dell, or you can sell it through HP. It's hosted in the cloud. It's ubiquitous. People are now using it in robot systems, robot and human robots, in self-driving cars. They are compatible in architecture. Pretty crazy.

Brad Gerstner:

This is too crazy.

Huang Renxun:

I don't want you to leave with the impression that I didn't answer the question. In fact, I did. When we talk about layered foundations, what I mean is the way we think. We are just doing things differently. Yes, as a company, we want to understand the situation, and I am very knowledgeable about everything around the company and the ecosystem, right?

I know everyone is doing other things, what they are doing. Sometimes this is not favorable to us, sometimes it is. I am very clear about this, but it doesn't change the company's goal. Yes, the company's sole goal is to build a platform architecture that can be ubiquitous. That's our goal.

We will not try to take shares from anyone. NVIDIA is a market maker, not a share taker. If you look at the slides our company hasn't shown, you'll find that this company doesn't talk about market share for a day, not internally. What we talk about is how we create the next thing?

What's the next problem we can solve in this flywheel? How can we better serve people? How can we shorten the flywheel that used to take about a year to about a month? Yes. What is the speed of light? Isn't it?

So we are considering all these different things, but one thing we won't, we won't, we don't have all the details on everything, but we are certain that our mission is very unique. The only question is whether this mission is necessary. Does it make sense? All companies, all great companies should have this as the core. It's about what you're doing, right?

Of course. The only question is, is it necessary? Is it valuable? Yes. Is it impactful? Is it helpful to people? I'm sure you are a developer, you are a generative AI startup, you are about to decide how to become a company.

One choice you don't have to make is which A6 I support? If you only support CUDA, you can go anywhere. You can always change your mind later. But we are the gateway to the world of AI, aren't we?

Once you decide to join our platform, you can postpone all other decisions. You can always build your own foundation later. We don't oppose that. We won't be angry about it. When I work with all GCP, GCP Azure, we show them our roadmap years in advance.

They have not shown us their basic roadmap, and this has never offended us. Does that make sense? We create, we are as one. If you have a unique goal, your goal is meaningful, your mission is valuable to you and others, then you can be transparent. Please note that my roadmap is transparent at GTC. My roadmap is deeper for our friends at Azure, AWS, and other companies. We have no problem doing any of these things, even if they are building their own assets.

Brad Gerstner:

I think when people look at the business, you recently said the demand for Blackwell has been crazy. You said one of the most challenging parts of the job is emotionally telling people 'no' in a world lacking the computational power you can produce and provide. But critics said these things. Hold on a moment. They said it's like Cisco in 2000, we are overbuilding fiber. This will be a cycle of prosperity and depression. I think back to the dinner we had at the beginning of 23. At that dinner in January 23, Nvidia's forecast was that the revenue in 2023 would reach $26 billion. You hit $60 billion.

Huang Renxun:

Let the facts be revealed. This is the biggest forecast failure in world history. Yes. At least we can admit that.

GPUs are playing an increasingly important role in AI computing.

Brad Gerstner:

That's right, we were very excited on November 22 because we had people like Mustafa from Inflection, without people from Character coming to our office to talk about investing in their company. They said, well, if you can't invest in our company, buy Nvidia, because everyone in the world is trying to get Nvidia chips to build these world-changing applications. Of course, the Cambrian moment happened on ChatGPT. Even so, these 25 analysts are still focused on crypto winners so much so that they cannot imagine what is happening in the world. So, in the end, the scale is larger. Put in very plain English, the demand for Blackwell is crazy, and as long as you can foresee, it will continue this way. Of course, the future is unknown and unpredictable. But why are critics so wrong, thinking it won't overbuild like Cisco did in 2000.

Huang Renxun:

The best way to think about the future is to start from first principles, right? Okay, so, what is the first principle of what we are doing about the problem? First, what are we doing? The first thing we are doing is reinventing computing, isn't it? As we just mentioned, the future of computing will be highly machine learning. Yes, highly machine learning. Okay, almost everything we do, almost every application, Word, Excel, Powerpoint, Photoshop, Premier, AutoCAD, your favorite applications, are all manually designed. I assure you, in the future, it will be highly machine learning. Right? So all these tools will be like that, most importantly, you will have machines, agents to help you use them. Okay. So now we know this is a fact. Right? We have reinvented computing. We're not turning back. The entire computing technology stack is being reinvented. Okay. Now that we have done that, we said software will be different. What software can write will be different. The way we use software will also be different. So let's acknowledge that now. So these are the basic facts for now. Yes.

The question now is, what will happen? Let's look back at past home computing. 1 trillion dollars were invested in past computing. We see, just open the door, look at the data centers, look at them. Are these computers the future you want? The answer is no. You have all those CPUs. We know what they can do, what they can't do. We only know we have 1 trillion dollars worth of data centers that need modernizing. So now, as we speak, if we are to modernize these old things in the next four to five years. That’s not unreasonable.

So we have a trend, you are talking to those who need to modernize it. Yes, they are modernizing it on GPUs. That's it.

Let's do the test again. You have 50 billion dollars in capital expenditure. Do you like to spend option A, option B, to build capital expenditure for the future, right?

Or build capital expenditure like in the past, now that you already have past capital expenditure, right? Yes, correct. It's already there. It hasn't improved much anyway. Moore's Law has basically ended. So why rebuild it?

We'll take out 50 billion dollars and invest in generative AI, right? So now your company becomes better. Right? How much of this 50 billion dollars would you invest now? Well, I would invest 100% of the 50 billion dollars because I already have the infrastructure from the past four years.

So now, from someone reasoning from first principles, this is what they are doing. Smart people are doing smart things. Now, the second part is this. So we have a trillion dollars worth of capacity. Go for it, Bill.

Infrastructure worth trillions of dollars. Probably around $150 billion. Okay. So we have $1 trillion worth of infrastructure to be built in the next four to five years. Well, the second thing we observe is that the way software is written is different, but the way software is used is also different.

In the future, we will have proxies. Our company will have digital employees. In your inbox, you'll see these small dots on the faces of these short icons. In the future, things will mean low icons of AIS. Right? I will send these to them.

I no longer program computers in C++, I program AI with prompts. Right? Now, this is no different from when I chatted with you this morning.

Before coming here, I wrote a lot of emails. Of course, I was guiding my team. I would describe the background, describe the basic constraints I knew, describe their tasks. I would leave enough space, give enough direction for them to understand what I need. I would try to explain as clearly as possible what the results should be, but I left enough room for ambiguity, a little creative space, so they could surprise me.

Right? This is no different from the way I prompt AI today. Yes, this is exactly how I propose AI. So, on top of the modernized infrastructure, there will be a new infrastructure. This new infrastructure will be an AI factory that operates these digital beings 24/7.

We will provide these devices to companies all over the world. We will own them in the factories, we will own them in autonomous systems. Right? So there's a whole layer of computing structures. This whole layer I call the AI factory, a world that must be manufactured, but doesn't exist at all today.

So the question is, how big is this. Currently unknown. It could be trillions of dollars. I know the current situation, but when we sit down to build, the beauty of it is that the modern architecture of this new datacenter is the same as the architecture of the AI factory. This is a good thing.

Brad Gerstner:

Can you be clear that you have a trillion old things? You must modernize. You have at least a trillion dollars of new AI workloads coming. Yes, your revenue this year will reach $125 billion. Someone once told you that the market cap of this company will never exceed $1 billion. What reason do you have sitting here today? Yes, if you only have $125 billion in a TAM of trillions, your future revenue will not be two or three times now. Is there a reason your revenue hasn't increased? No.

Huang Renxun:

As you know, not everything is like that, companies are only limited by the size of the fishpond, goldfish bowls can only be so big. So the question is, what is our fishpond? What is our pond? This requires a lot of imagination, that's why market makers consider the future without creating new fishponds. Looking back and trying to grab market share makes it difficult to understand this. Yes. Share gainers can only be this big. Of course. Market makers can be very big. Of course.

So, I think the good luck our company has is that from the very beginning of the company, we had to create a market to swim in it. People didn't realize it at the time, but now they have, and we are at the starting point of creating the 3D gaming PC market. We basically invented this market, as well as all the ecosystems and graphics card ecosystems, we invented all of it. So, there is a need to invent a new market to serve it in the future, which is a very comfortable thing for us.

Huang Renxun: I am happy about OpenAI's success

Brad Gerstner:

As is well known, OpenAI raised $6.5 billion with a valuation of $150 billion this week. We all participated.

Huang Renxun:

Yes, I am truly happy for them, really happy that they have come together. Yes, they have done something great, and the team has also performed well.

Brad Gerstner:

Reportedly, their revenue or operating income for this year will reach around 5 billion USD, possibly reaching 10 billion USD next year. If you look at the business today, its revenue is approximately double that of Google's IPO. They have 0.25 billion, yes, an average of 0.25 billion users per week, which we estimate is double that of Google's IPO. If you look at the company's P/E ratio, if you believe they will reach 10 billion USD next year, then it is about 15 times the expected revenue, similar to the P/E ratios of Google and Meta at their IPOs. Imagine a company with zero revenue and zero average users 22 months ago.

Tell us about the importance of OpenAI as a partner for you, and the power of OpenAI in driving public awareness and use of AI.

Huang Renxun:

Well, this is one of the most important companies of our time, a purely AI company pursuing the AGI vision. What its definition is, doesn't really matter to me. I almost don't think the definition or timing is important at all. What I do know is that AI will have a capability roadmap over time. And this capability roadmap will be very magnificent and fascinating. In this process, even before it reaches anyone's definition of AGI, we will make full use of it.

What you need to do now, while we are talking, is to talk to digital biologists, climate tech researchers, materials researchers, physicists, astrophysicists, quantum chemists. You can talk to video game designers, manufacturing engineers, robot experts. Choose your favorite. Whatever industry you choose, you need to delve into it, talk to important people, ask them how AI has completely changed your way of working. Collect these data points, then ask yourself how skeptical you want to be. Because they are not talking about AI's conceptual advantages. They are talking about using AI in the future. Now, in agriculture tech, materials tech, climate tech, choose your tech, choose your field of science. They are progressing. AI is helping them advance their work.

Now, as we said, every industry, every company, every height, every university. Incredible. Right? Absolutely. It's going to change business in some way. We know this. I mean, we know how real it is.

Today. It is happening. It is happening. So, I think the awakening of ChatGPT triggered it, which is absolutely incredible. I like their speed and their unique goals driving the development in this field, it's really important.

Brad Gerstner:

They built an economic engine that can fund the next frontier of models. I think Silicon Valley is forming a consensus that the entire model layer, commoditized Llama, enables many to build models at very low prices. So, in the early days, we had many model companies. These, features, tones, and cohesiveness are all listed.

Many question whether these companies can build escape velocity on the economic engine to continue funding the next generation. My own feeling is, this is why you see consolidation. OpenAI clearly achieved velocity. They can finance their own future. I am not sure many other companies can do the same. Is this a fair assessment of the current state of the model layer? We will, like in many other markets, consolidate this onto market leaders who can afford it, they have economic engines and applications to keep investing.

Merely having a powerful GPU does not guarantee a company's success in the AI field

Huang Renxun:

First, there is a fundamental distinction between models and AI. Yes. Models are an essential element. Yes. For AI, it is necessary but not sufficient. Yes. So, AI is a capability, but what is it used for, right? Then what is its application, right? The AI driving a software car has to do with human robots, but is not the same, the latter is related to chatbot AI, but is not the same.

So you have to understand taxonomy. Yes, the taxonomy of the stack. At every layer of the stack, there will be opportunities, but not every layer of the stack provides limitless opportunities for everyone.

Now, what I just said is that you replace the term model with GPU. In fact, this is a great observation from our company 32 years ago, that there is a fundamental difference between GPU, graphics chips or GPUs, and accelerated computing. Accelerated computing is different from what we do in ai infrastructure. They are related, but not entirely the same. They are layered on top of each other. They are not entirely the same. And each of these abstract layers requires completely different skills.

People who are truly good at building GPUs do not know how to become an accelerated computing company. I can give you an example, there are many people manufacturing GPUs. I don't know which one came later, we invented GPUs, but you know we are not, we are not the only company manufacturing GPUs today, right? GPUs are everywhere, but they are not accelerated computing companies. Many people do this. Their accelerators can accelerate applications, but this is different from being an accelerated computing company. For example, a very specialized AI application, right, can be a very successful thing, right?

Brad Gerstner:

This is MTIA (Mata's next-generation AI accelerator chip).

Huang Renxun:

Yes. But it may not be the kind of impactful and capable company. So you have to decide what kind of person you want to be. All these different fields may have opportunities. But just like building a company, you have to pay attention to the changes in the ecosystem and what will become commoditized over time, understand what is function, what is product, yes, what is a company. Okay. What I just said, well, you can think about this question in many different ways.

xAI and the Memphis supercomputer cluster have entered the era of "0.2 million to 0.3 million GPU clusters"

Brad Gerstner:

Of course, there is a new entrant who is rich, intelligent, and ambitious. That is xAI. Yes, indeed. Moreover, there are reports that you and Larry Ellison (Oracle's founder) and Musk had dinner together. They convinced you to abandon 100,000 H100 chips. They went to Memphis and within a few months set up a large coherent super cluster.

Huang Renxun:

Three points, don't equate them, okay? Yes, I had dinner with them.

Brad Gerstner:

Do you think they have the ability to build this super cluster? There are rumors that they want another one hundred thousand H200, right, to expand the scale of this super cluster. First, let's talk about X and their ambitions and achievements, but at the same time, have we reached the era of 0.2 million to 0.3 million GPU clusters?

Huang Renxun:

The answer is yes. First, acknowledge the achievements. From the moment of concept to the datacenter being ready for Nvidia to install our equipment there, to us powering it up, connecting it, and conducting the first training, everything is worthwhile.

Huang Renxun:

Okay. So the first part is to build a huge factory in such a short time, with water cooling, electricity supply, permits obtained - what I mean is, it's like Superman. Yes, as far as I know, there's only one person in the world who can do this. I mean, Musk's understanding of large system engineering, construction, and resource allocation is unparalleled. Yes, it's truly incredible. Of course, his engineering team is also excellent. I mean, the software team is great, the network team is great, the infrastructure team is great. Musk has a deep understanding of this.

From the moment we decided to start planning with the engineering team, network team, infrastructure computing team, and software team, all the preparations were made in advance. Then all the infrastructure, all the logistics, the number of technical equipment brought in that day, video infrastructure, computing infrastructure, and all the technology needed for training, everything remained undecided for 19 days, what do you want? Done.

Stepping back, do you know how long 19 days are? Are 19 days a few weeks, right? If you see it in person, the amount of technology is unbelievable. All the wiring and networks, the network of NVIDIA equipment is very different from that of a hyperscale data center. Okay, how many wires does one node need? The back of the computers is full of wires, and integrating this heap of technology and all the software together is really incredible.

So, I think what Musk and the X team have done, I am very grateful for him acknowledging the engineering work, planning, and so on that we did with him. But what they have achieved is unique, never seen before. Just from this perspective. One hundred thousand GPUs, as a cluster, could easily become the fastest supercomputer on Earth. Supercomputers you build typically require three years of planning. Then they deliver the equipment, and it takes a year to get them all up and running. Yes, we're talking about 19 days.

Clark Tang:

What credit does NVIDIA deserve?

Huang Renxun:

Everything is running smoothly. Yes, of course, there are piles of X algorithms, X frameworks, X stacks, and so on. We say we have a lot of reverse integration to do, but the planning is excellent. Just pre-planning.

Large-scale distributed computing is an important direction for the future development of AI.

Brad Gerstner:

One end is correct. Musk is one end. Yes, but when you answer this question, you start by saying, Yes, here we have 20 to 0.3 million GPU clusters. Yes, correct. Can this scale to 0.5 million? Can it scale to 1 million? Does your product demand depend on it scaling to 2 million?

Huang Renxun:

The last part is negative. My feeling is that distributed training must be effective. My feeling is that distributed computing will be invented. Some form of federated learning and distributed computing, asynchronous distributed computing will be discovered.

I am very enthusiastic and optimistic about this, of course, to be aware that the scaling law used to be about pre-training. Now we have shifted to multimodal, we have shifted to synthetic data generation, post-training has now expanded incredibly. Synthetic data generation, reward systems, based on reinforcement learning, and now inference scaling has peaked. A model has gone through an incredible 10,000 internal reasoning steps before giving you an answer.

This may not be unreasonable. It may have already completed tree search. It may have already undergone reinforcement learning on this basis. It may, it may have done some simulation, certainly a lot of reflection, maybe looked up some data, checked some information, right? So his background might be quite large. What I mean is, this type of intelligence is. Well, this is what we do. This is what we do. Right? So, for the capability, this kind of expansion, I just did the calculation and compounded it with model size and computation size quadrupling each year.

On the other hand, demand continues to grow in terms of usage. Do we think we need millions of GPUs? No doubt about it. Yes, now this is certain. So the question is, how do we build it from a datacenter perspective? It is largely about whether it is once a few gigawatts or 250 megawatts. My feeling is, you will get both at the same time.

Clark Tang:

I always think analysts focus on current architectural bets, but one of the biggest takeaways from this conversation for me is that you're considering the entire ecosystem and many years into the future. So, as Nvidia is just expanding or scaling up, it's to meet future demands. This doesn't mean you can only rely on a world with half a million or even a million GPU clusters. When distributed training emerges, you will write software to implement it.

Huang Renxun:

We developed Megatron seven years ago. Yes, these extensions for large training tasks will happen. So we invented Megatron, so all ongoing model parallelism, all breakthroughs in distributed training, all batching, and all that stuff is because we did the early work, and now we're doing the early work for the next generation.

AI is changing the way we work.

Brad Gerstner:

So let's talk about strawberries and o1. I think naming them o1 is cool. It means recruiting the best and brightest in the world and bringing them to the USA. I know we're all passionate about this. So I love the idea of creating a model of thought that takes us to the next level of expanding intelligence, right, paying tribute to the fact that it's these people who came to the USA through immigration that made us who we are today, bringing their collective wisdom to America.

Huang Renxun:

Of course. There's also extraterrestrial intelligence.

Brad Gerstner:

Of course. This is led by our friend Noam Brown. The importance of reasoning time as a new carrier for expanding intelligence is separate from simply building larger models.

Huang Renxun:

This is a big deal. This is a big deal. I think a lot of intelligence cannot be completed a priori. Right. A lot of computations, even a lot of computations cannot be reordered. What I mean is, unordered execution can be prioritized, many things can only be completed at runtime.

So, whether you are thinking from a computer science perspective or an intelligent perspective, too many things require context. Environment, right. And the quality, the type of answer you are looking for. Sometimes, a quick answer is enough. It depends on the consequences and impact of the answer. It depends on the nature of the answer's utility. So, some answers, take a night, some answers need a week.

Yes. Right? So I can totally imagine sending a prompt to my AI, telling it to consider overnight. Don't tell me immediately. I want you to consider all night, and then tell me tomorrow. What is your best answer and reasoning for me. So I think from a product perspective, the quality now, the segmentation of intelligence. There will be one-time versions. Of course. Some will need only five minutes.

Right? And humans. So if you're willing, we will become a huge workforce. Some of them are digital humans in AI, some are biological humans, and I hope some will even be super robots.

Brad Gerstner:

From a business perspective, this is something that is seriously misunderstood. You just described a company whose output is equivalent to having a workforce of 0.15 million people, but you achieved it with only 0.05 million people. That's right. Now, you didn't say I want to fire all employees. No. You are still increasing the number of employees in the organization, but the output of the organization will increase significantly.

Huang Renxun:

This is often misunderstood. AI is not me. AI will not change every job. AI will have a huge impact on how people work. Let's admit this. AI has the potential to bring incredible benefits. It also has the potential to cause harm. We must build safe AI. Yes, let's lay this foundation. Yes. Good.

Huang Renxun:

The part that people overlook is that when companies use AI to improve productivity, it is likely to result in better earnings or better growth, or both. When this happens, the next email from the CEO is most likely not about layoffs.

Brad Gerstner:

Of course it's an announcement, because you are growing.

Huang Renxun:

The reason is that we have more ideas, we can explore, we need people to help us carefully consider before automation. So in terms of automation, AI can help us achieve that. Obviously, it will also help us think, but we still need to figure out what problems I want to solve. There are tens of trillions of problems we can solve. So, the company needs to identify what problems to solve, choose these ideas, and find ways to automate and scale. Therefore, as we become more productive, we will hire more people. People forget that, if you go back in time, obviously we have more ideas today than 200 years ago. That's why GDP is bigger, and there are more jobs. Even though we are frantically automating at the base.

Brad Gerstner:

This is a very important point of this period, we are entering a time where almost all human productivity, almost all human prosperity is a byproduct of automation. The past 200 years of technology. What I mean is, you can look at the creative destruction of Adam Smith and Joseph Schumpeter, you can look at the chart of per capita GDP growth over the past 200 years, now it's accelerating.

Yes, this makes me think of this issue. If you look at the 90s, the United States' productivity growth rate was about 2.5% to 3% per year, right? Then in 2010, it slowed down to about 1.8%. And the past decade has been the slowest decade of productivity growth. So this is our fixed amount of labor and capital or output, it has actually been the slowest on record.

Many people are debating this reason. But if the world is really as you describe it, that we are harnessing and creating intelligence, are we on the verge of a sharp expansion in human productivity?

Huang Renxun:

This is our hope. This is our hope. Of course, we live in this world, so we have direct evidence.

We have direct evidence that either isolated cases or individual researchers who can use AI to explore science on an unimaginably large scale. That's productivity. One hundred percent measure productivity, or we're designing such incredible chips at such a high speed. The complexity of chips and computer complexity we are building is growing exponentially, and the company's employee base is not the standard for measuring productivity, right.

The software we are developing is getting better and better because we are using AI and supercomputers to assist us. The number of employees is almost linearly increasing. Another manifestation of productivity.

So, I can delve into research, I can sample and review many different industries. I can inspect personally. Yes, you're right. Business. Exactly.

So I can, of course, you can't, we can't, we may overfit. But its art is of course to summarize what we observe and whether it will be reflected in other industries.

Without a doubt, AI is the most valuable commodity known in the world. Now we are going to mass produce it. We, we, all of us must excel, what happens if you are surrounded by these AIs, they perform so well, much better than you. When I think back, this is my life. I have 60 direct subordinates.

They are world-class in their own fields, and they do better than me. Much better than me. I have no difficulty interacting with them, and I can effortlessly design them. I can also effortlessly program them. So I think what people need to learn is that they will all become CEOs.

They will all become CEOs of AI agents. They have the ability to be creative, uh, some knowledge, and how to reason, how to break down problems, so you can program these AIs to help you achieve goals like mine. That is running a company.

AI security requires multi-party efforts

Brad Gerstner:

Now. You mentioned something, which is the lack of coordination, safe AI. You mentioned the tragedy unfolding in the Middle East. We have a lot of autonomy, and a lot of AI is being used around the world. So let's talk about bad actors, secure AI, and coordination with Washington. How do you feel today? Are we on the right path? Do we have enough level of coordination? I think Mark Zuckerberg once said that the way to defeat bad AI is to make good AI better. How would you describe your view on how we ensure a positive net benefit for humanity, rather than living in this dystopian world.

Huang Renxun:

Discussion about security is indeed very important and good. Yes, abstract views, the conceptual view of AI as a huge neural network, is not as good, right. Okay. The reason being, as we all know, AI and large language models are related but not the same thing. I think a lot of things are being done very well. First, open source models so that the entire research community, every industry, and every company can be involved in AI and learn how to apply this capability. Very good.

Second, people underestimate the number of technical efforts dedicated to inventing AI to ensure AI safety. Yes, AI can process data, carry information, train, and create. AI is created to coordinate AI, generate synthetic data to expand AI's knowledge, and reduce its illusions. All are being created for vectorizing or graphifying or any other informing AI, protecting AI to monitor other AI AI systems, and the safe AI created by these AI systems is being praised, right?

Brad Gerstner:

So we've established.

Huang Renxun:

So. We are building all this. Yes, in the entire industry, methodologies, red teams, processes, model cards, evaluation systems, benchmarking systems, all of these, all of these are being built at an incredible pace. I wonder, celebrate. Do you understand? Yes, you know.

Brad Gerstner:

And, no, no, no government regulations say you have to do this. Yes, today participants building these AIs in this field are taking these key issues seriously and coordinating around best practices. Exactly.

Huang Renxun:

So this has not been fully appreciated, nor fully understood. Yes. Someone needs to, everyone needs to start talking about AI, it's an AI system, it's an engineering system, it's carefully designed, built from first principles, rigorously tested, and so on. Remember, AI is a capability that can be applied. I don't think it's necessary to regulate important technologies, but also not to over-regulate, so some regulation must be for the majority of applications. All the different ecosystems that already regulate technological applications must now regulate the technology applications integrated with AI.

So, I think, don't misunderstand, don't overlook the plethora of regulations that must be launched worldwide for AI. Don't rely on just one cosmic galaxy. An AI committee may be able to do this, as the establishment of all these different institutions serves a purpose. The establishment of all these different regulatory bodies serves a purpose. Going back to the original principle, I would.

The opposition between open source and closed source is wrong

Brad Gerstner:

You have launched a very important, very large, and very powerful open-source model.

Huang Renxun:

Nemotron.

Brad Gerstner:

Yes, it is clear that Meta has made a significant contribution to open source. I find that when I read Twitter, there are many discussions about open and closed. How do you view open source, your own open-source model, and if it can keep up with the forefront? This is the first question. The second question is, you know, having open-source and closed-source models, do they drive business operations, is this your view of the future? Do these two things create healthy tension for security?

Huang Renxun:

Open source and closed source are related to security, but not only security. For example, having closed-source models is absolutely not wrong, they are the engine of the economic model necessary to sustain innovation. Well, I completely agree with this. I think the opposition between closed and open is wrong.

Because openness is a necessary condition for activating many industries, now, if we don't have open source, how can all these different scientific fields be activated, activating AI? Because they must develop AI specific to their own areas, they must use open-source models to develop their AI, creating AI for specific areas. They are related, not, I repeat, not the same. Just because you have an open-source model does not mean you have AI. So you must have that open-source model to create AI. So financial services, medical care, transportation, these industries, scientific fields, are now able to be realized because of open source.

Brad Gerstner:

Unbelievable. Have you seen a huge demand for your open-source models?

Huang Renxun:

Our open-source model? First. Llama downloads. Obviously, yes, the work that Mark and the team have done is amazing. Beyond imagination. Yes. It has fully activated and engaged every industry, every scientific field.

Sure. The reason we do Nemotron is to generate synthetic data. Intuitively, an AI will somehow sit there in a loop and generate data to learn itself. This sounds fragile. How many times can you go around this infinite loop, this loop is questionable. However, the image in my mind is a bit like locking a super intelligent person in a padded room, closing the door for about a month, what comes out may not be a smarter person. So, so, but you can have two or three people sitting together, we have different AIs, we have different knowledge distributions, we can carry out quality assurance back and forth. The three of us can all become smarter.

Therefore, you can let AI models exchange, interact, pass back and forth, discuss reinforcement learning, synthetic data generation, etc., this idea makes intuitive sense, can propose suggestions and make sense. Therefore, our Nemotron 350B model is the best reward system model in the world. Therefore, this is the best critique.

Interesting. This is a great model that can enhance other models. Therefore, no matter how good other people's models are, I would recommend using Nemotron 340B to enhance and improve them. We have seen Llama get better, making all other models better.

Brad Gerstner:

As someone who delivered the DGX1 in 2016, this has truly been an incredible journey. Your journey is both incredible and unbelievable. It's truly remarkable how you survived in the early days. You delivered the first DGX1 in 2016. We ushered in the Cambrian moment in 2022.

So let me ask you a question I often wonder about – how long can you continue your current job with 60 direct reports? You are everywhere. You are driving this revolution. Are you having fun? Is there something else you would rather be doing?

Huang Renxun:

This is a question about the past hour and a half. The answer is "I really enjoy it." Great times. I can't imagine doing anything else I'd rather be doing. Let's see. I think we shouldn't leave people with the impression that our work is always fun. My work is not always fun, and I don't expect it to be always fun. Did I once expect it to be fun all the time? I think it's always important.

Yes, I don't take myself too seriously. I take my job very seriously. I take our responsibilities very seriously. I take our contributions and our moments very seriously.

Is it always fun? No. But have I always enjoyed it? Yes. Like all things, whether it's family, friends, or kids. Is it always fun? No. Do we always love it? Absolutely.

So I think, how long can I do this? The real question is, how long can I stay relevant? That's the most important thing, and the answer to that question will only be how will I keep on learning? Today I'm more optimistic. I say this not just because of our topic today. I'm more optimistic about my relevance and ability to keep learning because of AI. I use it every day, I don't know, but I believe you all use it too. I use it almost every day.

I don't have a study that doesn't involve AI. Yes, not a question, even if I know the answer, I'll check it with AI multiple times. Yes, surprisingly, the next two or three questions I ask reveal some things I didn't know. You choose your topics. You choose your topics. I think AI is a teacher.

AI is an assistant, AI is a partner, can brainstorm with me, check my work, guys, this is completely revolutionary. I am an information worker. I output information. So I think their contribution to society is remarkable. So I think, if that's the case, if I can maintain this relevance and continue to contribute, I know this job is important enough, yes, I want to keep pursuing it, my quality of life is incredible. So I will.

Brad Gerstner:

I can't imagine that you and I have been working in this field for decades, I can't imagine missing this moment. This is the most important moment of our careers. We are very grateful for this partnership.

Huang Renxun:

Looking forward to the next ten years.

Brad Gerstner:

Thought partnership. Yes, you make things smarter. Thank you. I think you as part of the leadership are really important, right, this will optimistically and safely lead everything forward. So thank you.

Huang Renxun:

Being with you all. I am really happy. Really. Thank you.

The translation is provided by third-party software.

The above content is for informational or educational purposes only and does not constitute any investment advice related to Futu. Although we strive to ensure the truthfulness, accuracy, and originality of all such content, we cannot guarantee it.

黄仁勋最新万字访谈：AGI即将来临，AI将彻底改变生产力

Huang Renxun's latest 10,000-word interview: AGI is coming soon, AI will completely change productivity.

AGI和AI助理的演变

英伟达的竞争护城河

英伟达正在构建一个完整的AI计算平台，包括硬件、软件和生态系统

GPU在AI计算中扮演着越来越重要的角色

黄仁勋：我为OpenAI的成功感到高兴

仅仅拥有强大的GPU并不能保证一家公司在AI领域取得成功

xAI和孟菲斯超级计算机集群已经到了“20万到30万个GPU集群的时代”

大规模分布式计算是未来AI发展的重要方向

AI改变了工作方式

AI安全需要多方共同努力

开源与不开源的对立是错误的

Evolution of AGI and AI assistants

NVIDIA's competitive moat

NVIDIA is building a complete AI computing platform, including hardware, software, and ecosystem.

GPUs are playing an increasingly important role in AI computing.

Huang Renxun: I am happy about OpenAI's success

Merely having a powerful GPU does not guarantee a company's success in the AI field

xAI and the Memphis supercomputer cluster have entered the era of "0.2 million to 0.3 million GPU clusters"

Large-scale distributed computing is an important direction for the future development of AI.

AI is changing the way we work.

AI security requires multi-party efforts

The opposition between open source and closed source is wrong

Risk Disclaimer

Statement