**Sarah Guo:** Jensen,非常感谢你今天来参加我们的节目。
**Sarah Guo:** Jensen, thanks so much for joining us today.
**Jensen Huang:** 太高兴能和你们在一起了。多么不可思议的一年啊。
**Jensen Huang:** So great to have you guys. What an amazing year.
**Sarah Guo:** 真的是不可思议的一年。
**Sarah Guo:** What a year.
**Jensen Huang:** 光明节快乐,圣诞快乐,新年也快到了。
**Jensen Huang:** Happy Hanukkah, merry Christmas, happy new year coming up.
**Elad Gil:** 是啊,节日快乐。
**Elad Gil:** Yep. Happy holidays.
**Sarah Guo:** 所以,2025 年发生了这么多事情,你又身处这场风暴的正中心,当你回顾的时候,什么让你最惊讶,或者你觉得最大的变化是什么?
**Sarah Guo:** So, uh, with everything that's happened in 2025, um, and you know, being in the middle of the vortex with it, what do you reflect on and say like this surprised you most or this is the biggest change?
**Jensen Huang:** 让我想想。有些事情并没有让我惊讶,比如说 scaling laws(扩展定律)并没有让我惊讶,因为我们已经知道这一点了。技术进步本身也没有让我惊讶。我对 grounding(事实基准)的改进感到欣慰。我对 reasoning(推理能力)的改进感到欣慰。我对所有模型和搜索的连接感到欣慰。我很高兴看到现在有了路由器(routers)放在这些模型前面,可以根据答案的置信度去做必要的研究,从而整体提升回答的质量和准确性。
我对此感到非常自豪。我认为整个行业解决了人们对 AI 最大的质疑之一,就是幻觉(hallucination)问题,以及生成无意义内容的问题。我觉得今年整个行业——从语言到视觉到机器人到自动驾驶——在推理能力的应用和答案的事实基准方面,都取得了巨大的飞跃。你们觉得今年怎么样?
**Jensen Huang:** Let's see. There there's some things that didn't surprise me like for example the scaling laws didn't surprise me because we already knew about that. The technology advancement didn't surprise me. I was pleased with the improvements of grounding. I was pleased with the improvements of reasoning. I was pleased with uh uh the connection of all of the models to to to search. I'm pleased that it that uh there are now routers that are in front of these models so that it could depending on the confidence of the answers go off and do necessary research and and just generally improve the quality and the accuracy of answers. I'm hugely proud of that. I think the whole industry addressed one of the biggest skeptical responses of AI which is hallucination and um generating gibberish and all of that stuff. I I thought that this year the whole industry everything from every every field from language to vision to robotics to self-driving cars the the application of reasoning and the grounding of the of of of the answers. Um big big leaps would you guys say this year?
**Sarah Guo:** 巨大的飞跃。我的意思是,像 Open Evidence 这样的产品用于医疗信息,医生们现在真的把它当作可信赖的资源在用;Harvey 用于法律领域。你真的开始看到 AI 成为一种可信赖的工具或合作方,让专家们能更好地完成他们的工作。
**Sarah Guo:** Huge. I mean things like open evidence too for medical information where doctors are now really using that as a trusted resource like you Harvey for legal you're really starting to see AI emerge as one of these things that's become a trusted tool or counterparty for you know experts to actually be able to do what they do much better.
**Jensen Huang:** 没错。所以在很多方面,我是预料到的,但我仍然为之高兴。我为此感到自豪。我为整个行业在这个领域的工作感到自豪。我非常欣慰,而且坦率地说有点惊讶,token 生成速率,特别是推理 token,增长得如此之快——好几个指数级同时在增长。而且我非常高兴看到这些 token 现在是有利可图的,人们在生成盈利的 token。我今天听到有人说 Open Evidence——说到他们——有 90% 的毛利率。我的意思是,这些是非常有利润的 token。
**Jensen Huang:** That's that's right. And so so in a lot of ways I was expecting it but I'm still pleased by it. I'm proud of it. I'm proud of all of the industry's work in this area. I'm really pleased and and uh uh and probably a little bit surprised in fact that token generation rate for inference especially reasoning tokens are growing so fast several exponentials at the same times it seems and uh and I'm so pleased that that these tokens are now profitable that people are generating I heard somebody hurt today that that open evidence speaking of them 90% gross margins I mean those are very profitable tokens.
**Sarah Guo:** 是的。
**Sarah Guo:** Yeah.
**Jensen Huang:** 所以他们显然在做非常有利可图、非常有价值的工作。Cursor 的利润率很好。Claude 的企业利润率很好。OpenAI 的企业利润率也很好。总之,看到我们现在生成的 token 质量足够好、价值足够高、人们愿意为之付出真金白银,这真的很棒。所以我觉得这些都是今年非常好的基础。
当然,一些其他的话题——与中国相关的话题确实占了我今年大量的时间。地缘政治,技术对每个国家的重要性。今年我花在全球各地出差的时间可能比我这辈子所有时间加起来还多。你知道我今年的平均海拔可能在 17000 英尺左右。所以能在地面上和你们在一起真好。所以我觉得地缘政治、AI 对所有国家的重要性,这些都值得以后再详谈。当然,我花了很多时间在出口管制上,确保我们的策略是有层次的、有根据的,能够促进国家安全,同时认识到国家安全各个方面的重要性。围绕这些有很多对话。当然还有大量关于就业的讨论,AI 的影响,能源,劳动力短缺。天哪,我们是不是什么都聊过了?是的。所有话题都和 AI 相关。
**Jensen Huang:** And so they're obviously doing very profitable, very valuable work. Cursor, their margins are great. Uh Claude's margins are great for the enterprise use of OpenAI. Their margins are great. Um so anyways, it's really terrific to see that that um we're now generating tokens that are sufficiently good, so good in value that that people are willing to pay good money for. And so I I think these are are really great grounding for the year. I mean some of the things that the narrative that that um uh of course the conversation with China really really you know occupied a lot of my my time this year. Geopolitics uh the importance of technology in each one of the countries. Uh I spent more time traveling around the world this year than just about any time in the h all of my life combined. You know my average elevation this year is probably about 17,000 ft. You know so so it's nice to be here on the ground with you guys. Um and so so I think uh geopolitics the importance of AI to all the nations uh all worth talking about later. You know of course I spent a lot of time on expert control and and making sure that our strategy is nuanced and uh really grounded and um uh promotes national security but recognizing the importance of various uh various facets of national security. Um a lot of conversations around that. Um, you know, of course, of course, uh, lots of conversation about jobs, the impact of AI, uh, energy, um, uh, labor shortage. I mean, boy, we covered everything, did we? Yeah. Everything was AI.
**Sarah Guo:** 所有话题都和 AI 相关。是的,太不可思议了。
**Sarah Guo:** Everything was AI. Yeah, it was incredible.
**Jensen Huang:** 是的,AI 绝对是所有这些主题的风暴中心。
**Jensen Huang:** Yeah, AI was definitely the center of the storm for like every one of those themes.
**Elad Gil:** 也许我们可以先从就业这个话题开始,因为当我看传统 AI 社区——甚至在 scaling 起来之前、在 AI 真正开始起作用之前——就有一种强烈的末日论调,而且很奇怪的是,最努力推动这个领域发展的人往往是最悲观的人,这非常矛盾——你为什么两件事同时做呢?我觉得这种叙事已经占据了一部分媒体和其他渠道,尽管我们认为 AI 在医疗、教育、生产力等方面有很多非常积极的作用。总的来说,每当我们经历一次技术变革,重要的岗位会发生变化,但你仍然会有更多的工作。
**Elad Gil:** Maybe one we can start with actually um is jobs because or there jobs and employment because when I look at the traditional AI community even before things were scaling and even before AI was really working there was a strong sort of doomsday component in the people working on AI oddly enough right the people who were most trying to push the field forward were often the people who are most pessimistic which is very odd why would you do both at once and I feel like that narrative has taken over some subset of media or some set of other things despite all the things that we think are very positive about what AI has done That's going to help with healthcare, with education, with productivity, with all these other areas. And in in general, whenever we have a technology shift, you have a shift in terms of the jobs that are important, but you still have more jobs.
**Jensen Huang:** 没错。
**Jensen Huang:** That's right.
**Elad Gil:** 你能谈谈你是怎么看待就业和工作的吗?人们在说什么,你认为真正的叙事是什么?
**Elad Gil:** Could you talk about how you think about employment and jobs and sort of what people are saying and what you think the real narrative is there?
**Jensen Huang:** 也许我可以从三个时间节点来谈。现在。
**Jensen Huang:** Maybe what I'll do is I'll I'll ground it on uh three points in space, three points in time. now.
**Sarah Guo:** 嗯。
**Sarah Guo:** Mhm.
**Jensen Huang:** 非常近的未来,然后是更远的某个时间点,也许还有一些反面叙事。关于近期就业,有一些别的角度需要考虑。最重要的一点是,AI 是软件,但它不是预录制的软件,你知道的。比如,Excel 是由几百名工程师编写的。他们编译好,就是预录制的,然后原样分发好几年。而 AI 的情况是,因为它要考虑上下文——你问了什么、世界上正在发生什么、各种上下文信息——它每一次都是第一次生成每一个 token。
这意味着每次你使用这个软件,我们所做的一切,AI 都是有史以来第一次被生成的。就像智能本身一样,我们今天的对话依赖于一些基本事实和一些知识,但每一个词都是第一次在这里生成的。AI 真正独特的地方在于,它每一次生成这些 token 都需要这些计算机。我称之为 AI 工厂,因为它在生产将被全世界使用的 token。有人可能会说它也是基础设施的一部分。之所以说它是基础设施,是因为它显然影响每一个应用程序,被每一个公司使用,被每一个行业、每一个国家使用。因此,它就像能源和互联网一样是基础设施。
正因为如此,以及生成这些 token 所需的计算量——这在以前从未发生过——因为我们需要这些工厂,三个新产业应运而生。准确地说,三种新型工厂需要建设。第一,我们需要建更多的芯片工厂。台积电在建,SK 海力士在建大量工厂,所以我们需要更多芯片工厂。第二,我们需要更多的计算机工厂。这些计算机非常不同。这些是世界上前所未见的超级计算机。Grace Blackwell 看起来和以往任何计算机都不一样。整个机架就是一个 GPU。所以我们需要新的超级计算机工厂。然后我们还需要新的 AI 工厂。这三种工厂目前正在美国大规模建设,遍布美国各地,这是第一次。
建筑工人、水管工、电工、技术人员、网络工程师——支撑这个新产业所需的技术劳动力数量将是巨大的。让我们面对现实吧。我太高兴听到电工们的薪水翻倍了。他们像我们一样出差。我们出商务差,他们也在出商务差。看到这三种工厂正在创造如此多的就业机会,真的很棒。
下一个方面是 AI 对就业的近期影响。我最喜欢的一个例子是——我很喜欢 Jeff Hinton——他大约五六七年前说过,五年之内 AI 将彻底颠覆放射学,每一个放射学应用都将由 AI 驱动,放射科医生将不再被需要,他建议第一个不要进入的职业就是放射科。他完全正确——100% 的放射学应用现在都是 AI 驱动的。这完全正确,大约八年过去了,AI 已经完全渗透了放射学。然而,有趣的是,放射科医生的数量增加了。
所以问题就变成了为什么。这就是工作的任务和目的之间的区别。一份工作有任务,也有目的。对于放射科医生来说,任务是研究扫描片,但目的是诊断疾病——
**Jensen Huang:** Uh maybe uh uh very near future and then some some point out out in the distance and and maybe maybe some counternarratives. Um something else to think about with respect to jobs in the near term. Uh one of the most important things is that that AI is not just AI is software but it's not pre-recorded software as you know. For example, Excel was written by several hundred engineers. They compiled it. It's pre-recorded and then they distribute it as is for several years. In the case of AI, because it takes into the context, what you asked of it, what's happening in the world, right? Contextual information, it generates every single token for the first time, every time. Which means every time you use the software and and everything that we do, AI is being generated for the first time ever. Just like intelligence, our conversation today relies on some, you know, ground truth and some knowledge and but it's every single word is being generated for the first time here. The thing that's really really quite unique about AI is that it needs these computers to generate these tokens every single time. I call them AI factories because it's producing tokens that will be, you know, used all over the world. Now, some people would say it's also part of infrastructure. The reason why it's infrastructure is because obviously it affects every single application. It's used in every single company. It's used in every single industry every single country. Therefore, it's part infrastructure like energy and and internet. Now, because of that and the amount of computers that's necessary to generate these tokens and it's never happened before and because we need these factories, three new industries have emerged. Number one, well, three new type of plants have to be created. Number one, we have to build a lot more chip plants. Mhm. TSMC is building, right? SKH Highix building a lot more plants and so we need more chip plants. We need more computer plants. These computers are very different. These are supercomputers that the world's never seen before. Right. Grace Blackwell looks like a very different type of computer than anything that's ever been made. And entire rack is one GPU. And so we need new supercomputer plants. And then we need new AI factories. These three plants are currently being met being built in the United States at very large scale quite broadly all over the United States for the very first time. The number of construction workers, plumbers, electricians, technicians, network engineers, you know, right? The the number of the skilled labor that's necessary to support this new industry in the near term, it'll be enormous. Let's just face it. Uh I'm [clears throat] so excited to hear that electricians are seeing their paychecks double. They're being they're being paid to travels like like us. We go on business trips. They're going on business trips. And so it's really terrific to see, you know, that this these three industries are now three types of plants, factories are just creating so much so much jobs. The next part is the the near-term impact of AI on jobs. And one of my favorites is um I love Jeff Hinton. uh he said uh you know some five six seven years ago that in five years time uh AI will completely revolutionize radiology that every single radiology application will be powered by AI and that radiologists uh will no longer be needed and that he would advise this the first profession not to go into is radiology and he's absolutely right 100% of radiology applications are now AI powered. That's completely true and in some eight years time it is now completely pervaded uh uh radiology. However, what's interesting is that the number of radiologists increased and so now the question is why and this is where the difference between task versus purpose of a job. A job has tasks and has purpose. And in the case of a radiologist, the task is to study scans, but the purpose is to diagnose disease and to research
**Sarah Guo:** 还有做研究。
**Sarah Guo:** and and that exactly and they're doing research.
**Jensen Huang:** 没错,他们在做研究。在他们的情况下,他们能够研究更多的扫描片、更深入地研究,能够申请更多的扫描检查,更好地诊断疾病,医院更高效了,能接待更多患者,这让他们赚更多钱,让他们想雇更多的放射科医生。所以问题是:你工作的目的是什么,而不是你工作中做的任务是什么?你知道,我大部分时间都在打字。[喷笑] 这是我的任务,但我的目的显然不是打字。所以如果有人能用 AI 来自动化我大量的打字工作,我真的很感激,这帮了大忙。但这并没有真的让我变得不那么忙。在很多方面,我变得更忙了,因为我能做更多的工作了。所以我认为这是第二个需要考虑的点——任务与工作的目的。
**Jensen Huang:** And so in the case in their case, the fact that they're able to study more scans more deeply, um they're able to uh request more scans, do a better job diagnosing disease, the hospital's more productive, they can have more patients, which allows them to make more money, which allows them to want to hire more radiologists. And so the question is what is the purpose of the job versus what is the task that you do in your job? And and as you know, I spend most of my day typing. [snorts] That's my task, but my purpose is obviously not typing. And so the fact that somebody could use AI to automate a lot of my typing, and I really appreciate that, and it helps a lot. Um, it hasn't really made me, if you will, less busy. In a lot of ways, I become more busy because I'm able to do more work. So, I think that that's the second part to consider is the task versus the purpose of the job.
**Sarah Guo:** 这个例子真的很有说服力,因为我嫂子 Erin 实际上在斯坦福负责核医学,对吧?所以她在放射科,随着所有这些技术进步的到来,这些医生真的欢迎它,他们每天工作 20 个小时,试图做更多研究、服务更多患者。
**Sarah Guo:** This example really strikes home because my my sister-in-law Erin actually leads um in nuclear medicine at Stanford, right? So, she's in radiology and with all the technology advancements that are coming, these doctors really welcome it and they are working 20 hours a day trying to do more research and serve more patients.
**Jensen Huang:** 没错。
**Jensen Huang:** Exactly.
**Sarah Guo:** 我觉得在基础设施投资创造的就业多样性之外,人们常常忽略的一点是,社会对各种商品实际上存在多大的潜在需求——比如更好的医疗服务。我不认为任何人觉得"嗯,我们已经达到了美国医疗或全球医疗的巅峰了"。我们越能让这些人高效工作,就会有越多的需求。
**Sarah Guo:** And and I think one thing that is often missed beyond the sort of um uh diversity of jobs being created by this investment in infrastructure is actually how much latent demand there is for different goods that we we need in society like better healthcare. I don't think anybody feels like you know what we have reached the the tiptop uh mountaintop of like what American healthcare or global healthcare could be and um the more we can make these people productive the more demand there will be
**Jensen Huang:** 完全正确。如果 Nvidia 更高效了,结果不是裁员,而是我们做更多的事情。
**Jensen Huang:** that's exactly right if I if Nvidia was more productive it doesn't result in layoffs it results in us doing more more things
**Sarah Guo:** 我今天见到了你们的新入职班,你们似乎每周都在招人。
**Sarah Guo:** I met your new hire class today you seem to be hiring every week anyway yeah
**Jensen Huang:** 没错没错。我们越高效,就能探索越多的想法,结果就是更多增长,我们变得更有利润,就可以追求更多的想法。所以我觉得你说得完全正确,如果这个世界的问题已经全部被定义好了、没有其他问题需要解决,那么生产力确实会缩减经济。但显然它会扩大经济。
我觉得下一个要考虑的部分是,人们说"天哪,所有这些机器人会抢走工作"。但我们清楚地知道,我们没有足够的工厂工人。我们的经济实际上受限于工厂工人的数量。大多数企业在留住工人方面非常困难。我们也知道全世界的卡车司机严重短缺。原因是人们不想做那些需要跨越全国、每天晚上住在不同地方的工作。人们想待在自己的城市,和家人在一起。所以我认为第一点是,机器人系统将帮助我们弥补劳动力短缺的缺口,这个缺口非常严重,而且由于人口老龄化还在加剧。这不仅是美国的问题,全世界都一样,你们知道的。
所以我们会弥补劳动力短缺。但人们忘记的第二点是,其他领域也存在短缺。
**Jensen Huang:** that's exactly right right the the more productive we are the more uh ideas we can explore uh the more growth as as a result the more profitable we become which allows us to pursue more ideas and so I think you're you're absolutely right that that if if the job if if your if your life if the world the problems is literally already specified and there's no other problem to solve then productivity would actually reduce the economy but it's clearly going to increase the e economy I think that the Next part that I would consider is, you know, people say, gosh, all of these robots that we're talking about, it's going to take away jobs. As as we know very clearly, we don't have enough factory workers. Our economy is actually limited by the number of factory workers we have. Most people are are having a very hard time retaining their workers. Um, we also know that the number of truck drivers in the world is severely short. And the reason for that is people don't want those jobs where you have to travel across the country and live in different parts of the world, different parts of the country, you know, every single night. So people want to stay in their town, stay with their families. And so I think that I think the first part is that having robotic systems is going to allow us to cover the labor shortage gap which is really really severe and getting worse because of aging population. This is this is not only United States, it's all over the world as you guys know. And so we're going to cover the labor shortage. But the second part that people forget and and as a result we'll go there are shortages as well in other places that people talk about AI being relevant.
**Elad Gil:** 会计就是一个有短缺的例子,护理也是。你可以列举很多其他行业说,好吧,这里有缺口,AI 正在帮助填补这些缺口。
**Elad Gil:** Accounting would be an example where there's shortages there. Nursing is another example. So you know you can you can go through multiple other industries and say okay there's gaps right and AI is trying to help fill those gaps.
**Jensen Huang:** 完全正确。所以自动化将帮助我们填补劳动力缺口。人们还忘了,当我们有汽车的时候,我们需要机修工来维护汽车。如果你看看今天路上已经有的无人出租车(robotaxi),这花了 10 年才走到今天。看看所有的维护团队、所有这些运营中心,你需要照顾这些无人出租车。现在想象一下我们有 10 亿个机器人。
**Jensen Huang:** That's exactly right. And so so um automation is going to help us increase and solve the the the the labor gap. Now people also don't don't remember that when we have cars, we need mechanics to take care of our cars. And if you look at the robo taxis that are that are even on the streets today, it's taken 10 years for that to happen. Look at all the maintenance crews and all of the the the various, you know, hubs that they're in where you have to take care of these robo taxis and just imagine we have a billion robots.
**Sarah Guo:** 嗯。
**Sarah Guo:** Mhm.
**Jensen Huang:** 那将是地球上最大的维修产业。所以我认为很多人没有想清楚这一点——
**Jensen Huang:** It's going to be the largest repair industry on the planet. So I I think a lot of people don't they they just have to think through
**Elad Gil:** 这就是你说的,当我们创造这类自动化的时候,我们同时创造了其他的工作。现在看看 AI 正在创造多少工作。
**Elad Gil:** and this is the part where you said um when we create this type of automation, we create this other job. Right now look at AI is creating so many jobs. Mhm.
**Sarah Guo:** 嗯。
**Sarah Guo:** Mhm.
**Elad Gil:** AI 产业正在创造一波就业热潮。
**Elad Gil:** The AI industry is creating a boom of jobs.
**Sarah Guo:** 我认为核心挑战之一是,人们很容易做简单的线性推断——比如"有工具帮助律师更高效,那就要取代律师了"。但实际上需要更进一步的推理才能看到:经济中对 AI 基础设施的一切都有巨大的吸力;对所有那些我们存在缺口的地方也有巨大的潜在需求。我觉得很多政策制定者关注的是"我们不能替换或减少现有的",但实际上在我们还没有满足的领域存在着更大的需求——
**Sarah Guo:** I think one of the core challenges here is it's very easy to draw a straight line of extrapolation from like oh you know uh there are tools that help lawyers be more productive. It's going to replace the lawyers but it's actually it takes like a step of incremental reasoning to say there's a sucking sound in the economy for everything in AI infrastructure. there's actually a sucking sound toward all of this demand that is latent in the places where we have gaps where um I think a lot of policy makers have focused on you know we can't replace or reduce what we have when it's really there's there's far more demand in what we actually are not
**Jensen Huang:** 就拿律师来说,律师的目的和律师的任务是什么?阅读合同、撰写合同不是律师的目的。律师的目的是帮你解决冲突。那不只是读合同写合同。律师的目的是保护你。那也不只是读合同写合同。所以我觉得真的非常重要的是回到工作的目的是什么,而不是我们用来执行工作的那个随时间变化的任务是什么。
**Jensen Huang:** and in the case of lawyer what's the what's the purpose of the lawyer versus the task of the lawyer reading a contract writing a contract is not the purpose of the lawyer the purpose of the lawyer is to help you resolve conflict And that's more than reading a contract. It's more than writing a contract. The purpose is to protect you. That's more than reading a contract. It's more than writing a contract. And so I think just it's really really important to go back to what is the purpose of the job versus the task that we use, you know, to perform that job that changes over time.
**Elad Gil:** 是的。今年另一个你提到的大主题,我觉得非常值得聊一聊的是中国,特别是中国开源的崛起。现在一些在基准测试中得分最高的模型是中国的开源模型——Qwen、DeepSeek 等等。在闭源方面仍然主要是美国的模型做得更好,但开源方面中国做得非常好。你一直是开源的倡导者,能分享一下你对中国 AI 开源崛起以及美国在开源和自身产业方面应该做什么的看法吗?
**Elad Gil:** Yeah. The other big theme of the year that you mentioned that I think is really important to touch upon is both uh China is sort of in the rise of Chinese open source in particular where you know some of the highest scoring models against benchmarks now are Chinese models on the open source side on the closer side it's still a lot of the US models but things like Quinn Deepseek etc are doing very well you've long been a proponent for open source in general could you could you share views about both China emerging for AI for open source and what the US should be doing in terms of both open source as well as its own industries
**Jensen Huang:** 当你面对这些复杂的、相互关联的、相互依赖的问题网络时——这团大大的问题交织——回到一个框架去理解我们在讨论什么,总是好的。
在 AI 的情况下,AI 是什么?当然,AI 的技术和能力本质上是关于自动化。这是第一次实现智能的自动化。你可以把它和机电一体化(megatronics)技术结合起来,赋予它实体形态,让它执行任务。所以这是 AI 自动化。
但让 AI 成为可能的技术栈是什么?功能栈是什么?最简单的思考方式是,它大概是一个五层蛋糕。最底层是能源。它把能源转化为我刚才描述的输出。下一层是芯片。再上一层是基础设施——包括硬件和软件,这就是你需要土地、电力和建筑的地方,数据中心在这里,编排用的软件栈也在这里——所以是软件和硬件。再上一层是大家都在想的那个层——AI,也就是模型。我们知道这些,但理解"AI 是一个模型系统"真的很有帮助。AI 是一种理解信息的技术,而信息包含人类信息,所以我们经常把 AI 想成聊天机器人。但别忘了还有生物信息、化学信息、各种物理信息、金融信息、医疗信息,各种模态、各种类型的信息。AI 真的非常广泛。当然,人类语言是很多事物的基础,但它不是一切的本质。因为你知道,生物分子不懂英语,它们懂的是别的东西。蛋白质不懂英语,它们懂的是别的东西。
然后最上层是应用,取决于行业。你刚才提到了 Open Evidence,提到了 Harvey,还有 Cursor,还有各种各样的应用。Full Self-Driving(完全自动驾驶)实际上是一个 AI 应用,被装在一辆机械汽车里。Figure 是一个 AI 应用,被装在一个机械人形体里。
所以这个五层栈是一种思考方式。然后另一种思考方式是,AI 确实非常多元。当你有了这个框架——技术能力是什么、如何构建技术、以及它有多多元——你就可以回过头来思考:开源有多重要?
没有开源的话,你知道,今天那些前沿模型,领先的实验室选择了闭源的应用模式,这完全没问题。公司如何决定自己的商业模式,说到底是他们自己的事,他们必须计算什么是获得投资回报的最佳方式,以便他们能扩大规模、取得更好的进展。他们做出的任何决定都很好。
但另一方面,没有开源的话,你知道,创业公司会很艰难。那些处于不同行业的公司——无论是制造业、交通运输还是医疗——没有开源的话,所有这些 AI 工作都会被扼杀。他们需要有一些预训练好的东西,需要一些关于推理的基础技术,在此基础上他们可以去适配、fine-tune、训练自己的 AI 模型,做成他们需要的领域和应用。人们真正忽略的是开源对所有这些行业的难以置信的渗透性和重要性。
一些百年老企业,我合作过的,在工业领域、在医疗领域——没有开源,它们会被扼杀,根本做不了这些事情。
**Jensen Huang:** when you Think about these complicated interconnected dependent um networks of problems. These this you know big goop of a mesh of problems it's always good to to go back and find a framework for what it is that we're talking about. In the case of AI um what is AI? Well, of course, the technology of AI and the capability, the capabilities of AI is about automation. It's about automation of intelligence for the very first time. And you could combine it with megatronics technology to embody that megatronics and and make it perform tasks. So that's what's AI automation. But what what is the stack that makes AI possible? What's the technology stack? functional stack. And of course the e the easiest way to think about that is it's kind of like a fivey year five year five layer cake which is at the lowest level is energy. Um it transforms energy to the output that I just described. The next layer is chips. The next layer is infrastructure and that infrastructure is both hardware software right this is where land power and shell this is where construction is data centers are the software stack you know for orchestrating the so it's software and hardware the layer above that is where everybody thinks about which is AI which is the models we know this but it's really helpful to understand that AI is a system of models and AI is a um a techn technology that understands information and there's human information and so we often times think about AI as a chatbot but remember there's biological information there's chemical information there's physical information of all kinds there's financial information there's healthcare information there's f there's information of all modalities all kinds AI is really really broad and of course human language is at the foundation of of many things but it's not the essence of everything because as you know you know biology molecules don't understand English they understand something else right proteins don't understand English they understand something else I think the next layer the important thing is is uh that's where the AI models are but there's a whole the AI is very very diverse and then the the layer above that is is applications and it depends on the industry and you already mentioned open evidence there you mentioned Harvey there's cursor there's all kinds of right there's all kinds of applications full self-driving is really an application, an AI application that is embodied into a mechanical car and figure is a AI application that has been embodied into a mechanical human. And so, so you got all these different applications. Well, this five layer stack is one way of thinking about it. And then the next way of thinking about I just mentioned is AI is really diverse. When you now have this framework of what the the technology capabilities are, how to how to build the technology and how diverse it is, then you can come back and think about okay, let's ask the question, how important is open source? Well, without open source, you know, today, of course, the frontier models, the the the leading labs have chosen to to use a closed source um application approach, which is just fine. you know what people decide to do with their business models is is really in the final analysis. It's their business and they have to they have to calculate what is the best way for them to get the return on investment so that they could scale up and and make better advances. Um however they they made that calculus is fantastic. On the other hand, uh without open source, as you know, startups would be challenged, uh companies that are in in uh uh different industries, whether it's manufacturing or transportation or um it could be in healthcare. Without open source today, all of that AI work would be suffocated. And so, they just need to have something that's pre-trained. They need to have some fundamental technology about reasoning. from that they could all adapt, fine-tune, you know, train their AI models into exactly the domain and application they want. And so what people really really miss is just the incredible pervasiveness and the importance of open source to all of these industries. large companies uh without without open source some of some of 100-year-old companies that I work with in in industrial spaces in healthcare spaces they would be suffocated they wouldn't be able to do that
**Elad Gil:** 开源在这一点上已经驱动了我们所有的数据中心,驱动了全球电信业的很大一部分——从 Android 到其他设备——它已经驱动了——
**Elad Gil:** open source at this point is driving all of our data centers is driving a big chunk of telefan in the world in terms of Android or other devices it's driving exactly
**Jensen Huang:** 没错,就像你说的,大量的工业应用。所以它已经无处不在了。我觉得更大的问题是——没有开源的话,高等教育怎么办?
**Jensen Huang:** you know to your point a lot of the industrial applications so it's already pervasive and I think the big question is open source without open source higher ed
**Elad Gil:** 高等教育就做不成了。
**Elad Gil:** higher ed wouldn't happen
**Jensen Huang:** 教育、研究、创业公司——名单还可以继续列下去。我们整天都在讨论那个最顶端的部分——最显眼的部分,最有新闻价值的部分——但在那之下是如此重要的开源 AI 空间。无论我们做出什么政策决定,都不要伤害那个创新飞轮。所以我花了大量时间教育政策制定者,帮助他们理解:无论你决定什么,无论你做什么,不要忘记开源。无论你决定什么,无论你做什么,不要忘记生物领域。
**Jensen Huang:** education research startups I mean the list goes on, you know, and so so we talk we talk all day long about the tip but the most visible part of that the most the part that's most newsworthy maybe but underneath that is such an important space of open source AI and whatever we decide to do with policies do not damage that innovation flywheel. So I spent a lot of time uh educating educating uh uh policy makers to help them understand whatever you decide whatever you do don't forget open source. Whatever you decide whatever you do don't forget biology.
**Sarah Guo:** 我觉得这里值得回应的反面叙事是,本质上有人认为应该有一个垄断的垂直整合公司和一个垄断性的资产——一个模型做所有事情——我们不能把这个皇冠上的宝石拱手让给其他国家或非美国公司。而你的论点是,我们实际上需要 AI 应用的巨大多样性,美国的优势——或者说任何主权国家的优势——在于整个技术栈,对吧?在于交付其中任何一个环节的能力。
**Sarah Guo:** I think the counternarrative here that is worth addressing is that essentially like you know there should be a monolithic vertical player and monolithic asset in the like one model that does it all and that we can't give away that crown jewel to other countries or non-American companies and and your your argument is like we actually need this huge diversity of AI applications and and the American advantage is actually or any any sovereign advantage is in the whole stack right? The capability to deliver any piece of it.
**Jensen Huang:** 我想也许有一天我们会有上帝 AI(God AI)。
**Jensen Huang:** I guess someday we will have God AI.
**Sarah Guo:** 但那一天是什么时候?
**Sarah Guo:** But when is that day?
**Jensen Huang:** 但那一天可能是在圣经尺度上的时间,你知道,我觉得是银河尺度上的。从我们今天所在的位置直接跳到上帝 AI 是没有意义的。我不认为任何公司实际上认为自己距离上帝 AI 很近。我也不认为任何研究人员有合理的能力去创造上帝 AI。能够极其出色地理解人类语言、基因组语言、分子语言、蛋白质语言、氨基酸语言、物理语言——那种上帝 AI 是不存在的。然而我们有大量需要 AI 的行业。
如果用最简单的话来说,AI 就是下一个计算机产业。你给我举一个不需要计算机的公司、行业或国家的例子?
**Jensen Huang:** But but that someday that someday is probably on biblical scales, you know, I think galactic scales. Um I I think it's it's not helpful to go from where we are today to God AI. And um I don't think any company practically believes they're anywhere near God AI. And nor nor do I do I see any researchers having any reasonable ability to create god AI. The ability to h understand human language and genome language and molecular language and protein language and amino acid language and physics language all supremely well. That god AI just doesn't exist. And and yet we have a lot of industries that need AI.
**Sarah Guo:** 嗯。
**Sarah Guo:** Mhm.
**Jensen Huang:** 我们不必都坐等上帝 AI 才能往前走,对吧?上帝 AI 下周不会出现。我对此相当确定。上帝 AI 明年也不会出现。但全世界下周、明年、下一个十年都需要继续前进。我认为那种一个巨无霸公司、国家拥有上帝 AI 的想法是——没有帮助的。没有帮助的。太极端了。
事实上,如果你要把它推到那个程度,那我们干脆全部停下来算了。还有什么意义呢?为什么还要政府?我的意思是,为什么他们还在制定政策?上帝 AI 会聪明到绕过任何政策。那还有什么意义呢?所以我认为我们应该把事情拉回到地面上,开始务实地思考,运用常识。
**Jensen Huang:** AI is if if you will at the simplistic level, it's just the next computer industry. And give me an example of a company, an industry, a nation who doesn't need computers. Mhm. And we all don't have to wait around for God AI for us to advance, right? So God AI is not showing up next week. I'm fairly certain of that. Okay. And God [clears throat] AI god AI is not not going to show up next year, but the whole world needs to move forward next week, next year, next decade. I think that that the idea of a monolithic gigantic company, country, nation, state that has got AI is just it's unhelpful. It's unhelpful. It's too extreme. Then in fact, if you want to take it to that level, then we ought to just all stop everything. What's the point of having even governments? I mean, why why why are they doing policies? God AI is going to be smart enough to avert, you know, work around any policy. And so, what's the point? And so, I I think that that we ought to bring things back to the ground ground level and start thinking about things practically and and use common sense.
**Elad Gil:** 这似乎是今天对话的一个大主题——有很多被放出来的东西,如果你真的仔细想想,是非常极端的。就业方面——没有人能再工作了。上帝 AI 会解决一切问题。因为 XYZ 原因我们不应该有开源,尽管开源已经驱动了我们大部分产业。
**Elad Gil:** This seems to be like a big theme in general in terms of this conversation where there's been a lot that's been kind of put out there that seems very extreme if you actually think about it. It's the jobs and employment. Nobody's going to be able to work again. It's God AI is going to solve every problem. It's we shouldn't have open source for XYZ reason despite open source powering much of our industries already.
**Jensen Huang:** 没错。
**Jensen Huang:** That's right.
**Elad Gil:** 所以总的来说,2025 年的一个主题可能是,公众对 AI 有很多极端的描述,如果你仔细审视,它们在短期内根本不合逻辑。
**Elad Gil:** And so it seems like in general maybe one of the themes of 2025 was there's a lot of extremes that were sort of painted in the public with AI that if you look at them very closely don't really follow a logical change in terms of happening anytime soon.
**Jensen Huang:** 是啊。所以进行这样的对话真的很重要。坦率地说,(那些极端叙事)非常有害。我认为一些非常受尊敬的人造成了很多伤害,他们描绘了末日论的叙事、世界末日的叙事、科幻小说的叙事。你知道,我理解我们很多人是在科幻小说中长大的,也享受科幻小说。但这没有帮助。这对人们没有帮助。对行业没有帮助。对社会没有帮助。对政府没有帮助。
政府中有很多人显然对技术不那么熟悉、不那么自如。当有博士学位的人、有 CEO 头衔的人去政府那里描述这些世界末日的场景、极度反乌托邦的未来时,你不得不问自己:这种叙事的目的是什么?他们的意图是什么?他们希望什么?他们为什么要跟政府说这些事情——是为了创造监管来扼杀创业公司吗?[清喉咙]
**Jensen Huang:** Yeah. And so it's it's it sounds like it's really important to have this conversation. Extremely hurtful frankly. And I I think we've done a lot of damage uh with very wellrespected people um who have who have painted a doom doomer narrative um end of the world narrative science fiction narrative and um you know and I and I appreciate that that many of us grew up in and enjoyed science fiction. Um but I but it's not helpful. It's not helpful to people. It's not helpful to the industry. It's not helpful to society. It's not helpful to the governments. Mhm. There are a lot of many people in the government who obviously aren't as familiar with as as comfortable with the technology and when PhDs of this and CEOs of that goes to governments and explain and describe these end of the world scenarios and extremely extremely dystopian future the future. Um, you have to ask yourself, you know, what is the purpose of that narrative and what is their what are their intentions and what do they hope? Why are they why are they talking to governments about these things to create regulations to suffocate startups? [clears throat]
**Elad Gil:** 他们这么做的原因是什么?你觉得这只是监管俘获(regulatory capture),他们试图阻止新创业公司出现并有效竞争,还是你觉得这些对话的目的是什么?
**Elad Gil:** For what reason would they be doing that, you know, and so and do you think that's just regulatory capture where they're trying to prevent uh new startups from showing up and being able to compete effectively or what do you think is the goal of some of these conversations?
**Jensen Huang:** 你知道,我没法猜测他们在想什么。我知道人们担心的是监管俘获。作为一种政策、一种做法,我不认为公司应该去政府那里为针对其他公司和其他行业的监管做游说。仅从实践角度看,他们的意图显然是高度矛盾的,他们的意图显然不完全是为了社会的最佳利益。我的意思是,他们显然是 CEO,显然代表着公司,显然在为自己做游说。
所以我觉得如果我们都能回到现在的位置,想想技术将走向何方。你看,就在一年前,正如我们开头讨论的,最让人自豪的时刻之一是行业能够非常积极地投资推进 AI 技术,而不是被拖慢。记得仅仅两年前,人们还在讨论要让行业慢下来。但当我们快速推进的时候,我们解决了什么?我们解决了 grounding,解决了推理,解决了研究。所有这些技术都被应用于改善 AI 的功能,而——
**Jensen Huang:** you know, I I can't I can't um uh guess what they what they have in mind. I know that the concern is regulatory capture. As a policy, as a practice, I don't think companies had to go to um governments to advocate for the regulation on other companies and other industries. just in practice their their intentions are clearly deeply conflicted and and uh their intentions are clearly you know not completely in the best interest of society. I mean they're obviously CEOs are obviously companies and obviously they're advocating for themselves and so so I think if we can all come back to where are we today and think about where the technology is going to be. I mean look lit literally in one year's time as we were talking about in the beginning uh some of the most proud moments is when the industry was able to invest very aggressively in advancing AI technology instead of being slowed down. Remember just two years ago people were talking about slowing the industry down but as we advanced quickly what did we solve? We solved grounding, we solved reasoning. We solved research. All of that technology was applied for good improving the functionality of the AI not you know yet the end has not come
**Sarah Guo:** 世界末日还没有到来。
**Sarah Guo:** yet the end has not come it's become more useful it's become more functional it's become able to do what we ask it to do you know and so the first the first part of the safety of a product is that it perform as advertised
**Jensen Huang:** 世界末日还没有到来。AI 变得更有用了,功能更强了,能做到我们要求它做的事了。所以产品安全的第一部分是它按照宣传的那样工作。安全的第一部分就是性能。就像汽车安全的第一部分不是某人会跳进车里把它当导弹用。汽车的第一部分是它按照宣传的那样工作。99.999% 的时间按照宣传工作。所以需要大量的技术来让那辆车或那个 AI 按照宣传工作。我很高兴过去两三年行业在提升 AI 的功能性方面投入了这么多。
我觉得如果我们展望未来 10 年,我们还有大量的工作要做来让它按照宣传工作。与此同时,你们两位投资了这么多生态系统中的公司——你们看到很多公司在做合成数据生成,让 AI 更有根据、更多样化、更少偏见、更安全。你们投资了大量网络安全公司,用 AI 来做网络安全。
人们说 AI 的边际成本将大幅下降——确实如此——因此 AI 会变得危险。恰恰相反。如果 AI 的边际成本大幅下降,那么一个 AI 将被数百万个 AI 监控。越来越多的 AI 将会互相监控。人们不能忘记,一个 AI 不会是一个孤立的 agent。很可能这个 AI 会被监控它的 agent 所包围。这就像如果维护社会安全的边际成本更低了,我们在每个角落都有警察一样。
**Jensen Huang:** the first part of safety is performance that it's is supposed like the first part of safety of a car isn't that some person is going to jump into the car and use it as a missile. The first part of the car is it works as advertised. Mhm. 99.999% of the time working as advertised. And so it takes a lot of technology to make that car or make that AI work as advertised. And I'm really glad that in the last couple two three years the industry has invested so much in enhancing the functionality of the AI as advertised. And I think if if we were to to look at the next 10 years, we have so much work to do to make it work as advertised. Meanwhile, as as you know, you both of you invest so much in in the in the ecosystem, you see so many companies being built for um synthetic data generation so that the AIs could be more grounded uh more diverse uh less biased more safe uh you're investing in a whole bunch of companies in cyber security using AI for cyber security you right people think that there's this AI um the marginal cost of AI is going to go go down significantly and it is and therefore the AI is going to be dangerous. It's exactly the opposite. If the marginal cost of AI is going to go down significantly, that one AI is going to be monitored by millions of AIS. Mhm. And more and more AI is going to be monitoring monitoring each other. People don't can't forget that an AI is not going to be an agent by itself. It's likely the AI is going to be surrounded by agents monitoring it. And so it's no different than if the if the marginal cost of of keeping society safe was lower. We have police in every corner. So
**Elad Gil:** 我们之前聊到的一个话题是 AI 的成本以及它如何在下降。我觉得在 2024 年,如果你看 GPT-4 等效模型每百万 token 的成本,下降了超过 100 倍。我团队里有人做了这个分析。所以成本在以相当惊人的速度和幅度下降,部分原因是你们在 Nvidia 层面推动的所有进步,但也包括整个技术栈的效率提升。
**Elad Gil:** So one thing that that we were talking about a little bit earlier was just the cost of AI and how it's been coming down. And so I I think um in 2024 the the cost of GPT4 equivalent models if you look at a million tokens it came down over 100x. Um you know somebody in my team did this analysis to show that. Uh so the costs are dropping pretty dramatically and very rapidly and part of it is all the advancements you all have been driving on and the Nvidia level but also just across the stack getting big efficiency gains.
**Jensen Huang:** 是的。
**Jensen Huang:** Yeah.
**Elad Gil:** 与此同时,模型公司在谈论成本如何上升,构建这些东西需要巨大的资本门槛。你怎么看训练成本和推理成本随时间的变化,以及这对普通终端用户或试图竞争的普通创业公司意味着什么?
**Elad Gil:** Um at the same time model companies are talking about how the costs are rising how there's enormous sort of capital modes to building these things out. How do you think about cost of training and cost of inference over time and what that means for the average end user or the average startup company trying to compete or people trying to do more in this industry?
**Jensen Huang:** 我忘了具体数据,但你知道 Andrej Karpathy 估算过构建第一个聊天机器人的成本——
**Jensen Huang:** I forget the statistic that but but you know Andre Andre Cararpathy um estimated the cost of building the first chatbt I think
**Elad Gil:** 和现在比——我觉得现在在 PC 上就能做了。
**Elad Gil:** versus now I think you could do that on the PC now.
**Jensen Huang:** 是的。是的。现在大概几万美元,甚至更少。
**Jensen Huang:** Yeah. Yeah. It's probably tens of thousands of dollars at this point or maybe even less.
**Elad Gil:** 对。所以基本上不花什么钱。
**Elad Gil:** Right. And so it costs nothing.
**Jensen Huang:** 对,而且他有一个开源项目,你一个周末就能做出来。
**Jensen Huang:** Mhm. And and he has an open source project that you can do in a weekend.
**Elad Gil:** 真的吗?好吧。太不可思议了。对。我们说的可是三年前的事情。
**Elad Gil:** Oh, is that right? Okay. That's incredible. Right. We're talking about three years. Mhm.
**Jensen Huang:** 嗯。
**Jensen Huang:** Mhm.
**Elad Gil:** 人们说需要花几十亿美元,需要建超级计算机,需要融几十亿美元才能做到的事情,现在——
**Elad Gil:** What people people said cost billions of dollars um supercomputers built raising billions of dollars in order to do all that now
**Jensen Huang:** 一个周末在 PC 上就能做到。所以这告诉你 AI 的成本效率在多快地提升——
**Jensen Huang:** cost you know something that you can do on a weekend on a PC. And so that tells you something about how quickly we're making making AI more cost effective
**Elad Gil:** 好吧,抱歉,可能还不完全是 PC。
**Elad Gil:** or Spark sorry probably not quite a PC.
**Jensen Huang:** 好吧,不完全是 PC。是的。我们每年都在改进架构和性能。我记得第一个 GPT 是在 Volta 上训练的。然后是 Ampere,而且最初的突破都没有用到 Hopper。然后当然是 Hopper 用了过去两三年,而我们已经在 Blackwell 上运行了大约一年半了。每一代架构都在改进,晶体管数量在增加,容量每一代都在大幅提升——从计算角度来看,非常轻松地每年都在进步。所有这些结合在一起,每年获得 5 到 10 倍的提升并不罕见。而 Rubin 马上就要来了。所以我们看到的是每年 5 到 10 倍。复合起来是不可思议的。摩尔定律是每一年半翻一倍,五年是 10 倍,10 年是 100 倍。而在 AI 的情况下,10 年大概是 10 万到 100 万倍。好吧,这还只是硬件。
然后下一层是算法层和模型层。所有这些结合在一起——如果你告诉我在 10 年的时间跨度里,我们将把 token 生成的成本降低大约 10 亿倍,我不会感到惊讶。
**Jensen Huang:** Okay. Not quite a PC. Yeah. We're improving our architecture and performance um every single year. The first GBTU I think was trained on Voltus. Mhm. And then uh Ampear um you know and and it wasn't I think the first breakthroughs none of it included Hopper. Mhm. And um of course Hopper the last couple two three years and um uh we're off in Blackwell for the last year and a half or so. And um every single one of these generations the architecture improves and of course the number of transistors go up and uh the capacity goes up every single generation very easily every every single year from a computing perspective. The combination of all that getting 5 to 10x every single year is not unusual. And here comes Reuben just around the corner. And so we're seeing 5 to 10x every single year. Well compounded it's incredible. Moore's law was two times every year and a half and over the course of five years is 10x over the course of 10 years is 100x in the in the in the case of AI over the course of 10 years is probably 100,000 to a millionx okay and that's just the hardware then the next layer is the algorithm layer and the model layer the combination of all that the fact that if you were to tell me that in the cost in the in the in in the span of you know 10 years we're going to reduce the cost of token generation about a billion times. I would not be surprised.
**Sarah Guo:** 嗯。
**Sarah Guo:** Mhm.
**Jensen Huang:** 好的。这就是 AI 的"token 经济学"(tokconomics)。在训练方面,成本下降没有那么激进,但也很接近。如果你说每年提升 2 到 3 倍,10 年下来也是不可思议的。但重要的观点是,当有人说训练某个东西花了 1 亿美元或 5 亿美元,好吧,明年就是十分之一的成本。后年又是十分之一。
**Jensen Huang:** Okay. And so that's the tokconomics of of of AI. On the training side, it's not quite as aggressive in in cost reduction, but it's close. If you were to say that that every single year we're increasing by two or 3x over the course of 10 years, incredible. But the important idea is when somebody says it cost $und00 million to train something or half a billion dollars to train something. Well, next year it's 10 times less. Next year it's 10 times.
**Elad Gil:** 但对于要扩大规模的人来说,反面论点是,我们每年会扩大 10 倍或 100 倍,试图用规模来抵消成本下降——然后其他人跟不上。
**Elad Gil:** For people to scale these things up, though, right? So the counter argument is, well, we'll just get bigger every year by 10x or 100x or, you know, we'll try to offset that decrease in cost by scale and others can't keep up.
**Jensen Huang:** 是的。但实际上发生的是——这就是关键所在——规模确实扩大了 10 倍,但计算负担并没有增加 10 倍,因为你获得了三件事的复合效益。硬件在进步,训练算法在进步,模型架构也在进步,而且我们在互相学习中获益。
坦率地说,DeepSeek 可能是过去两年硅谷研究者们阅读最多的论文。
**Jensen Huang:** Yeah. But really what's happening is is you're and and this is where come in as you know the scale went up by a factor of 10 but the computational burden did not go up by a factor of 10 because you're getting the compounded benefits of all three things. The hardware is going up the the algorithms of the training models are going up and of course the model architecture is going up and we're getting the benefit of learning from each other. This is, you know, let's face it, Deep Seek was probably the single most important paper that most Silicon Valley researchers read from in the last couple years.
**Sarah Guo:** 它是多年来唯一感觉像是前沿且开放的东西。
**Sarah Guo:** It was the only thing that felt frontier that was open.
**Jensen Huang:** 没错。这就是开源的价值——发布这些论文。DeepSeek 实实在在地惠及了美国的创业公司和美国的 AI 实验室——
**Jensen Huang:** That's right. In years, the value of open source again putting out these papers. Literally, Deep Seek benefited American startups and American AI labs all over
**Sarah Guo:** 还有基础设施公司。
**Sarah Guo:** and infrastructure companies
**Jensen Huang:** 到处都是。可能是去年对美国 AI 最大的单一贡献。如果你大声说出来,当然人们会——有点不安——美国 AI 实际上在从其他国家的 AI 中学习和受益。但为什么这会令人惊讶呢?你知道,全美国的 AI 研究人员中有很多中国裔,来自不同国家。我们从每个国家受益。我们从每个研究人员受益。世界上所有的想法不必都来自美国。
所以我觉得回到你最初的问题,确实有些关于 AI 成本的叙事是在吓唬所有人退出市场。"除了我们没有人应该做预训练,除了我们没有人应该训练这些前沿模型"。但由于模型、算法和计算栈的创新,AI 的成本实际上每年在下降远超 10 倍。所以如果你只落后一年甚至六个月,你完全可以保持很接近的距离。
**Jensen Huang:** and infrastructure company all over. probably the single greatest contribution to American AI last year. And so if you said this out loud, of course, you know, people kind of shudder um that we're uh American AI is actually getting learning from and benefiting from uh AI from other nation. But why would that be surprising? You know, AI researchers in all over America, all over America are uh Chinese natives and come from different countries. We benefit from every country. become benefit from every researcher and no all of the world's ideas don't have to come from the United States and so I I think um back to your your original question it is the case that you know some of the narratives around around the cost of AI is about scaring everybody out of the market you know nobody ought to do pre-training but us nobody should do you know training these frontier models but us because the because of innovation of models algorithms and the computing stack, the cost of AI is actually decreasing well more than 10x every single year. And so if you're just one year behind or even six months behind, you could you could really stay close.
**Sarah Guo:** 我觉得 2025 年有一个明显不同的感觉——Ilya 最近说过我们又回到了研究的时代,而不是 scaling 的时代。我觉得两件事同时在发生,顺便说一句。每个人也在试图在多个维度上 scale。
**Sarah Guo:** And I think one thing that felt very different to me about 2025 is um Ilia uh said recently that uh you know we're in the age of research again versus an age of scaling. I think both things are happening by the way. Everybody is also trying to scale on multiple dimensions.
**Jensen Huang:** 是的,没错。两者都在发生。
**Jensen Huang:** Yeah, exactly. both are happening.
**Sarah Guo:** 落后 6 个月或者在 100 集群 vs 200K 集群,如果你在对称竞争的话确实重要。但现在你有来自前沿实验室的人或者最顶尖的人,他们对如何从这里推进有非常不同的想法,或者在研究多样性的问题。我觉得这和 24 年的感觉不一样,24 年很多精力都集中在预训练规模和 LLM 上。
**Sarah Guo:** You know, being 6 months behind or being at 100 versus a 200k cluster, I think matters if you are competing symmetrically, but now you have people from Frontier Labs or um at the very top of the game who have very different ideas about how to progress from here or who are working on diversity of problems, right? Uh and and I I think that felt different from 24 maybe where there was a lot of energy focused on just pre-training scale and LLM.
**Jensen Huang:** 是的。还有其他几个动态。随着市场增长,每个模型可以选择垂直领域或细分市场来做差异化。有人可以选择做更好的编程。有人可以选择做得更易于使用,成为更好的消费者产品。这些模型的多样性。结果是,你可能可以在一个利基领域实现飞跃,而不必在其他所有方面都做到最好,仍然对市场非常有价值。不再需要煮沸整个海洋了。
两年前,因为它被叫做预训练——pre-training——人们说"好吧,预训练结束了"。首先,预训练没有结束。但预训练的要点是为训练做准备。这就是为什么它叫预训练(pre-training)——为真正的训练做准备。现在我们称之为后训练(post-training)。有点奇怪。我觉得那就是训练,但预训练是预训练,因此后面的就是训练。
训练,正如我们都知道的,是计算规模直接转化为智能的地方。现在训练一个模型所需的数据实际上相当小。也许只需要可验证的结果。现在它真的是算法密集型、计算密集型的。而且你不必在生活中什么都擅长——就像我们所有人一样,我们没有时间把所有东西都学得一样好。我们选择一个专长,把所有精力集中在上面,然后在某件事上变得超人般地厉害。所以我觉得 AI 实验室也会开始这样做。它们会开始分化到不同的细分领域,创业公司也会这样做。
**Jensen Huang:** Yeah. And several several other dynamics. Um, as the market grows, each one of these models could choose to have verticals or segments where they want to differentiate. Somebody could decide to be a better coder. Somebody could decide to be just better at being easier to be accessible so that it could be a greater consumer product. You know, the diversity of these models. As a result, you could you could probably make a niche leap without having to be great at everything else and still be super valuable to the market. It's no longer necessary to boil the entire ocean. The f two years ago, because it was called pre-training pre, you know, people people said, well, you know, pre-training is over. First of all, pre-training is not over. But the point of pre-training is to train yourself for training. That's why it's called pre-training to prepare yourself to do the real training. And now we call it post-training. It's kind of weird. I I think it's just training, but pre-training is pre-training and therefore it's training. Training as you as as we all know is is where uh compute scaling directly translates to intelligence. You you've you've largely now now this the the data the the data necessary to train a model is actually pretty small. Maybe it's just the verifiable results. Now it's really algorithmic, very compute intensive and so and you don't have to be good at everything in life as you know just like all of us we don't we could decide because we don't have time to learn everything equally well. We decide to choose a specialty and focus all of our energy on it and we become superhuman or incredibly good at something that other people are not. And so I think AI labs are going to start doing the same. They're going to start bifurcating into various segments and over time you're gonna and startups will do the same.
**Elad Gil:** 它们会找到一个微观利基,拿一些开源的东西,然后在上面做到极其出色。
**Elad Gil:** They'll find a micro niche and they'll take something open and then be incredibly good at it.
**Sarah Guo:** 我觉得这里最乐观的观点之一实际上是,这些微观利基相当有价值。我在和 Andrej 聊天——因为我一直在和很多人聊他们对明年的预测。当然我们也会问你的。他问我:"给一个去年回头看很有先见之明的预测的例子。"事后来看一切都很容易,但我的回答是:编程将成为第一个作为 AI 原生应用达到 10 亿美元年经常性收入(ARR)的应用级业务。我觉得如果你用旧世界的观点来看——你会相信两种叙事之一:一是单一模型做所有事情,一切都会被吸入某个巨无霸体内。
**Sarah Guo:** Well, I think one of the most optimistic views here is uh actually that these microniches are quite valuable, right? I was talking to Andre um because I've been talking to a lot of people about their predictions for next year. We'll ask you yours as well of course. Um um but he asked you know what is a what's an example of a prediction that would have been preient last year uh and my answer everything's easy in retrospect is that coding would be the first application level business that gets to a billion of AR as an AI native app right and I I think if you taken an old world view of this um you would have believed like one of two narratives right one is uh single model does everything and it'll all just be subsumed into something monolithic
**Jensen Huang:** 嗯。
**Jensen Huang:** Mhm.
**Sarah Guo:** 二是开发者工具永远做不大。好吧,那取决于开发者工具有多有价值。现在我觉得更多人理解了,软件工程是一个利基,而且对它的需求比以往任何时候都大。但我觉得我们会看到更多类似的情况。
**Sarah Guo:** And two is that developer tools never get very big, right? Well, kind of depends on how valuable the developer tool is. Now, I think many more people understand software engineering is in a niche and there's more demand than ever for it, but I think we'll see more like that.
**Jensen Huang:** 还有一个有意思的,我们这里在用 Cursor,而且我们广泛地使用它。每一个工程师都在用。而且工程师的数量,你刚才也提到了,我们现在招人的速度简直不可思议。对吧。周一是"来 Nvidia 上班日",为什么会这样呢?这又回到了目的和任务。软件工程师的目的是解决已知的问题并发现新问题去解决。编程是任务之一。
所以如果目的不是编程,如果你的目的仅仅就是编程——别人告诉你做什么你就写代码——好吧,也许你会被 AI 取代。但我们大多数的软件工程师——所有的软件——他们的目标是解决问题。而我们公司有太多问题了,还有太多未被发现的问题。所以他们越有时间去探索未被发现的问题,公司就越好。没有什么比他们全都不用写代码、只是在解决问题更让我开心的了。你明白我的意思吗?
所以我觉得这个目的 vs 任务的框架对每个人都很有用。比如说,一个服务员,他的工作不是记下点单。他的工作不是这个。实际上他的工作是让我们有一个很棒的用餐体验。如果某个 AI 在帮忙记点单,甚至送餐,他的工作仍然是帮我们有一个很棒的体验。他们会相应地重新塑造自己的工作。
所以关于计算成本的问题真的很重要。让我回到一个点——我们为什么如此执着于可编程架构而不是固定架构。记得很久以前,有 CNN 芯片出来了,人们说 Nvidia 完了。然后有 transformer 芯片出来了,Nvidia 完了。
**Jensen Huang:** Also interesting, uh we are using we we use cursor here and we use cursor pervasively here. Every engineer uses it and the number of engineers, you just mentioned it, the number of people we're hiring today is just incredible. Yep. Right. Monday is come to work at Nvidia day and and um uh why is that? Uh this is now the purpose and the task. The purpose of a software engineer is to solve known problems and to find new problems to solve. Coding is one of the tasks. And so if the purpose is not coding, if your purpose literally is coding, somebody tells you what to do, you code it. All right? Maybe you're going to get replaced by the AI. But most of our software engineers, all of our software, their goal is to solve problems. And it turns out we have so many problems in the company and we have so many undiscovered problems. And so the more time they have to go explore undiscovered problems, the better off we are as a company. Nothing would give me more joy than if none of them are coding at all. They're just solving problems. You see what I'm saying? And so I I think that this framework of purpose versus task is really good for everybody to apply. For example, somebody who's a waiter, their job is to not to take the order. That's not their job. As it turns out, their job is so that we have a great experience. And if somebody if some AI is taking the order, their job or even delivering the food, their job is still helping us have a great experience. They they would reshape their jobs accordingly. And so so I think the um the question about about cost of compute um uh is really important. Let's let let me come back to one the the reason why we are so dedicated to a programmable architecture versus a fixed architect. Remember a long time ago uh a CNN chip came along and they said Nvidia is done.
**Elad Gil:** 人们还在这么做呢。是的。
**Elad Gil:** And then and then a transformer chip came and Nvidia was done. People are still trying that. Yes.
**Jensen Huang:** 对。这些专用 ASIC 的好处当然是能很好地完成一项工作。transformer 是一种更通用的 AI 网络,但如你所知,transformer 的物种在快速增长——注意力机制如何思考上下文、扩散(diffusion)vs 自回归(auto-regressive)、混合 SSM(状态空间模型)-transformer 架构——比如我们刚刚发布的 Neotron 就是一种新的混合 SSM。所以 transformer 的架构实际上在快速变化,未来几年可能会发生巨大变化。
所以我们致力于一个灵活的架构。一方面是因为摩尔定律基本上已经结束了,晶体管的提升每年只有百分之十几。但我们希望每年有几百倍的提升。所以好处实际上全在算法上,而一个能支持任何算法的架构很可能是最好的——因为晶体管并没有进步那么多。
所以我们对可编程性的执着首先是出于这个原因——我们对算法创新有极大的乐观。第二个原因是,通过保护这个架构,我们的装机量非常大。当一个软件工程师想优化算法时,他想确保它不是只能在某一小片云或某一小个栈上运行。他想让它在尽可能多的计算机上运行。所以正因为我们保护了架构兼容性,flash attention 到处都能运行,SSM 到处都能运行,diffusion 到处都能运行,自回归到处都能运行。无论你想做什么都行。CNN 仍然到处能运行。LSTM 仍然到处能运行。
这个架构——向后兼容以支持庞大的装机量、可编程以面向未来——在我们推动技术进步方面真的非常重要。所有这些都在推动成本下降。[清喉咙] 我非常自豪,我们最新的创新 MVLink72 使我们成为世界上成本最低的 token 生成机器,优势巨大。原因是(这些问题)真的非常难。人们没有想到——4-bit 推理训练可能更容易,但推理时在上面生成 token 是极其困难的。
**Jensen Huang:** Yeah. NP and and the benefit of these dedicated AS6 of course it could perform a job really really well and transformers is a much more universal AI network but the transformer as you know the species of it is growing incredibly the attention mechanism the attention mechanism how it thinks about context diffusion versus auto reggressive these hybrid SSM transformation hybrid SSM for example Neotron we just announced a new hybrid SM SM and and so the architecture of transformer is in fact changing very rapidly and over the next several years it's likely to change tremendously and so we we dedicate ourselves to an architecture that's flexible for this reason so that we can on the one hand adapt with remember because MOS law is largely over transistor benefit is only tens 10% maybe a couple of years and yet we would like to have hundreds of X every year and so the benefit is actually all in algorithms and an architecture that enables any algorithm is likely going to be the best one right because the transistor didn't it didn't advance that much and so I I think the the our dedication to programmability is number one for that reason we have so much optimism for innovation and algorithms and iteration software that we protect our programmability for that reason the second thing is is by protecting this architecture our installed base is really large. When a software engineer wants to optimize their algorithm, they want to make sure that it doesn't run on just one this one little cloud or this one little stack. They want it to run on as many mo on as many computers as possible. So the the fact that we protect our architecture compatibility then flash attention runs everywhere. So SSM run everywhere, diffusion runs everywhere, auto reggression runs everywhere. Just depending it doesn't matter what you want to do. CNN still run everywhere. LSTM still runs everywhere. And so that this this architecture that is architecturally compatible so that we have a large installed base programmable for the future is really important in the way that we help to advance and as a result all of this drives the cost down [clears throat] and and I'm super proud that that um uh our latest innovation MVLink72 we're the lowest cost token generation machine in the world by enormous amounts and the reason for that is because are really really hard and so you know people didn't expect that um that forees it's probably easier to train but for inference is incredibly hard to generate tokens on but as as cost drop usually you open up new applications or new verticals that become more and more accessible
**Elad Gil:** 但随着成本下降,你通常会打开新的应用或新的垂直领域,变得越来越容易触达。我们刚才谈到了编程——Cursor、Cognition 和其他公司在过去一年真正从中受益。你对下一个突破性行业或新应用有什么想法或预测吗?在 2026 年你最兴奋的领域是什么?有没有一两个你觉得会——
**Elad Gil:** and we talked a little bit about coding like cursor and cognition and other companies that are really benefiting from that in this last year do you have any thoughts or predictions in terms of what the next breakthrough industries will be or new applications or areas that you're most excited about coming in 26 in particular like Are there one or two things that you think will
**Jensen Huang:** 因为两三件事情的结合,我觉得好几个行业将要迎来它们的 ChatGPT 时刻。我相信多模态(multi-modality)和超长上下文将当然会带来非常酷的聊天机器人。但基础架构加上合成数据生成的突破,将帮助创造数字生物学(digital biology)的 ChatGPT 时刻。那个时刻正在到来。
**Jensen Huang:** because of three things I because of because of a couple two three things I I think I think several industries are going to are going to experience their chat moment. Um I believe that multi-modality and um very long context is going to enable of course really really cool chat bots. Um but the basic architecture that in combination with breakthroughs in synthetic data generation is going to help create the chat GPT moment for digital biology. That moment is coming.
**Elad Gil:** 你说的数字生物学,具体指的是蛋白质折叠、蛋白质结合还是蛋白质诊断等其他方面?
**Elad Gil:** And by digital biology, do you specifically mean other aspects of like protein folding or protein binding or protein diagnosis? I see proteins.
**Jensen Huang:** 我看到的是蛋白质。我觉得我们在蛋白质理解方面已经做得很好了。现在多蛋白质理解正在上线,我们最近创建了一个叫 La Prina 的模型。它是开源的。它是用于多蛋白质理解、表征学习(representation learning)和生成的。所以蛋白质理解正在非常快速地推进。现在蛋白质生成将要快速推进。蛋白质的 ChatGPT 时刻。
**Jensen Huang:** I think we're good at protein understanding. Mhm. Now multi-proin understanding is coming online and we recently created a model called LA prina. It's open. Um it's for multi-proin um understanding and and represent representation learning and generation. Uh so so I think that the protein understanding is is advancing very quickly. Now protein generation is going to advance very quickly. Chat GPD moment proteins.
**Sarah Guo:** 是的。有很多有意思的公司在做端到端的分子设计,比如 Chai。
**Sarah Guo:** Yeah. There are a lot of interesting companies working on molecule design in endtoend way like chai.
**Jensen Huang:** 没错。然后当然是化学理解和化学生成,然后是蛋白质-化学构象的理解和生成。对吧?所有这些结合在一起——ChatGPT 时刻、生成式 AI 时刻——都在为数字生物学汇聚。
**Jensen Huang:** Exactly. And then and then of course chemical understanding and chemical generation and then protein chemical confirmation understanding and generation. Is that right? And so that combination the chat GBT moment the generative AI moment all of that stuff is coming together for for um digital biology
**Sarah Guo:** 说到你关于新行业的观点,我的思考方式是也要投资于 AI 的输入端。所有这些生物、化学、材料科学的事情,它们需要真实世界的数据生成和实验。那也是新的基础设施。
**Sarah Guo:** and to your to your point about like new industries or you know the way I think about it is like investing in the inputs for this AI as well. All of these things around biology and chemistry and material science, they require real world data generation and experimentation, right? And that's new infrastructure too.
**Jensen Huang:** 新的基础设施。合成数据将非常重要,因为他们的数据太稀疏了。他们没有像人类语言那么多的数据。真正的突破将是当我们能训练一个世界基础模型、一个蛋白质基础模型、一个细胞基础模型。我对这两者都非常兴奋。一旦我们有了基础模型,我们的理解能力、生成能力,那个数据飞轮就会真正起飞。
我兴奋的第二个领域——推理当然在语言方面取得了巨大的突破,但由于推理的进步,汽车也将表现得更好。不再只是感知汽车和规划汽车,它们将成为推理汽车。这些车将一直在思考。当它们遇到从未遇到过的情况时,它们可以把它分解成它们遇到过的情况,然后构建一个推理系统来导航通过。所以 AI 中分布外(out-of-distribution)的部分将在很大程度上被推理系统解决。结合生成式 AI、多模态视觉-语言-动作模型和推理系统,我觉得我们将看到人形机器人或多形态机器人的重大突破。
**Jensen Huang:** New infrastructure, uh, synthetic data is going to be really important because they just have such sparse, right? Spar sparity of data and they just don't have as much as human language. And there the the real breakthrough is going to be when we can train a a world foundation model, a foundation model for proteins, a foundation model for cells. I'm I'm very excited about both of those things. Once we have a a foundation model, our understanding capability, our generative capability, that data flywheel is really going to take off. The this this the second area that I'm excited about, um, of course, reasoning made huge breakthroughs in language, but because of reasoning, cars are going to be able to perform better. So, instead of just perception cars and planning cars, they're going to be reasoning cars. So, these cars are going to be thinking all the time. And when they come up they come up to a circumstance they they've never encountered before they can break it down into circumstances they have encountered it before and construct a reason reasoning system for how to navigate through it. And so the out of domain out of you know out of distribution part of AI is going to very much be be addressed by reasoning systems or and as a result we could do more things than we were taught to do between uh generative AI uh and um multimodal uh you know vision language action models and reasoning systems. I think we're going to see big breakthroughs in human robots or multi-mbodiment robots.
**Elad Gil:** 你觉得时间框架是什么?因为如果你看自动驾驶的类比——自动驾驶技术过去基于非常不同类型的神经网络,过去两三年发生了巨大的转变——
**Elad Gil:** you know does what do you think what do you think is a time frame for that because if you look at the self-driving analog and obviously self-driving technologies were based on very different types of neural networks than what we're using today in terms of you know there's been a big swap over the last two three years in terms of how we do a lot there
**Jensen Huang:** 我们开始得太早了。自动驾驶汽车实际上经历了四个时代。第一个时代是智能传感器连接到汽车上——Mobileye 时代。甚至 Waymo 最早期也是,你在用智能传感器加上大量人工工程化的算法,极端的测绘——
**Jensen Huang:** we started too soon self-driving cars really had four eras era was smart sensors connected into a car the mobile [clears throat] eye era the mobile eye era and even even the very earliest days of of Yeah. Even the earliest days of Whimo, the the um you're talk you're using smart sensors um a lot of human engineered algorithms and education severe mapping as far
**Sarah Guo:** 极端的地图测绘。
**Sarah Guo:** extreme mapping
**Jensen Huang:** 然后不同的感知和规划系统。没错。所以你本质上是在创建一辆在数字轨道上行驶的车,和迪士尼乐园的轨道没有区别。那是第一代。
第二代——在那一代中你有感知、世界模型和规划这些模块,每个模块都受到其技术的限制。感知首先受到深度学习的影响,然后传播到整个流程。但那个系统太脆弱了,它只知道如何执行你教它的东西。
现在我们到了端到端模型(end-to-end models)的阶段,然后下一步将是端到端模型的进一步发展。所以这大概是四个时代。在很多方面,如果我们大概三年前才开始做自动驾驶汽车,我们可能会在完全相同的位置。
**Jensen Huang:** mapping and then different systems for planning and perception. Exactly. And so so you're essentially creating a car that is driving on digital rails, right? It's no different than than the rails at Disneyland. There are digital rails. And so that's the first generation. the second generation. Um and during that generation you have perception, world model and planning. Mh. And and the these modules um and each one of these modules have the limits of their technology and and perception was first imple was was first affected by deep learning uh first and then and then uh and then it propagated through the pipeline. And so that but that system was too brittle and it only knows how to perform what you taught it. And now where we are are endtoend models and then and then where we're going to go next are end to end models. There you go. So that those are kind of the four eras in a lot of ways. If we would have started self-driving cars probably three years ago, we would probably be exactly the same place.
**Sarah Guo:** 那些在自动驾驶领域工作的朋友们可就惨了。
**Sarah Guo:** All our poor friends who were working in self-driving. Yeah.
**Jensen Huang:** 我倒不介意。我在这上面已经干了 10 年了。Nvidia 的自动驾驶栈,顺便说一下,安全性世界排名第一。我们刚在上周拿到了那个评级。第二名是 Tesla。所以我非常自豪两家美国公司在榜首——
**Jensen Huang:** And and I don't I don't mind it. I've been working on on it for 10 years. Nvidia's self-driving car stack, by the way, number one rated safety in the world today. Number one, we just got we just got that rating today uh last week. And number two is Tesla. So, I'm very proud that two American companies are up on the
**Sarah Guo:** 那么从机器人的角度来看,你觉得因为我们在现代已经建立了所有这些技术,机器人不会有同样的 10 到 15 年的周期?
**Sarah Guo:** Are you um So, from a robotics perspective, you think because we've already built all these sorts of technologies in the modern era, robotics won't have the same 10, 15 years. That's right.
**Jensen Huang:** 没错。我对机器人要乐观得多,因为我们已经推进了基础技术。现在人们在想人形机器人。人形机器人有很多挑战。有所有机电一体化方面的挑战。比如说,如果机器人重 300 磅,那可不好。要是它摔倒了怎么办?要和小孩互动怎么办?等等。你有各种挑战需要应对。我确信我们会解决这些问题。但记住,用于人形机器人的基础技术同样可以用于抓取放置(pick-and-place)机器人。
**Jensen Huang:** I'm much more optimistic with robotics because we we've kind of advanced foundational technology. Now, you know, people are thinking about human robotics. Human robotics has a lot of challenges. I mean, there's all the megatronics challenges there. You know, like for example, it's not helpful if the robot weighs 300 lb and what happens if it falls over and interacting with kids and so on so forth. And so, so you got all kinds of challenges to deal with. I'm certain that we're going to we're going to solve those. But remember the fundamental technology that goes into a human robot robot can go into a pick and place robot.
**Elad Gil:** 关于机器人我一直很好奇的一件事是——如果你看自动驾驶领域谁赢了或者被认为在赢,基本上是现有的大公司,对吧?Waymo、Tesla。你提到了 Nvidia 的安全评级。都是在这个领域工作了很长时间的人。需要大量资本。有供应链、硬件和所有这些额外的复杂性。你觉得机器人领域会是同样的情况吗?赢家基本上会是 Tesla 的 Optimus 和其他在这个行业里待了很久并且有这些先发优势的人?还是你觉得创业公司有空间?
**Elad Gil:** Um it could be it could be um how do you think about one thing I've been curious about for robotics in particular is if I look at who won or who who who's perceived as winning in self-driving. It's largely incumbents, right? It's Whimo, it's Tesla. You mentioned uh the safety rating Nvidia's gotten. And so it's people who've been working on this for a long time. It took a lot of capital. It was really intensive to get there. You have supply chain, you have hardware, you have all this extra complexity. Do you think the same thing will be true in robotics? Are the winners basically going to be Tesla with Optimus and other people who have both been in the industry for a while but also have all those sort of incumbent effects? Do you think there's room for startups?
**Jensen Huang:** 他们会是领导者之一,而且肯定是一个重要的。但所有会动的东西都将变成机器人。所有会动的东西。而"所有会动的东西"是一个非常大的空间。不全是人形机器人。而且每个 AI 都将是多形态的——就像人类一样,我们自己就是多形态 AI——我们可以坐进车里,操控那辆车。我们可以拿起网球拍,操控它。我们可以拿起筷子,操控它。
**Jensen Huang:** They will be one of the leader one of the one of them and and and surely a major one. Um but everything that moves will be robotic. Everything that moves will be robotic. And everything that moves is a very large space. It's not all human or robot. And yet every AI will be multi-mbodiment meaning you know just like just like a human with our m our multi-mbodiment AI ourselves we could sit in a car and embody that we could pick up a tennis racket embody that we could pick up a chopstick embody that and so we could embody the
**Elad Gil:** 人是通用的,对吧,可以做所有这些事情。
**Elad Gil:** people are general purpose right they can do all these things
**Jensen Huang:** 没错。所以 AI 也将变成通用的。你有单臂抓取放置,也许是双臂,也可以是六臂。你会有各种不同的尺寸和形状。可以是履带式的。可以是挖掘机。可以是各种东西。AI 将进入这些实体,就像建筑工人操控挖掘机、操控拖拉机一样。
**Jensen Huang:** exactly and so AIS are going to become general purpose so you have one arm pick and place maybe it's two arms pick and place could be six arms pick and place, you know. So, so I think you're going to have all kinds of different sizes and shapes. It could be a caterpillar. It could be, you know, it could be an excavator. It could be all kinds of stuff. And so AI will embody those just as a just as a a construction worker embodies an excavator embodies a tractor. You know, they you know,
**Elad Gil:** 会不会只有少数几家公司做所有东西的实体化,还是你说的更多是会有利基应用?
**Elad Gil:** could there be a small number of companies then that do the embodiment for everything or are you saying more there's going to be niche applications?
**Jensen Huang:** 你肯定会看到很多软件公司,那些软件公司可以服务很多不同的垂直领域。但每个垂直领域仍然会有解决方案提供商来把它落地,变成真正好用的东西。道理是这样的:对于面向消费者的 AI,如果它 90% 的时间好用,你就很开心了,觉得太神奇了。如果 80% 的时间好用,你也满意。但对于大多数工业和物理 AI,如果它 90% 的时间好用,没有人在乎那 90%。他们只在乎那 10% 出问题的时候。基本上就是 100% 不满意。所以你得把它做到 99.99999%。核心技术也许能让你到 99%。然后一个垂直解决方案提供商,比如 Caterpillar 或者其他公司,他们可以把核心技术做到 99.999%。
**Jensen Huang:** You should definitely see a lot of software companies and then those that software company could serve a lot of a lot of different verticals but each one of the verticals will still have solution providers that then grounds it all turns it into something that works perfectly. Does it make sense? Because in the case of AI for consumers if it works 90% of the time you're delighted you you're you know you're mind blown. If it works 80% of the time you're satisfied. In the case of most industrial and physical AIs, if it works 90% of the time, nobody cares about that. They only care about the 10% that it fails. Basically, you know, 100% dissatisfaction. And so, you got to take it to 99.99999. So, the core technology might be able to get get you to 99%. And then a vertical solution provider like a Caterpillar or somebody, they could take that core technology and make it 99.999% great.
**Elad Gil:** 你觉得最早期会是这样吗?因为在这么不成熟的市场里,通向市场最快的路径之一可能就是完全垂直化,因为你可以完全控制迭代速度——
**Elad Gil:** Do you think that's what happens like earliest on because in in markets that are this immature it seems one of the fastest paths to market could be full verticalization right because you just have control of iteration speed
**Jensen Huang:** 对于通用技术来说,垂直化的困难在于你没有足够的研发规模来构建通用技术。当然,开源在这方面帮助巨大。这也是为什么你将会看到未来几年 AI 垂直机会的大爆发。
我的预测是,未来五年,让人兴奋的将是垂直化。注意,我们对 Open Evidence 感到兴奋,对 Harvey 感到兴奋,对 Cursor 感到兴奋。Cursor 是个水平产品但又算是一种水平垂直。所以我对所有垂直领域都超级兴奋。
很多人说"AI 会变得这么好,上帝 AI 会变得这么好,所有这些封装公司都会过时"。这完全忽略了关键点。有人能谈论外科医生的生活,是因为他们从来没当过外科医生。有人做 AI 然后谈论会计师和税务专家的生活,是因为他们从来没当过税务专家。有人能谈论当杂工的生活而不当杂工,是因为他们从来没当过杂工。所以我觉得人们需要更有同理心一些,去理解工作的深层复杂性,真正理解工作的目的。很多时候技术解决的是任务,不是目的。
**Jensen Huang:** the different the the difficulty difficulty of of verticalization for technology that that is general purpose is that you don't have the R&D scale to build a general purpose technology. Now, of course, open source helps that tremendously, which is the reason why you're going to see a, you know, a a big surge of vertical opportunities in AI in the next several years. My my prediction would be over the course of the next five years, the excitement is going to be verticalization. Notice we we're excited about Open Evidence, we're excited about Harvey, we're excited about Cursor. cursor is is a horizontal but it's kind of a horizontal vertical you know and so um I'm I'm super excited about all the verticals you know a lot of people said yeah AI is gonna get so god AI is going to get so good that all these rapper companies are going to be obsolete it's just it misses the big point you know the reason why you could talk about the reason why somebody can talk talk about somebody is creating technology could talk about the life of a surgeon is because they've never been a surgeon the reason why somebody who builds at AI and talk talks about the life of a accountant and a tax, you know, a tax expert because they've never been a tax expert, you know, and so so I I think they just the reason why somebody could talk about being a bus boy without being a bus boy is they never been a bus boy. And so so I I think you you you've got to be a little bit more empathetic about the depth of the complexity of the work and and tr try to truly understand the purpose of the work. Often times the the technology addresses the task, it doesn't address the purpose.