Nvidia CEO Declares AGI Has Arrived as Company Targets $10 Trillion Valuation

英伟达CEO称AGI已至，公司目标估值10万亿美元 黄仁勋表示，能够推出十亿美元级服务的自主智能体意味着通用人工智能已经到来，并将推动向生成式Token工厂的必然转变。 在这场深入广泛的对话中，英伟达CEO黄仁勋探讨了公司如何在机架级工程和极致协同设计的推动下，从专用图形加速器演变为市值4万亿美元的AI巨头。他概述了AI的四大新扩展定律，借助合成数据的兴起重新审视了数据短缺问题，并详细介绍了大幅降低Token成本的硬件突破。此外，黄仁勋还坦诚分享了他的扁平化管理风格、基于第一性原理的“光速”工程哲学，以及由自主智能体驱动的数万亿美元生成式未来的必然趋势。 NVIDIA 正在从芯片级设计转向全机柜“极限协同设计”架构；CEO 黄仁勋采用直接管理 60 多名工程师的扁平化结构，摒弃一对一单聊，通过全员开放研讨攻关以突破分布式计算瓶颈。 (0:33) 早期在消费级 GeForce 显卡中全面普及 CUDA 的决策导致制造成本飙升 50% 并抹平了毛利润，公司市值一度跌至 15 亿美元，历经长达十年的补贴与深耕才奠定计算平台生态。 (10:42) AI 扩展定律已演化为预训练、后训练、测试时推理和智能体四大支柱；合成数据正逐步消除高质量人类数据短缺的瓶颈，而注重探索与推理的测试时扩展是极度依赖算力的核心工作负载。 (22:39) 由于现代机柜每架重达两至三吨，NVIDIA 已将超级计算机集成和每周高达 1 吉瓦的电力测试前移至供应链；单套 Vera Rubin Pod 拥有 1100 多颗 GPU，可提供 60 Exaflops 的极致性能。 (43:49)

NVIDIA CEO Jensen Huang argues that computing has fundamentally transitioned from file retrieval to real-time generative intelligence, transforming data centers into revenue-generating "AI factories." He contends that NVIDIA's market position is underpinned by full-stack, rack-scale co-design, the entrenched CUDA software ecosystem, and an ongoing expansion across pre-training, test-time, and agentic scaling laws. Extreme Rack-Scale Co-Design and Supply Chain Assembly (0:33) Huang asserts that modern artificial intelligence workloads have outgrown individual accelerators, compelling NVIDIA to transition from chip-level engineering to "extreme co-design" across CPUs, GPUs, networking, memory, cooling, power delivery, and physical data center racks. To bypass distributed computing bottlenecks framed by Amdahl's Law, the company developed tightly coupled architectures like the Vera Rubin pod—a system combining seven chip types, five rack types, and 40 racks connected via NVLink 72 to run multi-trillion-parameter mixture-of-experts models as a unified domain. This architectural shift has relocated supercomputer assembly from individual client data centers directly into the manufacturing supply chain. Dense systems are now manufactured and shipped as complete multi-ton units, requiring testing facilities to draw up to a gigawatt of power weekly prior to shipping. Huang explains that this engineering approach is governed by a "speed of light" framework, which models every hardware metric, manufacturing step, and software pipeline against the theoretical limits of physics rather than accepting incremental efficiency gains. Four Scaling Laws and Compute-Intensive Reasoning (22:39) Huang dismisses warnings that AI progress will stall due to a scarcity of human-generated training data, stating that scaling has expanded across four pillars: pre-training, post-training, test-time, and agentic scaling. He contends that synthetic data generated during post-training is bypassing data supply limits, making training data generation bounded only by available compute capacity. Furthermore, Huang disputes forecasts that inference will commoditize onto lightweight, low-power chips. Because advanced inference relies on test-time scaling—requiring search, planning, task decomposition, and first-principles exploration—he characterizes inference as computationally demanding reasoning rather than static pattern recall. Autonomous agentic systems amplify these compute requirements by performing continuous research and tool interaction, creating a loop where synthetic outputs continually refine baseline models while token generation costs drop by roughly an order of magnitude each year. Moats in CUDA, Ecosystems, and TSMC Orchestration (75:02) When evaluating NVIDIA's competitive moat, Huang emphasizes the two-decade install base of the CUDA platform over raw transistor performance. Supported by millions of external developers and 43,000 employees, CUDA provides developers with cross-platform reach spanning hyper-scale cloud providers, enterprise data centers, autonomous vehicles, and orbital satellites, alongside an assurance of backward compatibility and sustained optimization. This software layer is reinforced by NVIDIA's horizontal distribution and its operational relationship with TSMC. Huang states that TSMC's primary advantage is manufacturing orchestration—managing dynamic volume fluctuations, emergency wafer runs, and packaging technologies such as CoWoS—built on decades of trust rather than restrictive written contracts. This manufacturing foundation allows NVIDIA to deliver fully integrated proprietary stacks while remaining deployed across competing cloud and enterprise ecosystems. Flat Management, Task Automation, and Agentic Labor (3:16) To direct cross-disciplinary engineering, Huang operates a flat corporate structure with over 60 direct reports, eliminating traditional one-on-one meetings in favor of open group reasoning sessions. Strategic decisions and engineering trade-offs are vetted transparently across disciplines to socialize concepts incrementally, ensuring organization-wide alignment before executing major capital or technical transitions. Addressing fears of workforce displacement, Huang draws a historical parallel to radiology: despite computer vision achieving superhuman diagnostic capabilities around 2020, total radiologist employment expanded because automated scans increased clinical throughput. He argues that software engineering will experience a similar trajectory, as coding represents a task rather than the core purpose of collaborative problem-solving. In Huang's view, natural language specification tools will expand the global developer base from roughly 30 million to 1 billion individuals, positioning autonomous agents as operational multipliers rather than autonomous corporate replacements. While Huang views NVIDIA's continued expansion into a multi-trillion-dollar factory model as inevitable, his thesis depends on the continuous availability of massive industrial electrical power, global supply-chain stability, and sustained commercial demand for token generation. Unresolved questions remain regarding how geopolitical tensions, power grid constraints, and physical packaging limits will affect the pace of this multi-gigawatt infrastructure buildout.

英伟达首席执行官黄仁勋在访谈中指出，人工智能的发展正推动计算架构从单一芯片加速转向涵盖机架与数据中心的极端协同设计，并促使传统数据中心演变为实时生成Token的AI工厂。他主张，随着扩展定律延伸至合成数据与高算力推理，结合CUDA积累的生态信任与物理极限工程哲学，英伟达正通过重构计算底层基础设施推动产业生产力与组织模式的深刻变革。 计算架构跃迁：从单芯片到AI工厂 (0:33) 黄仁勋阐述了计算体系由芯片级优化向数据中心级全栈协同设计转移的技术动因。他指出，由于现代前沿模型规模已无法容纳于单台设备或单块GPU中，提升单个芯片性能已不足以突破阿姆达尔定律在分布式计算中带来的扩展损耗。为了让计算集群的性能增速超越物理机器的堆叠规模，工程体系必须对算法、数据及模型进行重构与切分，从而催生了将GPU、CPU、高速互连、供电、液冷、机架以至数据中心进行端到端整合的“极端协同设计”。如NVLink 72与Vera Rubin机架架构，其设计目标是支持4万亿至10万亿参数的混合专家（MoE）模型在单一计算域内无缝运行。 这一工程跃进反映了计算底层模式的范式转移：计算已从过去依靠预录文件和推荐引擎检索数据的存储仓库，彻底演进为根据上下文实时生成数字商品（Token）的AI工厂。黄仁勋预测，这一转型使数据中心产出直接与企业顶线收入挂钩，并在算力推动下驱动Token生成成本以每年约一个数量级的速度持续下降。 扩展定律演化与算力密集型推理 (22:39) 针对高质量人类数据耗尽将导致AI发展停滞的论调，黄仁勋认为数据规模的物理边界正在被算力本身攻破。他表示，人类自身的信息传递在本质上大多属于对已有知识的消化、加工与重构，具有合成属性；随着模型成熟，后训练阶段正大规模引入由基础事实强化的合成数据，使得训练瓶颈从人类数据供给彻底转向可用算力规模。 在此基础上，黄仁勋将扩展定律归纳为四大支柱：预训练、后训练、测试时间（推理）扩展以及智能体（Agentic）扩展。他驳斥了推理计算门槛低、易被轻量化芯片商品化的行业预测，直言“推理即思考，而思考极其困难”。面对未知经验时，测试时间扩展高度依赖搜索、推演与第一性原理规划，必然需要密集算力支撑。此外，由子智能体构成的闭环工作流能够自主调用工具与文件执行研究，其产生的交互体验数据再次回流至预训练与后训练体系，形成了智能随算力扩张而持续递增的自强化闭环。 供应链前置集成与物理极限工程哲学 (43:49) 为缓解先进光刻机、CoWoS封装与高带宽内存（HBM）等全产业链的供给制约，英伟达打破了传统在数据中心内部现场组装超级计算机的惯例，将系统级集成与调试全面移至供应链工厂制造环节。以两至三吨重的NVLink-72整机架为例，大规模生产周期要求制造与测试工厂自身在短期内具备每周吉瓦级的电力供给。面对这种指数级扩张，黄仁勋频繁与半导体高管协调，在无正式书面合同的前提下依靠深厚信任推动上下游进行数十亿美元规模的前置资本开支。 在底层研发方法论上，黄仁勋坚持30年前确立的“光速”基准哲学。该体系将所有数学吞吐、内存带宽、功耗、物料成本与生产周期直接对标物理定律的绝对极限，以此作为反推工程可行性的基准。他批判仅追求边际收益的渐进优化思维，要求团队从零推演理论最优解。面对每周需制造约200个Pod的Vera Rubin复杂系统，英伟达坚持“在必要前提下保持复杂，但尽可能追求极简”的设计原则，坚决剔除一切无法证明其必要性的冗余构造。 CUDA生态壁垒与扁平化推演体系 (75:02) 黄仁勋将英伟达的根本壁垒归结为其二十年来建立的CUDA平台安装基数以及跨全行业的横向覆盖。该架构目前已全面部署于主流公有云、主权超级计算机、边缘基站乃至轨道卫星中。开发者选择CUDA不仅获得了庞大的通用硬件触达能力，更获得了平台持续升级维护的确定性，这种确定性沉淀为行业内的“信任速度”。为促进整体生态的技术探索，英伟达亦开源了Nemotron等前沿模型架构，以此反哺底层硬件的协同设计，并协助医药、生物等外部实体工业发展专有模型。 为确保组织在极端复杂的软硬件交叉领域维持高敏捷度，英伟达构建了极度扁平的架构，由黄仁勋直接管理60多位具有深厚工程背景的高管。公司完全取消传统的一对一汇报，采用全员参与的公开讨论机制，将每个具体技术难题置于全体领导层面前推演，促使跨部门专家即时发现依赖矛盾并协同纠偏。面对外部庞大的地缘与商业压力，黄仁勋依托结构化拆解问题与“系统性遗忘”化解焦虑，并通过在日常会议中公开即时推演逻辑来完成组织传承，而非依赖传统的接班人流程。 职业范式重塑与AGI边界现实 (118:28) 针对白领群体与软件工程行业的自动化失业焦虑，黄仁勋提出了区分“职业本质使命”与“具体执行任务”的分析视角。他以医学影像领域为例指出：尽管计算机视觉早在数年前便达到超越人类专家的水平，但全球放射科医生人数不降反升，原因在于工具降本增效放大了医院诊疗通量，进而催生了更多以诊治患者为核心使命的人才需求。同理，软件工程师的核心价值在于协作界定问题与设计系统架构，代码仅是实现手段；随着交互门槛降低，具备编程能力的人群预计将从3000万专业人员扩展至10亿普通大众。 关于通用人工智能（AGI）的实现周期，黄仁勋指出这取决于评判标准：若将AGI定义为AI智能体能够端到端创立并运营一家短期估值超10亿美元的企业，此类现象在当下已属可能；但若指望分布式自主智能体完全自发组织并建立如英伟达般规模庞大且持久的实体工业企业，其概率在当前依然为零。同时他认为，硅基芯片虽然能够统计建模与解析情绪，但本质上无法产生人类的主观情感与内在意识体验。 黄仁勋描绘了一幅由算力扩张驱动各行业重塑的技术前景，其逻辑建立在算力需求指数级增长以及生成式计算替代检索计算的不可逆趋势上。然而，超大规模数据中心制造与部署对全球电网负荷施加的物理压力、百亿美元级前置投资对下游商业回报周期的严苛要求，以及空间在轨计算面临的热辐射耗散等工程极限，依然构成了这一宏大愿景落地过程中必须直面的客观挑战。

[3:36] Jensen Huang: No conversation is ever one person [10:42] Jensen Huang: Install base defines an architecture. Not... Everything else is secondary [15:38] Jensen Huang: I manifest a future and that future is so convincing, there's no way it won't happen [19:16] Jensen Huang: Jensen, what took you so long? [25:15] Jensen Huang: inference is thinking, and I think thinking is hard [38:40] Jensen Huang: no company in history has ever grown at a scale that we're growing while accelerating that growth [43:49] Jensen Huang: NVLink-72 literally builds supercomputers in the supply chain and ships 'em two, three tons at a time per rack. It used to be—they used to come in parts and we used to assemble 'em inside the data center. But that's impossible now because NVLink-72 is so dense. [52:44] Jensen Huang: The speed of light is not just about the speed. The speed of light is, is my shorthand for what's what's the limit of what physics can do. And so every single, everything, everything that we do is compared against the speed of light. [59:30] Jensen Huang: we need things to be as complex as necessary, but as simple as possible [98:15] Jensen Huang: then stop crying about it [101:47] Jensen Huang: How hard can it be? [115:10] Jensen Huang: I think it's now. I think we've achieved AGI. [121:03] Jensen Huang: I didn't care how many lines of code they wrote [133:34] Jensen Huang: what is it about a dishwasher that allows that dishwasher to sit in the middle of superhumans? [137:21] Jensen Huang: I really don't wanna die. [144:31] Alan Kay: The best way to predict the future is to invent it.

[0:00] - The following is a conversation with Jensen Huang, [0:03] CEO of NVIDIA, one of the most important and influential [0:07] companies in the history of human civilization. [0:11] NVIDIA is the engine powering the AI revolution, and a lot [0:15] of its success can be directly attributed to Jensen's sheer force of [0:19] will and his many brilliant bets and decisions as a leader, engineer, and innovator. [0:26] This is Lex Fridman Podcast. And now dear friends, here's Jensen Huang. [0:33] You've propelled NVIDIA into a new era in [0:37] AI, moving beyond its focus on chip scale design to now rack scale [0:41] design. And I think it's fair to say that winning for NVIDIA for a [0:45] long time used to be about building the best GPU possible, and you still [0:49] do, but now you've expanded that to extreme co-design of [0:53] GPU, CPU memory, networking, storage, power cooling, [0:57] software, the rack itself, the pod that you've announced, and even the [1:01] data center. So let's talk about extreme co-design. What is the [1:05] hardest part of co-designing system with that many complex [1:08] components and design variables? [1:11] - Yeah, thanks for that question. [1:12] So first of all, the reason why extreme co-design is necessary is because the [1:16] problem no longer fits inside one computer to be accelerated by one GPU. [1:24] The problem that you're trying to solve is you would like to go faster [1:28] than the number of computers that you add. So you added [1:32] you know, 10,000 computers, but you would like it to go a million times [1:36] faster. Then all of a sudden you have to take the algorithm, [1:43] you have to break up the algorithm, you have to refactor it, [1:46] you have to shard the pipeline, you have to shard the data, you have to shard the [1:50] model. Now all of a sudden when you distribute the problem this way, [1:56] not just scaling up the problem, but you're distributing the problem, [2:00] then everything gets in the way. This is the Amdahl's Law [2:04] problem where the amount of speed up you have for something [2:09] depends on how much of the total workload it is. And so [2:13] if computation represents 50% of the problem, and I sped up computation infinitely [2:21] like a million times, [2:23] you know, I only sped up the total workload by a factor of two. [2:27] Now all of a sudden, not only do you have to distribute a [2:30] computation, you have to, you know, shard the pipeline somehow. [2:34] You also have to solve the networking problem [2:38] because you've got all of these computers are all connected together. [2:42] And so distributed computing at the scale that we do, [2:47] the CPU is a problem, the GPU is a problem, the networking is a problem, the [2:51] switching is a problem. And distributing the workload across all [2:55] these computers is a problem. [2:57] It's just a massively complex computer science problem. And so we just [3:00] gotta bring every technology to bear. Otherwise, we scale up linearly or we [3:09] scale up based on the capabilities of Moore's Law, which has [3:13] largely slowed because Dennard scaling has slowed. [3:16] - I'm sure there's trade-offs there. Plus you have a complete disparate [3:20] disciplines here. I'm sure you have specialists in each one of these high bandwidth [3:24] memory, the network and the NVLink, the NICs, the optics and the [3:28] copper that you're doing, the power delivery, the cooling, all of that. I mean, there's like world [3:31] experts in each of those. How do you get 'em in a room together to figure out- [3:34] - That's why my staff is so large. [3:37] - What's the process—can you take me through the process of the specialists and the [3:41] generalists? Like how do you put together the rack when you know the [3:45] s- the set of things you have to shove into a rack together? [3:48] Like what does that process look like of designing it all together? [3:51] - Yeah. There's the first question, which is: what is extreme co-design? [3:55] You're, you, we're optimizing across the entire stack of software [3:59] from architectures to chips, to systems, to system software, to the [4:02] algorithms, to the applications. That's one layer. The second thing that you and [4:06] I just talked about is goes beyond CPUs and GPUs and [4:10] networking chips and scale up switches and scale out switches. [4:15] And then of course, you gotta include power and cooling and all [4:19] of that because, you know, all these computers are extremely, [4:23] extremely power hungry. They do a lot of work and they're very [4:27] energy efficient, but they in aggregate still consume a lot of [4:31] power. And so that's one. The first question is, what is it? [4:34] The second question is, why is it, and we just spoke about the reason, you [4:38] know you want to distribute the workload so that you can exceed [4:42] the benefit of just increasing the number of computers. [4:47] And then the third question is, how is it, how do you do it? [4:51] And that's the, that's kind of the miracle of this [4:55] company. You know, when you're designing a computer, you have to have operating system of [4:59] computers. When you're designing a company, [5:02] you should first think about what is it that you want the company to produce. You know, I see [5:06] a lot of companies organization charts, and they all look the same. [5:09] Hamburger organization charts, soft organization charts, and car [5:13] company organization charts. They all look the same. [5:16] And it doesn't make any sense to me. You know, the goal of a company is to be the [5:20] company is to be the machinery, the mechanism, the system that produces [5:25] the output. And that output is the product that we like to create. [5:29] It is also designed, the architecture of the company should reflect [5:33] the environment by which it exists. It almost indirectly [5:37] says what you should do with the organization. My direct staff is 60 people. [5:43] You know, I don't have one-on-ones with 'em because it's impossible. You can't have, you can't have 60 people [5:47] on your staff if you're, you know, gonna get work done and- [5:51] - So you still have 60 reports. You still have across- [5:53] - More, yeah. [5:54] - More. And most stars at least have a foot in engineering. [5:59] - Almost all of them. [6:01] There's experts in memory, there's experts in CPUs, there's experts in [6:05] optical. All, all— [6:06] - That's incredible. [6:06] - Yeah, GPUs and— Architecture, algorithms, design, um— [6:11] - So, you constantly have an eye on the entire stack, and you're having to, like, [6:15] intense discussions about the designer of the entire stack? [6:18] - And no conversation is ever one person. That's why I don't do [6:22] one-on-ones. We present a problem [6:24] and all of us attack it. You know, because we're doing extreme co-design. [6:30] And literally, the company is doing extreme co-design all the time. [6:33] - So, even if you're talking about a particular component, like cooling, [6:38] networking, everybody's listening in? [6:40] - Yeah, exactly. [6:41] - And they can contribute, "Well, this doesn't work for the power distribution. [6:44] This doesn't-" [6:45] - Exactly. [6:45] - "... This doesn't work for the memory. This doesn't work for this." [6:49] - Exactly. And whoever wants to tune out, tune out. You know what I'm saying? [6:54] And the reason for that is because the people who are on the staff, they know [6:58] when to pay attention. There's supposed... You know, it's something they could have [7:01] contributed to, they didn't contribute to, "I'm going to call them out." You know? [7:04] And so, "Hey, come on, let's get in here." [7:07] - So, as you mentioned, NVIDIA is this company that's adapting to the environment. [7:11] So, at which point can you say, did the environment change and [7:15] began adapting sort of secretly- [7:19] ... in the early days from GPU for gaming, maybe the [7:23] early deep learning revolution to we're now going to start thinking of it as an AI [7:27] factory? What does NVIDIA do? It produces AI, let's build a factory that makes AI. [7:32] - Uh, I could, I c- you, you could- I could reason through what just systematically. [7:35] We started out as a, as an accelerator company. [7:39] But the problem with accelerators is that the application domain's too narrow. [7:43] It has the benefit of being incredibly optimized for the [7:47] job. You know, any specialist has that benefit. The problem with [7:51] intense specialization is that, of course, your market reach is narrower, [7:57] but that's, that's even fine. The problem is, the market size also [8:03] dictates your R&D capacity. And your R&D capacity ultimately dictates [8:10] the influence and impact that you can possibly have in computing. And so, [8:14] when we first started out as an accelerator, very specific accelerator, [8:19] we always knew that had- that was going to be our first step. We had to find a [8:23] way to become accelerated computing. But the problem is, when you become a [8:27] computing company, [8:29] it's too general purpose and it takes away from your specialization. The [8:33] tur- I connected two words that are actually [8:37] have fundamental tension. The better computing company we become, the [8:41] worse we became as a specialist. The more of a specialist, the less [8:45] capacity we have to do overall computing. And so, [8:49] that... And I connected those two words together on purpose, that the company [8:53] has to find that really narrow path, step by step by step, to expand our [9:01] aperture of computing, but not give up on the most important [9:05] specialization that we had. Okay, so the first step that we took beyond [9:09] acceleration was, we invented a programmable pixel shader. [9:13] So, that was the first step towards programmability. You know, it was our [9:17] first journey towards moving into the world of computing. The second thing that we [9:21] did was we created we put [9:25] FP32 into our shaders. That FP32 step, IEEE-compatible FP32, was a huge step [9:33] in the direction of computing. It was the reason why all of the people who were [9:39] working on, on stream processors and, you know, [9:43] other types of data flow processors discovered us. And they said, "Hey, all of [9:47] a sudden, you know, we might be able to use this GPU that's incredibly computationally [9:51] intensive, and it's now, you know, compliant with IEEE." [9:55] I can take my software that I was writing, you know, previously on [9:58] CPUs, and I can, you know, see about, you know, using the GPU for that. [10:04] And which led us to create, put C on top of [10:08] FP32, what's called, we call Cg. The Cg path [10:12] took us to eventually CUDA. CUDA, step by step by step [10:18] We... Well, putting CUDA on GeForce, that was [10:22] a strategic decision that was very, very hard to do, because it cost the [10:25] company enormous amounts of our profits, and we couldn't afford it [10:29] at the time. But we did it anyways because we wanted to be a computing [10:33] company. A computing company has a computing [10:37] architecture. A computing architecture has to be compatible across all of [10:41] the chips that we build. [10:42] - Can you take me through that decision? So, putting CUDA on GeForce, could not afford to [10:46] do? Can you explain that decision? Why, why boldly choose to do that anyway? [10:52] Can you explain that decision? [10:53] - Yeah, excellent. That was, that was the first... I would say that that was [10:58] the first strategic decision that is as close to an existential threat. [11:06] - For people who don't know, it turned out to be, spoiler alert, [11:10] one of the most incredibly brilliant decisions ever made [11:14] by a company. So, CUDA turned out to be [11:18] an incredible foundation for computation in this AI infrastructure world. So- [11:23] - Thank you [11:24] - ... just setting the context. It turned out to be a good decision. [11:27] - Yeah, it turned out to have been a good decision. I think the... So, here's the way it [11:31] went. So, we invented this thing called CUDA, and [11:35] It expanded the aperture of applications [11:38] that, that we can accelerate with our accelerator. The question is, how do we, [11:44] how do we attract developers to CUDA? [11:48] Because a computing platform is all about developers. And developers [11:54] don't come to a computing platform just because, you know, it [11:58] could perform something interesting. They come to a computing platform because the [12:02] install base is large. [12:04] Because a developer, like anybody else, wants to develop software that [12:07] reaches a lot of people. So, the install base is, in fact, [12:11] the single most important part of an architecture. The [12:14] architecture could attract enormous amounts of criticism. For example, [12:20] no architecture has ever attracted more criticism than the [12:23] x86.... you know, as a less than, less than elegant architecture, but yet it is the [12:31] defining architecture of today. It gives you an example that in fact [12:35] so many RISC architectures which were beautifully architected, [12:42] incredibly well-designed by some of the brightest computer scientists in the [12:46] world, largely failed. And so I've given you two [12:49] examples where one is, you know, one is elegant, the other one's [12:53] barely aesthetic, and so yet x86 survived and the reason for- [12:58] - Install base is everything. [12:59] - Install base defines an architecture. Not... Everything else [13:03] is secondary, okay? And so there were other architectures at the time. [13:07] CUDA came out, OpenCL was here. There were... You know, there's several other competing [13:11] architectures. But the thing that... The decision that we made that was good [13:15] was we said, "Hey, look, ultimately it's about, [13:19] Install base and what is the best way we could get a new [13:23] computing architecture into the world?" [13:26] By that timeframe, GeForce had become successful. We were already [13:30] selling millions and millions of GeForce GPUs a year, and we [13:33] said, "You know, we ought to put CUDA on GeForce [13:38] and put it into every single PC whether customers use it or not, [13:43] and use it as a starting point of cultivating our install base." [13:48] Meanwhile, we'll go and attract developers, and [13:52] we went to universities and wrote books and taught [13:56] classes and put CUDA everywhere. And eventually people discover... [14:00] And at the time, the PC was the primary computing vehicle. There was no [14:04] cloud, and we could put a supercomputer in the hands of every [14:07] researcher in school, every scientist, you know, every engineering school, every... [14:11] or every student in school, and eventually something amazing will happen. [14:15] Well, the problem was CUDA increased our cost [14:19] of that GPU, which is a consumer product, so tremendously, [14:23] it completely consumed all of the company's gross profit dollars. [14:29] And so at the time, the company was probably, you know, worth, I don't know, at the time, [14:33] eight... [14:36] Was it like $8 billion or something? Like six, $7 billion or something like that. [14:40] After we launched CUDA, I recognized that it was going to add so much cost, [14:47] but it was something we believed in. [14:50] You know, our market cap went down to like one and a half billion dollars. And so [14:54] we were down, we were down there for a while and we [14:58] clawed our way way back slowly, but we carried [15:01] CUDA on GeForce. I always say that NVIDIA is the house that [15:05] GeForce built, because it was GeForce that took CUDA out to [15:09] everybody. Researchers, scientists, they discovered CUDA on GeForce because they were [15:17] all, you know... Many of 'em were gamers. Many of them built their own [15:21] PCs anyways. In a university lab, many of them built [15:25] clusters themselves, you know, using PC components. And, [15:29] and so that, you know, that's kind of how we got going. [15:31] - And then that became the platform and the foundation for the deep learning revolution. [15:35] - That was also another great, great observation. Yeah. [15:38] - That existential moment, do you remember... Like, what were those meetings [15:42] like? What were those discussions like, deciding as a company, risking everything? [15:48] - Well I had to make it clear to the board what we're trying to do, and [15:56] the management team knew our gross margins were gonna get crushed. [16:00] So you could imagine a world where GeForce would carry [16:04] the burden of CUDA and none of the gamers would [16:08] appreciate it and none of the gamers would pay for it. You know, they only pay [16:12] certain price and it doesn't matter what your cost is. And so the... You know, [16:16] we, we increased our cost by 50% and that con- [16:20] consumed... And we were a 35% gross margin company, and so it, [16:24] it was a... It was quite a difficult decision to make. But [16:28] you could imagine that someday this would go into workstations and it would go into [16:31] supercomputers and, and in those segments, maybe we can capture more margin. [16:37] so you could, you could reason your way into being able to afford this, [16:42] But it still took... It took a decade. [16:45] - But that, but that's more of, like, conversation with the board convincing them, but you [16:48] psychologically- ... as NVIDIA's continued to make bold bets [16:56] that predict the future, and in part, especially now, define the future. [17:01] So I'm almost looking for wisdom about [17:05] how you're able to make those decisions, to make leaps- ... like that as a company. [17:14] - Well, first of all I'm informed by a lot of curiosity. [17:21] At some point, there's a reasoning system that, [17:26] that convinces me, so clearly this outcome will happen. That this will happen. [17:34] And so I believe it in my mind, and when I believe it in my mind, [17:38] you know, you know how it is. You manifest a future and that [17:42] future is so convincing, there's no way it won't happen. [17:48] There's a lot of suffering in between, but you've gotta believe what you believe. [17:52] - So you envision the future- [17:55] ... and you essentially, from a sort of engineering perspective, manifest it? [17:59] - Yeah. And you, you reason about how to get there. You reason about why [18:03] it, it must exist. And, you know, I reason... We all [18:09] reason it here. The management team would reason about it. All the people that I... [18:13] We spend a lot of time reasoning about it. The thing, the thing that... The next [18:17] part of it is probably a skill thing, which is, [18:20] you know, oftentimes in leadership the leadership [18:24] stays quiet or they learn about something, and then they do some manifesto, [18:28] and it's a brand-new year, and somehow at the end of the year, next year, [18:32] we're gonna have a brand-new plan. Big huge layoff this way, big huge [18:36] organization change this way, new mission [18:39] statement... brand new logos you know, that kind of stuff. [18:43] Um, we've just never, I never do things that way. When I learn about something [18:48] and it's starting to influence how I think, I'll make it very clear to [18:52] everybody near me that, you know, this, this is interesting. [18:57] This is going to make a difference. This is going to impact that. [19:01] And I reason about things step by step by step. Oftentimes, I've already made up [19:05] my mind, but I'll take every possible opportunity, [19:09] external information, new insights, new discoveries, [19:13] New engineering, you know, revelations, new [19:16] milestones developed, I'll take those opportunities and I'll use [19:20] it to shape everybody else's [19:24] belief system. And I'm doing that literally every single day. [19:28] I'm doing that with my board, I'm doing that with my management team, I'm doing that with [19:32] my employees. I'm trying to shape [19:35] their belief systems such that when I come the day I say, "Hey, let's buy Mellanox," [19:43] it's completely obvious to everybody that we absolutely should. [19:48] On the day that I said, "Hey guys, let's [19:52] go all in on deep learning," and let me tell you why. I've [19:56] already been laying down the bricks to different organizations [20:00] inside the company. Every organization and everybody, [20:05] many of the people might have heard everything. [20:08] Most of the company hears, of course, pieces of it. [20:11] And on the day that I announce it, [20:14] everybody's kind of bought in to many pieces of it. [20:19] And in a lot of ways, I like to announce these things, and I imagine, [20:25] that the employees are kind of saying, "You know, Jensen, what took you so long?" [20:30] And in fact, I've been shaping their belief system for some time, and therefore [20:34] leadership. Sometimes it looks like you're leading from behind, [20:39] but you've been shaping their, you know, to the point where on the day that I declared it, [20:43] 100% buy-in. But that's what you want. You want to bring everybody [20:47] along. You know, otherwise, we announce something about deep learning and everybody goes, "What [20:51] are you talking about?" You know, you announce something about let's go all in [20:55] on this thing, and your management team, your [20:59] board, your employees, your customers, they're kind of like, "Where's this coming from? [21:02] You know, this is insane." And so, so GTC [21:06] effect, if you go back in time, you look at the keynotes, [21:11] I'm also shaping the belief system of my partners in the [21:15] industry and I'm using that to shape, you know, the belief [21:19] system of my own employees. And so by the time that I announce something, [21:24] like for example, we just announced Grok. [21:27] We've been late... I've been talking about the stepping stones for two and a half [21:32] years. [21:33] You just go back and go, "Oh my gosh, they've been talking about it for two and a [21:37] half years." And so I've been laying the foundation step by step by step, [21:41] so when the time comes you announce it, everybody's saying, "You know, what took you so long?" [21:44] - But it's not just inside the company. You're shaping the landscape, the broader global landscape [21:48] of innovation. Like, putting those ideas out there, you really are manifesting [21:52] reality. [21:53] - We don't build computers. We actually don't build clouds. We don't... As [21:57] it turns out, we're a computing platform company. And so nobody can [22:00] buy anything from us. That's the weird thing. You know, we vertically [22:06] design, vertically integrate to design and optimize, but then [22:10] we open up the entire platform at every single layer [22:14] to be integrated into other companies' products and services and [22:18] clouds and supercomputers and OEM computers [22:21] and so the amazing thing is, I can't do what I [22:25] do without having convinced them first. And so most of [22:29] GTC is about manifesting a future that by the [22:33] time that we... My product is ready, they're going, "What took you so long?" [22:39] - Yeah. So one of the things you've been a believer for a long [22:43] time is scaling laws, broadly defined. So are [22:47] you still a believer in the scaling laws? [22:49] - Yeah, yeah. Yeah, we have more scaling laws now. [22:51] - So I think you've outlined four of them with pre-training, post-training, test [22:55] time, and agentic scaling. What do you think, when you think about the future, [23:01] deep future and the near-term future, what are the blockers [23:06] that you're most concerned about that keep you up at night that you have to overcome [23:10] in order to keep scaling? [23:12] - Well, we can go back and reflect on what people thought were blockers. [23:16] So in the beginning, we were the first... The pre-training scaling [23:20] law. You know, people thought well rightfully [23:24] so, that the amount of data that we have, high-quality data that we [23:28] have will limit the intelligence that we achieve. And that scaling law [23:32] was an important, very important scaling law. The larger the model, the [23:35] correspondingly more data results in a better... With a [23:39] results in a smarter AI. And so that was pre-training. And [23:43] Ilya Sutskever, Ilya said, "We're out of data," or something like that. "Pre-training is [23:47] over," or something like that. The industry panicked, [23:51] you know, that this is the end of AI. And of course, of course that, that's [23:56] obviously not true. We're gonna keep on scaling the amount of data that we [24:00] have to, to train with. A lot of that data is probably gonna be synthetic, [24:04] and that also confused people, you know? And what people don't [24:08] realize is they've kind of forgotten that most of the data [24:12] that, that we are training that we teach each other with, inform [24:15] each other with, is synthetic. You know, I... [24:19] It's synthetic because it didn't come out of nature. [24:23] You created it. I'm consuming it. I modify it, augment it, [24:30] I regenerate it, somebody else consumes it. And so, so [24:33] we've now reached a level where AI is able to take ground truth, augment it...... [24:43] Enhance it, synthetically generate an enormous amount of [24:46] data. And that part of post-training [24:50] continues to scale, and so the amount of data that we could use [24:53] that is human generated will be smaller, and smaller, and smaller. [24:57] The amount of data that we use to train model, [25:02] is going to continue to scale to the point where we're no longer limited... [25:07] Training is no longer limited by... Data is now limited by compute. And the [25:11] reason for that is most of the data is synthetic. Then the next phase [25:15] is test time, and [25:19] I still remember people telling me that, "Inference? Oh, yeah, that's [25:23] easy. Pre-training, that's hard." These are giant systems that people are talking [25:26] about. Inference must be easy. And so inference chips are gonna be little tiny [25:30] chips, and [25:32] ... you know, they're not, they're not like NVIDIA's chips. Oh, those are gonna be [25:35] complicated and expensive, and, you know, we could make... And this is- [25:39] in the future, inference is gonna be the biggest market, and it's gonna be easy, and [25:43] we're gonna commoditize it. You know, everybody can build their own chips. And, [25:47] and that was always illogical to me because inference is thinking, [25:53] and I think thinking is hard. Thinking is way harder than reading. [25:59] You know, pre-training is just [26:01] memorization and generalization, you know, and looking for patterns in relationships. [26:05] You're reading and reading, versus thinking, reasoning, solving problems, taking [26:14] unexplored experiences, new experiences, and [26:18] breaking it down into... Decomposing it into, you [26:22] know, solvable pieces that we then go off, either through first [26:26] principle reasoning, or, you know, through previous examples, [26:30] prior experiences. You know, or just uh, [26:34] exploration and search and, you know, trying different [26:38] things. And that whole process of, of test time scaling, [26:44] Inference, is really about thinking. And it's about [26:47] reasoning, it's about planning, it's about search, it's about... And so how could that [26:51] possibly be compute light? And we were absolutely right about [26:55] that. You know, so test time scaling is intensely compute intensive. [27:01] Then the question is, okay, now we're at inference and we're at test time scaling, what's beyond that? [27:06] Well, obviously, we have now created, you know, one agentic person, [27:12] and that one agentic person has a large language model that we've now [27:16] now, you know, developed. But during test time, that agentic [27:20] system goes off and does research and bangs on [27:24] databases, and it goes out and, you know, uses tools, [27:27] and one of the most important things it does is spins off and spawns off a whole [27:31] bunch of sub-agents. Which means we're now creating large teams. [27:36] It's so much easier to scale NVIDIA by [27:40] hiring more employees than it is to scale myself. [27:44] And so the next scaling law is the agentic scaling law. It's kind of [27:47] like multiplying AI. [27:51] Multiplying AI, we could spin off agents as fast as you want to spin off agents. [27:55] And so, you know, I... You know, I have four scaling laws. [27:59] And as we use the agentic systems, they're gonna create a [28:03] lot more data, they're gonna create a lot of experiences. Some of it we're gonna say, [28:07] "Wow, this is really good. We ought to memorize this." [28:12] That data set then comes all the way back to pre-training. [28:15] We memorize and generalize it. We then refine it and fine-tune it [28:20] back into post-training. Then we enhance it even more with [28:24] test time, you know, and the agents, agentic systems, [28:28] you know, put it out to the industry. And so this loop, this [28:32] cycle, is gonna go on and on and on. It kinda comes down to basically [28:37] intelligence is gonna scale by one thing, and that's compute. [28:41] - But there's a tricky thing there that you have to anticipate and predict, which [28:45] is some of these components, it requires [28:49] different kind of hardware to really do it optimally. So you have [28:53] to anticipate where the AI innovation's going to lead. For example, a mixture of- [28:57] - Perfect [28:58] - ... experts with sparsity. [28:59] - Perfect. [29:00] - With hardware, you can't just pivot on a week's notice. You [29:04] have to anticipate what that's going to look like. It has some- [29:06] - So good [29:07] - ... that's so scary and difficult to do, right? [29:09] - For example, These AI model architectures are being invented about once every six [29:15] months. Right? And system architectures and hardware architectures [29:23] kind of every three years. [29:26] And so you need to anticipate what likely is going to happen, [29:30] you know, two, three years from now. [29:33] And there's a couple ways that you could do that. First of all, we could do research internally ourselves, and that's one [29:37] of the reasons why we have basic research, we have applied research. [29:40] We create our own models. And so we have hands-on life [29:44] experience right here. This is part of the co-design that I'm talking about. [29:48] We're also the only AI company in the world that works with literally every AI company in the world. [29:51] And so to the extent that we can, we try to get a sense [29:55] of what are the challenges that people are experiencing. [29:59] - So you're listening to the whispers across the industry, the AI labs. [30:02] - That's right. You got to listen and learn from everybody. And have a... [30:06] And then the last part is to have an architecture that's [30:10] flexible, that can adapt and move with the wind. And one of the [30:14] benefits of, of CUDA is that it's, you know, on the one hand, [30:18] an incredible accelerator. On the other hand, it's really flexible. [30:22] And so that balance, incredible balance between [30:26] specialization, otherwise we can't accelerate the CPU, [30:31] versus generalization, so that we can adapt with changing algorithms, [30:35] that's really, really important. That's the reason why CUDA has [30:39] been so resilient on the one hand, and yet we [30:43] continue to enhance it. We're at CUDA 13.2, and so we're [30:47] evolving the architecture so fast that we can stay [30:51] with, you know, with the modern algorithms. [30:55] For example... When mixtures of experts came out, [30:59] That's the reason why we had NVLink 72 instead of NVLink eight. [31:04] We could now take an entire four trillion, 10 trillion parameter model [31:08] and put it in one computing domain as if it's running on one GPU. [31:13] Um, people probably didn't notice, I said it, but [31:19] if you look at the architecture of [31:22] the Grace Blackwell racks, it was completely focused on [31:26] doing one thing, processing the LLM. [31:30] All of a sudden, one year later, you're looking at a Vera Rubin rack. [31:34] It has storage accelerators. It has this incredible new CPU called Vera. [31:41] It has Vera Rubin and NVLink 72 to run the [31:45] LLMs. It also has this new additional rack called Rock. [31:49] And so this entire rack system is completely [31:53] different than the previous one, and it's got all these new [31:57] components in it. And the reason for that is because the last one was designed to run [32:02] MoE large language models, inference. And this one is to run agents and agents [32:08] bang on tools, and- [32:10] - Obviously, the design of the system had to have been done before Claude Code, [32:17] Codex, OpenClaw. So you were anticipating the future, essentially. [32:21] And that comes from what? From the whispers, from understanding what all the state- [32:25] - No [32:25] - ... of the art is about? [32:26] - No, it's easier than that. You just reason about it. First of all, you just reason. [32:34] no matter, no matter what happens, at some point in order [32:38] for that large language model to be a digital worker... Let's just, [32:42] let's just use that metaphor. Let's say that we want the LLM to be a [32:46] digital worker. What does that have to do? [32:49] It has to access ground truth. That's our file system. It has [32:53] to be able to do research. It doesn't know everything. We don't have... And I don't wanna [32:57] wait until this AI becomes, you know, universally [33:01] smart about everything, past, present, and future before I [33:05] make it useful. And so therefore, I might as well let it go do research. [33:09] It's obviously, if it wants to help me, it's gotta use my tools. [33:13] You know, a lot of people would say, "You know AI is gonna completely [33:17] destroy software. We don't need software anymore. We don't even need tools anymore." That's [33:20] ridiculous. Let's use the... Let's use a thought experiment. [33:25] And you could just sit there, enjoy a glass of whiskey, and, [33:29] and think about all these things, and it would become completely obvious. Like, if I [33:33] were to create the most amazing- the [33:37] most amazing agent that we can imagine in the next 10 years. [33:41] Let's say it'd be a humanoid robot. If that humanoid robot were to be created, [33:46] is it more likely that the humanoid robot comes into my house [33:50] and uses the tools that I have to do the work that it needs to do? [33:54] Or does this hand turns into a 10-pound hammer in [33:58] one instance, turns into a scalpel in another instance, [34:02] and in order to boil water, it beams, you know, microwaves out [34:06] of its fingers? You know, or is it more likely just to use a microwave, you [34:10] know? And the first time it goes up to the microwave, [34:13] it probably doesn't know how to use it. But that's okay. It's connected to the [34:17] internet. It reads the manual of this [34:21] microwave, reads it, instantly becomes an expert. [34:24] And so it uses it. And so I think the... I just [34:28] described, in fact, almost all of the properties of OpenClaw. [34:35] You know, that it's gonna use tools, that it's gonna access files, it's gonna be able to do [34:38] research. It has I/O subsystem. And when you're done [34:42] reasoning through it, reasoning about it in that way, [34:46] Then you say, "Oh, my gosh, the impact to the future of computing is deeply [34:53] profound." And the reason for that is, I think we've just reinvented the computer. [34:57] And then now you say, "Okay, when did we reason about that? When did we [35:01] reason about OpenClaw?" If you take the OpenClaw schematic that I used at GTC, [35:08] you'll find it two years ago. Literally, two years ago at GTC, I was talking about [35:15] agentic systems that exactly reflect OpenClaw today. [35:21] And, of course, the confluence of [35:25] many things had to happen. First of all, we needed Claude and [35:29] GPT and, you know, all of these models to reach a level of [35:33] capability. So their innovation and their breakthroughs and their continued [35:36] advances was really important. [35:39] And then, of course, somebody had to create an open source, you [35:43] know project that was sufficiently robust, you know, and sufficiently complete [35:51] and that we can all put to work. And I think OpenClaw [35:55] did for, did for agentic systems what ChatGPT did for [35:59] generative systems. And I just think it's a very big deal. [36:02] - Yeah, it's a really special moment. I'm not exactly sure why it [36:05] captured so much of the world's attention, but it did, [36:09] more than Claude Code and Codex and so on. [36:12] - Because consumers could reach it. [36:13] - Sure, yeah. But there's also so much of this is vibes. [36:17] And Peter, I had a podcast with him, he's a [36:21] wonderful human being. So part of it is also the humans that represent the thing. [36:25] - Yeah, no doubt. [36:25] - Part of it is memes and the— 'Cause we're all trying to figure it out. [36:28] There's really serious and complicated security concerns [36:31] about when you have such powerful technology, how do you hand over your data [36:35] so they can do useful stuff? But then there's scary things associated with that. [36:39] And we, as a civilization, as individual people and as a civilization, figuring out how [36:43] to find that right balance. [36:44] - Yeah, we jumped on it right away and we sent a bunch of security experts this way. [36:49] And we did this thing called OpenShell. It's already been [36:53] integrated into, into OpenClaw. [36:55] - And NVIDIA put forward NemoClaw. [36:58] - Yep, exactly. [36:59] - They install super easy. It makes sure that it's secure. [37:03] - We give you two out of three rights. Agentic systems can access [37:07] sensitive information, it can execute code, and it can communicate [37:10] externally. We could keep things safe if we gave you two out of those three [37:17] capabilities at any time, but not all three. [37:21] And out of those two out of three capabilities, we also give you access control [37:25] based on whatever rights that you're given by enterprise. [37:29] And then we connect it to a policy engine that all these enterprises already have. [37:33] And so we're going to try to do our best to help OpenClaw become a better claw. [37:40] - So you eloquently explained how we have a long history of [37:44] blockers that we thought were going to be blockers, and we overcame them. But now looking into the [37:48] future, what do you think might be the blockers now that it's clear that agents [37:52] will be everywhere? [37:53] So obviously we're going to need compute. So what is going to be the [37:57] blocker for that scaling? [37:59] - Power is a concern, but it's not the only concern. But that's the reason why [38:04] we're pushing so hard on extreme co-design, so that we can improve the tokens [38:11] per second per watt orders of magnitude [38:15] every single year. And so in the last 10 years, Moore's [38:19] Law would have progressed computing about 100 times in the last [38:23] 10 years. We progressed and scaled up computing by [38:27] a million times in the last 10 years. And so we're gonna keep on doing that [38:31] through extreme co-design. So energy efficiency, perf per watt, [38:36] completely affects the revenues of a company. It [38:40] affects the revenues of a factory. And we're just going to [38:44] push that to the limit so that we can keep on driving token costs [38:47] down as fast as we can. You know, the [38:51] our computer price is going up, but our token generation [38:56] effectiveness is going up so much faster that token cost is [38:59] coming down. It's just coming down an order of magnitude every year. [39:04] - So power, that's an interesting one. So the way to try [39:07] to get around the power blocker is to try to, with the tokens per [39:11] second per watt, try to make it more and more efficient. [39:14] Of course, there's the question of how do we get more power. [39:16] - We should also get more power. [39:17] - That's a really complicated one. You've talked about small modular nuclear power plants. [39:21] There's all kinds of ideas for energy. How much does it keep you [39:25] up at night? The bottlenecks in the supply chain of AI, [39:29] like ASML with EUV lithography machines, TSMC with advanced [39:33] packaging like CoWoS, and SK Hynix with the high bandwidth memory? [39:38] - All the time, and we're working on it all the time. No company in history [39:44] has ever grown at a scale that we're growing while [39:47] accelerating that growth. It's incredible. [39:50] And it's hard for people to even understand this. [39:53] In the overall world of AI computing, we're increasing [39:57] share. And so supply chain, upstream and downstream, are really important to us. [40:04] I spend a lot of time informing all the [40:08] CEOs that I work with, what are the dynamics that's going to cause, [40:12] The growth to continue or even accelerate? It's part of the reasons [40:16] why to the entire right-hand side of me were CEOs of [40:22] practically the entire IT industry upstream and practically the entire [40:29] infrastructure industry downstream. [40:32] And they were all... There were several hundred CEOs. [40:36] And I don't think there's ever been keynotes where several hundred CEOs show up. [40:40] And part of it is, I'm telling them about our business condition now. [40:46] I'm telling them about the growth drivers in the very near future and what's [40:49] happening. And I'm also describing where are we going to go next [40:53] so that they could use all of this information and all of the dynamics that are [40:57] here to inform how they want to invest. And [41:01] so I inform them that way like I inform my own employees. [41:06] And then of course, then I make trips out to them [41:09] and make sure that, "Hey, listen, I want you to know this quarter, this [41:13] coming year, this next year, these things are going to happen." [41:17] And if you look at the CEOs of the DRAM industry, [41:21] The number one DRAM in the world was DDR memory for CPUs in data centers. [41:30] About three years ago, I was able to convince several of the CEOs that [41:36] even though at the time HBM memory was used quite scarcely, you know, and, [41:40] and barely by supercomputers that this was going to be a [41:44] mainstream memory for data centers in the future. And at first it sounded [41:48] ridiculous, but several of the CEOs believed me and decided [41:52] to invest in building HBM memories. Another memory was rather odd to put into a data [41:59] center, is the low power memories that we use for cell phones. [42:03] And we wanted them to adapt them for supercomputers in the data [42:07] center. And they go, "Cell phone memory for supercomputers?" [42:11] And I explained to them why. Well, look at these two memories, [42:14] LPDDR5, HBM4. The volumes are [42:18] so incredible. All three of them had record years in history, and these [42:22] are, these are 45-year companies. And so, [42:27] you know, I... That's part of my job, is to inform and shape, inspire, you know. [42:36] - So you're not just manifesting the future and maybe [42:39] inspiring NVIDIA, the different engineers of the company, you're, [42:44] you're manifesting the supply chain of the future. So you're having conversations with [42:48] TSMC, with ASML. [42:50] - Upstream, downstream. [42:51] - Upstream, downstream. So that's the thing. [42:53] - GEV, Caterpillar. Yeah, that's downstream from us. Yeah, yeah, there you go. [42:59] - Yeah, the whole thing. I mean, but that's so... [43:02] There's so much incredibly difficult engineering that [43:05] happens in the entire semiconductor industry, and it just feels scary [43:12] how intricate the supply chain is, how many components there [43:16] are, but it works somehow. [43:18] - Exactly, the deep science. The deep engineering, the incredible [43:22] manufacturing, and so much of the manufacturing is already robotics, [43:26] but we have a couple of hundred suppliers that contribute [43:30] the technology that goes into our 1.3 million component rack. [43:36] Each rack is 1.3, one and a half million components. There are 200 suppliers [43:43] across the Vera Rubin rack. [43:45] - So it's interesting that you don't list that as the thing that keeps you up at night in the list of [43:48] blockers. [43:49] - But I'm doing, I'm doing all the things necessary to- [43:52] - Okay [43:52] - ... yeah, see? [43:54] I can go to sleep because I checked it off. I said, okay, you know, I go, I [43:58] yeah, I can go to sleep and I go, well, let's see, what re- let's reason [44:02] about this. What's important for us? [44:03] Um, because okay, let's reason about this. Because we [44:08] changed the system architecture from the original DGX-I that [44:12] you remembered to NVLink-72 rack scale computing- [44:16] ... what's gonna... What does that, what does that mean? [44:19] What does that mean to software? What does that mean to engineering? What does that [44:23] mean to how we design and test? How, and what does that mean to the supply [44:27] chain? Well, one of the things that it meant was we moved supercomputer, [44:33] supercomputer integration at the data center [44:37] into supercomputer manufacturing in the supply chain. If you're doing that, [44:44] you also have to recognize you're gonna move one... And if, [44:48] if you're, if you're, you know, total footprint of [44:52] whatever data center you're gonna build, let's say you would like to have, you know, [44:58] 50 gigawatts of supercomputers that are running simultaneously, [45:02] and it takes one week to manufacture that 50 [45:06] gigawatts of supercomputers, then each week in the [45:10] supply chain, the supercomputers are gonna need a gigawatt of power. And so, [45:14] so we're gonna need the supply chain to increase the amount of power it has [45:18] to build, test, to build and test the [45:21] supercomputers in the supply chain before I ship it. [45:25] Well, NVLink-72 literally builds supercomputers in the supply chain and [45:28] ships 'em two, three tons at a time per rack. [45:32] It used to be—they used to come in parts and we used to assemble 'em [45:36] inside the data center. But that's impossible now because NVLink-72 is so dense. [45:41] And so that's an example. And I would have to go and to, you know, [45:45] I've... Fly into the supply chain, go meet my partners saying, "Hey," I said, "guess [45:48] what? So here's what I'm going to do with... This is the way we used to build [45:52] our DGXs. We're gonna build them this way. This is gonna be so much better because [45:56] we're going to need 'em for inference." The market for inference is, you know, [46:02] coming. The inflection point for inference is coming. It's gonna be a big market. [46:05] And so I first explain to them what's going on, why it's gonna happen, and [46:09] then I ask 'em to make several billion dollars of capital investments [46:15] each. And because they, you know, they trust me and I'm very [46:21] respectful of 'em, and I give 'em every opportunity to question me and [46:25] I spend time to explain things to people and I reason about it. I draw on pictures [46:29] and I reason about it in first principles. And by the time [46:33] I'm done with them, they know what to do. [46:35] - So it's a lot of it is about relationships and building a shared view of the future. [46:41] Uh, but do you worry about certain bottlenecks? I mean, what are the [46:45] biggest bottlenecks in the supply chain? Are you, are you worried about ASML's EUV [46:49] tooling? Are you, are you worried about the packaging, CoWoS packaging of [46:53] TSMC, about how fast it could scale? Like you said, you're [46:57] not only growing incredibly fast, you're accelerating your growth. So [47:01] it feels like everybody in the supply chain, and those are [47:05] certainly bottlenecks, would have to scale up. Are you having conversations with them, [47:09] like, how can you scale up faster? Do you worry about it? [47:13] - No. [47:13] - Okay. [47:14] - Because, because I told 'em what I needed. They understood what I [47:18] need. They told me what they're gonna go do, and I believe them what they're going to do. [47:22] - Interesting. [47:22] That's great to hear. So maybe if we can just linger on the power for a little [47:26] bit. What are your hopes for how to solve the energy problem? [47:30] - One of the areas, Lex, that I'm [47:34] that I would love, I would love us to talk about and just get the message out, [47:38] you know our power grid is designed [47:44] for the worst case condition with some margin. Well, [47:50] 99% of the time we're nowhere near the worst case condition because the worst case condition [47:54] is a few days in the winter, a few days in the summer, [47:57] and extreme weather. Most of the time we're nowhere [48:01] near the worst case condition and we're probably running around, call it [48:05] 60% of peak. And so 99% of the time, our power grid [48:13] has excess power, and they're just sitting idle, but they have to be there [48:17] sitting idle because just in case, when the time comes, hospitals have to [48:21] be powered and, you know, infrastructure has to be powered and airports have to run and so on and so [48:25] forth. And so the question that I have is whether we could go and, [48:31] Help them understand and create contractual agreements [48:35] and design computer architecture systems, data centers, such that when they need, [48:42] The maximum power for infrastructure in society, [48:46] that the data centers would get less. [48:49] But that's in a very rare instance anyways. And during that time, we either have a [48:53] backup generator for that little part of it, or we just have our computers shift the [48:57] workload somewhere else, or we have the computers just run [49:00] slower. You know, we could degrade our performance, [49:03] reduce our power consumption and provide [49:07] for a, you know, slightly longer latency response, you know, when [49:10] somebody asks for, you know, asks for an answer. And so I think that [49:14] that way of using computers, of building data centers, [49:18] instead of expecting 100% uptime-... [49:22] and these contracts that are really, really quite rigorous, it's [49:26] putting a lot of pressure on the grid to be able to... Now, they're gonna [49:30] have to increase from their maximum. I just wanna use their excess. [49:35] It's just sitting there. [49:36] - Yeah, that's not talked about enough. So what's stopping there? [49:40] Is it regulation? Is it bureaucracy? [49:43] - I think it's a three-way problem. It starts with the end [49:47] customer. The end customer puts requirements [49:51] on the data centers that they can never [49:55] not be available, okay? So that the end customer expects perfection. [50:00] Now, in order to deliver that perfection, you need a combination of [50:04] backup generators and your grid power [50:07] supplier to deliver on perfection. And so everybody's gotta have six nines. [50:13] Well, I think first of all, [50:15] right now, we ought to have everybody understand that when the customer asks for these [50:19] things, you have somebody in your data center operations [50:23] team disconnected from the CEO. I bet the CEO doesn't know this. I'm [50:27] gonna talk to all the CEOs. The CEOs are probably [50:30] not paying any attention to the contracts that are being signed, and [50:34] so everybody wants to sign the best contract, of course. And they go down to [50:39] cloud service providers, and the contract, the... The two contract [50:42] negotiators that are... You, I could just see them now. [50:46] You know, negotiating these multi-year contracts. Both sides want, [50:50] you know, the best contract. As a result, [50:55] the CSPs then have to go down to the utilities, and they expect the nine, [50:59] the six nines. And so I think the first thing is just make [51:02] sure that all of the customers, the CEOs and the customers [51:06] realize what they're asking for. [51:09] Now, the second thing is we have to build data centers that gracefully degrade. [51:13] And so if the power, if the utility, if the grid tells us, "Listen, we're gonna have [51:17] to back you down to about 80%," we're gonna say, "That's no problem at all." [51:21] We're just gonna move our workload around. We're gonna make sure that data's never [51:25] lost, but we can reduce the computing rate and use less [51:30] energy. The quality of service degrades a little bit. For the critical [51:34] workloads, I shift that somewhere else right away [51:38] so I don't have that problem, and so, you know, whoever, whichever data [51:42] center still has 100% uptime, and so... [51:44] - How difficult of an engineering problem is that, that smart, dynamic allocation of power in a data [51:48] center? [51:49] - As soon as you could specify, you could engineer it. U- beautifully put. [51:54] So long as it obeys the laws of physics on first principles, I think we're good. [51:58] - What was the third thing you were mentioning? Um... [52:00] - So the second thing is the data centers. And the third thing is we [52:04] need the utilities to also recognize that this is an opportunity- [52:10] ... and instead of saying, "Look it's gonna take me five years [52:14] to increase my grid capability," uh, if [52:18] you, if you have, if you're willing to take power of this level of guarantee, [52:25] I can make them available for you next month and at this price. [52:29] And so if utilities also offered more segments of power delivery promises, [52:36] then I think everybody will figure out what to do with it. [52:39] Yeah, but there's just way too much waste in the grid right now. We should go after it. [52:44] - Uh, you've highly lauded Elon and xAI's [52:47] accomplishment in Memphis, in building Colossus supercomputer, [52:53] probably in record time in just four months. It's now at 200,000 [52:57] GPUs and growing very quickly. Is there something that you could speak [53:01] to the understand about his approach that's instructive to, broadly [53:05] to all the data center creators that's [53:08] that enabled that kind of accomplishment? His approach to engineering, his approach to the [53:12] whole management of construction, everything? [53:15] - First of all, Elon is deep in so many different topics. Um, [53:20] Yet he's also a really good systems thinker. [53:24] And so he's able to think through multiple disciplines, and, [53:29] and he obviously pushes things, [53:34] questions everything, where they're, number one, is it necessary? [53:39] Number two, does it have to be done this way? And then numbers, you know, does it have, [53:43] does it have to take this long? And, and so, so he, he has, he has the a- he [53:49] has the ability to question everything, [53:53] To the point where everything is down to its minimal amount [53:57] that's necessary, you can't take anything else out. And yet the, [54:04] the necessary capabilities of the product remains, you [54:08] know? And so he's, he is as minimalist as you could possibly imagine, [54:12] and he does it at a system scale. I think... I also love the fact that he is [54:19] he is represented. He is present at the point of action. [54:25] You know, he'll just go there. If there's a problem, he'll just go there [54:29] and then, "Show me the problem." You know, when you do all of this in combination, [54:34] you overcome a lot of previous, "This is just the way we do it." [54:38] "Um, you know, I'm waiting for them. [54:42] Uh," You know, I mean, it's just, everybody has a lot of excuses. And [54:46] so, and then the last thing is when you act personally with [54:50] so much urgency it causes everybody else to act with urgency, [54:54] you know? And every supplier has a lot of customers going on. Every supplier [54:58] has a lot of projects going on, [55:00] and he made it his... He makes it his business that he's [55:04] the top priority of everybody else's, you know, projects. And so he does that by [55:08] demonstrating it. [55:09] - Yeah, I've been in a bunch of those meetings. It's just, it's fun to watch, 'cause really, not enough [55:13] people ask the question like, "Okay, so, can this be done a lot faster, and how? [55:20] Why does it have to take this long?" [55:21] - Yeah, right. [55:22] - And then in the... That becomes an engineering question often. And yes, I think when [55:26] you get the ground truth of actually... I remember, [55:29] one of the times I was hanging out with him, he literally is going through the [55:33] entire process of how to plug in cables into a rack. He's, [55:37] he's working with an engineer on the ground that's doing that task, and he's [55:41] just trying to understand what does that process look like so it can be less [55:44] error-prone. And just building up that intuition from every single task involved in, [55:50] putting together a data center— ...you start to immediately get a sense at the [55:56] detailed scale and at the broad systems scale of [56:00] where the inefficiencies are, and so you can make it more and more and more efficient. [56:04] Plus you have the big hammer of being able to say, "Let's do it totally different-" [56:08] - Yeah. That's right. [56:09] - "... and remove all possible blockers." [56:10] - That's right. [56:11] - Is there parallels in the NVIDIA Extreme Systems co-design approach that you [56:15] see in the way Elon approaches systems engineering? [56:18] - Well, first of all, co-design is an ultimate systems engineering problem. [56:22] And so we approach, we approach the work that we do from that first, from that [56:26] The other thing that we do and this is, this is a philosophy, a thought, [56:35] a state of mind, I guess, a method, [56:39] That I started 30 years ago, and it's called the [56:43] speed of light. The speed of light is not just about the speed. The speed of light is, is [56:46] my shorthand for what's what's the [56:51] limit of what physics can do. And so every single, everything, [56:55] everything that we do is compared against the speed of light. Memory speed, [56:59] Math speed, power, cost, time, effort, number of people, [57:06] manufacturing cycle time. And when you think about latency versus throughput, [57:12] When you think about cost versus throughput, cost [57:16] versus capacity, all of these things, You test against the speed of light to [57:23] achieve all of these different constraints separately. And then when you consider it [57:30] together, you know you have to make compromises because a [57:34] system that achieves extremely low latency versus a cheap, a system that [57:37] achieves very high throughput are architected fundamentally differently. [57:42] But you want to know what's the speed of light of a system that [57:45] achieves high throughput, what's the speed of light of a system that [57:50] achieves low latency? [57:52] And then when you think about the total system, you can make trade-offs. And so [57:56] I force everybody to think about what's the, what the first- the first principles, [58:00] the limits- ... the physical limits, For everything before we, you know, [58:06] before we do anything. And we test everything [58:10] against that. And so that's a good frame of mind. [58:13] I don't love the other methods, which is continuous improvement. [58:19] The problem with continuous improvement, it... [58:23] First of all, you should engineer something from first principles at the speed, [58:27] you know, with speed of light thinking. Limit it only by physical limits, [58:32] and physics limits. And after that, of course you would improve it over time. Um, [58:40] but I don't like going into a problem and somebody says, "Hey, you know, it takes 74 days [58:44] to do this today-" "... Right now. And we can do it for you in 72 days." [58:50] You know, I'd rather strip it all back to zero- [58:52] ... and say, "First of all, explain to me why 74 days in the first place. [58:56] And let's note, let's think about what's possible today. And if I were [59:00] to build it completely from scratch, you know, how long would it take?" [59:04] Oftentimes, you'd be surprised. It might come to six days. [59:08] Now, the rest of the six days, the 74, could be [59:12] very well-reasoned and compromises, and, you know, cost [59:16] reductions, and all kinds of different things. But at least you know what they are. [59:20] And then now that you know that six days is possible, [59:24] then the conversation from 74 to six, surprisingly much more effective. [59:30] - In such incredibly complex systems that you're working with, is simplicity sometimes [59:34] a good heuristic to reach for? I mean, if I can just... [59:40] I mean, the pod, the Vera Rubin pod that you announced is just incredible. [59:44] We're talking about seven chips, seven chip types, five purpose-built [59:48] rack types, 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, [59:56] over 1,100 Rubin GPUs, 60 exaflops, 10 petabytes per second of scale bandwidth. [60:01] That's all just one... [60:03] - That's just one pod. [60:04] - That's just one pod. [60:06] - Yeah, that's just one pod. [60:07] - I mean, in- ... so you have the... And then even the NVL 72 rack [60:11] alone is 1.3 million components, [60:15] 1300 chips, 4,000 pods crammed into a single 19-inch wide rack. [60:19] - And Lex, we're probably gonna have to crank out about 200 of these pods a week, [60:24] just to put it in perspective. [60:25] - The amount of different components, I suppose simplicity is impossible, [60:30] but is that a metric that you kind of reach for in trying to design things? [60:35] - You know, the phrase, the phrase that I use most often is, we- [60:39] we need things to be as complex as necessary, but as simple as possible. [60:43] And so the question is, is all that complexity there [60:47] necessary? And we ought to test for that. [60:50] And we got to challenge that. And then after that, everything else above it, [60:54] you know, is gratuitous. [60:56] - But it's still almost incredible. Semiconductor industry broadly, but what NVIDIA is [61:00] doing, [61:03] some of the greatest engineering in history. So these systems are just truly, [61:08] truly marvels of engineering. [61:10] - It is the most complex computer the world has ever made. [61:13] - Yeah, the engineering teams, I mean- [61:14] ... I don't, it's not a competition, but I don't know. If it was like an Olympics of [61:18] engineering teams, I mean, TSMC does incredible engineering. Like I said, [61:22] ASML at every scale, but NVIDIA is gonna give them a run for their money. [61:27] Just incredible, incredible teams. [61:28] - Well, it's gold medalists in every single, in every single sport, all [61:32] assembled right here. [61:33] - And have to work together. [61:35] And report directly to you. This is wonderful. You recently traveled to China. [61:41] so it's interesting to ask you China's been incredibly [61:45] successful in building up its technology sector. What do you [61:49] understand about how China's able to, over the [61:53] past 10 years, build so many incredible world-class companies, [61:57] world-class engineering teams, and just this technology ecosystem- [62:01] ... that produces so many incredible products? [62:05] - A whole bunch of reasons for, well, first of all, let's start with some facts. [62:09] 50% of the world's AI researchers are Chinese, plus or minus, and [62:16] they're mostly in China [62:18] still. We have many of them here, but there's amazing researchers still in China. [62:23] They, their tech industry showed up at precisely the right time. [62:30] At the time of the mobile cloud era their way of [62:33] contributing was software, and so this is a country's incredible science [62:37] and math. Really well-educated kids. [62:42] Their tech industry was created during the era of software. [62:48] They're very comfortable with modern software. [62:52] China is not one giant economic country. [62:56] It's got many provinces and cities with mayors all [63:00] competing with each other. That's the reason why there's so many EV [63:03] companies. That's the reason why there's so many AI companies. That's the reason why there's so [63:07] many, every company you could imagine, they all create some of them. And as [63:15] a result, they have insane competition internally. [63:19] And, you know, what remains is an incredible company. [63:24] They also have a social culture where it's family first, friends second, and [63:32] company third. And so, the amount of conversation that goes back and [63:42] forth between... They're essentially open source all the time. [63:47] So the fact that they contribute more to open source is so sensible [63:51] because they're probably, "What are we protecting?" You know, my engineers, [63:55] their brothers are in that company, their friends are in that company, and they're all [63:59] schoolmates. You know, the schoolmate concept. There's a, you know, one [64:03] schoolmate, you're brother for life. And so they, [64:08] they share knowledge very, very quickly. [64:12] And so there's no sense keeping technology hidden. You might as well put it on [64:16] open source. And so the open source community then amplifies, [64:20] accelerates the innovation process. So you get this rapid, incredible great talent, [64:26] rapid innovation because of open source and just, you know, the nature of friends, [64:32] and insane competition. Among comp- [64:36] among the company, what emerges is incredible stuff. And so [64:41] this is the fastest innovating [64:44] country in the world today, and this is something that has everything that, everything that I've just [64:48] said is fundamental to just how the kids were grown, [64:52] the fact that they have excellent education, the fact that they, parents [64:56] want them to do well in school, the fact that they, their culture is that [64:59] way. These are, you know, these are just the thing about their country, [65:04] and they showed up at precisely the time when technology is going through that [65:07] exponential. [65:09] - Plus culturally, it's pretty cool to be an engineer. [65:12] It connects to all the components that you're mentioning... [65:16] - It's a builder nation. [65:18] - It's a builder nation. [65:19] - Yeah, it's a builder nation. Our country's leaders, [65:22] incredible, but they're mostly lawyers. Their country's [65:25] leaders, and because we're, they're trying to keep us safe, rule of law, governing, [65:31] their country was built out of poverty. And so [65:37] most of their leaders are incredible engineers. Some of the brightest minds. [65:43] - To take a small tangent, because you mentioned open source, I have to [65:47] go to Perplexity here, who you have been a fan of a long time. [65:51] - Love it, yeah. [65:52] - And thank you for releasing open source Nemotron 3 Super, [65:56] which you can also use inside Perplexity to look stuff up. [65:59] Now, which is 120 billion parameter open weight MoE model. [66:05] What's your vision with open source? So you mentioned China with [66:12] DeepSeek and MiniMax, with all these companies really [66:16] pushing forward the open source AI movement, and NVIDIA is really leading the [66:21] way in close to state-of-the-art open source LLMs. What's your vision there? [66:28] - First, if we're gonna be a great AI computing company, we have to understand how AI [66:35] models are evolving. One of the things that I love about Nemotron 3 is it's not a, [66:40] just a pure transformer model, it's transformer and SSMs. And we were early in, [66:48] Developing the conditional GANs, which, [66:52] that progressive GANs, which led step-by-step to diffusion. And so, [66:56] The fact that we're doing basic research [66:59] in model architecture and in different domains [67:03] gives us visibility into, you know, what kind of computing [67:07] systems would do a good job for future models. And so it is part [67:11] of our extreme co-design strategy. Second, um, I think we rightfully recognize [67:19] that on the one hand, we want world-class [67:23] models as products, and they should be proprietary. [67:27] On the other hand, we also want AI to diffuse into every industry and every country, [67:34] every researcher, every student. [67:37] And if everything is proprietary, it's hard to do research [67:41] and it's hard to innovate on top of, around, with. [67:46] And so...Open source is fundamentally necessary for many industries to join the AI [67:53] revolution. NVIDIA has the scale and we have the motives [67:58] to not only skills, scale, and motivation to build and continue to [68:06] build these AI models for as long as we shall live. [68:10] And so therefore, we ought to do that. We can open up, we can [68:13] activate every industry, every researcher, [68:17] you know, every country to be able to join the AI revolution. [68:21] There's the third reason, which is from that, to [68:25] recognizing that AI is not just language. These AIs will likely use tools and [68:33] models and sub-agents that were [68:36] trained on other modalities of information. Maybe it's [68:40] biology or chemistry or you know, [68:44] laws of physics, or you know, fluids and thermodynamics, and not [68:48] all of it is in language structure. [68:50] And so somebody has to go make sure that weather prediction, [68:55] biology, AI, AI for biology, physical AI, all of that stuff stays, can be pushed [69:03] to the limits and pushed to the frontier. We don't build cars, but we wanna make sure [69:07] every car company has access to great models. We don't, [69:11] discover drugs, but I wanna make sure that Lilly has the world's best [69:15] biology AI systems, so that they can go use it for discovering drugs. [69:19] And so these three fundamental reasons, both in, [69:23] recognizing that AI is not just language, that AI is really broad, [69:27] that we wanna engage everybody into the world of AI, and then also co-design of AI. [69:32] - Well, I have to say, once again, thank you for open sourcing, [69:36] really truly open sourcing Nemotron 3 and ... [69:39] - Yeah, I appreciate you were saying that. We open sourced the models, we open sourced the weights, we open [69:43] sourced the data, we open sourced how we created it. Yeah, it's pretty amazing. [69:48] - Uh, it's really, it's really incredible. [69:51] You're originally from Taiwan and have a close relationship with TSMC. [69:56] So I have to ask, TSMC I think [70:00] also is a legendary company in terms of the engineering teams, in terms of the [70:03] incredible engineering work that they do. What [70:07] what do you understand about TSMC culture and their approach that [70:10] explains how they're able to achieve this singular [70:14] unmatched success in everything they're doing with semiconductors? [70:19] - You know, first of all, the deepest misunderstanding about TSMC is that [70:29] their technology is all they have. That somehow they have a really great [70:36] transistor, and if somebody shows up another transistor, game over. [70:41] It's the technology [70:43] and of course, you know, I don't mean just the transistor, the metallization [70:47] systems, the packaging, the 3D packaging, the silicon photonics, [70:51] the, you know, all of the technology that they have. That technology is really what [70:55] makes the company special. Their technology makes the company special. [70:59] But their ability to orchestrate [71:04] the demands, the dynamic demands of hundreds of companies in the world as [71:11] they're moving up, shifting out, [71:14] you know, increasing, decreasing, push, pushing out, pulling [71:17] in changing from customer to customer, Wafer starting, wafer stopping, [71:26] Emergency wafer starts, you know, all of this [71:30] dynamics of the world's complexity as the world is [71:34] shape-shifting all the time, and somehow they're [71:38] running a factory with high throughput, high yields, [71:43] really great costs, excellent customer service. [71:46] They take their work, they take their promises seriously. They, when your [71:50] wafer, because they know that they're helping you run your company, when the wafers, [71:54] when the wafers were promised to show up, the wafers show up, you know, so that you could [71:58] run your company appropriately. And so their system, their [72:02] manufacturing system is completely miraculous, I would say. Then [72:06] the second thing is their culture. This culture is simultaneously, [72:10] Technology focused on one hand, advancing technology, [72:14] simultaneously customer service oriented on the other hand. A lot of [72:19] companies are very customer service oriented, but they're not very [72:23] technology excellent. They're not at the bleeding edge of technology. [72:27] There are a lot of companies who are tech, at the bleeding edge of technology, but they're not the best customer [72:31] service oriented company. And so it just depends on somehow they've, [72:34] they've balanced these two and they're world-class at both. [72:39] And then probably the third thing is the technology that I most [72:43] value in them that they created this, you know, this, [72:48] Intangible called trust. I trust them to put my company on top of them. That's a [72:54] very big deal. [72:55] - When they trust, I mean, there's a really close relationship there that you've established, and that trust [72:59] is established based on many years of performance, but there's human [73:03] relationships involved there as well. [73:05] - Three decades, I don't know how many tens, hundreds of billions of dollars [73:09] of business we've done through them, and we don't have a contract. [73:14] That's pretty great. [73:15] - Amazing. Okay, there's this story [73:18] ... ... That in 2013, the founders of TSMC, Morris Chang [73:22] offered you the chance to become TSMC's chief executive, [73:26] And you said you already had a job. Is this story true? [73:30] - Story is true. I didn't dismiss it. Um but I was deeply honored and, and of course, [73:38] of course uh, I knew then as I know now, TSMC is one of the [73:42] most consequential companies in history. [73:45] And, and Morris is one of the highest regarded executive and, [73:50] and business and personal friend that I've had in my life. [73:55] And, um ...Uh, for him to ask is, uh um, I was humbled and, and really honored. [74:05] But, but the work that I'm doing here is really important, and I've seen, [74:09] you know, in my mind's eye, what NVIDIA [74:13] was going to be and what the impact that we could have. [74:17] And uh, it was really important work. [74:20] And it's my responsibility, you know, my sole responsibility to make this [74:24] happen. And so I uh, I declined it, You know, [74:31] not, not because it wasn't an incredible offer. [74:34] It's an unbelievable offer but, but I simply couldn't take it. [74:38] - I think NVIDIA, both NVIDIA and TSMC are two [74:42] of the greatest companies in the history of human civilization. And running either [74:46] one, I'm sure, is incredibly complicated effort and takes... [74:50] You have to truly be all in. [74:52] Uh, everybody at every scale, not just at the CEO level. Everybody is really [74:56] truly all in- [74:57] - Yeah. Yeah, no doubt [74:59] - ... To, to accomplish this kind of complexity. [75:00] - So now I can help both companies. [75:02] - Exactly. So NVIDIA is now the most valuable company in the world. [75:08] I have to ask, what is the NVIDIA's biggest moat, [75:13] as the folks in the tech sector say? [75:15] The edge you have that protects you from the competition. [75:20] - Our single most important property as a company [75:27] is the install base of our computing platform. [75:31] Our single most important thing today is our, is the install [75:36] base of CUDA. Now, the reason why, [75:39] uh, 20 years ago, of course, there was no install base. [75:46] But what makes... And if somebody came up with a [75:50] GUDA or TUDA it wouldn't make any difference at [75:53] all. And the reason for that is because it's [75:57] never been just about the technology. The technology, of course, was [76:01] incredible visionary. But it's the fact that the company was dedicated to it, [76:07] stuck with it, expanded its reach. Um, [76:11] it wasn't three people that made CUDA successful. It was [76:14] 43,000 people that made CUDA successful. And the several [76:18] million developers that believed in us, [76:21] That trusted that we were going to continue to make CUDA 1, 2, 3, 13, [76:27] that they decided to port and dedicate their software on top of it, their [76:31] mountain of software on top of it. And so the install base [76:34] is the number one most important advantage. [76:38] That install base, when you amplify it with the velocity of our [76:42] execution at the scale that we're talking about, no [76:45] company in history had ever built systems of this [76:49] complexity, period. And then to build it once a year is impossible. [76:56] And that velocity combined with the install base, [77:02] in the developer's mind, you just go now, take the developer's mind. [77:06] From the developer's perspective, if I support CUDA, [77:11] tomorrow it'll be 10 times better. I just have to wait six months on average. [77:16] Not only that, if I develop it on CUDA, I reach a few hundred million people, [77:22] computers. I'm in every cloud, I'm in every computer company, I'm in every single [77:27] industry, I'm in every single country. [77:31] So if I create an open source package and I put it on CUDA first, [77:36] I get these both attributes simultaneously. And not only that, [77:43] I trust 100% that NVIDIA is going to keep CUDA around and maintain it and [77:50] improve it and keep optimizing the libraries for as long as they shall live. [77:57] You could take that to the bank, and that last part, trust. [78:00] You put all that stuff together, if I were a developer today, [78:04] I would target CUDA first. I would target CUDA most. [78:09] And that's the reason that I think in the final analysis is [78:13] our first, that's even our first- [78:16] - core advantage. Our second one is our ecosystem. [78:21] The fact that we vertically integrated this incredibly complex system, [78:25] but we integrate it horizontally into every single company's computers. [78:30] - We're into Google Cloud, we're into Amazon, we're in Azure. You know, we're ramping up [78:34] AWS like crazy right now. We're in new companies like [78:38] CoreWeave and Nscale. We're in supercomputers at Lilly. [78:43] We're in enterprise computers. We're at the edge in radio base [78:46] stations. You know, I mean, it's just crazy. One architecture is in [78:50] all these different systems. We're in cars, we're in robots, we're in satellites, we're out in [78:54] space. And so the fact that you have this one architecture and the [78:58] ecosystem is so broad, it basically covers every single industry in the world. [79:03] - Well, how does the, how does the CUDA install [79:06] base evolve into the future with AI factories as a [79:10] moat? What do you... Do you think it's possible that NVIDIA of the future is all about [79:14] the AI factory? [79:16] - Well, the unit of computing used to be GPU to us. [79:20] Then it became a computer, then it became a cluster. Now [79:24] it's an entire AI factory. When I see a computer, when I see what NVIDIA [79:28] builds, in the old days, I would, you know, I visualize the chip. [79:32] And then when I announced the new product, new generation, like, [79:36] "Ladies and gentlemen, we're announcing Ampere today," I'd pick up the chip. [79:40] That was my mental model- [79:41] ... of what I was building. Today, I wouldn't... Picking up [79:45] the chip is kind of still adorable. But it's adorable. It's not my mental [79:51] model of what I'm doing. My mental model is this giant gigawatt [79:56] thing that has power generations connected to the grid. It's got cooling [80:01] systems and networking of incredible monstrosity, you [80:05] know. 10,000 people are in there trying to install it, [80:09] hundreds of networking engineers in there, thousands of engineers behind it trying [80:12] to power it up. [80:14] You know, powering up one of those factories, as you know, it's not somebody going, [80:18] "It's on now." It takes thousands of people to bring it up. [80:22] - So mentally, you're actually... When you're thinking about a single unit of compute, you're like [80:27] literally, when you go to bed at night, you're thinking now about collection of racks, [80:31] so pods, not individual chips. [80:33] - Entire infrastructure. And I'm hoping my next click is when I'm thinking about building [80:37] computers, it's, you know, planetary scale. That'll be the next click. [80:42] - Well, what do you think about the space angle that Elon has [80:45] talked about, doing compute in space for solving some [80:49] of the... It makes some of the energy issues in terms of scaling energy easier. [80:56] - Cooling issues is not easy. Yeah. [80:58] - Cooling. Well, there's a large number of engineering complexities involved with [81:02] that. So, so what... You know, NVIDIA has also announced that you're [81:07] already thinking about that. [81:09] - Yeah, we're already there. NVIDIA GPUs are the first GPUs in space. [81:14] And I didn't realize it, it was so interesting [81:18] to... I would have declared it maybe. [81:20] We're in space. You know, little, little astronaut suit on one of our GPUs. [81:27] But we've been in space. It's the right place to do a lot of imaging. [81:32] You know, because those satellites have, have really high resolution imaging systems, [81:36] and they're sweeping the Earth, you know, continuously now. And, [81:40] um, you want, you know, centimeter scale, imaging [81:43] imaging that is done continuously for the world, so that, [81:47] you know, you'll basically have real time telemetry of [81:50] everything. You don't wanna beam that back down to earth. It's just, you know, [81:57] petabytes and petabytes of data. You gotta just do AI right there at the edge, [82:01] throw away everything you don't need, you've seen before, didn't change, and then just [82:05] keep the stuff that, that you need. And so AI had to be done at the edge. [82:09] Obviously we have, we have a 24/7 solar, if we put it at the polars. And uh, [82:19] but, you know, there's no conduction, no convection. [82:23] And so, you know, you're pretty much just radiation. [82:26] And uh, but, you know, space is big. I guess, you know, we're just gonna put [82:30] big, giant radiators out there. [82:32] - How crazy of an idea do you think it is? Like is this, is this five years out, 10 [82:36] years out, 20 years out? So we're talking about blockers for AI scaling. [82:41] - You know, I'm just so much more practical. I look for where, where, um [82:46] my next, next bucket of opportunities are first. [82:51] Meanwhile, I'm cultivating space. And so I send [82:55] engineers to go work on the problem. We're starting to... We're [82:58] learning a lot about it. How do we deal with radiation? How do we deal with [83:03] degrading performance? How do we deal with a continuous, [83:07] Testing and attestation of, of de- defects? [83:10] And you know, how do we deal with redundancy? And how do we [83:14] degrade gracefully and things like that? And so we could, we could do a... [83:18] What about software? How do you think about software and redundancy and [83:23] performance out in space? Make it so that the computer [83:27] never breaks, it just gets slower, you know. [83:31] And I... So we could start doing a lot of engineer [83:34] exploration upfront. But in the meantime, my, my favorite answer [83:38] is ge- eliminate waste. [83:41] You know, we've got all that idle power, I want to evacuate it as fast as possible. [83:47] - Yeah. There's a lot of low-hanging fruit here on Earth- [83:50] ... That we can utilize for the AI scaling. Quick pause. [83:56] Quick 30-second thank you to our sponsors. Check them out in the description. [84:00] It really is the best way to support this podcast. Go to lexfridman.com/sponsors. [84:06] We got Perplexity for curiosity-driven knowledge exploration, [84:10] Shopify for selling stuff online, LMNT for electrolytes, [84:15] Fin for customer service AI agents, and Quo [84:19] for a phone system, like calls, texts, contacts, for your business. [84:23] Choose wisely, my friends. And now, back to my conversation with Jensen Huang. [84:30] Do you think NVIDIA may be worth 10 trillion at some point? [84:36] Let's ask it this way. What does the future [84:40] of the world look like where that's true? [84:45] - I think that NVIDIA's growth is Extremely likely, and in my mind, inevitable. And [84:55] let me explain why. We're the largest computer company in history. [85:00] That alone should beg the question, why? [85:04] And the reason of course... Two reasons. First, two foundational technical reasons. [85:10] The first reason is that computing went from being a retrieval-based, file [85:14] retrieval system. Almost everything is a file... We pre-write something, we [85:21] pre-record something. You know, we draw something, we put it on the web, we put it [85:24] in a file. And we use a recommender system, some smart [85:28] filter, to figure out what to retrieve for you. And so we were a [85:32] pre-recording, human pre-recording, and file retrieving [85:35] system. That's what a computer is, largely. [85:39] To now, AI computers are contextually aware, [85:43] which means that it has to process and generate tokens in real time. [85:47] So we went from a retrieval-based computing system to a generative-based [85:51] computing system. [85:53] We're gonna need a lot more processing in this new world than in the old [85:57] world. We need a lot of storage in the old world. We need a lot [86:01] of computation in this new world. And so, [86:06] so that's the first part of it. We fundamentally changed computing and the way [86:10] how computing is done. The only thing that would cause it to go [86:14] back...... is if this way of computation, [86:18] this way of computing generating information that's contextually relevant, [86:22] situationally aware, that is grounded on new insight before it [86:27] generates information, this computation-intensive way of doing computing [86:33] would only go back if it's not effective. [86:36] So if... For the last 10, 15 years while working on deep learning, [86:40] if at any single moment [86:43] I would have come to the conclusion that, "You know what? This is [86:47] not gonna work out. [86:49] I think this is a dead end." Or, "It's not gonna scale, it's not gonna solve this modality, not gonna be [86:53] used in this application." Then, of course, I would feel very differently about it, [86:57] but I think the last five years has given me more confidence [87:02] than the last ten year, previous ten years. The second idea, [87:07] is computers, because it was a storage system, it was largely a warehouse. [87:14] We're now building factories. Warehouses don't make much money. [87:20] Factories directly correlates with the company's revenues. And so, [87:28] the computer did two things. Not only did it change the way it did it, [87:33] its purpose in the world changed. It's no longer a computer, it's a factory. [87:39] It's a factory, it's used for generation of revenues. [87:45] We're now seeing not only is this factory generating [87:49] products, commodities that people want to consume, [87:53] we're seeing that the commodities are so [87:55] interesting, so valuable, so, to so many different audiences [88:00] that the tokens are starting to segment, like iPhones. Mm-hmm. [88:04] You have free tokens, [88:06] you have premium tokens, and you have several tokens in the middle. [88:10] Yeah. And so intelligence, as it turns out, [88:13] you know, it's a scalable product. There's extremely high intelligence [88:17] products, tokens that you could... that are used for specialized things, people be [88:21] willing to pay. You know, the idea that somebody's willing to [88:24] pay $1000 per million tokens is just around the corner. It's not if, it's only when. [88:32] And so now we're seeing that the [88:35] commodity that this factory makes is actually valuable, and is revenue [88:39] generating and profit generating. How, now the question is how [88:43] many of these factories can, does the world need? [88:48] How much, how many tokens does the world need? And [88:54] how much is society willing to pay for these tokens? And [89:00] what would happen to the world's economy [89:02] if the productivity were to improve so substantially? [89:07] What would happen... Are we, are we gonna discover new drugs, new products, [89:11] new services? And so when you take these things in combination, I am absolutely [89:16] certain that the world's GDP is going to accelerate in growth. [89:22] I'm absolutely certain the percentage of that GDP that will be used for computation [89:30] will be 100 times more than the past... because it's no [89:34] longer a storage unit. It's a product generation unit. [89:39] And so when you look at it in that context [89:42] and then you back into what is NVIDIA's, what does NVIDIA [89:46] does NVIDIA do and how much of that [89:50] new economics, new industry would we have to benefit [89:54] to address, I think we're gonna be a lot, lot bigger. [89:58] And then the rest of it, to me, is is it [90:02] possible for NVIDIA to be a, you know, $3 trillion [90:06] revenues company in the near future? The answer is of course yes. And the [90:10] reason for that is because it's not limited by any physical limits. [90:15] There's nothing that I see that says, you know, gosh, [90:19] $3 trillion is not possible. And [90:23] as it turns out, NVIDIA's supply chain is, the burden is shared by 200 companies. [90:31] And the fact that we scale out on the backs with the partnership of this ecosystem, [90:38] the question is do we have the energy to do so? [90:41] And surely we will have the energy to do so. And so all of these things combined, [90:49] that number is just a number, you know? And I still remember, NVIDIA [90:53] was a, NVIDIA was a, the first time we crossed a billion dollars, [90:58] I was reminded of a CEO who told me, "You know, Jensen, it's theoretically [91:02] impossible for a fabless semiconductor company to exceed a billion dollars." [91:07] And I won't bore you with why, but the end, [91:11] of course it's illogical and there's a lot of evidence we're not. And then, [91:15] somebody told me, "You know, Jensen, you'll never be more than $25 billion [91:21] because of some other company." Somebody told me that, "You'll never be, you know, [91:24] because..." And so the, those aren't first principled thinking. [91:32] And the simple way to think about that is what is it that we make [91:38] and how large is the opportunity that we can create? [91:42] Now, NVIDIA is not in the market share business. Almost everything that I just talked [91:46] about don't exist. That's the part that's hard. [91:51] You know, if NVIDIA was a $10 billion [91:55] company trying to take market share, then it's easy to see [91:59] for shareholders that, oh, yeah, if they could just [92:02] take 10% share, they could be this much [92:05] larger. But it's hard for people to imagine how [92:09] large we could be because there's nobody I could take share from. [92:12] You know? And so I think that that's one of the [92:16] challenges.... for the world is, is the imagination of the [92:20] future. But I got plenty of time, and I'll keep reasoning about it, and I'll keep talking [92:24] about it, and every single GTC will become more and more real. [92:27] You know, and then more and more people will talk about it, and one of these days, you [92:31] know, we'll get there. But I'm 100% we'll get there. [92:34] - Yeah, this view of you know, token factories essentially, this token per [92:39] second per watt, and every token having value. Like [92:43] it's an actual thing that brings value, and it brings different kinds of value, [92:47] different amounts of value to different people with value. That's the actual product, is really [92:51] could be loosely thought of as the token. And so you have a bunch of token factories. [92:54] And then it's very easy, first principles, to imagine a future, given all [92:58] the potential things that AI can solve, that you're going to need an [93:01] exponential number more of token factories. [93:06] - And what's really interesting, the reason why I was so excited about it, the [93:10] iPhone of tokens arrived. [93:11] - What do you call it? Wait, are you saying OpenClaw's iPhone? That's interesting. [93:15] - Agents. [93:16] - Yeah, agents. True. [93:18] - Agents in general. [93:19] The iPhone of tokens arrived. It is the fastest-growing application in [93:23] history. It went straight up. Went straight up. [93:26] - That says something. [93:27] - Yep, there's no question OpenClaw is the iPhone of tokens. [93:31] - Is there something truly, as you know, [93:34] something truly special happening from about December, [93:38] where people have really woke up to the power of Claude Code of Codex, of OpenClaw. [93:44] Um, I mean, I'm embarrassed to admit that on the way here in the airport, I've ... [93:52] It's the first time I've done this in public. I was programming, quote unquote, [93:57] by talking to my laptop. [93:59] And I was embarrassed because I was pretending like I'm talking to a human colleague. [94:04] Uh, I'm not sure how I feel about the future where everybody- [94:07] ... is walking around talking to their AI, but it's such an [94:11] efficient way to get stuff done. [94:13] - And it's more likely that your AI is bothering you all the time. [94:18] And the reason for that is because it's getting stuff done so fast. [94:22] It's reporting back to you, "I got that done." [94:24] "You know, what do you want me to do next?" You know, it... That's the part that I think- [94:28] ... most people don't realize is they're The person who's gonna be chatting with [94:32] them, texting them most, is their claws or lobster. [94:37] - What an incredible future. [94:39] Uh, I read that you attribute a lot of your success to your ability to work harder than anyone [94:43] and withstand more suffering than anyone. [94:46] So we can list many of the things that entails. I mean, dealing with failure, [94:53] the cost and engineering problems we've talked about. The human [94:57] problems, uncertainty, responsibility, exhaustion, embarrassment, [95:01] the near-death company moments that you've mentioned, [95:04] But also the pressure. Now, as the CEO of this [95:10] company that economies and nations strategize around plan their, [95:18] Financial allocations around, plan their AI [95:21] infrastructure around, how do you deal with this much pressure? [95:26] What gives you strength, given how many nations and peoples depend on you? [95:38] - I'm conscious about the fact that, [95:42] NVIDIA's success is very important to United States. [95:47] We generate enormous amounts of tax revenues. [95:51] We established technology leadership for our nation. Technology [95:54] leadership is important for national security. [95:57] National security not just in one aspect of national security, all aspects [96:01] of national security. When our country's more prosperous, [96:05] we could do a better job with domestic policies and helping social benefits. [96:11] Because we're generating so much re-industrialization in the United States, [96:15] we're creating mountains of jobs. We're helping shift, how we, [96:23] how we build things back to the United States in so many [96:27] different plants, chips, computers, and of course, these AI [96:31] factories. I'm completely aware [96:34] that, that... And I have the benefit, and this is a real [96:40] a real gift with mainstream investors, [96:45] teachers, policemen who have somehow, for whatever reason, [96:51] invested in NVIDIA or because they watched Jim Cramer bought [96:55] some stock and now are millionaires. And I am completely aware of [97:02] that circumstance. I'm aware of the circumstance that NVIDIA, [97:07] is central to a very large network of [97:13] ecosystem partners behind us and downstream from us. [97:16] And so the way I deal with that is exactly what I just did. I reason about [97:23] what is... what is it that we're doing? [97:26] What is it causing? What's the impact that has on other people [97:29] benefit, you know, positively or even, [97:33] even through great burden, for example, to supply chain? [97:38] And the question is, therefore, what are you gonna do about it? [97:44] In almost everything that I feel, I break it down, I reason about, "Okay, [97:50] "what's the circumstance? What has changed? What's hard? [97:54] And what am I gonna do about it?" And I break it down, decompose the problem, [98:00] and the decomposition of these circumstances [98:07] turns it into manageable things that I can do. And the only thing after that [98:11] I could do is, "Did you do it? Did you either do it or did you [98:15] get somebody else to do it? And if you didn't do it, [98:18] you reasoned that you need to do it, and you didn't do it, and you didn't get anybody else to do it, [98:23] then stop crying about it." You know? And so- ... so I'm fairly [98:31] Tough on myself. And, but I also break things down so that, [98:35] so that I don't panic. I can go [98:39] to sleep because I've made the list of things that needed to be done, and [98:43] I've made sure that everything that could put our company in harm's way, [98:47] could put my partners in harm's way, put our industry in harm's way, I've told [98:51] somebody. Everything that I feel [98:57] could put anybody in harm's way, I've told someone. And I've told [99:00] that someone who could do something about it. And so I've gotten it off my chest [99:05] or I'm doing something about it. And so after that, Lex, what else can you do? [99:10] - So given all the insane, intense amount of suffering [99:15] on the journey of building up NVIDIA, have you hit low points psychologically? [99:22] - Oh, yeah. Oh, yeah. Sure. All the time. All the time. [99:27] - And there— [99:27] - All the time. [99:27] - ... you just break down the problem into pieces? [99:31] - Yeah. Yeah. [99:31] - See what you could do about it? [99:33] - And, and part of, and, you know, Lex, part of it, part of it is forgetting. [99:39] One of the most important attributes of AI learning, as you know, is, [99:43] right? Systematic forgetting. You, you need to know when to forget [99:47] some things. You can't memorize everything. You can't keep everything and, and, [99:51] you know, you, you want to— you don't want to carry everything. One of the things that I do very [99:54] quickly is decompose the problem, I reason about the problem, and I [99:58] share the load with it. When I say I tell everybody, [100:01] I'm essentially sharing that burden. As quickly as possible. [100:06] Whatever worries me, tell somebody else. Don't just keep it. You know, [100:10] don't freak them out. Decompose the problem into smaller parts and get [100:16] people to, so, and, and inspire them to be able to go do something about it. [100:20] But part of it is just, just forgetting. You know, like, [100:24] a lot of it is you gotta be tough on yourself. You know, just come on, stop [100:28] crying about it. Let's get going. You know? And, and then you get out of bed. And then [100:32] the other part is, is you, you're attracted to the [100:36] next shiny light, the next future, you know, the next opportunity, the [100:40] next, "Okay, that's behind us. Let— what's next?" [100:43] It's a lot, I think, you know, you watch this with great athletes. They, [100:47] they just worry about the next point. [100:50] The last point is behind them. The embarrassment, the, you know— the setback. [100:56] You know, and, and then, and because I do so much of my job publicly, [101:01] you know? Lex, you do a fair amount of your job publicly too. And so, so I do a lot of my [101:05] job publicly. And so you know, I say a lot of [101:09] things that, that seem sensible at the time or funny at the time, [101:13] mostly it's just because it's funny to me at the time. And then, you know, you [101:17] reflect on it, it's less funny, but, but... [101:20] - Yeah. No, trust me, I know. But you basically allow yourself to [101:23] be pulled by the light of the future. Forget the past and just keep- [101:27] - That's right. [101:28] - ... keep working towards that. I mean, you did say, there's this kind of [101:32] famous thing you said that if you knew how hard it would be [101:38] to build NVIDIA it turned out to be, what is it? A [101:42] million times more hard than you anticipated that you wouldn't do it. [101:46] - Yeah, right. [101:47] - Um, but isn't... You know, when I hear that, [101:51] that's probably true about everything worth doing, right? [101:53] - Exactly. That is, by the way, what I was trying to explain, is that there's [101:59] a, there's a incredible superpower of being, being have a, the mind of a child. [102:07] You know? And I say to myself oftentimes when I look at something, [102:11] and almost, almost everything, My first thought is, "How hard can it be?" [102:19] You know? And so you get yourself into that mode, how hard could [102:23] it be? And, and nobody's ever done it. It [102:27] looks gigantic. It's gonna cost hundreds of billions of dollars. [102:31] It's gonna take, you know, all this... And you just go, "Yeah, but how hard could it be?" [102:35] You know? How hard could it be? [102:37] And so you gotta get yourself into that state of mind. You don't [102:41] wanna, you don't wanna actually over simulate [102:45] everything and all the setbacks and all the trials and tribulations and all [102:48] the disappointments. You don't wanna simulate all that in advance. You don't wanna know that. [102:53] You don't, you wanna go into a new experience thinking it's gonna be perfect, [102:56] it's gonna be great, it's gonna be incredibly fun. And then while you're there, [103:01] you know, you need to have, [103:03] you need to have endurance, you need to have grit, so that when the setbacks [103:07] actually happened, and those setbacks are gonna surprise you, the [103:10] disappointments are gonna surprise you, you know, the embarrassments are [103:14] gonna surprise you, the humiliations are gonna surprise you. [103:17] You just can't let... Now you just gotta turn on the other bit, which is just forget [103:21] about it. Move on, keep moving. And, and to the extent that, [103:27] to the extent that my assumptions [103:31] about the future and why the future is gonna manifest, [103:36] so long as those assumptions and that input [103:40] doesn't change or didn't change materially, then I should [103:43] expect that the output won't change. And so my simulated output of the future [103:49] is still gonna happen. And if it's still gonna happen, [103:53] I'm still gonna go after it. I believe it's gonna, you know, and so there's a combination [103:57] of two or three human characteristics, [104:01] the ability to go into a, into an experience fresh-minded, [104:05] the ability to forget the setbacks, the ability to believe in yourself, [104:11] you know, to believe what you believe and stay, stay true to that belief. [104:16] But you're constantly reevaluating. This combination of three, four, five things [104:23] I think is, is really important for resilience. And, [104:27] and... and, you know, I'm fortunate that, [104:32] that whatever life experiences led to this, I've got kind of those [104:36] four, five things. You know, I'm always curious, always learning. [104:40] I'm always learning from everybody, you know? I'm always asking my... [104:44] And because I'm humble about everything, I'm always thinking, [104:48] "Gosh, they did that so nicely. They did that so wonderfully." You know, [104:53] I wonder what they're thinking through. How do they... You know, so I'm [104:56] simulating everybody. In a lot of ways, you know, I'm emulating almost everybody [105:00] I watch, right? You're empathetic towards everything that they [105:04] do that you're observing and respect. And, and so you're [105:08] constantly learning and, you know. [105:10] - You're now one of the wealthiest people on Earth. One of the [105:14] most successful humans on Earth. [105:18] Is it harder to be humble and to be able to... Do you feel the [105:21] effect of money and power and fame in making it harder for you to [105:28] sort of be wrong in your own head? Enough [105:32] to hear out an opinion of somebody else when they [105:36] disagree with you and learn from them? Those kinds of things. [105:41] - Um, surprisingly, no. And I would, I would actually go the other way. [105:46] Because I do so much of my work publicly, when I'm wrong, [105:51] pretty much everybody sees it. [105:53] - You get humbled. Fair enough. [105:55] - And when I'm wrong, when I'm wrong or it didn't turn out that [105:59] way or you know, I mean, most of the things that, that I say outside I'm fairly [106:07] certain about. And the reason for that is because, because it's gonna [106:11] impact somebody else and I want to be quite concerned about that and quite, [106:14] circumspect about that. For stuff that, that I'm reasoning about inside a meeting, [106:20] you know a lot of things could turn out differently. And [106:23] so, but it doesn't ever stop me from reasoning. The way that the way that I [106:27] manage and lead, I, [106:30] you know, I'm constantly reasoning in front of people. And even when I'm talking to you, you can [106:33] kind of see me kind of reasoning through things. [106:35] And I want to make sure that you understand what I'm saying not because I told you- [106:40] ... because I'm so humble about what I'm about to tell you. [106:43] I kind of show you the steps that I got there. [106:46] And then you can decide whether you believe what I said in the end. And so I'm doing that [106:50] all day long in meetings. [106:52] With all of my employees, I'm constantly reasoning through, "Let me tell you what, how I [106:56] see it." And then I reason through it. It gives everybody the [106:59] opportunity to intercept and say, "I disagree with that part." [107:04] The nice thing about reasoning through things and letting people interact with [107:08] it is that they don't have to disagree with your outcome. [107:12] They can disagree with your reasoning steps. [107:15] And they could pull me in different directions, and then we can reason forward. [107:19] And so we're kind of, you know, collective [107:25] path searching method. And it's really fantastic. [107:29] - Yeah, you have this [107:30] way about you of ... When you're explaining stuff, I can feel you [107:34] actually reasoning on the spot about it with a [107:38] constant open-mindedness where you could ... I could feel like I could [107:42] steer your thinking. [107:44] And that's a, that's really beautiful that you've been able to maintain that after so many [107:48] years of success, and pain. [107:50] I think sometimes pain makes you close, closes you down a bit. [107:55] - Mm-hmm. Yeah. [107:56] - And I think to maintain- [107:57] - Yeah. Tolerance for embarrassment, I think is... [107:59] - Yes, that's ... The tolerance ... I mean, that's a real thing. [108:03] Is many years of embarrassing yourself. Even those meetings [108:07] knowing that there's people around you where you declared one idea and [108:11] it was shown that that idea was wrong- [108:13] ... and be able to admit that and to grow from that. That's not, that's very difficult on a human [108:17] level. [108:17] - Yeah. Well, you know. They knew that recently my first [108:21] job was, you know, cleaning toilets, so. [108:25] - I'm glad you maintained that same spirit of Denny's [108:29] the, the work. I mean, that, that was beautiful. Your whole journey from, starting from Denny's is a [108:33] beautiful one. Let me ask you about video games. So I'm a big gaming fan. [108:41] So I have to say thank you to NVIDIA for many years of incredible graphics. [108:47] - By the way, GeForce is our still, to this day- [108:51] ... our number one marketing strategy. [108:55] Right. People learn about NVIDIA while they're in their teenage years. [109:00] And then they go to college and they know who NVIDIA is and then in beginning [109:04] is just, you know, playing Call of Duty, you know? You know, Fortnite. And then later [109:08] they're using CUDA, and then later they're using NVIDIA and, you know, Blender and [109:14] Dassault and Autodesk. [109:16] - Yeah. I mean, I should say I mentioned to a friend that I'm [109:20] talking with you. He said, "Oh, they make great gaming GPUs." [109:25] - Yeah, exactly. [109:26] - It's like- [109:26] - Exactly. [109:27] - You know, there's more to it, but, yeah, yeah, [109:31] people really love the ... It really brought a lot of joy to a lot of people. The, [109:35] the, the hardware really brings these worlds to life. [109:38] There was some controversy around this with DLSS 5. [109:44] Can you explain to me the drama around this? Uh, I guess [109:48] people, the gamers online were concerned that it makes games look like AI slop. [109:54] Uh, what do you think of this drama? [109:56] - Yeah. Uh, I think their perspective makes sense [110:01] and I could see where they're coming from, because I don't love AI slop myself. [110:05] You know, all of the AI generated content increasingly, [110:10] um, looks similar and they're all beautiful, [110:14] and I can... So I can... I'm empathetic towards what they're thinking. [110:18] That's just not what DLSS 5 is trying to do. I showed several examples of it. [110:24] But DLSS 5 is [110:27] 3D conditioned, 3D guided. It's ground truth structure data guided. And so the [110:34] artist determined the geometry. We are completely truthful.... to [110:40] the geometry maintains in every single frame. [110:43] It's conditioned by the textures, the artistry of the artist. And so every single [110:50] frame, it enhances but it doesn't change [110:54] anything. Now, the question is, the question about enhancing, [110:59] DLSS 5 also lets, because it's, [111:02] the system is open, you could train your own models to determine, [111:06] and you could even in the future prompt it. You know, I want it to be a toon [111:10] shader. I want it to look like this kinda, you know, so you can give it even [111:14] an example. And it would generate in the style of that, [111:19] all consistent with the artistry, you know, the [111:22] style, the intent of the artist. And so all of that is done [111:29] for the artist, so that they can create something that is more beautiful, [111:34] But still in the style that they want. I think that [111:38] they got the impression that the games are gonna [111:42] come out the way the games are shipped the way they do, and then [111:46] we're gonna post-process it. That's not what DLSS is intended to do. [111:50] DLSS is integrated with the artist, and so it's, [111:54] it's about giving the artist the tool of AI, the tool of [111:58] generative AI. They could decide not to use it, you know? [112:01] - I think people are very sensitive to human faces. [112:04] And we're now living in this moment, which I think is a beautiful one, which [112:08] is people are sensitive to AI slop. [112:11] It puts a mirror to ourselves to help us realize that what we seek is [112:15] imperfections. What we seek is sometimes not perfect graphics. [112:20] It helps us understand what we find compelling in the worlds we create. [112:25] And that's beautiful. And as long as it's tools that help us create those worlds- [112:28] - Yeah, that's right [112:28] - ... it's wonderful. [112:29] - That's right. Yet, yet another tool, and they want the generative, [112:33] models to generate the opposite of photo real. [112:39] Yeah, it'll do that too. And so it's just yet another tool. I think the [112:43] the gamers might also appreciate that, that in the last couple of years, [112:50] we introduced skin shaders [112:55] to the game developers. And many of those games have skin shaders [112:59] that include subsurface scattering that make skin look more skin-like. And so the [113:06] industries, you know, game developers are looking for more and more and more [113:10] tools to express their art. And so this [113:14] is just yet more, one more tool, and they get to decide what to use. [113:16] - Ridiculous question. What do you think is the greatest or most influential [113:20] game ever made? Maybe from NVIDIA's perspective? [113:24] - Doom. [113:25] - Doom, unquestionably. That was the start of the 3D. [113:28] - I would say Doom, from an art, the intersection of the [113:32] cultural implication as well as the industry, turning a PC into a gaming device. [113:38] That was a very important moment. Now, now of course, flight simulation companies were [113:42] before it. And but they just didn't have the popularity that Doom did to have made [113:48] the industry turn the PC from an office automation tool into a personal computer for [113:55] families and gamers and things like that. And so Doom was really impactful there. [113:59] From an actual game technology perspective, I would say Virtua Fighter. [114:03] And so we're great friends with both of them, you know? [114:07] - And then there's games more recently, I mean, Cyberpunk 2077, [114:12] really nice GPU-accelerated graphics. Like- [114:16] - Fully ray traced. [114:17] - Fully ray traced. Also, I like, I personally, I'm a huge fan of Skyrim, [114:22] uh, Elder Scrolls, and the, you know, it's, it's been released a long, long time [114:26] ago, but people release mods and- [114:29] - We love mods [114:30] - ... they create these, these inc- I mean, it- ... it's like a different game and it just [114:34] allows me to replay the game over and over and get i- [114:37] It makes you realize that you can re- experience in a [114:41] totally new way the world you already love. So- [114:45] - That's right [114:45] - ... I do that all the time. One of my favorite things is just walk across Skyrim. [114:48] - Uh, we created this thing called RTX Mod. Yeah, it's a modding tool. [114:53] - Awesome. [114:53] - It allows the community to inject the latest technology into an old game. [115:00] - Of course, like what makes a great video game is not just graphics, it's also story and [115:04] character development, but- [115:06] - That's right [115:06] - ... beautiful graphics can add to the immersion. [115:10] The feeling like it's another place you're transported to. [115:15] Ah, what you said, I think accurately, that the AGI timeline [115:22] question rests on your definition of AGI. [115:26] So, let me ask you about possible timelines here. [115:31] Let's, this ridiculous definition perhaps of what AGI is, [115:35] but an AI system that's able to essentially do your job. So, run, no, [115:44] start, grow, and run a successful technology company that's worth- [115:52] - A good one or a one? [115:54] - No. It has to- It has to be worth more than a billion, [115:58] more, more than a billion dollars. So, [116:02] you know, you know how hard it is to do all those components. [116:06] So, how far are we away from that? So, we're [116:10] talking about OpenClaw that does all the [116:14] incredibly complex stuff that are required to [116:18] to, first of all, innovate, to find customers, to sell to them, to [116:22] manage, to build a team of [116:26] some agents, some humans, all that kind of stuff. Is this five, 10, [116:29] 15, 20 years away? [116:31] - I think it's now. I think we've achieved AGI. [116:35] - Do you think you could have a company run by an AI system like this? [116:37] - Possible, and the reason for that is this. You said a billion, and you didn't [116:41] say forever. And so for example, uh... It is not out of the question [116:48] that a Claw was able to create a web service, some interesting little app that [116:59] all of a sudden, you know, a few billion people used for 50 cents, and then [117:07] it went out of business again shortly after. Now, we saw a whole bunch of those type of companies during the [117:11] internet era, and most of those websites were not anything more [117:17] sophisticated than what OpenClaw could generate today. [117:20] - Interesting. Achieve virality and monetize that virality. [117:23] - Yeah. It's just that I don't know what it is, but I couldn't have predicted any of [117:27] those companies at the time either, you know? And - [117:30] - You're gonna get a lot of people excited with that statement. [117:33] It's like, what do you mean? I can- I can just, uh - ... launch an agent and [117:37] make a lot of money. [117:38] - Well, by the way, it's happening right now, right? You know that when, when you go to China [117:42] you're gonna see, you're gonna see a whole bunch of people [117:46] teaching their, getting their Claws to try to go out and look for jobs and, you know, [117:51] do work, make money. [117:53] And I'm not, I'm not actually... I wouldn't be surprised if some [117:57] social thing happened or somebody created a, a [118:01] digital influencer, super, super cute or [118:04] some social application that, you know, feeds your little [118:08] Tamagotchi or something like that, and, and it become an out of the [118:12] blue an instant success. A lot of people use it [118:16] for a couple of months and it kind of dies away. Now, the odds of, [118:22] you know, 100,000 of those agents, Building NVIDIA is zero percent. [118:28] And then the one part that I will, I won't do, [118:32] And I, I want to make sure we all do, is to [118:36] recognize that people are really worried about their jobs. [118:40] And I just want to remind them that the purpose of your job [118:47] and the tasks and tools that you use to do your [118:50] job are related, not the same. I've been doing my job for [118:54] 33 years. I'm the longest running tech CEO in the world, 34 years. [118:58] And the tools that I've used to do my job have changed [119:03] continuously in the last 34 years, and sometimes quite [119:07] dramatically, you know, over the course of a couple, two, three [119:10] years. And the one story that I really [119:14] wanna make sure that everybody hears is the story that the [119:18] first job that computer scientists said, AI [119:22] researchers said was gonna go away was radiology. [119:25] Because computer vision was going to achieve superhuman levels, and it did. [119:30] CV... Computer vision was superhuman in 2019, 20, maybe a little bit later, 2020? [119:39] Okay? And so it's been a long time since computer vision has been superhuman. And [119:43] so the prediction was radiologists would go away because studying [119:47] radiology scans was a thing of the past. AI will do that. Well, [119:52] they were absolutely right. [119:54] Computer vision is completely superhuman. Every radiology [119:58] platform and package today is driven by AI, [120:02] and yet the number of radiologists grew. And so [120:06] the question is why? And we now have a shortage of radiologists in the world. [120:11] And so, one, the alarmist warning went too far and it scared people from [120:20] doing this profession that is so important to society. [120:23] And so it did harm. Now, why was it wrong? The reason why is because [120:29] the purpose of a radiologist, the purpose is to diagnose disease [120:33] and help patients and doctors diagnose disease. [120:38] And because we're able to study scans at so much faster [120:41] now, you could study more scans, you could diagnose better, you could, [120:47] you could inpatient faster, [120:50] you can see people more. The hospitals are making more money. You have [120:54] more patients in the hospital. You need more radiologists. I mean, the [120:58] amazing thing is, it's so obvious this was gonna happen. [121:03] The number of software engineers at NVIDIA is gonna grow, not decline. [121:08] And the reason for that is because the purpose of a software engineer [121:12] and the task of a software engineer coding are related, not the same. [121:16] I wanted my software engineers to solve problems. I didn't care how many lines of code they [121:20] wrote, you know? But their job, their purpose of their job didn't change. Solving [121:26] problems, working as a team, diagnosing problems, evaluating the result, [121:32] looking for new problems to solve, innovation, connecting [121:35] dots. You know, none of that stuff is gonna go away. [121:39] - Do you think it's possible that... Let's even take coding. Do you think the number of [121:43] programmers in the world might increase, not decrease? [121:45] - Yes. And the reason for that is this. What is the definition of coding? [121:52] I believe it is... The definition of coding, as of today, is simply specifying, [121:58] specification, and maybe if you want to be [122:02] rather directive, you could even give it an architecture of the software [122:06] that you wanted to write. So the question is, how many people could do that? [122:11] Describe a specification for a computer to go... Telling the [122:14] computer what to go build. How many people? I think we just went from 30 [122:18] million to probably 1 billion. And so every carpenter in the future will be [122:26] a coder, except a carpenter with AI is also an architect. [122:33] They've just increased the value that they could deliver to the customer. Their, [122:37] their artistry just elevated tremendously. [122:43] I believe that every accountant is, you know, also your financial analyst, [122:47] also your financial advisor. So, all of these professions have [122:51] just been elevated.... and, and if I were a carpenter, I sees a, [122:55] I see AI, I would just completely go berserk. [122:58] You know, the services I can bring to my clients if I were a plumber, [123:03] completely go berserk. [123:04] - And the people that are currently programmers and software engineers, I think [123:08] they're at the cutting edge of understanding intuitively how to communicate [123:15] with the agents using natural language in order to design the best kind of software. [123:20] - That's right, exactly. [123:20] - So over time they'll converge, but I think [123:24] there's still value in getting, I think learning how to program, [123:28] like learning what programming languages are. The old, the old kind of [123:32] programming, what are good practices for [123:36] programming languages, what are design principles for programming- [123:39] - That's right [123:40] - ... Languages for large software systems? [123:43] - And the reason for that, [123:45] Lex, and you know, as you're saying for the audience, I think [123:49] the goal of, the goal of specification, the artistry of [123:53] specification, the goal and the artistry of it, [123:56] Is going to depend on what problem you're trying to solve. [124:01] When I'm thinking, when I'm thinking about giving the company strategies and [124:06] formulating corporate directions and things that we should do, [124:11] I describe it at a level that is sufficiently [124:16] specific that people generally understand the direction [124:21] and it's actionable. It's specific enough that they can take action on it, [124:26] but I under specify it on purpose, so that [124:30] enable 43,000 amazing people to make it even better than I imagined. [124:36] And so when I'm working with engineers and when I'm working with people, [124:41] I think about who, what problem am I trying to solve? Who am I working with? [124:47] And the level of specification, the level of architecture definition [124:54] relates to that. And, and so [124:59] everybody's going to have to learn how, where in the spectrum of coding they want [125:03] to be. Writing a specification is coding. [125:06] And so you might decide to be quite prescriptive because there's a very [125:10] specific outcome you're looking for. You might decide that, you know, [125:14] this is an area you want to be much more exploratory, and so you might under [125:18] specify and enable you to go back and forth with the [125:21] AI to even push your own boundaries of creativity. [125:25] And so this artistry of where you are in the spectrum, this is the future of coding. [125:31] - But just to linger on it outside of coding, I think a lot of people, rightfully so, [125:36] are worried about their jobs, have a lot of anxiety about their jobs, especially in [125:40] the white-collar sector. I don't think any of us know what to do, [125:50] With tumultuous times that always come when automations and new technology [125:54] arrives. And I just... First of all, I think [126:01] we all need to have compassion and the responsibility to feel sort of the burden [126:06] of what the actual suffering feels like for individual people and families that lose their [126:10] job. I think whenever you have transformative technology [126:14] like that's coming with artificial intelligence, there's going to be a lot of pain, [126:18] and I don't know what to do about that pain. Hopefully, it creates much [126:22] more opportunities for those same people, for the same kind of job as the tooling [126:30] evolves and makes them more productive and makes them more fun, [126:34] hopefully, as it does in the programming. I've been having so much fun programming, [126:38] I have to say. Like, I've never had this much fun. So hopefully it makes their job, [126:42] automates the boring parts and makes the creative parts, [126:46] the ones that the human beings are responsible for. But still there's going to be [126:50] a lot of pain and suffering. [126:51] - So my first recommendation before... And this is now how I deal [126:55] with anxiety. In fact, we just talked about it earlier. [126:59] Enormous anxiety about the future, enormous anxiety about the pressure, enormous [127:03] anxiety about uncertainty, I first break it down, and then I'm gonna tell myself, [127:10] "Okay, there are some things you can do something about, there's some things you can't do anything [127:14] about. But for the stuff that you can do something about, let's reason [127:18] about it and let's go do it." [127:20] If we were to hire a new college graduate today, and I have a choice between two, [127:26] one that have, that is no clue what AI is [127:31] and one that is expert in using AI, I would hire the one who's expert in using AI. [127:38] If I had an accountant, a marketing person, [127:42] the one that is expert in using AI, supply chain, [127:46] customer service, a salesperson, business development, a lawyer, [127:52] I would hire the one who is expert in using AI. [127:55] And so I would advise that every college student, [127:59] every teacher should encourage their student to be, to go use [128:03] AI. Every college student should graduate and be an [128:07] expert in AI. And everybody, if you're a carpenter, if [128:11] you're, you know, electrician, go use AI. [128:15] Go see what it can do to transform your current job, [128:19] elevate yourself. If I were a farmer, I would absolutely use AI. [128:24] If I were a pharmacist, I would use AI. I wanna see how, what it [128:28] could do to elevate my job so that I could be the innovator to [128:33] revolutionize this industry myself. [128:36] And so that would be the first thing that I would do. And then I would [128:39] also, I would also help them... [128:44] it is the case that the technology will dislocate and will eliminate many tasks [128:52] if... And because it will automate it, if your job is the [128:56] If your job is the task, then you're very highly going to be disrupted. [129:02] If your job's purpose includes you, certain tasks- [129:08] ... then it's vital that you go learn how to use AI to automate those tasks. [129:12] And then there's the world of spectrum in between. [129:14] - And by the way, the beautiful thing about AI, so the chatbot versions, [129:21] is you can break down... You have anxiety [129:24] and you can break down the problem by talking to it. [129:28] Like, I've recently... It's really just incredible how much you can think [129:32] through your life's problems, and through... And I don't mean, like, therapy problems. [129:35] I mean, like, very practically, "Okay, I'm worried about my..." Literally, [129:39] "I'm worried about my job. What are the skills? What are the steps I need to take? How do I get better at [129:43] AI?" Everything you just said, you could literally ask and it's going to give you- [129:47] ... a point-by-point plan. I mean, it's just a great life coach, period. This- [129:51] - I don't know how to use AI, and the AI goes, "Well, let me show you." [129:54] - Exactly. It's very meta, but it's- [129:57] It's kind of incredible. So people definitely should- [130:00] - You can't walk up to Excel and say, "I don't know how to use Excel." You're done. [130:03] - I mean, that's really what AI has done for me in all walks of life, [130:07] is that initial friction of being a beginner of using a thing for the first time. [130:11] I can literally ask about any single thing, [130:15] "What are the first steps I need to take?" [130:16] - That's right. [130:17] - And that handholding that it does, removing the friction [130:21] of all the experiences that the world offers is... You [130:25] know, like I mentioned to you offline, you mentioned, "I'm going to China [130:29] and Taiwan." [130:30] - So awesome. [130:31] - Just ask, "Where do I-" [130:32] - So excited for you. [130:32] - "Where do I—what do—" "You know, where do I go? How do I..." All of those questions— [130:35] ... immediately answered, and it's beautiful. [130:37] - Well, when you, when you go to Taiwan, just ask AI- ... "What are [130:41] Jensen's favorite restaurants in Taiwan?" And it'll actually- [130:45] - You don't know? [130:45] - Oh, yeah. [130:46] - Is it accurate? Okay. [130:47] - Yeah. [130:47] - All right. [130:47] - It's all over Taiwan. [130:50] - Well, you're a rockstar over there. And like we also mentioned [130:54] offline, maybe our paths will cross, which would be really wonderful in computing. [130:58] - COMPUTEX. NVIDIA GTC Taiwan. [131:01] - Uh, do you think there's some things about human nature, about human consciousness [131:08] that is fundamentally non-computational? Maybe something a chip, no matter [131:14] how powerful, can never replicate? [131:18] - I don't know if the chip will ever get nervous. And that's the, you know, of course, [131:22] the conditions by which that causes anxiety or nervousness or whatever emotion. Um, [131:31] I believe that AI will be able to recognize those [131:36] and understand those. I don't think my chips will feel those. [131:41] And therefore, the... How that anxiety, how that [131:45] feeling, how that excitement, how that, how that, you know... [131:50] All of those feelings manifest in human performance. For example, [131:54] extremely amazing human performance, athletic performance, you know, [131:59] average or lesser than average. That entire [132:02] spectrum of human performance that comes out of [132:06] exactly the same circumstances for different people, [132:10] manifesting a different outcome, manifesting a different performance. I don't think [132:17] there's anything about anything that we're building that would suggest [132:22] that two different computers [132:24] being presented with all of exactly the same context would perfo- [132:29] Of course, it would produce statistically different outcomes, but it's not because [132:33] it felt different. [132:34] - Yeah, the subjective... [132:36] Boy, there's something truly special about the subjective experience [132:41] that we humans feel. Like I mentioned to you, I was pretty nervous talking to you. [132:47] Like I mentioned to you, that, the hope, the fear, the anxiety, [132:51] and just life itself, the richness of life. How amazing everything is. How [132:55] deeply we fall in love, how deeply our hearts get broken, [132:59] how afraid we are of death and how much pain we feel when our [133:03] loved ones pass away. All of that, the whole thing. I know it's very hard to- [133:09] ... think AI being able to... A computational device being able to do that. [133:13] But there's so many mysteries about this whole thing that we're yet to [133:17] uncover, that I am open to be surprised. I've been surprised a lot over the past- [133:23] ... few months and few years. Scaling can create some [133:27] incredible miracles in the space of intelligence. [133:30] Has been truly marvelous to watch, so I'm open to surprise. [133:34] - And it's just really important to break down what is [133:37] intelligence. You know, the word, that word we use all the time, it's [133:41] not a mysterious word. Intelligence has a meaning, you know? [133:46] And it's a system that... You know, it's something that we [133:50] do that includes perception and understanding and [133:54] reasoning and the ability to do plan. And, you know, that, that [133:58] loop, that loop, is the... Fundamentally what intelligence is. [134:04] Intelligence is not one word that is exactly equal to humanity. [134:11] And that's, I think it's really important to separate the two. We have two words for [134:14] that. I'm not... I don't over-fantasize about, and I don't [134:20] over-romanticize about intelligence. Intelligence [134:23] is... And people have heard me say it before, I actually think [134:27] intelligence is a commodity. I'm surrounded by intelligent people. [134:33] And I'm surrounded by intelligent people more intelligent than I am in each one of the [134:37] spaces that they're in. And yet, [134:40] I have a role in that circle. It's actually kind of interesting. [134:45] They're more educated than I am. They went to better schools than I did. [134:51] They're deeper than, in any of the fields that they're in. All of 'em. [134:55] I have 60 of 'em. They're all superhuman to me. [134:59] And somehow, I'm sitting in the middle orchestrating all 60 of 'em. [135:03] And so you gotta ask yourself... Uh, what is, what is [135:06] it about a dishwasher that allows that [135:10] dishwasher to sit in the middle of superhumans? Does that make sense? [135:15] And so, but that's my point. My point is intelligence is a functional thing. [135:22] Humanity is not a, not specified functionally. It's a much, much bigger word. [135:29] And, and our life experience, our tolerance for [135:33] pain, our determination, those are, those are different words than intelligence. [135:39] And so the thing that I wanna help the audience understand, [135:43] if I could give them one thing, is, is [135:46] intelligence is a word that we've elevated to a very high form over time. [135:50] - The, the word we should really elevate is humanity. [135:53] - Character, humanity. [135:55] - All those things. [135:55] - All of those things. Compassion, generosity, [136:00] all of the things that you said just now, [136:03] I believe those are superhuman powers. And that now [136:07] intelligence is gonna be commoditized. [136:09] Because we've spoken about it, the most important thing is your education. [136:13] The most... Now, even, even when they said the most important thing is your education, [136:17] when you went to school, there's more than just knowledge that you gained. [136:22] And so, but unfortunately, our society has put everything into one single word, [136:28] and life is more than one word. And I'm just telling you, my life would suggest [136:34] that being lower on the intelligence curve than everybody around me [136:41] doesn't change the fact I'm the most successful. And so, and I [136:45] think, I think that kind of is I'm trying to hopefully inspire everybody [136:49] else that don't let this democratization of [136:52] intelligence, this commoditization of intelligence, [136:57] you know, cause you anxiety. You should be inspired by that. [137:00] - Yeah. I think AI will help us celebrate humans more. And certainly [137:07] humanity and human first, and I, I think what makes this world incredible is humans [137:13] forever will be so, and just AI is this incredible tool that makes us- [137:18] - That's exactly right. [137:18] - ... humans more powerful. [137:20] - That's exactly right. [137:21] - Uh, so much of the success of NVIDIA [137:25] and the lives of millions of people that I mentioned depend on you. [137:31] But you're just one human, like we mentioned, [137:35] a mortal like all of us. Do you think about your mortality? Are you afraid of death? [137:42] - I really don't wanna die. Um, I have a great life. I have a great family. [137:48] Uh, I have really important work. Uh, this is, this is not a once in a, [137:58] once in a lifetime experience suggests [138:02] that it has been experienced by many people, just not one [138:05] person. This is a once in a humanity experience, what I'm going through. [138:11] Uh, NVIDIA is one of the most consequential technology companies in history. We're doing [138:15] very important work. I take it very seriously. Um, [138:20] And so some of the, some of the things that, that of course are, [138:24] are practical things, like how do we think about succession planning? And, [138:31] I'm famous in saying that I don't believe in succession planning. [138:36] - Man. [138:36] - And the reason, the reason for that, the reason for that isn't because I'm [138:40] immortal. The reason for that is because if you're worried about [138:46] succession planning, [138:48] if you're worried, all that anxiety of succession planning, then what should you do about it? [138:52] Then you break it all the way back down. [138:54] The most important thing you should do today, if you care about the future of your [138:57] company, post you, is to pass on knowledge, information, insight, skills, [139:04] experience as often and continuously as you can, which is the reason [139:08] why I continuously reason about everything in front of my team. [139:13] Every single meeting is about a reasoning meeting. [139:16] Every moment I spend inside a company, outside a company is [139:20] about passing on knowledge to people as fast as I can. Nothing I [139:24] learn ever sits on my desk longer than, you know, [139:28] a fraction of a second. I'm passing that information, that knowl- oh my gosh, this is [139:32] cool. Before I even finish learning all of it myself, I'm [139:36] already pointing it to somebody else. "Get on this. This is so cool. You're gonna wanna, you're [139:40] gonna wanna learn this." And so I'm constantly passing [139:44] knowledge, empowering people, elevating the capability of [139:48] everybody around me, so that the outcome that I, that I seek, [139:55] that I hope for, is that I die on the job, you know? And, [139:59] and hopefully I die on the job instantaneously, you know? [140:03] And there's no long periods of suffering, you know? It's, uh — [140:06] - Well, from a fan perspective given your, your extremely, [140:13] um, your enormous positive impact on civilization, of course, I hope you keep [140:18] going. But also it's just fun to watch what NVIDIA is doing, you [140:22] know. It's just the rate of innovation. And I'm a huge fan of engineering. [140:26] There's so much incredible engineering being continuously being done by [140:30] NVIDIA. It's just fun to watch. It's a celebration of humanity, a celebration [140:34] of great builders, a celebration of great engineering. So, it represents [140:37] something special. So I hope you and NVIDIA keep going. What gives you hope [140:44] about this whole thing we got going on, about humanity, about the future of [140:47] humanity? When you look out, when you think about the future quite a bit, when you [140:51] look out 10, 20, 50, 100 years from now, what gives you hope? [140:56] - I've always had a great confidence in the kindness, uh, the generosity, [141:09] uh... the compassion, the human capacity. I've always been [141:16] extremely confident of that. Sometimes more so than I should. [141:25] And, and I get taken advantage of, but it doesn't, it doesn't [141:29] ever cause me not to. I start with, always, [141:34] That people want to do good. People want to, um help others. And, [141:44] vastly, I am proven right. Constantly proven right. And, and [141:51] often it exceeds my expectations. [141:56] And, and so I have complete confidence in the human capacity. I think the, [142:03] the thing that, the things that give me incredible hope, [142:07] Is what I see as, as I extrapolate, as I, what I see now is possible, and as I [142:14] extrapolate, Based on the things that we're doing, what will very likely happen. [142:22] And, and that there's so many things that we wanna solve. [142:27] There's so many problems we wanna solve. There's so many [142:31] things that we wanna build. There's so many good things that we wanna do [142:35] that are now within our reach, and within the reach of my, my lifetime. [142:40] You just can't possibly not be romantic about that. You know what I'm saying? [142:46] - What an exciting time to be alive. Like, truly- [142:49] - How can- [142:49] - ... truly so. [142:50] - How can you not be romantic about, about that? The, the [142:54] fact that, that there is a, there, [142:58] it's a reasonable thing to expect the end of disease. It's [143:02] a reasonable thing to expect. It's a reasonable thing to expect [143:06] that pollution will be drastically reduced. [143:09] It's a reasonable thing to expect that traveling at the speed of light [143:15] is actually in our future. [143:17] And then, you know, not, not for long distances, but short distances. [143:21] You know, and people ask me how. Well, first of all, very soon, I'm gonna put a [143:25] humanoid on a spaceship, and it's gonna be, you know, my humanoid, [143:29] and, and we're gonna send it out as soon, you know, as soon as possible, and it's [143:33] gonna keep improving and enhancing along the flight. And then when it's time, [143:39] all of the, all of my consciousness has already been, you know so much of [143:43] my life has been uploaded in the internet. Take all my inbox, take everything that I've [143:46] done, everything I've said. You know, it's been collected and becoming my AI. [143:51] And I'm just, you know, when the time comes, you know, we'll just send that at [143:55] the speed of light, catch up with my robot. [144:00] - Oh, that's brilliant. I mean, but for me, that's sorta application-focused. [144:04] But also, for me, the curiosity- ... Maxing perspective, [144:09] I just, all of those mysteries. There's so much- [144:12] ... fascinating scientific questions there. [144:14] - Understanding the biological machine is right around the corner. It's, it's [144:18] not 10 years. It's five years probably. [144:20] - And then your biological machine, the human mind and cracking physics, [144:24] theoretical physics open. It's so exciting. [144:26] - Explaining consciousness, that one would be awesome. [144:29] - And it's all within our reach. [144:31] Jensen, thank you so much for everything you've done over the years. Thank you for everything you're doing [144:35] for the world. Thank you for being who you are. I can tell [144:39] you're a great human being, and I wish you incredible success [144:45] this year. I can't wait. As a fan, I can't wait to see what you do next, and hopefully [144:49] I'll see you in Taiwan and thank you so much for talking today. [144:52] - Thank you, Lex. I had a great time. And also, if I could just say one more thing. [144:56] - Yes. [144:57] - And thank you for all the interviews that you do, the depth, [145:01] the respect that you go through with and the research that you [145:05] do to reveal, you know, for all of us, [145:09] The amazing people that you've interviewed over the years. I've [145:13] enjoyed I've enjoyed them immensely. And as an [145:16] innovator, to have created this long form, [145:21] unbelievable, and yet, you know, it's just captivating. So [145:24] anyways, thank you for everything you do. [145:25] - It means the world. Thank you, Jensen. [145:27] - Thank you, Lex. [145:29] - Thank you for listening to this conversation with Jensen Huang. To support this [145:33] podcast, please check out our sponsors in the description, where you can also [145:36] find links to contact me, ask questions, give feedback, and so on. [145:42] And now, let me leave you with some words from Alan Kay. [145:46] "The best way to predict the future is to invent it." [145:51] Thank you for listening, and hope to see you next time.