Andreessen Horowitz Launches Machine Age Fund as AI Bottleneck Shifts South of the Model

Andreessen Horowitz推出机器时代基金，AI瓶颈转向模型底层 在扩展模型智能方面，资本投入如今已直接替代开发时间与人力工程，推动风投资金转向能源、物理硬件及高压直流数据中心基础设施。 讨论小组详细分析了AI开发如何从单纯的软件工程挑战演变为严峻的物理资源瓶颈，进而推动超大规模服务商资本支出激增，并导致严重的供应链短缺持续至2028年。他们探讨了推理模型与自主智能体如何促使Token消耗呈指数级增长，迫使传统数据中心全面向高压直流供电与液冷技术转型。此外，他们还讨论了吉瓦级数据中心的庞大能源需求、硬件初创公司在与Nvidia等现有巨头竞争共存中的战略市场机遇，以及为何复杂的硬件创新需要经验丰富的系统工程创始人。 a16z 宣布成立 Machine Age Fund 专注布局 AI 基础设施，指出技术发展的核心瓶颈已从基础模型算法全面转移到底层的芯片、内存、网络、电力供应及数据中心等物理层。 (0:50) 算力基础设施需求呈前所未有的爆发式增长，超大规模云服务商（Hyperscalers）年度资本支出预计将达 1 万亿美元，核心供应链关键零部件已全数被预订至 2028 年。 (4:00) 顶尖创始人团队中进军深科技与系统硬件的比例从以往的 3%–5% 激增至 20%–30%，AI 系统构建已从受布鲁克斯法则制约的传统软件工程，转变为资本可直接换取时间与算力性能的资源分配竞赛。 (16:25) 数据中心正面临严峻的物理与劳动力极限，单机架功耗从 5–10 kW 跃升至 100–150 kW，迫使系统转向 800 伏高压直流（DC）供电与液冷方案，而目前全美仅有 2% 的电工获得高压直流施工认证。 (28:08)

Andreessen Horowitz partners Ben Horowitz, Martin Casado, and Raghu Raghuram argue that the primary constraint in artificial intelligence has shifted from software algorithms to physical infrastructure, prompting a major reallocation of talent and capital. Announcing the firm's Machine Age Fund, they contend that exponential token demand is transforming software development into a capital- and resource-allocation challenge governed by hard physical limits in power, cooling, and silicon. Severe Physical Bottlenecks South of the Model (8:00) Raghuram and Horowitz note that while the semiconductor industry historically expands at 20% to 30% annually, AI compute token demand is growing at nearly 1,000%, triggering severe compounding shortages across silicon, high-bandwidth memory, and liquid cooling. Hardware lead times remain steep: developing new chip architectures takes three to four years, while building and powering a new data center takes four to five years. Memory suppliers report that existing backorders alone would consume three full years of total manufacturing capacity. Capital as a Direct Substitute for Engineering (16:25) Casado and Horowitz observe that AI development has diverged from Brooks's Law, which held that adding personnel cannot compress software engineering schedules. Instead, massive capital deployments directly into compute clusters allow well-funded entities to close multi-year competitive gaps in months. Hyperscaler capital expenditures reflect this shift, with spending projected to rise from roughly $700 billion this year to $1 trillion collectively next year. Every GPU produced is reportedly pre-sold and deployed immediately, with key supply chain components committed through 2028. Token consumption is multiplying further as systems shift from simple query-and-response interfaces to autonomous agents, long-thinking models, and agentic workflows. Raghuram and Casado note that moving to multi-agent loops increases compute consumption from hundreds of tokens per query to hundreds of thousands per task, establishing an autocatalytic loop where AI is increasingly deployed to design and optimize its own underlying hardware and software layers. Grid Shortages and Gigawatt Data Center Realities Power infrastructure has become an acute operational barrier. By 2028, newly built data centers are projected to require approximately 44 gigawatts of incremental power against only 25 gigawatts of expected electrical grid additions. Casado contextualizes a single gigawatt as equal to the power required for approximately 50,000 homes, noting that entire cities like Flagstaff, Arizona, consume less than one gigawatt in total. Raghuram points out that construction labor, transformer shortages, turbine backlogs, and regulatory permitting cycles prevent utilities from rapidly scaling to meet demand. At the rack level, power density is jumping from legacy 5–10 kW levels to 100–150 kW per rack. This transition renders traditional alternating current (AC) and air cooling obsolete, forcing facilities to implement 800-volt direct current (DC) architectures, reinforced concrete to bear immense rack weight, and sound-proofing walls. Horowitz highlights that only 2% of US electricians are certified to work on high-voltage DC power, while domestic permitting bottlenecks are already prompting some US startups to locate infrastructure abroad in Mexico and Australia. Niche Specialization and the Custom Hardware Opening Addressing whether incumbents like Nvidia will capture all market value, the panel argues that the massive scale of the industry creates ample room for startups. Drawing on an analogy from Alex Rampell and Dan Rose, Horowitz explains that incumbents must focus on multi-billion-dollar "gold bricks," leaving specialized high-margin "silver bricks" for challengers. Casado compares the sector to early automotive manufacturing, where Henry Ford's attempt at total vertical integration at the Rouge River plant eventually gave way to a fragmented ecosystem of specialized component suppliers. Economic incentives also favor custom Application-Specific Integrated Circuits (ASICs). Casado estimates that training a frontier model costs between $3 billion and $5 billion, requiring up to $10 billion in inference revenue to justify the outlay. In this regime, achieving a 20% efficiency gain in inference yields roughly $2 billion in savings—an economic return large enough to justify fabricating bespoke silicon tailored exclusively to execute a single, fixed-weight model architecture. The Return of Experienced Systems Builders (45:35) The capital intensity and operational friction of hardware have fundamentally shifted the founder profile away from early-career consumer software developers. Raghuram and Horowitz state that hardware founders must be seasoned "systems thinkers" capable of simultaneously managing silicon architecture, packaging, foundry supply chains, and complex manufacturing allocations. Hardware startups routinely require hundreds of millions of dollars upfront before shipping a functional commercial product. Because academic programs and corporate engineering defocused hardware for nearly two decades in favor of high-level software, deep technical talent remains scarce. However, the speakers note that resource-starved AI labs are now willing to sign commercial contracts with hardware startups well before functional silicon is delivered. Casado anticipates that this influx of deep-tech capital will produce a "SpaceX effect," training an alumni diaspora of systems engineers who will seed the next generation of industrial and computing companies. Horowitz frames the ultimate 5-to-10-year goal of the Machine Age Fund as establishing American abundance across chips, memory, and clean energy to secure technological leadership. However, whether the industry can overcome compounding physical lead times, regulatory grid bottlenecks, and severe electrical labor shortages quickly enough to match 1,000% annual token growth remains an open structural question.

a16z投资合伙人团队指出，人工智能的演进瓶颈已从模型算法全面下移至电力、散热、芯片与数据中心等“模型以南”的物理基础设施。随着软件开发从人月工程转向资本与底层资源的直接置换，顶级创业者与风险投资正大规模涌入重资产的深度技术与系统硬件领域。 资源堆叠取代工程律与推理消耗激增 (13:50) Martin Casado与Ben Horowitz分析认为，传统软件开发长期受制于布鲁克斯法则（人月神话），单纯增加工程人员往往伴随收益递减；但在当前AI周期中，直接投入资本换取算力集群能够有效跨越研发时间壁垒，诸如xAI与月之暗面等团队通过数以十亿美元计的算力投入即可快速跻身前列。软件行业的核心约束因而从工程师团队的吞吐能力，转变为对GPU、电力及资本等物理资源的分配能力。 Raghu Raghuram与Casado进一步指出，随着AI从单轮问答演进为具备深度推理、自我强化学习和长期自主运行的智能体，单次任务的Token消耗量已从数百个剧增至数十万个，全球算力Token需求正以近1000%的年化速度爆炸式增长。伴随大模型开始反向介入GPU内核编写等底层系统优化，这种自动催化循环导致推理与算力需求呈现无天然上限的指数级扩张。 超大规模支出与反转的芯片供应链 (4:00) 针对AI产业是否存在投机泡沫的争议，a16z团队列举市场数据指出其底层需求高度真实。超大规模云厂商基于前沿实验室及企业端的实际消耗，年度资本支出预计将从当年的约7000亿美元攀升至次年的近1万亿美元。不同于历史上芯片性能提高伴随价格下跌的常规曲线，当前AI芯片在二级市场的转售溢价最高可达原采购价的四倍。 供应链在全产业链路全面承压。核心硬件零部件的产能排期已被预订至2027至2028年，业内甚至频繁出现耗时数日竞标数千张GPU现货的局面。在斯坦福Hot Chips大会上，顶尖存储芯片厂商透露其现存积压订单需耗费整整三年满负荷产能方能交付完毕。Horowitz与Casado强调，与1990年代末缺乏终端流量消化的“暗光纤”泡沫不同，当前出厂的每一张GPU均处于完全售罄并即刻通电运转的状态。 数据中心逼近物理墙与电力重构 (28:08) 物理基础设施正面临全方位的断裂危机。单机架功耗从传统数据中心的5至10千瓦跃升至100至150千瓦，迫使机房全面舍弃交流电并转向800伏高压直流电（DC），散热方案亦由风冷全面转向环保液冷。由于超高密度机架重量过大，传统建筑结构面临垮塌风险，对特殊加固混凝土与隔音减震的要求成倍上升。与此同时，高压电工短缺成为严峻瓶颈，全美目前仅有2%的持证电工具备800伏直流作业资质，倒逼Meta等企业自建免费技术培训项目。 更深层的制约在于宏观电力供给的断层。Erik Torenberg与Casado指出，预计到2028年新建数据中心将产生约44吉瓦（GW）的额外电力需求，而同期电网新增供给预计仅为25吉瓦（1吉瓦相当于一座中型城市或5万户住宅的用电总量）。Raghuram说明，高压变压器与涡轮机等关键重工业装备面临数年的交付周期与严苛审批，电力短缺已迫使部分美国初创公司转向墨西哥或澳大利亚等海外区域部署计算集群。 定制架构机遇与资深创业者崛起 (24:47) 在硬件架构层面，Casado提出了针对前沿大模型的定制芯片（ASIC）经济学模型：训练单颗前沿模型的成本已达30亿至50亿美元，对应在全生命周期内需产生上百亿美元的推理收入；在此规模下，哪怕针对特定模型固定权重打造专用芯片仅带来20%的推理效率提升，就能节省20亿美元成本，足以覆盖定制硬件研发支出。尽管英伟达牢牢把控着90%的主流大宗市场（即“金砖”），但在万亿美元规模的生态中，针对专用系统、新型互联与内存层级的差异化方案（即“银砖”）为初创企业留出了巨大的商业空间。 这种高资本壁垒的产业特征深刻改变了创业团队的画像。Raghuram与Horowitz指出，硬件系统公司在交付首款商用产品前往往需要数亿美元的初始融资，创始人必须同时通晓芯片架构、晶圆制造与跨国供应链调度。这使得以往由二十余岁年轻程序员主导的纯软件轻资产模式不再适用，深耕高带宽内存、高压电气工程与复杂系统集成的资深产业老兵成为当前硬件创新的主力。 a16z设立“机器时代基金”（Machine Age Fund）旨在推动芯片、内存、清洁能源与数据中心的全栈充裕以确立长期技术优势，但也承认当前算力扩张受制于重型制造3至5年的长周期、电网审批与扩建滞后以及特种劳动力短缺等硬性物理约束；初创企业与云巨头能否突破物理制造与电网基建的极限，依然是AI产业实现下一阶段指数级扩张悬而未决的关键问题。

## Background 本期视频由著名风投机构 a16z 于 2026 年 8 月 29 日在 YouTube 平台发布。节目中，合伙人 Ben Horowitz、Martin Casado、Raghu Raghuram 与主持人 Erik Torenberg 宣布推出全新的“机器时代基金”（Machine Age Fund），并深入探讨了 AI 创新瓶颈如何从算法模型转向底层芯片、内存、互连网络、电力与数据中心等全套物理基础设施。 ## TL;DR 对话指出，随着推理模型与数字员工（Agent）爆发式消耗 Token，AI 的演进瓶颈已全面转移至“模型以南”的物理硬件与资源供给层。超大规模云厂商的资本开支迈向万亿美元，供应链排期普遍锁定至 2028 年，不仅芯片面临物理极限，电网容量、变压器短缺、800V 高压直流电以及机房承重等现实问题也变得极为尖锐。AI 颠覆了软件工程的《人月神话》，让资本与算力能够直接折现为智能，进而倒逼顶尖创业者从单纯的软件获客转向对全栈硬件与系统工程的深度重构。 ## Atmosphere & style 整场对话充满硬核的技术洞察与厚重的历史纵深感，嘉宾们摒弃了浮躁的软件叙事，以第一性原理冷静剖析从铜矿开采、持证高压电工短缺到加固混凝土建材的物理约束。合伙人之间互动默契且谈笑风生，既有对“爱迪生直流电时刻”和魔幻二手内存拍卖的生动调侃，又在谈及 44 吉瓦能源鸿沟与全球技术主导权时流露出强烈的危机感与使命感。整场对谈透露出一种资深系统老兵重掌主战场的宏大格局，让听众真切感受到一场重回物理世界与硬核工程的工业革命脉搏。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) a16z 宣布成立“机器时代基金”（Machine Age Fund），聚焦于 AI 模型底层从芯片、内存、网络一直延伸到底层电力与物理开采的全栈计算机科学基础设施重构。 - [02:40](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=160s) 顶尖创业团队攻坚硬核硬件的融资比例从过去的 3%~5% 激增至 20%~30% 以上，行业核心痛点从纯软件时代的“获客增长”转向 AI 时代的“计算效率与毛利率”。 - [04:23](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=263s) 全球超大规模云厂商（Hyperscalers）资本支出预计从 2024 年的 7000 亿美元攀升至 2025 年的 1 万亿美元，全产业链核心硬件产能已被全面预订至 2028 年。 - [34:04](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2044s) 到 2028 年新增数据中心的电力缺口高达 44 吉瓦（GW），而同期全美电网预计仅能增供约 25 吉瓦，高压变压器与涡轮机等物理供应链交期严重受阻。 - [47:33](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2853s) AI 基础设施的高难度与极度复杂的供应链，打破了“年轻软件极客”的神话，推动具备多系统统筹能力的资深行业老兵与软硬件全栈创始人成为核心主力。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) AI 瓶颈全面南移：前沿大模型算法本身的迭代已不再是唯一瓶颈，整个产业的制约正向“模型以南”的芯片、互连、机架液冷、电力乃至关键矿产等物理基础设施层转移。 - [04:23](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=263s) 供需失衡不是泡沫：不同于 1990 年代铺设“暗光纤”的投机热潮，当前算力硬件呈现未出厂即售罄、二手 GPU 甚至出现原价 4 倍拍卖的超强真实需求特征。 - [09:52](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=592s) 软硬件增速结构性脱节：传统硬件及数据中心基础设施维持 20%~30% 的年复合增长率已属不易，而下游 AI 模型与 Token 消耗需求呈现超 100% 乃至 1,000% 的爆发式增长。 - [13:39](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=819s) 算力缩放法则打破《人月神话》：传统软件工程中人月不可互换，而 AI 体系打破了纯工程调节阀，百亿美元的充沛资本与算力集群可以直接折现并跨越时间差距生成前沿智能。 - [18:05](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1085s) 智能体（Agent）驱动算力需求指数级裂变：从单次问答交互演进到类似数字员工的自主 Agent 与“计算机操作”，执行单项复杂任务所需的 Token 数量呈数量级递增。 - [25:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1506s) 模型静态权重催生 ASIC 经济账：前沿模型训练成本高达数十亿美元，巨大的推理回本诉求使得为特定静态模型架构定制专属专用芯片（ASIC）以节省 20% 成本具备了极高的商业合理性。 - [28:07](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1687s) 机架功率与供电发生质变：单机架功率密度从传统 5~10 kW 暴增至 100~150 kW，液冷成为绝对标配，并迫使数据中心全面转向极度危险的 800V 高压直流电（DC）供电体系。 - [31:43](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1903s) 劳动力与物理基建断层：全美仅约 2% 的持证电工具备高压直流电操作资质，加固钢筋混凝土等建材因机架超重承重要求而成本飙升，物理交付周期无法通过软件思维压缩。 - [36:41](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2201s) 本土审批受阻导致算力外流：由于美国本土面临电网接入审批与变压器紧缺等硬约束，部分创业团队正被迫将 GPU 算力集群转移部署至墨西哥、澳大利亚等海外国家。 - [38:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2286s) “机器时代”命名的深层逻辑：AI 本质是人类已有智慧沉淀的缓存与机器智能，决定下一次智能突破上限的不再是纯软件层微调，而是承载计算的底层物理机器的品质。 - [39:45](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2385s) 巨头无法赢家通吃：即便英伟达等万亿巨头占据核心基本盘，初创公司只要凭借第一性原理在能效比（Tokens/Watt）或单位成本产出（Tokens/$/Sec）上实现 10 倍跨越，撕开 5% 的细分市场即可诞生百亿美元企业。 - [42:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2520s) 产业链重回裂变周期：随着计算工作负载的多样化，AI 基础设施不会走向福特式的高度垂直整合，而是必然在计算、网络、供电与集群编排等环节形成分化的专业供应商生态。 - [49:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2946s) “SpaceX 效应”重振硬件人才：过去二十年科技界过度偏向纯软件，如今从重资产硬科技外溢出来的系统工程人才，正在与野心勃勃的年轻创始人结合，填补硬件人才代际断层。 - [52:32](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=3152s) 基础设施充沛度关乎技术主导权：五到十年内实现清洁电力、自研芯片与环保数据中心的计算资源大丰饶，是维系开放价值体系与全球科技领导地位的核心基石。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) 机器时代基金：从芯片到铜矿的全栈重构 a16z 宣布成立全新的“机器时代基金”（Machine Age Fund）。主持人 Erik Torenberg 引用了 Marc Andreessen 的评述开启对话：“这是我人生中见识过的最大技术革命，显然比互联网还要宏大。能与此相提并论的参照物是微处理器、蒸汽机、电力，或者甚至是轮子的发明。” Ben Horowitz 深入剖析了成立该基金的战略初衷。他指出，随着人类迎来史上最重要的全新技术，以往习以为常的计算基础设施已经完全无法适配当前的需求。这绝不仅仅是更换芯片或重写系统软件那么简单，而是一场涵盖电力生产与供应、全新物理散热架构乃至铜线材料替代方案的全方位变革。Ben 强调，过去解决问题靠纯粹的工程构架，而现在面对的则是彻底的“物理与资源限制”。 Raghu Raghuram 紧接着补充说明，过去在 IT 领域谈及基础设施，大家的视野局限在服务器、存储阵列和网络设备上；然而在 AI 驱动的新周期里，技术链路直接下探到了底层的物理开采——“一路延伸到开采铜的矿井深处”。更关键的转折点在于，过去三年模型自身能力飞速进化，模型算法本身已不再是核心瓶颈；相反，模型的快速进化让所有瓶颈都转移到了模型的底层——也就是 Raghu 所定义的“模型以南”（south of the model）的所有系统与物理层。 ```mermaid flowchart TD accTitle: AI基础设施层级瓶颈转移 accDescr: 展示从上层模型到底层物理资源的基础设施瓶颈转移路径 subgraph Upper ["模型与智能层"] M["前沿大模型 / 智能体"] end subgraph Mid ["传统IT基础设施"] S["计算芯片 (GPU/ASIC)"] N["网络互连与高带宽内存 (HBM)"] SW["系统底层软件"] end subgraph Deep ["深度物理基础设施 (模型以南)"] P["电力供给与变压器"] C["机架液冷与散热系统"] R["物理原材料 (铜矿 / 关键稀有材料)"] end Upper --> Mid Mid --> Deep ``` --- ### [02:40](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=160s) 创始人集体涌向硬件：从软件泡沫到物理效率 Martin Casado 分享了一组关于创业者行为变化的观察数据。风险投资机构向来习惯跟随最敏锐的创业者，而在过去一到两年里，顶尖创业团队攻坚硬核硬件问题的比例迎来了爆发式增长。过去向 a16z 寻求融资的顶级项目中，硬件类可能勉强占到 3% 至 5%（Ben 和 Raghu 甚至打趣说 5% 都算高估了，实际可能只有 3%）；而如今，这类交易的比例已经飙升到 20% 甚至 30% 以上。创业者群体显然比 VC 更加敏锐地捕捉到了创新的主战场。 谈及驱动这一巨大转向的宏观背景，Martin 分析认为，当前市场对 AI 的需求几乎是“无限”的，因而整个供应链都处于极度紧绷的状态。在传统软件时代，创业公司核心焦虑的是增长与获客（“我能不能把东西卖出去”）；但在 AI 时代，需求与增长天然存在，核心焦虑反倒变成了商业模式的毛利率与运行效率。如果缺乏极高的效率，算力成本将吞噬所有利润，而这一效率瓶颈本质上受制于硬件的物理极限。现有的计算系统最初根本不是为了承载 AI 工作负载而设计的，整个行业不得不从底层核心组件进行推倒重构。 - **硬核创业占比跃升**：顶级创始人攻坚硬件复杂课题的项目占比从过去的 3%~5% 激增至 20%~30% 以上。 - **商业关注点转移**：行业关注点从“获客增长”（Growth）转向“计算效率与毛利率”（Margin & Efficiency）。 - **工作负载脱节**：传统数据中心架构无法承受 AI 计算的极高密度与能耗，底层重构成必由之路。 --- ### [04:23](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=263s) 供需失衡的实证：不是泡沫，是万亿美元资本支出与算力哄抢 针对外界关于这是否又是一场炒作周期（Hype Cycle）的疑问，Raghu Raghuram 摆出了无可辩驳的产业端核心数据。首先，最具判断力的买家正在开出史上最惊人的采购订单：超大规模云厂商（Hyperscalers）的资本支出（CapEx）正在狂飙。2024 年大型云厂商的资本开支约为 7000 亿美元，而到 2025 年预计将达到创纪录的 1 万亿美元。由于云厂商直接对接前沿实验室（Frontier Labs）、AI 原生创业公司、跨国企业及全球市场，他们拥有一览无余的需求全景能见度，他们的疯狂追加 CapEx 彻底坐实了真实需求的强劲。 与此同时，应用端和前沿实验室的营收增长创下了科技行业历史上的最快纪录。更反常的经济信号来自硬件价格：在计算机产业历史上，芯片价格从来都是随着制程演进而持续下行，但现在的 GPU 和算力硬件价格曲线却出现了史无前例的逆势暴涨。算力供应紧张到甚至出现了为期数天、专门竞拍几千张二手 GPU 的极端拍卖会。此外，工作负载的消耗机制发生了质的飞跃——从前简单的聊天交互可能只需要上百个 Token，而如今复杂的智能体（Agents）执行任务动辄消耗成千上万个 Token，这种由任务复杂度和知识工作者全员采纳带来的双重放大效应，让算力需求呈现指数级裂变。 | 评估维度 | 传统科技周期 (如1990年代互联网热潮) | 当前 AI 机器时代周期 | |:--- |:--- |:--- | | **资本开支 (CapEx)** | 投机性铺设，如数以万里的“暗光纤”（Dark Fiber）埋地闲置 | 云厂商 CapEx 直奔 1 万亿美元，所有硬件“未出厂即售罄” | | **价格趋势** | 芯片/带宽硬件单位成本与售价规律性单调下降 | 出现反常溢价，二手 GPU 拍卖溢价甚至达到原价的 4 倍 | | **供需能见度** | 基础设施超前建设，缺乏足够的用户与真实高带宽应用 | 全球供应链已被预订至 2027~2028 年，需求消耗呈指数级激增 | | **物理瓶颈** | 主要是工程实现与用户采纳速度 | 遭遇严苛的电力、冷却、变压器短缺及政治审批阻力 | --- ### [07:07](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=427s) 产能锁定至 2028 年：物理资源枯竭与魔幻的二手内存 对话深入探讨了关键元器件供应已全面排期至 2027 年乃至 2028 年所带来的魔幻产业现实。Martin 和 Ben 回忆起 1990 年代末的互联网基础设施泡沫，指出当年的“暗光纤”铺设大量属于前置投机，缺乏下游真正高带宽应用（如当时还不成熟的在线视频）和海量用户支撑；而今天的 AI 算力供应链则是生产出来的每一颗 GPU 在下线之前就已被全部预购一空，市场上甚至有人以原售价的 4 倍转卖二手 GPU。 除了算力芯片本身，整个生态系统在电力、散热、场地等方面全线告急，并伴随着复杂的政治阻力和审批瓶颈。Martin 讲述了一个生动的行业真实花絮：某家大型上市公司的 CFO 过去出于保守态度一直极其抗拒上云，数据中心里囤积了大量的自建服务器；近期他们在做内部固定资产清点盘库时惊讶地发现，其旧服务器里闲置的高规格内存芯片因为全行业极度缺货而身价暴涨，光是把这批存量内存卖掉套现的资金，就足以全额支付他们将整个公司系统完整迁移到云端的所有费用。 Raghu 最后提到了正在斯坦福大学举办的半导体行业旗舰峰会 Hot Chips 上的重磅信息：全球领军的内存大厂在会上公开表示，仅凭市场上“现存”的既有订单需求，就需要该厂商开足全部产能满负荷生产整整三年才能勉强交付完毕——这甚至还没有算上未来不断追加的新增需求。算力、内存、机架电力、液冷冷却全方位告急，已成为全行业共同面对的硬核现实。 ### [09:52](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=592s) 算力缺口的本质：软硬件增速错配与物理极限 探讨的起点落在一个核心疑问上：当前严重的算力紧缺，究竟是因为人们低估了模型的能力，还是根本无法提前预见这种爆炸性需求？Ben Horowitz 与 Martin Casado 认为，即便拥有一台能预见未来的“完美神谕”，行业也根本不可能提前建出足够的产能。生成式 AI 爆发至今仅仅四到五年，而实体硬件的周期天然受到物理世界的约束：芯片设计与制造周期通常需要 3 到 4 年，数据中心的选址、平整土地、破土动工与交付需要 4 到 5 年，更不用说还要寻找甚至自建配套的发电设施与电网接入。 Raghu Raghuram 进一步指出了严重的“增速错配”：传统的机器学习与硬件基础设施产业如果能维持 20% 到 30% 的年复合增长率，就已经堪称优秀；然而它如今对接的却是一个以三位数（100%+）爆发式增长的 AI 软件产业。面对这一现象，投资界也在反思：为什么类似 Machine Age Fund 这样的基金没有在五到七年前设立？Martin 回顾了计算史上的几次重大范式转移： - **大型机走向客户端-服务器（Client-Server）**：涌现出一批独立的新硬件与系统公司。 - **互联网时代的崛起**：孕育了 Cisco（思科）和 Juniper。 - **超大规模云数据中心（Hyperscaler Data Centers）**：即便伴随着云巨头的垂直整合，也走出了 Arista 这样的网络新贵。 过去每个计算周期虽然都有新硬件公司冒头，但通常只局限于单一组件（例如出现一家新芯片公司或一家新交换机公司），技术变革相对局部。然而当下，不仅对智能的需求呈垂直向上飙升，而且现存的所有硬件体系架构原本都是为完全不同的计算时代设计的。无论在铜缆传输、功耗还是散热层面，现有技术均已逼近设计之初的物理极限。要想跟上 token 需求每年近 1,000% 的增速，必须全方位推进底层技术突破与架构重构。 ```mermaid flowchart TD accTitle: Hardware Cycle vs AI Software Demand Mismatch accDescr: 硬件基础设施周期与 AI 软件爆发式需求之间的结构性脱节关系图 subgraph HardwareCycle["硬件与物理基础设施周期（增速约 20%-30%）"] A["芯片研发周期<br/>(3-4 年)"] B["数据中心建设与供电<br/>(4-5 年)"] C["物理极限瓶颈<br/>(铜互连、散热与电力)"] A --> C B --> C end subgraph SoftwareDemand["AI 软件与智能需求（增速达三位数至 1000%）"] D["企业级与消费者端普及"] E["Token 消耗量呈指数级暴涨"] D --> E end C -.->|"严重产能供给脱节"| E ``` --- ### [13:39](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=819s) 算力缩放法则打破《人月神话》：资本可直接折现为智能 当对话转向需求端时，讨论集中在为什么 Token 消耗会随着模型演进而呈指数级跃升：从聊天机器人（Chatbots）、推理模型（Reasoning），再到自主 Agent 及多 Agent 协作系统，每一个技术台阶都会让解决单项任务所需的 Token 数量成数量级倍增。正如嘉宾们打趣所言："没有人比 AI 更喜欢消耗 AI"（"Nobody likes to use AI more than AI"）。无论是强化学习（RL）、思维链（Chain of Thought, CoT）还是长程运行的 Agent，其扩缩容（Scaling）的底层支柱全部建立在海量的推理（Inference）计算之上。 更为根本的范式转变在于研发规律的变化。Martin 指出，在过去的软件工程时代，做产品是一个典型的“工程问题”。如果你试图靠盲目增加工程师来加速交付，系统就会撞上弗雷德里克·布鲁克斯在《人月神话》（The Mythical Man-Month）中揭示的物理铁律——人月不可互换，“九个女人不能在一个月内生出一个孩子”。然而在当今的 AI 体系中，传统的工程调节阀失效了。现在的逻辑纯粹变成了“资源驱动”：向系统注入庞大的资本、GPU 和电力，系统便直接产出相应水准的智能。 Ben 对此进行了极其尖锐的对比：过去十年中，风险投资界普遍弥漫着一种怀疑情绪，认为给初创公司投太多钱会导致估值泡沫和执行力崩塌；如果落后竞争对手两年时间，指望靠招聘 1,000 名工程师去追赶往往只会把公司彻底搞垮。但在今天，这种追赶“真的起作用了”——只不过手段不是招聘十万名工程师，而是直接砸下 30 亿美元点亮一座顶级算力集群。无论是 xAI 的 Grok 从近乎空白迅速杀入第一梯队，还是 Kimi 等模型的崛起，都证明了充沛的资本与算力能直接跨越原有的时间先发优势。智能的上限不再受限于工程排期，而是受限于物理世界的供给构建能力。 --- ### [18:05](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1085s) 从编码到数字员工：不可阻挡的算力需求浪潮 随着 ChatGPT 坐拥近 10 亿周活跃用户、约 3,000 万开发者正在占据可观的算力资源，未来的计算需求又将如何演进？Raghu 拆解了 AI 需求的递进扩散图景： 1. **轻量与日常应用（Casual Apps）**：以初期的 ChatGPT 为代表，主要满足对话和日常咨询。 2. **专业编码领域（Professional Coding）**：AI 成为一线程序员的核心生产力工具，显著推高了复杂推理的计算频次。 3. **知识工作者辅助（Knowledge Workers）**：编码工具的成功模式正在向全球超过 10 亿的知识工作者迁移，驱动工作流自动化。 4. **后台深度自主 Agent（Back Office Agents）**：彻底接管企业后台操作，每一轮工作流流转都在后台消耗巨额的 Token。 Ben 与 Martin 进一步指出了当下最新的驱动力——以 GrokBot 及各类具备“计算机操作（Computer Use）”能力的 Agent。这类技术本质上就像是有一个虚拟的人类员工坐在电脑屏幕前敲击键盘、点击鼠标。Martin 分享了自己的亲身体验：就在刚过去的周末，他直接让这类 Agent 帮他登录各个账户，把拖延已久、懒得手动处理的一大堆服务绑定的信用卡信息挨个更新完毕。从单纯的文本对话，到代理人操纵软件，再到具身智能（Embodied AI）与物理机器人，每一波应用形态的扩展，都将对底层基础设施提出不可估量的算力需求。 ### [19:50](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1190s) 蒸汽机时刻：长达数十年的算力大扩张 当 AI 不仅局限于写代码，而是真正开始像人类一样操作电脑、取消繁琐订阅时，其本质就演变成了在计算机内部凭空创造出数以亿计的“数字知识工作者”。正如马克·安德森（Marc Andreessen）所指出的，当前的技术突破最恰当的类比是蒸汽机或电力的发明。人类刚刚发现了一种可以将算力转化为有效工作的新范式。目前的语言交互、代码生成以及初期的电脑操作（Computer Use）还只是起步阶段，未来三四十年里，只要一个领域具有清晰的奖励信号（Reward Signal），人类就会不断把算力砸向其中——无论是前沿科学、材料科学、生物工程还是创意产业。 在传统软件开发中，最大的瓶颈始终是“软件工程师的产能”。而现在，这一核心瓶颈正在被彻底击碎。尽管瓶颈会随之向系统的其他环节转移，并在不同层面催生出新的复杂性，但整个行业正处于一场长周期算力革命的起点。这种对算力的巨大渴求，不仅不会在短时间内消退，反而将持续数十年之久。 ### [20:55](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1255s) 从搜索框到独立员工：AI Agent 的演化与管理范式 科技行业在理解“AI 如何介入人类生活”这一命题上经历了三个阶段的认知跃升。最早阶段，行业只是机械地把 AI 当作功能插件（Feature）或搜索框，用户通过聊天窗口与模型一来一回地对话。随后，市场上出现了类似“个人代理”的形态，它作为使用者的能力延伸，共享用户的密码与 API 密钥，替用户执行个人操作。而真正带来范式质变的，是类似 GrokBot 的新突破——它不再附属于任何单个自然人，而是一个拥有独立身份的“实体员工”（Employee）。这个虚拟员工拥有自己专属的电脑、独立的浏览器，凭借顶尖的模型推理能力，能够像人类雇员一样独立处理高阶任务，比如自主通读邮箱、梳理优先级并在采取行动前主动找人类确认。 ```mermaid flowchart TD accTitle: AI 交互范式的三代跃迁 accDescr: 展示 AI 从早期的搜索对话工具，演化为共享凭证的个人副驾，最终成为拥有独立系统环境的数字员工。 subgraph G1["第一代：功能组件"] A["对话 / 搜索条"] --> B["单轮或多轮问答"] end subgraph G2["第二代：能力延伸"] C["个人副驾驶 Copilot"] --> D["共享个人账户与密码"] end subgraph G3["第三代：独立数字员工"] E["虚拟员工 GrokBot"] --> F["独立电脑与浏览器环境"] F --> G["自主拆解并执行高阶复杂任务"] end G1 --> G2 G2 --> G3 ``` 本·霍洛维茨（Ben Horowitz）指出，面对这批崭新的数字雇员，企业面临着漫长的组织适应与文化磨合。过去人类花了几十年摸索如何管理普通员工，如今管理 Agent 同样存在陡峭的学习曲线：它们可能会在漫无目的的任务中疯狂燃烧 Token、耗费巨额资金却一事无成；它们同样会遗忘上下文、胡编乱造，甚至引发严重的系统安全隐患。企业当下的目标绝非“借由 Agent 彻底自动化所有流程并把人类开除”，而是探索如何让人类员工与数字员工协同，在防止智能体失控搞砸业务的前提下，让人类获得超越极限的“超人能力”。事实证明，将智能体真正视作“独立自然人”来管理和融入团队，是目前在组织内部推动 Agent 落地最稳健、最持久的方式。 ### [25:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1506s) 逆向推演与底层重构：为模型量身定制专用芯片（ASIC） 当前的数据中心、芯片微架构、系统级软件以及配套机房基础设施，最初都不是为了大模型而设计的。若要设计纯粹面向 AI 的基础设施，必须从最基础的计算本质进行端到端重构。以推理引擎为例，其底层核心主要是密集的矩阵乘法（Matrix Multiplications）计算与海量内存吞吐，随着新 Token 的不断生成，系统对显存层级、芯片内互连、跨芯片通信乃至跨数据中心传输的带宽与能耗都提出了严苛的约束。当下的硬件初创公司创始人正沿着计算、内存、连接、功耗到散热的每一道断裂点进行逐层拆解与重新设计。 从商业回报的角度审视，算力硬件正在迎来前所未有的范式重塑： - **前沿模型训练成本**：目前训练一个前瞻性前沿模型（Frontier Model）大约需要投入 30 亿至 50 亿美元。 - **推理回本门槛**：为了让这笔高昂的训练资本支出成立，该模型在推理阶段至少需要赚回两倍的资金，即达到 100 亿美元的收入规模。 - **专用芯片（ASIC）的经济账**：在 100 亿美元的推理规模下，如果能通过针对特定模型架构的定制化优化提升 20% 的运行效率，就能直接省下 20 亿美元。而这 20 亿美元的节省空间，已足够覆盖专门为该模型定制一款专用 ASIC 芯片的完整研发成本。 - **静态权重与专用化**：不同于状态极其动态、逻辑多变的传统软件，大模型的权重是完全静态固定的。历史上从未出现过单一价值高达 50 亿美元的纯数字产物，这一特质正在对底层硬件提出极具针对性的极限制程与定制化诉求。 ### [28:07](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1687s) 机架功率暴增与直流电供电的“爱迪生时刻” 随着计算密度的激增，数据中心物理层面的物理定律正在面临极限考验。单个机架（Rack）的功耗需求正从传统数据中心的 5 至 10 千瓦（kW）一路飙升至 100 到 150 千瓦，计算密度激增了约 70 倍。传统的风冷散热已经彻底无法支撑这种散热负荷，液冷（Liquid Cooling）成为了不可逆转的刚性标配。 更为极端的变化发生在电力输送形式上。在单机架高达 100 到 150 千瓦的功率密度下，传统的交流电（AC）供电体系已经完全失效，数据中心必须全面转向高压直流电（DC）供电。直流电不仅自身需要复杂的专属冷却方案，而且高压直流电在物理操作上极其危险。本·霍洛维茨风趣地调侃了这一历史讽刺：在当年的“电流之战”中，托马斯·爱迪生为了推销自己的直流电，大肆宣传交流电有多危险，甚至当众用交流电电死马匹来做恐吓演示；然而历史兜兜转转，如今真正具备极限输电能力、却又极度危险且需要小心驾驭的，反而是这种极高功率的直流电网络。基础设施层面的深刻重构，正在从底层电力与散热开始全面铺开。 ### [29:23](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1763s) 物理现实的重击：液冷、超重机架与钢筋混凝土 谈及数据中心的底层架构，业界早已完成了从风冷到液冷的范式转换，当前任何最前沿的数据中心采用液冷都已是板上钉钉的事实。然而，仅仅拥有液冷远远不够，在当前的政治与监管环境下，液冷必须做到完全环保，直流供电（DC power）也必须回馈社会电网而非单纯消耗。过去有大约 10% 的“不良”数据中心浪费了大量水资源——虽然网上常有人拿它们与加州的开心果或杏仁种植用水量作对比，但数据中心在水效上确实有巨大提升空间；更不用说那些只进不出、宛如“电网寄生虫”的设施。这种粗放建设模式已经走过单向门，必须彻底终结，这就对工程设计提出了极高的全新要求。 除了水电平衡，AI 机架极高的功率密度直接引发了连锁物理反应：地板承重结构必须彻底推倒重来，以支撑成倍剧增的超重机柜；运行时巨大的噪音污染也迫使建筑必须配备更厚的墙体，否则周围居民和州政府绝不可能批准。当下一代架构（如代号 Feynman 的算力集群）落地时，现有数据中心绝大部分都将因无法适配而彻底报废。具有讽刺意味的是，大家都在盯紧内存价格的飞涨，却没注意到数据中心造价中飙升最快的成本之一其实是“钢筋混凝土”（reinforced concrete）。 ```mermaid flowchart TD accTitle: 数据中心物理与电力瓶颈演进 accDescr: 展示从AI计算密度提升到建筑结构、高压电工以及电网建设等维度的连锁制约瓶颈。 subgraph Hardware["机架与设施改造"] A["高密度AI机架/液冷系统"] --> B["地面高承重需求 (钢筋混凝土)"] A --> C["降噪隔音厚墙"] A --> D["800V 高压直流机架"] end subgraph Workforce["劳动力缺口"] D --> E["高压直流认证电工不足<br/>(全美仅约2%电工持证)"] E --> F["Meta等巨头启动免费岗位培训"] end subgraph Grid["电网与建设制约"] G["44 GW 激增电力需求"] --> H["电网仅能增供 25 GW"] H --> I["变压器、燃气轮机短缺"] I --> J["环评审批与政策壁垒"] end ``` - **高密度机架的建筑压力**：极重的高密机柜对地基与楼板承重提出了极端要求，直接拉动加固钢筋混凝土建材价格暴涨。 - **环境噪音控制**：数据中心散热与运行轰鸣声巨大，缺乏超厚隔音墙将直接通不过地方环保审批。 --- ### [31:43](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1903s) 800V高压危险与全美仅2%的持证电工缺口 当新一代数据中心开始将 800V 高压直流电直接输送至机架时，作业环境的危险性呈指数级上升。眼下最严峻的瓶颈之一在于全美严重缺乏能处理机房内 800V 电力系统的高素质电气承包商。在全美持证电工群体中，真正获得高压直流电（DC power）操作认证的比例仅有可怜的 2% 左右。为了自救，Meta 等云厂商甚至设立了专属的培训项目与免费“就业培训班”，从零培训高压电工。大众总在焦虑“AI 正在抢走所有人的工作”，现实却是 AI 基础设施正在催生大量高薪的专业电工需求。 与劳动力短缺并行的另一个趋势是数据中心自动化。各大云巨头正在紧锣密鼓地测试实体机器人，试图让机器人来接管机房内的硬件组装、服务器上下架与物理巡检。随着 AI 模型的演进，数据中心内部的自动化与机器人渗透率将加速攀升。 - **高压电工人才断层**：800V 直流电直入机架，全美仅约 2% 的电工掌握资质，催生科技巨头自建培训学校。 - **机房维护自动化**：面对高危作业与熟练工短缺，各大 Hyperscaler 均在加速引入组装与搬运机器人。 --- ### [32:53](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=1973s) Machine Age Fund 的定位：立足计算机科学硬核底层 Martin Casado 明确界定了新设立的 Machine Age Fund 的投资边界：该基金全力聚焦于“计算机科学基础设施”（computer science infrastructure），即承载 AI 模型运行的一切底层技术栈——从芯片、算力互联网络、存储，一直向下延伸至电力管理与输送。 当谈及该基金与 a16z 旗下 American Dynamism 基金在机器人领域的区分时，Martin 解释称，AI 最大的范式突破在于赋予了计算机感知和操作物理世界的能力（看、听、说）。这种能力催生出全新的承载平台，俗称的“边缘设备”（Edge Device）定义太宽泛，它可以是手机、笔记本或 CDN，也可以是具身移动设备。作为底层基础设施投资人，该基金不碰重监管或垂直行业应用，但凡是能够作为新型计算机科学平台、推动 AI 能力向物理世界延伸的硬核软硬件，都是其重点布局的方向。 - **核心投资标的**：芯片、网络互联、高通量存储及机房电能调度等一切支撑算力运转的计算机科学基础。 - **排除范围**：回避高度管制的重工业及行业垂直方案，专注赋能 AI 感知物理世界的通用软硬件计算平台。 --- ### [34:04](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2044s) 2028年44吉瓦缺口：为什么扩建不是写代码 讨论转向宏观能源鸿沟：预计到 2028 年，新增数据中心的电力缺口高达 44 吉瓦（GW），而同期全美电网预计只能额外提供约 25 吉瓦。面对“1 吉瓦到底意味着什么”的追问，Martin 生动地解释道，1 吉瓦足以覆盖大片足球场大小的设施，相当于 5 万户家庭的日常用电，这几乎等同于他从小长大的亚利桑那州弗拉格斯塔夫（Flagstaff，常住人口约 4 到 6 万）整座城市的总能耗；也就是说，1 吉瓦电量足够为整座城市全天候提供照明与空调制冷。 面对“公用事业公司和云巨头为何不能加快建设”的疑问，嘉宾们给出了冷峻的现实剖析： 1. **物理审批与法律门槛**：新建数据中心需要审批许可（permits），接入电网需要漫长而惨烈的各方竞价与监管博弈。 2. **硬件供应链断档**：自建发电面临全行业的硬件短缺，高压变压器（transformers）与燃气轮机（turbines）的交付周期长达数年。 3. **不可压缩的物理交付周期**：这不是敲敲代码、让工程师通宵加班赶工就能解决的软件问题，物理世界的供应链周期极难被压缩。在电力与场地需求每年暴涨 10 倍的背景下，算力供给的增长速度被物理定律死死卡住。 - **44 GW vs 25 GW**：2028 年新增数据中心用电缺口达 44 吉瓦，电网预期供给仅有 25 吉瓦，缺口难以弥合。 - **1 吉瓦（GW）具象化概念**：相当于约 50,000 户家庭或一座 4-6 万人中型城市的整座城市日常总用电量。 - **物理硬约束**：变压器、涡轮机等关键电网设备出现全球性交期延宕，无法靠软件式的“周末加班突击”来缩短交付周期。 --- ### [36:41](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2201s) 出海算力荒与共生电网模式 由于本土审批繁琐与电网准入受阻，算力饥渴正迫使许多初创公司将 GPU 集群部署在墨西哥或澳大利亚等海外国家。这种在美国本土变相“封杀”数据中心的阻滞机制，反而在无意中将宝贵的基建就业与长期经济红利拱手送给其他国家。 针对数据中心与社区的对立，健康的演进方向应当是建立一套“反哺标准”：合规的数据中心必须回馈社区，做到没有噪音污染、不争夺民用水资源，并为当地创造就业。实际上，部分现有数据中心已经实现了这一点：它们通过自建发电设施，白天全负荷运转时将多余电量反哺给州公共电网，到了夜间全城用电低谷、电厂产生富余峰值电量时再借电运行。由于城市白昼用电激增、夜间骤降，而数据中心的负荷昼夜平稳，二者完全可以形成“能源共生”（symbiotic relationship），甚至让所在区域的居民电价逐年下降。 - **GPU集群外流**：由于美国本土建设受阻，算力项目不得不转移至墨西哥、澳大利亚等国落地。 - **昼夜电力互补模型**：数据中心利用日间自主发电反哺城市、夜间吸收电网冗余电量，形成削峰填谷的互利生态。 --- ### [38:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2286s) 为什么叫“机器时代”（Machine Age）？ 在讨论基金命名时，团队一致认同“人工智能”（Artificial Intelligence）这个词从根本上就带偏了方向，更准确的称谓应当是“机器智能”（Machine Intelligence）。原因在于： - **AI 并非人类思维模式的重现**：它本质上是人类过往思想沉淀的数据缓存（cache）。目前人类根本无法做到在毫无先验知识的情况下，把一个 AI 放进物理世界就能让它从零重构出一套完整语言；我们所做的，是让它在人类已有的一切认知成果上进行消化学习并输出成果。 - **脱离 70 年的历史包袱**：“AI”一词在计算机科学界沿用了 70 年，混杂了太多科幻小说的浪漫想象和尼克·博斯特罗姆（Nick Bostrom）等哲学家的末日叙事，带有沉重的认知包袱。 - **软件吞噬世界之后的终极反讽**：曾经高呼“软件吞噬世界”的投资机构，如今真切见证了资本不论投入多少亿，最终都被底层的物理机器死死限制住。 下一次智能维度的根本性突破，不再取决于上层纯软件模型的微调，而是完全依赖于承载计算的“底层机器的品质”。将这个时代命名为“机器时代”，正是对硬件在这一波革命中压倒性决定作用的最强致敬。 > “曾经宣扬‘软件正在吞噬世界’的人们，如今真真切切地走到了这样一个境地：你把巨额资金砸进去，最后卡死你的却是最底层的物理机器。” ### [39:45](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2385s) 巨头阴影下的破局：为什么大厂无法赢家通吃 面对在 AI 基础设施领域动辄数十亿美元的资本开支与巨额沉没成本，主持人提出了一个行业普遍关心的核心疑问：在英伟达（Nvidia）等万亿美元巨头盘踞的当下，新初创公司是否已经错过了以实质性规模突围的历史窗口？ Raghu Raghuram 明确指出，当前市场要延续指数级增长，就必须在关键指标上实现 10 倍（10x）的飞跃——无论是“每美元每秒生成的 Token 数”（tokens/second/$）、“每瓦特 Token 数”（tokens/watt），还是“每机架 Token 数”（tokens/rack）。这种级别的跨越无法单靠渐进式改良，必须依赖具有第一性原理思维的创始人带来颠覆性创新。 ```mermaid flowchart TD accTitle: 市场规模与初创空间演进 accDescr: 展现市场扩张后催生的利基市场与初创企业切入逻辑。 subgraph Incumbent["万亿巨头聚焦区"] A["多万亿美元核心市场 (Nvidia 等)"] --> B["占据 90% 绝对核心增长盘"] B --> C["专注金砖业务：难以顾及非核心细分需求"] end subgraph Startup["初创突破空间"] D["5% 边缘创新空间"] --> E["规模达数百亿美元的大型独角兽"] E --> F["第一性原理重构系统 (芯片/网络/存储/电力)"] end A -. 溢出与分化.-> D ``` Martin Casado 补充阐述了“市场法则”（law of markets）：即便现有芯片巨头的市值达到数万亿美元，只要初创企业能撕开其中 5% 的缝隙，就能诞生出规模极其庞大、甚至年营收惊人的独立企业。英伟达当前专注于驱动自身 90% 增长的核心业务，根本无暇兼顾边缘需求。Ben Horowitz 分享了一个生动的类比：当年合伙人 Alex Rampell 创立 TrialPay 并试图向 Facebook 推销服务时，时任企业发展负责人的 Dan Rose 就直言：“Alex，你听上去能捡到很多银砖，但我手里的金砖多到根本捡不完，我最不可能做的就是分心去看一块银砖。”这正是英伟达目前的真实写照。 - 关键效率衡量维度：Tokens/Sec/$（成本效率）、Tokens/Watt（能效比）、Tokens/Rack（机架密度）。 - 巨头防线以外的生存空间：哪怕只拿下万亿美元盘子的 5%，也足以支撑起一家巨型非上市独角兽。 - 历史规律重现：云计算早期行业也曾质疑“亚马逊 AWS 会做一切”，微机时代质疑“微软会做一切”，但最终都在边缘地带催生了庞大的创新生态。 --- ### [42:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2520s) 从福特工厂到多变工作负载：市场的垂直整合与裂变周期 在探讨大模型实验室与基础设施演变时，Ben Horowitz 提到了历史上的垂直整合极限。1913 年福特汽车在密歇根州建立著名的红河工厂（Rouge River Plant），试图实现绝对闭环：一端运进水、煤炭和橡胶树，另一端直接开出整车；亨利·福特甚至在亚马逊丛林建立名为“福特兰迪亚”（Fordlandia）的企业城镇来开采橡胶，引入全套美式文化与露天音乐台，最终却因强硬推行打卡作息引发强烈抵触而走向失败。 现代工业的发展规律证明：随着市场急剧扩张，产业链必然走向分化（fragmentation），形成多层级的供应商网络；唯有在增长停滞后，才会进入并购或被颠覆的整合周期。现在的 AI 基础设施同样正在经历剧烈裂变。不仅下游使用场景在爆炸式增长，即使是推理（Inference）环节，也从曾经单一简单的架构演化为极其错综复杂的异构形态。 此外，Ben 强调了商业逻辑层面的底层质变：过去纯纯软件模式下，只要商业模式运转起来，利润率天然就会非常丰厚；但在 AI 时代，底层的硬件级系统级优化直接决定了商业利润空间的上限，软硬件结合带来的杠杆效应达到了前所未有的高度。 --- ### [44:46](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2686s) Machine Age 投资图谱与全栈系统思维 谈及“机器时代基金”（Machine Age Fund）的投资布局，Raghu Raghuram 描绘了一张跨越软硬件的完整技术栈图景。在当前阶段，单纯做一块计算芯片已经无法成立，必须交付端到端的整机系统（full system）： - **计算芯片（Compute Chips）**：从单一通用加速器向系统化设计演进。 - **存储互联与网络（Memory & Networking）**：攻克内存墙与集群通信延迟瓶颈。 - **供电与电力电子（Power Chips & Power Delivery）**：优化芯片级和机架级的电能传输转换。 - **自动化运维与集群编排软件（System Software & Fleet Automation）**：向上承接超大规模硬件集群的管理与调度。 在这些细分领域中，每一个都有潜力诞生出独立的上市公司。然而，这类硬件系统型公司的投资与运营规律与传统 SaaS 软件迥然不同。它们的早期轮次规模惊人，往往在做出首款实体产品之前就需要数亿美元的巨额资本注入，面临极高的技术交付与执行风险。 更为关键的是创始人的画像演变：优秀的硬件初创创始人必须是“全系统创始人”（systems founders），既不能只是单点的学术研究员，也不能只是纯粹的计算机科学家。他们必须从架构定义的第一天起，就通盘考量芯片制造、晶圆代工产能、上下游供应链约束以及量产可行性。正如 Raghu 所评价的：“黄仁勋（Jensen Huang）就是这个领域的迈克尔·乔丹（Michael Jordan）。” --- ### [47:33](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2853s) 宏观气候的剧变与“资深老兵”创始人的回归 Martin Casado 重点分析了支撑这波基础设施创业浪潮的两大宏观环境变化： 1. **头部 AI 实验室的极度饥渴**：前沿实验室对算力和能效的需求迫切到了甚至愿意在初创公司硬件流片或交付之前，就直接签署采购合同与战略协议的程度。这与五年前初创硬件厂商连 Google 等巨头门槛都进不去的局面截然不同。 2. **资本充裕度大幅改善**：风险投资界对底层计算重构达成共识，后续巨额融资渠道畅通，为长周期、高资本强度的硬件创新提供了流动性保障。 围绕创始人年龄结构的变化，主持人 Erik 引用了 Stripe 创始人 Patrick Collison 的观察：硅谷似乎越来越少见到像年轻时的马克·扎克伯格（Mark Zuckerberg）或比尔·盖茨（Bill Gates）那样在大学宿舍就能创办千亿美元企业的 20 岁出头年轻人。 对此，Ben Horowitz 与 Raghu 认为这一现象非常合理：当创业命题转向供应链极度复杂、工艺制造容错率极低、技术链路盘根错节的深科技领域时，行业经验（Experience）便成为了无可替代的核心壁垒。无论是在内存领域深耕数十年的技术专家，还是像埃隆·马斯克（Elon Musk）或特拉维斯·卡兰尼克（Travis Kalanick）这样具备复杂现实系统统筹能力的创业者，都在印证这一代底层浪潮属于具备深厚积淀的行业老兵。 ### [49:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2946s) 硬件创新的门槛：软件极客、资深老兵与“SpaceX 效应” 探讨顶尖创始人的背景时，本·霍洛维茨（Ben Horowitz）指出，许多传奇创始人在年轻刚出道时做的都是纯软件公司。即便是最顶尖的天才，也需要经历早期创办公司、打磨技术的过程，积累了足够的经验之后，才能“毕业”并跨入那些变量极其繁杂、更为精密深奥的物理与硬件领域。创办一家公司本身就极具挑战性，如果你对产品了如指掌，这尚且是一场硬仗；但如果你并不完全精通产品，必须一边学习深奥的物理工程一边学习管理公司，对于初出茅庐的年轻创业者来说，这种学习曲线陡峭到近乎致命。霍洛维茨以新一代年轻创业者迈克尔（Michael Truell，Cursor 联合创始人）为例，他极其聪明且才华横溢，但他打造的是纯软件层面的 AI 工具；而像埃隆·马斯克（Elon Musk）或特拉维斯·卡兰尼克（Travis Kalanick）这样的人物，则是凭借数十年摸爬滚打的经验，才足以驾驭重度硬件和现实世界的复杂工程。“迈克尔十年后或许也能做到这一点，但在今天，直接上手硬件基础设施会无比艰难。” 马丁·卡萨多（Martin Casado）补充道，学术界和整个科技产业在过去二十年里几乎完全忽略了硬件领域，硬件从来都不是所谓的热门增长赛道。过去二十年的聚光灯和风口完全被应用软件、消费互联网和高层网络占据，导致大学和大型科技企业中具备底层芯片和系统实战经验的人才严重匮乏——“你根本不可能随便找个实习就能去造一颗芯片”。然而，这种脱节正在被打破。马斯克对这个时代最伟大的贡献，除了 SpaceX 本身之外，正是从 SpaceX 走出、正在重塑整个国防与工业体系的创业者军团。硬件与计算机科学领域即将迎来同样的变革浪潮，一批在顶尖 AI 基础设施企业锤炼过的年轻工程师将蜕变为新一代硬件创始人。拉古·拉古拉姆（Raghu Raghuram）对此表示认同，并指出他们投资的一家基础设施企业就是由两位二十出头的年轻人创办的，但只要走进其办公室，就会看到身旁坐满了经验老到的资深系统专家。创始人自己不一定非要是资历极深的老将，但必须具备吸引顶级老兵并与他们无缝协作的能力。 ```mermaid flowchart LR accTitle: Machine Age 时代创始团队的双轮驱动模型 accDescr: 展示年轻创始人的敏锐视野与资深硬件老兵工程经验在AI基础设施中的结合 subgraph CoreTeam ["创始团队画像"] direction TB YF["年轻领军者<br/>(野心、软件直觉、迭代速度)"] VE["资深工程老兵<br/>(芯片/系统/散热/电力落地经验)"] end subgraph Ecosystem ["人才供给与行业红利"] direction TB SPX["工业黄埔军校效应<br/>(SpaceX、现代硬科技外溢人才)"] HW["硬件复兴<br/>(打破20年软件偏执，重回物理系统)"] end CoreTeam --> Ecosystem Ecosystem --> Output["打造新一代 AI 机器时代基础设施"] ``` - **经验曲线与试错代价**：在纯软件时代，试错成本极低，代码可以瞬间部署；但在硬件和复杂物理系统中，一个封装错误或散热设计失误可能导致数百万美元的流片报废和长达半年的进度推迟，极度考验团队的工程底蕴。 - **20 年硬件断层**：自 2000 年代互联网爆发以来，最聪明的人才和顶尖高校毕业生大多涌入移动端、SaaS 和纯算法，底层硬件工程人才出现了一整代的断档。 - **SpaceX 效应的蔓延**：马斯克旗下企业培养了数以千计懂得如何在严苛物理限制下进行极限工程创新的工程师，如今这类人才正在外溢至 AI 基础设施、先进能源与计算硬件领域。 --- ### [51:42](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=3102s) Machine Age Fund 的底色：深植于合伙人基因的硬核基础设施 主持人埃里克·托伦伯格（Erik Torenberg）提出疑问：a16z 正在推出规模如此庞大的新基金（Machine Age Fund），却没有宣布招募任何新的普通合伙人（GPs），这是出于什么考量？霍洛维茨和卡萨多相视一笑，解释道，团队之所以不需要招募新人，是因为他们多年前就是做底层系统与硬件技术出身的。在很长一段时间里，由于市场风口完全偏向应用层软件，合伙人们过去的这些硬核工程经验一直处于“休眠与潜伏”状态，甚至在过去还要刻意克制自己对硬件的偏爱。 卡萨多打趣道，过去的合伙人会议上大家甚至需要被提醒“别去碰硬件”，因为硬件的研发周期太漫长、太痛苦。然而，对底层的迷恋早已融入整个团队的 DNA。纵观 a16z 过往的投资版图，早在市场将其定义为风口之前，他们就已经在极早期开出支票支持硬核科技企业，例如 SpaceX、安杜里尔（Anduril）、卫星通信公司 Astranis 以及自动驾驶公司 Waymo。如今成立专门的 Machine Age Fund，并非为了扩充团队的能力圈，而是将团队深处原本就具备的底层系统认知与当前不可逆转的物理瓶颈全面对接。 - **隐性能力的全面释放**：从早期网络虚拟化公司 Nicira（由马丁·卡萨多创办并被 VMware 收购）到拉古·拉古拉姆在 VMware 负责的大规模虚拟化与数据中心软件，核心团队对底层基础设施有着长达数十年的实操理解。 - **硬科技早期下注记录**：包括 SpaceX（可重复使用火箭）、Anduril（国防自主硬件）、Astranis（小型地球静止轨道通信卫星）和 Waymo（自动驾驶硬件集成）在内的案例，验证了其对重资产物理系统的理解力。 --- ### [52:32](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=3152s) 算力丰饶的未来：美国基础设施之争与技术领导权 面对托伦伯格关于“如果这只基金达成预期，五到十年后的世界会是什么样”的提问，霍洛维茨给出了兼具技术乐观主义与地缘政治清醒的回答。他展望道，最佳的结果是美国在基础设施之战中赢得最终胜利——届时全美将遍布大量超级环保、极致高效的数据中心，迎来芯片、内存与清洁电力的极大丰饶（abundance）。这种计算资源的丰裕，将彻底解放人工智能的推理与物理改造能力。 霍洛维茨动情地谈到了技术背后的底层价值观。他强调，美国之所以是一个特殊的地方，不仅在于它对本土公民的意义，更在于它对全世界任何渴望做出超越自我之贡献的人而言，是“世界上唯一一个可以让你空手而来、却能在此成就非凡事业的地方”。这种向所有人开放的机遇机制和开放价值观，完全依赖于技术上的绝对领先地位。霍洛维茨直言不讳地警告：“如果我们输掉了技术领导权，我们就会进入另一个时代；那个时代将由另一个国家主导，而他们对待技术、个人与自由的价值观，将完全不同。”在对这一愿景的共识与致谢声中，全场对机器时代基础设施大浪潮的讨论圆满收官。 - **五至十年核心愿景**：摆脱长达数年的供应链断供与容量枯竭，实现超高能效数据中心网络、自研芯片生态以及吉瓦级清洁电网供给的计算“大丰饶”。 - **技术领先即价值观领先**：AI 基础设施不仅是商业和资本的回报工具，更是全球技术规则和自由创新土壤的基石；技术制高点的旁落将不可避免地导致底层制度与价值观主导权的转移。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) 这是我人生中见识过的最大技术革命，显然比互联网还要宏大。能与此相提并论的参照物是微处理器、蒸汽机、电力，或者甚至是轮子的发明。 — Marc Andreessen > [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) 一路延伸到开采铜的矿井深处 — Raghu Raghuram > [00:00](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=0s) 模型以南 — Raghu Raghuram > [13:39](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=819s) 没有人比 AI 更喜欢消耗 AI — Ben Horowitz / Martin Casado > [38:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2286s) 曾经宣扬‘软件正在吞噬世界’的人们，如今真真切切地走到了这样一个境地：你把巨额资金砸进去，最后卡死你的却是最底层的物理机器。 — Ben Horowitz > [39:45](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2385s) Alex，你听上去能捡到很多银砖，但我手里的金砖多到根本捡不完，我最不可能做的就是分心去看一块银砖。 — Dan Rose > [44:46](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2686s) 黄仁勋（Jensen Huang）就是这个领域的迈克尔·乔丹（Michael Jordan）。 — Raghu Raghuram > [49:06](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=2946s) 你根本不可能随便找个实习就能去造一颗芯片 — Martin Casado > [52:32](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=3152s) 世界上唯一一个可以让你空手而来、却能在此成就非凡事业的地方 — Ben Horowitz > [52:32](https://www.youtube.com/watch?v=Zx1Ec8LWFeM&t=3152s) 如果我们输掉了技术领导权，我们就会进入另一个时代；那个时代将由另一个国家主导，而他们对待技术、个人与自由的价值观，将完全不同。 — Ben Horowitz ## Mentioned a16z, Machine Age Fund, Ben Horowitz, Martin Casado, Raghu Raghuram, Erik Torenberg, Marc Andreessen, Alex Rampell, Dan Rose, Jensen Huang, Patrick Collison, Michael Truell, Elon Musk, Travis Kalanick, Mark Zuckerberg, Bill Gates, Nick Bostrom, Thomas Edison, Henry Ford, Frederick Brooks, Nvidia, Meta, Google, SpaceX, Anduril, Astranis, Waymo, VMware, Nicira, TrialPay, Cursor, Cisco, Juniper, Arista, ChatGPT, Grok, GrokBot, Kimi, xAI, Hot Chips, 《人月神话》, Rouge River Plant, Fordlandia, American Dynamism Fund ## Video description Ben Horowitz, Martin Casado, Raghu Raghuram, and Erik Torenberg discuss the launch of a16z's new Machine Age Fund and the infrastructure buildout behind AI, from chips, memory, and networking to power, cooling, and data centers. Why a dedicated fund now? The group argues that the bottleneck in AI is increasingly shifting from the models themselves to everything beneath them. Hyperscaler CapEx is surging, critical components are booked years in advance, and each new generation of reasoning and agents requires dramatically more compute. They unpack why this cycle looks different from previous infrastructure booms and how AI is turning problems once constrained by engineering into problems that can increasingly be attacked with capital and compute. They also explore where the next generation of infrastructure companies could emerge, why founders are returning to hard technical problems across hardware and systems, and what it will take to rebuild the computing stack for the Machine Age. Timestamps: 00:00 - Intro 00:50 - Introducing the Machine Age Fund 02:00 - Why Founder Interest in Hardware Just 4x'd 04:00 - How Do We Know Demand Isn't a Hype Cycle? 07:00 - Sold Out to 2028: The Unprecedented Supply Crunch 10:00 - What's Actually Bottlenecked Right Now 14:00 - Tokens, Scaling & Why There's No Natural Regulator 19:00 - Agents as a New Kind of Employee: The GrokBot Moment 25:00 - What "AI-Designed" Infrastructure Actually Looks Like 28:00 - Rack Power, Liquid Cooling & the Data Center Redesign 34:00 - 44 Gigawatts by 2028: Why Building Faster Is So Hard 38:00 - Why "Machine Age" Is the Right Name 40:00 - Won't Incumbents Like Nvidia Take Everything? 48:00 - The Founder Profile: Why Hardware Needs Experience Resources: Read more about the Machine Age Fund : https://www.a16z.news/p/the-machine-age-fund Follow Ben Horowitz on X: https://x.com/bhorowitz Follow Raghu Raghuram on X: https://x.com/RaghuRaghuram Follow Martin Casado on X: https://x.com/martin_casado Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=Zx1Ec8LWFeM)

[0:00] Marc Andreessen: 这是我人生中见识过的最大技术革命，显然比互联网还要宏大。能与此相提并论的参照物是微处理器、蒸汽机、电力，或者甚至是轮子的发明。 [0:00] Raghu Raghuram: 一路延伸到开采铜的矿井深处 [0:00] Raghu Raghuram: 模型以南 [13:39] Ben Horowitz / Martin Casado: 没有人比 AI 更喜欢消耗 AI [38:06] Ben Horowitz: 曾经宣扬‘软件正在吞噬世界’的人们，如今真真切切地走到了这样一个境地：你把巨额资金砸进去，最后卡死你的却是最底层的物理机器。 [39:45] Dan Rose: Alex，你听上去能捡到很多银砖，但我手里的金砖多到根本捡不完，我最不可能做的就是分心去看一块银砖。 [44:46] Raghu Raghuram: 黄仁勋（Jensen Huang）就是这个领域的迈克尔·乔丹（Michael Jordan）。 [49:06] Martin Casado: 你根本不可能随便找个实习就能去造一颗芯片 [52:32] Ben Horowitz: 世界上唯一一个可以让你空手而来、却能在此成就非凡事业的地方 [52:32] Ben Horowitz: 如果我们输掉了技术领导权，我们就会进入另一个时代；那个时代将由另一个国家主导，而他们对待技术、个人与自由的价值观，将完全不同。

[0:00] Ben Horowitz: Not only do we need new chips, new system software, we need new ways of doing power, we need to replace copper... it's absolutely everything. [8:00] Martin Casado: I was talking to a CFO of a large public company who had historically been very resistant about going into the cloud... they did an inventory check and realized that the memory in their existing servers had increased so much in value, it could fund their entire migration to the cloud. [13:50] Martin Casado: It used to be when you built something, it was an engineering problem... here it feels like it really is a resource limitation. We're pouring a ton of money into systems, and then those systems are producing a result. Any problem that you have can be solved with enough infrastructure—basically GPUs, power, and money. [19:04] Ben Horowitz: We've spent many, many, many years figuring out how to work with our kind of regular human employees, and now we've got these other kinds of employees... I don't want to sit up here and say I've cracked the code... we're much more going like, 'Okay, how do we make all our humans superhuman without like wrecking the place because the bots got out of control?' [28:08] Ben Horowitz: AI is taking all the jobs... [but] AI is going to create a lot of new electricians. [38:06] Martin Casado: It's not how humans think necessarily... It is a cache of how humans thought—it's a collection of human thoughts. To date, we don't know how to take an AI with no knowledge, put it out in the world, and have it reconstruct language. We've built something that can learn off of everything we've already learned and use that productively. [39:45] Dan Rose: Alex, that's great. It sounds like you can collect a lot of silver bricks. But I have so many gold bricks I can't even pick them all up—so the last thing I'm doing is looking at a silver brick.

[0:00] We have a whole new technology. That's [0:02] the most important technology ever. And [0:04] you need a whole new infrastructure. [0:05] >> Normally when we talk about the [0:06] infrastructure world, we're talking [0:08] about the servers on the storage and the [0:09] network. Here it goes all the way down [0:11] to the mines of copper min. That's how [0:14] widespread this thing is going to be. [0:15] >> It used to be when you built something, [0:17] it was an engineering problem. And here [0:18] it feels like it really is a resource [0:20] limitation. So whether it's tokens or [0:22] not, we're pouring a ton of money into [0:24] systems and then those systems are [0:26] producing a result. And right now we're [0:28] bottlenecked on the systems ability to [0:30] actually match the resources we're [0:31] pouring into them. [0:32] >> The leading memory vendor said the [0:34] demand they have today will take them 3 [0:36] years of capacity to supply. [0:38] >> If this fund does what we think it will [0:39] do, how do we see the world in 5 to 10 [0:41] years? [0:42] >> America wins in the infrastructure and [0:44] that would be awesome. [0:49] >> Ben Martin Ragu, welcome. [0:52] >> Thank you. [0:52] >> All right. Thank you. [0:53] >> I want to start with a Mark quote to [0:55] introduce this new fund. This is the [0:57] biggest technological revolution of my [0:59] lifetime. This is clearly bigger than [1:00] the internet. The comps on this are the [1:02] microprocessor, the steam engine, and [1:04] electricity, or maybe the wheel. Guys, [1:07] the machine age fund. Please introduce [1:09] it. Ben, start us off. [1:12] >> Well, um, basically what's uh happened [1:16] is we have a whole new technology that's [1:18] the most important technology ever. And [1:21] what happens every time um there's a [1:24] dramatic new way of using all of the [1:29] things that we love infrastructure um [1:31] you need a whole new infrastructure and [1:34] never has it been more high impact as it [1:38] is on this one. So not only do we need [1:42] new chips, new system software, we need [1:45] new ways of doing power, we need to [1:48] replace copper. I mean like it's [1:50] absolutely everything. So it's a very [1:53] exciting time. So you know particularly [1:55] for the kind of hardware aspects of of [1:59] this new era um we needed a new [2:02] approach. [2:03] >> Yeah I would agree. I mean normally when [2:05] we at least in the computing when we [2:07] talk about the infrastructure world we [2:09] talking about the servers and the [2:12] storage in the network here it goes all [2:14] the way down to the mines copper mines. [2:16] That's how widespread this thing is [2:18] going to be. [2:19] Um that's number one and number two I [2:23] think what we have seen over the last 3 [2:25] years is the steady increase of the [2:27] capabilities of the models where the [2:29] model is no longer the bottleneck and in [2:32] fact using AI these models are getting [2:34] better faster and faster and faster. Now [2:36] the bottleneck is all what I call south [2:38] of the model and so that's why we need [2:40] to work on that. You know, the only [2:42] thing I'd add very quickly is like we [2:43] tend to follow founders and we've been [2:46] watching over the last couple of years [2:48] is the number of very strong teams going [2:50] after complex hardware problems has [2:52] increased. I don't know the actual [2:54] numbers but I was trying to estimate it [2:55] over the weekend. So I think we'd get [2:57] you maybe 5% of the deals from top [2:59] founders would come in would be hardwood [3:00] before. Now I would say north of 20% or [3:02] 30% right now. So like the founder [3:04] community which tends to be much smarter [3:05] than the VC community has identified [3:07] this as a very active area for [3:09] innovation and they're responding. I [3:10] think 5% is probably generous. [3:12] >> Yeah, it's very low. Is very low. Yeah. [3:14] 3%. Yeah. [3:15] >> And explain some of the macro conditions [3:17] that have led to this this trans this [3:19] change in terms of the surplus of [3:20] founders pursuing these idea. Like what [3:21] are they seeing that's that's enabled? [3:23] >> Well, I mean the obvious is like you [3:25] know the the demand for AI is basically [3:27] infinite and as a result of that every [3:30] part of the supply chain is under under [3:32] duress. I mean everything including like [3:34] materials used to make things like like [3:37] memory. Um it's also very interesting [3:39] whether there's something unique about [3:41] AI. Um which because the demand is [3:44] infinite and growth is infinite. Um uh [3:47] what you tend to worry about is the [3:48] margin of companies which is how [3:50] efficient it is. Like normally you worry [3:52] about growth like can I just you know [3:54] can I just get people to buy this stuff. [3:55] You don't have to worry about that here. [3:56] The question is is can you do this in a [3:57] way that's profitable? And a lot of the [4:00] um efficiencies are actually strictly a [4:02] physical limitation of hardware. And so [4:04] even the business model of the AI wave [4:06] is really putting a lot of stress on the [4:09] existing systems because they weren't [4:10] built for AI. They weren't built for [4:11] those workloads. And I think there's [4:12] just this you know this global [4:15] observation that we actually need to [4:17] change the core components to get that [4:19] efficiency to help drive the growth and [4:21] to drive the value of the businesses. [4:23] >> Yeah. And how did we know that demand is [4:26] actually outpacing supply here rather [4:27] than this being you know another hype [4:29] cycle? Well, I mean, [4:32] >> yeah, there are any number of cities [4:34] today. Um, firstly, it is that some of [4:38] the smartest judges of demand are [4:42] cutting huge purchase orders. I mean, if [4:44] you look at the hyperscalers, right? [4:46] >> Yeah. [4:47] >> Their capex spend has been exploding. [4:50] Next year supposedly, it's going to [4:52] reach a trillion dollars collectively [4:54] across the big hyperscalers. This year, [4:56] it's about 700 billion dollars, right? [5:00] And if you think about the hyperscalers [5:02] position in the industry, they see [5:04] demand from everywhere, right? They see [5:06] obviously the frontier labs um wanting [5:10] their compute, they see the AI native [5:12] companies, they see the enterprise, they [5:15] see the US geography, the international [5:17] geography. So if anybody has visibility, [5:20] it is down and they've been jacking up [5:23] their capex like it's never been seen [5:26] before, right? So that's a clear clear [5:28] sign. And secondly, if you look at the [5:30] companies that we see on a day-to-day [5:32] basis, they are all ripping. All the [5:34] application companies, the growth is [5:36] insane. The frontier labs, the growth is [5:39] insane. It's been documented. So I would [5:42] say on the demand side, the signals have [5:44] never been clearer that this is not a [5:47] hype. It's to top it all, all of it is [5:50] just and prices are going up. [5:53] >> Like we've never seen prices go like up [5:55] on chips. [5:56] >> You keep your prices went down. They [5:57] always go down. Yeah. It always goes [5:59] down. If you look at the price curve, it [6:02] it went like this and then went back [6:04] this. [6:04] >> Yeah. [6:05] >> And we know only like 5 10% of the [6:07] address for the market to stop today. [6:09] >> I mean, the the the supply, if you look, [6:11] if you look at the supply across the [6:12] board, it's basically all booked out to [6:14] 2028. [6:16] I mean, it's so bad we've actually seen [6:18] multi-day auctions for a few thousand [6:21] GPUs. [6:22] Um, you know, the other side of that, of [6:24] course, is demand. And as Ragu said, [6:26] we've seen the fastest growing companies [6:27] we've seen in the history of the [6:28] industry. [6:29] >> But also the unit of work that AI can [6:32] do, [6:34] the value of that unit of work keeps [6:36] increasing. But underneath the covers, [6:38] the number of tokens that are consumed [6:40] is going by orders of magnitude, right? [6:43] If it's one token for I mean 100 tokens [6:46] for chat or an agent, it's thousands of [6:49] tokens, right? So you got expansion of [6:51] both sides of demand. One is the unit of [6:53] work is becoming more and more [6:55] consumptive [6:56] >> of tokens and then secondly the number [6:57] of people therefore that are going to be [6:59] benefit it's not just the developers [7:01] it's going to be all knowledge workers [7:03] and then all of beyond that so that's [7:05] that's what we see [7:07] >> you said the key components in supply [7:10] are sold out to 2027 maybe in 2028 [7:12] >> what does it mean for an entire industry [7:14] to be sold out that that far like [7:17] >> I don't know this has ever happened [7:18] before do you guys recall I mean [7:20] remember in the Um, in in in the [7:23] internet days when we were doing massive [7:26] buildout, the majority that was actually [7:27] being put in the ground was speculative [7:29] and was dark. Remember the dark fiber? [7:31] And here basically every GPU that's [7:34] being created is already pre-sold. So [7:36] >> yeah, we weren't quite there. I mean [7:37] there was [7:41] there was a lack of bandwidth [7:46] like in the 989 time frame, but [7:51] there wasn't there there wasn't that [7:53] much real demand for it because there [7:54] just weren't that many people on the [7:55] internet. So like it was a two-sided [7:58] thing and the companies were all rushing [8:01] there and needed more bandwidth [8:02] theoretically, but there weren't [8:05] the users on the other side to consume [8:07] it necessarily. And then to really [8:10] consume a lot of bandwidth, you have to [8:12] do high bandwidth things like video, [8:14] which weren't really viable for, you [8:18] know, a number of reasons that that had [8:20] nothing to do with how much bandwidth [8:21] was in the data center. So it smelled [8:26] similar, but it wasn't this. This is [8:28] like we're flat out and people are [8:31] reselling GPUs for four times what they [8:35] bought them for and this kind of thing. [8:36] Like it's just not. And then, [8:39] >> you know, we're we're also out of power [8:42] and cooling. Uh and then on top of that, [8:46] it's really hard to build because [8:48] there's these incredible political [8:49] headwinds going into it. So, it's it's [8:52] really unprecedented in my career. uh [8:55] that we've had anything anything like [8:58] this. No, [8:59] >> I want to give you a quick anecdote. So, [9:01] this I was talking to a CFO of a large [9:03] company, large public company who had [9:04] historically been very resistant about [9:06] going into the cloud. So, they had a lot [9:07] of servers and they were doing an [9:09] inventory uh check and they realized [9:11] that the memory in their servers had [9:13] increased so much it could uh fund the [9:15] entire migration to the cloud. So, I [9:17] just feel like we're in a very unusual [9:20] situation. [9:21] >> Yeah, that's right. We're out of many [9:22] things. power, cooling, memory, GPUs, [9:26] like you name it, we're out of it. [9:28] >> Yeah. So, the flagship uh conference for [9:31] the industry is the one called hot chips [9:34] is going on in Stanford and the leading [9:38] memory vendor said the demand they have [9:42] today it will take them 3 years of [9:44] capacity to supply it. It's just today [9:49] it's not even future demand. [9:51] So in terms of being about everything [9:52] simultaneously, is it because people [9:54] just underestimated how good the models [9:56] would be, how how useful they would be, [9:59] they just couldn't have foreseen the [10:00] demand? [10:01] >> Well, I don't even think it's that. I [10:03] mean, this stuff came out of nowhere, [10:04] right? We we're only four years into [10:05] this. So even if we had a perfect [10:07] oracle, once it started working, I don't [10:09] think [10:10] >> we could have built the capacity. [10:11] >> We could have built the capacity. [10:12] There's no way. And we're talking about [10:15] like chip cycles, which tend to be three [10:16] to four years. We're talking about [10:18] breaking ground and building data [10:19] centers which is you know four to five [10:21] years we're talking [10:22] >> and connecting breaking down and [10:23] building them and having uh power [10:25] source. So like you either have to build [10:27] your own power or like usually both [10:30] you've got to build your own power and [10:32] have a power source which is not easy. [10:35] Yeah. the ML industry that's used if [10:38] it's growing at 20 30% it's a great [10:40] growth rate right and it's being [10:42] connected to an AI software industry [10:45] that's like triple digits is the base [10:47] you know [10:48] >> so you can see the disconnect right so [10:49] it's it's just wide bank [10:52] >> and so why didn't this fund exist you [10:54] know five years ago or seven years ago [10:56] or why was it not a great category to to [10:58] invest in in in the same way Brad [11:02] >> well I would say we're probably I'd like [11:05] to think for just in time, but you know, [11:07] we probably would have been uh well [11:09] suited to have it at least a couple [11:11] years ago. [11:12] >> I will say you could actually point on [11:14] on on basically every epoch to uh an [11:18] independent company that came up, right? [11:19] Clearly the move from the client from uh [11:21] in mainframe to the client server, we [11:22] saw a bunch of companies come up. Um the [11:24] move to the internet, this we got Cisco [11:26] and Juniper. uh even in the mega data [11:28] centers which by the way was largely [11:30] driven by the incumbent cloud providers [11:32] verticalizing you saw the uh arising of [11:34] Arista so there has been the ability to [11:37] invest in you know silicon and hardware [11:40] but it's been relatively minor because [11:42] the change has been relatively minor [11:43] like one chip company one switch company [11:46] where here everything is and so I I [11:48] agree with Ben we're probably you know [11:50] we probably could have started a little [11:52] bit earlier um but the amount of change [11:54] is so high now that it's just an obvious [11:56] thing [11:57] And the other thing is the demand for [12:00] intelligence is so vertical um with [12:06] really no end in sight. I mean cuz every [12:10] company that's adopted it is growing [12:12] very fast in its usage and then most [12:16] companies haven't adopted it to a high [12:18] degree and then consumers are just [12:22] getting started. [12:24] And so it's going to probably the demand [12:26] for tokens is probably going to grow [12:28] close to a,000% a year, which you cannot [12:32] grow supply that fast. Like like we're [12:36] not [12:37] >> like the amount of just work we're going [12:40] to have to do across the board to get to [12:43] the point where we can grow like infr at [12:47] that kind of rate is is pretty vast. So [12:50] I think there's a lot of investing [12:52] opportunity on the way. And by the way, [12:54] the other thing is like all the [12:55] architectures [12:57] of the hardware systems were built for a [13:00] whole different era of computing. And so [13:04] more than just we need more capacity, we [13:08] need capacity to build there there's [13:10] lots of opportunities to build different [13:12] kinds of infrastructure. Yeah, they're [13:15] they're all reaching their the physics [13:19] limits for what they were designed, [13:21] right? Like what Ben was talking about [13:24] copper and so on and so forth. And you [13:26] could go across every one of these [13:28] categories and you could find, okay, [13:30] this is the limit of this type of [13:31] technology. So now you got to get some [13:36] technical breakthroughs to get to the [13:38] next one. [13:39] >> Yeah. I want to dive deeper on the [13:40] demand side for a second. As we've moved [13:42] from chat bots to reasoning to agents to [13:45] to multi- aents, each step has [13:48] multiplied the number of tokens a single [13:50] task takes up by orders of increasing [13:53] orders. [13:53] >> Yeah, nobody likes to use AI more than [13:55] AI. [13:57] >> So, um why does that keep happening [14:00] instead of a leveling off? Do you just [14:01] see that happening you know indefinitely [14:04] just continuing to [14:05] >> uh well so there's a couple as that the [14:07] first one is is for sure right now if [14:09] you look at like the the way we're [14:12] achieving scaling the way we're doing it [14:14] is through a lot of inference so through [14:16] a lot of token right if you think about [14:18] what RL is you know it's it's a lot of [14:20] inference if you think about chain of [14:21] thought it's a lot of inference u think [14:24] longunning agents of course it's a lot [14:26] of inference and so that's just [14:27] basically been one of the approaches [14:29] that we've been using to um uh to [14:32] scaling um I think if you want to step [14:34] back and say kind of what is the macro [14:35] trend here it used to be when you built [14:38] something it was an engineering problem [14:39] and you throw a bunch of engineers at it [14:41] and that doesn't scale and that would [14:42] have a natural law of engineering [14:44] physics uh which is what the mythical [14:46] manmouth came from and here it feels [14:48] like it really is a resource limitation [14:50] so whether it's tokens or not we're [14:52] pouring a ton of money into systems and [14:54] then those systems are producing a [14:56] result and right now we're bottlenecked [14:58] on those position those systems ability [15:00] to actually uh match the resources we're [15:02] pouring into them. And so I think like [15:04] tokens right now is probably where we [15:05] are on the scaling curve, but we don't [15:07] have a natural regulator like [15:09] engineering like we did before. So I [15:10] think we should expect this to continue [15:12] and we have to build a supply to support [15:14] it. [15:14] >> Yeah. Like the simple way to think about [15:16] it is [15:18] any problem that you have can be solved [15:22] with enough [15:24] infrastructure [15:25] >> basically [15:26] >> GPUs and power and money. Uh and so [15:30] until we run out of problems, we're not [15:32] going to run out of demand. And that's [15:34] the that's the uh challenge. I [15:38] >> mean AI's answer to getting better and [15:41] better is to use more AI, right? [15:43] inference is one basic building block [15:46] that it keeps using over and over and [15:47] over again and so that's why these [15:49] tokens multiply at each step. [15:51] >> Yeah, even the autoc catalytic effect so [15:53] even the idea of using AI to create more [15:55] AI like creating a GPU kernel of course [15:57] is just using more AI um as part of the [16:00] process. So again, one way that we think [16:04] about it is in the past money would come [16:06] in, you have an engineering problem, we [16:09] know that it takes two years, normally [16:10] fails, you know, it's a national [16:12] governor and then you get the product on [16:13] the other end. This there's there's [16:15] there's nothing between the money going [16:17] in and then the hardware, you know, [16:19] creating intelligence. And so now we're [16:21] just limited by our ability to create [16:23] supply. It's a very very different [16:25] >> be more GPUs and our sol. [16:28] >> Yeah, that's right. [16:28] >> So that's the cycle. as long as you have [16:30] the money, the GPUs, and the data, you [16:32] know, for the foreseeable future, you'll [16:35] be able to scale these things. [16:36] >> And it's it's fascinating because, you [16:38] know, over the last decade, it feels [16:39] like there so many, you know, people and [16:41] the pervasive sentiment was there's too [16:43] much money going to startups. We're [16:45] overfunding these startups. There's too [16:46] much money in in in in venture capital. [16:49] Say more Ben about what that means [16:50] because there there used to be this um [16:52] sort of skepticism that the more money [16:54] you put into into the industry that you [16:56] know there would be bigger outcomes and [16:58] now you know we were saying at the [17:00] offsite that there's to some degree the [17:01] the market is as big as we collectively [17:04] contribute to it. [17:05] >> Yeah. So, this is this look, the one [17:07] thing we all knew with in startup world [17:11] is that if I have a two-year lead on you [17:15] and you try and catch me by hiring a [17:17] thousand engineers, you're going to [17:18] wreck your company. Like, that never [17:20] works. It's a mythical man month. Nine [17:22] women can't have a baby in a month. That [17:24] like that's it. Like, that never works. [17:28] Okay, now that works. [17:31] But it's not hiring a hundred thousand [17:32] engineers. It's taking $3 billion [17:36] and like lighting up a magnificent [17:40] cluster and then all of a sudden, you [17:42] know, whatever Grock can come out of [17:44] nowhere and like, oh, all of a sudden [17:46] it's real or or a Kimmy or or what have [17:48] you. [17:50] It's just like these leads um you can [17:53] throw money at the problem and you can [17:54] throw money at almost any problem and [17:57] that works. And so that is just [18:00] completely different than anything we've [18:03] ever lived through. So we're, by the [18:05] way, we're all psychologically adjusting [18:07] to this. The Chad GBT app has a billion [18:11] weekly activives. There's about 30 [18:13] million uh developers um who are using, [18:17] you know, relatively a big portion of [18:18] compute demands. How do we think about [18:20] compute compute demands needs now and in [18:23] the future in light of what people are [18:24] actually doing with AI? [18:26] >> That's the progression, right? So Chadic [18:28] was a casual app and it's coding for [18:30] professionals right now using coding you [18:34] have now built amazing tools for [18:36] knowledge workers so that's the next [18:38] frontier and now there are over a [18:41] billion knowledge workers in the line [18:43] right and with that it's a long ways to [18:47] go for that demand and by the way the [18:50] work that they do all this workflow and [18:52] automation and so on and then you get to [18:54] the back office which is all the agents [18:56] So progressively each of these things [18:58] unlocks [19:00] uh I would say order of magnitude more [19:02] demand. I mean just at the start of this [19:04] >> well and now you have Grockbot which is [19:08] kind of uh you know what uh happened [19:11] with coding is kind of happening with [19:14] all use of computer viaot and so we're [19:18] in a whole another wave of demand and [19:21] most certainly there's going to be more [19:23] to come. Uh so it does seem quite [19:28] unlimited at the moment and we haven't [19:29] even gotten into embodied AI or robots [19:33] uh which are going to be another source [19:35] of demand. [19:36] >> Martinez is expert but my understanding [19:39] is that part uses computer use which is [19:41] just like [19:42] >> human being sitting inside the computer [19:44] typing away. [19:45] >> I literally used it over the weekend to [19:47] to update my credit card with a bunch of [19:49] services that I'd been like lazy to do [19:50] and cancel a bunch of subscriptions. I [19:52] mean, this is not coding or whatever. [19:54] This is true computer use. [19:55] >> All of a sudden, you're creating like [19:57] half a billion knowledge workers except [19:58] they're all sitting inside of the [20:00] computer doing what? [20:02] >> I I do I do think that that that Mark [20:04] Mark Andre is right. It's like the right [20:05] analog here is like the steam engine or [20:07] electricity in the following way. Like [20:09] we've we've introduced this new thing [20:11] that you can turn to work and there are [20:14] some very obvious applications now, but [20:16] there's probably 30 40 years of throwing [20:19] computer at problems. anything with a [20:20] clear reward signal and we we're just [20:22] starting like we've got language and [20:23] code that's it and just starting [20:25] computer use but like what else are we [20:26] looking at we're looking at uh in terms [20:28] of science materials biology I mean of [20:31] course creativity is a massive use and [20:34] so listen we're at the very very early [20:35] part of a very long journey and we've [20:38] removed this key bottleneck which is you [20:41] know traditional software engineering [20:43] now of course you know bottlenecks will [20:44] move and there'll be kind of more [20:45] complexity elsewhere but I think we're [20:47] at a very early in a very long run of [20:49] throwing computer problem. So let's [20:50] expect you know this compute need to [20:53] persist for decades. [20:55] >> But because we mentioned it u Marty talk [20:57] about grabbot um because we um we're [20:59] talking at the offsite about how you [21:01] know what what struck you about it. [21:02] Obviously we're involved in every [21:03] possible way you could you could be [21:04] involved but what um yeah what what did [21:07] you find so interesting about it? So I [21:08] think we've I think we've as an industry [21:10] gone through kind of multiple [21:12] realizations for how AI enters our [21:14] lives, right? And and uh very early on [21:16] we're like okay well you add AI to a [21:19] product and it's like whatever it's like [21:21] a search bar and then you kind of you [21:23] know you you do chat with it and it [21:25] chats back because that's kind of the [21:26] traditional way to do it. Um uh and then [21:29] Open Clock kind of showed up and that [21:31] was it earlier in the year. And with [21:32] Open I say, "Okay, well maybe like it [21:35] just being like Google but better. Maybe [21:37] that's not the full embodiment of it. [21:38] How about we'll have it be a standalone [21:39] thing but it'll be an extension of you [21:41] and it'll share your keys and it'll know [21:44] your passwords and it'll just kind of do [21:45] stuff that you would do, right? So it's [21:47] kind of an extension of you but it's [21:48] more like a human, an extension of you." [21:50] And then what I think Rockbot got really [21:53] right is no, how about it's actually an [21:55] employee. So now you have this thing [21:56] that's an entity and it doesn't have [21:59] like special access to your keys or [22:01] whatever. It has its own computer and it [22:02] has its own browser and because these [22:04] are the smartest models in the world, it [22:06] can do whatever an employee can do. And [22:08] it's kind of interesting because now [22:10] actually if I if I want something done [22:12] my first thing I I think is like well [22:13] can do it for me and and often the [22:17] answer is yes even if it's something you [22:18] wouldn't you know expect it. So, the [22:20] obvious ones are like whatever. It'll [22:21] like manage my calendar. It'll like book [22:24] a meeting, but there's also non-obvious [22:26] ones as well. Like, so, for example, [22:27] I'll have it um uh read through my email [22:30] and do triage. And and I did I don't [22:33] tell it how to do that, but it will know [22:36] to check with me before actually doing [22:38] the triage. So, like these things are [22:39] sophisticated enough that you can give a [22:41] relatively high level task and it'll do [22:42] kind of, you know, like sophisticated [22:45] things as a result. [22:46] >> Ben, I know you're thinking a lot a lot [22:47] about this and how this you know, works [22:48] in the organization. you think a lot [22:49] about culture of course. What are your [22:51] thoughts here? [22:53] Well, I mean, I think if you just look [22:55] at us, um, [22:59] you know, it's like having a new kind of [23:02] employee and [23:04] there's going to be a lot of them and we [23:07] have to just like, you know, with our we [23:10] spent many, many, many years figuring [23:11] out how to work with our kind of regular [23:14] human employees and now we've got these [23:16] other kinds of employees [23:18] and, you know, there is a learning curve [23:21] with them. So um they can burn a lot of [23:25] tokens and spend a lot of money and get [23:27] nothing productive done. Um they can [23:30] forget stuff. They can make stuff up. Um [23:34] you know they can have good behavior. [23:36] They can have bad behavior. They can [23:38] >> they can create security problems. So [23:41] like there there there's all those [23:43] aspects to it, but like they can also be [23:46] like super duper productive. And so I [23:50] think figuring out how to integrate them [23:52] in, have them work nicely with the [23:53] people that they're working with, um, [23:55] the actual humans, uh, is all something [23:58] that we're learning how to do. I mean, I [24:00] I don't want to sit up here and say I've [24:02] cracked the code. We've got this more of [24:04] a sloop and [24:06] the whole firm is just completely [24:08] automated now, and I'm going to slowly [24:10] get rid of all the humans because I can. [24:12] Like, that's not at all where we are. [24:14] were much more going like, okay, how do [24:16] we make all our humans superhuman um [24:19] without like wrecking the place um [24:21] because the bots got out of control? [24:24] >> Yeah. And it's interesting because we've [24:26] done we've tried a couple of different [24:28] ways as how best to get agents into the [24:31] system if you will and eventually [24:36] it was Martin's insights just treat them [24:38] as people and get it done. And that's [24:40] what we're doing. that's turned out to [24:42] be the most durable way of getting this [24:45] thing going inside of an organization. [24:47] >> I want to go back to the supply side and [24:49] go deeper into the the bottlenecks. You [24:51] we were talking about how you know in [24:53] terms of the data centers, the chip [24:54] architecture, system software, [24:56] facilities themselves that none of them [24:57] were designed with with AI in mind. What [25:00] would it look like for them to be [25:01] designed with AI? Like what is sort of [25:03] the mental model for thinking about what [25:05] that could mean? [25:06] >> Yeah. So um I mean if you [25:10] start with a statement that you just say [25:11] hey original model [25:14] um of infrastructure on any of these [25:17] models has to change you can go se [25:21] category by column and see how it bricks [25:24] right and then you start uh unlocking [25:26] the bottlenecks in each one of these [25:27] things. So eventually you have to get to [25:31] a system where if you look at what an [25:33] inference engine does, right? It takes [25:36] up a lot of memory, it generates new [25:38] tokens along with the compute. And so [25:41] you can just think about how do I [25:42] optimize all of this? What does the [25:44] memory need to be? What does the compute [25:46] need to be? How do they need to talk to [25:47] each other? How much power does each of [25:49] them need? And if they need all of these [25:51] power, how do you cool each of these? [25:53] Right? And then how do you put the [25:55] collections of these things together? [25:57] That is the exercise that's underway in [25:59] the industry right now with a lot of the [26:00] founders. So they're breaking down the [26:02] problem into its fundamental components [26:04] and saying what is the exact nature of [26:07] the compute that's getting done. Okay, [26:09] it's how it's going to be matrix [26:10] multiplications. How do I optimize my [26:13] compute around that kind of a scenario [26:16] and then they all need very [26:19] progressively to generate these tokens? [26:21] What is the best way of hierarchically [26:22] arranging this memory right and then how [26:25] does the power consume I mean then you [26:27] got to connect it together what are the [26:28] ways of connecting it on the same chip [26:31] but across chips and across data centers [26:33] how much power does each of these data [26:35] transmission take so you have to [26:37] progressively break it all down and [26:39] rebuild it from these fundamental [26:40] building blocks and that's what we see [26:42] underway and that's where we see the [26:43] opportunity [26:44] >> let me give you an interesting mental [26:46] model to think about how the landscape's [26:48] changed so um uh so today to build a [26:51] frontier model costs let's say $3 to5 [26:54] billion right so and let's you know and [26:58] and that's to train it and so the [27:00] inference has to pay back at least that [27:02] of course right you know in order for [27:03] any of this stuff to be viable let's say [27:05] two times that so let's say that now [27:07] inferred has to to make $10 billion so [27:12] if you can save 20% of efficiency on [27:14] that that's $2 billion and you can [27:17] easily build an ASIC for $2 billion [27:20] Right. So, so we've actually gotten to [27:22] this interesting point in the industry [27:24] where it actually makes sense to build [27:27] an ASIC per model just because the [27:30] amount of capital investment in that [27:31] model and then unlike traditional [27:33] software, traditional software has a lot [27:34] of state and a lot of you know is very [27:37] dynamic. These models are fixed. The [27:38] model weights are fixed. And so we don't [27:41] know if the world goes to per model A6. [27:44] But it gives you a great mental model of [27:46] how you would evolve the architecture to [27:48] be far more bespoke for these massive [27:52] capital investments we're doing. Like I [27:53] don't think in the history of the [27:54] industry we've ever created a digital [27:56] artifact with something like $5 billion [28:00] that went directly into that artifact. [28:01] And so you know like this I think is [28:03] going to put the greatest demands on [28:05] hardware that we've ever seen. Well, to [28:08] that end, rack power requirements are [28:10] moving from roughly 5 to 10 kilowatts to [28:13] 100 to 50 kilowatts. Compute density is [28:15] climbing something like 70x. [28:17] >> Cooling is moving from air to liquid as [28:19] a requirement. What are the investment [28:21] opportunities as a result of this? [28:24] >> Well, first of all, when you get to that [28:26] level of power per rack, AC power [28:29] doesn't work anymore. So, like that's a [28:32] pretty wild thing. Um, so now now you're [28:36] into DC power, which by the way also [28:38] requires its own cooling. Um, and is [28:42] like, [28:44] uh, by the way, super dangerous. [28:47] Uh, [28:48] which is kind of ironic because this [28:50] was, um, Edison promoted DC power by [28:54] claiming how dangerous AC power was in [28:56] demonstrating it by like electrocuting [28:59] animals and things. [29:00] >> The horse. Yeah. [29:01] >> The horse. Yeah. [29:03] So but he was right but around his own [29:06] kind of power which is extremely [29:09] powerful is the good news. Um so you [29:11] know just starting with power uh yeah [29:15] that's going to be like very very [29:17] different I think with cooling. So and [29:20] this gets into [29:23] so yes we're going air cooling to liquid [29:25] cooling. I think we're already at liquid [29:27] cooling for any state-of-the-art data [29:28] center. Like that's already kind of a [29:30] done thing. [29:31] >> But it gets into okay, you know, given [29:34] the political environment and so forth, [29:37] you like liquid cooling isn't enough. [29:40] It's got to be eco-friendly liquid [29:42] cooling. Um, and you know, kind of DC [29:46] power is not enough. It's got to be um [29:49] power that contributes uh to the power [29:52] of society, not takes away from it. And [29:54] so you have data centers who have been [29:58] behaving badly. Um small percentage [30:01] actually probably 10% [30:04] wasting a lot of water. Um you know not [30:07] as much as pistachios or almonds and so [30:09] forth as people demonstrate on the [30:11] internet but like they could be a lot [30:12] more efficient with that. Uh and then [30:14] there are ones that you know kind of are [30:18] uh parasites of power and don't [30:19] contribute power back. I think all [30:22] that's going to end. it's going to have [30:24] to end just because like like that we [30:27] we've kind of gone through a one-way [30:28] door on that. Uh so that requires like a [30:32] level of engineering um that you know [30:35] many haven't invested in yet. So that's [30:37] coming. Uh and then you know like if [30:42] racks are that dense um there are other [30:46] things that like the way the floors are [30:49] designed have to support that kind of [30:51] weight. uh you know that kind of thing [30:54] is is actually for real. Um and then you [30:59] know I I think that there's [31:02] you know you just need a lot of [31:04] everything and so there's going to be [31:06] you know kind also by the things are [31:08] really loud so you have to build the [31:10] data center with thicker walls or you're [31:12] going to disturb the peace in the [31:14] neighborhood which is not going to be [31:15] acceptable like I don't think any [31:16] state's going to allow that. [31:18] >> And so a lot of the ways people have [31:20] architected and designed [31:22] the buildings themselves are already [31:25] completely obsolete. Like once we get to [31:27] Fineman, um a much smaller percentage of [31:31] the data centers that we have today [31:32] work. In fact, I mean everybody talks [31:35] about memory prices, but one of the [31:37] fastest areas where prices is increasing [31:39] is reinforced concrete from for their [31:42] sake. [31:43] >> The other thing that happens when these [31:45] data centers are uh sending 800 volts to [31:49] the whack is it's become so dangerous [31:53] number one, but secondly, we don't have [31:55] enough electrical contractors that have [31:58] the expertise to deal with the 800 rules [32:01] inside the data center. because this is [32:04] high voltage. Only 2% of electrical [32:08] engineer or uh electricians in the US [32:11] have been certified on DC power. So like [32:14] that gives you an idea. Now Meta's got a [32:16] whole program to train people up and so [32:18] forth which is great. It's like a new [32:19] job court where they train people for [32:21] free to do this job. But you know it's [32:24] funny AI is taking all the jobs. AI is [32:26] going to create a lot of new [32:28] electricians. [32:29] >> Yeah. I think we're just doing something [32:31] in space too. the big guys that own the [32:33] big cloud data big data centers, they [32:37] all are furiously experimenting with the [32:41] robots, right? To do the work of [32:43] assembling or putting servers into the [32:45] data center, etc. And so you will see [32:48] that increasing as a result of the [32:52] evolution in AI. [32:53] >> By the way, to be to be clear on the um [32:55] on the actual fund that we're raising, [32:58] our focus is on computer science [32:59] infrastructure. So anything a model runs [33:02] on that's computer science right so [33:04] think you know chips network [33:06] interconnect storage all the way down [33:10] probably to the electricity [33:11] >> yeah and and saying more about the [33:14] robotics arm in terms of what we'll be [33:16] doing versus maybe American dynamism or [33:17] how to think about that [33:18] >> yeah yeah for sure so um you know again [33:21] we we think that any platform that that [33:23] AI will run on like one of the the the [33:26] great breakthroughs that AI does is it [33:28] allows computers to interact with the [33:30] physical world, right? It can see, it [33:32] can hear, it can talk, right? And this [33:34] means new platforms, right? And the [33:36] simplest way people say edge device, but [33:37] that doesn't really mean anything, [33:39] right? I mean, it could be a mobile [33:40] device, it could be a CDN, it could be a [33:42] laptop, but it also could be an [33:44] embodied, you know, device that goes [33:47] around. And so again, we we um as you [33:50] know, as infrastructure focused [33:53] investors don't do heavy regulated [33:55] industries um or more verticalized [33:57] industries, but any sort of computer [33:59] science platform that's going to push AI [34:01] further out, we're quite interested in. [34:04] >> Yeah. Going back to the data centers, by [34:06] 2028, new data centers are going to need [34:08] something like 44 gawatts of additional [34:10] power against maybe 25 gawatt of [34:12] expected grid additions. [34:15] >> Hold on, hold on. We use that word [34:16] gigawatt. [34:17] >> No, it's like it's like we'll have 100 [34:20] gigawatt, [34:22] >> Martine. What's a gigawatt? [34:23] >> I mean I mean how big is it? It's [34:25] multiple football fields. I mean it's [34:27] massive. It's 50,000 50,000 people. [34:31] >> What do you mean? What is it? [34:32] >> Like like [34:33] >> the equivalent is like 50,000 houses. [34:36] >> City [34:36] >> 50,000 50,000 homes. 50,000 homes. [34:40] I I I grew up I grew up in Flagstaff, [34:42] Arizona, [34:43] which is a town of 40 to 60,000 people, [34:46] depending on the universities. We have [34:47] less than a gigawatt of power [34:48] consumption. I mean, this is [34:49] >> so you can basically light up and air [34:52] condition your entire town for a [34:54] gigawatt. [34:54] >> Yeah. I mean, this is [34:55] >> just throwing them around. [34:58] >> No, but by the way, everybody talks [34:59] about the gigawatt. There's very few [35:00] gigawatt data centers that are actually [35:02] up. I mean, we've got a long way to go, [35:04] right? But then why can't utilities and [35:06] hyperscalers just build faster? [35:09] Oh, there's so many things there. Well, [35:11] there's first of all, right now, you [35:13] need humans to build them. So, there [35:15] there's just like the regular [35:17] construction, but much more than that, [35:19] you need [35:21] permits. [35:23] Um, you need access to power uh that you [35:27] can plug in. So you're either you're [35:29] doing a combination of you've got to get [35:32] access to power which is a massive kind [35:35] of regulatory [35:37] bidding struggle. There's very limited [35:40] kind of amounts and things you can tap [35:42] into in terms of n natural gas power [35:45] grids, what have you. Um but then you [35:47] also have to build your own power and [35:50] guess what? We've got shortages of [35:52] transformers and turbines and everything [35:54] that goes into that. So, it's just, [35:59] you know, like you've got to get all [36:00] that stuff. It's not this is not a [36:02] software problem. It's not just like a [36:04] bunch of engineers can't you like work [36:06] weekends and that type of stuff like [36:08] this is not that that works anyway, but [36:11] uh [36:12] there are real bottlenecks in this and [36:14] these these lead times are not that easy [36:19] to compress. And look, we have the best [36:21] minds in the world trying to figure out [36:23] how to compress them. Uh but [36:27] I it's not easy. It's not easy and the [36:30] demand is not slowing down. So we're [36:32] already behind. The demand is growing, [36:35] you know, 10x a year right now. And you [36:39] the supply just can't grow that fast. [36:41] >> By the way, it is so bad that right now [36:43] if we have new companies going for GPUs, [36:45] it's often in Mexico or Australia or [36:47] another country just because it is so [36:50] difficult in the United States. Yeah, [36:52] we're creating huge both job and [36:55] long-term economic opportunity in other [36:58] countries by uh banning data centers [37:00] here. I think look the the right answer [37:02] would be to set a standard where a data [37:06] center contributes [37:08] back to the community like that power [37:12] gets better, there's no noise, there's [37:14] no water issue, um and it's adding jobs [37:18] like that ought to be the standard. [37:19] Yeah. And then everybody ought to be [37:20] just held to that center. And by the [37:22] way, like there are there are data [37:24] centers that do that now. Like that's [37:26] not a you know like a a futuristic dream [37:29] or something. Rates energy rates have [37:32] gone down like every year they're there. [37:34] And the reason is they provide their own [37:37] power. They give power to to the state [37:41] during the day and then at night they [37:44] borrow power from the state when the [37:46] state doesn't need it because you're [37:47] always the the way power plants work is [37:51] you're always generating peak uh [37:54] capacity and since a data center has [37:56] steady capacity during day and night and [37:58] a city goes way up in the day and way [38:00] down at night um that's a symbiotic [38:02] relationship. zooming out. Uh why do we [38:06] think the you know we we've been batting [38:08] around the name for for a little bit. Uh [38:10] why do we think machine age is a is a [38:13] compelling term for for what we're doing [38:14] here? [38:16] >> Well, listen, let let me let me take it. [38:18] So, the first one is I think Ben's [38:19] absolutely right. Artificial [38:20] intelligence was the wrong word. Like we [38:22] shouldn't have called it it's machine [38:23] intelligence. Um [38:26] >> say more about that. Why is that? [38:27] >> Uh because it's not how humans think [38:31] necessarily, right? I mean, it is a [38:33] cache of how humans thought is a [38:35] collection of humans thoughts. But like [38:37] to date, we don't know how to take a um [38:40] an AI with no knowledge and put it in [38:42] out in the world and have it reconstruct [38:44] language, right? Like that's not what [38:45] we've done, right? We've we've built a [38:47] something that can learn off of [38:48] everything we've already learned and [38:49] then use that in a productive way. And [38:52] listen, a a AI is a general term that [38:54] goes back 70 years in computer science [38:58] formally that applies to many different [39:00] things. And of course it's got a lot of [39:01] baggage either from science fiction or [39:03] from you know Nick Bstrom who wrote [39:05] about it or whatever. And so so the [39:07] first one is just an acknowledgement [39:08] like this really is machine [39:09] intelligence. And then you want to [39:11] emphasize the machine part of it. I mean [39:12] there's kind of this deep irony and this [39:14] is from you know the uh the software's [39:16] eating the world people that you know [39:18] you've really come to a place where you [39:19] pour money into something and then [39:21] you're limited by the actual machines [39:23] below it. And so I think it is a kind of [39:24] a nod to like the hardware component is [39:26] so significant in this wave and and we [39:28] want to acknowledge that. [39:30] >> Yeah. I think that's what's going to [39:31] create the next breakthroughs in [39:33] intelligence is the quality of the [39:36] machines underneath. And so that's [39:38] that's basically a reason for the name. [39:40] >> It's also a cool name. [39:42] >> That sounds good. [39:44] >> Futuristic. [39:45] >> Yeah. the um given how much has been [39:48] spent on AI infrastructure to date and [39:50] how capex intensive these businesses can [39:52] be are we past the point where new [39:54] companies can break break in at at at [39:56] sort of material levels uh you know you [39:59] know in why not incumbents like Nvidia [40:01] Core etc just take the line share of [40:02] these markets [40:05] they're all doing well there's no [40:06] question about it right but to our [40:08] discussion earlier when you're get you [40:13] need funly new innovations to keep the [40:15] growth both continuing or the pace of [40:18] improvement continuing whether it's [40:19] tokens per second per dollar or tokens [40:23] per watt [40:25] uh tokens per rack right um or power you [40:29] take any metric if you want to have a [40:32] 10x on those metrics you got to have new [40:34] innovation and new innovation [40:36] traditionally comes from brilliant [40:38] founders thinking about solving the [40:40] problem from first principles in a [40:42] different way right and that's what's [40:44] needed here for the next jump in [40:46] innovation. [40:47] >> I mean this is the law of markets, [40:49] right? I mean let's assume that the the [40:51] existing silicon incumbents are multi- [40:53] trillion dollars in market cap which is [40:55] absolutely the case. [40:57] >> Even [40:58] 5% of that is a massive private company, [41:01] massive private company, right? We're [41:03] talking, you know, annual. Um and you [41:06] could say, well, but Nvidia could do [41:07] that. They could, but why would they if [41:09] they're focused on things that are in [41:11] the 90% which is also driving the same [41:13] amount of growth? You always ask these [41:15] questions. We ask these questions during [41:16] the cloud days, right? Like why wouldn't [41:18] Amazon did this? You would ask these [41:19] questions during the Microsoft days. Why [41:20] wouldn't Microsoft do this? There's a [41:22] very natural law of markets is once you [41:24] get to a certain scale, there's [41:25] tremendous opportunity for innovation um [41:28] at at the margins. [41:29] >> Yeah, there's a funny uh quote from our [41:31] partner Alex Rmpelle. He had this [41:33] startup called Trial Pay and he was [41:35] trying to sell it to or sell the it his [41:38] services to to Meta and uh then Facebook [41:43] and Dan Rose who was the head of corp [41:45] dev at the time said Alex that's great [41:48] you're it sounds like you can collect a [41:50] lot of silver bricks but I'm like I have [41:53] so many gold bricks I can't even pick [41:54] them all up so the last thing I'm doing [41:56] is looking at a silver brick and I think [41:58] Nvidia is in that position [41:59] >> 100% yeah we were talking about as [42:02] relates to the model providers that if [42:03] you're, you know, in the sweet spot of [42:05] of what open airropic can can do, you [42:07] know, one of their main sort of interest [42:09] areas, that might be a tough place to [42:10] be, but anything outside of those maybe, [42:12] you know, three to five areas might uh, [42:14] you know, as as markets expand, they [42:18] fragment, right? And it happens all the [42:20] time. And remember in the early days of [42:22] Ford, there was the 1913, there was the [42:24] Rouge River plant. Literally this, you [42:26] know, this was like made cars like in [42:28] went like water, coal, and rubber trees [42:31] and out came cars. [42:32] >> By the way, he bought a whole [42:35] >> rubber [42:36] plantation in the Amazon jungle, [42:39] >> right? And and there's a great book [42:40] called Ford Landia where so because he [42:43] wanted to like own like the complete [42:44] vertical thing where he created this [42:48] city called Ford Landia in the Amazon [42:52] jungle um which had like was all [42:55] Americanized band stands and ice cream [42:57] and all this kind of stuff. And it [42:59] actually worked for a while until uh he [43:01] made people like show up to things on [43:03] time and then they were like screw this [43:05] get the out of here. [43:07] So, so now if you look at the car [43:08] industry, of course, there's multiple [43:10] levels of supplier and there's a bunch [43:11] of companies and this always happens. [43:13] So, you know, as markets expand, they [43:14] fragment. There's a lot of and then and [43:16] then once that growth slows down, they [43:18] tend to consolidate. The consolidation [43:19] can either be acquisition or it can be [43:22] like new challengers rise up and that [43:24] that is the, you know, everlasting cycle [43:26] of private markets. [43:27] >> Yeah. the use cases are are multiplying [43:29] and there's no way like if you're the [43:32] biggest company you can get to the [43:34] biggest use cases but there's so many [43:35] use cases and all as Martin was saying [43:39] very valuable use cases that it's just [43:41] very hard to get to in a great way [43:43] >> yeah even inference used to be one [43:45] simple architecture right and it no [43:48] longer is like it's so complex now so [43:51] it's inevitable that you can optimize [43:53] things in a different way [43:54] >> by the way here's a very interesting [43:55] thing like people don't um uh often [43:58] don't understand that like the margins [44:00] kind of fell out of the standard way of [44:02] doing the technology with software, [44:05] right? Like it wasn't really a [44:06] technology problem. Like once you got [44:08] the business working, you tended to have [44:09] pretty good margins because that's just [44:11] kind of how software works. certainly [44:12] when you shipped it but even as a [44:13] service uh and that's not necessarily [44:15] the case with AIS we may actually be [44:18] entering an era where the optimization [44:21] in the hardware is absolutely meaningful [44:24] to the upside of the business in a way [44:25] that we haven't seen in the past so [44:27] there's a lot of opportunity here let's [44:30] get deeper in talking about the types of [44:32] companies we'll be investing in um maybe [44:34] we could start by either illustrating [44:36] the the the subsectors or if we can talk [44:39] about a few um or a couple investments [44:41] that we've I know there's some that [44:43] haven't been announced yet, but Robert, [44:44] do you want to take us down? [44:46] >> Yeah, I mean the sub sectors as we've [44:48] been talking all is every one of these [44:49] categories, right? The obvious ones are [44:53] um uh computer chips, but these days [44:56] it's not enough to build a chip. You [44:57] need to build a full system, right? And [44:59] then therefore, what goes into the [45:01] system? There's potentially memory [45:03] innovation. There is potentially [45:05] networking innovation. There's [45:07] potentially power chips and so on and so [45:09] forth. So each one of these categories [45:11] are categories where you can see public [45:14] company style companies emerging and [45:16] those are all things that we are looking [45:18] into. Um and then once you put it all [45:21] together there's a layer of software [45:22] around it to automate all of these [45:24] things to manage these fleets and so on [45:27] and so forth. So that is another [45:29] important area. So these things keep [45:32] building on each other but every one of [45:33] these categories is important. [45:35] >> Talk about what's different about these [45:37] kinds of companies from from the usual [45:39] company. I mean one thing you can tell [45:40] from the companies we announced is their [45:41] first rounds have been massive. You know [45:43] hundreds of millions. Is it a different [45:44] kind of founder or what else is [45:46] different as we think about just the [45:48] practice of you know building and [45:49] investing these kinds of businesses [45:50] relative to our traditional software? [45:53] >> Well I I think the big thing is you hit [45:55] on one of the big things which is a lot [45:56] of money goes in um before they get to a [46:00] product. Um and that's just kind of the [46:02] nature of it. Now that's true on big [46:04] models too, but I I would say [46:08] that's a little more of a known path. Uh [46:11] whereas this has got a little more risk [46:14] and a little more money than uh than [46:16] some of the other things that we've [46:18] done. But [46:19] >> um and you know, look, a lot of a lot of [46:23] the chip founders um are here from the [46:26] past. [46:27] >> You know, like the guys who know how to [46:29] make memory, [46:31] they're not young. Yeah, you know, so [46:35] it's uh you know that part is different [46:37] too, but it's kind of exciting, you [46:38] know. [46:39] >> Yeah. The other thing about these [46:40] founders, [46:42] they have all got to be [46:45] systems founders. So what I mean by that [46:48] is you can't just be a researcher or a [46:50] great computer scientist, right? You got [46:53] to be able to architect and design the [46:55] chip or the system, whatever it is. Then [46:57] you got to think about how is this thing [46:59] going to actually get manufactured, [47:00] right? who's going to be supplying this [47:03] and a whole bunch of these downstream [47:05] things which normally if you're building [47:08] software you don't have to think about [47:09] all of these things. So the to really [47:13] the best founders um [47:16] and of course Jensen is the Michael [47:19] Jordan of this right they think the [47:21] entire ecosystem right from the get- go [47:24] before they start designing the chain [47:25] right because of the nature of the [47:27] bottlenecks and all these things that [47:28] have to come together. So that's that's [47:30] a big characteristic that is different. [47:33] >> There there's two environmental factors [47:34] that are important too. The first one is [47:35] the labs are so desperate that they will [47:38] engage with startups. And so like we [47:39] actually have quite a bit of signal [47:41] early on because they're you labs are [47:44] inking deals with companies before they [47:46] actually have hardware available. And [47:47] that's a big big shift than you know 5 [47:50] years ago, right? Like you just didn't [47:51] go and you know sell your kind of janky [47:54] hardware thing to Google or whatever. So [47:55] that that's a shift. The second one is [47:57] the the capital availability has [47:59] loosened up a lot. I think there's [48:00] general consensus that that you know it [48:03] is the time that to reshape this stuff [48:05] and so follow on rounds there's a lot of [48:06] capital available which you know of [48:08] course you want to be investing into [48:09] areas where there's capital available [48:10] and so the atmospherics are also just [48:12] different. Patrick Carlson, you know, [48:14] remarked a few years ago said, "Hey, it [48:16] feels like there's less younger founders [48:19] today, you know, in the way that you [48:20] know, Zuck, you know, in college [48:22] building next Facebook or or Gates, um, [48:24] you know, in the same way with Microsoft [48:27] and of course, you know, the Michael [48:28] Trolls of the world, there there's [48:29] still, you know, some young founders [48:30] building iconic companies, but it does [48:31] seem, you know, to your point that [48:33] there's more older founders building [48:36] these these these companies or or less [48:39] 20-year-olds. I'm I'm curious if you [48:41] resonates and why. Well, I think it's [48:43] Ragu's point that if [48:47] if you're building something that has [48:50] like a very complicated supply chain, [48:52] has to manufacture things um and is [48:55] technically complicated that you know [48:58] some experience helps uh and you know if [49:03] you look at Elon or Travis Kalanick [49:06] their companies when they were young [49:08] were software companies. It wasn't until [49:10] they got like a lot even those guys the [49:14] best guys um needed some experience in [49:19] building a companies building technology [49:21] and so forth to kind of graduate to the [49:23] much more kind of complicated or [49:27] I would say elaborate domains you know [49:29] there's just much more there many more [49:32] moving parts in these things and so look [49:35] when you're learning how to build a [49:38] company it's hard enough if you [49:39] completely understand the product. If [49:41] you don't completely understand the [49:43] product and have to learn it while you [49:44] build the company, um that's just such a [49:48] steep learning curve for a brand new [49:49] entrepreneur. So, I think that what [49:51] we're seeing is you see Michael on the [49:53] one hand, um who is a very young guy, [49:56] brilliant, but what he built was kind of [49:59] a pure [50:01] software AI thing. [50:03] >> Yeah. And then on the other end, you [50:04] have like an Elon or a Travis who can [50:07] who's got enough experience. I think [50:09] Michael could probably do that, you [50:11] know, 10 years from now, but today that [50:14] would have been hard. [50:15] >> And it's important to remember like it's [50:17] been defocused by the entire industry [50:18] and academia for the last 20 years, [50:20] right? It just there just hasn't been [50:22] the same opportunity. Like it's been [50:24] there, but like it's never been a growth [50:26] area. The growth areas have been, you [50:27] know, software, um, networking, things [50:30] like that. And so I also think we just [50:32] have a posity of people coming out of [50:33] the universities or having experience at [50:36] large companies that have have done [50:37] this. I there just not that many. [50:39] >> Like you don't go intern and like build [50:41] a chip. So but a lot of that's changing [50:44] now. like listen we're going to create a [50:45] whole generation of of you know founders [50:48] that come from these new companies that [50:50] will know how to do this and you know [50:51] they'll be hired in much more junior and [50:52] like I would say actually one of the [50:54] greatest legacies of Elon towards this [50:56] is is of course he's created these great [50:58] companies but the amount of [50:59] entrepreneurs that have come out of [51:00] SpaceX that that have are changing the [51:03] entire industrial complex maybe even [51:05] greater um legacy than than the [51:08] companies themselves and I think we're [51:09] going to see the same thing uh for for [51:11] computer science and hardware [51:12] >> y as a matter of All of our investments [51:15] was started by uh two founders in their [51:18] 20ies. But if you go watch one of their [51:20] offices, you see the experienced people [51:22] as well. So it's ideal combination here. [51:24] >> Yeah. [51:25] >> Yeah. Yeah. It doesn't necessarily have [51:27] to be the founder with experience, but [51:28] that founder better be able to tap into [51:32] that experience with it. [51:34] >> Yeah. Yeah. Well, and then be able to [51:35] work with them and and and [51:38] they have to be good and and all these [51:40] kinds of things. It's complicated. [51:42] Speaking of experience, this is a a big [51:45] new fund we're we're launching and [51:48] there's no new GPS. Um yeah, we're sort [51:50] of collecting. It's because you guys [51:52] have a lot of experience and the rest of [51:54] the group, you know, in in this field [51:55] that it's been kind of latent and uh [51:57] dormant. Yeah. Well, it's kind of funny. [51:59] I think we almost had to be warned [52:00] against it almost just because like our [52:02] backgrounds are from is hard. And I [52:04] think the reason that we needed a [52:05] reminder is because all of us have spent [52:07] so much in our careers existence and [52:08] hardware. We're kind of drawn to that. [52:10] And so listen, we've been clearly [52:11] invested [52:13] um in harbor over other years, right? [52:14] We're in SpaceX, we're in Andrew and [52:16] all. These are very early checks. We're [52:18] in astronomics, we're in Whimo, you [52:20] know, so we even even early on we did a [52:22] number of those investments. Um but like [52:25] you know this is because it's so much in [52:28] our DNA and so I don't think this is [52:29] necessarily need to increase the team [52:31] competencies just bonus. [52:32] >> If this fund does what we think it will [52:34] do, [52:35] >> how do we see the world changing or [52:37] looking like in 5 to 10 years? Well, you [52:39] know, hopefully uh America wins in the [52:42] infrastructure game. Um and we have lots [52:46] of like super eco-friendly efficient [52:49] data centers out there and lots and lots [52:52] an abundance of chips and abundance of [52:53] memory and abundance of power. Uh and [52:57] you know that would be awesome. Uh, and [53:00] I think we look we we, you know, it goes [53:02] back to like we really think uh America [53:06] is a special place and um we're [53:09] important not only to everybody here but [53:11] anybody in the world who wants to kind [53:14] of make a contribution and do something [53:16] bigger than themselves that it's kind of [53:18] the best place to come with nothing and [53:22] do something profound. So we'd like to [53:25] keep that going and I I think that [53:27] doesn't continue to go if we lose our [53:31] lead in technology. Think I think we'll [53:33] be in another era and it'll be another [53:34] country and maybe they have a different [53:36] set of values around that. [53:38] >> Thanks. That's a wrap. [53:39] >> Great for seeing me fun. Martine Beni. [53:42] Thank you. Thank you. [53:42] >> Thank you. Thanks. Yeah.