A16z's George Says AI Compute Absorbs Billions That Broke SaaS

a16z合伙人George称AI算力正吸纳曾拖垮SaaS的数十亿资金 与SaaS时代巨额融资导致企业人员冗余不同，直接投入AI算力集群的资本能够可预测地提升模型性能，从而使后期成长期风投有望斩获全行业70%的回报。 David George与Jack Altman指出，当前的AI浪潮是一个具有拓展性的“兼备型”周期，而非零和博弈；目前空前的企业营收仅由全球极少数开发者推动，而更广泛的知识工作渗透率仍不足5%。他们详细阐述了持续存在的供给端算力短缺、推理模型的兴起以及前线部署工程，如何让垂直应用层在基础实验室主导原始Token消耗的同时蓬勃发展。在生成式软件之外，他们还将自动驾驶与机器人技术视为一场即将到来的经济转型，其规模有望超越此前25万亿美元的技术浪潮。他们总结道，后期风投已转向支持那些具备宏大叙事能力且充满使命感的创始人，这类创始人能够将资本转化为武器，实现超越静态财务模型的规模化扩张。 当前生成式AI热潮的企业营收高度集中于软件工程师精英群体：全球虽有3000万开发者，但排名前100万的核心重度用户贡献了当下绝大部分企业支出。 (2:18) 全球数据中心电力容量至2028年已全部售罄，由于持久的测试期算力与推理需求，即便规模从4吉瓦翻倍至8吉瓦，也将被即刻消化。 (4:45) 与SaaS时代过剩资本导致人员编制膨胀和营销泛滥不同，AI资本能够通过算力扩展直接转化为模型能力，而不会破坏组织的运营效率。 (35:00) 涵盖AI、自动驾驶、机器人、美国活力（American Dynamism）及医疗健康的更广泛技术浪潮，预计将超越上一轮25万亿美元的周期，这一趋势正由安全性比人类高出10至14倍的自动驾驶汽车等物理世界部署所驱动。 (25:34)

David George of Andreessen Horowitz and Jack Altman argue that an emerging technological super-cycle spanning artificial intelligence, autonomous vehicles, and robotics will surpass the prior $25 trillion cloud-and-mobile wave. They contend that current enterprise AI adoption represents less than 5% of potential knowledge workers, positioning the market for multi-year expansion driven by compute infrastructure and late-stage capital deployment. The $25 Trillion Super-Cycle and Autonomy (25:34) The previous technology wave—combining mobile computing, cloud software, social networks, and e-commerce—generated roughly $25 trillion in public and private market capitalization. George and Altman project that the current cycle, which integrates generative AI, autonomous vehicles, robotics, AI-driven healthcare, and defense modernization, will comfortably exceed that benchmark. Within physical AI, autonomous driving represents a massive underappreciated shift: consumer vehicle ownership averages approximately $0.80 per mile compared to more than $2.00 per mile for human-driven rideshare services, giving autonomous fleets significant room to expand the mobility market. Safety data indicates autonomous fleets from Waymo are already 10 to 14 times safer than human drivers, an advantage George expects to expand to 20 times alongside order-of-magnitude cost declines. While Waymo operates fewer than 10,000 vehicles nationwide today, George anticipates manual driving will eventually be phased out over a multi-decade horizon across 250 million registered US passenger cars. Similarly, physical robotics is positioned to address an even larger total market than large language models, beginning with structured factory automation—such as Mind Robotics deploying inside Rivian assembly plants—before reaching broader consumer applications. Workforce Diffusion and Severe Compute Bottlenecks (2:18) Record top-line revenue at leading AI companies is currently propelled by a tiny sliver of the workforce: the world's 30 million software engineers, with spending intensely concentrated in an elite cohort of roughly 1 million developers. George estimates that enterprise diffusion across the broader pool of 1.5 billion global knowledge workers sits below 5%. This gap indicates substantial enterprise demand ahead as industries like legal technology begin adopting the software engineering playbook roughly twelve months behind. On the supply side, data center compute capacity is effectively sold out through 2028, insulating the sector from near-term oversupply even as operational footprints expand. George dismisses the likelihood of demand collapse, arguing that test-time compute requirements and inference scaling will sustain massive compute consumption. Consequently, high-growth technology companies are focusing on front-end product capabilities that offer 10x to 100x top-line upside rather than treating AI merely as an operational cost-reduction measure. Frontier Horizontal Focus and Vertical Moats (13:46) The top three frontier labs currently capture more than 95% of enterprise AI spending (roughly $125 billion across leading players), but their commercial strategies concentrate primarily on coding tools and horizontal productivity platforms aimed at office software suites. Because the labs risk destabilizing their developer ecosystems by competing directly in niche sectors, high-value vertical markets remain open to specialized software providers. In legal tech, corporate clients are already mandating that law firms deploy vertical tools like Harvey, while top law firms commit hundreds of millions toward modernizing tech stacks. Vertical application defensibility relies on domain-specific workflows and forward deployed engineering (FDE) teams rather than base-model ownership. Furthermore, application architectures are evolving toward tiered model economics: developers increasingly direct standard, high-volume workloads to cost-effective 'N-1' (one generation behind frontier) or open-source models to protect gross margins, reserving premium frontier tokens exclusively for novel, high-reasoning tasks. Capital Deployment, Valuation Flywheels, and Firm Scale (35:00) In venture capital, private return distributions are shifting from an even split between early and late stages toward a 70/30 distribution favoring Series C and beyond as companies stay private longer. Unlike the SaaS era, where flooding mature software firms with capital caused operational bloat and marketing waste, AI capital deployed into training and inference infrastructure directly and predictably improves underlying model performance. George notes that growth-stage alpha stems from backing 'model busters'—founders at companies like Databricks or SpaceX whose compounding product launches outpace static valuation models. A compelling market narrative also provides structural economic advantages. High valuations lower the cost of capital, reduce equity dilution, and enhance talent acquisition and enterprise credibility—a dynamic demonstrated by Palantir's conversion of public enthusiasm into commercial enterprise contracts. To support portfolio companies across this complex macro cycle, venture firms like a16z have scaled operationally, employing over 650 platform professionals to secure market share across leading technology domains. George and Altman conclude that the principal risk in growth-stage venture is underestimating the scale and compounding durability of exceptional founders and infrastructure buildouts. Nonetheless, enterprise diffusion remains constrained by high-touch implementation demands, brand safety parameters, and complex legacy system integrations, leaving a substantial gap between raw model capabilities and realized enterprise return on investment.

Andreessen Horowitz 合伙人 David George 与投资人 Jack Altman 认为，当前科技正迈入由人工智能、自动驾驶与机器人等驱动的万亿美元级超级周期，AI 市场的非零和扩张机制使基础设施与垂类软件具备并存发展的巨大经济空间。 算力供需瓶颈与知识工作者市场渗透 (0:00) 当前 AI 领域的顶级营收纪录主要由极少数人群驱动。在以 OpenAI、Anthropic 和 Cursor 等为代表的营收飙升中，绝大部分企业端支出集中于全球 3000 万名软件工程师中最顶尖的约 100 万人。David George 指出，目前企业级知识工作中生成式工作流的渗透率尚不足 5%，这意味着全球 15 亿知识工作者市场仍未被充分挖掘。随着法律等行业在技术采纳上紧随其后（滞后编程约 12 个月），未来五年企业需求预计将持续旺盛。 在供应端，算力与基础设施链条受到物理层面的刚性制约。George 说明，数据中心容量在 2028 年前已基本预订一空，即使运营规模从约 4 吉瓦翻倍至 8 吉瓦，新增算力也会迅速被市场消化。针对“需求是否会停滞”的质疑，George 评估认为，知识工作由于复杂工作流造成的落地滞后只是暂时的工程与数据治理问题，加之推理阶段的测试时计算（test-time compute）对模型表现的显著提升，推理算力规模未来将远超训练算力，算力供求基本面具有极高的确定性。 前沿模型主导地位与垂直应用软件护城河 (10:00) 在企业 AI 支出分配中，前沿实验室占据了超过 95% 的资金流，前三家机构吸纳了约 1250 亿美元。重度用户表现出对顶尖算力的明确偏好，愿意为高价值推理场景支付高达 10 倍的 Token 溢价。然而，前沿实验室的商业战略集中在编程及与其相关的工具链，以及替代通用办公套件的横向知识工具，缺乏动机深入数百个细分垂直行业构建定制化解决方案。 这为垂直应用软件提供了广阔的生存空间。以法律领域的 Harvey 为例，机构客户的态度已从禁止使用转向明确要求律所采纳，推理模型的出现更是精准契合了法律严密逻辑的需求。George 强调，垂直软件的真正护城河在于贴合实际工作流的“最后一公里”工程以及部署驻场工程师（FDE），而非底层模型权重。同时，企业正在形成分层调用架构：利用 N-1 代或开源模型处理常规、大通量任务以保护毛利率，仅将核心高难度逻辑交给高价前沿模型。 消费端拟物化瓶颈与物理自主系统爆发 (20:00) 消费级 AI 虽已拥有数以亿计的用户，但多数仍停留在将对话框当作“更快的谷歌搜索”这一拟物化阶段。George 指出，真正的商业拐点在于系统从被动响应转向主动执行的代理模式，并在多模态与低延迟语音交互成熟后，诞生融合订阅与新型原生广告的大型平台。历史上这类直接服务大众消费者的超级入口往往能捕获极其可观的经济价值。 在物理世界，自动驾驶与机器人正构成更大的经济驱动力。George 援引数据指出，Waymo 的全自动驾驶车队目前在安全性上已达人类驾驶员的 10 至 14 倍，全美每年约 1700 万辆新车销售和 2.5 亿辆保有量构成了巨大的消费级转化底盘。此外，机器人市场潜力预计将超越大语言模型。例如 a16z 投资的 Mind Robotics 通过直接进驻 Rivian 工厂组装线，在受控制造场景下积累即时收入和数据闭环，预示着工业机器人正逐步逼近突破性的落地时刻。 资本周期演变与溢价叙事的竞争壁垒 (35:00) 在风险投资层面，成长期投资的回报分布正在从早期的各占一半转向约 70% 集中于后期轮次。与 SaaS 时代盲目注资导致人员冗余和营销内耗不同，AI 时代的资本能直接转化为训练与推理算力并换取确定性的性能提升。George 评估当前产品周期高达 9 或 10 分（满分 10 分），尽管估值狂热使资本周期降至约 6 分，但核心企业的复合增长率依然打破了传统财务模型的测算上限，如 Databricks 和 SpaceX 展现出的持续爆发力。 同时，“叙事”（Narrative）与公众声誉在成长期已成为关键竞争武器。强势的叙事能够直接转化为更高的估值倍数，从而降低股权稀释成本、提供充裕的资本弹药，并帮助企业锁定顶尖人才与企业客户的信任。如同 Palantir 凭借领袖形象建立起深厚的机构信任，在 AI 时代，善于直接面向公众沟通的创始人与核心运营团队正赋予企业超越常规财务指标的结构性优势。 George 预测，科技占整体公开市场市值的比重有望从目前的三分之一向历史工业高点的更高份额靠拢，新一波由 AI 与自主化驱动的创新将超越此前移动与云计算时代创造的 25 万亿美元市值。但企业级深层系统的整合摩擦、昂贵的基础设施资本支出以及劳动力重构的社会适应速度，仍是该超级周期必须面对的现实限制。

## Background 本期内容来自播客节目《Uncapped》，由主持人 Jack Altman 与 a16z 普通合伙人 David George 共同录制，于 2026 年 10 月 4 日在 RSS 及各大音频平台发布。节目主要围绕人工智能的生态之争展开，探讨前沿大模型与开源生态、底层算力与垂直应用如何实现共存，并展望自动驾驶、具身智能及超越上一代 25 万亿美元规模的宏观科技超级周期。 ## TL;DR David George 指出，AI 领域的诸多争议并非零和博弈，其核心范式是底座模型与垂直应用的“共存”（And）。当前 AI 商业化主要由极少数核心开发者支撑，广阔的白领知识工作与消费级市场仍有巨大渗透空间；随着算力瓶颈持续与推理需求激增，高价前沿模型与低成本开源模型将走向工作流分工。与此同时，AI 与全自动驾驶、机器人及医疗科技的交汇，正在开启一个规模远超 25 万亿美元的新超级周期。面对极其强劲的“9-10分”产品周期，成长型投资逻辑已被重构：大额资本直接转化为算力杠杆，而极具远见的“弥赛亚式创始人”及其掌控叙事与氛围感的能力，构成了公司最坚固的商业护城河。 ## Atmosphere & style 整场对话洋溢着硅谷顶尖投资人之间高密度的智识碰撞与极其浓厚的科技乐观主义氛围。Jack Altman 提问精准敏锐且善于刨根问底，David George 则以从容自信、宏大而极富洞察力的表达风格化解市场普遍焦虑。两人在冷静拆解财务模型与出行经济学账本的同时，不时流露出对技术奇点的敬畏与兴奋，既有老练投资人审视商业周期的犀利务实，又充满见证历史性技术革命正在发生的澎湃张力。 ## Bulletin - `00:00` 当前 AI 基础设施建设资本开支已占 GDP 约 3%，OpenAI、Anthropic 与 Cursor 合计创造约 120 亿美元年化收入，其新增月收入规模已超过除 AWS 外的多数云厂商。 - `00:00` B2B 领域的 AI 商业化极度集中于全球约 3000 万程序员中的前 100 万核心开发者，在面向全球 15 亿知识工作者的大盘中，实际渗透率仍不足 5%。 - `10:00` 前沿模型实验室聚集了市场超 95% 的收入，仅 OpenAI 和 Anthropic 两家累计融资即达 3500 亿美元，但其第一方产品主要聚焦于编程辐射圈和横向办公平台，为垂直垂直软件留出了充足空间。 - `25:31` AI 与全自动驾驶、通用具身机器人、AI 生物医疗以及国防科技等浪潮的并发叠加，预计将创造远超上一代移动与云计算浪潮所带来的 25 万亿美元市值。 - `35:00` 成长期风险投资的逻辑正在发生范式转变：在 AI 时代，大额资本可直接转化为算力和模型性能阶跃，超额回报将由顶级创始人驱动的营收端指数级突破所主导。 ## Key takeaways - `00:00` AI 生态的核心格局并非非此即彼的单选题，开源模型与前沿闭源实验室、芯片巨头与下游应用层呈现出并存扩张的“And”格局。 - `05:00` 算力供给呈现极度紧缺状态，全行业 4 吉瓦甚至更多容量上线即被迅速消化，新增数据中心订单排期已延后至 2028 年。 - `05:00` 推理阶段的计算规模扩展法则（Inference-time reasoning）表明，模型推理过程中的思考越多输出表现越好，这意味着推理算力消耗在未来只增不减。 - `13:17` 前沿模型实验室的产品精力受限于通往 AGI 的核心路径，垂直行业应用（如法律 AI 领域的 Harvey）凭借极致的最后一英里交互和沉重的线下销售交付构建起坚固护城河。 - `18:02` 随着总 Token 消耗量爆发式增长 50 至 100 倍，应用层为捍卫毛利率，会自然将高频且成熟的日常业务分流至低成本的 N-1 代或开源模型上。 - `20:00` 即使技术方案完全成熟（如 AI 客服已碾压传统方案两年），企业级落地的物理法则与销售交付周期依然限制着技术的快速扩散。 - `21:05` 消费端 AI 目前大多数成年人仍处于把 ChatGPT 当搜索引擎用的“拟物化阶段”，下一代重大范式转变将是从被动响应迈向原生多模态的主动型私人助手。 - `23:18` 消费级“杀手级产品”拥有覆盖数十亿人的终极市场空间，未来商业模式将由高阶订阅与尚未被发明出的全新原生内容推荐/广告形态共同驱动。 - `27:26` Waymo 等系统的实测安全性已达人类驾驶员的 10 至 14 倍，使得继续让人类驾驶在道德上面临严峻审视；无人化降本将重构出行价格弹性并带来至少 10 倍的网约车市场扩张。 - `30:00` 实体机器人正在工厂与工业产线等可控场景中通过“首发客户部署 + 真实作业数据回流反哺”建立飞轮，未来有望迎来超越语言模型体量的“ChatGPT 时刻”。 - `35:00` 私人市场的投资回报有一半甚至更多比例产生于 C 轮及以后的成长轮次，且 AI 领域的资本能直接转化为模型能力的规模化提升，打破了上一代 SaaS 资本效率递减的宿命。 - `40:00` 成长型投资的超额收益几乎全部源于营收端的超预期爆发，顶尖创始人（如 Databricks 的 Ali Ghodsi）往往具备跨越产品周期并打破传统财务模型上限的能力。 - `43:20` 在全行业注意力高度集中的社交媒体时代，“氛围感（Vibes）”与掌舵者的叙事能力已成为降低融资稀释、吸纳全球顶级人才和建立客户信任的硬核商业杠杆。 - `45:00` 兼具“弥赛亚式创始人”个人号召力与大众热爱产品的企业（如 Palantir、SpaceX、Tesla）在资本和商业端享有极高的估值溢价与不可替代性。 - `50:00` 科技行业占整体资本市场的比重目前仅略高于三分之一，远未达到历史上铁路或金融行业巅峰时期的 80% 份额，科技市值仍具备巨大的渗透与膨胀空间。 ## Detailed walkthrough ### `00:00` AI 生态的核心命题不是二选一，而是“共存” Jack Altman 提到他的合伙人 Eric Vishra 最近在播客《Invest Like the Best》上表达过一个观点：“一切都会成真”（it's all going to work）。市场总在纠结各种非此即彼的对立：是开源模型胜出还是前沿闭源模型胜出？是英伟达独大还是新兴芯片公司崛起？但 Eric 认为大家问错了问题，这并不是单选题，而是两者都会成功。a16z 合伙人 David George 完全赞同这一视角，并将其精炼为一个字——“And”（既……又……）。在 David 看来：“If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and.”（如果你的问题前提是这个会成功还是那个会成功，AI 领域的答案很可能是‘二者皆然’。）现在是有史以来作为科技市场投资者最令人兴奋的时代，尽管紧跟这一节奏极其困难。 从基础设施建设的维度来看，当前 AI 的资本开支规模已经超越了当年的铁路建设时期，占 GDP 的比重达到约 3%。David George 预测，这一支出规模未来还会增长十倍。这并不意味着它会占到 GDP 的 30%，因为这是历经数年拉动的全社会 GDP 增长过程，但全行业最终将投入数万亿美元。眼下，OpenAI、Anthropic 以及像 Cursor 这类企业展现出了商业史上前所未见的最快增长速度：它们合计创造了约 120 亿美元的年化收入，每个月新增的收入规模甚至超过了除亚马逊 AWS 之外的所有云厂商（Hyperscalers）。 ```mermaid flowchart TD accTitle: AI 商业化现状与未来渗透漏斗 accDescr: 展示 AI 从目前极小规模的核心开发者向海量知识工作者扩展的层级结构。 subgraph Current ["当前收入支柱（极度集中）"] TopDev["头部顶级程序员（约 100 万人）<br/>贡献绝大部分极高支出"] AllDev["全球开发者人群（约 3000 万人）<br/>贡献当前 B2B 核心收入"] TopDev --> AllDev end subgraph Expansion ["未来扩张空间（渗透率不足 5%）"] EarlyAdoption["早期延伸职能（法律等领域，落后约 12 个月）"] BroadKnowledge["全球广义知识工作者（约 15 亿人）<br/>市场潜力呈幂律级放大"] AllDev --> EarlyAdoption EarlyAdoption --> BroadKnowledge end ``` 这种惊人增长的背后存在一个极度反直觉的现实：如此庞大的商业化收入，仅仅建立在极少数付费人群之上。全球虽然已有上亿乃至十亿级别的消费者在使用 AI，但当前 B2B 领域的绝大部分收入其实仅来自于全球约 3000 万开发者，而在开发者群体内部又遵循严苛的幂律分布（Power Law）——其中最顶尖的 100 万程序员贡献了极其庞大的开销。相比于全球 15 亿的知识工作者大盘，AI 目前在企服经济中的渗透率甚至不足 5%。除了编程和文本写作，以及刚刚起步、大约落后编程 12 个月的法律行业之外，广义知识工作几乎尚未被触及。需求端呈现完全无法满足的状态，这意味着未来五年的市场空间将极其巨大。 --- ### `05:00` 算力供给瓶颈与企业端 ROI 的终极验证 在供给侧，瓶颈无处不在。供应链中每个环节的卡点导致数据中心容量极其紧缺，甚至要排期到 2028 年才能拿到新增容量。Jack Altman 追问，目前行业容量大约为 4 吉瓦（GW），如果立刻上线 8 吉瓦的容量，是否也会被瞬间吞噬？David 肯定地表示这些容量都会被消耗殆尽，无论是 OpenAI、Anthropic 还是 xAI（SpaceX/Musk 体系）的算力，只要上线就会被完全利用。 为了审视这一乐观假设的边界，Jack 提出：在什么情况下，算力从 4 吉瓦扩张到 20 吉瓦时，Token 会不再被完全消化、甚至价格崩溃？David George 指出可能存在两种打破预期的极端假设，但他认为发生的概率极低： 1. **编程经验不可复制**：编程领域的突破只是特例，由于企业工作流程与商业系统难以打通，AI 无法有效渗透其他白领知识工作。David 认为这只是数据收集与模型应用工程的问题，而非能力壁垒。 2. **算法效率产生跃迁式突破**：团队可能在模型算法上取得质的飞跃，使训练与推理所需的算力大幅降低数个数量级。人类大脑构建认知体系所需的数据量确实比当前大模型训练少两到三个数量级，但即使训练效率突破，推理（Inference）市场本身的规模也远大于训练；而且目前所有的工程事实均证明：“The more you reason, at the model, at the time of inference, the better the results get.”（你在推理阶段让模型进行的思考推导越多，输出的结果就越好。）因此推理端的算力消耗只增不减。 Jack Altman 进一步探讨了近期市场上的核心担忧：目前的 AI 营收绝大多数归属于“Token 的卖家”（基础设施与模型提供方），而非“Token 的买家”（最终使用模型的企业）。如果买方投入巨资调用模型却无法转化为实打实的生产力飞跃与企业收益，这种资本循环就会断裂。David 回应称，这种悲观论调的前提是“更高效率地写出更多代码不会带来切实的商业价值”，而这一前提根本立不住脚。尽管一年前曾有研究声称使用 AI 编程会让开发效率下降 20%，但那是 GPT-4 早期甚至更早版本的情况。如今全球顶尖程序员的生产方式已经彻底转变，像 Stripe 这样的前沿科技公司 CEO 们都明确感知到写代码效率提升带来的极高回报。尽管像宝洁（Procter & Gamble）这样的传统巨头目前可能还没有获得巨大的效率提升，但像 Databricks、Palantir 以及各大前沿实验室正在通过深度介入的“贴身服务”（hand-holding）模式，帮助传统企业逐步释放出巨大的商业价值。 - 关键指标与事实整理： - AI 基础设施目前占 GDP 比重约为 3%，未来整体规模预计将实现 10 倍增长（许多万亿美元级别）。 - OpenAI、Anthropic 与 Cursor 等核心玩家年化收入合计约 120 亿美元，月增量超过除 AWS 外的多数云厂商。 - 全球知识工作者总数约 15 亿人，而当前驱动绝大多数商业化收入的仅为 3000 万程序员中的前 100 万高净值核心开发者。 - 数据中心电力供应严重受限，当前上线容量约 4 吉瓦（GW），全产业链供给短缺导致新订单排期延至 2028 年。 ### `10:00` 企业的资源聚焦与前沿模型实验室的绝对主导 在思考现代科技企业的资源分配时，David George 和 Jack Altman 探讨了一个核心分歧：企业的技术努力究竟应该投向前台（开发更多新产品）还是后台（优化现有业务的效率）？David George 坚信，当前世界上最优秀的公司（例如 Stripe）都极度聚焦于前台。在他看来：“如果聚焦于财务端，你对自己的未来抱有开阔的视野，而如果只关注成本端，就像是在做空自己的未来。”（“It's like latent. So it's like shorting your own future if you are just focused on the cost side, whereas you're taking an expansive view of what you can become if you're focused on the finance side.”）如果押对下一个产品，收入可能会带来 10 倍乃至 100 倍的增长，而成本端即便极致优化，空间也是有限的，那种机会随时都处于潜伏状态。 与此同时，两人谈及 OpenRouter 等项目的思路，认为未来经济将在美元和 Token 之间来回摆动，Token 将成为一种全新的货币形态。当人们争论开源模型与前沿实验室（Frontier Labs）谁输谁赢时，David George 强调这并非零和博弈——只要整体 Token 使用量增长 100 倍，整个生态的所有层级都会变得极其庞大。在讨论应用层之前，必须先看清前沿实验室的绝对统治力：仅排名前三的实验室就占据了市场 95% 以上的收入（约 1250 亿美元规模），而下游生态与之相比几乎只是微不足道的尾数。单是 OpenAI 和 Anthropic 两家公司，融资规模就超过了 3500 亿美元，拥有无可比拟的规模优势。 重度用户的实际选择（revealed preference）证明，前沿模型依然是绝对首选，即便前沿 Token 的价格贵上 10 倍，用户也心甘情愿买单。这一方面是因为解决问题所释放的消费者剩余极高；另一方面，前沿实验室在第一方产品（First-party Products）的打磨上非常出色，例如 Codex、Claude Code、Cursor 等工具，模型与运行框架（harness）紧密结合，提供了难以替代的极致体验。Jack Altman 补充指出，随着模型能力进一步跃迁，前沿模型所承担的任务价值将更为惊人：“如果编程值 X，那攻克癌症值多少？攻克癌症的价值将是一个天文数字。” - 前沿实验室吸纳了企业级市场超过 95% 的资金，仅 OpenAI 与 Anthropic 两家累计融资即达 3500 亿美元。 - 用户甘愿为前沿模型支付 10 倍溢价，首要动力来自于“工作被彻底搞定”带来的巨大价值剩余。 - 编程领域对前沿能力的需求尤其刚性，第一方工具将模型逻辑与产品脚手架紧密闭环。 --- ### `13:17` 前沿实验室的产品边界与垂直应用的护城河 面对前沿实验室不断推出第一方产品的势头，成长阶段的应用层投资逻辑该如何演变？David George 提出了自己的研判框架：前沿实验室绝不会在所有领域全面出击，它们的第一方产品会高度集中在两个核心战场：一是编程本身及其“辐射半径”（blast radius）；二是面向全球 15 亿知识工作者的通用横向工具（类似于 Microsoft Office 和 Google Workspace 系列）。从实验室与数据供应商的合作动向来看，他们的精力正牢牢锚定在这些横向通用领域。 在通用知识工作工具之外，垂直行业软件迎来了巨大的爆发期。以 a16z 领投的法律 AI 独角兽 Harvey 为例，法律市场正处于爆发拐点。虽然社交平台 X 上的舆论曾充满怀疑——18 个月前人们还在拿幻觉问题大做文章，甚至当柯克兰律所（Kirkland & Ellis）宣布斥资 5 亿美元自建技术栈时引发行业担忧——但如今端侧大客户已经开始反向倒逼律所：“客户会说：‘我们既看重交付品质，也看重成本控制，因此你们必须使用 Harvey。’”（“They're like, We care about it for product and we care about it for cost. And so you must use it.”）法律的普及节奏大概比编程晚 12 个月，其本质是具有数学严密性、高度逻辑化且判例完备的语言工作，推理模型（Reasoning Models）的突破恰恰与法律推理完美契合。 ```mermaid flowchart TD accTitle: Frontier Labs Focus vs Application Layer Moats accDescr: 前沿大模型实验室的精力分配与垂直应用层护城河架构图 subgraph FrontierLabs["前沿大模型实验室的核心阵地"] direction TB Code["编程开发及周边辐射区<br/>(Codex, Claude Code)"] Horizontal["横向知识工作者平台<br/>(对标 Office / Google Apps)"] end subgraph AppMoats["垂直行业应用护城河 (例如 Harvey / Legora)"] direction TB LastMile["最后一英里交互细节<br/>(极度贴合专业工作流)"] GTM["线下高难度 GTM 与交付<br/>(现场部署、合规与客户信任)"] end FrontierLabs -.->|"难以顾及 / 优先级靠后"| AppMoats ``` Jack Altman 指出，所有伟大的平台型企业（如微软、Salesforce）都明白一个道理：如果平台亲自下场吞噬生态中的零碎应用，就会彻底失去生态开发者的信任，真正的平台必须让构建于其上的产业产出远超平台自身的营收。垂直行业虽然不是实验室通往 AGI 的必由之路，但本身却是价值数百亿美元的庞大市场。垂直应用的壁垒不在于基础算法，而在于两点：极度贴合专业用户习惯的最后产品细节，以及极其沉重的线下 GTM（Go-to-Market）拓展与合规交付能力。前沿实验室受限于精力和定位，不可能在每一个细分垂直领域面面俱到。 - 法律 AI 领域的渗透进程目前大约比编程工具滞后 12 个月，推理模型的出现成为关键拐点。 - 即使柯克兰律所宣称投入 5 亿美元自研系统，也没有阻挡垂直标准化软件在企业端的刚性渗透。 - 软件护城河正重构为两项能力：打磨到极致的最后一英里产品体验，以及深扎行业底层的线下销售与交付能力。 --- ### `18:02` “生态全赢”格局与 N-1 代开源模型的成本分流 不仅在垂直领域，甚至直接处在编程“辐射半径”核心区域的产品，依然展现出惊人的生命力。David George 提到，他们是 Replit 的最大投资方，同时也看好 Lovable 和 Cursor 等产品。即便前沿实验室的 Codex 或 Claude Code 原生具备网页构建和代码生成能力，但 Replit、Cursor 等平台凭借极致的独立交互形态和开发者体验，业务表现极其抢眼。这印证了“一切都会行得通（It's all going to work）”的投资判断：市场空间不仅能支撑通用底座，还能让处于辐射范围内的优秀独立工具大放异彩。 ```mermaid flowchart LR accTitle: Frontier vs N minus 1 Workload Distribution accDescr: 高价值前沿模型与低成本开源模型在工作负载中的分流结构 HighValue["高价值 / 复杂推理任务<br/>(前沿模型 Frontier)"] CostDriven["高频 / 成本敏感型流水线<br/>(开源 / N-1 代模型)"] Workload["爆发式增长的 Token 需求 (50x - 100x)"] --> HighValue Workload --> CostDriven CostDriven -->|"直接关乎毛利率"| AppEconomics["应用层商业模式优化<br/>(例如成熟的客服工单处理)"] ``` 在模型消耗层面，未来也将呈现鲜明的双轨并进格局： 1. **对前沿技术的渴求毋庸置疑**：高价值场景与复杂任务将持续依赖最尖端模型。 2. **对 N-1 代与开源技术的胃口将空前庞大**：当整个社会的 Token 消耗总量增长 50 倍甚至更高时，算力成本对应用层公司而言就是毛利率（Gross Margin）的命脉。一旦前沿 Token 与次级 Token 出现高达百倍的价差，企业便会极其理性地精打细算。 正如 Jack Altman 指出，次世代的 N-1 代开源模型很快就会超越此前的顶尖模型（如 GPT-5 或同代前沿）。这种趋势在应用层最前沿已经显现——对于绝大多数日常场景（如标准客户支持工单），早在两三年前就已经完全不需要最高规格的前沿模型来处理了，分流与降本正在自然发生。 - 前沿编程工具与独立应用（Replit、Cursor、Lovable）正在证明各自独特的共存空间。 - 随着总使用量呈几十倍放大，应用层公司为了捍卫毛利率，会主动将成熟工作流迁移至低成本的 N-1 代或开源模型上。 - 算力分层已经常态化：极高价值推理锁定前沿模型，常规业务流水线则由性价比更高的次级模型承载。 ### `20:00` 技术扩散的现实与企业级落地的物理法则 谈及 AI 基础设施的抽象层争夺战，David George 认为应用层公司、模型实验室的原生产品，以及像 Databricks、Palantir 或 Snowflake 这样的数据巨头都会参与竞争。但他强调用户最终会采用这些模式的各种变体，而每一种路径都有生存空间。Jack Altman 随即提出了一个鲜明的反差案例：客户支持（Customer Support）。在客服领域，技术在相当长一段时间内已经“好得绰绰有余”，但在整个市场的渗透率依然极低。这证明了即使技术完全就绪，真正的技术扩散（diffusion）仍需经历漫长的数年。 David George 赞同这一观点，指出尽管过去两年来 AI 在客服场景中完全符合传统企业软件“更好、更快、更便宜”（better, faster, and cheaper）的标准，相关创业公司的增长速度也打破了历史纪录，但企业软件仍需遵循不可违背的“物理法则”。销售流程需要推进，系统必须实施，还要处理复杂的脏乱集成、企业合规规则与品牌指引。正因如此，前沿模型实验室绝不可能去构建这些繁琐的企业销售与交付团队，这为垂直领域的创业公司留出了巨大的生态位，形成了一个经典的“实验室与应用共存”（and）的局面。 - **抽象层竞争者**：包括 AI 原生应用公司、大模型实验室的第一方产品，以及 Databricks、Palantir、Snowflake 等成熟数据平台。 - **客服渗透率反差**：技术能力虽已彻底碾压传统方案两年之久，但实际企业渗透率极低，说明落地瓶颈在于销售周期与系统集成。 - **企业落地的“物理法则”**：必须面对定制化集成、合规安全审核以及品牌准则要求，底层大模型实验室无力亲自构建这种服务体系。 ### `21:05` C端 AI 的拟物阶段与下一代“主动型助手” 话题转向消费者（Consumer）市场。Jack Altman 回忆道，ChatGPT 的爆发不仅拉开了整个人工智能新时代的序幕，也是第一个真正席卷大众的消费级时刻，让听了十年 AI 概念的普通人第一次切身体会到前所未有的产品体验。如今四年过去，David George 观察到，虽然已有数以亿计甚至超过十亿的用户在重度使用 ChatGPT，但对绝大多数成年人而言，其使用方式仍然处于“拟物化（skeuomorphic）模式”——大家仅仅把它当成一个更加强大的搜索引擎。 与成年人不同，下一代年轻人展现出了完全原生的使用习惯。Jack 提到自己 10 岁的儿子已经能用 AI 写出完整的说唱歌曲并制作音乐，David 直言“这不是插曲，这就是未来（This is the future）”。年轻一代从未经历过搜索引擎的思维定式，他们天然地向 AI 索取任何问题的答案或直接交付工作成果。David 预言，消费端最大的范式转变将是从“被动响应（reactive）”演进为“主动执行（proactive）”。随着语音交互能力因摆脱弱模型依赖而支持打断与自然对话，以及能够代表用户采取行动的自动化 Agent 技术不断成熟，消费级产品必将重塑。 ```mermaid flowchart LR accTitle: 消费级 AI 范式转变路径 accDescr: 描绘从传统的被动响应搜索替代品到具备多模态与自主行动能力的主动型私人助手的演进路线。 subgraph Phase1["当前阶段：拟物化搜索替代"] A["文本交互 / 简单问答"] --> B["被动响应 (Reactive)"] end subgraph Phase2["下一代阶段：原生多模态"] C["自然语音交互 (支持打断)"] --> D["多模态感知"] end subgraph Phase3["终极愿景：通用私人助手"] E["跨系统自主行动 (Agent)"] --> F["主动式协助 (Proactive)"] end Phase1 --> Phase2 --> Phase3 ``` - David George 尖锐地指出了当前普通用户的局限：“We're in skeuomorphic mode for the vast majority of consumers, which is they are doing what they used to do on a search engine inside ChatGPT.”（对于绝大多数消费者来说，我们仍处于拟物化模式中，他们所做的只是在 ChatGPT 里面做以前在搜索引擎上做的事。） - **原生代际差异**：从未被传统搜索训练过的新一代少年儿童，直接将 AI 视为能够交付最终成品的生产与创作伙伴。 - **语音与行动能力的突破**：早先的语音交互因依赖弱模型而存在假象，现已实现支持自然打断的拟人化体验；以 OpenClaw、Grok 等为代表的技术正推动其向主动行动阶段迈进。 ### `23:18` 消费级商业模式：被低估的“杀手级产品”与广告形态革命 David 观察到当前科技界存在一个有趣的盲区：由于编程工具（Coding）爆发式的收入增长，所有人的目光都集中在代码领域，几乎没人讨论消费端。然而，根据历史技术周期的规律，最具价值的市场皇冠始终属于能够覆盖数十亿人群的“杀手级消费产品”。这些产品将产生巨大的消费者剩余（consumer surplus），并孕育出世界上最强大的商业模式。 对于未来消费级 AI 的盈利形式，David 认为将是订阅制与全新广告模式的结合。正如在报纸分类广告时代向人们描述移动端的“信息流（Feed）”会让人难以理解一样，AI 时代的原生内容消费形态和相匹配的原生广告格式目前尚未被真正发明出来。当 AI 真正演化为每个人都离不开的、可信赖的个人助手时，即便 90% 的价值作为剩余留给消费者，仅凭剩下的 10% 价值捕获，就足以让数以亿计的用户心甘情愿支付高额费用。David 坦言自己正在积极尝试通过自建自动化流程让自己变得更“原生”，虽然目前 10 个自动化中只有 2 个能带来实质价值，但随着模型能力在未来 5 到 10 年内提升 10 到 100 倍，私人助手的潜力将彻底释放。 - David 强调了消费市场的终极规模：“But if historical technology markets are in a guide, the big prize is going to be the killer consumer products because you can address billions of people with them.”（如果历史技术市场能够提供任何指导的话，终极的大奖将是杀手级消费产品，因为你可以用它们服务数十亿人。） - **商业模式构想**：由高客单价的高阶订阅制与尚未被发明出来的“下一代原生广告/推荐格式”共同驱动。 - **价值分配模型**：10 倍的价值将转化为用户端的体验剩余，而留在企业端的那 10% 捕获份额，足以支撑起体量庞大的巨头。 ### `25:31` 科技浪潮叠加：从 25 万亿美金到自动驾驶的颠覆性拐点 谈及个人消费级自主技术，Jack Altman 分享了他新买特斯拉后对全自动驾驶（FSD）的震撼体验：输入目的地后，汽车就能自动完成全行程并解决各种复杂状况，这种体验让他深感大多数没体验过特斯拉或 Waymo 的普通人根本没有意识到自动驾驶已经多么成熟。David George 顺此将视角拉升至更大的宏观技术浪潮，指出 AI 的耀眼光芒在当下掩盖了其他正在并行的重大技术革命。 回顾上一代由智能手机、社交网络、电商、SaaS 以及云计算协同驱动的产品周期，科技界创造了高达 25 万亿美元的市值。而如今摆在眼前的，是人工智能、全自动驾驶、通用机器人、AI 生物医疗以及以国防现代化为代表的“美国活力（American Dynamism）”赛道的罕见叠加。David 确信，这组新浪潮创造的总价值将远远超越过去的 25 万亿美元，当前正是身处科技与投资市场最令人兴奋的时代。 ```mermaid flowchart TD accTitle: 科技周期市值浪潮对比 accDescr: 对比上一代移动与云计算浪潮所创造的25万亿市值与当下AI及物理自主系统叠加浪潮的潜力。 subgraph PreviousCycle["上一轮超级周期 (~25 万亿美元)"] direction LR P1["智能手机 + 社交网络"] P2["电子商务"] P3["SaaS + 云计算"] end subgraph CurrentCycle["当前超级周期 (远超 25 万亿美元)"] direction LR C1["人工智能 (AI)"] C2["全自动驾驶 (Autonomy)"] C3["具身智能与机器人 (Robotics)"] C4["AI + 医疗生物 (Bio/Health)"] C5["国防现代化 (American Dynamism)"] end PreviousCycle -.->|"范式跃迁与规模倍增"| CurrentCycle ``` - **上一代技术周期的丰碑**：移动通信、社交、电商与云基础设施共同孕育了约 25 万亿美元的市场规模。 - **多技术并发的黄金时代**：AI、自主驾驶、机器人、生物医疗与国防现代化的交汇，构成了前所未有的价值爆发期。 ### `27:26` 自动驾驶的经济学账本：10倍市场扩张与“不让人类开车”的道德必然 聚焦到自动驾驶领域，David 强调汽车工业与日常出行是全球最大的支柱产业之一，但自动驾驶的潜力被严重低估了。回顾出租车时代，其行驶里程占全美总行驶里程的比例不足 0.1%；Uber 和 Lyft 等网约车平台的出现不仅革新了商业模式，更将市场规模扩大了 10 倍（例如旧金山出租车市场原本仅为 1 亿美元，网约车进驻三年内便激增至 10 亿美元）。David 预言，自动驾驶在网约车领域将迎来至少又一次 10 倍的市场膨胀。 从安全性与道德伦理来看，现有的充分数据表明，Waymo 等自动驾驶系统的安全性已经是人类驾驶员的 10 到 14 倍。面对这一悬殊差距，Jack Altman 直言继续让人类驾驶汽车将演变为一种“野蛮行径”（barbaric），未来历史如果回望不开放自动驾驶进入经济体系的决策，必将给予严厉批判。而在经济账面上，由于出行市场具有极高的价格弹性，自动驾驶完全有清晰的路径将成本压缩到现有网约车甚至私家车之下，从而彻底重构交通出行经济学。 | 出行维度 | 传统私家车拥有成本 | 现有网约车 (Uber/Lyft) | 自动驾驶出行 (Waymo/Tesla 愿景) | |:--- |:--- |:--- |:--- | | **每英里成本** | 约 0.80 美元（全包折旧、保险、油费） | 约 2.00 美元以上 | 极具弹性，目标成本将远低于当前网约车并逼近私家车 | | **安全性水平** | 基准线（事故与伤亡主要来源） | 与普通人类驾驶水平相当 | 比人类驾驶员安全 **10 至 14 倍** | | **市场规模演进** | 占据绝大多数里程比例 | 较传统出租车时代实现 **10 倍扩容** | 预计在网约车基础上再实现**至少 10 倍扩张** | - David George 断言自动驾驶将重现爆发式扩张：“Autonomous driving, we're going to look back 10 years from now and say, oh, there was another obvious 10xing that would happen there. At least, just on the ride-hail side.”（十年后回过头看自动驾驶，我们会说，哦，那里显然又发生了一次 10 倍的扩张。至少在网约车这一端是如此。） - **安全与伦理反转**：自动驾驶系统被证实比人类司机安全 10 到 14 倍，使得阻碍自动驾驶落地的行政与社会阻力在道德上变得站不住脚。 - **高弹性需求曲线**：每英里综合出行成本一旦通过无人化从 2 美元以上压低至私家车的 80 美分以下，将引爆巨大的增量出行需求。 ### `30:00` 自动驾驶的万亿美元市场与机器人领域的“ChatGPT 时刻” David George 和 Jack Altman 深入探讨了自动驾驶与实体机器人的商业潜力与发展前景。在自动驾驶出行服务方面，David 指出当前的自动驾驶系统安全性已经达到人类驾驶的 14 倍，并且很快就会达到 20 倍，同时成本将大幅降低，带来超过 10 倍的综合体验提升。当谈到个人乘用车市场时，全自动驾驶的价值同样极其庞大：仅在美国，每年新车销量约为 1,700 万辆，存量汽车保有量约 2.5 亿辆。Jack 询问如果私家车拥有全自动驾驶功能消费者愿意为此支付多少，两人一致认为至少在 10,000 美元以上，无论采取一次性买断还是按月订阅形式，一旦用户体验过自动驾驶就很难再倒退回手动驾驶时代。放眼整个西方世界乃至逐步降低成本普及到发展中市场，David 直言在未来 50 年内，整个交通出行结构将彻底重塑，甚至“没有人会再自己开车了”。 尽管像 Waymo 这样的无人出租车在旧金山等重点城市给人的感觉已经无处不在，但实际上全美目前的 Waymo 车辆总数还不足 10,000 辆。该领域最终不会容纳数十家竞争者，除了 Waymo 和 Tesla 之外，只会有少数几家新锐玩家脱颖而出并瓜分这一万亿级市场。随后，两人将话题转向实体机器人。David 和 Jack 均认为，虽然离家庭保姆机器人进门可能还比较遥远，但在工厂、仓库等环境可控、任务重复度高、安全性有保障的工业场景中，机器人技术正在飞速落地。例如 a16z 领投了由 Rivian 创始人 RJ Scaringe 发起的新机器人项目（Mind Robotics），该项目直接将机器人部署在 Rivian 的工厂流水线上参与整车装配，从而建立起“自带首发客户 + 获得即时收入 + 真实作业数据反哺模型”的闭环飞轮。 ```mermaid flowchart LR accTitle: 机器人与自动化工厂的数据闭环飞轮 accDescr: 展示实体机器人在工业场景中落地，并通过现场部署与真实工作数据反哺模型的商业循环机制。 subgraph S1["工业落地"] A["工厂装配现场<br/>(如 Rivian 产线)"] B["执行高重复/安全任务"] end subgraph S2["商业与数据飞轮"] C["即时产生商业营收"] D["真实作业数据回流"] E["驱动模型迭代进化"] end A --> B B --> C B --> D D --> E E -->|"提升作业能力"| A ``` - 自动驾驶安全性已达人类基准的 14 倍，未来可达 20 倍，整体综合成本下降将带来 10 倍以上的价值跃迁。 - 全美年新车销售量达 1,700 万辆，汽车总保有量达 2.5 亿辆，按每台车 10,000 美元的全自动驾驶软件溢价估算，市场体量极为惊人。 - 目前全美实际运营的 Waymo 车辆总数不足 10,000 辆，规模仍处于极早期爆发前夕。 - David George 预测实体机器人将迎来自身的质变节点并超越纯语言模型：“我们将在机器人领域迎来类似 ChatGPT 的时刻。我不知道具体什么时候发生，但我希望是在未来五年之内，届时它的市场规模会比语言模型还要大。” - 面对颠覆性技术对就业市场的冲击，Jack 和 David 认为虽然需要关注劳动力转型，但正如过去 70 年间产生的前所未有的新职业（占今日职业的约 60%）一样，这场变革必然会带来前所未有的全社会生产力大爆发。 ### `35:00` 资本周期与产品周期：为什么成长型投资迎来范式重构 Jack Altman 提到 Benchmark 破天荒地募集了其历史上的第一支成长型基金（Growth Fund），David George 随即戏称这是“欢迎来到黑暗面（Welcome to the dark side）”。Jack 解释了 Benchmark 的底层逻辑：当今超级公司的价值创造和终局回报正在向成长阶段转移，超高回报倍数不再仅仅局限在种子轮或早期轮次。David 对此深表赞同，并给出了具体的数据支撑：公开市场与科技巨头构成了约 5 万亿美元的巨大体量，占据标普 500 指数的约 20%。a16z 过去的量化分析显示，私人市场的投资回报有一半产生于种子轮到 B 轮之间，而另外一半则来自 C 轮及以后的成长轮次。如果一家机构不深度参与成长轮投资，就等于放弃了整个私人市场回报的一半；更何况如今优秀公司保持私有化的时间大幅拉长，未来这一回报比例甚至可能倾斜至 70% 对 30%。 David 进一步剖析了 AI 时代与上一代 SaaS 时代的根本差异：当今的极端幂律（Power Law）使得成长型投资具有独特的飞轮效应。在过去以 Salesforce、Workday、ServiceNow 为代表的 SaaS 或消费互联网时代，盲目投入巨额资本（如软银愿景基金时期的激进打法）往往会适得其反——SaaS 公司拿到百亿美元无法有效转化，只能无节制雇佣销售或工程师，消费级公司则只能将资金白白浪费在 Google 和 Facebook 的买量广告上。但在 AI 时代，大额资本可以被直接转化为算力并用于训练前沿大模型，资本规模直接等同于模型性能的阶跃式提升。 关于投资周期，David 提出了将世界划分为“产品周期（Product Cycle）”和“资本周期（Capital Cycle）”的 1 到 10 分评价框架： | 评估维度 | 2010 年前后 | 2021 年高点 | 当前时期（AI 时代） | |:--- |:--- |:--- |:--- | | **产品周期打分** | **8 / 10**（云原生与移动互联网起步） | **1 / 10**（后移动与 SaaS 晚期，增量受限） | **9 - 10 / 10**（AI、自动驾驶、机器人全面爆发） | | **资本周期打分** | **较优**（SaaS 估值仅 4 倍市销率） | **极端非理性**（资产严重高估） | **~6 / 10**（估值略有溢价，但企业增速极快） | - “如果你不从 C 轮开始激进布局，那你绝对是在把一半的回报拱手让人，”David George 论述成长型基金的重要性时指出，随着企业更晚甚至更少过早上市，未来的回报重心将进一步向中后期集中。 - 谈及资本投入在 AI 领域的异化优势，David 明确指出：“在 AI 领域，这种情况非常独特，因为你往里面砸更多的钱，它就真的能单纯地规模化并且变得更强。这和以往的周期完全不同。” - 2021 年的产品周期在事后复盘时其实只有 1 分，属于 SaaS 与移动互联网增长消耗殆尽的晚期，但当时的资本市场却陷入了盲目狂热；而如今处于极佳的 9 到 10 分产品大周期，即使当前资本环境评分为 6 分左右，底层的技术巨浪依然足以在未来 10 年驱动出惊人的风险投资回报。 ### `40:00` 成长型投资的核心：相信创始人与打破财务模型的上限 David George 谈到当前的产品周期极为强劲，评分可以达到“9 分或 10 分（满分 10 分）”。在谈及投资哲学时，他引用了 Thrive Capital 等机构的一种观点：一种逆向思维是去相信完全没人相信的事情；另一种逆向思维则是去相信大家都相信的事，但你的坚信程度是其他人的十倍。这正是 a16z Growth 基金的座右铭——在成长期，花时间去思考“如果一切顺利，事情会发展到何种惊人的地步”。他强调，他们的业务逻辑是以合理的价格投资极其优秀的公司，而投资获得的超额收益（Alpha）永远来自营收端爆发式的增长，而不是纠结于利润率是否能挤出几个点。 ```mermaid flowchart TD accTitle: 超级创始人的复利飞轮 accDescr: 展示具有远见的创始人如何通过跨越产品周期为公司带来指数级增长并打破传统估值模型。 subgraph Core["核心资产"] A["远见型创始人<br/>（Ali Ghodsi / Elon Musk）"] end subgraph Execution["战略推进"] B["解决下一个核心命题<br/>（持续研发新产品）"] C["踩准技术周期红利<br/>（如全面拥抱 AI）"] end subgraph Outcome["市场表现"] D["营收持续逆势加速<br/>（打破财务模型预测）"] E["开创全新商业模式<br/>（如 Starlink / 轨道算力）"] end A --> B B --> C C --> D D --> E E -->|"反哺资本与信任"| A ``` 市场往往无法充分认知到某些公司能在极其漫长的时间跨度内持续高速增长。David 举了 Databricks 创始人兼 CEO Ali Ghodsi 的例子，并创造了“破模者（Model Busters）”这个词。a16z 早期由 Ben Horowitz 领投了 Series A，之后在设立成长基金之初，便在约 60 亿美元估值时领投了其成长轮。David 甚至开玩笑说，Databricks 融资轮次可能要像体育场阶梯座位排号一样一路融到“AA 轮”，一直保持私有化。七年过去，Ali 不仅持续孵化出新产品，在 AI 浪潮来袭时更使 Databricks 完美占据卡位，营收反而进一步加速。 这种对创始人的下注同样体现在 SpaceX 身上。在 a16z 投资时，Starlink 甚至还没有公开发布（GA），早期经济模型极差、看起来根本走不通，而如今它已成为人类商业史上最优秀的业务板块之一，甚至还在规划未来的轨道算力。正如 David George 所总结的： > “对于后期风险投资而言，那些创始人本身就是资产类别；如果支持他们，他们会把所有这些惊人的事情都琢磨明白并做出来。” - 成长型投资的超额回报几乎全部来自营收规模的超预期扩大，而非对利润率等边际指标的财务微调。 - 真正顶级的创始人拥有连续开创新业务的能力，其最终发展规模往往远远超越任何传统财务模型（Spreadsheet）所能推演的上限。 --- ### `43:20` 叙事即战略：为什么“氛围感（Vibes）”具有极高的商业价值 Jack Altman 指出，a16z 从创立之初就深刻理解“叙事（Narrative）”的威力。早在 2009 年媒体环境不同的年代，他们就懂得发声与讲故事的重要性；如今在社交媒体直接触达大众的时代，叙事声量更是被放大到了前所未有的程度。新公司上线精美短视频，后期巨头通过媒体掌控话语权，整个创投圈似乎都活在推特（X）的“氛围感（Vibes）”中。 David 分享了一个发生在 2018 年加入 a16z 前几个月的趣事：当时他与 Ben Horowitz 闭门讨论 Growth 基金的组建策略，David 提出了投资团队和组织架构构想。当 Ben 问他“你的营销策略是什么”时，来自传统保守机构的 David 回答说成长投资通常不需要营销策略，Ben 当场直截了当地斥责道：“那是我这辈子听过最蠢的狗屁想法（that’s the dumbest fucking idea I’ve ever heard）。”Ben 解释道，当年主流 VC 从不营销自己，而 a16z 率先做了且大获成功，如今做成长基金为何反而不用？ 这让 David 彻底认同了叙事与“氛围感”的硬核商业价值。他将氛围感直接挂钩为公司的竞争杠杆： 1. **更低稀释与更大估值溢价**：高估值直接降低融资稀释率，在算力和资本可以直接转化为产品竞争力的 AI 时代，充足的资金池本身就是战略武器。 2. **人才引力场**：全行业的优秀工程师与潜在员工天天活跃在 X 上，强大的叙事构成了招聘与留存顶级人才的核心吸铁石。 3. **客户层面的信任背书**：在技术颠覆期，强大的品牌声浪会让大企业高管坚信“找他们交付 AI 绝对不会犯错”。 --- ### `45:00` “弥赛亚式创始人”与超级叙事在 AI 时代的价值爆发 David 与 Jack 探讨了叙事在公开市场与非公市场的具体投射。Palantir 及其 CEO Alex Karp 就是最典型的范例。即使许多公开市场的散户甚至不清楚 Palantir 核心的“本体（Ontology）”到底代表什么，甚至普通散户单纯因为“好氛围（Good vibes）”而大量持有，其股票依然长期享有极高倍数溢价。如今随着商业化增长全线爆发，Palantir 迅速成为全美 CEO 眼中落地实施 AI 唯一值得信赖的合作伙伴。估值优势进而演化为雄厚的人才吸引力与资本弹药库。 ```mermaid flowchart LR accTitle: 弥赛亚式创始人的资本与业务放大效应 accDescr: 说明具备强大魅力的创始人如何通过叙事在资本、员工和客户三端形成正反馈循环。 subgraph Founder["领袖核心"] M["弥赛亚式创始人<br/>（Elon / Karp / Palmer）"] end subgraph Multipliers["三大杠杆"] C["资本端：极高估值与忠诚散户支持"] T["人才端：强大的招聘与留存号召力"] B["商业端：企业客户视其为行业指路人"] end subgraph Advantage["战略终局"] Adv["构建超强资本护城河与不可替代性"] end M --> C M --> T M --> B C --> Adv T --> Adv B --> Adv ``` Jack 提到了某对冲基金投资人的著名做空戒律：“我从不做空弥赛亚式的创始人（Messianic founder），也从不做空大众真正热爱的产品。”当一家公司同时拥有这两者时，其估值就能摆脱常理长期居高不下。David 指出，除了 Elon Musk、Alex Karp 以及 Anduril 的 Palmer Luckey 这样的创始人具备这种魔力外，那些默默无闻的普通 SaaS 公司 CEO 根本无法吸引到同等量级的注意力与资源倾斜。 叙事能力甚至已经渗透到组织中层。David 举例提到 OpenAI 的 Thibaut（Thibaut Sidaner），指出其在 X 上坦诚而直接的沟通方式，不仅在投资人和员工间营造出极佳的氛围，更在客户侧直接驱动了实际的采用与商业转化。David 总结道，现在他们每天都会提醒被投创始人： > “你必须直接发声并掌控叙事，你是公司的门面，必须讲出一个令人信服的故事；如果你不这么做，就算技术同样出色，大众也只会把目光投向别人。” - 目前整个科技市场几乎以月为周期，紧密维系在两到三家未上市 AI 巨头的实际表现与预期之上。 - 随着这些头部 AI 公司迈向 IPO，市场各方对于预期管理和叙事权力的博弈已经进入白热化阶段。 ### `50:00` 风投机构的规模上限与科技市值的未来扩张 Jack Altman 提出了最后一个问题。他提到自己之前分别在 Zoom 上采访过 Marc Andreessen 和 Ben Horowitz，能极其清晰地感受到他们身上那种宏大的野心、规模感以及思维格局。他开玩笑说，在节目录制前还聊到 Benchmark 的团队规模非常精简，而 Andreessen Horowitz（a16z）在团队体量上则非常庞大。鉴于未来创造的市值可能远不止 25 万亿美元，Jack 询问 David George：风险投资机构到底能做到多大？这种扩张是否存在瓶颈或上限？ David George 表示，a16z 内部一直在讨论这个问题，因为风投业务中既有具备可扩展性（scalable）的部分，也有不具备扩展性的部分。他引用了 Jack 的一句话并赞同道：“如果你想走得快，那就保持小巧；如果你想走得远，你必须做大。”（“If you want to go fast, be small; if you want to go far, you got to be big.”）尽管庞大的体量会带来更多的组织管理开销（organizational overhead），但 a16z 认为这能让他们走得更远。这种扩张既体现在投资团队和投资策略的规模化上，也体现在 a16z 组建的拥有 650 名员工的平台服务团队上，他们全力以赴赋能被投企业。 关于科技行业的总盘子究竟能有多大，David 指出，目前科技行业大概占整个经济体总市值的略高于三分之一。然而纵观历史，铁路、金融服务等支柱产业在其巅峰时期曾占据整个资本市场高达 80% 的份额。David 认为科技行业在整体市值中的占比完全没有停止上升的理由。全球市值排名前十的公司中，有八家都是起源于美国、曾获风险投资支持的科技巨头，而且在当前价位下它们甚至“看起来依然很便宜”，堪称人类商业史上最优秀的商业模式。紧随其后的是 OpenAI、Anthropic 和 SpaceX 等新一代巨头，它们的增长速度甚至更快，未来的体量也会更大；更不用说未来十年内即将到来的自主化与具身智能等重大浪潮中，许多核心玩家甚至都尚未入局。 - **历史产业占比对比**：过去铁路（Railroads）与金融服务（Financial Services）等行业在巅峰期可占到全市场市值的约 80%，科技行业目前约占三分之一强，仍有巨大的渗透空间。 - **顶级科技企业地位**：全球前十大最有价值公司中有 8 家是曾获 VC 支持的美国科技公司，兼具强大的盈利能力与估值性价比。 - **下一波巨头接棒**：OpenAI、Anthropic、SpaceX 等正在快速追赶，其爆发势头比以往巨头更迅猛，同时未来十年的新型趋势公司尚未完全登上历史舞台。 ### `52:30` a16z 的市场份额雄心与垂直赛道布局 谈到 a16z 自身的野心与市场份额，David George 坦言，如果从成长型投资（growth land）的真实市场框架来看，a16z 目前的市场份额大约在 20% 左右。机构的理想状态并不是盲目追求更高的整体百分比，而是“在 100 件最好的事情中投中前 20 件”，牢牢占据顶级资产中 20% 到 25% 的份额。因此，他们追求的不是无节制地扩大自身份额，而是顺应未来庞大的宏观市场增长大潮。正如 David 所强调的：“哪怕仅仅是跟上这股浪潮，本身就已经非常困难了。”（“Even keeping up with it is going to be hard.”） ```mermaid flowchart TD accTitle: a16z 投资赛道与宏观增长架构 accDescr: 展示 a16z 围绕前沿科技设立的细分赛道，以及通过顺应宏观科技浪潮锁定顶级资产份额的战略。 subgraph Macro["宏观趋势驱动"] M1["科技占全市场市值持续攀升"] M2["新一代 AI、算力与自主化技术爆发"] end subgraph Sectors["a16z 核心科技细分板块"] S1["底层基础设施 (Infrastructure)"] S2["应用层 (Applications)"] S3["美国活力 (American Dynamism)"] S4["生物医疗 (BioHealth)"] S5["加密资产 (Crypto)"] end subgraph Objective["机构核心战略"] O1["保持顶级资产 20%-25% 渗透率"] O2["650人平台服务团队深度赋能"] O3["捕捉并驾驭下一轮创新浪潮"] end Macro --> Sectors Sectors --> Objective ``` 为了实现这一目标，a16z 针对各个高度专注且差异鲜明的垂直科技领域建立了独立的专业团队，涵盖底层基础设施（Infrastructure）、应用层（Applications）、“美国活力”（American Dynamism，包含国防科技与先进制造等）、生物医疗（BioHealth）以及加密货币（Crypto）。各个领域的需求和打法截然不同，随着全新技术浪潮的不断涌现，a16z 致力于深度参与这些前沿变革，共同分享经济规模增长带来的巨大红利。对谈最后，Jack Altman 感谢了 David 的深度分享，节目在友好的互致谢意与官方免责声明中圆满结束。 - **市场份额策略**：重点并非单纯扩大管理资产份额，而是锁定头部市场中前 20% 至 25% 的最优质创业公司。 - **专注的细分阵地**：包括基础设施、应用层、American Dynamism、生物医疗、加密货币等离散板块，分别配备专业团队提供深度支持。 ## Notable quotes > `00:00` If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and. — David George > `05:00` The more you reason, at the model, at the time of inference, the better the results get. — David George > `10:00` It's like latent. So it's like shorting your own future if you are just focused on the cost side, whereas you're taking an expansive view of what you can become if you're focused on the finance side. — David George > `21:05` We're in skeuomorphic mode for the vast majority of consumers, which is they are doing what they used to do on a search engine inside ChatGPT. — David George > `23:18` But if historical technology markets are in a guide, the big prize is going to be the killer consumer products because you can address billions of people with them. — David George > `27:26` Autonomous driving, we're going to look back 10 years from now and say, oh, there was another obvious 10xing that would happen there. At least, just on the ride-hail side. — David George > `30:00` 我们将在机器人领域迎来类似 ChatGPT 的时刻。我不知道具体什么时候发生，但我希望是在未来五年之内，届时它的市场规模会比语言模型还要大。 — David George > `35:00` 在 AI 领域，这种情况非常独特，因为你往里面砸更多的钱，它就真的能单纯地规模化并且变得更强。这和以往的周期完全不同。 — David George ## Mentioned David George, Jack Altman, Eric Vishra, OpenAI, Anthropic, Cursor, Amazon AWS, xAI, Stripe, Databricks, Palantir, 宝洁 (Procter & Gamble), OpenRouter, Harvey, Kirkland & Ellis, Microsoft, Salesforce, Replit, Lovable, Snowflake, Tesla, Waymo, Rivian, RJ Scaringe, Mind Robotics, Benchmark, Workday, ServiceNow, SoftBank Vision Fund, Google, Facebook, Thrive Capital, Ali Ghodsi, Elon Musk, Ben Horowitz, SpaceX, Starlink, Anduril, Palmer Luckey, Alex Karp, Thibaut Sidaner, Marc Andreessen, Andreessen Horowitz (a16z) ## Video description a16z General Partner David George joins Jack Altman on Uncapped to make the case that many of the biggest debates in AI are framed the wrong way. Frontier models or open source? Labs or applications? David’s answer is often “and.” With AI adoption still concentrated among a relatively small group of heavy users, he argues there could be room for multiple layers of the stack to grow at once. David and Jack discuss why demand for compute continues to outstrip supply, why applications can thrive even as frontier labs expand into new products, and why consumer AI may still be at the beginning of its biggest shift, from reactive chatbots to proactive assistants that can act on our behalf. They also zoom out to autonomy, robotics, healthcare, and the next generation of technology companies, before turning to venture itself: why David believes today’s product cycle is unusually strong, how capital can accelerate AI companies in ways it couldn’t during the SaaS era, and why founders and narrative become increasingly important as companies scale. This conversation originally appeared on Jack Altman’s Uncapped. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/f7d0e381-40ba-4655-bdfe-afe618282935/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=f7d0e381-40ba-4655-bdfe-afe618282935&feed=JGE3yC0V)

[0:00] David George: If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and. [5:00] David George: The more you reason, at the model, at the time of inference, the better the results get. [10:00] David George: It's like latent. So it's like shorting your own future if you are just focused on the cost side, whereas you're taking an expansive view of what you can become if you're focused on the finance side. [21:05] David George: We're in skeuomorphic mode for the vast majority of consumers, which is they are doing what they used to do on a search engine inside ChatGPT. [23:18] David George: But if historical technology markets are in a guide, the big prize is going to be the killer consumer products because you can address billions of people with them. [27:26] David George: Autonomous driving, we're going to look back 10 years from now and say, oh, there was another obvious 10xing that would happen there. At least, just on the ride-hail side. [30:00] David George: 我们将在机器人领域迎来类似 ChatGPT 的时刻。我不知道具体什么时候发生，但我希望是在未来五年之内，届时它的市场规模会比语言模型还要大。 [35:00] David George: 在 AI 领域，这种情况非常独特，因为你往里面砸更多的钱，它就真的能单纯地规模化并且变得更强。这和以往的周期完全不同。

[30:00] David George: We will have a ChatGPT moment for robotics. [45:00] David George: A higher valuation means less dilution, more ability to raise capital, and you can use that as a weapon. [23:18] David George: If historical technology markets are any guide, the big prize is going to be the killer consumer products because you can address billions of people with them. [0:00] David George: If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and. [32:55] David George: The history of all these technology breakthroughs is that we find higher-order things to do.

[0:00] If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and. Like, it's just the most exciting time, you know, to ever be an investor and in the technology markets. Keeping up with it is going to be hard. Keeping up with it is going to be hard. Frontier models or open source? Labs or applications? David George thinks the answer to many of AI's biggest either or questions is simply and. general partner David George to discuss why he believes we're still early in the diffusion of AI across the economy. They get into why today's AI revenues are still concentrated among a relatively small number of heavy users, what happens as AI reaches the broader knowledge workforce, and why David expects frontier models, open models, and application companies to grow alongside one another. They also look beyond AI to autonomy and robotics, why consumer AI could for investing when both the companies and the markets they're entering can become much larger than expected. All right, David, I'm really excited to do this. I've been really looking forward to this. So thanks for coming by. Yeah, great to hang with you. So one of the topics that's been on my mind a lot that I think you've got really good thoughts on is something that my partner, Eric Vishra, said recently on the Invest Like the Best podcast. And it was basically, he was like, it's all going to work. And it was in the context of, you know, people question, is it going to be open source or frontier? And he's like, the open source is going to work. The frontier is going to work. Is it going to be NVIDIA or is it going to be new chips companies? And he's like, there's going to be both. And you can kind of like go on and on through, you know, the stack and AI. And Eric's point was basically like, people are asking the wrong question of just like, is it this or it's that? And he's like, it's both. And I'm curious how you think about this at Andreessen and sort of in your seat. Yeah, of course. So I am very, very closely aligned with Eric on this point that it's all going to work. I actually like the way that he framed it up. I framed it up slightly differently, which is just like the answer to all this is and. And so like if you're answering, if your question, if the premise of your question is, you know, is this going to be successful or that going to be successful? The answer in AI is probably and. I think it's helpful to start just like, where are we? So right now, from an infrastructure build out standpoint, we just surpassed the railroads. Right. It's like 3% now or something like that. Yeah, exactly. And look, it's probably going to get harder and harder to build data centers, which we could talk about later. But I think that's going to 10x. So right now, what's happening is basically… As a percentage of GDP, you think? No, I think hopefully this will drive GDP growth. And this is going to take many years. So it won't 10x in one year. So it won't be 30%. But we'll be spending many trillions. But we will spend many trillions, right? So all of this success, which is the greatest, fastest success we've ever seen in business, which is basically opening I and Thropic together. cursor, they're, call it like $120 billion of revenue, and they add more revenue per month than all the hyperscalers, except for Amazon. It's crazy. Now, all of that is on the back of just actually a small amount of people paying for stuff. So it's like 30 million coders in the world who are the vast majority of the revenue. There's 1.5 billion knowledge workers. But basically all of this, which is adding more revenue than the best businesses that have ever been created, is on the back of a small group of people. That's actually a really interesting point, that it's over a billion users at these things. And we think about it that way. But it's actually, you're saying it's 30 million are spending basically all money. 30 million. So there's a billion consumer users. And by the way, we should talk about consumer because we're nowhere in consumer right now, even though there's a billion users. But in the enterprise, basically, the vast majority of the revenue is coming from just a small amount of coders. And this is like in power terms. Even within that 30 million, I'm sure it's a power law. It's a massive power law. I'm sure you've talked to your companies. The top million of the 30 are spending. The top million of the 30 are spending just a massive amount. And so if you take an optimistic view of what that means, the potential is so much greater than what has actually diffused into everything beyond those million people. And so, again, there's one and a half billion knowledge workers, and we're just talking about B2B. So it's just a massive amount that is going to happen in what I think is going to be like the next five years. Right now, demand is totally insatiable. Right. in the enterprise, basically what's working, it's like coding, writing stuff, you're starting to see early signs in some functions beyond that. Like legal is one where we've kind of seen this take off. I think it's like 12 months behind. But for broad knowledge work, for the one and a half billion knowledge workers in the world, we're like nowhere. So you just start with that as the premise, which is, you know, we're already seeing like scale that's never been seen before. And, you know, we're like scratching the surface and diffusion. It's like less than 5% diffused into the B2B economy. That means we're going to get really, really big if we're right about this sort of diffusing and being applicable broadly. So that's so that's one piece on the supply side. we are massively, massively constrained. - Massively. - Like basically in everything in the supply chain. [5:00] So basically, you can't get data center capacity until 2028 because basically there's bottlenecks in every step of the supply chain. So that is where we are in the supply and demand. And I think it's important to lay that groundwork because... it is like the crux of why I think everything is going to work. Yes. Right? Like all of the sort of like steps in the ladder are going to work. Talking about the supply side for a second, so it's four gigawatts right now. Is your, is sort of like your kind of view at the moment that if we had eight gigawatts live and online, it would all get used as well? Yeah, it would all get used. Yeah, yeah, yeah. And look, maybe like it would be different prices or something. Like we could talk about pricing of what's happening for tokens. Yes. But yeah, I think, you know, if you, if you get like more OpenAI, Ananthropic or SpaceX AI capacity live, it generally gets used. So just to sort of like check the premises along the way, and obviously, as you know, I fully agree with what you're saying. But just to check it along the way, what would have to happen for it not to be the case that as gigawatts come online, all the tokens keep getting consumed? Is there any world where, is there anything that could update our priors, you know, that four gigawatts becomes 20 and they don't all get used at the same or even higher prices? I think there's two forms. where this could take hold. One is, um, Like, What happened with coding is anomalous relative to everything else. in B2B or knowledge work. And because of the difficulty of injecting commerce from knowledge work from companies, the rest of white-collar work just doesn't really have major impact. I think that is extremely unlikely. I think it's sort of like a data collection problem and an application of the models problem. The other would be, we do have teams that we've backed that are working on algorithmic breakthroughs. And so perhaps you would get an algorithmic breakthrough rather than having to consume massive, yeah, efficiency goes way up, both in training and in inference. I think that's unlikely. I think it's possible we could get breakthroughs in training. Like, I don't know, you've probably seen the stats on what it took you and I to build our knowledge base. It's like, you know, whatever, two or three orders of magnitude, less than what the models actually train on. So there seems to be an opportunity there. you still will then have inference, which is an absolutely massive market that's going to be much bigger than training anyways. And so, yeah, we could get efficiency gains in inference, but I think what we've learned so far is the more you reason, at the model, at the time of inference, the better the results get. And so I think that'll continue to hold in knowledge work broadly. So I struggle to find a way for this not to work. Now, if we went to from five to a hundred in a quick amount of time and the products aren't built to actually take that on, sure. But we're not going to have that problem because it's going to be really hard to get this stuff online. I'm curious what you think about this. So I I also can't easily see it, but the best factor that I've heard was a conversation I had with a friend where we were basically saying, all this stuff has come online. And so at first, a few years ago, there was the big build out and you saw we're putting hundreds of billions in at the time. And then the models get a little bit better and you start deploying them and you keep moving through the pipeline. And now you see revenues growing like crazy on the back of AI. But the point in this conversation was... Most of the revenues that we're talking about are sellers of tokens, not buyers of tokens. And so the next thing we kind of need to see is data around the buyers of tokens actually getting these huge efficiency gains. In other words, a hyperscaler. Yeah, like what's the ROI at the enterprise level? Yes. Of writing all this more code. Yes. Is your business getting better by writing all this? Yeah, or said another way, it would be a bad outcome if the sellers of tokens were making lots of revenue and the buyers of tokens were not making lots of revenue. Yeah, yeah, yeah. point. I think that whole argument is predicated on the fact that there won't actually be tangible, productive use of writing more code far more efficiently that is better. I just think that's fundamentally wrong. Yeah, and I haven't looked. I know a year ago, there was a study, or maybe even a little more, but something like a year ago, there was a study that in AI coding, it was actually, you got 20% less effective by using AI code, but this was before the big four, five. Yeah. My guess is if you re-ran that study, you would not see that anymore. And like the best coders in the world have like fully flipped over on this. And, um, you know, if you talk to like the CEOs of the cutting edge companies, I think they would tell you that they feel much more productive. Like I, you know, spending time with Stripe, like they feel like there's very high returns to them being able to be a lot more efficient in writing code. Um, you know, like is... Like Procter & Gamble getting big efficiency gains from what they're spending on AI today? Probably not yet, but I think they will. They probably need a lot of hand-holding. That's probably the answer. I think the labs and Databricks and Palantir, companies like that, are probably thinking about it the right way where you have to do a lot of hand-holding to produce value, but you will produce a lot of value. [10:00] For those companies. And then there's companies like Stripe. I don't know, one of the things that we're thinking a lot about right now is where is that effort actually going to? Is it going to front-end building more products? Or is it going to back-end trying to make your business more efficient? And the best companies in the world right now are singularly focused on front-end. You can produce 10x, 100x more revenue, potentially, if you nail the next product. Whereas on the cost side, you can only get so much more efficient. And that opportunity will kind of always be there. It's like latent. So it's like shorting your own future if you are just focused on the cost side, whereas you're taking an expansive view of what you can become if you're focused on the finance side. By the way, the open router idea was very good there too, just to be like, we're going to own this at the ALA as well. Yeah. That's very smart. Yeah, look, and I think they're right that the economy is going to oscillate between using dollars and tokens and it's going to be a new form of currency. Totally. So we're obviously big fans of those guys. Also to the point of it's all going to work. We talk about open source a lot in relation to the frontier labs. And I guess the point there is if the total token usage just goes up by 100x it just like these things are all going to get so big They all going to get so big And so I wanted to lay the groundwork first before making the case for which of the individual layers of the stack are going to be successful because I think it paints a picture where we're going to 20x or 50x over the next 10 years and just how much work is being done with this stuff. And by the way, we're just talking about enterprise right now. We should talk about consumer, but just an enterprise. So if you look at the Frontier Labs, Today, the Frontier Labs are like 95% plus of dollars, like the $125 billion or so across three of them. It's like they're consuming all of the revenue in the market, basically, everything else is a rounding error. They have raised more than the entire downstream ecosystem. Just OpenAI and Anthropic have raised $350 billion. So they do have massive scale benefits. And right now, what's revealed in the preferences of users is users are opting to use Frontier models. And I think there's a bunch of reasons why that's the case. Right now, people are willing to pay 10x more for Frontier tokens. I think in coding, that's definitely going to persist for some period of time. There reasons why you could see a case that people will cost and use below the frontier over time But right now the revealed preference is everyone who is a heavy user of this stuff wants to use the frontier. Also, it seems to me like if that happens, the frontier should be at even more valuable tasks that are worth even more. Even more. So like curing cancer. Yeah. If coding was worth X, what's curing cancer? Curing cancer is worth some astronomical. Yes, of course. And so right now, I think there's a couple reasons why people are using the Frontier stuff. One is just there's so much user or consumer surplus to having this work done for you that you're willing to pay the incremental amount for that Frontier. Does that make sense? Yes. Secondly, they've done a pretty good job with first-party products. So people really like using Codex and Cloud Code and Cursor. And I think increasingly we'll use stuff like GrockBot. that, it's going to be really valuable to use the first party stuff. There's a bunch of reasons why that's the case. They've done a good job with product. I also think they have very tightly coupled their harnesses with the models so that the experience you have with those products is actually very, very good. And so subbing out some other model where you don't have the harness and you don't have the product and don have the interface no one doing that today I curious your thoughts on how do you and obviously it very hard to know but how do you sort of mentally picture what their incentives should be on making more first-party products? And I'm asking this in relation to, as you think about investing in applications at the growth stage, and you think like, what's going to be worth it for them to build a first-party app in? And will that matter? Or again, does it all win? How do you think about their incentives and how, if at all, it relates to your application investing? Yeah, so we've done a lot of application investing. So again, this is like in the theme of and, right? We think a lot of things are going to work. On the application side, my framework is the big labs are going to be very focused on first-party products that are obviously in coding and within the blast radius of coding. And so you can interpret the blast radius of coding however you like, but it's probably some things around it. That's one piece. And then secondly, they're going to be focused on first-party products that appeal horizontally to all 1.5 billion knowledge workers. So think Microsoft Office and Google Apps. And they would like to have those as first-party products. We know from their work that they're doing, the work that they do with data vendors, they are very focused in those areas. So think all the Microsoft Office products. [15:00] Outside of that, There's a question of like, what number of priority is it for them? So, you know, we're, large investors in Harvey. And so, you know, Harvey, I feel like, you know, legal is, I don't know, it's very interesting to people on X. But, um, But it's a market that's absolutely in takeoff. So if you take an optimistic view of it, there's many differences, which we could talk about. But it's probably like 12 months behind coding, just in terms of diffusion into the workforce. And it's now gotten to the point where end clients are demanding that their law firms use RV. They're like, We care about it for product and we care about it for cost. And so you must use it. So it used to be this thing where people would talk about like hallucinations and And then now it's just like, no, everyone has it. and so you know It's probably like somewhere between number six through... By the way, that one, it really is funny how much it gets people on X. And we know this too, obviously. Yeah, for Legora. Yeah, and it's like, I think people, A, were surprised that these companies can grow this fast. in legal, which you would not have thought would be an early adopter. And then you've got a lot of doubt, because they're like, oh, well, wasn't it hallucinating and stuff? It's like, yeah, like 18 months ago. Yeah, 18 months ago, and they made a smart bet, which is like, what happens when models get 100x better? Yeah. And it's going to work. And legal is actually just a perfect thing, too, because it's language, but it's kind of mathematical. Yes, exactly. It's extremely logical. It's extremely well-documented in terms of case law. And the big breakthrough that we saw, which I'm sure is the same case with Legora, is reasoning models. And so lawyering is reasoning. And so, yes, it makes sense. But no, I agree. That doesn't seem like the next thing. picking off random companies, then you lose the trust. Yeah, then the ecosystem, yeah, of course. I think all the big platforms, I'm sure Microsoft had this. Salesforce had it. Yeah, they all had to decide, like, is it worth pissing everybody off to go, like, eat some random thing? Yeah, the old Microsoft thing, like, you know, a platform is only a platform if all of the things built on top of it generate more revenue than it. That's right. Right? And so I don't think that that's totally lost in the labs. Yeah. Now, it might still be priority number six. And there was a major validation event, which was when Kirkland came out and said they were going to spend $500 million to build their own technology stack. People were like, oh man, what's that mean for Harvey? So then you've got these bubble ones where they're... It's great. It's not a niche, but it is not the path to AGI. Yeah, exactly. It is going to be worth many, many billions. And there's two things that are really important about it. One, the last details of the product really matter to the users. And then two, it requires real go-to-market and on-the-ground efforts. And so those two things, everyone talks about what are now the modes of software or applications. FPs. Yeah, those things in the ground, making sure they work is a big thing. And so will the labs go do that in a verticalized way? At law, probably not. Whatever, there's a cloud legal product. They just can't be everywhere all the time. They can't do all of the things. are informed by that. It's not to say that we won't do things in that blast radius or in that knowledge worker horizontal category, but we do ask the question a lot more about what is the product roadmap, what's the right to win. In the applications beyond those, I think there's going to be hundreds of flowers that bloom because there's going to be a massive unlock. Back to the point of it's all going to work, I think even the stuff in the blast radius It's probably going to work. It's probably going to work. Yeah. We're large investors. We're the biggest investors in Replit. It's working very, very well. Yeah. Replit and Lovable. And you wouldn't think that these things, you would think that you're like, well, aren't Codex and Cloud Code going to be able to make a website? And the answer is, yeah, they can. And also, these things are unbelievable. They're very good. And Cursor also. And that is dead in the center of the blast radius. And there's a huge market of people who use it. And so, yeah, it's all going to work. I think... you know, with the model there, like, The appetite for frontier technology, I just think it's undisputed. It's just going to be there. The appetite for N-1 or open source technology is also going to be huge. It's going to be massive because if you 50x the usage that happens, people will care about cost. And maybe there will be a point where if cost is... 100x higher for Frontier, people really start paying attention to it. Well, an N-1 is going to be really good soon. It's going to be very good. N-1 will be better than 5-6. Yeah, take... I don't know when, but by soon. Yeah, and so you can see it actually at the cutting edge of the application companies, right? Who care about their own cost because that's their gross margin. And so, how long has it been since most customer support tickets required the highest-end model? Yeah. Like three years or something, right? the sort of [20:00] abstraction layer following how much value is required. And there's going to be a lot of people vying for the abstraction layer. like the application companies, obviously the labs with their first-party products, companies like Databricks or Palantir or Snowflake, they're going to vie for it. But I think people are going to use like, every variation of that. And it's all going to work. By the way, the customer support example is such a good one to me of... how long diffusion is going to take. Because the technology has been more than good enough for a little while, and we still, I don't know what percent penetration it has. but tiny. And that's like the clearest example to me of the technology is fully there. It's just going to take many years to get there. Yeah, the old enterprise thing of better, faster, cheaper. It's been better, faster, and cheaper. It's one on everything. For two years. And the companies are doing great. The companies are growing faster than any other application companies before them. There's still some laws of physics here. But yeah, I mean, you got to sell it and you got to implement it. And it's like, there's a lot of messy integrations you have to do and rules that you have to follow, brand guidelines. Yeah. Again, the labs aren't going to build that motion. Yeah, like there will be companies built on top of them. So that's like a classic and. So this is all enterprise. What about consumers? So, you know, like I think, I mean, obviously ChatGPT was sort of like the first consumer moment. And, you know, it was also the whole AI moment. But the first AI moment was a consumer moment. And like, that's what caught everybody's attention was for the first time people were using this thing. And they had been hearing about AI for, I don't know, 10 years. And then all of a sudden you have this thing. And even the first version of ChatGPT, you were like, oh, I've never seen that before. But, you know, Now talk to me four years later or whatever it is. What's the progress of consumer and how do you see it? Yeah, I mean, there's a billion people using it, right? And there's more than a billion people using it. Obsessively. Obsessively. But the instantiation of the product and the use case is very basic. We're in skeuomorphic mode for the vast majority of consumers, which is they are doing what they used to do on a search engine. inside ChatGPT. And that's the extent of the use case of the vast majority of those. Except for the kids. The kids use it better than that. Yeah, my son, total aside, a 10-year-old is writing full-on rap songs and creating music. No, it's not an aside. This is the future. This is the future. But this is incredible. It tells you the difference where there's a large cohort of people who are going to probably always Google, use this as a Google replacement. But that's not how the kids are going to use it. native. You can see it with the kids now. My kids were never trained on search engines. It's a whole new thing where they get it to give them answers to anything and give them work product of anything. The big shift is going to be when it goes from being reactive to proactive. And, you know, we're like just scratching the surface on that. Um, It's going to be multimodal, so it's not going to be just sitting doing this. The voice stuff actually is very, very good now. There was a bit of a head fake because the OpenAI voice stuff was not very good. for a period of time, and it had to rely on a dumber model. And now it doesn't, and it can handle interruptions and things like that. So like a natural human interaction form now exists. And capabilities for doing proactive work on your behalf. and action taking is like, we're just now there. OpenClaw was the first moment. I think GrokBot is a moment now. And I think that'll be the future of consumer use. It's so funny to me that no one talks about consumer right now. Everyone is obsessed with coding, rightfully so, given what's happened with the revenue ramps of the companies. But if historical technology markets are in a guide, the big prize is going to be the killer consumer products because you can address billions of people with them. So I think those are going to get a lot better. I'm very optimistic. And those are going to come with the world's best business models too. There's going to be a lot of value generated. There's going to be a ton of consumer surplus. What would that look like? I think it'll be advertising or some greater form of advertising. The thing I always say is if you try to describe the form factor of the feed, to people during classifieds at newspapers, it would make no sense. So we've yet to figure out what the native consumption model is going to be and then what the native ad format is going to be that follows that. Do you think it's ads and subscriptions? Yeah, definitely. Yeah, I think it's both. You think that it'll be a situation where hundreds of millions of people might be willing to pay today surprising amounts for this thing? Yes, absolutely. And again, I think... And that's when it's like a personal assistant. That's when it's a personal assistant for everybody. And so, again, I think my framework is like, 10x more value will just accrete to the surplus of the users. But there's going to be a lot of value created for that 10%. And so that accretes to the companies. And so it'll be subscriptions. It'll be ads or some new form that isn't quite what we think of today as ads yet been invented, and it's going to be a new form of consumption. [25:00] And so I'm very optimistic. The thing that you're most excited about in consumer is something that gets to an assistant for everybody. Yes. Yeah. And like take action on your behalf in a trustworthy way. You know, like... Hey. I've started building all these automations for myself. I'm aggressively trying to make myself more native. And I get like real value out of like, Two out of ten of them. But there's a lot more that I think will be done. And the models will get... 10x to 100x better over the next 5 to 10 years. So think about what that's going to resolve in. One of the other big... I fully agree with this. One of the other big personal consumer, you know, autonomy, AI things I think about a lot is just self-driving cars and autonomy is, I just feel like, I mean, you know, I just got a Tesla for the first time and like, I got it, you know, I... It is insane. It is like the best product because you get in the car, you type in where you want to go, and it takes you. You don't have to think about anything else. The whole ride gets you out of any weird thing. It can park anywhere. And I don't think most people have realized how good self-driving is yet unless you have a self-driving Tesla. Yeah, or if you live in a place where you can get a Waymo. Like basically that's like it's a small piece of people. Yeah, so it's like you've got these companies that can do it now, Waymo and Tesla, but it has not diffused yet. And so A, I don't think people realize... I don't think the people who have done it realize, but I don't think everybody realizes how good autonomy is right now. But then the other most important thing is I don't think that it has turned into the business model that it needs to turn into yet. So I'm curious your take on this. I know you're an investor and you have thought a lot about this, but where are we on the autonomy curve? Yeah. It's so funny to think about these other major technological tailwinds that are happening because AI just supersedes everything in our minds right now. Totally. I always find it helpful to just take a step back and just say, okay, what are the ways? Product cycles drive our business. And so what are the big waves that we're catching right now? Autonomous driving is one. But the previous generation of these was like, Mobile phones plus social. e-commerce, SaaS, cloud, kind of all coinciding. And that produced like 25 trillion of market cap. And like, it's like a ton of, that's like a lot, right? And, you know, like, what do we have staring at us in the face right now? Like, we've got AI, we've got autonomy. We've got robotics. We've got... AI's application to health. Yeah, I was going to say bio. We have the American dynamism category, which is like modernization of defense. The combination of all of those, I'm probably going to say, is way more than 25. It's just the most exciting time to ever be an investor and in the technology markets. Autonomy specifically, I think it's just very... underappreciated. The auto industry and people driving cars is one of the biggest industries in the world, obviously. Right now, okay, so if you rewind all the way to taxi driving, right? That was like less than 0.1%. of the overall miles traveled in the US. There were a bunch of reasons why that was the case. Then the ride share companies came along. And, you know, I was an investor in Uber. You know, obviously you guys were too. The big thing that happened obviously was like a better business model. But it was really like the 10Xing of the market size. So like, you know, taxis were 100 million bucks in San Francisco. And then they became a billion dollars like three years into Uber and Lyft. And so it 10X the market size because you delivered a much better product. Thank you. Autonomous driving, we're going to look back 10 years from now and say, oh, there was another obvious 10xing that would happen there. At least. At least, just on the ride-hail side. So if you have a product that is 14 times safer than a human driver. Then it's barbaric to drive. It's actually so, this is my view now. Don't you think, if you're like, how many people die a year from a human driver, if it's safer, isn't that, because isn't the data pretty abundant? and Wemos and they're like, between 10 and 14 times safer than a human driver. So it is barbaric to not actually allow for these things to diffuse into your economy. So I think history will look back on people and judge them very poorly if they don't let it happen. But for the consumer side, just start with that. If you can drive in something that's 10 to 14 times safer, you would pay a massive premium. Yeah. I think there's a very clear case for how you actually get the cost below the cost of Uber and Lyft today. Um, so right now owning a consumer car equates to like 80 cents per mile, like fully loaded depreciation, insurance, gas, all that stuff. Um, like that's your like cost of an incremental mile. Um, so. Riding in an Uber or Lyft is probably like two bucks and change. per mile. Um, This market is like massively elastic. So like if you can get the cost of a Waymo or a Tesla or whatever, [30:00] ideal service, which I think there'll be others, down below that. like there's going to be an explosion of demand. So it's 14 times safer. It's only going to get better. By the way, it's going to get safer. It's going to be 20 times safer soon. And, you know, it's going to be less than the cost. Like you will see a 10X plus. Then you talk about personal car ownership. And so there's 17 million cars sold a year in the U.S. new. And by the way, there's like 250 million cars in the US. just in the U.S., like, How much would you pay for full autonomous features for a passenger car that you owned? A lot. It's like double. $10,000. Yeah, like minimum $10,000. Yeah. Right? And so— I mean, once you're used to it, it is— Yeah, you can't not have it. It's really hard to go back. So you could sell it up front. You could do a subscription. But like call it minimum $10,000 a car, 17 million cars. Then you have all of the Western world, and then eventually it'll get cost- compelling enough to reach developing markets. So it's going to be one of these things where 50 years from now, it like the whole composition. Yeah. No one will be driving. Like the composition of everything looks different. It takes time because you got to produce the cars. Um, People are always struck like, how many Waymos there are, right? Like Waymos, if you're in one of the cities where they are, they feel ubiquitous. Yeah. There's like less than 10,000 in the US, in the whole US. And so it's one of those areas where there's Waymo, there's Tesla. I think there'll be other players. We've backed a couple that I think have a real shot at it. And it's not going to be 10 players, and it's going to be a massive market. And so it's easy to lose sight of markets like that because it's like the growth of AI. But those are going to be huge, huge markets that we're super excited about. There's nothing quite going yet, but you see the glimmers of it now. You go into the labs that are working on it, and the startups that are doing it, obviously Tesla's very good at it, but it's clearly going to work. It's clearly going to work. Yeah, and it'll be a bigger market than language when it does work, because it will have B2B applications, it'll have consumer applications. I think we're probably pretty far off from having robots in our house. Yeah. But for like, you know, in a factory, like relatively defined use cases with like, redundant tasks with safe environments like We see it. We invested in, one of my partners, a lot of investment in mind robotics. And so Sarah did this deal, and it's the founder of Rivian. And he basically was like, first of all, he's an exceptional founder and has produced a great company with great products. But he's going to put robots on the factory floor. And so they'll do manufacturing assembly work on Rivians. today. And then it's an embedded customer, you know, you kind of have it and you can not only get revenue from that, but you can also have the models learn from the work that you're doing. And so you have this feedback loop in the field that you can actually work on. Um, you know, there's other stuff that is kind of picks and shovel stuff that we're excited about, but like, we will have a chat GPT moment for robotics. I don't know when it's going to be. Yeah. Like, I think it's within five years, hopefully. Um, there's very, very smart teams working on it. it'll be bigger than language. Yeah, and it almost has to be when you think about... Think about the industrial impact of it, right? And this is one of these things where it's like, the physics doesn't prevent it from happening, and you can see the glimmers. So this will take, I agree with you, this is going to take more than 10 years to get all the way there, but I just don't see why it wouldn't all work. Yeah, I think it's going to work. Yeah, I think it's going to work, and I think we'll have them everywhere. And I think, by the way, the byproduct of this is... That we'll have massive productivity booms. I think that's going to be the economic impact. The quality of life will go up. We'll have to think hard about job impacts and stuff like that. Our workers are trained to make a transition, especially if it happens fast. But the history of all these technology breakthroughs is that we find... higher order things to do. Like 60, you know, whatever, 60% of jobs didn't exist 70 years ago like that we have today. So like we'll have a similar transition period that happens. But I think it's going to be a massive productivity boom. It's going to be awesome. So I want to talk about capital and financing a little bit. We at Benchmark just raised our first growth fund in history. And, you know, the baseline for that was something Welcome to the dark side. Yeah, exactly. It's like the forbidden fruit. And the thinking for us behind it was something along the lines of venture returns now happen at growth in a way that they didn't used to. And this conversation that we've just had really supports it because the outcomes are that big. And the things that run really run and it's all going to work and all of these things. the really big multiples can happen all along the way. Yeah. [35:00] I know you believe that. You're preaching the choir. I want to connect this though. I would like to hear your thinking on that. And I'd also like to hear, we just talked about the product cycle and where we are, but there's also this related, but not exactly the same thing of the capital cycle and the capital following the products and investing behind them and all those things. So I'd just be curious to hear your take on where we're at from a capital perspective, from a growth investing perspective. Yeah, of course. So I, As it relates to like why go do growth and like, is it a compelling category? Obviously you, you have, you have seen the light, um, but, the dark, it's, it's great. Um, The market that we're in is like a $5 trillion market. So it's like, you know, whatever, it's like 20% of the S&P 500. So if you're not actually, you know, playing the full, you know, the full sort of gamut of that in the private markets, like you're probably leaving a lot of money on the table. Like we did this analysis a long time ago that basically showed half of private market returns get generated between the seed and the B. And then half of returns get generated from the C+. And so if you're not aggressively playing from the C plus, you're definitely leaving off half. And by the way, that was during an era when companies went, public earlier. So now that companies are staying private longer, if you could fast forward five years, the ratio is going to be even more extreme. It'd probably be 70-30 or something in favor of the late stuff. You can also get venture outcomes because the power law is so crazy strong today. And so when you get it right, it can really go right. And we're in this dynamic market where the power law is more extreme than in the past. And I think it's a result of a few things. One is just, you know, we're early cycle. And so, you know, the early companies can establish themselves as leaders earlier. Secondly, this one in AI is unique because if you throw more money at it, it can just scale up. up and get better. That is different than previous cycles. If you threw endless amounts of money at ServiceNow or Workday or Salesforce, like during the SaaS era, they would get all messed up. Like we saw this experiment with like the vision fund, right? Like, You throw money at these things and the companies get screwed up. or if it was a consumer company they would spend it on Google and Facebook and that's not a good thing. Was that the mechanism for why the Vision Fund was dumping too much money in companies that didn't need it or was it hitting the cycle at too late of a time or was it both? I think it was both. It was like Well, first of all, I don't think it was a bad idea. I think it was just wrong time of the cycle. And then... we were not in an environment where if you put more money into something, it got better. Yeah. Just automatically, like with, with no scale limitations. Like that is what happens with AI. Yes. Like it used to be, you know, if you put, you know, a $40 billion fundraise into, you know, Salesforce like What are they going to do? Go hire a bunch of sales reps and engineers? It just wouldn't work. Whereas if you put $40 billion towards training, These models, they get much better. And so that is very unique in the moment that we're in right now. I think about the world as like, product cycles and capital cycles. And, um, I tell my team, like the ideal state of the world would be, you're in a great position on the product cycle, like skill one to 10, 10 being good one being not good. Like in 2021, we were at like a one, which is like, we didn't realize it at the time. Um, but that was like late SAS, late cloud. It was post mobile phone growth. Like all of the incremental consumer time spent was going to the major, you know, the major something new. So product cycle was like a one out of 10. And in 2010, it was like a 8 out of 10. Very good. So... Right now, product cycle is like a 9 or 10 out of 10. For all those reasons, I'd lay it out. Like AI autonomy, American dynamism, robotics, biohealth. Like all these things are going to happen over the next 10 years. Capital cycle, you know, your ideal would be like a 10 on product cycle and like, you know, a 10 on capital cycle, meaning valuations are really low. They almost never coincide because there's like excitement about the trend. It was like the best part of the product cycle ever and no one else saw it but you. pretty good. And the capital markets were a little more benign. And some of the crazy stuff and run-ups hadn't happened yet. That was, I don't know, like a seven. And 2010, I don't You could buy SaaS companies and B2B marketplace companies for you know, whatever, four times revenue or something. That was pretty good too. Um, Right now, capital cycle-wise, like people see the excitement, and you can still find ways to do good investments, and the companies are growing so fast and they're so dynamic, it's fine. but we're probably at a six or something. It's okay, but not great. But the thing that drives our business over 10 years, whether it's venture [40:00] or growth is the product cycles. - Yeah. - And that's like a nine or 10 out of 10 right now. - Yeah. It's sort of like, I think this, I can't remember if this is, I heard this first from Thrive or somebody, but it was like, you know, one way to be contrarian is to believe a thing nobody else believes. The other way is to believe the same thing everybody else believes, but believe it 10 times more. - Yeah. - Or something like that. And that's sort of like, that's a little bit our motto as a firm, I would say, is just, you know, especially at the growth stage, incrementally believe what could go right, and spend the time thinking about what could go right. I always say, like, our business is we, like, pay fair prices for great companies. Um, and, like, the alpha you get is always on the revenue side. Yeah. It's not like, is this going to end up a higher margin business than I'm, like, it's like, and so that's like market insights, product insights, and it's, um, um, The market doesn't appreciate sometimes the things that can really, really, really grow fast for a long period of time. Databricks is an example of this. Yeah. By the way, I got to have dinner with Ollie recently, and oh my god. what an amazing CEO. Like, So I like... coined this phrase like model busters or whatever, which is basically like the first time we invested out of the growth fund, obviously, you know, Ben did the series A and then we did basically every letter. And I think we're on like M now. I like joke with Ali that, um, you know, we're going to be like stadium seating and get to like double A, you know, like, you know, we're going to just stay private. I don't, I don't think so. Yeah. But, um, uh, but the first one that we did out of the growth fund, um, was right when we started the growth fund and we did a round at like $6 Ali has figured out Seven years later, his, his revenue has accelerated. It's like he has figured out more new products. And like our observation when things really go right is, it's often the founder that figures out the next thing. And, you know, his business is now positioned to benefit from AI. That's why it's growing so fast. Um, and our bet is he'll figure out the next thing and the next thing and the next thing. And he's great. And like, Those founders like that, as you guys have a growth business, like late-stage venture business, those founders are the asset class. If you back them, they're going to figure out all these amazing things to do. When we did SpaceX, Starlink wasn't even GA yet. And they had theories, and the economics were terrible, and it didn't work. Here we are. That is one of the best business segments ever created. going to have orbital compute, which is also going to work. Those founders are the things that you have to bet on, and you know, the outcomes of the companies can be much larger than you would ever put in a financial model. And that is the thesis that we have. One of the things that I really look up to about Andreessen is I think you got the sort of importance of narrative right from the beginning. I think about even when the firm was started, there was such an understanding that At the time, media was very important, but there was just an understanding that stories matter and that the way you present matters. And being loud and in the conversation is really important. And I feel like, obviously, from 2009 or something, since then, obviously, things have changed a lot. And I think in a very happy update, we've had sort of the Lululification of everything where people are going direct and owning their own narrative and you get more authenticity. a lot of new media going and all these other things. But I think Andreessen has understood the importance of narrative. And it seems to me that narrative is louder than ever. Right now, for companies, it's at all stages. It's like companies launch and you see these beautiful videos and companies at the later stages have their investors creating stories for them in various ways and they're owning media. And it's all just loud. But I think it's important. And I'm curious how you think about narrative now at this point in the cycle when we're all living on Yeah, look, this is just like the vibes, right? Like the vibes, the vibes matter tremendously. But they matter. They matter a lot. I actually have a funny story from a meeting that I was having with Ben in 2018. So this was, I don't know, three or four months before I joined. And we were kind of hashing out, you know, what is our strategy going to be? And so I had all these ideas of like, these are the new teams we need to build and this is what the investment team looks like. And he's like, well, what's your marketing strategy? And I was like, we don't really need a marketing strategy. It's not really a thing in growth. And he just looked at me. and said, that's the dumbest fucking idea I've ever heard. And venture firms didn't market themselves And then we did, and it worked really well. So like, why would you not do it here? And so I came along, you know, I came from a place that was a lot more conservative, and so I've sort of seen the light. Vibes matter. Let's just talk about why the vibes matter. So the vibes matter. [45:00] because, um... of fundraising. of your valuation, Um, for retaining your employees and for hiring new employees among other reasons. And, and, and In some markets, the vibes matter for your customers too. but at a minimum, it's those things. If you look at And this is basic, but why does a higher valuation matter? A higher valuation means less dilution, more ability to raise capital, and you can use that as a weapon. And in a world where capital actually makes your company better, it's worth more. Yeah, exactly, it's worth more. It was worth less when it was not the thing. Yeah, and now every investor... and every employee and potential employee is living on X and consuming the vibes. So if you look in the public markets, the companies that have done a great job of this, there's a direct line to their valuation, and their life gets a lot easier with a higher valuation. Famously right now, it's Palantir and Karp. And he's got a cult following. You know what's amazing on that one is that he... I mean, Palantir had a huge multiple before the recent... I think the growth has inflected recently. Yeah, it's massively accelerated. Massively accelerated. But it had the premium before it had that. Yes. And there's something interesting there. Yeah. And like... By the way, to tell you how much its vibes Like, No one knows what ontology means. No, I don't know. It's not even like a thing, right? And... I think a lot of people don't even really know what Palantir does. Yeah, for sure. Including big retail holders of the stock. A friend of mine who has a huge position, and I'm like, explain it to me. He's like, good vibes. But what has happened, right? They're massively accelerated in commercial, and they're now seen among CEOs as a trusted place to go implement your AI. So it has actually given them a direct benefit on the commercial side, because they're seen as the ones that will navigate the situation for you, and they're trusted. The valuation is a huge benefit for them, and then it's massive for them retaining and hiring employees. Similar to the Elon companies. It allows them to have a bigger war chest, to buy companies, to raise more capital, to have loyal following. So you can look at the high retail ownership stocks. And there's a correlation, if they're doing well, a correlation to the valuation. It's funny, I heard from a hedge fund guy this theory on shorting stocks. And his theory on shorting stocks was, I never short a Messianic founder, and I never short a product that people love. The inverse of that is if you have a Messianic founder, and there is a product that people love, you can actually like trade very high and have all these benefits, right? Right now in AI, obviously you said why the capital side matters and valuation side matters. It is incredible that the whole market is like hanging by a thread on the performance of two, now kind of three private companies on a monthly basis. I was just thinking as you were talking that the narrative, however strong the narrative is now, the importance of it all is about to go hyperbolic as these companies get public. You can even see it. There's this funny bit I mentioned to you. You can see investors and people who are on the opposite side kind of jockeying about trying to set expectations around It's not like the insiders are all pumping crazy numbers. It's actually like expectation management. But the amount of calories going into controlling these stories is just going to be wild. Yeah, and there are... It's not even just the CEO. So like I talked about how the asset class is the founder, and I talked about CARP, and I talked about Elon. Ideally if you have a dynamic founder who can be the face of the company, it's like Palmer with Andrel, right? Like, That will benefit you in many ways. - Yeah. - Again, if you have a product that people love, that Messianic founder thing is like a massive asset, and will make people care at every layer that matters, right? Capital side, employee side, customer side. And if you don't have that, if you're some no-name SaaS CEO, it's very hard to command people's attention and make them care. It also matters for people down in the organization, right? So right now, I think the person who's doing the best job of this in the market is Thibaut from OpenAI. And he's very direct. And not only is it feeding good vibes among the investor community and good vibes among employees, it's driving a lot of behavior on the customer side, too. And so I think when done right, it can have massive benefits. And it's not to be underestimated. It's something we talk to our founders about all the time now. It's like, you've got to go direct. own the narrative. You're the face of the company. You need to tell a compelling story. And if you don't, Yeah, some people care, but a lot of people aren't going to care. Yes, and someone else that's almost as good, but not quite as good. [50:00] Good product can come along and tell the story better. Yeah, exactly. Yeah, exactly. Yeah. My last question, and I think I asked, I had both Mark and Ben on Zoom. at different times and They both... were You could just feel the ambition and the scale and the mindset around all of it. And this shows up in a lot of ways. I was just joking to you before we sat down that Benchmark's pretty small, Andreessen's pretty big in terms of its team and all these things. I guess my question is how far, given that we're looking at not 25 trillion, but something how far can this all go? How big can a venture firm get? What rate limiters, if any, are there on what you can do? How do you think about what's left above the already large size you guys are at? Yeah. Look, we talk about this all the time because there's pieces of our business that are scalable and there's pieces of our business that are not, right? And something that you said, which I think is super smart, is if you want to go fast, be small if you want to go far you got to be big, right? And so there are things that we do that may cause, you know, more organizational overhead, but allow us, we think, to go further. And part of that is scaling our investment approach and scaling our team that does investing. And then a big part of that is the 650 employees that we have that help our companies. So how big can it get? Yeah, $25 trillion in market cap creation. Tech is... Call it like a third... to a little higher percent of the overall market cap. in the economy. If you look at previous eras. as a percentage of the overall market, different industries were actually much larger in the past. So like, and financial services and stuff. At their peaks, they would be like 80% of the market. I don't see a reason why tech will not continue to ascend as a percentage of the market cap. Like right now... eight of the top 10 most valuable companies in the world are, previously venture-backed technology companies in the US. And they all kind of look cheap. And they're probably the best businesses ever created. And then, by the way, we have OpenAI and Anthropic and SpaceX on their heels that are growing faster and that are going to get bigger. And then, by the way, if you go back to those big trends that I see coming over the next 10 years, we don't even have the players on the board yet for those. So in terms of I see no limit on how big it can get. In terms of our share and our ambitions as a firm, - Yeah. we probably have like 20% market share. We measure this at least in growth land. If you look at a real frame of our market, we probably have 20% market share. Now, ideally, we would have 20% you know, the top 20%, right? Like we would do the 20 best of the 100 things. But there's kind of no limit on like how much more we could scale. We just want to be doing the top 20 to 25% of things. So I think you're going to see like massive macro market growth, um, I think – Our share could grow a little bit, but that's not a goal. I think it's more like we can ride the wave really well. I mean, even keeping up with it is going to be hard. Keeping up with it is going to be hard, right? And so our ambitions are to be involved in the best companies across all the areas that we focus on. They're all technology focused, but they're pretty discreet, right? There's infrastructure, there's applications, there's American dynamism, there's biohealth, crypto. Those all require different things. But there's going to be new waves. we can like be part of those new waves and then ride the market growth. And that's kind of the sum of the goal. Love it. David, this was great. Thanks a bunch for making the time. I really enjoyed it. Yeah, me too, man. Thanks. Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, X at A16Z and subscribe to our sub stack at A16Z.substack.com. Thanks again for listening, and I'll see you in the next episode. This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product. This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with A16Z. Such advertisements, or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy. [55:00] Bye.