Benchmark's Vishria Says AI Compute Limits End Zero-Sum Tech Views

Benchmark合伙人Vishria：AI算力限制终结科技零和视角 能源供应瓶颈与软件护城河的转变，将使独立推理及芯片厂商与科技巨头并存共荣，而非被其吞噬。 在本期节目中，Eric Vishria剖析了云计算的演进如何启示当今的AI格局，并驳斥了有关AI价值捕获的零和假设。他指出，成功的AI初创企业在模型能力的参差边缘上运作犹如“搭建沙堡”，并解释了为何传统SaaS打法和企业销售模式在AI时代失效。此外，Vishria详细探讨了AI规模化所面临的物理限制（尤其是能源瓶颈），同时分享了他在Benchmark的投资框架、押注Cerebras等硬件公司的经验教训，以及为何现实中的体制摩擦能够避免AI引发突发性的劳动力崩溃。 模型服务基础设施中的深度系统工程，使Fireworks AI等专业平台在使用相同硬件时，能比标准云服务商实现高达5倍的速度和吞吐量。 (2:20) 担忧AI基础设施或基础模型市场将沦为单一赢家的零和博弈，与早期对AWS的误解如出一辙；不断扩大的市场自然能够支撑庞大的独立软件和基础设施生态。 (7:40) 企业采用AI的速度远快于最初的云迁移，因为企业管理者既将AI视为生死存亡的风险，又视其为迫在眉睫的效率机遇。 (9:01) 服务传统企业客户的AI初创公司必须扮演“AI夏尔巴人”的角色，将前沿研究转化为切实可行的企业工作流程。 (11:03) 现代AI工程需要“建造沙堡”——即构建灵活的产品架构，以便随着底层模型能力的迅速演进，每隔几个月便可快速重写。 (13:05)

Benchmark General Partner Eric Vishria argues that the expanding artificial intelligence market is non-zero-sum, challenging both legacy SaaS playbooks and deterministic fears of immediate disruption across the software and infrastructure landscape. Non-Zero-Sum Expansion and Infrastructure Realities (7:40) Vishria contends that the assumption of a single AI foundation model lab capturing all economic value mirrors the historical fallacy that AWS would eliminate independent software. Just as Snowflake, Datadog, and MongoDB thrived on top of early cloud providers, physical bottlenecks—such as power generation, memory, data centers, and specialized hardware—prevent any single entity from absorbing total demand. For compute providers, systems-level optimization provides defensible margins rather than commodity pass-through economics. For instance, Fireworks AI achieves up to a 5x speed advantage and superior throughput over hyperscalers using identical hardware and open-source models. Erosion of Moats and the Inverted SaaS Playbook (17:13) Legacy software businesses face structural multiple compression, often falling from 30 times revenue in 2021 to 3 to 6 times revenue unless they integrate core AI capabilities. Historical barriers to exit, such as database migration lock-in, are dissolving because autonomous coding agents easily translate specifications and eliminate switching friction. Consequently, winning software demands high agility, fast iteration, and continuous reinvention rather than rigid annual execution plans. This shift also renders conventional quota-capacity sales models obsolete. Instead of sales representatives pushing demand through fixed $1.2M to $2.5M quotas, standout AI tools benefit from intense inbound pull, enabling individual reps or technical founders to manage tens of millions in demand. At the product level, teams must adopt Bret Taylor's 'sandcastle' approach: engineering modular systems designed to be discarded and rewritten every few months as underlying model capabilities shift, which requires builders to re-engage deeply with low-level technical mechanics. Hardware Realities, Physical Bottlenecks, and Compute Constraints (27:53) Scaling compute faces severe physical limitations, where electricity and power generation act as the fundamental ceiling converting energy into intelligence. In hardware development, execution differs drastically from software: while an architectural block diagram represents 80% of software completion, it accounts for only 2% in hardware due to complex packaging, thermal dissipation limits, and intricate supply chains involving high-bandwidth memory and fabrication partners. Startups such as Cerebras illustrate the heavy capital demands and regulatory vulnerabilities (such as CFIUS reviews) intrinsic to competing against established semiconductor architectures. Real-World Friction Versus Deterministic Trajectories (62:16) Vishria challenges deterministic predictions of mass technological unemployment, citing Geoffrey Hinton's flawed 2016 prediction that AI would quickly eliminate radiologists. Although models can outperform humans on isolated scan reads, real-world deployment is slowed by fragmented multi-modal datasets, medical liability, and regulatory reimbursement structures. Furthermore, lowering the cost per diagnostic read triggers Jevons' Paradox: exponential increases in total demand for imaging expand the need for practitioners rather than reducing it. Vishria notes that pure technologists routinely extrapolate software capability in a vacuum while underestimating the institutional inertia, incentives, and systemic friction inherent in physical and human workflows. Vishria concludes that while the macro AI pie will expand to support hyperscalers, edge infrastructure, and specialized platforms alike, high-level tailwinds will not protect non-differentiated products. Startups must build with radical technical distinctiveness while navigating physical energy limits, real-world institutional friction, and fast-moving capability frontiers.

Benchmark合伙人埃里克·维什里亚（Eric Vishria）指出，人工智能并非赢家通吃的零和赛局，传统软件工程与销售范式正在被彻底颠覆，而能源供应与物理世界摩擦则是决定技术落地节奏的关键制约。 非零和博弈与AI基础设施分层 (7:40) 维什里亚回顾云计算历史指出，市场在2014年前后曾误判AWS会消灭所有独立企业软件公司，但随后Snowflake、Datadog和Cloudflare等独立企业依然在竞争中崛起为千亿美元级别巨头。他认为当前AI领域同样存在单一前沿实验室（如OpenAI或Anthropic）将吃下整个市场的零和思维陷阱。受电力供应、数据中心、芯片产能与算法演进等物理边界制约，没有任何单一供应商能够独占全球需求，超大规模云厂商、新型云平台、专业推理托管商与边缘计算将长期并存。 在基础设施层面，维什里亚以其投资的Fireworks AI为例反驳了“模型托管仅是低毛利同质化代工”的论调。在采用完全相同的英伟达硬件与开源模型时，Fireworks依靠底层系统工程实现了相比主流公有云平台高达5倍的运行速度以及显著的吞吐量优势。这表明在2万亿至4万亿参数规模的旗舰开源模型服务中，深厚的系统级技术专长能够创造出坚实的效率壁垒与盈利空间。 SaaS护城河瓦解与沙堡工程范式 (13:05) 维什里亚分析认为，AI编程代理（如Claude与Codex）能够无疲劳地完成繁重的接口转换工作，使过去因极高迁移成本而构筑的数据库等软件护城河迅速瓦解。受此冲击，公开市场SaaS公司的估值倍数已从2021年峰值的30倍市销率（P/S）压缩至当前的约6倍，若无法有效整合AI能力则面临进一步滑落至3倍的风险。与此同时，传统依赖划定销售区域和人员配额的自上而下推销模式正在失效，在具备颠覆性体验的AI产品面前，能够打通技术前沿与客户痛点的技术型创始人展现出强大的主动吸引需求能力，单人销售产能甚至可达数千万美元。 在产品构建方面，维什里亚强调了Bret Taylor提出的“沙堡”理念，即AI工程团队必须打破过去耗费数年构建十年前端架构的“石头城堡”思维。由于底层基础模型每隔几个月就会产生突破性涌现能力，像Cursor和Sierra这类顶尖团队选择紧贴模型能力的前沿边缘（jagged edge），每隔半年便主动淘汰并重构自身代码库。这一变革促使产品经理必须放弃仅关注用户痛点而不干预技术细节的旧习惯，重新深潜入技术底层去掌握大语言模型的机械机理与失效边界。 物理硬约束：能源瓶颈与芯片现实 (27:53) 探讨AI规模化扩张的边界时，维什里亚认为终极瓶颈既非芯片亦非存储，而是电力能源。AI的扩张公式本质上由“能源转化为算力、算力转化为智能”构成，能源短缺或电价上涨将直接推高Token生成成本，进而对智能供给造成物理压制。维什里亚指出中国正在建设规模远超美国的能源基础设施，并呼吁对天然气、太阳能、核能及电网基础设施进行全面投资，以避免算力成本发生结构性飙升。 在硬件工程方面，维什里亚结合Benchmark在2016年投资晶圆级芯片公司Cerebras的历程指出，软件设计中逻辑架构图完成即意味着80%的工程就绪，但在物理芯片中仅代表完成了2%。硬件创业必须直面封装物理极限、台积电以外30多家核心供应商的供应链协同以及散热熔解等极端物理风险。类似的技术落地障碍也出现在机器人领域：与坐拥海量网络文本的大模型不同，物理交互缺乏现成的数据语料库，Sunday Robotics等团队必须通过定制遥操作手套等专用硬件自建高保真数据集，才能借助小样本微调与强化学习让机器人完成泛化操作。 投资筛选法则与技术落地摩擦 (62:16) 在风险投资方法论上，维什里亚在Benchmark推行极度精简的投资节奏，过去12年间仅投资了18家企业。其决策遵循严格的三项心智过滤标准：能否问心无愧地劝说挚友将该公司作为毕生事业（终身事业测试）、是否极具热情在周六深夜接听创始人来电（绿钮测试），以及最根本的战略自省——“如果我们是对的，这真的重要吗？”。伴随全球技术市场空间呈指数级放大，Benchmark重启了成长基金，因为百倍现金回报已不再局限于极早期阶段，在10亿美元估值入场同样存在创造超额回报的可能。 针对行业内的恐慌情绪，维什里亚以杰弗里·辛顿（Geoffrey Hinton）在2016年断言“应立即停止培养放射科医生”为例，指出了纯技术专家对真实社会系统的误判。以Benchmark投资的放射学AI企业New Lantern为例，现实中不仅存在跨模态数据严重割裂的障碍，还面临医疗事故责任认定与医保报销体制等制度惯性；更重要的是，根据杰文斯悖论（Jevons' Paradox），读片成本下降反而刺激了影像检查需求爆发式增长，导致医疗系统对放射科医生的实际需求不降反升。这表明将纯技术能力简单线性外推至社会结果往往会严重脱离实际。 维什里亚总结指出，在技术范式剧烈转换的时期，创业团队与投资者必须在微观工程上保持极度敏捷，同时对现实产业机制与物理约束保持清醒；当前的核心未决议题在于前沿基础模型实验室与芯片硬件巨头之间将如何切分长远经济价值，以及数百家估值严重承压的未上市SaaS企业将如何完成资本破局。

## Background 本期节目来自播客频道 Colossus 旗下的《Invest Like the Best》，于 2026 年 8 月 11 日发布。节目邀请了顶级风险投资机构 Benchmark 的普通合伙人 Eric Vishria，围绕他十年来在软件与硬件领域的深耕投资经历，深入探讨了云计算历史经验对当下 AI 浪潮的映射、底层工程护城河以及创始人和投资人制胜标准的演进。 ## TL;DR Eric Vishria 结合 Fireworks、Sierra 与 Cerebras 等核心投资案例，深入剖析了当下 AI 浪潮的本质：大模型推理远非同质化通道业务，而是高度依赖底层极端工程优化的稀缺能力。他通过复盘云计算早期遭遇怀疑到“AWS 吞噬论”的破灭，指出零和博弈思维的荒谬，强调 AI 竞争前沿已彻底转移，传统 SaaS 的护城河与销售配额模型正在崩溃，未来的扩张终极约束在于能源。与此同时，无论是具身智能的数据荒漠还是硬件芯片的物理极限，创业者与投资人均需具备“富有成效的无知”与苏格拉底式的深层追问，在多极化生态中将战略推至极致。 ## Atmosphere & style 整场对话沉稳、睿智且极具思辨张力，洋溢着硅谷顶级风投教父级的内行对谈氛围。Vishria 的叙事既有鸟瞰计算史三十年沧桑演进的宏大格局，又有亲历芯片“熔毁”边缘与董事会残酷博弈的鲜活细节。语气冷静克制却充满信念感，在破除行业流行神话与浮躁共识时刀刀见血，展现出高认知密度带来的清醒与从容。 ## Bulletin - `02:20` 大模型推理并非简单的大宗商品业务，Fireworks 借助专有底层编译与系统优化，在同等 NVIDIA 算力与开源模型下实现了主流云厂商 5 倍的速度与数倍的吞吐量。 - `07:40` 传统“AWS 吞噬一切”或“大模型实验室独占 98% 价值”的零和博弈悲观论断并不成立；受物理电力、基础设施建设及芯片供应链等硬约束限制，AI 市场蛋糕将呈指数级膨胀并演变为多极化赢家生态。 - `17:13` 传统 SaaS 赖以生存的配额产能销售模型与高黏性护城河正在瓦解，AI 驱动的极速代码重构让数据库迁移等高壁垒场景变得轻而易举，企业必须拥抱 AI 转型以避免估值倍数严重压缩。 - `01:03:35` 杰弗里·辛顿 2016 年关于“停止培养放射科医生”的预言彻底破产，揭示了单纯技术推论忽视了体制责任、数据异质性及“杰文斯悖论”带来的真实需求爆发。 ## Key takeaways - `02:20` 开源大模型推理绝非通道型大宗商品，底层的系统工程与编译调度优化能拉开数倍的吞吐量与成本差距。 - `05:42` 历史表明技术革命从不走向单一垄断，当年的云计算催生了 Snowflake、Datadog 等独立巨头，AI 也不会被单一实验室通吃。 - `07:40` 零和博弈思维低估了技术爆发带来的总体增量，物理电力与供应链限制注定 AI 演进为多极并存格局。 - `09:01` AI 采纳轨迹重演早期云计算历程，正从早期数字原生团队的激进采用走向大型传统企业的加速探索，“AI 向导”（AI Sherpa）具备极高商业价值。 - `11:30` 现代产品构建从“修坚固城堡”转为“堆敏捷沙堡”，团队必须深刻理解模型的“锯齿状能力边界”，主动拥抱快速重构。 - `17:13` AI Agent 极大削弱了代码改写与接口绑定的黏性，传统数据库等基础设施的胜负手转向冷启动速度、零到无限的弹性与极端成本效益。 - `22:10` 传统企业软件按部就班的年度计划和配额销售模型正在失效，具有魔力体验的 AI 原生应用正在凭借狂热的拉式需求实现极高的单兵人效。 - `27:53` 将算力转化为智能的终极瓶颈不是资本，而是以电网、电厂和核能为代表的能源硬约束。 - `30:00` 硬件加速深度学习受限于三大物理维度：核心数量、核心互连通信带宽以及存储与计算的物理距离，Cerebras 借晶圆级芯片将其推向极致。 - `34:30` 在 GPU 之外，专为大型语言模型所生成代码设计的高能效第六代 CPU 架构正成为前沿硬科技的新机会。 - `36:18` 投资底层硬科技需要“富有成效的无知”（Productive Naivete），若只盯着 99% 的失败先验，就会错失改变世界格局的非凡机会。 - `40:00` 具身智能亟需摆脱受控环境，通过传感器垂直整合采集高价值物理数据以冷启动预训练模型，再经由强化学习涌现通用操作能力。 - `45:58` 卓越的董事会伙伴核心是苏格拉底式的发问者与认知陪练，通过每年协助做出好 1% 的决策在长周期中产生巨大复利。 - `50:00` 投资前需经受三道拷问：能否真诚说服顶尖牛人将其作为毕生事业、周六晚九点是否乐于接听创始人电话、做成了是否真正重要。 - `51:13` 由于终端超级赢家体量被空前撑大，早期阶段与高现金回报倍数（MOIC）的等价关系外溢至成长阶段，促使 Benchmark 设立成长基金。 - `58:38` 大量未上市及 2021 年高估值上市的 SaaS 企业面临流动性锁死与估值倍数严重压缩，市场注意力和资本正在全面流向 AI 原生公司。 - `01:02:16` 生态各层皆有巨大空间，端侧推理、边缘节点与云端超大模型将协同共存，商业模式亦将从订阅制迈向按成果付费（Outcome-based）。 - `01:03:35` 辛顿对放射科医生的误判印证了技术确定论的盲区：AI 降低成本后引发“杰文斯悖论”，叠加责任伦理与复杂体制，促使专业需求不降反升。 ## Detailed walkthrough ### `02:20` 透视 Fireworks：大模型推理远非大宗商品 在 Benchmark，合伙人们每年出手投资的项目极度克制，但每一个成功的标的——比如 AI 领域的 Sierra 和 Fireworks——都成为洞察整个世界底层秩序重组的绝佳透镜。当被问及通过 Fireworks 看到了哪些违背外界直觉的行业真相时，Eric Vishria 直截了当地指出：**这些超大规模模型在工程底层的运行难度被严重低估了**。当今主流前沿模型动辄拥有 2 万亿、3 万亿乃至 4 万亿参数，把它们平稳跑起来已经极其困难，而要想“高效”地运行，更是难如登天。 外界通常误以为开源模型推理是一种通道型的大宗商品业务（commodity pass-through）。无论是 AWS、Azure、GCP 等传统云巨头，还是各类新型算力云（NeoClouds），亦或是 Fireworks 和 Together 这样的专业模型服务商，大家在开发者集群里跑的都是公开可见的开源模型，底层挂载的也都是同样的英伟达硬件。然而，Fireworks 展现出的速度性能却能达到传统云厂商的 5 倍左右。除了肉眼可见的延迟和速度优势， Fireworks 在吞吐量（throughput）层面的效率更是云厂商的数倍——尽管吞吐量对外并不透明，但它直接决定了推理服务的底层单位经济学。 这种技术鸿沟引发了一个极具戏剧性的商业反差：Fireworks 本身也是租用底层硬件并承担着云厂商的基础利润率，但在叠加上自身专有的推理优化系统后，不仅能够提供碾压性的性能，还能保持健康的盈利。"等一下，面对同样的开源模型和同样的英伟达硬件，怎么可能出现 5 倍的速度差距和数倍的吞吐量鸿沟？" 这彻底打破了“推理即大宗商品”的论调。这背后依赖的是极度稀缺且高度特定的底层系统工程专长，绝非简单的买卡铺算力。 ```mermaid flowchart TD accTitle: AI推理能力对比架构 accDescr: 对比传统云厂商与Fireworks在相同硬件与模型下的性能差异 subgraph HardwareModel["基础资源层（同质化）"] Nvidia["NVIDIA GPU 算力集群"] OpenModels["2T-4T 参数开源大模型"] Nvidia --- OpenModels end subgraph CloudInfra["传统大型云厂商 (AWS / Azure / GCP)"] StandardStack["通用虚拟化与调度栈"] CloudResult["基准性能<br/>（低吞吐量 / 高延迟成本）"] StandardStack --> CloudResult end subgraph FireworksInfra["Fireworks 专业推理引擎"] DeepSystems["专有编译与吞吐优化系统架构"] FWResult["5x 速度领先<br/>数倍吞吐量提升 & 独立盈利空间"] DeepSystems --> FWResult end HardwareModel --> StandardStack HardwareModel --> DeepSystems ``` --- ### `05:42` 历史重演：从 2006 的无人看好到 2014 的“AWS 吞噬一切” 面对市场关于“规模效应最终会把大模型推理收敛为一两家寡头垄断的大宗商品市场”的担忧，Eric 翻开了云计算的历史账本。2006 年，亚马逊推出了 AWS 的前两大基石产品：计算服务 EC2 与存储服务 S3。在 2006 年底至 2007 年间，杰夫·贝佐斯（Jeff Bezos）在股东信中大篇幅阐释 AWS 的战略意义与远景。然而，当时顶级投资机构的反应却极其冷淡：如果把当时硅谷最聪明的 30 位投资人关进一间屋子，询问 AWS 是否能成为一门拥有持久高毛利、非大宗商品的优质生意，投票结果恐怕是 0 票赞成、30 票反对，所有人当时都认定它只是廉价出让算力的公用事业通道。 然而，仅仅过了数年，到了 Eric Vishria 于 2014 年加入风投行业时，市场风向却彻底倒转为另一种极端的恐慌叙事：“天呐，AWS 将吞噬整个软件产业（AWS is going to eat everything）。”彼时业内普遍认为，AWS 不仅会吃掉数据库和基础设施，还会将应用层也彻底扫平。传统的 SaaS 与企业软件之所以受资本青睐，是因为它们拥有类似年金的持续订阅收入和高达 85% 的毛利率；而市场传言亚马逊可以承受以 8% 的微薄毛利提供同质服务，这意味着传统软件公司的生存空间似乎将被完全抹杀。 事实却给出了极具讽刺意味的回击。从 2014 年至今的十年里，企业级市场不仅没有被 AWS 独家垄断，反而诞生了一大批超级独立软件巨头： - **基础架构层**：Snowflake 顶着亚马逊 Redshift 的直接竞争，在 AWS 的基础设施之上建立起庞大业务，完成了“在亚马逊的地盘上超越亚马逊（out Amazon-ing Amazon on Amazon）”的壮举；此外，Confluent、Elastic、MongoDB、Databricks 等公司也在各自赛道全面开花。 - **应用监控层**：亚马逊在早期就推出了原生的云监控竞品，但专注于该领域的 Datadog 依然一路高歌猛进，最终成长为市值千亿美元的巨头。 --- ### `07:40` 零和博弈陷阱与多极化赢家生态 2014 年“AWS 吞噬论”之所以破产，最根本的原因在于投资者普遍陷入了**零和博弈思维（Zero-sum thinking）**。悲观者完全没有预料到整个企业计算市场的增量空间究竟有多么庞大——"如果这一切真的全部跑通了呢？（What if it all works?）" 当技术革命真正爆发时，市场蛋糕的膨胀速度远远超出了单个厂商的吞吐极限，即便强如亚马逊，其组织和交付能力也无法单独承载并消耗掉整个行业的巨大需求。 不仅独立软件服务商崛起，就连云计算平台本身也没有走向单一垄断。2014 年时，微软 Azure 和谷歌 GCP 在市场上几乎无足轻重，而到了今天，它们双双成长为年收入数百亿美元的巨擘，与 AWS 共同构成了约 40%、30%、20% 的三足鼎立寡头格局。即便在这三朵主流公有云之外，像 Cloudflare 这样从边缘计算切入的差异化云厂商，同样长成了千亿美元体量的全新生态位。 这一历史轨迹为当下的 AI 竞争提供了直接的镜鉴。当下业内充斥着类似的恐慌与神化叙事——例如断言“Anthropic 或 OpenAI 将包揽一切应用与基建”。在 Eric 看来，这种论调在商业底层上依然站不住脚。与当年单纯依赖代码和服务器扩容的云服务相比，今天 AI 模型的扩展（Scaling）受制于极其复杂的物理链条： - 底层电力与能源供应的锁死； - 变电站与数据中心机房（Shell）的建设周期； - 芯片（GPU/ASIC）与高带宽内存（HBM）的严苛供应链； - 顶层算法架构的快速演化与适配。 物理约束使得单一厂商更加不可能独自吃下整个市场，未来的生成式 AI 格局同样大概率收敛于几大寡头并存、并在其外围孕育出数家千亿美元垂直技术冠军的生态体系。 --- ### `09:01` 云计算与 AI 的采纳曲线映射：从数字原生到传统企业 从企业需求端的采纳节奏来看，当前的 AI 扩散曲线也与早期的云计算如出一辙。回溯到 2010 年至 2011 年（即 AWS 面世后的三到四年），华尔街大行、保险巨头以及各类传统蓝筹企业对云计算依然抱有极其强烈的怀疑态度，甚至断言绝不可能将核心业务数据迁移上云。 在那个阶段，云平台早期惊人的用量增长几乎完全是由“数字原生代（Digital Natives）”驱动的： - 一批全新的、从第一天起就原生架构在云上的初创公司开始爆发。其中最为典型的案例是 Snapchat：Snap 完全构建在谷歌云平台（GCP）之上，在 2012 至 2014 年间，仅 Snapchat 一家公司的用量就一度占据了 GCP 总业务量的近 40%。 - 这一景象与今天 AI 原生应用的爆发完全重合——正如当前 Cursor 这类现象级 AI 工具，在短时间内迅速吃掉了各底层模型提供商极高比例的 API 调用量与算力份额。 直到 2014 至 2016 年前后，云计算的稳定性、合规性与成本优势在经历了近十年的验证后，最保守的银行、大型金融服务机构与保险财团才终于恍然大悟，开启了迟到但不可逆转的大规模上云浪潮。企业级 AI 的渗透当前正处于极为相似的拐点前夕。 ### `10:00` 企业的 AI 焦虑与“AI 向导”（AI Sherpa）的角色 在回顾云计算的普及历程时，企业起初普遍持怀疑和轻视态度，直到顶级工程师因渴望使用更便捷的平台而流失，企业才被迫正视云计算的价值。与当年的迟缓相比，当前大型传统企业（Blue Chip Enterprise）对 AI 的态度发生了根本性转变。尽管纽约传统金融机构的 AI 落地深度远不能与 Cursor 这样的技术原生团队相提并论，但企业高管们对 AI 抱有极高的探索欲望：他们正在积极划拨预算、组织内部实验并展开高频讨论，没有任何人敢对这项技术等闲视之。 企业普遍将 AI 视为比当年的云计算具有更大颠覆潜力的机遇，同时也视其为更致命的生存威胁。由于当年在云计算浪潮中吃过动作迟缓的亏，企业这次下定决心紧跟节奏。然而，从硅谷走入现实中的大型传统企业，会发现它们在技术采纳速度和落地障碍上仍面临巨大鸿沟。这种割裂孕育了巨大的商业空间：初创公司不能只做一个单纯的软件供应商，而应充当企业的“AI 向导”（AI Sherpa）。能够横跨前沿 AI 技术与复杂企业现实两座大山、为客户引路的团队，将拥有极高的商业价值。 - 传统企业在云计算时代曾因忽视开发者体验而遭遇招聘危机；在 AI 时代，企业从一开始就展现出极强的主动性与危机感。 - “AI 向导”定位的核心在于连接硅谷快速演进的模型能力与传统企业迟缓的业务系统，帮助其克服吸收门槛。 --- ### `11:30` Sierra 的启示：贴近底层模型与“锯齿状能力边界” 如果说 Fireworks 代表了底层的 AI 推理基础设施，那么由 Bret Taylor 创立、Benchmark 合伙人 Peter Fenton 投资的 Sierra（两人已有长达 20 年的合作经历，这是他们联手打造的第三家公司）则代表了应用层的前沿探索。Sierra 团队虽然扎根于企业级服务领域，但其骨子里是极具硬核技术背景的极客。他们精准践行了“AI 向导”策略：从极易实现自动化的客户服务切入，随后推出 Horizon 等长运行代理（Long-running agents），逐步接管更复杂的企业流程。 表面上看 Sierra 是一家企业级应用软件公司，但其研发过程极度贴近底层模型。团队不仅在做真实的 AI 前沿实验，还在紧密围绕模型的底层能力、评估框架（Harnesses）构建 Agent。他们深刻理解模型能力的“锯齿状边缘”（The Jagged Edge of AI Capabilities）——AI 的能力曲线并不像人类直觉中那样平滑渐进，它在某些极度复杂的任务上表现惊人，却在某些极其简单的常识上漏洞百出。优秀的应用公司必须能够精准看清这个锯齿状的边界，并围绕其凹凸不平的特性进行针对性工程封装与弥合。 ```mermaid flowchart TD accTitle: 传统产品研发与 AI 时代研发范式对比 accDescr: 对比传统产品管理模式与当前以锯齿状能力边界为核心的敏捷沙堡式迭代模式 subgraph Traditional ["传统研发范式（建造坚固城堡）"] direction TB T1["产品经理洞察客户需求"] --> T2["撰写不干预技术实现的 PRD"] T2 --> T3["交接给工程师进行代码落地"] T3 --> T4["打造可用数年的长久固定架构"] end subgraph AINative ["AI 时代范式（堆砌并迭代沙堡）"] direction TB A1["深刻洞察模型的锯齿状能力边界"] --> A2["紧跟模型每四周一次的突现能力"] A2 --> A3["结合客户痛点快速补齐能力缺口"] A3 --> A4["主动淘汰半年前代码（沙堡冲刷重塑）"] end ``` --- ### `13:05` 从“修城堡”到“堆沙堡”：产品开发的逻辑彻底颠覆 正如 Cursor 从最初的 IDE 界面演进到 Tab 键自动补全，再跨越到全自主的 Agentic 编码一样，前沿团队在持续主动淘汰自己半年前写出的代码。Bret Taylor 将这种状态形象地比喻为“堆沙堡”（Sandcastles）：“我们过去是在建造坚固的城堡，而现在我们是在建造随时会被海浪冲垮的沙堡。”如果创始人依然抱持着传统手艺人心态，非要打下坚如磐石的地基、铺设完美的砖块并期望产品维持上百年，在 AI 时代注定会被淘汰。随着基础模型每隔三四周就突现出新的能力，开发者必须主动拥抱这种沙堡式的不确定性，不断在模型的最新边界与客户痛点之间填补缝隙。 这种变化彻底颠覆了过去几十年的传统产品开发方法论。过去的教科书要求产品经理（PM）只负责深入理解客户需求，严禁干涉技术实现或指定具体工程方案，只需将需求翻译给工程师即可。但在当下，这种传统模式已完全失效。现如今的产品构建者必须极其精通模型特性的细微差别——明确知道它们在哪些场景下强大，在哪些环节会崩溃——并将这种对模型锯齿边界的理解与客户真实问题深度熔接。Cursor 创始人 Michael 及其团队之所以能一骑绝尘，正是因为他们从一开始就对锯齿状能力边界了如指掌，并不断随模型能力的漂移重塑产品。 这一趋势带来了一个看似违背直觉的结论：尽管大众鼓吹 AI 会抹平技术壁垒，但现实中“具备硬核技术理解力的回报反而大幅飙升”。决定未来胜负的不再是传统的 PM、设计师或工程师职位头衔，而是团队是否具备以下三项核心特质： - **理解客户痛点的能力**：能精准识别真实业务环境中的刚需。 - **卓越的产品品味（Taste）**：在交互与架构上具备敏锐的取舍直觉。 - **对模型锯齿状能力边界的深刻认知与好奇心**：能洞悉底层能力的极限并随时推倒重构产品。 --- ### `17:13` 竞争前沿的迁移：以数据库为例看 SaaS 护城河的崩塌 外界有一种流行的论调，认为 SaaS 的黄昏是因为未来每个人都可以通过“氛围编码”（Vibe Coding）随手写出自己的软件，但这一观点完全偏离了重点。SaaS 公司真正面临的系统性危机在于：**软件行业的“竞争前沿”（Competitive Frontier）发生了彻底位移，过去创业者精心构筑的取胜法宝与护城河已经失效。** 以软件行业长期以来利润率最高、黏性最强的数据库领域为例： - **过去的护城河**：应用开发者必须针对特定的数据库接口（Database Interface）手写大量代码；随着应用运行，业务数据不断累积膨胀。将一套核心业务应用从 Oracle 或 SQL Server 迁移到另一个数据库，往往是一项耗时数年、风险极高、所有工程师都避之不及的灾难级巨型工程。因此，数据库厂商能够享有极高的定价权和惊人的毛利。 - **AI 时代的规则重写**：首先，现在不再是人类工程师编写针对数据库接口的代码，而是由 Claude 或 Codex 这类模型代劳；其次，数据库接口具有极高、极清晰的规范性（Well-specified），而 AI 天生最擅长处理高度规范化的任务；最后，AI Agent 不会因为枯燥重复的代码改写与格式转换而产生疲劳感。 - **护城河的崩解**：过去被视为软件工程“第一大禁忌”的数据库迁移，如今在 AI Agent 的驱动下已变成轻描淡写的琐事，只需投入极小的算力成本即可自动完成迁移。 数据库本身并不会消失，世界依然需要数据库，但成为一家卓越数据库公司的评价标准已经彻底颠覆。由于开发门槛大幅降低，未来新启动的实验性应用数量将呈指数级爆发。因此，新一代数据库的杀手级特性不再是绑定特定接口，而是能否实现“极速冷启动与用完即毁”、能否平滑支撑从“零负载”瞬间扩展至超大规模，以及是否具备极端的成本效益。 ### `20:00` 估值折价与执行陷阱：传统 SaaS 的生死抉择 在底层数据库与基础设施层，胜负的标准正在发生根本转变。过去企业采购软件，往往是为特定用户场景定制数据库规格、采购商业许可证并在固定硬件上运行；如今，迭代速度被无限推高，成本、零到无穷大的弹性扩展能力以及工作负载的可移植性成了终极裁判。几年前，面向传统 SaaS 公司的警钟就已经敲响：摆在面前的选择只有两个——要么拥抱 AI，要么接受 3 倍市销率（P/S）的估值惩罚。在 2021 年的高峰期，SaaS 公司的估值倍数普遍高达 30 倍，而如今许多上市 SaaS 企业只能在 6 倍甚至更低的水平挣扎。即使一家公司过去几年营收增长了 4 倍、实现了盈亏平衡，估值倍数萎缩 6 倍带来的惨烈现实依然是：业务做得更大，企业反而更不值钱。 这种范式转移在董事会会议室里催生了一种极其反直觉的残酷现实。Eric Vishria 直言，自己曾当面警告管理团队：“你们按部就班达成年度计划的每一天，实际上都在摧毁股东价值。”过去整个科技行业信奉的法则是制定严密计划、雷厉风行地执行、达成或超越预期，以此持续滚雪球并沉淀股权价值；然而在技术底层剧烈动荡的今天，死守旧计划无异于自掘坟墓。正如投资人好友 Ann Miura-Ko（Anne Lee Skates）所观察到的，许多管理着数亿美元规模业务的 CEO 们陷入了严重的惯性陷阱：他们习惯在早 8 点到晚 5 点处理传统既有业务，仅在晚 5 点到 8 点抽空思考 AI；而他们真正应该做的恰恰是彻底倒置这种时间分配。真正的新一代赢家——如 Fireworks 的 Lin Qiao、Baseten 或 Modal 等团队的创始人——对评估指标（eval）和业务演进极度敏锐，绝不受制于过去的路径依赖，随时根据底层模型的发展重塑业务形态。 ```mermaid flowchart TD accTitle: SaaS 业务模式对比 accDescr: 对比传统 SaaS 线性爬坡与 AI 时代颠覆式迭代的思维路径差异 subgraph OldWay ["传统 SaaS 登山模式 (陷入价值摧毁)"] A1["制定年度确定性计划"] --> A2["按部就班凶狠执行"] A2 --> A3["实现 1.5M-2.5M 销售配额"] A3 --> A4["估值遭遇倍数压缩 (30x 降至 3-6x)"] end subgraph NewWay ["AI 原生动态适应 (创造股权价值)"] B1["颠覆时间分配: 全力以赴投入 AI"] --> B2["根据底层模型迭代极速调整"] B2 --> B3["创始人填补前沿能力缺口 (拉动式需求)"] B3 --> B4["单兵撬动数千万美元魔法级增长"] end ``` ### `22:10` 传统销售法则的崩溃：从配额产能到“售卖魔法” 许多从上一代顶尖 SaaS 企业走出来的高管，带着光鲜的履历跨入快速扩张的 AI 原生企业，却遭遇了惨烈的失败。这种行业内普遍存在的“水土不服”，根源在于思维框架与业务实际需求的严重错位。长期以来，企业软件销售完全建立在“配额产能模型”（Quota Capacity Model）之上：初期每个销售代表背负 120 万至 150 万美元配额，内销代表（ISR）约为 75 万至 85 万美元，随着企业规模扩大，大客户销售配额提升至 250 万美元；财务模型以此为基准打折预测，并机械化地按照西海岸、东海岸、中部地区划分销售战区。 然而，这套逻辑的底层假设是“推式需求”（Pushing Demand），即销售人员必须费力向客户推销产品。但在当下的 AI 浪潮中，客户面对的是一种“纯粹的魔法”（Selling Magic），市场呈现出极其狂热的“拉式需求”（Pulling Demand）。当一家公司率先拿出真正具备魔力的 AI 产品时，传统的产能模型瞬间失效： - 销售代表不再受制于 200 万美元的配额上限，在新型 AI 公司中，单个销售甚至能够做到 1000 万、2000 万、3000 万乃至 5000 万美元的年产出。 - 带着传统经验空降的高管往往忙于画组织架构、计算人头配额和划分区域，却完全看不见真正的交付瓶颈与需求阻力。 - 当前最顶尖的销售本质上是“创始人本身”——他们做的事情极其质朴却不可替代：在模型参差不齐的“参差边缘”（jagged edge）与客户实际业务能力之间搭建桥梁。 - 无论是在云计算早期还是如今的 AI 时代，从业者犯下的最大错误都是严重低估了市场的潜在体量，而这次 AI 市场的广度远超以往。 ### `27:53` 能源瓶颈：算力转化为智能的终极约束 随着各界对智能的需求展现出近乎无限的吞吐胃口，整个科技界陷入了一种喘不过气的扩张节奏中：算法代理（Agents）从不知疲倦，但现实世界、组织和资本却急需两三个月来消化巨量变革。尽管人们常讨论资本是否会被巨大的基础设施开销耗尽，但在 Eric Vishria 看来，真正的硬约束不是美元，也不是宏观经济周期，而是最基础的物理法则——能源。 从根本原理来看，大模型的作用机制十分纯粹：以极高的效率将算力（Compute）转化为智能（Intelligence）。既然人类对智能的需求深不见底，对算力的渴求就将无限放大；而算力最底层的输入端只有能源。 Vishria 警示道，外界往往过度聚焦于模型蒸馏或某些开源进展，却忽视了极其严峻的能源建设现实：中国在下一年并网投产的能源规模预计达到美国的近 10 倍。如果美国在能源供应上落后，由于算力短缺或电力成本畸高，势必导致 Token 产出受限或单价暴涨，最终反噬智能的供给规模。面对天然气轮机、太阳能、核电以及稀土供应链等交织复杂的产业链瓶颈，政府和产业界唯一的出路是摒弃单一路径偏见，全面推动所有能源形式（核能、天然气、光伏等）的激进投资与建设。正是对底层算力极限的深刻着迷，也直接引出了他对 Cerebras 这一颠覆性芯片历程的深入思考。 ### `30:00` 投资 Cerebras 与硬件加速的三大物理极限 2016 年左右，Benchmark 投资了 Cerebras，这也是 Eric Vishria 首次涉足极其艰难的底层硬件投资领域。如今硬科技投资机会遍地开花，但在当时这完全是一次极具风险的孤例。Vishria 直言硬件投资极其痛苦，但也生动展现了创业初期某种“无知”（naivete）的必要性。当时 Cerebras 团队只有 5 位联合创始人带着一份 PPT 来 Pitch，Benchmark 已经有整整十年没有碰过半导体芯片公司。Vishria 最初甚至根本不想去听这场汇报，但作为合伙人又不得不去。然而，创始人翻开的第一页幻灯片就彻底抓住了他：**“GPU 在深度学习任务上其实很烂，它只是碰巧比 CPU 快了 100 倍而已。”** 当时的时代背景完全不同：Transformer 架构尚未问世，OpenAI 还只是一家名不见经传的前沿科研实验室，英伟达（NVIDIA）当时的市值仅约 400 亿美元（而非如今的几万亿），谷歌的 TPU 甚至都还没对外公布。Vishria 在此前的 18 个月里一直在考察深度学习在安全、医学影像等各类垂直软件领域的应用，试图弄清楚这项技术究竟会颠覆什么。当听到 Cerebras 对 GPU 架构底层缺陷的剖析时，他顿时豁然开朗。从周三初次会面到周一合伙人会议，中间经历了密集的多轮沟通，Benchmark 迅速建立起巨大的投资决心。 在硬件层面加速深度学习，归根结底只有三个维度的物理手段。Cerebras 的核心主张非常激进，就是将这三个维度全部推向物理与逻辑的极限： ```mermaid flowchart TD accTitle: 加速深度学习硬件的三大物理维度 accDescr: 展示硬件加速深度学习仅有的三条路径，以及晶圆级芯片如何将其推至极致。 subgraph Hardware_Limits["硬件加速三大物理维度"] A["1. 增加计算核心数量<br/>(Increase Cores)"] B["2. 增强核心间通信带宽<br/>(Core-to-Core Communication)"] C["3. 缩短存储与算力物理距离<br/>(Memory Near Compute)"] end subgraph Cerebras_Solution["Cerebras 晶圆级芯片 (Wafer-Scale Engine)"] D["450,000 个核心集成于整片晶圆"] E["晶圆内互连消除片间延迟与损耗"] F["20GB 板载 SRAM，彻底避免跨芯片内存访问"] end A --> D B --> E C --> F ``` - **三大硬件加速维度**：无论过去还是现在，加速深度学习在硬件上只有三件事可做：一是增加计算核心数量；二是提升核心之间的通信带宽；三是把内存搬到离计算核心尽可能近的地方。 - **晶圆级极限设计**：Cerebras 提出直接制造一整块晶圆级芯片（Wafer-Scale Chip）。在第一代（约 7 纳米制程）上堆叠了多达 45 万个核心，板载约 20GB 的超高速 SRAM，彻底避免了脱离芯片去读取外部内存的延迟，同时由于所有核心处于同一晶圆上，核心之间的互连速度达到了物理制造的极值。 --- ### `32:45` 硬件创业的残酷现实：从理论“天花板”到供应链物理法则 软件领域与硬件领域之间存在一条巨大的认知鸿沟。在软件创业中，如果你画出了逻辑架构块图、讲清了底层运行机制，往往意味着产品研发已经完成了 80%，接下来主要依赖于 Go-to-Market（市场落地）的商业执行力。然而在硬件行业，拿出逻辑架构图只相当于完成了漫长征程的 2%。 硬件创业者不仅要直面冷冰冰的物理定律，还要对抗极端复杂的全球供应链。除了行业皆知的代工厂台积电（TSMC）之外，还有多达 30 余家关键供应商决定着芯片制造、封装与集成的成败。Cerebras 经历了长达数年的艰苦推进：2016 年启动研发，直到 2019 年才拿到第一批工程样片；样片回来后绝非大功告成，而是必须经历严苛复杂的“唤醒流程”（Bring-up），其中包含多达 14 个环环相扣的测试步骤；直到 2020 年左右，团队才终于拥有了第一个真正能稳定运转的系统。 半导体开发中存在一个残酷的“顶线理论”（Roofline Model）： - **仿真即最高天花板**：在芯片设计初期的仿真模拟（Sims）数据，代表了硬件在物理上能达到的绝对性能极限（Roofline）。 - **现实软件层的层层折损**：一旦硬件物理流片归来，所有的现实因素——软件栈、编译器（Compiler）、计算核心算子（Kernels）等——都会无情地蚕食硬件理论性能。 - **漫长的打磨战役**：刚点亮硬件系统时，系统可能只能发挥出物理天花板 10% 的算力水平。工程团队往往需要在后续数年间不分昼夜地重写底层代码，才能一点一点将实际性能逼近那条理想中的天花板。 --- ### `34:30` 五代计算浪潮的演进法则与新兴的“第六代 CPU 机会” 回溯计算历史，每一代颠覆性算力架构的诞生，都源于全新工作负载（Workload）引入了全新的系统约束条件，并由此诞生出千亿美元级的巨头企业。如果 AI 无法成为一个体量庞大的新型计算负载，专用芯片的尝试就终将沦为边缘探索： | 计算浪潮一代 | 代表性新兴工作负载 | 引入的核心技术约束 | 诞生的百亿/千亿美元巨头 | |:--- |:--- |:--- |:--- | | **第一代：通用计算** | 通用多用途计算任务 (Multi-purpose Compute) | 逻辑分支控制与指令调度 | Intel（CPU 霸权） | | **第二代：图形处理** | 图形渲染与密集像素运算 (Graphics Rendering) | 大规模数据级并行性 (Massive Parallelism) | NVIDIA（GPU 时代） | | **第三代：网络互联** | 数据包路由与交换通信 (Networking / Routing) | 极低延迟处理 (Ultra-low Latency) | Broadcom（博通）、Avago（安华高） | | **第四代：移动计算** | 电池供电与便携式计算设备 (Mobile Devices) | 极致功耗比与能效限制 (Power Efficiency) | Qualcomm（高通）、Arm | | **第五代：人工智能** | 深度神经网络层级反向传播 (Deep Learning / LLM) | 核心间通信墙与网络互连带宽约束 | NVIDIA、Google (TPU)、Cerebras、Groq 等 | Vishria 当年坚定投资的核心逻辑，就在于确认了 AI 绝非伪需求，且它带来的全新瓶颈正是“通信受限”（Communication Bound）——GPU 虽然拥有强大的并行算力，却无法从根本上解决跨芯片、多层神经网络之间的通信传输墙。 紧接着，Vishria 透露了一个极具前瞻性的判断：在当前的 AI 加速芯片（GPU/ASIC）竞争格局之外，**第六代计算浪潮正在孕育，那就是针对 AI 时代的“新型 CPU 架构”**。目前的底层现实是：大型语言模型（LLM）虽然在 GPU 和加速芯片上运行，但它们生成的海量代码最终必须在 CPU 上执行。现阶段这些生成代码依然跑在拖着数十年历史兼容包袱的传统 CPU 架构之上。Vishria 透露 Benchmark 已经隐秘投资了一家未公开的新型 CPU 公司，致力于剥离传统 CPU 沉重的历史包袱，打造面向 AI 生成代码时代的全新架构。 --- ### `36:18` 芯片熔毁危机与“富有成效的无知”（Productive Naivete） 硬件初创公司的生死往往在一线之间。Vishria 回忆起 2019 年 Cerebras 的一次董事会会议，当时公司已经累计融资了约 5 亿美元。在现场测试中，那台庞大的芯片系统几乎当场“熔化”烧毁（"It's fucking melty"）。面对眼前的危机，Vishria 当时内心近乎绝望，暗想：“这下完了，我们要把这 5 亿美元全赔光了，这套方案根本行不通。” 然而 Cerebras 的工程团队展现出了超越常人的韧性与工程天赋，奇迹般地攻克了散热与物理封装难题。谈及此，Vishria 动情地表示，正是这样的瞬间让人为身为一名风险投资人而感到自豪：风险投资的核心绝非沉迷于金融数字游戏，而是作为一种媒介去支持那些“试图实现人类 50 年来未能攻克之难关”的团队。无论 Cerebras 最终能否获得商业全胜，这样重大的雄心探索本身就配得上风险投资。 这种敢于冒险的心态被称为**“富有成效的无知”（Productive Naivete）**。在芯片或底层硬科技领域，如果投资人拥有行业专家的全部经验，往往会被历史经验束缚而不敢开枪——因为专家们总能列出 47 条“年轻人不可能造出芯片”、“硬件胜率太低”、“物理法则不允许”的客观理由。 - **99% 的大概率偏见**：在风险投资中，所有唱衰项目的经验法则在 19/20 甚至是 99/100 的情况下都是正确的，这也是传统投资人倾向于规避风险的原因。 - **Bruce Dunlevie 的核心问询**：Benchmark 资深合伙人 Bruce Dunlevie 始终要求团队反向思考：“如果它成功了，会是怎样的景象？（What could go right?）我们是否看得清成功的路径？” - **风险投资的底层辩证法**：正如投资界流传的一句箴言：“如果一个项目失败了，往往是因为合伙人们提前预警过的所有理由应验了；但如果它奇迹般地成功了，一定是因为那些致命的反对理由最终都变得不再重要。” --- ### `39:19` 机器人技术的瓶颈与新前沿 对话随后转入另一个具备毁灭级商业潜力的领域——机器人技术（Robotics）。业界公认，如果通用机器人技术真正跑通，其商业价值与对社会结构的重塑将远远超过眼下经历的一切技术变革。但长久以来，家用管家式机器人（例如在家里帮忙叠衣服）就像核聚变一样，始终被戏称为“永远还需要 10 年”的画饼技术。 Vishria 分析指出，传统机器人技术其实早已大规模普及： - **已解决的确定性场景**：在汽车装配流水线、高度自动化的工业制造线上，传统机器人早已随处可见。 - **环境与任务约束**：这类古典机器人之所以能够规模化，是因为它们被严格限制在“受控环境”（Controlled Environments）中处理“高度重复性任务”（Repetitive Tasks）。在这一闭环场景下，工程问题在过去几十年中已基本被工业界彻底攻克。而机器人技术迈向下一阶段的核心挑战，正是如何摆脱受控环境，走向充满动态不可预测性的非受控真实世界。 ### `40:00` 具身智能的数据荒漠与垂直整合破局 机器人要在现实世界中处理细微多变的物理任务，核心不仅在于机械结构，更在于驱动机器人的 AI 模型。然而，当前具身智能面临的最大瓶颈是“缺乏互联网级规模的数据”来冷启动整个训练流程。大语言模型（LLM）的成功建立在抓取海量人类互联网知识的基础之上，但现实物理世界中根本不存在与之等价的即成数据集。虽然行业尝试了合成视频、物理仿真（Simulation）以及遥操作（Teleoperation）等方案，但仅靠人工遥操作去堆砌出媲美互联网规模的数据无异于天方夜谭。 针对数据匮乏，当下前沿机器人团队的破局思路是“抓取最高价值的高密度数据”。互联网本身充斥着大量垃圾信息（Slop data），模型公司在训练 LLM 时也是对维基百科、GitHub、Reddit 精品内容赋予更高权重，而降低普通杂讯的权重。AI 机器人公司同样可以从最高质量的物理交互数据切入，先打通底层的预训练模型基座（Pre-trained base）；之后只需在后训练（Post-training）阶段利用强化学习（RL）注入极少量的辅助示例，模型便能展现出惊人的泛化能力。 这种技术路径在历史上早有呼应。自动驾驶本质上就是“AI 叠加特定形态的机器人”，Waymo 与特斯拉（Tesla）的崛起均依赖于高度垂直整合的架构：自建车队或量产车采集专属高质数据，回传训练自家专有模型，从而打通端到端的闭环体验。在机器人领域，Benchmark 投资的 Sunday Robotics 也采用了完全相同的垂直整合哲学：研发与机器手结构精准匹配的手套，让人类示范数据与机械本体获得极高的数据迁移度（Transferability），建立起坚实的基础模型底座。 ```mermaid flowchart TD accTitle: 具身智能训练管线流程 accDescr: 展示从定制传感器采集高价值数据，到垂直整合预训练基座，再到RL后训练与涌现泛化的完整闭环。 subgraph DataCollection["数据获取与预训练"] A["高价值物理交互数据<br/>(例如: 专用传感器/定制手套采集)"] --> B["垂直整合架构<br/>(软件、模型与本体紧密协同)"] B --> C["高保真预训练基座模型"] end subgraph PostTraining["后训练与能力拓展"] C --> D["强化学习 (RL) 与微调"] E["少量下游场景辅助样本"] --> D D --> F["通用操作能力涌现<br/>(如: 任意衣物折叠与复杂分拣)"] end ``` ### `42:47` 地下室里的折叠机器人与“应用场景”之辩 谈及对 Sunday Robotics 的直观震撼，Eric Vishria 回忆起投资时合伙人 Peter 安排的一次拜访。最初在斯坦福大学实验室的地下室里，创始人拿出的原型甚至是用“纸板拼凑的简陋手套”。但不久后，当团队再次来到他们新办公室的地下室时，眼前的场景令人目瞪口呆：现场有十几台机器人正在完全自主地折叠各种形态各异的衣物。这并非预设剧本的演示（Demo），而是机器人自主尝试、试错与调整的过程，旁边甚至有工程师用尺子精准测量折叠后的平整度，以建立极其严苛的基准评估体系（Evaluation criteria）。 面对主持人关于“是否担心折叠衣服只是解决方案在生硬寻找需求（Solution in search of a problem）”的质疑，Eric Vishria 直言并不担忧，因为大语言模型的发展史已经证明了场景会不断泛化： - **历史镜像**：早期 LLM 的早期支持者早已预见代码生成和自动化 AI 研究将引向自我改进（Recursive Self-Improvement），但在大众应用层面，最早爆发的爆款（如 Jasper）仅仅是帮人写营销文案和文章。 - **任务特征**：折叠衣物是一项极其完美的训练场景——它形态任意、物理接触复杂、需要极其精细灵巧的手部微调（Dexterous manipulation），同时又完全没有严苛的时间敏锐度（就算花上三倍时间慢慢折也无所谓）。 - **核心本质**：初期的具体任务究竟是什么根本不重要，真正重要的是底层的预训练模型是否足够强大；一旦基座模型成熟，机器人解决多样化任务的能力将会呈指数级不断倍增。 ### `45:58` 顶级董事伙伴：不仅是投资人，更是认知陪练 在硅谷创始人和同行眼中，Eric Vishria 被公认为最卓越的董事会伙伴（Board partner）之一。对于为何能获得如此评价，Eric 剖析了自己的投资哲学：在进入董事会时，他永远把自己定位为“伙伴第一，投资人第二”。市场上常有“投资级（Investment-grade）机会”——数据好看、投进去大概率稳赚不赔，纯粹的投资人会果断开枪，但他作为“伙伴”必须直接放弃，因为缺乏与创始人之间深层的相互吸引与化学反应。 这种深度的合伙关系在顺风期并不显眼，却在企业陷入泥潭（Grind）时发挥决定性作用。以生命科学 SaaS 独角兽 Benchling 为例，其创始人 Saji 与 Eric 合作多年，公司业务曾顺风顺水，前六年报表里甚至从来没有“客户流失（Churn）”这一统计栏。然而，随着后续生物医药一级市场遭遇寒冬，公司在短短 12 个月内承受了相当于过去 7 年总和的客户流失。在最煎熬的逆境期，合伙人不仅要坚信企业能突围，更要享受共同战斗的过程（Joy of the game）。 关于如何发挥董事会成员的实际价值，Eric Vishria 分享了他的核心观察与工作法则： - **不同风格的先贤**：风险投资领域有很多成功的模板，例如 Michael Moritz 曾是记者作家，John Doerr 是顶级销售型领袖，Bill Gurley 是硬核工程师出身，而 Peter 则是纯粹的职业风险投资家。 - **苏格拉底式的发问**：优秀的创始人心中通常早已有了答案，甚至知道自己该怎么做；董事的核心职责不是发号施令，而是通过提出精准、尖锐的问题，帮助他们理清思绪、强化信念（Conviction），并把战略更清晰地表达出来。 - **1% 的复利奇迹**：陪伴的价值在于一年中帮创始人多做出两三次“好上 1% 到 2%”的优质决策。这种细微的认知校准在长达十年的创业马拉松中持续复利，最终便会沉淀为巨大的胜势。 ### `50:00` 投资前的三道灵魂拷问：人生追求、周六夜电与“做成了有人在乎吗” 在决定是否投资一家公司前，埃里克·维什里亚（Eric Vishria）会在内心做一套极其严苛的思想实验。第一道坎是“招募顶级人才测试”：他会问自己，能否坦诚、理性且毫无违心之感地去说服一位顶尖人才离开舒适区加入这家初创公司，并向对方证明这足以成为其一生的事业？如果自己都无法做到在理智上对自己完全诚实，那这笔投资就绝不该做，因为这说明项目本身根本不具备这种量级的号召力。只要一个项目具备改变世界的潜力，且创始人足够特别，具体做什么赛道其实没那么重要。 第二道坎则是极具人情味的“周六晚九点绿键测试”（The Green Button Test）：如果这位创始人在周六晚上 9 点打来电话，自己会不会毫不犹豫地按下绿色接听键？这关乎合伙人与创始人之间深层的化学反应与心理契约，且这种信任必须是双向的。第三道坎则是冷酷的商业价值拷问：“即便做对了，这件事真的重要吗？”维什里亚提到，许多创业点子在商业上完全可以做成，他前一周还跟一位创始人直言：“我相信你能做成一家极其出色的公司，但你绝对不该拿风险投资。”如果哪怕你全盘做对了也根本没人真正在意，就无法沉淀出足够庞大的股权价值。 - **三道核心检验问题**： - **人生事业测试**：能否凭良心说服顶级牛人将此作为毕生追求（Life's work）而加入？ - **绿键测试**：周六晚 9 点创始人的来电，是否愿意毫不犹豫按下接听键？ - **重要性测试**：“If we're right, does it matter?”（做成了有人在乎吗？若无人在意，就无法创造巨额股权价值）。 --- ### `51:13` 风险投资的范式转移：Benchmark 为何破例募集成长基金 当前科技领域的赌注规模与资本杠杆远超以往，最直接的体现便是融资金额激增与估值飙升。维什里亚剖析了 Benchmark 打破多年惯例募集成长基金背后的逻辑。他回顾了耶鲁捐赠基金主管大卫·斯文森（David Swensen）确立的 LP 投资范式：机构配置风险投资，根本不是为了跑赢纳斯达克 3 到 5 个百分点，而是为了追求极端疯狂的现金回报倍数（Cash-on-cash multiples / MOIC）。 在风投行业的大部分历史中，“早期投资”与“超高回报倍数”几乎是完全重叠的同心圆——要拿百倍回报，唯一的路就是投早期。然而近几年这一底层逻辑发生了质变：终端市场空间被极大地撑开，百亿美元乃至千亿美元级的结果频频出现，使得“高现金回报倍数”的圆环已经溢出了早期投资的边界。尽管能赚取 100 倍回报的后期公司依然极其罕见，但在早期之外，确实出现了能提供惊人回报的确定性机会。维什里亚承认 Benchmark 迈出这步可能“晚了几年”，但该做就必须去做，同时 Benchmark 最核心的高信念度、高承诺度和紧密合伙制文化绝不能变。 ```mermaid flowchart LR accTitle: VC回报范式的演变 accDescr: 展示从历史上的早期与高倍数重叠，到当下成长阶段也出现高回报倍数的机会外溢。 subgraph Past ["历史范式 (过去数十年)"] direction TB A1["早期投资 (Early Stage)"] <--> A2["高现金回报倍数 (High MOIC)"] A3["二者几乎完全等同重叠"] end subgraph Present ["当下范式 (超级终端市场)"] direction TB B1["超大市场与终局上限"] --> B2["早期投资圈"] B1 --> B3["成长阶段超级赢家"] B2 & B3 --> B4["高现金回报倍数机会扩大"] end ``` 在 Benchmark 内部讨论是否设立成长基金时，争论极其激烈。有人质疑这是否只是重演 1999 年互联网泡沫或 2020 年市场狂热时的“非理性线性外推”，也有合伙人主张应严守能力圈。维什里亚指出，过去几年内部最大的痛点在于：团队曾多次对某些非早期项目有着极度准确的商业直觉，却因为条条框框限制在“盒子之外（Outside the box）”而眼睁睁放弃，事后看这种固步自封显然是愚蠢的。要破局的关键不在于贪图管理规模，而在于必须搭建起具备全新评估心智和战斗力的专门团队，去狙击那些极其罕见、可能成为脱缰野马般的超级标的。 --- ### `55:39` 顶级赢家的启示：Cerebras 的波折、软硬件差异与“高尔夫球论” 从亲自见证的 20 倍乃至 100 倍回报案列中，维什里亚提炼出了核心规律：与极其特别的人共事、极度努力、聪明工作，然后等待运气的降临。很多决定命运的因素纯属时机，完全不受公司掌控。以芯片独角兽 Cerebras 为例，团队在 2024 年就曾尝试推进上市，当时预期的估值要低得多，但因美国外资投资委员会（CFIUS）的审查等宏观地缘政治因素被迫搁浅；直到 2024 年 5 月再次冲刺时，这 18 个月的推移改变了一切——不仅大模型推理需求爆发，Cerebras 自身的推理性能也全面跑通，时机的变化带来了天壤之别的结果。 维什里亚着重强调了软件与硬件公司的根本差别： - **软件公司**：除依托 AWS 等云基础设施外，底层技术栈与产品走向几乎全在掌控之中，能够完全把握自身命运。 - **硬件公司**：深嵌在庞大且脆弱的全球供应链中，高带宽内存（HBM）、DRAM、台积电（TSMC）晶圆代工，环环跨越敏感的地缘边界，极易受国际政治与监管波及，不可控变量极多。 谈及努力与运气的关系，维什里亚分享了自己 20 岁在投行工作时的经历。当时马克·安德森（Marc Andreessen）与本·霍洛维茨（Ben Horowitz）刚创办处于隐身模式的 LoudCloud，霍洛维茨给了维什里亚一个做其助理的 Offer。当时一位年仅 25 岁却如同长辈般的投资银行经理问他：“你打高尔夫吗？”在维什里亚表示不打后，对方给出了一个影响他至今的隐喻：打三杆洞（Par 3）时，日复一日刻苦练习，把每一记击球不断打到旗杆附近，这叫努力和聪明；而最终那一记一杆进洞（Hole-in-one），纯粹是运气。人无法掌控运气，但可以通过一次又一次把球打到洞口极近的地方，来人为扩大运气降临的概率。这也正是 Benchmark 风格的写照：每位合伙人每年只投 1 到 2 家公司，维什里亚在 Benchmark 的 12 年间总共仅投资了 18 家企业，保持极高胜负欲与利益绑定，静待非凡发生。 --- ### `58:38` 走向公开市场：从大学联赛到职业赛场，以及 SaaS 的估值大逃杀 关于企业为何以及何时应该上市，前一晚 Benchmark 的周一例行合伙人晚宴恰好邀请了一位管理数百亿美元估值非上市公司的 CEO，这让维什里亚对这一话题深有感触。他认为，上市能赋予企业完全不同层级的武器库：极高的透明度带来的公众信任、可用于并购扩张的股票货币，以及公开市场无与伦比的持续融资能力。维什里亚特别指出，像前沿大模型实验室（Foundation Model Labs）这类巨头最终必然会走向公开市场，除了庞大的资本饥渴外，让全美和全世界公开审视其发展轨迹以获取社会信任，本身就是极有益的机制。 维什里亚将上市比作大学运动员进军职业联赛：职业赛场当然更艰难、对手更强壮、节奏更快、外界审视与压力也更严苛，但想要追求更高水准的竞争，就必须“Go pro”。虽然市场上存在极少数靠着自身充沛自由现金流和顶尖执行力长年无需上市的庞然大物，但对绝大多数企业而言，走向公开市场才是正道。 ```mermaid flowchart TD accTitle: 软件与AI时代的IPO及资本困境 accDescr: 对比2021年SaaS上市公司估值压缩、大量未上市SaaS流动性锁死与AI原生公司虹吸资金的现状。 subgraph SaaS_Struggle ["传统 SaaS 的估值困境"] A["2021年 IPO 潮 (30x 市销率)"] --> B["业务规模增长 4 倍"] B --> C["估值倍数严重压缩至 6x"] C --> D["股价仍大幅破发/处于水下"] E["500+ 家未上市中型 SaaS (ARR $1亿-$5亿)"] --> F["员工与早期投资人缺乏二级退市通道"] F --> G["流动性完全锁死 (Stuck)"] end subgraph AI_Shift ["AI 原生企业的资本虹吸"] H["AI-Native 新兴企业"] --> I["超高增长率展现绝对统治力"] I --> J["吸干市场全部流动性与关注度"] end SaaS_Struggle -.->|资金与预期转移| AI_Shift ``` 反观资本周期的残酷窗口期，维什里亚列出了一组惨烈的对比数据： - **2021 年上市的一大批 SaaS 企业**：当年往往以 30 倍市销率（P/S）狂飙上市，随后遭遇极其严重的估值倍数压缩；哪怕过去几年公司实际营收规模翻了 4 倍，市销率却被无情杀到 6 倍，导致股价至今深陷水下。 - **500 多家估值在 1 亿至 5 亿美元之间的未上市 SaaS 公司**：既无法上市，也没有年度股权回购（Tender offers），员工期权无法套现，早期投资人困在局中进退维谷。维什里亚坦言这些软件公司不至于全部死掉，但残酷的现实是，AI 原生企业正凭借惊人的爆炸性增长率，吸走公开与私募市场上的全部资金与注意力。 ### `01:00:00` Benchmark 合伙人内部的激烈辩论：AI 价值将流向何方？ 在 Benchmark 的合伙人周会上，内部辩论始终保持着极其健康且高强度的碰撞。当主持人问及当前合伙人们争论最激烈的话题时，埃里克·维什里亚（Eric Vishria）直言，核心争论点完全围绕着 AI 的价值分配：在基础大模型（Foundation Models）、基础设施层（Infra）与上层应用（Apps）构成的生态中，价值究竟会在哪里沉淀？各自的护城河（Moats）到底是什么？不仅是技术架构，商业模式的创新同样处于风暴中心。 维什里亚指出，很多人忽视了 SaaS 之所以能够击败传统本地软件（On-premise Software）的本质原因——它不仅是一种更高效的软件交付方式，更是一场深刻的“商业模式创新”。订阅制（Subscription）在软件厂商与客户之间创造了一种双赢格局。如今在 AI 浪潮中，这种底层商业范式正在重演，最具潜力的演进方向便是转向“按成果付费”（Selling by Outcome）。然而，伴随这一探讨的深层疑问是：巨额价值是否最终会被大模型实验室（Labs）尽数收割？抑或是被底层半导体巨头（Semis）全盘通吃？ ```mermaid flowchart TD accTitle: AI生态价值分配的多维辩论 accDescr: 展示Benchmark内部围绕基础模型、芯片、基础设施和应用层价值沉淀的争论维度 subgraph Hardware_Infra["底层与基础设施"] A["半导体芯片巨头与初创公司<br/>(NVIDIA / Cerebras)"] B["云厂商与推理平台<br/>(CSPs / Neo-clouds / Fireworks)"] end subgraph Models_Apps["模型与业务创新"] C["前沿大模型实验室<br/>(Labs: 价值是否被其垄断?)"] D["应用层与成果付费模式<br/>(Outcome-based Apps)"] end A <--> B B <--> C C <--> D ``` - **SaaS 的真正启示**：SaaS 改变世界的关键不仅在于云端交付，更在于重塑了利益机制的订阅制；AI 时代的等价变革正在走向以实际产出/结果为基准的收费架构。 - **价值流向的疑问**：产业界普遍焦虑于价值究竟会沉淀于顶层实验室，还是被垄断算力的半导体企业吞噬。 --- ### `01:02:16` 破除零和博弈思维：为什么说“所有环节都能跑通” 面对“谁会吃掉最大蛋糕”的提问，维什里亚抛出了一个旗帜鲜明的逆向观点：“我认为这一切都能跑得通（It all works）。”当下的市场充斥着极其狭隘的零和博弈思维（Zero-Sum Thinking）——人们习惯于将市场视为一个固定大小的饼，不断计算谁分走多少，甚至充斥着极端论调，认为“Anthropic 这类大模型公司将吞下 98% 的产业价值，顺便包揽所有的药物研发”。维什里亚对此不以为然，直接反驳道：“得了吧，现实绝不会是那样。” 在维什里亚描绘的未来图景中，生态的各个层级都拥有巨大的生存空间：传统的超大规模云服务商（CSPs）会活得很好，部分新兴的 Neo-clouds 也能大放异彩；以 Fireworks 为代表的高性能推理平台大有可为，NVIDIA 以及一部分芯片初创公司会持续盈利；与此同时，端侧手机上的边缘推理（Edge Inference）、网络接入点（POPs）的近边缘推理、以及数据中心内的超大模型将会并存。但他特别强调，**“整个生态能跑通”绝不等于“其中的每一家公司都能成功”**。事实恰恰相反，各层的大多数平庸企业都会被淘汰出局。在当下环境里，企业比以往任何时候都更需要真正的技术与产品差异化，创业者必须将自己的战略假设推演至“逻辑极致”（Logical Extreme），才能在洗牌中立于不败之地。 - **多元并存的推理形态**：手机端侧本地计算、边缘 POP 节点分发与数据中心集群将协同运作，而非单一计算范式垄断。 - **分层繁荣不等于普惠成功**：赛道广阔并不代表进入者皆能分一杯羹；缺乏底层技术护城河的模仿型企业绝大多数将面临倒闭。 --- ### `01:03:35` 杰弗里·辛顿对放射科的误判：技术确定论与现实世界的摩擦力 维什里亚随后谈到了科技界普遍存在的认知偏差：许多极其聪明的顶级专家，往往容易陷入确定性的技术推论，例如笃信 AI 将迅速导致“大规模失业”。他举了一个最具代表性的生动案例——深度学习先驱杰弗里·辛顿（Geoffrey Hinton）在 2016 年曾公开发表著名论断：“我们现在应该停止培养放射科医生，因为 AI 很快会在读片上全面超越人类。”维什里亚调侃道：“辛顿比我聪明三个数量级，但在这个预测上，他错得不能再离谱了。” 辛顿的技术出发点在逻辑上完全正确——AI 确实能够在医疗影像识别上达到甚至超越人类的水平，后续的科研也印证了这一点（Benchmark 也投资了该领域的初创公司 New Lantern）。然而，辛顿的预测之所以在现实中彻底破产，是因为他忽视了真实世界复杂的结构性阻力与经济规律： 1. **多模态与分散的数据壁垒**：现实中不存在一套高度整合的全能放射学训练集。初创企业通常只能在单一垂直病种（如肺部胸片 CT）上训练模型，但现实中一位普通放射科医生每天需要审查几十种不同类型的影像——涵盖全身各部位的 X 光片、CT、核磁共振（MRI）等。只精通单一部位 CT 的 AI 插件，在医生每天处理的几十个多样化病例中，只能提供微乎其微的辅助价值。 2. **医疗体制与责任绑定的现实摩擦**：整个医疗健康体系是围绕着“医生出具诊断报告并据此获得医保与商业保险报销”来构建的。一旦发生漏诊或误诊，将面临严重的医疗事故诉讼（Medical Malpractice）。AI 无法承担法律责任，制度与保险机制在短时间内根本无法绕开人类医生。 3. **“杰文斯悖论”（Jevons Paradox）的重演**：在很长一段过渡期内，放射科将处于人机协同的“副驾驶”（Co-pilot）模式，医生与 AI 相互校验，从而提升读片通量。更为关键的是，随着 AI 降低了单次影像诊断的成本，根据杰文斯悖论，整个社会对影像检查的需求反而呈现爆发式增长。结果不仅没有消灭岗位，社会反而需要*更多*而不是更少的放射科医生。 维什里亚总结道，如果只看到技术的潜在上限，却脱离现实世界的落地环境、制度摩擦与经济激励机制去推导，即便拥有最顶尖的技术认知，也会得出完全错误的社会结论。当下科技界对“AI 导致全面大失业”的恐慌叙事，几乎完全踏入了当年辛顿对放射科医生预测的同一思维陷阱。 ## Notable quotes > `02:20` 等一下，面对同样的开源模型和同样的英伟达硬件，怎么可能出现 5 倍的速度差距和数倍的吞吐量鸿沟？ — Eric Vishria > `05:42` 天呐，AWS 将吞噬整个软件产业（AWS is going to eat everything）。 — Eric Vishria > `07:40` 如果这一切真的全部跑通了呢？（What if it all works?） — Eric Vishria > `13:05` 我们过去是在建造坚固的城堡，而现在我们是在建造随时会被海浪冲垮的沙堡。 — Bret Taylor > `20:00` 你们按部就班达成年度计划的每一天，实际上都在摧毁股东价值。 — Eric Vishria > `30:00` GPU 在深度学习任务上其实很烂，它只是碰巧比 CPU 快了 100 倍而已。 — Cerebras 创始人 > `36:18` 这下完了，我们要把这 5 亿美元全赔光了，这套方案根本行不通。 — Eric Vishria > `36:18` 如果它成功了，会是怎样的景象？（What could go right?）我们是否看得清成功的路径？ — Bruce Dunlevie > `36:18` 如果一个项目失败了，往往是因为合伙人们提前预警过的所有理由应验了；但如果它奇迹般地成功了，一定是因为那些致命的反对理由最终都变得不再重要。 > `50:00` If we're right, does it matter?（做成了有人在乎吗？若无人在意，就无法创造巨额股权价值）。 — Eric Vishria > `01:03:35` 辛顿比我聪明三个数量级，但在这个预测上，他错得不能再离谱了。 — Eric Vishria ## Mentioned Eric Vishria, Benchmark, Fireworks, Sierra, Cerebras, AWS, Amazon, Jeff Bezos, Microsoft Azure, Google Cloud Platform (GCP), NVIDIA, Snowflake, Redshift, Confluent, Elastic, MongoDB, Databricks, Datadog, Cloudflare, Anthropic, OpenAI, Snapchat, Cursor, Bret Taylor, Peter Fenton, Michael (Cursor), Oracle, SQL Server, Claude, Codex, Ann Miura-Ko (Anne Lee Skates), Lin Qiao, Baseten, Modal, TSMC, Intel, Broadcom, Avago, Qualcomm, Arm, Groq, Bruce Dunlevie, Sunday Robotics, Waymo, Tesla, Jasper, Benchling, Michael Moritz, John Doerr, Bill Gurley, David Swensen, Marc Andreessen, Ben Horowitz, LoudCloud, CFIUS, Geoffrey Hinton, New Lantern ## Video description My guest today is Eric Vishria, a General Partner at Benchmark.&nbsp; Eric has spent his career in software and cloud, and few people know the history of these markets as well as he does. What makes him special is his ability to use that history to make sense of today.&nbsp; We discuss what the rise of AWS teaches us about AI, what he has learned from investing in Fireworks, Sierra, and Cerebras, and how the criteria for winning have changed for founders and investors.&nbsp; Please enjoy my conversation with Eric Vishria. For the full show notes, transcript, and links to mentioned content, check out the episode page ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠here⁠⁠⁠⁠⁠.&nbsp; ----- Become a Colossus member to get our quarterly print magazine and private audio experience, including exclusive profiles and early access to select episodes. Subscribe at ⁠colossus.com/subscribe⁠. ----- ⁠Ramp’s⁠ mission is to help companies manage their spend in a way that reduces expenses and frees up time for teams to work on more valuable projects. Go to⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ ⁠ramp.com/invest⁠⁠ to sign up for free and get a $250 welcome bonus. ----- Trusted by thousands of businesses, ⁠Vanta⁠ continuously monitors your security posture and streamlines audits so you can win enterprise deals and build customer trust without the traditional overhead. Invest Like the Best listeners get a special offer of $1,000 off Vanta when you go to ⁠vanta.com/invest⁠.&nbsp; ----- WorkOS⁠ is the infrastructure B2B and AI-native companies use to sell to enterprise. It covers everything enterprise security requires: SSO, SCIM, RBAC, Audit Logs, AI governance, and more. Trusted by 2,000+ fast-growing companies, including OpenAI, Anthropic, Cursor, and Vercel. ----- Rogo is the AI platform for finance. They're building agents for Wall Street that are trained to understand how bankers and investors actually do work: from diligence and modeling, to turning analysis into deliverables. To learn more, visit rogo.ai/invest. ----- ⁠Ridgeline⁠ has built a complete, real-time, modern operating system for investment managers. It handles trading, portfolio management, compliance, customer reporting, and much more through an all-in-one real-time cloud platform. Visit⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ridgeline.ai⁠. ----- Editing and post-production work for this episode was provided by The Podcast Consultant. Timestamps: (00:00:00) Welcome to Invest Like The Best (00:02:20) Learning the World Through Fireworks (00:05:42) AWS Was Going to Eat Everything (00:07:40) The Zero-Sum Thinking Trap (00:09:01) Comparing Cloud and AI Adoption (00:11:03) Becoming Enterprise's AI Sherpa (00:13:05) Building Sandcastles (00:14:55) The Return to Being Technical (00:17:13) The Shifting Competitive Frontier (00:22:10) Why the Old Playbook Fails (00:27:53) Energy as the Binding Constraint (00:29:38) The Cerebras Story (00:37:57) The Virtue of Productive Naivete (00:39:19) What Robotics Still Needs (00:45:58) What Makes a Great Board Partner (00:51:13) Raising A Growth Fund (00:55:39) What the Big Winners Taught Him (00:57:37) Hard Work Versus the Hole-in-One (00:58:38) The Best Reasons to Go Public (01:01:09) Debates Inside Benchmark (01:02:16) What If It All Works (01:03:35) What Geoff Hinton Got Wrong [Source](https://traffic.megaphone.fm/CLS6500696931.mp3)

[2:20] Eric Vishria: 等一下，面对同样的开源模型和同样的英伟达硬件，怎么可能出现 5 倍的速度差距和数倍的吞吐量鸿沟？ [5:42] Eric Vishria: 天呐，AWS 将吞噬整个软件产业（AWS is going to eat everything）。 [7:40] Eric Vishria: 如果这一切真的全部跑通了呢？（What if it all works?） [13:05] Bret Taylor: 我们过去是在建造坚固的城堡，而现在我们是在建造随时会被海浪冲垮的沙堡。 [20:00] Eric Vishria: 你们按部就班达成年度计划的每一天，实际上都在摧毁股东价值。 [30:00] Cerebras 创始人: GPU 在深度学习任务上其实很烂，它只是碰巧比 CPU 快了 100 倍而已。 [36:18] Eric Vishria: 这下完了，我们要把这 5 亿美元全赔光了，这套方案根本行不通。 [36:18] Bruce Dunlevie: 如果它成功了，会是怎样的景象？（What could go right?）我们是否看得清成功的路径？ [36:18] 如果一个项目失败了，往往是因为合伙人们提前预警过的所有理由应验了；但如果它奇迹般地成功了，一定是因为那些致命的反对理由最终都变得不再重要。 [50:00] Eric Vishria: If we're right, does it matter?（做成了有人在乎吗？若无人在意，就无法创造巨额股权价值）。 [63:35] Eric Vishria: 辛顿比我聪明三个数量级，但在这个预测上，他错得不能再离谱了。

[20:00] Eric Vishria: Get to AI or be worth three times revenue. [35:00] Eric Vishria: If it doesn't work, it'll be for all of the reasons that your partner said. If it does work, it will be because those reasons didn't matter. [30:00] Eric Vishria: In software, if you have that logical block diagram of why it works, you're 80% of the way there. In hardware, you're 2% of the way there. [40:00] Eric Vishria: You go down into the basement... and there was like a dozen robots just folding arbitrary laundry. It wasn't in the demo—it was just trial and error... And I was like, 'Oh my God, this is happening.' [50:00] Eric Vishria: Could I talk one of the people I care about into going to this company, and honestly, intellectually explain to them why this could be their life's work? [50:00] Eric Vishria: If we're right, does it matter? [55:00] Eric Vishria: With golf, you keep on practicing, keep getting the ball on a 3-par close to the pin... That's hard work. That's working smart. Getting the hole-in-one, that's luck.

[0:00] Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all run on Ramp. Mine does too, and so should yours. Learn more at ramp.com slash invest. They all use WorkOS. To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SCIM, RBAC, and audit logs. Instead of spending months building these mission-critical capabilities yourself, you can just use WorkOS APIs to gain all of them on day zero. That's why so many of the top AI teams you hear about already run on WorkOS. WorkOS is the fastest way to become enterprise-ready and stay focused on what matters most, your product. Visit WorkOS.com to get started. by Rogo is a personal finance agent that turns a single prompt into finished client ready work using your firm's own templates, context, and standards. Send Felix an email like take these comments and turn them for me or update my tracker with the context of these emails. And Felix sends back finished PowerPoint decks, Excel models, and sourced research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai.felix. Hello and welcome, everyone. I'm Patrick O'Shaughnessy, and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, stories, and strategies that will help you better invest both your time and your money. If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus along with all of our podcasts at colossus.com. Patrick O'Shaughnessy is the CEO of Positive Thank you. I love asking you and all your partners this every time we hang out, which is, okay, you've got these singular investments. You don't do that many investments each per year. And then the ones that go on to work so far, Sierra and Fireworks certainly are. You get to learn so much about the world through the lens of the company. So I'd actually love to do both of those, maybe starting with Fireworks. What do you know or what have you learned about the world and how it's reordering itself based on watching the world through the lens of fireworks that would maybe be surprising or interesting? One really interesting thing. is these models are big. These are 2 trillion, 3 trillion, 4 trillion parameter models. It turns out running those models... is damn hard. And running them efficiently is like super hard. The way to see this, and everybody can see this, which is... Everybody from AWS to Azure to GCP to the NeoClouds to the fireworks based entogethers of the world, like all of them. They all run the stock market. open source models that are available in their developer pools and everything else. And that's what it is. The performance difference for a fireworks versus a cloud provider. is like 5x. And that is just the speed performance. Then you add on top of that, throughput, which is not visible externally, only visible if you know the economics of these businesses. And you're like, wait a minute, this is the same open source model with the same NVIDIA hardware, and there's a 5x performance difference and a multiple x throughput difference? And the way to just simply understand that is… These companies are paying money. the margins of the cloud providers and running on top and making money. How can that be? My big takeaway on it was, wow, this stuff is actually really hard to run. It's just really hard to run. And there's a lot of expertise involved in doing that. And it's this very specific expertise that exists. And it is the kind of thing that when you look at it as an investor from the outside, and we should talk about early days of AWS, but when you look at it from the outside, it's like, this is a commodity. This is just as like pass through No, it turns out it isn't. It isn't at all. Do you think that's just a moment in time thing? And I'd love to just hear you riff on like cloud. You watch cloud very carefully and closely. You know a lot about it. And the adoption curve there versus how people use these things and the nature of those two businesses in comparison. It's tempting to say there will be one or two scale winners like there typically have been in a commodity market. Totally. Where cost to serve is everything and scale drives cost to serve down. And like that's the whole story. I think the AWS example is so good. Okay, so... 2016. [5:00] Six. You have S3 and EC2 watch, right? Their compute platform and their storage platform. Those are the first two AWS offerings in 2006. And you kind of start talking about it in late 2006 or whatever. 2006, 2007, I think the 2007 annual letter Bezos talks a lot about AWS and why it's important, why it's interesting and everything else. And the investor reaction, It's just not good. I think if you put 30 of the smartest investors at that time in a room, and ask them, what's the probability that this AWS business is a good business with durable long-term margins and like super interesting and everything not commodity. I think you would have gone zero for 30 with really smart people that you and I know who were around at that time in 07. Fast forward from 07 to 2014. I joined the venture business in 2014 and a really common narrative in 2014 was oh my God, AWS is going to eat everything. There's no enterprise opportunity left. It's going to eat databases and infrastructure, but it's going to eat the apps too. And they're going to offer it the cheapest and best. And we all have these like amazing SaaS and software businesses that we were involved with or investors. And part of the reason people loved them was they were annuities and they ran at 85% gross margins and everything else. And so it was just like, oh my God, AWS, Amazon can offer things at 8% gross margin. And like, they're just going to the whole thing, whole narrative. And think about from like 2014 to now. in enterprise. Of course, you have Snowflake, direct competitor to Amazon Redshift, ran on Amazon. You're out Amazon-ing Amazon on Amazon. But it's not just snowflake. You had Confluent and Elastic and Mongo. Databricks. All of these companies, amazing. That's the infrastructure layer. Then you have the whole app layer. in the app layer think about offerings that they offered at the beginning Datadoc, $100 billion company today. They had a competitive offer. And they did it. And of course, there was tons and tons of roadkill. There was tons of roadkill. They did run over a bunch of stuff. But even then, in 2014, the thesis was the view that AWS was going to eat everything was massively wrong, not because of all the examples I just mentioned. It was massively wrong because Azure and GCP were irrelevant then. And fast forward to 2026 and they're unbelievable businesses. Is AWS the biggest? I think it's like a 40, 30, 20 split. You ended up with an oligopoly of that. And even outside of those big three, you have Cloudflare, which is... a cloud provider in a different sort, which is another $100 billion company. So you have these smaller players that emerge as $100 billion companies outside of it. So what's the takeaway? The takeaway to me is, oh, there's just a bunch of zero sum thinking and not realizing like how big. What if it all works? What if it all works? It all worked. And of course, getting the relative winner right matters. And there was roadkill. And so it's all those things still matters. I'm not saying spray and pray. I'm not saying that at all. But I'm just saying the market was so big that one vendor could not scale and consume it all. And they just couldn't consume the whole industry. It's different now and like all these things. But just that whole notion right now with what you and I are seeing in AI sure feels like it rhymes. Anthropics going to do everything. Right. Really? Is that really right? To me, that view that it's just like, oh, this one company is going to eat it all doesn't hold. And I would tell you that scaling while cloud scaled very quickly, cloud did not scale anywhere close to as quick as what's happening right now. for that company to actually scale and deliver it. Scaling in this case requires a a ton of infrastructure buildup from energy, power, shell, chips, memory, obviously the algorithms and everything else on top. It feels to me like we're going to end up with an oligopoly. of winners. I really believe they will be these like $100 billion crazy smaller winners. It just feels like the same thing is kind of happening. Can you talk about it also from the demand side and compare it to how you watched cloud get adopted in enterprise versus how you're seeing AI get adopted now? So if you go back to like 2010, 2011, now you're like three, four years into AWS being an offering. Enterprises were super skeptical, traditional enterprise, blue chip enterprise. You had digital natives out here. You had new companies forming that were using the cloud. I think kind of famously, Snapchat was built on GCP. And I think at one point in this era, maybe 2012, 2013, 2014, Snapchat was like 40% of GCP. Those kinds of things were happening. That's like cursor being 30% of all the things. A hundred percent. Same thing. Same exact thing happened. Yeah. Same exact thing happened. Enterprises were very skeptical, I think, of cloud until by 2014, 2015, 2016, it was like, oh, yeah, yeah. Then I think the big banks and financial services and insurance companies and more conservative blue chip enterprise were like, oh, wait a minute, this is out. [10:00] actually different and we're gonna have to pay attention and It became an issue in recruiting for them because they couldn't get the best developers because the best developers wanted to work. on the easiest platforms and all these kinds of things happened. So the difference now feels profound to me. Because while it is 100% the case that blue chip enterprise AI is not well absorbed and well adopted and not the same thing as going to cursor and walking the halls of cursor versus walking the halls of a big computer. New York financial services firm. in terms of their AI use, they want it. They want to figure it out. They're running experiments. They are spending against it. They're trying to figure it out, talking about it, They're not being dismissive about it. And I think. They view it probably as more of an opportunity than they did the cloud in terms of the potential impact on their business. I think they probably view it more as a threat. Then they did. the cloud, and maybe they just also learn lessons from the cloud in terms of what's possible here. I feel that they will continue to try to figure it out and absorb it. Having said all that, I do think one of the more interesting things is, If you go from Silicon Valley out into the world and talk to these companies, these enterprise companies, and you realize what the pace of adoption is and what the barriers for adoption are and everything else, there's a ton of opportunity to help the enterprises get there. To be one of the things that I tell the companies I work on is, hey, let's be their AI Sherpa. If we're in that position to be their AI Sherpa, where we're crossing both worlds, that's very valuable. And I think that'll continue to work. What is Sierra teaching you about the adoption of this stuff? It's really interesting contrast to Fireworks, where Fireworks is the sort of infrastructure provider. Sierra is feeling its way through what are really cool things that we can do for end consumers, enable companies to do for end consumers, starting with customer service. I know with Horizon now going beyond that. Again, same question as for fireworks. What do you know that the world doesn't fully appreciate because of how you've seen that My partner, Peter and Brett, I think this is their third company working together. So it's like a 20 year relationship, which is just like an amazing and great place to start. And one of the things that I think makes Road to special and the team is they're technologists. But they actually have lived in enterprise world for a long time and really understand it and everything else. And I think actually he's personified this whole idea of, hey, let's be their AI Sherpa. Let's start with customer service. We're very automatable and be their AI Sherpa. And now we have these long-running agents with Horizon that can do more and more stuff. And one of the things that I think is most interesting about this to me is, yes, it's an application company. on the surface, but they're doing real AI work and they're experimenting with the models and they're building agents and they're very close to the metal of the models and the capability and the harnesses. And they're understanding the jagged edge of AI capabilities and which is very different than the human smooth arc that we understand intuitively. They understand that jagged edge of capability and they build around it. You saw the evolution of cursor from the IDE to like tab auto complete to a gentic work over and over and over again. They were... and obsoleting their work from six months ago. And it's what Brett Taylor calls like sandcastles. We used to be building castles. Now we're building sandcastles that get launched away. And that's if you don't think that way. Yeah, you have to embrace it because if you're an artisan and you're like, hey, I built these perfect foundation of castle and there's bricks and it's like perfect. And I really care about this and it's going to be here for 100 years. You're just not going to make it. So as you get emergent new properties in the models, which is, you know, every four weeks, you get new capabilities. They understand that jagged edge. They understand that application of that jagged edge. or that valley to their customer base. They're filling in those gaps and translating it. You can't just be superficially applying things. I think it's actually a complete inversion of how product development used to work. to how product development happens now. In product development, you'd also say like, hey, product manager, you're trained. Oh, the product manager should understand the technology, but like really shouldn't be thinking about implementation and shouldn't be specifying implementation and shouldn't be doing this and doing that. And the product manager's job is to really understand the customer and translate that problem to the engineer so that the engineers build a solution like that was traditional product management. Good luck doing that now. That's a horrible way to do it. You can't do it that way. You actually really need to understand the nuances of model capabilities, what they're great at and where they fail. And you need to understand the customer problem. and put those together. and bridge those gaps to build valuable solutions. I think that's one of the things that Michael and team at cursor did so well. from the very beginning is they really understood the jagged capability and built a product that allowed that translation from developer to that jagged capability. and kept on iterating on that as the edge changed. It's kind of ironic that all that sounds like the returns to being technical [15:00] are going up. Even as models supposedly are taking away technical edge. Well, I think there's two things. Yeah, 100%. And actually, I think there's this weird thing where... There was this whole discussion about the traditional roles of product manager and designer and engineer, whatever it is. I think what there really are. are people who understand customer problems, people who have taste and people who understand the jagged edge of AI capabilities and are curious about it. Those are the three things. If you've got those three, you're going to do wrong. If you've got those three things, you're going to do great. And it doesn't really matter if you are an engineer or a product manager or a designer, but if you have taste and understanding of customer problems and understanding the jagged edge, You're going to do well. Vanta automates security and compliance for over 16,000 fast-moving companies like Ramp, Cursor, and Harvey, keeping them audit-ready around the clock. It's the number one agentic trust platform, and it now helps companies like yours watch for the risks that show up between audits across your vendors, your AI tools, and your whole environment. Every new tool your team signs up for, every vendor that turns on AI features, is an opportunity for something to go wrong. and most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24/7 GRC engineer in the background, finding issues, drafting fixes for you, and cutting vendor assessment time by up to 50%. Whether you're a fast-growing startup or a global enterprise, Vanta helps you earn and prove trust. Invest like the best listeners get a special offer for $1,000 off. at vanta.com slash invest. Ridgeline is the first end-to-end system of record with embedded AI for investment management firms running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared to anything else in investment management software. Which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline's unified platform. If you're serious about your firm's AI strategy, Ridgeline should be part of that conversation. You can request a demo at Ridgeline.ai. you When we first did this so many years ago now, it's just crazy. It'd be fun to revisit some of the ideas that we talked about the first time with SaaS. But you use this term that I've used ever since, which you call it the competitive frontier, meaning like the things that will determine the winners and the losers. My partner, Bree, has this great idea that's stuck in my head, which is that everything is a jump ball right now. I'm really curious, in addition to this idea of sandcastles versus real castles, what else you're seeing here? amongst the people that are becoming competitive winners. Are the traits different for winners now, personality-wise or business strategy-wise or business model-wise versus what you learned in the SaaS era? I'm very dismissive of this idea that people are going to vibe code their own shit and whatever. That isn't the issue at all with SaaS companies. The issue with SaaS companies is competitive frontier completely shifted and everything they thought they were building against and it would make them win consistently. are not what's going to make them win. Let's say databases. Databases have been a phenomenal area for software for a long, long time. Great margins why you had app developers that would build against these specific database interfaces that existed for that database and you would have more and more data over time migrating an app from one database to another database was a giant project that was like very, very difficult to do. So these were unbelievably sticky businesses that you could generate a ton of margin in, right? Of course, you got Oracle and SQL Server and like, and a whole slew of smaller players like that did really, really well in databases. Well, let's think about that in the context of AI. Now, you don't have a developer building against the database interface. You have cloud or codecs building against the database interface, one. Two, the beauty of database interfaces is they're very, very well specified. Well, it turns out AI is very good at things that are very, very well specified. Three, agents don't get tired of monotonous work of translating one specification to another. So it turns out that now all of a sudden database migration, which used to be the number one thing you would not do in software, is kind of trivial. Yeah. It's just like, put some money against it, move it. So what changed? Well, what changes is the criteria. to be an amazing database company changed. It isn't that we don't need databases or that everyone's going to build their own database or whatever. That isn't what's going to happen. What's going to happen is the criteria changed. So now you're going to have a ton more applications that start. Obviously, as we're seeing everywhere, people are going to experiment a lot more because it's much cheaper to experiment. It's much cheaper to start a new application. So now you need databases that scale from basically zero usage to if it works all the way through, like that matters a lot more. Your cost matters a lot more. You want to be able to spin these things up, spin them down, tear them apart over and [20:00] over again so the iteration speed goes up and what you need on database. And ultimately, I think the cost becomes arbiter of this. So the cost and then this zero to infinity scaling and transportability and everything else around that become like the arbiters of who wins and who doesn't. That's really different than, hey, I specced this database for our user thing and I procured a license and I ran it on this kind of hardware and everything else. There have been elements, of course, of these things over time, but I think that just criteria changed. And I think one of the big messages to these SaaS companies... a few years ago was You have a choice. Get to AI or be worth three times revenue. Those are a hard message to hear. You're just like, get to AI or three times revenue. Those are your choices. We say three times revenue now because a lot of these public SaaS companies are trading at six times or whatever. But keep in mind, in 21, they were going for 30 times, like everybody. You're like three times, wow. I've grown 4X since then. This is a whole multiple compression is a bitch. I've grown 4X. The multiple has gone down by a factor of six. So I'm worth less, even though I've grown four XF years and I've gotten to break even all these things. So like that was the first message. But I think that even it became really visceral to me where you're in these meetings and you're like, hey. Every single day that you are hitting your plan, you are destroying equity value. Think about that. Our whole careers, we learned you lay out a plan, you execute against it relentlessly and violently. You hit your plan or you see your plan and you keep on building that. And that's how you build equity value. That's what the whole management teams learned. That's what the CEOs learned. Like, yeah, this is like everything learned. And now you're in here just every day you hit that plan. You're fucking up. You're fucking up. You're destroying value. And the point of saying that to them was to set them free. Another articulation of this from my friend Anne Leeskates was just the CEOs who were going through this transitory period and had a business that was at hundreds of millions. And they thought, you know, they're all ready. They were working on their business from 8 a.m. to 5 p.m. and then trying to do AI from 5 to 8 in the evenings. And what they needed to be doing, the inverse, the inverse. And it's so hard to do that because of all of the training and all of the muscle memory and all the inertia and everything that we've learned about. the hill that we were climbing. Like we're all hill climbing in a way. The success model was set the plan, execute the plan, build value, compound value, And it's like, oh, no, no, no, stop that. You got to completely invert. There's a very, very long way to get back to You're a question of what the profile or what the mentality of the winners are right now. But if we look at Brendan from McCore or Lynn from fireworks or, or Max or Brett, any of these people, they are so nimble. about what the eval is. What are they optimizing against? They are so nimble on all of it. And if you look at every one of those companies, the evolution of the business, it's just the business is constantly evolving. And they've done such an excellent job at that. I think that is very, very different than the way I was... taught. What's your sense of the disorienting nature of model progress? You're one step removed from that as an investor versus as a technical founder with your hands on the metal. Are we behaving differently than you would have three years ago or something because of the pace? Anytime that I'm talking to a founder about a problem in the company or what they're doing or move they're making or anything else, which is what I spend 80% of my day doing. I'm very... This is how we used to do it. This is what we would typically do. This would be the typical readout on this old school readout of why this candidate is better than this candidate. Let's reevaluate that in the context of today. Let's reevaluate that in the context of an unstable technology substrate. Let's reevaluate that. in the context of a business model that's growing this way versus that way. I've really started to question every assumption and every lesson that I learned before, which of it translates and which of it doesn't. That's a huge difference. I'll give you a really concrete example. which has been very disruptive inside of these scaling AI companies. So these scaling AI companies, we've had a lot of leaders come from the prior generation with great experience and everything else come to these scaling AI companies and completely flame out. And you see it across the industry. And so the question is why these are some of the best leaders from four or five years ago. They had all the lessons they learned at all. They're excellent, but somehow it's not translating. And there's some impedance mismatch between the founders, potentially the needs of the business and what these people are bringing. And I saw it really abruptly with a particular sales leader who we hired. He was like, hey, we can't hit. any of these things, 'cause the way [25:00] Software sales has been taught forever is a quota capacity model. You have a quota capacity model. Each rep does this. In the early days of a company, the quotas are 1.2, 1.5 million. Maybe the ISRs are at 750 or 850. And then over time, it scales up and enterprise gets to 2.5 million. And that's how all these financial models are built. You start with the quota capacity model, you take a discount on attainment. This is what we can do. Boom, boom, boom. Fundamentally, without realizing it, everybody was implementing something that was based on pushing demand, not pulling demand. And for so many of these companies, they're operating here, these customers, and there's a new AI enabled product that comes along and it's just fucking magic. So these companies are selling magic. Well, it turns out if you're selling magic and you're the first one there, you're going to sell a lot more than $2 million. The whole notion of a quota capacity model and that being working, It's not that it doesn't matter. It matters kind of, but it's definitely not the first order thing or constraint. So you have these execs come over with this, like, here's our math and here's the territory assignment and here's what we would do. And First you do West Coast and then you do East Coast and then you do Central, like all of these things. Oh, wait, no, no, it doesn't work like that at all. One of my big things that I started to realize is as I'm interviewing these folks and talking to them is just like, hey, you need to check everything at the door. Check it all, which is probably good practice anyway. But check all the baggage. Check everything that you learn and just learn this from first principles. How is it working? What are really the bottlenecks on delivery? What are the bottlenecks on demand? Because it turns out that in a lot of these companies, you have reps doing 10 or 20, 30. I saw 50 recently. It turns out that's different. What is like the best salesperson that you've seen that's doing it in a de novo way doing? Honestly, the best salesperson in any of these companies is a founder. And what they're doing is bridging the jagged edge to what the customer's capability is. And that's it. It sounds so simple, but it's not. But that's what they're doing. The market is just so big, just like we were talking about with cloud. I think the biggest mistake everybody made was they just undersized the market. And it turns out the market's just really, really big. And this market is bigger. It feels like right in this moment in time, actually just this morning, you and I are in this great group chat together where the discussion is... The demand for intelligence seems kind of unlimited, and it seems like the smarter the thing gets, the more demand there is. Maybe the bottleneck is just capital. The world just feels like it needs to take a breath. The RSI concept, if you apply it across technology, doesn't need to breathe. The agents don't get tired. But the world feels kind of like, oh man, it sure would be nice to have three months just to like digest this a little bit and for capital to form and evaluate its prospects. And the scale is getting so big. Does it feel to you like this can just keep going or are we just going to get tapped out of money that can be invested? in these things when it seems like we could consume like any amount of money to build any amount of stuff and serve any amount of inference I'm not a macro economist. I am worried about energy. If you think about the models as translating compute into intelligence, quite simply, what do models do? They very effectively translate compute into intelligence. What's the demand for intelligence? Well, it seems like a lot. So then it follows that we will continue to have more and more demand on compute. I'm saying compute, probably not chips, not storage, not whatever. But then, like, what do we need for compute? We need energy. a lot of energy. There's all this topic of distillation and Chinese open source and all these things. But the bigger thing to me is that I think China is bringing on 10 times as much energy next year as we are. And the U.S.? If energy – Is what you need for compute and is the bottleneck and there's unlimited demand for intelligence that it stands to reason that if we have a lot less energy then we will have a lot less intelligence or lawless tokens. or a lot more expensive tokens. If we have a lot more expensive tokens than supply-demand, you're going to end up with less. And that seems very bad to me. I think the energy bottleneck. However, that is, it'll manifest in 20 different ways. Gas turbines go up and down, natural gas get up and down, and solar, whatever, all these different things, rare earths. That to me is probably more concerning. And if I'm thinking about it from a regulatory perspective or government perspective, and I think the administration is doing some things around this, fostering investment and development of all energy. solar, nuclear, gas, whatever. Do it all. We should do it all. And it'll work itself out. This is one of these things where, yes, one will be relatively better than the other. And I don't know. And I'm not smart enough to predict which one's which, but it'll all work. Speaking of compute, I would love to hear the Cerebra story. I haven't heard you tell the full version of this. The reason I'm asking about it is I'm deeply interested in compute. I have bigger investments in compute and I'm fascinated by it. It's just like the most magical thing to watch happen. It's Mind boggling when you get close to one of these things, what humans have been able to do. on these chips and in these systems. [30:00] I think you invested in 2016 or thereabouts. I think it was your first foray into like extremely difficult hardware type investment. Shit's so hard. And now the world is full of opportunities like this. Whereas back then it was a one-off. Teach me everything you've learned about hardware investing through Cerebris. Mostly, it's really hard. It's an amazing example to me about... The naivete required. The company came in in 2016. It was five founders and a deck. I did not want to go to the pitch, but it's my job because I'm like, why are we going to go to hardware investment? Like it's crazy. Been 10 years since we've made a semi investment. I think basically the team was excellent. And then. The kind of first slide was just like GPUs actually suck for deep learning. They just happen to be 100 times better. than CPUs. You have to remember, this is pre-Transformer. OpenAI is this weird research lab at this time. NVIDIA was worth like $40 billion, not $4 trillion. The TPU hadn't been announced. None of that. So this is early. But the whole idea, as soon as he said it, I was like, oh, shit, of course. Of course. Like, why? I had spent the last 18 months trying to figure out applications of deep learning and looking at the security thing, looking at this medical imaging thing and like all this other stuff, thinking like, hey, there must be something here that's going to be really transformed by this stuff. Anyways, we go through this whole journey. We end up investing. which was amazing. We first met on Wednesday, part of meeting on Monday, had a bunch of meetings in between, just built a lot of conviction that this was a great swing. And I'll tell you what I understood. And I really just understood so little. But basically, there are three things that we know how to do to speed up deep learning and hardware still till this day. Increase the number of course. increase the communication between cores, bring the memory closer to the compute. Those are the three things. That's it. Those are only three dimensions that we know. and hardware. My articulation of what they said to me, honestly, all I understood was, let's just take all three of those things to their logical maximum. You have a wafer scale chip. At that time, you would have 450,000 cores on it. You'd have something like 20 gig of SRAM on the chip. So you never have to go off chip to get to memory. And because they were all in the same wafer, the communication between cores is maximized. So this is the best you could do on that process. And I think for the first chip was 7 nanometer or something. You're like, okay, that's it. That's what we do. And it turns out that in software, if you have that logical block diagram of like why it works and everything else, you're kind of 80% of the way there. And it's a matter of go to market execution. In hardware, you're like 2% of the way there. There's things like physics, entire supply chain of vendors. There's, of course, TSMC, which everyone knows, but it's not just TSMC. There's 30 other vendors that matter in putting all this stuff together and everything else. I didn't know any of that. Fast forward from 2016 to like, In 2019, I think they got their first parts back. You go through this like bring up and then it's like bring up. Oh, yes, we got a part back. And then it's like, yeah, you got to go through like we bring up and there's like 14 steps of bring up and everything else. And then by like 2020, we had our like first thing that works. Then it's just this march of actually getting it to work. One of the lessons that I've learned on this stuff is basically you go through all these Sims and everything else in hardware and semis in particular. That basically is your roofline. The best it's ever going to be is what that is. And then every bit of software and reality and compilers and kernels takes away from that roofline. You might start at 10% of the roofline. Once you bring it up, you guys are grinding for months and years to get closer and closer and closer to the roofline. It's really different. It's really hard. I'm astonishingly bullish if I kind of rewind. Part of the reason we made the investment was if you looked at the four prior generations of compute, In my lifetime, so you had CPUs, you had graphics, the networking, the mobile. There was a new workload each time so you had multi-purpose compute and it lived to the cpu You had massive parallelism led to the graphics processor. Graphics processor offered massive parallelism, which led to graphics. Then with networking, you needed really low latency chips. And so you had low latency chips. And then with mobile, you needed really power efficient chips. And in each case, we ended up with a new $100 billion company. The first question, going back to 2016, was, is AI that big a new workload? Because there have been many, many other attempts. for specialized chips for other things that really just didn't end up mattering. There's some fine outcomes, but they just didn't really end up mattering, et cetera. Right, exactly. And so it's just like, well, okay, well, you need something that's a really, really big workload. Okay, so that's one. We had a lot of conviction on that. And then two, was the nature of the workload, did it introduce a new constraint or problem in it? What I learned was basically the AI workload benefited from the parallelism of GPUs massively, but GPUs didn't solve the core-to-core communication, basically the layers of the network problem. [35:00] You're like, okay, wait a minute. There is a new constraint, which is communications, communication bound problem. then you're like, okay, is AI a new giant workload that is going to have specialized chips? Everything that I just said was the entire everything I knew at that time. That's obviously played out. And in each prior generation, we got Intel, we got NVIDIA, we got Broadcom and Vago, we got Qualcomm and ARM in each of these generations. And there will be these giant winners, standalone winners. Obviously, the TBU itself is a winner, training is a winner. You've had Grok and Cerebrus etch. It'll keep getting fought out. But I think that wound up being big. And actually, I think there's a new sixth one that's coming, Generation, which is, and I'm really excited we've made an investment that's unannounced in this, but... I think that for the first time in a long time, there's actually room for a new CPU approach. The thing that's happening right now, and you see this reflected in all the semi-stocks and everything else, is the LLMs, which are running on... accelerators and GPUs generate code. The code runs on CPUs. And right now it's running on classic CPEs we've had around forever. But there's a whole bunch of constraints on CPUs that have existed, and CPUs have dragged all this baggage forward. that you might not need to anymore. And so I'm actually really excited about that. Paul next category. The next category. Does the experience with Cerebris make you want to do a lot more investing in companies? No. But why not? In 2019, we're sitting in a board meeting. And this thing is melty. It's like fucking melty. Okay. And we've raised $500 million or something. And it's like, wait, what? It's melting or burning or something. And like, we're looking at it and I'm like, holy shit. I realized like $500 million isn't that much in today's era, but I was just like, we're going to lose all this money. This is not going to work. And what that team did, insane. They're built differently. And I have so much respect and thanks to them. for what they've done. There are efforts that make you really proud to be a venture capitalist because you're funding something that makes a difference and matters. I'm an investor because that's a means to work with companies, not because I fundamentally love investing or something like that. I like working with companies. That's my favorite part of it. Working with teams like that and companies like that is so special on these giant ambitious efforts. And I said this well before 2018 or whatever, is whether the cerebrus worked or didn't, I think it was an effort that was worth venture capital. That was the kind of thing you should do. You should try to build something that people have tried for 50 years and have been unable to, but now we think we could do. And there's a reason and application for it and everything else. I love that. I do like those kinds of things, joking aside. And we do have a robotics company. And then this CPU... project that we're talking about. I think these kinds of things are actually really fun and interesting and good use of vendor capital, but they definitely aren't easy. The productive naivete that you described, Rare, is probably a virtue that you didn't know more than you knew, otherwise you wouldn't have done it. This is certainly my experience with etch. Like in any of these fields, you ask experts, they're going to tell you, don't do it. It sucks. It's too hard. Base rate's too low. Young people can't do it for 47 reasons. All those things, yeah. Is there anywhere where that is just a bridge too far? That could be like bio or something like this, where you just are unwilling to invest if you're naive? You know what's so funny is the hard thing about investing is... Most of the time, all of these stereotypical statements are correct. They're not correct three times out of four. They're correct 19 times out of 20, maybe 99 times out of a hundred. Like they are correct. The thing that, Bruce, one of our founders, always says, what could go right? We have to ask ourselves, what could go right? And do we see that path? Yeah. Young people can't build chips, or we shouldn't do a no-semming company, or you shouldn't do this, or whatever. All of that stuff is actually totally right. Except when it isn't. Apparently, there's a saying that someone said to one of my partners, which was, if it doesn't work, it'll be for all of the reasons that your partner said. If it does work, it will be because those reasons didn't matter. It's just such a good example of this whole thing, which is, yeah, most of the time when we lay out the reasons that a company won't work, It's right. But then sometimes they just don't matter. You and I are both interested in robotics. It's not controversial that if robotics works, it might dwarf what we're currently living through. What do you think has to be true for it to work? Obviously, it's exciting. I want a robot in my house folding my laundry. It sounds great. It's kind of one of these classics. It's always 10 years away and it's been that way for a long time. What do you see happening? What has to happen for this to actually be a thing in the near to medium term? We've had classics robotics and they're all over the place and they're on assembly lines and manufacturing lines and all those things where you're doing repetitive tasks in controlled environments. Repetitive tasks in controlled environments, is like more or less solved and like that, you know, that'll continue [40:00] to happen. But having tweaked tasks in real world environments, that's where you need the AI. That's where you need the AI plus the robots. The trick with it all is you need a model that can do that. Of course, the first problem is there is no internet connection. scale data. to bootstrap the whole thing, right? LOMs were all bootstrapped on the internet, which is a ton of human knowledge. And the equivalent for that for robots doesn't exist. And people are trying different things with... videos and simulations and teleoperation. So like there's a lot of different ways. But if you kind of think about, take tele-op as an example, how much you need to tele-op robots to get to an internet scale data this first step is like getting a good set of data to bootstrap the model. One of the, I think, insights that you can have is with the Internet, there's like a lot of slop data. Even before AI generated all this stuff, there was also a bunch of junk data. And some data was more valuable than others. Like you may value, okay, certain things on Reddit more valuable than other things. You might value Wikipedia more than... Like other forums, you might value GitHub more than other things. And all of the model companies did that, right? They prioritized data that was more valuable and less valuable and ran that through the model. I think one of the most interesting things that these AI robotics companies are doing are saying, OK, well, let's just go after the high value data to start. If we go out for the high value data, then we can kind of bootstrap this model. Now, once you do that, then can you, through the pre-training process, get to a place where you can have... very small auxiliary examples of data. that you add in post-training and all of a sudden it works for that. That's the magic we have with LOMs is you have this giant pre-trained base and then you add a little magic on it in RL and post-training and you teach it a new thing that it wasn't in the pre-training set and you kind of go from there. And I think the exact same thing has happened in robotics. We're investors in Sunday Robotics, which is going out for this exact pipeline. It's a cool company. And they're doing household robots. But the key part before you get to household and all that stuff matters much less actually than can you get this training pipeline to work and how do you do that? And one of the lessons I learned and I look back, I try to learn from history because it doesn't repeat, but it rhymes. And if I look at autonomous vehicles as an example, which is their robots, basically their AI plus robots. You look at Waymo and you look at Tesla. They were both designed vertically integrated in their own way. You have a Tesla, you have a fleet of Teslas that everybody owned. It was collecting data on the Teslas and it was used to train the models to drive the Teslas. Same thing with Waymo. I think part of the lessons is they gather very, very high quality data. They did their pre-training and then they worked off of that. That was a simplification to allow you to get to a complete product or a complete solution, which of course is going to continue to improve and ultimately will generalize. I'm sure it will generalize in some way. So you'll be able to strap it on any car and everything else. So I think the same thing is happening in are using techniques where they're vertically integrating the robot and the model, the data collection around the robot. Sunday uses gloves that are designed with the robot hands. So they're perfect. So you get very good data transferability from one to another. You do this pre-training and you have a great pre-training data set. You have a good model. And then you start adding these examples and URL and post-train on top and you get cool emergent behavior. Do you have a most visceral We invested in Sunday the first time we saw them. Partner Peter arranged a demo. We went down into the basement at Stanford Lab and they had like this totally janky cardboard glove thing. You see a few evolutions of it. The last time we saw a demo, we went down in the basement. of their now office building. And there was like a dozen robots just folding arbitrary laundry. It wasn't in the demo. It was just trial and error. And then they have people like taking the clothes and then measuring them to make sure that they were folded properly and creating a rigorous baseline and evaluation criteria. And I was like, oh my God, this is happening. Do you ever worry about how to pick the right customer for these companies? Every one of these things folds laundry, which I don't think anyone likes folding laundry. So it seems like a good use case, but it feels like we don't actually understand the demand for what these things will be used to do. Do you worry about that? That we're sort of building solutions that will then be in searches of problems? I don't. And I'll explain why. And it's certainly informed by watching the LLM evolution. some of the people who were involved in the early LLM at [45:00] But I really understood that code was going to be important and obviously you know, the topic team have this perspective that you could get to RSI if you've got code generation going and automating AI research and whatnot. If you think about it, the first use cases were very much language oriented. They were very much essay writing and editing and marketing. I think the first application that really took off was Jasper, which was just writing marketing copy. I think it's going to evolve a lot, basically. I think laundry is kind of a good task because it's arbitrary, it's complex, it requires dexterous manipulation, it isn't time sensitive. If it takes three times longer, so be it, who cares, doesn't matter, just let it run all day. So I think it has some of those properties. I don't think the task is actually that important. What I think is much more important is, are you pre-training an amazing model than being able knowing then the task capability will just keep multiplying. Maybe this question will be annoying or slightly uncomfortable for you, but If you ask basically every founder and critically other investors of your type, almost everyone, if I ask like who's the best board partner, we'll say you. You come up way more often than anyone else that I've come across. And I'm curious why you think that is. What it is that you're doing that other people, the incentives are there to do a great job as a board partner. What do you think you're doing on the boards of these companies, partner with the founders, that's actually different from other really talented investors who are also nominally doing the same job but don't come up nearly as often when asked that question? One of the things that I've realized is we each are attracted to different types of entrepreneurs where we have chemistry. I'm an investor second, and I try to be a partner first. We'll see companies come in. We have one come in yesterday, and it's what I would call an investment-grade opportunity. You can invest. It probably works. You make money. It's good. And an investor would do that. A partner wouldn't because that's not sufficient for a partner. Unless you have real chemistry with that person, where you feel like you're going to be able to work together really effectively, And I'm going to learn a ton from them and they're going to learn something from me. Together, we're going to just feed each other's loops. Unless you feel that way, He can't be a partner. And so you pass on that. That's a really important like fit element to me. It kind of starts with this mutual selection. Actually, weirdly, they want to partner with us and I want to partner with them. I'm really like looking forward to working with them together. And I'll give you a really good example of where this comes into play for me. If I take Saji and Benchling, Benchling is life sciences, SaaS, companies absolutely crushed, done really, really well. And then, of course, you have this biotech crash and everything else. And the company became grindy. The company had never had any churn for the longest time, such that even on their reports for every SaaS company, you have this like, okay, gross ARR added, churn line, net ARR added, like everybody does the same thing. They never had a turn line, never reported it. For the first of like six years that I worked with the company. So then they got seven years of churn in like 12 months. Turns out life sucks when you get seven years of churn in 12 months. Through that grind and through it all and related to the whole, wait a minute, the goalpost moved. We have to do something different. Everything else. They kept thinking about how to apply AI for these biotech and pharma customers, which they're very close to. How can we make it better for them? How can we apply these models in their world in a way that they're excited about and continue to iterate? And it was grindy. And I was there for it. You're excited to work with that person Because like one, of course, you think it's a really special opportunity and there's a way out. There's a path and we can find it. But two, because of the joy of the game, the relationship and like everything else, like that's part of it. It's very different. I've seen different models and lots of different models of venture capital work. Moritz was a writer. Doerr was a sales guy. Gurley was an engineer. Peter's a career venture capitalist. They're all different. Working with them, it's like when I call these people, I learn something and they push back on me. And then I ask them questions. And what I've realized is like so much of my job is they know the answer. They know what they want to do. They know the answer and it's maybe asking questions. of them to maybe help solidify their conviction or solidify their articulation of what they want to do and why. And you just keep doing that. Through that process, hopefully we get 1% better a few times a year. We make a 1% better decision, 2% better decision a few times a year. And if you do that over a decade, That compounds to real results. One of the questions that I ask myself before making an investment is there are all these people that I care about through my life like you carry yours. [50:00] Could I... Talk one of them. into going to this company and honestly, intellectually, honestly to myself, Explain to them why this could be their life's work And if I can't do that, I should not invest. It just means that the project doesn't line up in that way. And so as long as we have one of those things, it doesn't matter that much what it is to me. It's just it's important. It could make a big dent. And if it can make a big dent and it's a special person, I'd love to work on. Are there any other questions like that that you ask yourself before investing? That's a particularly good one. The other question, if this person calls me at 9 p.m. on a Saturday night, will I pick up the phone? How's that the green button test? Yeah, yeah, yeah. It's a chemistry thing, and they have to feel the same way, obviously. One of the other ones... is if it's right, does it matter? Which is different than this first one, but there's so many things that we could be right on as a business. I looked at one last week and I told entrepreneurs, I really think that you can build an amazing company here and you just shouldn't raise venture capital. But there's like so many things that you can be right on, but they ultimately just don't matter. Like nobody cares. That's a better way to say it. If we're right, we'll let nobody care. If they won't care, then you're just not going to build enough equity value. That's another useful one. One of the coolest things that's happening right now is all of that that you just described has higher stakes and more leverage attached to it, which is manifested most simply in more dollars and higher prices. You and I have talked about this notion of what high multiple on invested capital investing is like and what it has been like and what it's moving into. You did something recently, which was you raised a growth fund for the first time in a long time. I think that is related to this concept. These companies need more capital. The prices are higher. The outcomes are bigger. Maybe we can earn the same multiple on a billion dollar entry price that we could on a 50 million dollar entry price 10 years ago or whatever. Can you talk through that evolution? Talk through the partnerships like way of thinking about it and talking about it? What you believe to be true that results in this decision to do this? Just go back to like, why did LPs starting with Swenson and everyone else, why did they start investing in venture capital? And fundamentally, it wasn't because they thought they could beat the NASDAQ or the index by like three percentage points a year, five or whatever. It's because there are situations where venture capital could drive these insane multiples on invested capital. from a financial perspective, that's what they're seeking. For the longest time for most of the industry's history, two things were synonymous. Early stage investing and high cash on cash multiples. The way to get high cash on cash multiples was to do early stage. That's it. Those two circles in the Venn diagram like almost perfectly overlap. The thing that's changed is recently, relatively recently, in the last few years, because outcomes have gotten so much bigger and these markets are bigger and everything else. The circle of high cash on cash multiple opportunities is bigger than just early stage. And it's not so, so big that there's a gazillion new companies in there that you can generate 100 Xs on. That's not true. But there are certainly... many outside of early stage where you can generate high returns. That's it. I think that's what we want to go after. You could argue we're a few years late. I think I'd take that criticism. But I think that opportunity exists on a go-forward basis. We should go do it. Everything else that we represent, which is the high conviction, high commitment, partnership, that has to still be there. As part of that discussion, what were the other sides of the debate, such as Maybe we would have said the same thing in 99 and halfway through 2020. As markets get exciting, the possibilities, we all do this extrapolation error. What were the counter arguments to like, let's despite all that still not? I think the biggest counter argument that really made this time right versus two years ago or whatever was. was you need the team that can do it. It's just a different mentality. There are differences in how you evaluate and think about things. All the other stuff or there's, you know, why not change and stay within your circle of competency and all those things are true. But that was like the biggest one. We had several examples over the last couple of years where we had, I think, the right intuition on a company or an opportunity. We didn't do it. Because it was outside the box. Because it was outside the box. That's obviously dumb. And I think it is quite different than a lot of... Than she does. We are really chasing these very rare special companies that have very high cash-on-cash opportunities. Where we think there just can be runaway successes and we can invest in them. Is there any lesson to be pulled from the many... let's call it 20 to 100 X's that you personally have observed. I mean, it's such a crazy amount of return. Obviously, it doesn't pencil in the beginning. You can't make something pencil if it was that clear. The price would be different. What have the 20 to 100 X's taught you in aggregate, if anything? Work with really special people. You want to work really hard, you want to work smart and get lucky, and you need to [55:00] at all, it really all has to come together. There's a lot of things that are timing dependent, you have no control over as a company. And you take the service example as a good one, which is, this is our second, we took it public this time in May, but we try to take it public in 2024. Yeah. And it would have been taken public at a much, much lower valuation. And it didn't work out because of CFIUS and all this stuff. So timing matters. The advancement of that 18 months made all the difference in the world. For a bunch of things that were honestly outside of our control, there were some things that were in our control. getting inference running and everything else. But there was a lot of stuff that was outside of her control. These are all the classic things. We've got to focus on what you can control. That's one of the Things that's really different than software companies. With software companies, aside from like building on AWS or whatever, you pretty much own your whole stack. And so you're really fully in control of your destiny in that way. With hardware companies, you don't. There's an entire supply chain and all this other stuff. HBM's a thing, DRAM's a thing, and TSMC's a thing, and a lot of those cross geopolitical borders. And so geopolitics gets involved, and that makes that complicated. That's a really big difference. And so you got to get lucky on the timing and macro and other stuff. But I think it really starts with working with these crazy people with unbounded opportunities. And if you work with these crazy special people on unbounded opportunities, You get lucky from time to time. You're bound to. When I was 20 years old. I was working at an investment bank. Ben Horowitz, Mark and Ben, had started LoudCloud. It was still in stealth, and Ben gave me an offer to be his assistant. I was talking to this associate who seemed like this elder at the time. He's probably 25. He said this thing to me, which stuck with me. He was like, do you golf? It's like classic banking question, do you golf? No, I don't like golf. But he's like, with golf, you keep on practicing, keep getting the ball on a three par, like close to the pen, close to the pen, close to the pen, close to the pen. You keep getting the ball close to the pen and keep practicing. That's hard work. That's like working smart. That's what you want to keep doing. Getting the hole in one, that's luck. I kind of love that framing I use with my kids, actually. What it says is like, yeah, there's luck involved and there really is luck involved, but there is actually a way to increase your luck. And the way they increase your luck is get a lot of balls close to the pen. Eventually one will drop. I'd like that. Each of us invests in one to two companies a year. I think in my 12 years, I've invested in 18 companies total. Which is a relatively small number. So it's a very high conviction and very high commitment. I have a lot of skin in the game. I believe in these companies. if you keep on working with these very special people in these opportunities, magic can happen. What have you learned about the best reasons and conditions for going public? Bedford does these Monday night dinners. And so we had a CEO last night, multi hundred billion dollar private company. And we had this whole conversation. So it's kind of fresh. I think that ultimately when you go public, You have a range of new opportunities in what you can do. Public trust, because there's some transparency that comes with being a public company. You obviously have a currency that you can do things with that ends up being there. You have an unbelievable ability to raise capital, which I think is why the labs will ultimately go. Although I think the trust thing is actually a really important element of why they should go. And it's beneficial to the world and to America. If they do go public, it's like, okay, yeah, see what's going on. Everyone can see it. I think that's like a really beneficial setup. There's another element of it, which is what does a collegiate athlete want to do? Go pro. They want to play at a higher level. Is it harder? Yeah, it's harder. Is the competition tougher? Yeah, the competition's tougher. They move faster, they're tougher, they're bigger, they're stronger, the stakes are bigger, the stage is bigger, the scrutiny is bigger. All of that's true. It's kind of the same thing with companies. There are a handful, and it really is a handful of companies. three four whatever that can get to this tremendous scale without going public because things have gone through their execution and excellence. Lots of free cash flow. They've lots of free cash flow, and they've done really well over a really long time. And I think that's fantastic. good for them, but in general for everyone else, get out there. The other thing that I would tell you is there are windows for a particular type of company. The SaaS companies that went public in 2021, A whole boat of them have struggled and it's been tough in the public markets because their stocks ripped to this multiple compression issue. They were trading at 30 times. They forexed in size, but now they're trading at six times. It turns out you're under still. That's a tough place to be. I will also tell you that there's 500 something, probably SaaS companies that are between 100 million and 500 million that are private. What happens? Those employees never got a chance to sell. Those employees don't have annual tenders. Those employees don't. have an opportunity to exit. Those investors don't have an opportunity to exit. They're stuck. I don't think they're all going away. And as I said, I don't think they're all getting vibe coded and everything else. But... ultimately The AI natives with their growth rates have sucked all [60:00] oxygen out of the room and all the interest and the window was missed. That's tough. What are the biggest debates right now inside of the partnership? I always love coming here and talking to you guys when there's something interesting going on because... You debate, it's healthy, can be really fun to watch, and I learn a lot from it. What are those debates today? There's a ton of debate around in this AI, infra, apps, apps. for foundational models, infra apps, ecosystem, where does value accrue and how does it accrue? And where are the modes? And how do we think about that? but also the business model innovation. I think one of the things people don't understand about why SaaS did so well versus traditional software was it wasn't just that it was a better delivery model and everything else. There was actual business model innovation on it. You really did have this subscription element that ended up being fantastic for both the company and the customers. It was a win-win situation. That same thing actually exists in AI and selling by outcome and that piece of it. But then wrapped up into that debate, discussion, is how much value just accrues. To the labs. Yes. How much of the value just accrues to the semis? That's a real discovery. What do you think? I am of the view that it all works. It's a very weird thing. Will the CSPs do well? Yes. Yes. Not all of them, but will some of these neoclods do well? Yes. Will the fireworks of the world do well? Yes. Will NVIDIA do well? Yes. Will these chip startups do well? Some set of them. Yes. Are we going to have edge inference on our phones? Yes. Are we going to have near edge interest on POPs? Yes. Are we going to have big models and data centers? Yes. There's so much zero-sum thinking. which is just like, okay, how do we cut up this pie? And they're going to eat this much. And like, oh, no, no, no, no, no. Anthropic or whomever is going to eat 98% of the value and they're going to do all the drug discovery. And I was like, come on. No, that's not what's going to happen. When I say like, I think everything's going to work and I listed off all these other things, it's really important to understand that doesn't mean that every company that's doing every one of those things is going to work. It actually means quite the opposite of that. Most companies in each of those areas are not going to work. And It's actually more important than ever to have real differentiation, to like really take each of these thoughts to their logical extreme and understand, wait a minute, you got to go all the way. on these things and really be differentiated on it. Is there anything that you have your eye on, whether it's in the funding market, in the technology world, anything at all that you really are watching carefully? It's actually funny to me. Some of the people are so, so smart. And yet they're in this tech world where they're like reaching these deterministic, almost conclusions of like mass unemployment and all of these different things. I'm going to give you a really concrete example, which I think is just so good. Take Jeff Hinton. in radiology. So I think it was 2016 where he was like, we should stop training radiologists. AI is going to do it all better. Jeff Hinton's three orders of magnitude is farther than I am. could not have been more wrong. But the actual thing that led him to make that statement or that conclusion was 100% correct. If you look at these radiology images, we should be able to train AI to do a better job reading these things than humans. And that's probably true. And actually, I think the studies and areas have shown that to be true. And we have an investment in a company called New Lantern, which is approaching this. But the big hurdle and the big thing that it articulated was like, wait a minute. First off, all of the aggregated training data set doesn't exist anywhere. What you see is companies going after like chest CTs or like very specific elements, but Your typical radiologist looks at a whole variety of things every single day, from x-rays to CTs to MRIs of all parts of the body and everything else. And so in AI climbing in specific areas like test CTs, is very marginally helpful. Because it's only doing that one thing, which could be one of 20 things or 40 scans that they read that day. Problem number one, you don't have the data, just like we talked about in robotics and everything else to train the IAC. Problem number two, the whole healthcare industry is oriented around reimbursing doctors for making readouts. How is that going to work? And there's liability associated with that. And there's repercussions of getting something wrong or missing something. And there's medical malpractice and everything else. How are we going to avoid that? And how are we going to get around that? Problem number two, real world stickiness. We're going to end up with this application where AI really does help radiologists. It helps radiologists get more and more higher, higher throughput. Because they I can do some parts and the radiologist does some parts and they're checking each other and everything else. And you do kind of weirdly end up in this co-pilot situation for some time. And then you're going to slowly have the AI read more and more of the scans and build up and build up and build up. But the actual duration to get from here to there is going to take a long time. In the ensuing time, we need more radiologists, not less, because, oh, by the way, everyone's getting more imaging than they used to get because the cost of images. [65:00] down in a Javon's Paradox kind of way. My point on it is you have someone very, very smart who really understands the capabilities, really understands what's happening, has the right data. But by not thinking of that data in the real world application comes to the wrong conclusion. And that's how I think of the unemployment thing. I think it's just it's almost the exact same setup. Eric, I love talking about markets and companies with you. An absolute blast. Thanks for the time. Thank you. If you enjoyed this episode, visit Colossus.com. Thank you. Outro Music. You know how small advantages compound over time? That's true in investing and just as true in how you run your company. Your spending system is your capital allocation strategy. Ramp makes it smarter by default. Better data, better decisions, better economics over time. See how at ramp.com slash invest. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com slash invest. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit RidgelineApps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your real outputs. Check them out at rogo.ai slash invest.