Maris Warns Google Cash Threatens OpenAI as Venture Returns Shrink

Maris警告：风投回报缩水之际，谷歌资金威胁OpenAI 规模庞大、达数十亿美元的超级基金人为推高了科技公司估值，并将后期下行风险转嫁给公共退休账户，而在现金回报方面，小型精品投资工具从数学逻辑上对其形成了碾压。 在本期节目中，Google Ventures创始人Bill Maris剖析了风险投资的结构性机制，从数学逻辑上论证了为何规模低于7.5亿美元的敏捷小型基金在回报上能持续超越数十亿美元的超级巨额基金。他回忆了早年在佛蒙特州艰苦创业的坚毅经历，阐述了新基金Section 32的投资理念，并就面对老牌科技巨头的掠夺性定价时未上市AI巨头所表现出的结构性脆弱发出预警。展望未来，Maris重点指出了AI基础设施和计算生物学领域的机遇，同时告诫称，后期一级市场的发展态势正在不公平地将下行风险转嫁给公众401(k)持有者。 Bill Maris携规模为1.5亿美元的精品基金Section 32重返风险投资界，该基金完全将财务回报作为衡量成功的唯一指标。 (0:00) 在人工智能的推动下，全球格局正处于一场前所未有的指数级变革的边缘，在接下来的一个世纪里，这场变革将以数量级的幅度重塑世界。 (0:00) Maris于1997年凭借通过信用卡透支购买的仅三台服务器创办了他的第一家网站托管公司，当时他在佛蒙特州忍受着严酷的室内零下严寒环境。 (0:33) 优秀的风投机构会搜寻那些在外界看来略显疯狂的创始人，因为这些人知晓一个外界尚不相信的关于未来的秘密。 (0:33) Google Ventures在2009年至2018年间通过运用数据科学和机器学习对数百万次投资组合模拟进行回测，实现了约4.1倍的整体回报。 (5:58)

Section 32 founder Bill Maris and fellow venture investors argue that outsized fund sizes and misaligned incentives distort venture capital returns, while emerging opportunities in artificial intelligence and deep tech demand infrastructure-level discipline rather than speculative capital. The Arithmetic Case for Smaller Venture Funds Maris asserts that small funds systematically outperform multi-billion-dollar mega-funds due to basic mathematical realities. Under historical baselines cited by Maris, funds under $750 million generated an average return of 4.76x and comprised 95% of top-decile performers, whereas funds exceeding $1 billion averaged 2.42x. Because roughly 75% of venture capital funds fail to return capital, Maris focuses solely on Distributed to Paid Capital (DPI). Section 32 maintains an average fund size of approximately $400 million, targeting a concentrated, artisanal investment model. To demonstrate the scale problem, Maris calculates that a $7 billion fund seeking a 3x return with 10% portfolio ownership requires $210 billion in aggregate exit value—a sum that regularly exceeds total yearly venture-backed M&A and IPO activity across the entire industry. By comparison, a $500 million fund targeting the same 3x outcome requires $15 billion in aggregate exit value. Maris rejects late-stage asset-accumulation models under Registered Investment Advisor structures as authentic venture capital, maintaining that true venture investing relies on hands-on company building. Misaligned Incentives and Public Market Exit Risks (16:47) Venture partner David Sacks observes that fund size dictates deployment strategy, with funds typically spreading capital across 20 to 25 investments. This framework leads mega-funds to artificially inflate early-stage company valuations merely to deploy large checks. Sacks points out that a $5 billion fund delivering a modest 1.01x return can market itself in the 75th percentile to raise subsequent vehicles, generating management fees that exceed the personal earnings of a disciplined manager delivering 3x returns on a $500 million vehicle. The panel warns that delaying initial public offerings exposes retail investors and retirement accounts to severe late-stage downside. By remaining private while valuations escalate, mega-cap tech startups concentrate early value creation within private circles. When these companies finally enter public portfolios and passive 401(k) index funds, retail investors face steep valuation corrections if projected future cash flows do not materialize. Maris adds that highly valued private AI firms like OpenAI and Anthropic remain vulnerable to incumbents like Google, which could pressure their unit economics by slashing API token prices. AI Infrastructure and the Gaming Evolution Analogy (19:00) Maris compares current artificial intelligence systems to the text-based adventure games of the 1980s, characterizing the sector as being at an "Atari command line stage" burdened by session resets and lack of persistent memory. He forecasts that the forty-year progression seen in video games—advancing from text interfaces to modern graphics engines—will compress into the next five years of AI development, resulting in ambient computing. Because base models face structural commoditization, Maris intentionally avoids investing directly in large foundation model providers. Instead, his investment thesis targets the foundational infrastructure required to reach a "PlayStation 10 stage," focusing capital on physics engines, specialized controllers, chips, and hardware platforms that support complex applications. Biotech Bottlenecks and U.S. Scientific Competitiveness (21:03) While computational biology offers promising discovery mechanisms, Maris stresses that identifying a compound in silico accounts for only about 5% of the total effort required to deliver an approved therapeutic. Because biology and strict Food and Drug Administration safety standards mandate human trials and dosage titration, biotech cannot scale along the frictionless exponential trajectories seen in pure software. Transformative acceleration in the field depends on developing validated, fully simulated human cells in silico. At the macro level, Maris cautions that the United States risks losing its technological lead due to funding reductions at institutions like the CDC and the NIH, an increasingly anti-science culture, and restrictive H-1B immigration rules. Meanwhile, foreign competitors like China are actively recruiting global scientific researchers and taking greater regulatory risks in clinical experimentation, raising long-term concerns over American research primacy. While AI advances and computational tools make deep tech sectors increasingly viable, venture capital faces unresolved structural friction: multi-billion-dollar fund economics continue to incentivize inflated private valuations, even as the public markets face an uncertain appetite for absorbing these late-stage assets.

前Google Ventures首席执行官兼Section 32创始人Bill Maris与对谈嘉宾指出，风险投资规模过大导致行业激励扭曲并严重压缩实际回报，而当前人工智能仍处于早期基建阶段，投资重点应转向底层工程与深科技，同时警惕高估值未上市公司向散户转嫁风险的结构性弊端。 小型基金的数学胜势与行业激励扭曲 (24:26) Maris指出，小规模基金在投资回报率上天然具备算术层面的数学优势。历史数据显示，规模低于7.5亿美元的基金平均回报达4.76倍，且占据了行业前10%顶尖表现基金的95%；相反，超过10亿美元的基金平均回报骤降至2.42倍。在10%股权假设下，一只5亿美元的基金实现3倍回报仅需被投企业创造150亿美元的总退出价值；而一只70亿美元的巨型基金实现同等回报则需要2100亿美元退出总值，这一数字往往直接超过全美风投行业全年的并购和IPO总和。Maris坚持将实际向LP返还的现金回报（DPI）视作风投成败的唯一标准。 对谈嘉宾David Sacks进一步补充，大型基金的管理费模式造成了深层机制扭曲。一只50亿美元的基金哪怕仅获得1.01倍的平庸回报，其普通合伙人（GP）依靠管理费赚取的收益也远超一只创造3倍回报的5亿美元基金。为了在有限的20至25家投资组合中强行配置庞大资本，巨型基金不惜向初创公司强加虚高估值（例如将原本合理估值1亿美元的公司人为抬高至40亿美元以注入2.5亿美元），导致早期企业资本结构失真并埋下长期隐患。 独角兽推迟上市与散户接盘风险 (16:47) 对话参与者对超级科技独角兽长期推迟上市的现象提出了严厉批评。Maris与嘉宾指出，部分创始人和顶级投资机构表面标榜“造福人类”，实质却将高增长阶段的财富增值封闭在极少数精英圈层内，直到估值见顶后才通过ETF和被动指数基金将溢价风险转嫁给普通401(k)退休金持有者。对谈方认为，如果科技公司真心希望惠及公众，理应尽早通过公开市场上市以实现财富分配的民主化。 此外，巨额未上市AI企业的估值体系显得尤为脆弱。Maris警告称，若Google等拥有雄厚现金流的在位科技巨头动用资本武器，将其Gemini API调用价格大幅下调80%，OpenAI和Anthropic等高估值私营公司的核心利润空间与商业模式将瞬间面临严峻压力。历史上约75%的风投基金处于亏损状态，面对纸面上极其庞大的账面浮盈，公开市场在解禁期满后是否愿意依据未来现金流折现买单，构成了未上市资产最终兑现的重大考验。 人工智能的雅达利阶段与底层算力基建 (19:00) 针对基础大模型之外的创新机遇，Maris将当前的人工智能生态比作20世纪80年代的文字冒险游戏（如《Zork》）和“雅达利命令行阶段”。早期文本交互反应迟缓、容错率低且缺乏情境一致性，与当前零售端AI模型频繁重置上下文、缺失长期记忆的表现如出一辙。但他预测，电子游戏行业耗费40年完成的技术跃迁，在AI领域将被压缩至未来5年内走完，最终推动技术迈入具备环境计算特征的“PlayStation 10阶段”。 基于该演进逻辑，Maris明确表示回避直接投资大语言基础模型本身。正如电子游戏的突破依靠控制器、物理引擎和GPU而非单纯改进文本剧本一样，AI革命的核心驱动力在于底层硬件平台、专用控制系统及物理模拟引擎。与此同时，AI与先进物理引擎的结合大幅压缩了深科技（Deep Tech）领域的研发周期，使其成为当下资本配置更具确定性的方向。 计算生物学的临床瓶颈与跨国科研竞争 (21:03) 在生命科学领域，曾创办Calico并投资多家长寿医疗企业的Maris指出，虽然社交网络频频宣扬肽类创新和癌症疫苗带来的“健康复兴”，但当前的医学成果本质上仍依赖十年前优秀的传统湿实验研究，而非近期的纯计算突破。Maris强调，在计算机中模拟筛选候选药物分子（in silico）仅占新药研发总工作量的5%；由于人体生物学的极端复杂性以及FDA严格的安全性与滴定监管要求，生物医药无法直接获得类似纯软件的指数级增长，只有在单细胞计算机模拟达到完全逼真后，产业瓶颈才可能真正突破。 与技术研发受阻并存的是全球科研竞争力格局的演变。Maris指出，美国监管长期将人体安全置于上市速度之上，而部分海外市场（如中国）则优先追求速度，并在克隆等前沿实验上展现出更为激进的姿态。受访者担忧，美国国内CDC与NIH等基础科研机构资金遭到削减，叠加反科学舆论环境和收紧的H-1B签证政策，正在驱使全球顶尖智力资源外流；而中国等国家正趁机通过激进的人才引进模式吸收国际脑力储备，威胁美国的基础研究优势。 本期访谈展现了风投资本在管理规模扩张下的结构异化与认知分歧。尽管AI与深科技展现出加速产业进化的潜力，但巨型基金的费用套利动机、未上市资产对公开市场的风险转嫁，以及生物医药对临床实验客观规律的依赖，都对纯粹的技术乐观主义构成了强力制约；万亿级账面繁荣能否最终转化为经得起现金流考验的实体价值，仍是市场面临的未解命题。

## Background 该内容由知名科技投资播客《All-In Podcast》于 2026 年 6 月 27 日发布在 YouTube 平台上。视频简介显示，本期节目邀请到 Google Ventures（GV）创始人兼 Section 32 掌舵人比尔·马里斯（Bill Maris）做客，深入探讨了他从科技创业到风投生涯的四堂关键课程、Google 打击生成式 AI 竞争对手的潜在策略、小型风投基金为何在数学逻辑上优于巨型基金，以及 AI 当前所处的“雅达利时代”等核心话题。 ## TL;DR 比尔·马里斯结合自身在 1997 年创立数据中心与领导 GV 的传奇履历，系统总结了敢于成为少数派、不与计算机科学对赌以及恪守专注的投资哲学。他通过详实的历史数据揭示出规模在 7.5 亿美元以下的小型基金在回报倍数与退出逻辑上完胜超大型基金的数学铁律，并抨击了当下风投机构沉迷管理费与推高估值的畸形激励机制。此外，马里斯尖锐剖析了 Google 凭借充沛现金流以低价 Token 打击 OpenAI 等竞品的杀手锏，并将当前脆弱的一问一答式大模型交互比作 AI 的“雅达利阶段”，强调未来五年的真正红利将集中在底层的环境计算、物理引擎与算力基础设施之上。 ## Atmosphere & style 现场弥漫着硅谷老兵之间既松弛风趣又充满火药味的思辨张力。马里斯的分享既有早期雷雨天在屋顶刷沥青的生动狼狈与草根叙事，又有拿数据模型解构沙丘路传统神话时的冷峻犀利。随着主持团队在后期资本逻辑与真实退出收益上展开连环追问，对话由轻松叙旧迅速转向对当今 AI 估值泡沫与管理费收割游戏的针砭时弊，整场对谈听来既具前瞻视野，又透着一股直戳行业痛点的真实与辛辣。 ## Bulletin - [10:21](https://www.youtube.com/watch?v=0umrMuUClC4&t=621s) 小规模风投基金回报率显著碾压巨型基金：在衡量实际现金分配（DPI）的前 10% 顶尖风投中，规模低于 7.5 亿美元的基金占 95%，平均回报达 4.76 倍，而 10 亿美元以上基金平均仅有 2.42 倍。 - [14:36](https://www.youtube.com/watch?v=0umrMuUClC4&t=876s) 若 Google 动用搜索业务的充沛现金流发起价格战，将 Gemini 的 Token 调用价格降价 80%，将对全行业 1 万亿美元基础设施开支承诺对应仅 600 亿美元营收的生成式 AI 独角兽构成降维打击。 - [19:09](https://www.youtube.com/watch?v=0umrMuUClC4&t=1149s) 当前生成式 AI 仍处于类似《Zork》文字交互的“雅达利阶段”，未来五年技术演进将极致压缩游戏业 40 年历程，真正的质变红利在于环境计算、物理引擎与底层硬件设施，而非一味堆砌模型参数。 - [24:26](https://www.youtube.com/watch?v=0umrMuUClC4&t=1466s) 风投行业激励机制严重异化：管理 50 亿美元巨型基金仅需 1.01 倍保本收益即可靠管理费轻松赚取巨资，迫使大基金开出数倍荒谬估值抢占额度，彻底扭曲了初创公司的健康成长空间。 ## Key takeaways - [04:38](https://www.youtube.com/watch?v=0umrMuUClC4&t=278s) 真正的创业远见在外人看来往往像“疯子”，唯有洞察到大众尚未相信的未来秘密，才能在暴风雨般的恶劣环境中生存下来并穿越周期。 - [05:58](https://www.youtube.com/watch?v=0umrMuUClC4&t=358s) 永远不要与计算机科学对赌：Google Ventures 早期依靠海量投资数据与机器学习模拟构建资产组合与规模上限，最终创造出远超行业顶级十分位（Top Decile）的投资业绩。 - [08:48](https://www.youtube.com/watch?v=0umrMuUClC4&t=528s) 风投行业唯一真实可信的硬指标只有 DPI（实际向 LP 分配的现金回报率），任何账面浮盈与未变现估值在二级市场检验前皆是脆弱的纸面数字。 - [10:21](https://www.youtube.com/watch?v=0umrMuUClC4&t=621s) 小基金在数学逻辑上天然优于大基金：5 亿美元基金实现 3 倍回报仅需旗下组合创造 150 亿美元退出市值，而 70 亿美元超级基金则需要 2100 亿美元退出体量，超出绝大多数年份全美 IPO 与并购承载力总和。 - [12:02](https://www.youtube.com/watch?v=0umrMuUClC4&t=722s) 超大规模后期投资更接近注册投资顾问（RIA）的 AUM 管理费收割游戏，缺乏严肃数据科学支撑其长期超额复利，偏离了集中时间与资本陪跑打磨业务的传统风投手艺。 - [14:36](https://www.youtube.com/watch?v=0umrMuUClC4&t=876s) 生成式 AI 面临巨大的资本畸形：全行业承担着 1 万亿美元的算力开支承诺，却仅对应 600 亿美元实际营收，一旦巨头通过 Token 价格战倾销，将对高估值独角兽造成毁灭性冲击。 - [16:20](https://www.youtube.com/watch?v=0umrMuUClC4&t=980s) 独角兽推迟 IPO 导致价值在私有化阶段被精英阶层垄断，而估值见顶后修改指数规则强推给 401(k) 养老金和散户接盘，构成了极大的二级市场接盘风险与社会不公。 - [19:53](https://www.youtube.com/watch?v=0umrMuUClC4&t=1193s) 大语言模型的单次回合制交互仅是 AI 的“雅达利时代”，未来五年环境计算将解决记忆缺失与上下文重置问题，底层计算架构与物理引擎比盲目追求大参数模型更具投资价值。 - [21:03](https://www.youtube.com/watch?v=0umrMuUClC4&t=1263s) 计算生物学无法简单复制软件业的指数级爆发：发现靶点仅占研发流程的 5%，FDA 临床毒理与人体试验不可跳过，关键转折点在于何时能实现细胞级别的高保真数字模拟。 - [22:51](https://www.youtube.com/watch?v=0umrMuUClC4&t=1371s) 美国基础科研预算干涸、反科学情绪蔓延及收紧 H-1B 签证政策，正在削弱其核心脑力资产储备，让全球科研竞争力面临外流危机。 - [24:26](https://www.youtube.com/watch?v=0umrMuUClC4&t=1466s) AI 与物理引擎的成熟大幅降低了深科技的试错门槛，使得人体健康与底层工程硬件等周期长、重资产的深科技蓝海迎来了爆发契机。 - [26:23](https://www.youtube.com/watch?v=0umrMuUClC4&t=1583s) “基金规模决定基金策略”的异化导致市场估值虚高，巨型基金为部署资本开出不切实际的高估值，不仅伤害了早期创业者，其后期狙击策略也已被历史数据证伪。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=0umrMuUClC4&t=0s) 从谷歌传奇到 Section 32：比尔·马里斯的人生四堂课 新闻短片回顾了比尔·马里斯（Bill Maris）在宣布退隐后重返投资界的消息。作为 Google Ventures（GV）的创立者兼首任 CEO，他为其新成立的风投基金 Section 32 成功募集了 1.5 亿美元。马里斯在受访时强调，较小的资金规模赋予了他极大的灵活性与筛选优势，无论是在所投公司还是团队招募上都能做到极致审慎；风投的唯一考核标准就是纯粹的财务回报，任何其他虚浮的指标都因无法精准衡量而注定失败。紧接着，马里斯登台感谢听众，并简要介绍了自己的履历：在创立 Section 32 之前，他不仅打造了 Google Ventures，还担任过 Google 特别项目副总裁，参与孵化了自动驾驶公司 Waymo、Google X 实验室以及长寿生命科学公司 Calico 等重量级项目。在更早之前，他则是一家网络托管与数据中心公司的创始人。 马里斯提到，回顾过去一百年科技带来的剧变，再对比 AI 时代未来百年即将迎来的指数级爆发，整个世界将发生数十个数量级的巨变。立足于这些非凡经历，他总结出贯穿自己职业生涯的四堂核心课程，并将时光拉回到 1997 年——彼时他刚刚大学毕业，手里攥着一个神经科学学位。 ### [01:25](https://www.youtube.com/watch?v=0umrMuUClC4&t=85s) 课一：在暴风雨中修屋顶——创业者的远见与疯狂 刚走出校园的马里斯机缘巧合进入华尔街工作，然而穿着西装、在酷暑中挤车通勤的生活让他感到无比痛苦与压抑。转变发生在一个平凡的工作日：马里斯偶然扫视办公室衣帽间的衣架下方，发现一件件挂着的西装外套底下藏着一台嗡嗡作响的服务器。当他好奇询问这是什么设备时，同事答道：“那是存放我们所有电子邮件和网站的地方。”那一刻，他宛如被顿悟的灵光击中：“既然你们能在衣柜里放自己的网站和邮箱，那如果我在我的衣柜里放，能放得下多少家公司的网站和邮箱？”他坚信自己透过未来的钥匙孔瞥见了互联网与数据中心的雏形，于是毫不犹豫地辞掉了华尔街的铁饭碗。 然而现实与理想有着巨大落差。1997 年所谓“最先进的数据中心”，其实就是马里斯位于佛蒙特州的出租屋公寓，创业资本全靠刷信用卡。他在一个房间里放了三台服务器（小号、中号、大号各一台，后来业务发展到五台），自己在另一个房间工作。由于机器散热惊人，房间酷热难耐，但他只要一把窗户打开，佛蒙特州的刺骨寒风便倒灌进来，室内温度骤降到正午放在桌上的水杯都能结上一层厚冰；到了晚上，他甚至只能用从家得宝（Home Depot）买来的粗糙地毯当被子御寒。最惊心动魄的一次发生在雷雨交加的深夜：老旧公寓屋顶突然漏水，眼看雨水就要倾泻到服务器上引发短路。马里斯打电话向房东求援，房东却冷漠地回了一句“漏水很正常”。走投无路之下，马里斯连夜赶到家得宝买了一桶沥青和一把拖把，在电闪雷鸣、暴风骤雨中爬上湿滑的屋顶补漏。由于不懂施工技巧，他没能遵循从远端往出口退着涂抹的常识，反而一路把自己涂抹困在了死角里。面临要么机器被淹短路、要么自己被雷劈死的极端处境，他硬是咬牙扛了过去，尽管他最后踩着沥青逃脱时，鞋子被永远粘在了那个佛蒙特州的屋顶上。 - **1997 年初创数据中心配置**：位于佛蒙特州出租公寓内，启动资金全部来自个人信用卡，由 3 台服务器（小、中、大号）起步，后扩充至 5 台。 - **极端创业环境**：夏季室内酷热，冬季开窗降温导致室内水杯结冰，夜间靠家得宝购买的粗地毯御寒。 ### [04:38](https://www.youtube.com/watch?v=0umrMuUClC4&t=278s) 课二：洞悉未来的人，在旁人眼里总是“疯子” 屋顶涂沥青的狼狈经历引出了第二堂课：“要看清未来，有时你必须显得有点疯狂（To see the future, sometimes you need to be a little bit insane）。”在外人看来，雷雨天爬上屋顶刷沥青的人与疯子无异，但正是这种偏执造就了穿越周期的洞察。为了具象化这种认知时差，马里斯展示了好友斯图尔特·巴特菲尔德（Stewart Butterfield，Slack 创始人）分享的三张美国总统就职典礼历史照片： 1. **1989 年老布什就职典礼**：茫茫人群中，只有零星一两个人举着传统胶卷相机拍照，其余人都在单纯用眼睛见证。 2. **2005 年小布什就职典礼**：时隔 16 年，现场几乎没有本质变化，依旧是稀疏的举相机动作。 3. **2009 年奥巴马就职典礼**：短短 4 年之后，人海演变成了数码屏幕的海洋，几乎人人手里都高举着相机或早期照相手机。而在密集的人群中，赫然出现了一位高高举起整台笔记本电脑的人——他正用笔记本摄像头对着讲台进行现场录屏甚至流媒体直播。 在当时周遭的人看来，举着笨重笔记本拍就职典礼的人同样像个疯子。然而马里斯指出，这位陌生人其实洞悉了某种其他人尚未相信的未来趋势（移动直播与即时多媒体记录）。这种“掌握着大众尚不相信的未来秘密”，正是他评判一名优秀创业者最核心的特质。 ### [05:58](https://www.youtube.com/watch?v=0umrMuUClC4&t=358s) 课三：不要与计算机科学对赌——用数据驱动重塑 Google Ventures 时光转至 2007 年，马里斯受邀加入 Google 并面临一项全新挑战：为 Google 建立一个正式的风险投资基金。当时的 Google 零星做过一些投资，但没有连贯的战略，也没有明确的预算规划。马里斯找来 Android 联合创始人里奇·米纳（Rich Miner）作为搭档，两人密集走访了沙丘路（Sand Hill Road）上的所有老牌风投机构，向一切愿意交流的人取经。吸取经验后，他们决定反其道而行之，拿出一套前所未有的方案： ```mermaid flowchart TD accTitle: Google Ventures 投资决策机制与运作模式 accDescr: 展示 Google Ventures 结合海量历史数据与机器学习进行基金构建和投资决策的流程 subgraph DataEngine ["数据驱动核心"] A["收集全行业海量历史风投数据"] --> B["机器学习建模与模拟仿真"] B --> C["测算最优资产组合配置"] B --> D["反推最优基金规模上限"] end subgraph Execution ["实战执行与业绩分化"] C & D --> E["GV 团队整体运作<br>(约 4.1x 回报率)"] C & D --> F["马里斯主导投资项目<br>(严格执行量化模型，回报更高)"] end E --> G["行业顶尖四分位 (Top Quartile) 水平"] F --> H["大幅超越行业第十百分位 (Top Decile) 水平"] ``` 整个方案的核心分为两步：第一，依托 Google 强大的技术基底，收集尽可能全面的全行业历史风险投资数据；第二，利用数据科学模型进行数百万次模拟和回测，以此来推演最理想的资产组合架构（Portfolio Construction）以及最佳的基金资产管理规模。马里斯回忆道，当时由于“AI”这个词在外界听来宛如科幻小说甚至会引发恐慌，Google 高管层严令禁止团队使用 AI 这个字眼，长达数年里团队只能将其称为“机器学习（Machine Learning）”。 这种用量化工具改造风投“艺术”的做法引发了内部高管的强烈质疑与外界嘲讽，在当时显得荒唐离奇。然而从 2009 年到 2018 年的历史业绩彻底打破了怀疑：在同周期全行业风投基金中，GV 的整体回报率（基于公开可查数据估算）达到了约 4.1 倍，稳居顶级四分位（Top Quartile）；而由马里斯亲自操刀、最严格遵循该数据科学模型的投资组合，业绩更是远远甩开行业排名前 10% 的顶级十分位（Top Decile）。这印证了他的第三堂课：“千万不要与计算机科学对赌（Don't bet against computer science）。”只要在恰当的时机将合适的技术应用到正确的痛点上，数据与算法必将给出最优解。 ### [08:48](https://www.youtube.com/watch?v=0umrMuUClC4&t=528s) 课四：风投的数学铁律——为什么小规模基金表现完胜巨型基金 2017 年，马里斯选择走出舒适区创办自己的独立基金 Section 32。身边的朋友再次认为他疯了，质疑他为何要离开 Google 那种“午餐免费、按摩充足的温暖子宫”。而当同行接受了他要自立门户的事实后，给出的建议几乎千篇一律：“尽可能募集更多资金，管理规模越大，管理费就越高，你就能活得更滋润。”马里斯断然拒绝了这种追求资产管理规模（AUM）最大化的建议。 截至当时，Section 32 已经先后募集并运营了 6 期基金，平均每期规模牢牢克制在约 4 亿美元左右，旗下捕获了包括网络安全巨头 CrowdStrike、前沿大模型独角兽 Cohere 以及加密平台 Coinbase 在内的标杆企业。这 6 期基金全部跻身行业排名前 10% 的 Top Decile。马里斯格外强调，风投界唯一真实可信的硬指标只有 DPI（即实际向有限合伙人分配的现金回报率），一切纸面估值皆是泡影。 这直接推导出了他的第四堂课，也是颠覆传统管理费导向的异端言论：“在风险投资中，小型基金的回报表现必然击败巨型基金。”马里斯强调这绝非个人主观偏见，而是由冷酷的数学逻辑和管理精力决定的规律： - **组织损耗与精力分散**：他曾亲自执掌百人规模、资金数十亿美元的庞大机构，深知庞大管理体系会带来极度严重的内部损耗与分心，投资人根本无法向创业者投入真正高密度的精力支持。 - **专注力与严苛选择标准**：规模适度的基金逼迫投资合伙人极其苛刻地挑选项目与人才，唯有保持精悍与专注，才能在深耕细分领域的同时，实现几何倍数的高额资本回报率。 ### [10:21](https://www.youtube.com/watch?v=0umrMuUClC4&t=621s) 基金规模的残酷算术：为什么小体量 VC 能碾压大型基金 在衡量真正资金分配回报率（DPI）的前 10% 顶尖风投基金中，规模小于 7.5 亿美元的基金平均回报倍数达到 4.76x，而规模超过 10 亿美元的大型基金平均回报仅有 2.42x。统计数据表明，在考察期内，规模低于 7.5 亿美元的基金占到了前 10% 顶尖梯队的 95%；一旦基金规模突破 7.5 亿美元门槛，回报率便会出现断崖式的“不连续压缩”。针对这一悬殊差异，Bill Maris 摆出了一个清晰的思维实验：风投行业排在 75% 分位点的基金基本都在亏钱，只有前四分之一的头部机构具备持续盈利能力。 假设一家机构管理着一只 5 亿美元的基金，按当下的平均持股比例 10% 计算，若想仅仅保本，被投企业就需要实现 50 亿美元的总退出估值；如果想在风投行业长期立足并将目标定为 3x 回报，旗下企业就必须创造 150 亿美元的退出市值。反观那些规模高达 70 亿美元的超级巨型基金，同样按 10% 持股比例推算，要实现 3x 回报，其组合公司必须完成惊人的 2100 亿美元退出总估值——这一数字已经超过了多数年份全美所有风投支持的并购（M&A）与 IPO 退出总和。 - 规模小于 7.5 亿美元的基金，在 DPI 顶级梯队中平均实现 4.76x 回报；规模大于 10 亿美元的基金平均仅为 2.42x。 - 低于 7.5 亿美元的基金占全行业前 10% 顶尖表现者的 95%。 - 5 亿美元基金（以 10% 股权计）保本需 50 亿美元退出，3x 回报需 150 亿美元退出。 - 70 亿美元巨型基金要拿到 3x 回报，需要高达 2100 亿美元的退出体量，严重超出大部分年份公开市场与并购的整体承载力。 ```mermaid flowchart TD accTitle: 基金规模与退出市值需求对比 accDescr: 展示5亿美元与70亿美元基金在3倍回报目标下所要求的企业退出市值差异 subgraph SmallFund ["5 亿美元规模基金"] A1["基金本金: $500M"] --> B1["保本退出市值 (10% 持股): $5B"] B1 --> C1["3x 回报目标退出市值: $15B"] C1 --> D1["属于历史常态市场可消纳空间"] end subgraph MegaFund ["70 亿美元超级基金"] A2["基金本金: $7B"] --> B2["保本退出市值 (10% 持股): $70B"] B2 --> C2["3x 回报目标退出市值: $210B"] C2 --> D2["超过多数年份全美全行业退出总和"] end ``` --- ### [12:02](https://www.youtube.com/watch?v=0umrMuUClC4&t=722s) 老友叙旧：从 Climate Corp 到 A-Halo 的投资哲思 演讲结束后，Bill Maris 与台上的老友们坐到了一起。David Friedberg 笑着揭开两人的渊源：当年 Maris 创办 Google Ventures（GV）时，Friedberg 创办的 Climate Corp 正是 GV 投资的第一家由前谷歌员工创立的公司。那笔交易最终以孟山都（Monsanto）斥资 10 亿美元收购圆满收官。在当时，10 亿美元还是一笔极其耀眼的巨资，而如今在疯狂的估值膨胀下，这个数字甚至被调侃成“仅仅相当于一轮种子轮或 A 轮融资”。两位老友眼下正准备在名为 A-Halo 的新项目上再次联手重演当年的成功。 随后，对话切入到更具交锋性的投资战略分歧。Friedberg 抛出了前一位讲者 Thomas Laffont 提出的后期复利逻辑：超大体量资本专注于下注已验证业务模式的后期巨头，数据表明从 10 亿涨到 100 亿、从 100 亿迈向万亿市值的确定性更高。Friedberg 由此发问：难道不能采用“杠铃策略”——既设立聚焦早期纯风投的小型工具，又配置押注确定性赢家的大型载体来双管齐下实现 3x 吗？Maris 直言，除了当下因数家数万亿美元退出预期交织而形成的这一“诡异历史瞬间”外，他尚未看到任何严肃的数据科学能支撑后期投资具有长期持续超额复利的结论。在他看来，很多机构本质上是在做注册投资顾问（RIA）的资产管理规模（AUM）收割游戏，这根本算不上纯粹的风险投资，真正的风投是一门手艺，需要投资者将极其集中的时间和资本倾注在创业者身上，陪跑帮助他们从零打磨业务。 --- ### [14:36](https://www.youtube.com/watch?v=0umrMuUClC4&t=876s) 巨头价格战与道德悖论：Google 如何降维打击 OpenAI 与 Anthropic 谈及当下生成式 AI 独角兽的生存困境，Maris 毫不留情地戳破了行业的估值泡沫。他特别指出，有些公司通篇充斥着“造福人类”（public benefit）的高尚言论，却在估值曲线上涨最诱人的阶段将巨大价值紧紧锁死在自身与少数精英投资人的小圈子里，等到泡沫见顶才说“我们要去造福全人类”。Maris 一针见血地讽刺道：“人类真正需要的是真金白银。” 紧接着，Maris 描绘了一个足以彻底摧毁 OpenAI 和 Anthropic 商业模式的推演场景：如果拥有源源不断造血现金牛的 Google 决定亮出獠牙，直接将 Gemini 的 Token 调用价格腰斩甚至砍掉 80%，市场会发生什么？在模型能力基本雷同的情况下，任何理性企业客户都没有理由去支付五倍溢价。“他们的利润率就是我的机会。”台上主持人补充道，Google 完全可以将资本与 Token 当作无情武器，效仿 Uber 当年烧钱砸市场的逻辑，将 Token 降到两折倾销以抢占消费者与企业端生态。当前全行业背负着 1 万亿美元的算力支出承诺，对应的却仅有可怜的 600 亿美元营收，如此巨额的亏空最后却试图推向公开市场，指望散户与大众来当接盘侠。 - 行业畸形现状：全行业 AI 基础设施开支承诺高达 1 万亿美元，而目前对应的实际业务年收入仅约 600 亿美元。 - Google 降维打击逻辑：凭借搜索业务的充沛现金流，通过官方降价 80% 将 Token 武器化，从而在短时间内将竞争对手推入现金流严重枯竭的超临界危机。 - 伪善指责：头部 AI 实验室一方面以非营利或公共福祉外衣自居，另一方面在最肥厚的价值创造期排斥大众参与，待到估值透支后再推向公开市场。 --- ### [16:20](https://www.youtube.com/watch?v=0umrMuUClC4&t=980s) 401(k) 接盘陷阱与被扭曲的二级市场逻辑 Maris 与主持人深入剖析了科技公司推迟 IPO、长期保持私有化背后的社会不公。99% 的普通公众无法在早期阶段参与财富创造，但随着这些估值数千亿甚至数万亿美元的超级独角兽修改甚至打破标普 500 指数的传统准入规则，美国的 401(k) 养老金账户、被动指数基金和 ETF 最终都不得不被动建仓买入。Maris 严厉警告，普通工薪阶层的养老金资产正在被推上风险极高的断崖，被迫接盘已经严重透支成长空间的昂贵资产，让财富进一步向本就不缺钱的精英阶层集中。 当被问及这轮洗牌对整个风投生态意味着什么时，主持人提到像 Founders Fund 的二期基金（Fund II）这类极少数神级基金，可能仅凭 2 亿美元的初始投资就能在 paper 上印出 1000 亿美元的逆天回报。Maris 指出，除去这极少数异类，风投的本质就是极端的双峰分布（75% 的基金都在亏损）。更严峻的现实在于，只要那 1000 亿美元的股票没有在公开市场上卖给真实的买家，它就只是一串停留在纸面上的浮盈。在未来现金流折现（DCF）的严苛审视下，无论是 SpaceX 还是 Anthropic，二级市场散户和机构在六个月禁售期解禁后是否愿意持续买单，仍是一个巨大的问号。 --- ### [19:09](https://www.youtube.com/watch?v=0umrMuUClC4&t=1149s) 人工智能的“雅达利时代”：从《Zork》文本交互看大模型雏形 在剥离了那些数万亿美元的巨型泡沫之后，Maris 转向了他真正看好的风投领域，并提出了著名的“游戏行业演进”类比。面对外界关于“AI 无处不在的未来究竟是末日毁灭还是乌托邦天堂”的极端争论，Maris 认为未来的真实走向既没有悲观者想象的那么恐怖，也不会像空想家描绘的那么完美。 为了看清 AI 当前所处的真实阶段，Maris 带领大家回顾了 20 世纪 80 年代早期的经典文字冒险游戏《Zork》以及《Planetfall》。那时的游戏系统极其脆弱稚嫩，完全依赖于僵硬死板的“一问一答”回合制交互：玩家输入“拿起灯”，如果系统判定词库里只有“灯笼”，游戏就会报错，玩家必须精准改成“拿起灯笼”，再输入“向北走”，然后静待计算机给出下一行响应。Maris 随后切换到幻灯片，向全场展示了目前市面上最尖端、最前沿的商业零售 AI 系统界面——它的交互本质依然是一个输入框、一段 prompt、等待计算机生成一段回复。这种极其脆弱、回合制的人机交互，表明现阶段的大语言模型不过是刚刚走进了人工智能的“雅达利（Atari）时代”或文字冒险游戏时代，距离真正成熟的交互革命才刚刚拉开序幕。 ### [19:53](https://www.youtube.com/watch?v=0umrMuUClC4&t=1193s) AI的“雅达利阶段”与底层基础设施投资机会 Bill Maris 将 AI 的发展路径与电子游戏工业的演进进行了深度类比。他指出，游戏产业从 1980 年代至今走过的数十年历程，将在未来大约 5 年内被极致压缩并在 AI 领域完整上演。今天的电子游戏早已告别过去的简陋模样，变得高度逼真（photo-realistic）、沉浸感极强；而在 AI 领域，伴随着环境计算（ambient computing）的普及，诸如当年经典文字冒险游戏《Zork》所面临的记忆缺失、上下文不一致以及会话重置等核心缺陷，都将被彻底攻克。 在回答关于未来投资方向的提问时，Bill Maris 明确表示自己并不打算投资那些盲目追求更大参数规模的基础模型。这正如游戏产业的质变并非单纯依靠“编造更好的剧本故事”，而是依赖于手柄控制器、物理引擎、GPU 等底层机械设备与技术栈的升级。他认为，当前我们仍处于 AI 的“雅达利与命令行阶段（Atari command line stage）”，但五年内整个行业将跃升至“PlayStation 10”的水准。因此，他押注的焦点在于支撑这场革命发生的所有基础平台与工程硬件。 ```mermaid flowchart LR accTitle: 游戏工业与AI技术演进类比 accDescr: 对比早期文本或命令行阶段到高度拟真沉浸阶段的技术驱动路径 subgraph Gaming["游戏产业演进 (40年历程)"] G1["早期文字游戏 (如 Zork)"] --> G2["控制器 / 物理引擎 / GPU"] --> G3["沉浸式拟真游戏"] end subgraph AI["AI 发展浪潮 (压缩至未来 5 年)"] A1["当前: 雅达利与命令行阶段<br/>(会话重置 / 缺少记忆 / 缺乏一致性)"] A2["底层机械设施<br/>(环境计算 / 控制器 / GPU / 物理引擎)"] A3["未来: PlayStation 10 水准<br/>(高度自洽的沉浸式智能体验)"] A1 --> A2 --> A3 end ``` - **核心观点**：不要将筹码全部押注在“越来越大的模型”上，推动质变的关键在于赋能模型发挥作用的底层平台与计算架构。 - **时间压缩**：游戏产业耗费约 40 年的技术蜕变，将在 AI 领域于未来 5 年内爆发式完成。 --- ### [21:03](https://www.youtube.com/watch?v=0umrMuUClC4&t=1263s) 生命科学的现实与计算生物学的瓶颈 谈及生命科学领域，Bill Maris 提到自己曾主导创立了抗衰老与长寿研究公司 Calico，并投资了 Flatiron Health 和 Vir Biotechnology 等知名企业。他依然坚信生命科学具有“造福人类与创造商业价值并存”的双重魅力（doing good and doing well）。然而，他也坦承目前团队在需要进行人体临床试验的传统药物研发领域投入极少，因为那是一个需要极其专门化技能的极窄投资赛道，他当前的兴趣主要聚集在计算生物学（computational biology）。 针对社交媒体上关于癌症疫苗、肽类化合物等“人类健康大爆发”的热议，讨论指出目前涌现的大多数突破性赢家并非由计算机算力驱动，而是基于十年前扎实的传统生物科学积累。Bill Maris 解释道，即便现在他投资了由 Blake Byers 和 Coinbase 创始人 Brian Armstrong 创办的长寿抗衰公司 NewLimit，发现一个有效化合物其实只占整个研发进程的 5%。“人体生物学机制与 FDA 监管极其复杂”，后续漫长的滴定、毒理与人体安全性测试无法简单跳过，因此生命科学难以像软件那样实现纯粹的指数级增长。真正的转折点，将取决于人类何时能够在计算机中实现对人体细胞的高拟真数字模拟（in silico simulation）。 - **投资案例**：Calico、Flatiron Health、Vir Biotechnology，以及当前的细胞重编程抗衰公司 NewLimit。 - **研发瓶颈**：药物发现仅占整体工作量的 5%，FDA 临床试验的安全性验证与人体复杂性构成了硬性时间障碍。 - **突破前提**：在计算机中构建完整且逼真的人体细胞模拟体系（realistic in silico simulation）。 --- ### [22:51](https://www.youtube.com/watch?v=0umrMuUClC4&t=1371s) 全球科研竞争、基础科研拨款与人才外流危机 针对资本和生物科技项目流向印度与中国的现象，Bill Maris 分析了中美在生物医药监管逻辑上的深层差异。美国长期以来将“人体安全性”置于“上市速度”之上，这种审慎虽带来制度成本，但也保护了生命安全；而部分国家采取了相反的权衡策略，尽管催生了克隆等激进前沿实验，但也伴随着潜在的人类生命代价。 更令 Bill Maris 担忧的是美国内部的科研环境恶化。美国疾病控制与预防中心（CDC）和美国国立卫生研究院（NIH）等关键机构遭遇削弱，叠加社会蔓延的“反科学情绪（anti-science vibe）”，导致基础科研资金大幅缩水。与此同时，收紧 H-1B 签证政策正在将全球顶尖的“脑力资产（brain trust / neurological reserves）”推向境外。中国推行类似二战后美国吸引科学家的“回形针行动（Operation Paperclip）”，正在积极吸收欧洲与印度的核心科研人才，美国正在丧失曾经引以为傲的人才蓄水池。 - **监管天平**：美国倾向于绝对的人体安全性（human safety），海外部分体系则偏向极快的市场落地速度（speed to market）。 - **制度隐患**：基础科研预算干涸、国家机构削弱及移民政策对 H-1B 的限制，正在削弱美国的长期科研优势。 --- ### [24:26](https://www.youtube.com/watch?v=0umrMuUClC4&t=1466s) 算力赋能深科技（Deep Tech）与大模型溢价泡沫 在讨论诸如 SpaceX 和特斯拉这类传统意义上周期长、重资产、高风险的“深科技（deep tech）”时，Bill Maris 强调，由于 AI 的加持与物理引擎的成熟，深科技创业正变得前所未有地可行且高效。人体生物学与大健康是全球最大的潜在市场（TAM），而连接这些应用的物理引擎、控制系统以及 GPU 等基础设施同样构成了庞大的深科技蓝海。 随后，对话切入风投行业的商业模式本质。针对 David Sacks 提出的“是坚持 5 亿美元的最佳基金规模，还是扩张为 a16z 那样的巨型平台”这一问题，Sacks 提出基金规模决定了基金策略（fund size determines fund strategy）——通常将总额分配给 20 到 25 家公司来构建分散投资组合。Sacks 思考当前是否应该避开嘈杂混乱的超早期轮次，直接以 5000 万美元单笔支票押注后期的成熟突破项目。 ```mermaid flowchart TD accTitle: 巨额基金与早期创业者的激励错位模型 accDescr: 展示巨型基金如何凭借管理费驱动抬高创业公司估值并扭曲市场机制 subgraph LP_GP["LP 与 GP 激励机制"] LP["大学捐赠基金 (如 Stanford)"] -->|"必须部署 2~5 亿美元"| BigFund["50 亿美元巨型基金"] BigFund -->|"管理费驱动 (只要 1.01x 回报)"| GPProfit["GP 净收益超越高倍数小基金"] end subgraph Startup_Valuation["创业者面临的估值异化"] Founder["OpenAI 离职研究员创业"] SmallVC["精品基金: $20M 投资额 / $100M 估值<br/>(稀释 20% 股权)"] BigVC["巨型基金: $250M 投资额 / $4B 估值<br/>(极其微小的股权稀释)"] Founder -->|"更理性但稀释多"| SmallVC Founder -->|"估值暴涨，难以拒绝"| BigVC end ``` Bill Maris 随即深刻剖析了风投行业当前畸形的激励机制（incentive structure）： 1. **LP 与巨型基金**：一只 50 亿美元的超大基金只要取得 1.01 倍的保本收益，账面数据就足以宣称自己排进行业前 75% 的分位线并轻松募集下一期，斯坦福大学等大型捐赠基金也不会为此问责投资经理，因为他们有着庞大的资本部署压力； 2. **GP 的赚钱逻辑**：管理 50 亿美元基金获得 1.01 倍收益的 GP，通过管理费赚取的真金白银远多于管理 5 亿美元基金打出 3 倍回报的 GP； 3. **创业者面对的资本诱惑**：一个 OpenAI 的研究员出来创业，精品基金可能提出以 1 亿美元估值投资 2000 万美元（稀释 20% 股权）；但巨额基金为了把资本花出去，会直接开出“给你 2.5 亿美元、按 40 亿美元估值”的荒谬条件。缺乏经验的年轻创业者几乎毫无抵抗力。 Bill Maris 总结道，这种被彻底扭曲的激励体系必然会迎来均值回归（pendulum will swing back）。历史数据明确表明，单纯指望在后期轮次以高昂估值“狙击”成熟企业的策略，从长期来看根本无法持续。在热烈的掌声中，主持人向 Bill Maris 的精彩分享致谢并结束了本次对话。 - **核心论断**：“基金规模决定基金策略”，规模过度膨胀必然导致资本为了规模化部署而主动推高项目估值。 - **机制扭曲**：巨型基金的商业模式本质上是管理费驱动，使得管理 50 亿美元拿 1.01 倍回报的利润优于 5 亿美元做 3 倍回报。 - **市场警示**：高估值直接扼杀了早期创业公司的健康成长空间，后期盲目狙击大项目的策略已被数据证伪。 ## Notable quotes > [01:25](https://www.youtube.com/watch?v=0umrMuUClC4&t=85s) 那是存放我们所有电子邮件和网站的地方。 — 比尔·马里斯同事 > [01:25](https://www.youtube.com/watch?v=0umrMuUClC4&t=85s) 既然你们能在衣柜里放自己的网站和邮箱，那如果我在我的衣柜里放，能放得下多少家公司的网站和邮箱？ — 比尔·马里斯 > [04:38](https://www.youtube.com/watch?v=0umrMuUClC4&t=278s) 要看清未来，有时你必须显得有点疯狂（To see the future, sometimes you need to be a little bit insane）。 — 比尔·马里斯 > [05:58](https://www.youtube.com/watch?v=0umrMuUClC4&t=358s) 千万不要与计算机科学对赌（Don't bet against computer science）。 — 比尔·马里斯 > [08:48](https://www.youtube.com/watch?v=0umrMuUClC4&t=528s) 在风险投资中，小型基金的回报表现必然击败巨型基金。 — 比尔·马里斯 > [14:36](https://www.youtube.com/watch?v=0umrMuUClC4&t=876s) 人类真正需要的是真金白银。 — 比尔·马里斯 > [14:36](https://www.youtube.com/watch?v=0umrMuUClC4&t=876s) 他们的利润率就是我的机会。 — 主持人 > [21:03](https://www.youtube.com/watch?v=0umrMuUClC4&t=1263s) 人体生物学机制与 FDA 监管极其复杂。 — 比尔·马里斯 > [24:26](https://www.youtube.com/watch?v=0umrMuUClC4&t=1466s) 基金规模决定基金策略（fund size determines fund strategy）。 — 大卫·萨克斯 ## Mentioned 比尔·马里斯 (Bill Maris), Section 32, Google Ventures (GV), Waymo, Google X, Calico, 斯图尔特·巴特菲尔德 (Stewart Butterfield), Slack, 里奇·米纳 (Rich Miner), CrowdStrike, Cohere, Coinbase, 大卫·弗里德伯格 (David Friedberg), Climate Corp, 孟山都 (Monsanto), A-Halo, 托马斯·拉丰 (Thomas Laffont), OpenAI, Anthropic, Google, Gemini, Uber, Founders Fund, SpaceX, Zork, Planetfall, Atari, Flatiron Health, Vir Biotechnology, 布莱克·拜尔斯 (Blake Byers), 布莱恩·阿姆斯特朗 (Brian Armstrong), NewLimit, FDA, CDC, NIH, 特斯拉 (Tesla), 大卫·萨克斯 (David Sacks), Andreessen Horowitz (a16z), 斯坦福大学 (Stanford University) ## Video description (0:00) Bill Maris joins the Besties! (0:33) Four critical lessons from a career in technology (5:58) Building Google Ventures with data and machine learning (9:51) Why small VC funds beat big ones on average (14:36) OpenAI's valuation problem and the AI price war (19:09) AI's "Atari Stage": what comes next? (25:23) VC's broken incentives and the future of deep tech Thanks to our partners for making this possible! EY - Agentic AI is introducing a new investment discipline. As AI shifts to consumption-based models, EY connects spend to enterprise value. https://www.ey.com/en_us/insights/ai/agentic-ai-token-costs?WT.mc_id=3501318&AA.tsrc=sponsorship NYSE - Thank you to our partner, the New York Stock Exchange - a modern marketplace and exchange for building the future. It all happens at the NYSE. https://www.nyse.com Plaud - Never miss a moment. Plaud, our official wearable AI note-taking partner at All-In Liquidity Summit, captured every insight. https://www.plaud.ai Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=0umrMuUClC4)

[1:25] 比尔·马里斯同事: 那是存放我们所有电子邮件和网站的地方。 [1:25] 比尔·马里斯: 既然你们能在衣柜里放自己的网站和邮箱，那如果我在我的衣柜里放，能放得下多少家公司的网站和邮箱？ [4:38] 比尔·马里斯: 要看清未来，有时你必须显得有点疯狂（To see the future, sometimes you need to be a little bit insane）。 [5:58] 比尔·马里斯: 千万不要与计算机科学对赌（Don't bet against computer science）。 [8:48] 比尔·马里斯: 在风险投资中，小型基金的回报表现必然击败巨型基金。 [14:36] 比尔·马里斯: 人类真正需要的是真金白银。 [14:36] 主持人: 他们的利润率就是我的机会。 [21:03] 比尔·马里斯: 人体生物学机制与 FDA 监管极其复杂。 [24:26] 大卫·萨克斯: 基金规模决定基金策略（fund size determines fund strategy）。

Bill Maris: water and computers and servers don't mix well Landlord: well, that happens sometimes Bill Maris: to see the future, sometimes you need to be a little bit insane Bill Maris: they know a secret about the future that most of us don't believe Google Senior Leadership: AI is science fiction Google Senior Leadership: a hundred years away Google Senior Leadership: freak people out Peers: warm womb of Google Bill Maris: that is not venture Founders and elite investors: benefit humanity [16:47] Venture funds: for the benefit of humanity [19:00] Bill Maris: Grab the lamp. Oh... I should have said lantern. Go north. [21:03] Bill Maris: do good, do well [21:03] Social media platforms: renaissance in human health [22:51] Bill Maris: gutting of the CDC and the NIH [22:51] Bill Maris: anti-science vibe [24:26] David Sacks: fund size determines fund strategy [24:26] David Sacks: A $5 billion venture fund that returns 1.01X gets to say that they are in the 75th percentile and can raise their next fund... [and] the GP... is going to make more money than Bill with his $500 million fund that returns 3X.

[0:02] After [music] saying he was out, now [0:04] Bill Maris is returning to the investing [0:06] world. The founding CEO of Google [0:07] Ventures has raised [music] $150 million [0:10] for his new fund called Section 32. [0:15] >> With a smaller fund, I have the [0:17] advantage to be very selective in the [0:20] companies that I invest in, the people [0:21] that I hire. We're going to invest for a [0:23] financial return. Any other metric is [0:25] impossible to measure and therefore [0:27] won't succeed. [0:28] >> Think of the change that [music] has [0:29] happened just in the last 100 years and [0:31] what's about to happen in the next 100 [0:33] years with the advent of AI. The world's [0:35] going to change by orders of magnitude. [0:37] >> Thank you very much for that warm [0:39] welcome. I am Bill Maris. I'm the [0:41] founder of Section 32. Prior to that, I [0:44] was the founder and CEO of Google [0:46] Ventures. I was also Google's Vice [0:49] President of Special Projects where I [0:51] incubated Waymo and Google X, Calico, [0:56] and many other [0:58] projects as well. And before that, I [1:00] founded a web hosting and data center [1:02] company, [1:03] which we're going to talk a little bit [1:04] about. [1:05] And [1:07] today I think I'm going to talk to you [1:08] about a few of the lessons I've learned [1:11] on these interesting experiences I've [1:13] had in life. So, we'll start We're going [1:15] to have four lessons I'm going to talk [1:17] about and we're going to go back to 1997 [1:20] to start when I was a fresh college [1:23] graduate. I had a degree in [1:24] neuroscience. [1:25] And I found myself on Wall Street. [1:28] Somehow managed to land a job there, but [1:30] I was miserable having to wear a suit [1:33] and trudge to work in the heat, but one [1:36] good thing came of that, which was I [1:39] looked in the closet of the office one [1:41] day and I saw a server. And I asked, [1:43] "Well, what is this [1:45] thing beneath our jackets?" And they [1:47] said, "Well, that's where our email and [1:50] websites [1:51] live." And And as can happen to many of [1:54] us, I I had a moment where I felt like I [1:56] was bathed in the light of inspiration [1:59] and and I thought [2:01] I thought [2:02] I think I've glimpsed the future. [2:04] Uh I I I I think I can maybe make a [2:06] business out of this because [2:08] if you can have our website and email in [2:11] your closet, how many websites and [2:13] emails could I put in my closet? So, I [2:15] immediately quit my job [2:17] I because I I had I had kind of glimpsed [2:20] through a keyhole and through that [2:22] keyhole I thought I saw the internet and [2:26] I saw a data center and it looked [2:28] something like this or maybe when I say [2:31] data center [2:32] you think of a something like this or [2:34] something like this, but in 1997 [2:37] the state-of-the-art data center looked [2:40] almost exactly like this. [2:42] We had three servers, [2:44] a small, medium, and large. Uh [2:47] business grew, we eventually had five [2:49] servers [2:51] and this isn't a data center at all. [2:52] This was my apartment where I founded [2:54] the company [2:55] with credit cards [2:57] and the servers lived in one room. [3:00] The work happened in the other room [3:03] and we get very hot in that room [3:06] and this was in Vermont. So, I opened [3:08] the windows and then we get very cold. [3:10] So cold in fact [3:12] that by noon if you had a glass of water [3:14] on your desk it would ice over. [3:17] You may think though [3:19] this isn't so bad, but but actually this [3:21] was also my apartment as well. This was [3:23] the bed [3:24] and you may look at that and think well [3:27] you've got a mattress and a nice pillow [3:29] and then look at that nice blanket, but [3:30] this is a rug I got from Home Depot to [3:32] keep myself warm on those nights and one [3:35] day [3:36] there was a thunderstorm. [3:37] The roof started to leak [3:39] and I knew I needed to do something [3:42] because water and computers and servers [3:44] don't mix well. So, so I called the [3:46] landlord and said [3:48] the roof's leaking. The landlord said, [3:50] well, that happens sometimes. [3:52] But I knew that I needed to do [3:54] something. So, when you don't know what [3:55] to do, you go to Home Depot. I got a [3:57] bucket of tar and a mop, and I went up [4:00] on the roof, [4:01] and there was lightning, and there was [4:03] rain, and I went up there and I I tarred [4:06] the roof. [4:07] And I did not glimpse the future in that [4:09] case because I didn't know when you're [4:12] tarring the roof that you should start [4:14] at the far corner and work towards the [4:15] door rather than [4:17] the reverse, and I tarred myself into a [4:19] corner, but the choice that I faced was [4:21] either [4:22] the servers get electrocuted, or perhaps [4:24] I get electrocuted, but as an [4:26] entrepreneur, I was willing to take that [4:28] risk, which, you know, news flash, I [4:30] survived. Uh my shoes, though, are still [4:32] stuck on that roof uh in Vermont, which [4:35] takes me uh to uh [4:38] lesson two, which is [4:40] to see the future, sometimes you need to [4:42] be a little bit insane. [4:44] Uh [4:45] it may appear to those around you that [4:48] you were tarring the roof in the [4:49] thunderstorm, [4:51] and to that point, I'm going to share a [4:52] few slides here that a friend named [4:55] Stewart Butterfield was kind enough to [4:57] share with me. [4:58] And here's the inauguration 1989, [5:01] >> [snorts] [5:01] >> and [5:02] there's someone taking a picture. [5:04] That makes sense, probably a film [5:05] camera, and [5:07] 2005, it it's not very different. [5:08] There's still someone back there taking [5:10] a picture, and then [5:12] let's go just 4 years later, to another [5:14] inauguration, [5:16] and if we look closely, it's quite a bit [5:18] different [5:19] because now everybody's got a camera. [5:22] Everybody's got a camera, and this was [5:23] kind of before cameras were mushed into [5:26] cell phones. It was kind of around that [5:28] time it was starting to happen, but but [5:29] that's not the most interesting thing [5:31] about this photo because in this crowd [5:33] is someone who, to his friends, I'm sure [5:35] seemed insane, who also did glimpse the [5:37] future. If we look closely, [5:40] this gentleman has decided to, I don't [5:42] know, live stream, or record the [5:44] inauguration on his laptop. Uh [5:47] he knew something that those around him [5:49] didn't know, which is one of the things [5:51] that I've always looked for in [5:52] entrepreneurs is they know a secret [5:55] about the future that most of us don't [5:57] believe. [5:58] Let's fast forward to 2007. [6:01] I find myself somehow at Google. [6:03] Uh and a challenge was given to me. [6:06] Uh the challenge was Google needs a [6:08] venture fund. [6:09] Uh we were starting to make some [6:11] investments. Uh we didn't have a [6:12] coherent strategy. There were no [6:14] budgets. I had to figure out what to do. [6:17] Uh so I [6:18] first found a friend, Rich Miner, who's [6:20] the co-founder of Android, [6:22] uh and he became my partner in crime as [6:24] we conceptualized what what could Google [6:26] Ventures be. Uh we went up and down Sand [6:29] Hill Road and we [6:31] we talked to [6:33] everyone. Anyone that was willing to [6:34] talk to us and have a conversation, we [6:36] were willing to talk to to see what we [6:38] what we could learn. [6:40] Uh [6:41] we came up with a plan. [6:43] Our plan was to obtain all the data of [6:46] venture that we could find. And being [6:48] Google, you can imagine it was a lot of [6:50] data. Historical data, you name it. [6:54] Uh [6:54] then we decided we would as step two use [6:57] AI, but at that time [6:59] Google would not let us use the term AI. [7:02] And this persisted for many years. [7:04] Bill, [7:05] AI is science fiction. It is it's a [7:07] hundred years away if it's ever going to [7:09] happen. [7:10] Uh let's stick to machine learning. By [7:12] the way, when you say AI, it freaks [7:14] people out. So stop freaking So we had [7:17] to call it machine learning [7:18] and we used machine learning to do two [7:21] things: design [7:22] the ideal portfolio construction by [7:25] running millions and millions of [7:26] simulations and back testing and all of [7:29] the things you can imagine that data [7:30] scientists would do. [7:33] And and to determine what the ideal fund [7:35] size would be. And people were excited. [7:39] Here's a headline from TechCrunch at the [7:41] time. [7:43] And and people that inside of Google [7:45] were also pretty excited. This is one of [7:46] the senior execs I later learned had [7:49] this to say. [7:51] And you know, I I have to admit it [7:54] seemed crazy. The plan seemed crazy at [7:56] the time, but let's look at how it [7:58] turned out. So over this time period [8:01] 2009 to 2018, top quartile VC returns [8:04] looked like this and top decile looked [8:06] like this. [8:08] Using publicly available information, [8:10] I'm not sharing any [8:11] non-public proprietary Google [8:13] information. We would estimate Google [8:15] Ventures returns at about 4.1x. [8:18] And I [8:20] adhered more closely [snorts] to the [8:22] strategy and the investments that I led [8:24] and the investments that I led turned [8:26] out like this, which takes me to lesson [8:29] three, which is don't bet against [8:31] computer science. I've seen it [8:34] happen many many times and many many [8:36] fields. If you apply the right kind of [8:37] computer science at the right time to [8:39] the right problem, [8:41] you will get to the right answers. I [8:43] would not bet against it. [8:45] Even if it looks like you're tarring the [8:46] roof in a thunderstorm. So let's fast [8:48] forward to 2017. [8:51] I decided to start my own fund. And [8:53] again, those around me said, "You're [8:55] insane. Why would you do that? You're in [8:57] the warm womb of Google. Lunch is free [8:59] and the massages are plenty and so [9:02] forth." But after the idea, you know, [9:04] sunk in, I the advice turned into raise [9:06] as much money as possible. You know, [9:09] that's the right way to run a fund. [9:11] You'll get a big management fee. You'll [9:13] be happy. Things are going to work out [9:14] really well for you. [9:16] And I thought about that relative to [9:18] everything I had done up to that point [9:19] and I decided to to not take that [9:22] advice. And [9:24] over the course of my time at Section [9:26] 32, we've had six funds. We've invested [9:29] in companies like CrowdStrike and Cohere [9:32] and Coinbase. [9:34] And all six of those funds have [9:37] averaged [9:38] about 400 million in size and all are [9:40] performing in their top decile. And to [9:43] the extent there is DPI to measure, [9:45] that's the only measure as far as I'm [9:46] concerned in venture that counts is DPI, [9:50] which takes me to lesson four. That this [9:52] is this will be heresy to some, but [9:54] small funds outperform large funds. This [9:57] is simply the math. This is not [10:00] an opinion I'm trying to convince you [10:01] of, but there are many reasons for this. [10:04] Smaller funds you can have more focus. [10:06] You have I mean I've I've already [10:07] managed a multi-billion dollar fund with [10:10] hundreds of employees. It's distracting. [10:12] You cannot give the attention to [10:14] founders that I would like to give. [10:17] It [10:18] There There are many reasons for this. [10:19] And if we look at [10:21] top decile performance of DPI, [10:24] um [10:25] funds smaller than 750 million average [10:28] return of 4.76x and funds larger than a [10:31] billion [10:32] 2.42x. [10:34] Funds below 750 million across that time [10:37] period represented 95% of top decile [10:40] performers with discontinuous return [10:42] compression above 750 million. [10:46] Why is this? There's a lot of reasons [10:48] for this. [10:50] You can use your own numbers, but I'll [10:51] just do a little thought experiment. If [10:53] you have a 500 million dollar fund [10:55] and let's say on average these days you [10:57] can own 10% of a company, [10:59] you need 5 billion dollars of exits to [11:02] get your money back. Let's just remind [11:04] ourselves that the 75th percentile of [11:06] venture loses money [11:08] and there is persistence of performance [11:10] of the top quartile. So So if you need 5 [11:14] billion to get your money back and and [11:15] if you want to be in this business for [11:17] the long term, let's say you set your [11:19] your goal at 3x, you you need to return [11:21] 15 billion dollars of exit value in your [11:24] companies. Now if you have a 7 billion [11:26] dollar fund and we do the same math [11:28] through [11:29] you know, you've got to return 210 [11:31] billion. 7 billion to to 70 * 3x is 210 [11:35] billion, which uh [11:38] exceeds the total venture-backed M&A and [11:40] IPO exit value in most years. [11:42] Uh [11:43] this year may be an exception, but I [11:45] that is something I'm looking forward to [11:46] talking about when we sit down. For [11:48] those of you we've crunched the numbers, [11:50] we've done all the math. Those are [11:51] Bill's four lessons for today. I hope [11:53] that they're somewhat useful. There's a [11:54] lot of stories behind all this and I'm [11:56] looking forward to talking about them [11:58] for a few minutes with the guys. Thanks [11:59] so much. [12:02] >> You guys are old friends. [12:03] >> Yes, we are. [12:03] >> We go way back. Well, Bill's when he [12:05] started Google Ventures [12:07] I was the first ex-Google [12:11] company you invested in. [12:12] >> That's correct. And [12:12] >> How did it go? [12:13] >> Climate Corp. A billion-dollar exit to [12:15] Monsanto. [12:16] >> What was your multiple? What was the [12:17] return? [12:18] >> Oof, I don't know. [12:19] >> It was actually good for you guys. [12:20] >> It was quite good, yes. [12:21] >> Yeah, you guys were in the B and the [12:22] >> Yeah. It was when billion dollars was a [12:24] a lot of money then. [12:26] >> That back then that was a good deal. [12:27] >> That would have been That would have [12:28] been the seed round. [12:29] >> Like an A round? [12:30] >> Yeah, that would have been your A round. [12:32] Now we're going to do it again with A [12:33] Halo, so [12:33] >> Now we're going to do it again with A [12:34] Halo. Um so [12:37] you know, I just want to juxtapose what [12:39] you said with what Thomas shared. [12:42] They've got a very large kind of capital [12:44] base that they invest and they're [12:46] investing significantly in these later [12:48] stage rounds of these well-proven [12:49] companies where it's, you know, the data [12:51] he shared is that you can get [12:53] significant multiples to get to that [12:55] next phase. You know, you're more likely [12:57] to go from a billion to 10 billion and [12:59] then you're more likely to go from 10 [13:00] million to 100 and 100 to a trillion, [13:02] trillion to whatever. [13:03] Um you know, doesn't that justify an [13:06] alternative strategy to what you're [13:07] saying of having smaller funds focused [13:10] on venture that you can maybe barbell [13:11] it, have smaller vehicles focused on [13:14] venture, and then very large vehicles [13:16] that bet on the sure things that have [13:19] that durability and that compounding [13:20] advantage, and you can kind of have the [13:22] two together both be 3x return. [13:24] >> So, my observation on that would be one, [13:27] I haven't seen the data science to [13:28] support that second conclusion of the [13:30] late stage companies that that can be an [13:32] an ongoing trend other than this one [13:35] moment, this weird moment in time with [13:37] these multi kind of trillion-dollar [13:39] exits that are coming. That that would [13:40] be kind of observation one. Two, [13:43] um would be at a certain point and this [13:46] is not a negative, it's just an [13:48] observation. If you're an RIA and [13:50] you're, you know, collecting assets, [13:52] that is not venture. You know, venture, [13:55] as I practice it at least, is a [13:56] different craft where you are making [13:59] concentrated bets of your time and [14:02] capital on entrepreneurs and helping [14:04] them build a business. And there's [14:07] nothing wrong with late stage investing. [14:09] However, I also have a an observation [14:12] that [14:13] I I a uh a bit of an objection to [14:17] uh companies that wrap themselves up in [14:20] public benefit [14:22] language and then [14:24] uh [14:25] keep the value creation uh to themselves [14:28] and an elite group of investors through [14:31] a big part of the curve and then say, [14:34] "Well, we're here to benefit humanity." [14:35] Well, what humanity needs is money. So, [14:37] it would it might be better to go public [14:39] sooner because we'll see how these [14:42] multi-trillion-dollar IPOs go. However, [14:45] if I'm Google and I don't speak for [14:47] Google and I decide to arbitrarily cut [14:50] the cost of, you know, [14:51] tokens to 80%. I'm going to cut them in [14:55] What happens to the business models of [14:56] Open AI and Anthropic at that point? [14:58] >> What happens? Tell us. [14:59] >> Actually, what you know, what does [15:00] happen? [15:00] >> well, if you're a company and you can go [15:02] to Google and Gemini and you can pay 80% [15:05] less [15:06] for that [15:08] basically identical product [15:10] why wouldn't you do that? And then the [15:14] compression and the pressure on those [15:17] other businesses goes super critical. [15:19] >> What are the chances that the other shoe [15:20] has fallen that [15:21] >> might happen. [15:22] >> If I were Google, that's what I'd do. [15:24] >> Walk us through the scenario where [15:25] Google decides with their war chest [15:28] with their money printing machine, you [15:30] know what? [15:31] Their margin is my opportunity. [15:34] I'm going to give tokens out 20 cents on [15:37] the dollar. Every time they lower their [15:38] price, I lower our price. [15:40] What happens [15:42] on the playing field? Walk us through [15:44] that. [15:45] >> Would that not be the rational thing for [15:47] >> It's clear they're going to do it. [15:49] >> Well, it may not [15:50] >> be a margin though to the [15:51] >> They they may be burning investor cash [15:52] sort of like an Uber type model, grab [15:54] market share, [15:56] >> Capital as a weapon, tokens as a weapon. [15:58] >> Token as a weapon, grab market share, [15:59] grab an install base on consumer and [16:01] enterprise. But fundamentally at some [16:03] point, you got to have cash generation. [16:06] >> So that's 100% possible. [16:08] It's 100% probable. [16:09] >> Look, I'll just it's been said before, a [16:11] trillion for spend commitments [16:13] on 60 billion dollars of revenue. And [16:15] now you're going to go to the public and [16:16] hope that retail is going to pick that [16:19] up. [16:20] >> Yeah, tell tell us about companies [16:23] staying private longer [16:25] and how unfair that is to the bottom [16:28] half of society who don't get to [16:29] participate in it. [16:30] >> for those 99% who are mostly not us, [16:33] right? So So your 401, you know, those [16:36] 401ks, those retirement plans to get [16:39] into those companies now which are [16:42] getting bizarre exceptions to S&P 500 [16:45] rules that they're all of the rules are [16:47] being broken. [16:48] Uh the passive funds, the uh ETFs are [16:51] going to have to pick them up. And where [16:53] do you think we are on that curve of [16:55] value creation? Could they go 3x from [16:57] here? Sure, but [16:58] >> So the [17:00] just to say it as plainly as possible, [17:03] we're going to force overpriced products [17:07] on the 401k holders of America who [17:09] didn't get to participate early. This is [17:11] your position that this is profoundly [17:13] unfair and creates more wealth creation [17:15] for the people who don't need it and it [17:17] makes the people's retirement accounts [17:18] the bag holders. [17:19] >> There's a lot of risk in that and my my [17:22] my objection is don't say you're doing [17:25] this for the benefit of humanity and do [17:27] the other thing. [17:28] >> Make the public's retirement accounts [17:30] the bag holders. [17:32] >> Or just say this is how we're running [17:33] our business and this isn't for the [17:35] benefit of humanity. [17:37] >> Bill, do you think that [17:39] what happens to venture? I asked Thomas [17:40] this question. [17:42] When these dollars get distributed, [17:44] there's going to be a handful of funds [17:46] that have ginormous returns. I mean just [17:49] unbelievably excessive. Founders two, [17:52] you know, is going to print a hundred [17:54] billion dollar return on two hundred [17:56] million dollars of invested capital. [17:58] But that's one fund in isolation. [18:00] >> Right. [18:00] >> Right. And there'll be a few. Your funds [18:02] when you were at GV are going to print [18:04] an enormous [18:05] upside. [18:08] And so if you don't look closely though [18:10] at beyond the averages, venture's going [18:12] to look incredible. If you look past the [18:14] averages, you're still going to look [18:15] extremely bimodal. A handful of winners [18:17] and a ton of losers. [18:19] >> How does that play out? [18:21] >> one, that's how venture is, right? 75% [18:23] of funds lose money. But two, in order [18:26] for Founders Fund or pick any fund to [18:28] get that hundred billion out, they have [18:29] to sell that stock to someone else. [18:31] Otherwise it's just on paper. So who's [18:33] the buyer for that? Is it is it retail? [18:36] Is it [18:37] you know, what [18:38] you've got to make a a business case in [18:40] the public market that can show that [18:42] this business is worth a discounted [18:44] value of its future cash flows. And so [18:47] whether it's SpaceX or Anthropic or so [18:49] forth, like can that case be made? We'll [18:51] see six months after or so. I know [18:54] they're playing with the [18:56] with the lockups to kind of drag that [18:57] out, but we'll we'll see what the public [18:59] market thinks of that. [19:00] >> Okay, [19:01] so we have we have this one set of [19:03] companies [19:04] and then there's everything else. What [19:06] do you like in the everything else [19:07] bucket as a venture investor? [19:09] >> So so I'm going to make an analogy to [19:11] the gaming industry. We all I get asked [19:14] and we all think about, "Well, what does [19:15] the future look like, you know, when [19:17] when AI is everywhere?" And and there's [19:19] doomers on one side and utopians [19:21] >> Zork? [19:21] >> on the other. That's Zork. I'm going to [19:23] get to that. Just just give me Bear with [19:24] me 30 seconds. It's probably not as bad [19:27] or as great as everyone says. So, let's [19:29] look at the gaming industry. So, I used [19:31] to play this game, Zork. There's one [19:32] called Planetfall back in the '80s. And [19:35] it was very brittle. It was turn [19:37] response, turn response. Grab the lamp. [19:40] Oh, I didn't It's a lantern. I should [19:42] have said lantern. Go north. And and you [19:44] wait for the computer to respond. Let's [19:46] show the most sophisticated retail [19:48] available AI system out there today on [19:50] the next slide. And tell me how [19:53] different it looks. So, so what what's [19:55] happened to the gaming industry from the [19:57] '80s to today is going to happen in AI, [19:59] but in the next like five years. So, [20:01] that will be compressed in terms of how [20:03] quickly that change happens. But, we [20:05] would all agree games are better today [20:08] than they were then. They're photo [20:09] realistic. You can like inhabit them. [20:11] And they're they're they're moving very [20:12] quickly. On the AI side, there'll be [20:15] ambient computing. There'll be uh [20:18] The problems that Zork had will be [20:20] solved for AI. Lack of memory, lack of [20:22] consistency, session resets, and and so [20:24] forth. How did we get there? You have To [20:27] answer your question, [20:28] I don't plan on investing in kind of [20:31] larger models, right? Just like wasn't [20:35] uh better stories that would make better [20:36] games. It was controllers and physics [20:39] engines and GPUs and and those are the [20:41] parts of the AI uh cycle that I'm [20:43] interested in, which is which is all the [20:45] platforms that need to be built to [20:47] >> machinery. [20:47] >> You're correct. That is going to make [20:49] this reality real in the next five [20:51] years. And it's not just bigger models. [20:53] I think we're at the Atari command line [20:56] stage of of AI. And we're going to get [20:59] to the, you know, PlayStation 10 stage [21:02] in the next five years. [21:03] >> You uh you also used to do a a of stuff [21:05] in life sciences. [21:06] >> Yeah. [21:07] >> Um not as much anymore. [21:09] Yeah. [21:10] >> My interest in life science, I founded [21:11] Calico and been very interested in that [21:13] space and we were investors in Flatiron [21:16] uh and Veer and lots of other companies. [21:19] Uh I'm very interested in that space [21:21] because it has a dual benefit of helping [21:24] people and also do good, do well. [21:25] >> Correct. However, uh the uh the [21:29] therapeutic space that requires human [21:31] clinical trials is a specialist [21:33] investment area that uh [21:36] we're not uh spending a lot of time on. [21:38] I'm very interested in computational [21:39] biology and and and those areas, which [21:42] is [21:43] >> if you just look on X that there's a [21:44] renaissance happening in human health. [21:46] I don't know if that's true, whether [21:47] it's cures for pancreatic cancer, cancer [21:50] vaccines, peptides, obviously. There's [21:52] just an explosion and a lot of it seems [21:54] to come back to computation. [21:56] Um [21:57] but this class of winners so far is not [21:59] really computationally driven. It was [22:00] just really good science 10 years ago. [22:03] >> Yeah. [22:03] >> And so do you think that we're about to [22:04] see this massive [22:06] >> I hope so. So I started Calico and and [22:09] again it was like fringe science [22:10] longevity at the time. And now we're [22:13] investors in uh um New Limit, which is [22:15] Blake Byers and uh and Brian Armstrong's [22:17] company and a number of other companies [22:19] in that space, which doesn't seem so [22:21] crazy anymore. However, because of the [22:25] human biology and the FDA, if you find a [22:28] compound and you think you've got [22:30] something, that's like 5% of the work. [22:33] We there's still all kinds of titrating [22:35] and safety testing that needs to go on. [22:37] And so I don't think it's going to go [22:39] quite as exponential as we would all [22:41] like it to. However, if we can achieve a [22:44] realistic simulation of a human cell in [22:46] silico, then you will see that [22:48] accelerate as well. We're not quite [22:50] there yet. [22:51] >> But generally we're seeing some might [22:53] say a flight of capital to India and [22:55] China right now. Are you seeing that [22:56] that their biotech path to market is [22:59] faster if you invest in firms that are [23:03] based offshore versus the US. [23:04] >> has always um uh [23:07] indexed on human safety over speed to [23:11] market, and that has cost us in some [23:13] ways. However, [23:16] some other countries are indexed in the [23:18] opposite direction, which costs lives in [23:20] that. So, there's a balance there, uh [23:22] but there are certainly [23:24] there's research going on in China and [23:26] other places, experiments in cloning and [23:28] all sorts of things that that as far as [23:30] I know aren't happening here. Uh so, [23:33] yes, and I think the gutting of the CDC [23:36] and the NIH and and anti-science [23:40] vibe that has now pervades this country [23:44] has driven a lot of mind share elsewhere [23:47] as funding is drying up for basic [23:50] research. [23:51] >> China's got their own paper clip model [23:53] now. They're recruiting some of the best [23:54] scientists from Europe and India, and [23:57] they're all emigrating to China [23:59] >> Yeah. [23:59] >> to go do work, and that used to be a [24:00] scientific pool that we used to access, [24:02] and we used to recruit. [24:04] >> And we're losing We really need the the [24:07] the neurological reserves here. [24:10] Uh and this business with [24:13] >> Or brain trust would be another way to [24:14] say that, but yeah. [24:15] >> well. But the the H- [24:17] the the the pushing out of H1B holders [24:21] like there's so much happening now that [24:22] it's causing it's just easier to go [24:24] elsewhere. That's not good for science. [24:26] >> What's your view on what's been called [24:28] deep tech for the last decade? These [24:30] traditionally long investment cycle, [24:33] capital intensive, high-risk like Elon [24:36] is one of the few entrepreneurs that has [24:38] successfully tackled uh deep tech [24:41] business model with SpaceX and Tesla. Is [24:45] this becoming a more tractable area for [24:48] entrepreneurs to activate and for [24:50] investors to invest in because of AI [24:52] enablement and physics engines and [24:54] >> Absolutely, cuz things are moving so [24:55] much faster. What [24:57] >> like that are you focused on investing [24:59] in? [24:59] >> I mean human biology and healthcare [25:02] that's probably the largest TAM in the [25:03] world. So super interested in that. And [25:06] then all of the others I I mentioned [25:08] that kind of underlay [25:10] the AI revolution which are the the [25:13] physics engines and the controllers and [25:15] the GPUs and the everything that is [25:17] going to take to to get us there. [25:19] >> I want to [25:20] bring I want to bring Sax and Freeburg [25:21] before we run out of time if it's [25:22] possible. Sax [25:24] I'm curious your thoughts on the venture [25:26] capital business. I think you've did [25:28] five craft funds or four? [25:30] >> Well, we've done four venture and two [25:32] growth. [25:32] >> I'm assuming you're going to be going [25:34] back into the venture business. But I'm [25:35] curious your take on when you started in [25:38] venture and when we started as [25:39] entrepreneurs 25 30 years ago, this was [25:42] a much different playing field. What are [25:44] your plans based on you know sort of [25:47] Bill's [25:48] um look at this and do you believe in [25:50] the $500 million fund sweet spot or do [25:53] you think you need to become Andreessen [25:55] Horowitz when you go back to the private [25:56] sector? [25:57] >> Well, I don't I don't think we need to [25:58] become Andreessen Horowitz um but um [26:02] you know, I I look I think fund size [26:05] determines fund strategy. [26:07] And the size of your fund cuz you're [26:09] going to divide your fund size by 20 to [26:12] 25 names to achieve some [26:15] portfolio diversification and [26:16] construction. That'll determine your [26:18] check size and that sort of determines [26:20] where you play in the market. [26:22] The thing that's spinning through my [26:23] head after Tom's presentation [26:26] is you know, are you better off just [26:29] focusing on you know, let's call it what [26:33] used to be called I don't know late [26:34] venture early growth. You know, you're [26:36] writing $50 million checks. You just [26:38] kind of wait for the breakouts as [26:41] opposed to playing in this really noisy [26:43] super early stage game. Well, I think [26:45] the problem [26:47] with that is we have to look at the [26:49] incentive structure of venture. So a [26:53] $5 billion venture fund that returns [26:55] 1.01 X gets to say that they are in the [26:58] 75th percentile and can raise their next [27:00] fund, and no one at the Stanford [27:01] endowment is going to get in trouble for [27:03] writing that check. They need to put two [27:05] or 500 million into a fund multiple [27:08] times. [27:09] So, so I understand that dynamic. So, [27:12] now let's look at the GP dynamic. Well, [27:14] if I have a $5 billion fund, I return [27:17] 1.01 X, I'm going to make more money [27:20] than Bill with his $500 million fund [27:22] that returns 3 X. [27:23] Okay? So, that's also a strange [27:25] incentive. So, now let's look at the [27:27] entrepreneur side. I am researcher X [27:30] from Open AI. I'm going to start a [27:32] company. [27:33] Bill says, "I'll give you $20 million at [27:36] a $100 million valuation. I want to buy [27:38] 20% of your company." [27:41] Giant fund Y, we're friends, it's a [27:43] different model, but giant fund Y says, [27:46] "Well, we have this giant fund. We need [27:48] to put 250 million in." [27:50] And then an entrepreneur says, "Well, [27:51] but my company's valuation is 100." No, [27:53] your valuation is now 4 billion. And [27:56] we'll give you 250 million for a percent [27:58] of your company. [27:59] They're going to take that deal every [28:01] day. [28:02] Unless you're a seasoned entrepreneur [28:05] who has kind of been down the road and [28:06] knows the pitfalls of that. And so, the [28:08] incentives are broken in all those ways, [28:11] and the pendulum will swing back. So, I [28:13] don't think just staying late stage and [28:16] waiting to snipe her at larger companies [28:19] will be a long-term The The data would [28:20] suggest that's not going to work in the [28:22] long term. [28:22] >> Okay, let's thank Bill. Amazing job. [28:25] >> Thank you. [28:26] Thanks, Bill. [28:27] >> [music] [28:35] [music] [28:41] [music]