「90% 的软件毛利已经死了」：Modal CEO Erik Bernhardsson 谈推理新云、开源见顶之争与专门化模型的千亿美元机会

信源说明：本文整理自 20VC（The Twenty Minute VC）主播 Harry Stebbings 与 Modal 联合创始人/CEO Erik Bernhardsson 的对谈（视频 ID 1dwyL6Iz_PU，约 62 分钟；转写共 32 段、覆盖 0s–3732s 无缺口，已逐段完整读取）。嘉宾身份与关键融资数字经 Axios、Accel 官网、Modal 官方博客及多家媒体交叉核实（详见正文「经核实」标注）；节目中其余数字与判断均为播客自述，以「Harry 称/嘉宾认为/节目引用」归属，本文未逐项独立核实、不为播客观点背书。转写由 Qwen3-ASR 生成，疑似错词按上下文校正并集中列于文末校正表，未确认专名保留原文并标「转写疑作」。相关背景可参见本站对推理算力格局的 earlier 分析：从 Cerebras 上市看 AI 算力新格局。

先说结论

这是 Modal 刚宣布 D 轮之后的一场信息密度很高的对谈：ARR 破 10 亿美元、年增 15 倍、Accel 领投 7.5 亿美元、投后估值 157.5 亿美元（经 Axios/Accel 官网/Modal 官方博客核实；Harry 在节目中称 D 轮「昨天」宣布，与公开报道的 2026 年 10 月 14 日相符）。嘉宾 Erik Bernhardsson 的核心判断有三层：其一，推理就是「新云」，但单卖 token 没有生意——零切换成本必然导向逐底竞争，赚钱靠沙箱、post-training、agentic 等平台产品集；其二，90% 软件毛利的时代已经结束——客户付的每一美元要穿过应用、模型、基础设施、云、英伟达多层栈，每层抽走 20–30%，Modal 的长期目标是 40–60% 而不是传统软件的 90%+；其三，比所有人预期大 10–100 倍的推理市场会碎裂成大量专门化子市场，其中有机会长出千亿美元公司——他用非文本模型 Jev（TypeSafe AI）做例证，并当场回答 Harry「85 亿美元估值你会不会投」：会。对更广泛的读者，这场对话还给出了一组可直接对照自身的基准：GPU 交期 3–4 个月的采购外推法、每月 30–40% 的需求增速、新毕业生被实验室以 2000 万美元以上的薪酬包挖走，以及「每一轮融资都当作最后一轮来融」的纪律。

一、卖 token 为什么没生意：推理是「新云」，赚钱靠平台和毛利栈

先把 Modal 的位置说清楚：它不做模型，做的是跑模型的地方——嘉宾称为「为 AI 时代解构并重构云」（960–1080s）。这套叙事的起点是一个朴素的商业判断：推理市场会变成「 enormously large」的「新云」，但推理本身不够（720–840s）。如果只是把 token 按美元卖出去，切换成本为零，客户随时换供应商，那就是「逐底竞争」（a race to the bottom）。所以 Modal 把推理与 post-training 沙箱、agentic 工作流、低延迟/批量推理等一整套产品捆在一起——用嘉宾的话说，「没人想买单轮把戏，人们要买一整套工具」——而支撑这整套产品的自研基础设施「非常难复制」。

毛利部分是全场最有结构感的一段（840–960s）。Harry 引入背景：当前推理服务商的毛利普遍在 25–40% 区间，能不能随着吃下更多技术栈走向 70%？嘉宾认为几乎没有任何公司今天还有「软件毛利」——十年前硅谷常见的 90%+ 毛利已经死了，而且原因「几乎是拓扑学的」：今天的技术栈是应用→模型→基础设施→云→英伟达，客户在栈顶付的一美元要流过很多层，每一层都要留 20–30% 的毛利。所以他给 Modal 的长期画像是 40–50–60%——「空间很大、钱很多，但你在把它分给栈里的很多层」。这个判断对投资者和创业者都是直接的估值锚：用传统 SaaS 的毛利结构去套 AI 基础设施公司，会系统性看错。（本站此前分析 Cerebras 上市后毛利率指引时也遇到过同一问题：推理硬件的毛利同样被供应链与租回设备挤压，见上文互链的分析框架。）

这条链完整地展开是：信号（纯 token 转售毛利 25–40%、零切换成本）→直接原因（商品化供给 + 无锁定 → 价格战）→谁得谁失（平台型厂商得溢价与 staying power，纯 API 转售商失利润率）→二阶影响（推理层并购与产品套件化加速，「压缩层」论失效）→观察指标（推理厂商的收入构成中沙箱/训练/workflow 等非 token 收入占比、客户续约形态）→读者启发（在 AI 栈的任何一层做生意，先问自己:客户离开你的摩擦力是什么）。

二、比预期大十倍到百倍的市场，正在碎裂成专门化机会

嘉宾承认，推理市场的规模超出了包括他在内所有人的预期——「比我想的大 10 倍或 100 倍」（840–960s）。而大市场会分裂：低延迟/高延迟、各类微调与 post-training、新硬件、「让你用少得多的资源做推理的古怪新技术」。他由此推出本期最出圈的判断：「专门化模型 + 推理市场，有机会造出千亿美元公司」（开场即抛出，1800–1920s 再次展开）。Harry 当场做投委会游戏：Jev（TypeSafe AI 的非文本「校准决策」模型——经 TechCrunch/Bloomberg 核实，2026 年 10 月 9 日 a16z 领投 8.7 亿美元、估值 75 亿美元）凭什么值千亿美元？嘉宾的答案不是基准分数，而是品牌、社区与「运动感」是新时代的巨大资产。Harry 追问「85 亿美元时你会投吗」，他答「会，可能退出、可能上市」。注意一个数字出入：节目说 85 亿，公开报道是 75 亿——本文按转写保留节目数字并标注差异，不替任何一方背书。

支撑「专门化」叙事的是 Modal 自己的供给端打法，这段几乎是一份 GPU 采购方法论（960–1200s）。三月份的英伟达 GTC 被 8 岁写代码、36 岁创业的这位瑞典人形容为「露天毒品市场，只不过货是 GPU」——500 家云厂商在场内狂奔询价。Modal 与 25 家 neo-cloud 合作、自称控制全球约 0.1% 的算力（自称约数），长期目标是把供给侧的碎片产能与需求侧的波动聚合起来高效调度。工程底座是彻底自研：不用 Kubernetes 和 Docker，自研容器文件、文件系统、容器运行时、调度器——数千付费客户的负载各自波动剧烈，聚合之后却是平滑曲线，这让 Modal 敢让客户「按 GPU 秒计费、不签三年锁定量」；Jev 上线当晚在 Modal 上一夜扩容，Modal 事先毫不知情。采购侧的方法是：GPU 交期 3–4 个月，按全体客户的聚合需求做直线/指数外推——过去 12 个月月环比增长约 30–40%，外推三到四个月就是 2–3 倍需求，照此买产能。他也纠正了一个流行误解：并不是全球缺货——英伟达每月出货量「疯狂」（节目称数十万块 Blackwell 级别），价格确实涨了，但「只要你愿意比别人付更多，就能买到很多 GPU」；Modal 结构上反而是 GPU 的多头（长协锁量、短租放出，价格上行可传导）。

硬件价格的时间轴上，嘉宾与被引用的黄仁勋观点一致：Harry 转述黄仁勋在伦敦的回答——供给终将赶上需求，「可能一年，可能五年」；嘉宾的口径是「硬件价格像重力，长期必然下降」，当前短缺造成的涨价可能再持续一两年甚至五年，但终会回落（1200–1320s）。芯片竞争格局他拒绝二选一：英伟达会继续增长，Cerebras、Etched（转写另有两家疑作专名保留原文）等专用芯片也会增长——市场大到来得及容纳所有人。Harry 问 Cerebras 股价跌三成怎么看，他说不炒股、只被动买指数。对 CUDA 护城河，他认为「有点被高估」：CUDA 本身是个挺难用的低层编程模型，真正的锁定是生态——每家开源推理引擎（vLLM、SGLang）与 PyTorch 都是英伟达优先；他自嘲「过去十五年我每年都说明年是非英伟达加速器之年，年年错」，Trainium、TPU、AMD 在追但仍落后（1440–1560s）。对 neo-cloud 向上进攻推理层（Nebius 创始人曾在本节目宣称要拿下推理市场），他礼貌地怀疑：很难同时做好核心生意和别人的市场——最难复刻的是每分钟放置百万级沙箱、秒级伸缩数千 GPU、数十万并发任务的整栈工程；而他对 CoreWeave/Nebius 最尊敬的其实是财务工程：锁长期客户合同→据此融资→买芯片→交付→数年还债，「这是一场资本游戏」（1560–1680s）。

三、开源推理见顶了吗：与 Jason Lemkin 的分歧，和「拥有智能」的边界

开源是本场火药味最浓的交锋点。Harry 转述 Jason Lemkin 的「见顶论」：前沿闭源厂商正在 bluntly 提供极便宜的产品，美国企业又不愿用中国模型，两股风叠加，开源 LLM 推理的使用量将平台期甚至下滑（2280–2400s）。嘉宾的回应干脆利落：「这是一种固定饼大小的思维」——未来十二个月及以上，开源与闭源都会增长 5 倍、10 倍、20 倍，争论谁更大没有意义。这场分歧背后是 Modal 自己的学费：他坦承曾以为开源推理「会是逐底竞争的商品化空间」而按兵不动，DeepSeek 发布后才连夜组队，相关团队从一两个人扩张到十二个月后的三十人，如今是开源推理的前三名提供商——「承认误判、快速补课」本身就是组织能力。他还点出一个结构性变化：开源模型正在集体转向 licensing 收费，能捕获价值就有钱训更好的模型，形成正循环；他并希望出现更多欧洲与美国模型——尊重中国实验室的能力，但「权力更分散对世界更好」（1560–1680s）。对欧洲的判断有两层：先说欧洲难积累巨额资本，随即自我纠正——欧洲的人才资本其实极多，真正缺的是把顶尖人才聚合进同一家公司的能力，而这恰是美国的强项（1680–1800s）。

「每家公司都该拥有自己的智能吗」这个应用层问题，嘉宾给的是两面答案（2880–3000s）：节目引用的两侧案例是 Harvey（称专门化模型很棒）与一家转写疑作「Lagora」的公司（称底层模型变得太快、该策略荒谬）。他的立场：开源微调的真实价值在于底座不会变、不会被拿走、可以持续调参——这解释了为什么总有人要自有模型；但拥有智能只是延迟、成本、性能、可 hack 性、相邻服务等诸多因素之一，不是普适答案。有趣的是 Harry 指出这个诚实对他生意不利（「每家公司都需要，而我们是帮你做的伙伴」才是利益一致的说法），嘉宾笑着举出反例：TypeSafe 的 Jev 是高度专有模型，照样跑在 Modal 上——Suno、Runway、TypeSafe 这类专有模型客户与开源微调客户两头都是 Modal 的生意，通用平台的中立性本身是护城河。模型路由（routing）同理：可以是好生意（他个人投了 OpenRouter 且「结局很好」），但长期更可能被吸收进综合平台（3120–3240s）。

后训练一段值得所有押注 Agent 的人细读（2520–2760s）。他认为 pre-training 向 post-training 的资源迁移「是新常态」，但post-training 有效的边界是反馈机制有限的任务：写代码（可以在沙箱里跑强化学习——很多实验室正在 Modal 沙箱上做这件事）、网络安全、数学（Lean 与 OpenAI 的形式化证明）、法律（evals）。对反馈回路更长的工作，RL 是否够得着，他明确存疑——工程师的工作时间尺度是周和月，「我猜我们终将撞到 RL 的某些限制，但我错了太多次，也许还会错」。顺着工程师话题，他给出本场最「反焦虑」的判断：极度看多软件工程。三十年编程史里，语言、云、框架每次提效都扩大了软件需求；外行最大的误解是把工程师当成「英语翻译成代码」的角色，实际工作是理解需求、在人之间做接口；抽象层一路从汇编→COBOL→C++→Python→如今的 agents 上移，工程师「未来十到二十年没问题」。作为佐证他报了 Modal 自己的账：150 名工程师、每月约 50 万美元花在编码 agent 上——Harry 直呼「低得让人担心，应该十倍、应该 token maxing」，并引用一家公司（转写疑作）「推理花费是人力薪资三倍」的说法；嘉宾笑答团队已经全员 agent-pilled（2760–2880s 处 Harry 亦引用 Claude Code「600 亿美元收入、主导 2 万亿估值」与 Benioff「3 亿美元、工程师薪资 5% 花在 Claude Code」的说法——均为节目引用，未独立核实）。他的结论沿同一机制：效率提升只会让 agent 吃掉更多工作流——「这是计算机与软件五十年的规律」。

四、品牌、人才与钱：新基础设施公司的第二战场

技术栈之外，这场对话有一整条关于组织与资本的暗线。护城河话题上，嘉宾先立规矩：护城河有严格定义（《Seven Powers》），早期公司根本没有护城河——他常回看 Ben Thompson 写 Snap 的「姜饼人策略」：Snap 没有护城河，只是跑得比别人快；今天几乎所有创业公司的真实防御力就是速度（480–600s）。Harry 顺势给出自己的投资三要素：顶级团队、产品决策速度（他举的例子是 Instinct 的功能交付速度——该产品经核实为 2026 年 8 月上线的邀请制个人 AI agent，WhatsApp/iMessage 即可唤起，Harry 自称在上面跑着一百多个自动化，从买演唱会门票到母亲的用药提醒）、以及品牌。品牌怎么造？嘉宾用 Jev 创始人的对照讲得极其生动：「人人都想克隆 Jev，但没人愿意穿上粉色皮草大衣上台、用 unhinged 的方式讲 Jevolution」——品牌=真实+不同=给公司造一个人格，再把它变成一场运动。他自省 Modal 缺一点「人格崇拜 cult」，可以更 loud；而早期 Modal 的获客飞轮是他本人：粉丝数到五万之前，他每天凌晨一点到两点半给每个新关注者手写感谢私信，一晚 50–100 条，其中一半后来成了 Modal 的天使投资人（1920–2040s）。

产品节奏的故事可能是对创业者最有用的一段（2040–2280s）：Modal 的第二产品沙箱 2023 年就上线了，前两年不温不火，直到 2025 年夏天模型（转写校正为 Sonnet 4.5）的代码生成能力突然成熟才爆炸，如今是 3 亿美元 run rate、约占公司收入三分之一——「第二产品用了四年才见成功」，而公司「五年到十亿收入」。更早的 serverless 函数产品 2021 年就发布，真正的杀手应用 Stable Diffusion 是他们没预料到的。他的方法论：基础设施公司要造「积木」，工程师会提各种随机需求，你要蒸馏出基本原语、相信自己的直觉，然后**「在游戏里待得够久，让时机来找你」**。未来的产品地图上，刚上线的训练产品他希望六到十二个月做到 1 亿美元 run rate，后面还有批量、强化学习、数据仓库、向量库、CI、agent harness——他自嘲这是「ADD 大脑」，所以必须无情排优先级、一次只做一两件。

人才与资本的部分几乎是给所有创业公司的对照题（3240–3480s）。人才市场「疯了」：新毕业生被大实验室以 2000 万美元以上的薪酬包挖走（开场白与快问快答两次出现），嘉宾承认那些年轻人「确实 exceptional」；Modal 的对策不是比现金，而是 95 分位薪资 + 股权上行（他顺势说「我们的股价看起来便宜」）+ 员工二级出售先例（上限约 10%、保留 guardrails）——让股权成为「看得见摸得着的硬通货」，招聘叙事就能打赢现金更多的实验室。融资纪律上，他从未做过 pitch deck，D 轮与 Accel 谈了一年多才谈条款；他的反直觉观点是宁可以更低估值融更少的钱、让 VC 对下一轮保持饥饿——VC 总想投得比公司需要的多，那只是稀释。快问快答里他把风险也说透了：没有任何 AI 类目「过热」，他真正紧张的是依赖持续输血的公司——「每一轮融资都当作最后一轮来融，要有通往商业化与收入的真实路径」。问他只投一家天使公司会选谁：turbopuffer（创始人 Simon 是丹麦人，「斯堪的纳维亚连线」），向量库可以长成一个很大的品类；过去十二个月最大的改观是沙箱市场的规模——「收入一直涨，我以为它总该停了，像对收入做逆向心理」。至于 Harry 用估值帮他算的账（若明年收入 40 亿、净利 12 亿，157.5 亿估值只相当于 3 倍收入/14 倍利润，「便宜得离谱」），他照单全收：「我不做这家公司，除非它能成为 generational company。」

五、回调必来：风险清单与观察指标

整场对话的自限与风险表述集中在三处。其一，嘉宾主动给出的「反自我利益信念」：几乎可以肯定会有 20–30% 的股市回调，届时「世界末日」的舆论会炸开——「但我对此完全没问题，因为长期机会大得要命」（3120–3360s，与开场白首尾呼应）。其二，GPU 消费侧的商业化滞后：大量公司还没赚钱却在烧大量 GPU（他举例 Instinct「据称还没有收入」是个好样本），这些公司迟早要开始赚钱，而这个滞后「可能要很久」——这是比「缺卡」更真实的供给侧风险（1320–1440s）。其三，资本约束：利率上行、数据中心建不过来，反而是「对过度建设的保险」；他判断未来五年不可能过度建设算力。给读者的观察清单可以很具体：Modal 的非 token 收入（沙箱/训练）占比走向；开源模型 licensing 化的收入落地；英伟达 GPU 交期与二手 H100 价格（硬件价格「重力」何时生效）；编码 agent 支出占工程师薪资的比例（对照 Benioff 的 5% 与「三倍薪资」的传闻）；以及新毕业生薪酬包是否继续上探——那是实验室资金优势最诚实的温度计。个人向的收尾也很「瑞典」：钱不必然带来幸福、他几乎不卖老股；对两个孩子「推但不猛推」；最期待的应用是 AI 制药——「给我所有 AI 生成的药，我想活得久一点」，虽然 FDA 会让它慢下来。

机制与启发

把全场内容压缩成四条可迁移的链条：

1. 毛利栈链条：AI 技术栈分层（应用→模型→基础设施→云→英伟达）→ 每层 20–30% 抽成 → 90% 软件毛利结构性消失 → 平台化（多产品集）成为推理层唯一逃离逐底竞争的路径 → 二阶影响：估值模型从 SaaS 倍数转向「层内份额×毛利带宽」→ 观察指标：非 token 收入占比、客户切换摩擦 → 启发：在栈上任一层创业，先回答「客户离开我的摩擦力是什么」。
2. 专门化链条：市场规模超预期 10–100 倍 → 子市场（低延迟、批量、微调、非文本模型如 Jev）足以各自撑起大公司 → 通用实验室与专门化玩家分层共存（非零和，英伟达与专用芯片同理）→ 二阶影响：应用层按功能「非捆绑」选模型成为默认架构 → 观察指标：专门化模型的定价与采用（Jev 类「校准决策」模型的收入曲线）→ 启发：别用「通用模型会碾压」一票否决专门化创业——历史（他十五年看错非英伟达加速器）同样警告反方向的过度自信。
3. 供需聚合链条：GPU 交期 3–4 个月 + 单客户需求波动大 → 聚合数千客户得到平滑需求曲线 → 聚合者可以「按 GPU 秒」零售弹性、吃效率价差 → neo-cloud 上攻推理层受制于整栈工程复杂度，胜负手在调度器/文件系统这类「无聊」资产 → 观察指标：GPU 交期、月度需求增速（Modal 为 30–40%）、硬件价格「重力」何时压过短缺溢价 → 启发：任何紧缺商品的聚合层都有同一门生意，前提是你自控调度与计费的底层。
4. 人才—资本链条：实验室以 2000 万美元级薪酬包虹吸应届生 → 现金竞赛创业公司必输 → 用 95 分位现金 + 股权上行 + 员工二级出售先例把股权做成「硬通货」→ 二阶影响：薪酬结构设计成为早期公司核心能力；同时「每轮当最后一轮融」的公司在回调中活下来 → 观察指标：应届生薪酬包、员工二级出售的普遍化、依赖输血公司的融资死亡率 → 启发：融资纪律与人才货币化是同一枚硬币——都关于让你的「纸面资产」变成真钱。

ASR 校正表

以下校正均基于上下文与公开资料核实，机械应用于文末附录；无法确认者保留原文并标注：

未采用条目及原因

• 童年细节（8 岁在 Mac Plus 上写游戏、16–17 岁编程竞赛、36–37 岁因缺自信晚创业）：已在正文以半句带过，未展开——个人史细节，与行业判断关联弱。
• Harry 自述曾可 8000 万美元出售 20VC、母亲劝留：未展开——主持人个人轶事，与主题无关。
• 窗帘决策、《辛普森》Wolfcastle 笑话、财富管理师玩笑、Swedish girlfriend 段落：未采用——纯氛围内容，无信息量。
• Harry「十九分钟 VC」的呼应、双方互相恭维：未采用——无实质主张。
• 挪威税制/主权基金「后经济社会」双关、钢琴育儿、瑞典「别觉得自己特别」文化讨论：各以一句带过或省略——文化背景类，对核心链条无增量。
• 融资细节「与 Rich 和 Gonzalo 谈了一年」：已并入正文（Accel 关系前置），人名拼写未独立核实。
• 「Instinct 不创收」表述：已在风险清单采用，标注为节目口头的「据称（allegedly）」，未独立核实。

转写原文（修正版）

以下为整份转写全文：已按上文「ASR 校正表」对可确认项做机械替换（替换清单与校正表同一份），「转写疑作」项保留原文；原始转写由 Qwen3-ASR 生成，共 32 段、覆盖 0s–3732s、无缺口。每段英文原文之后附该段完整中文译文（译文含「转写作 X」标注处，指附录英文原文中被校正的词）。

[0s - 120s] I think there is an opportunity to create a hundred billion dollar company with a lot of specialized models in inference. There’s going to be a pullback. I’m okay with that because I think the long term opportunity is so fucking big. It’s like so large. Today, I’m thrilled to welcome Erik, co-founder and CEO of Modal. Now, Modal just hit a billion in ARR. They just raised 750 million dollars led by Accel, and they’re priced at 15.75 billion dollars today. But today, we dig into the inference market, what to come, what no one sees that everyone should see, and so much more. We’ve lost new grads going to you know some of the labs for like 20 plus million. I mean, these are exceptional people, exceptional. I don’t want to sell this company because I would just go and start Modal too. Ready to go? Erik, listen, dude, it is so great to have you on the show. You know, I was sitting downstairs with you chatting before, and I was just like, you can feel when it’s going to be a good show because people are themselves, and that was an amazing conversation, and we’re going to have a lot of fun. So thank you for joining me. It’s so exciting to be here. I mean, I’ve listened to so many episodes. It’s it’s kind of surreal to be inside of it right now. That’s really awesome. As I said to you, you you could have texted me when you started the company, and maybe I would have been a little bit richer. Yeah, but you probably didn’t know about me then. No, I didn’t, but I would have welcomed investing in another Swede. Um, I wanted to start with. I was talking to Zach Frankel last night, and he said about your childhood and like early indicators of success and like exceptional ability. How did you fall in love with technology? And for me as an investor, what would have made me excited in seeing your childhood? I started coding 1992 or something like that. I was eight and just loved it. You know, it’s like one of those things. Like I realized you can control computers. I had a Mac Plus, you know, one megabyte of RAM or whatever. Um, and yeah, started making games and like I mean, they were terrible.

【译】 我认为，用大量专门化模型做推理，有机会造出一家千亿美元公司。之后会有回调，我接受，因为长期机会大得惊人，就是那么大。今天，我非常高兴地请到 Erik，Modal 的联合创始人兼 CEO。Modal 刚刚突破 10 亿美元 ARR，刚融了由 Accel（转写作 Excel）领投的 7.5 亿美元，如今估值 157.5 亿美元。今天我们深挖推理市场：接下来会怎样、有什么人人都该看到却没人看到的东西，还有更多。我们流失的新毕业生，去了某些实验室，拿的是 2000 万美元以上的薪酬包。这些人出类拔萃，真的出类拔萃。我不想卖这家公司，因为我会转头再去开一家 Modal。准备好了吗？Erik，老兄，能请到你来太棒了。刚才在楼下跟你聊天我就在想：当一个人完全做自己时，你能感觉到这会是期好节目——刚才那段对话太精彩了，我们会聊得很开心。谢谢你来。来到这里太兴奋了。我听过太多期节目，现在置身其中有点超现实。太好了。就像我跟你说过的，你创业的时候要是给我发条短信，我现在的身家可能更厚一点。是啊，不过你那时候大概还不认识我。确实不认识，但我会欢迎投资另一个瑞典人。我想从……昨晚我和 Zach Frankel 聊天，他谈到你的童年、成功的早期迹象和过人的天赋。你是怎么爱上技术的？作为一个投资人，在你的童年里看到什么会让我兴奋？我大概 1992 年开始写代码，那时我八岁，就是喜欢。你能控制计算机——我意识到这一点。我有一台 Mac Plus，一兆内存还是多少。然后开始做游戏——它们都很烂。

[120s - 240s] You know, it’s just. I think back at like, in a way, I’m like so grateful. I look back at like my childhood. My my friends, they were like collecting hockey cards or whatever. That was like their hobby. Like my hobby was programming. You know, and and now I make a lot of money from it. Isn’t that kind of crazy? Like you you know when you think about it, like the fact that I can make a lot of money like doing what what was my hobby since I was eight. It’s kind of nuts. So, I’m just so grateful about you know the fact that I fell in love with technology and coding, and then you know I started doing programming competitions when I was 16, 17, and so much fun. Modal was your first company, or is your first company? Why did it take so long? I think the consistent pattern in my life is like I never had quite enough confidence in myself. And a lot of it was that was missing was just the confidence, a little bit the money too. But so it wasn’t until I was like 36, 37, I started modal. And I think I don’t think I’ve ever wanted to necessarily start a company. It was like one of those things where like I just saw this big opportunity in the market. I almost thought to myself like Why is no one else doing this? Such an obvious idea. And and that led me to think like maybe I should start a company. And then you know five years later here we are. It’s gone pretty well. I think in hindsight maybe I should have started much earlier. Do you wish you had? I look at some of these like Stanford for kids, and they’re like 21, and like I kind of feel like they maybe have too much confidence. You know what I mean? Like they’re like you know think they’re going to take over the world, and they’re 21. I think I think the ideal age is maybe like I don’t know 32 or something like that. You have a little bit more experience. You know how to navigate people. You know how to hire people. You know how to fire people. Like all these things. I don’t I don’t know if I would have been a good founder like in my 20s. Before we get into the market and everything beneath that, we were chatting downstairs, and we often hear that founding companies is like eating glass. It’s hard. It’s painful. How do you feel about that? Dude, I hate that stuff. It’s like I don’t get it. Like you’re eating glass like every day. Like why are you doing it then? Like I don’t.

【译】 回想起来我满怀感激。我看我的童年：我的朋友们收藏冰球卡什么的，那是他们的爱好；我的爱好是编程。而现在我靠它赚了很多钱。你不觉得这很疯狂吗？八岁开始的爱好成了我赚钱的方式，有点不可思议。所以我非常感恩自己爱上了技术和编程。十六七岁的时候我开始参加编程竞赛，特别好玩。Modal 是你的第一家也是唯一一家公司吗？为什么等了这么久？我一贯的模式是：对自己总差那么一点信心。缺的主要就是自信，钱也算一点。所以直到三十六七岁我才创办 Modal。而且我从没特别「想」创业，只是看到了市场上的一个大机会。我甚至想：为什么没人做这个？这么明显的点子。这让我觉得也许我该开一家公司。五年过去，就成了现在这样，走得挺好。事后看，也许我该早得多开始。你希望当时更早吗？我看那些「斯坦福神童」，21 岁就觉得自己要接管世界——我反而觉得他们信心过头了。理想的创业年龄也许是 32 岁左右：多一点经验，知道怎么与人打交道、怎么招人、怎么裁人。我不确定二十多岁的我会是个好创始人。在进入市场话题之前——我们刚才在楼下聊过，常说创业像吃玻璃，又难又痛。你怎么看这个说法？老兄，我讨厌那套。每天吃玻璃？那你为什么还在做？我不理解。

[240s - 360s] No, you expect some like magic reward, like some big cookie in the end. Like I never understood that mindset. I think it’s like this mindset that a lot of people have. I’m going to suffer pain, and then there’s going to be this big reward in the end. I don’t know. I don’t think about it that way. I enjoy every day I go to work. Of course, there’s like hard times. Like there’s, you know, you have to have hard conversations. Maybe fire people. Maybe like do hard things. But like in the grand scheme of things, like I mean, such is life. And I enjoy life. I I enjoy building a company. And I don’t I don’t think of it as like in one one day we’re going to flip this company and be. I I hope to work on this for twenty, thirty, forty, fifty years. I’m having so much fun. I don’t want to sell this company because I would just go and start Modal too. You really wouldn’t. No, I don’t. It’s funny. I had the chance to sell twenty VC. The show’s always successful. We said earlier to share everything for like eighty million, and I own a hundred percent of it. And that for me is a lot. Yeah. And my mother went for a walk with me, and she said, “What would you do tomorrow? " I said, “I’d start the nineteen minute VC. " Same thing, yeah. She was like, “Don’t sell, yeah, yeah. " What would you do? I don’t think money necessarily makes you happy. I mean, a little bit. Like it’s good to have some money, but you know, be on a point. Like, you know, I like building stuff. You’re such a. I love technology. I love coding. So much fun. You’re such a sweet. Is that Swedish? Is that Swedish mentality? Maybe it is. Yeah. Dude, you never has an American come on before and said it’s not about money. I mean, I do care about money a little bit, but maybe you’re right because I do think in Swedish mentality, it’s like it’s like self actualization, like doing what you enjoy is like kind of the highest form of prestige in a certain way, and it could be music, it could be like poetry, whatever. Like it almost doesn’t matter. What do you do that you don’t enjoy? Yeah. I kind of hate fundraising, honestly. I mean, it’s fine. I don’t know. Like, I enjoy it when it’s like good people, but it’s like kind of challenging. You have to say no to a lot of people. I guess that’s a good problem to have. But you know, in general, it’s it’s. Oh, you have to say no to a lot of people. Oh, Erik, I know. Sorry. That’s a horrible thing to do. Oh, Harry, you can only have ten million, not fifty million. Yeah. Oh, you’ve never done a deck, have you? I I’ve never made a pitch deck. Yeah. Every every.

【译】 不——你期待什么魔法奖励、最后领一块大饼干？我从来不懂这种心态：先吃苦受罪，最后领一份大奖。我不这么想。我享受每天去上班。当然有难的时候：要做艰难的谈话，可能裁人，做困难的决定，但大局而言——这就是生活。我享受生活，享受建公司。我没想着哪天把公司一卖就翻篇。我希望在这件事上干二十年、三十年、四十年、五十年。我太开心了。我不想卖这家公司，因为我会转头再去开一家 Modal（转写作 Model Two）。你真不会卖？不卖。有意思，我有过出售 20VC 的机会。节目一直很成功，有人出价大约 8000 万美元买下全部，而我百分之百持股，对我来说是很大一笔钱。我母亲陪我散步时问我：「明天你会做什么？」我说：「我会创办《十九分钟 VC》。」一模一样。她说：「那就别卖。」你会怎么做？我认为钱不一定带来幸福。有一点用——有点钱是好事，但到某个点就够了。我喜欢造东西。你真是……我爱技术、爱编程，太好玩了。你真贴心。这是瑞典性格吗？也许是吧。老兄，从来没有美国人上节目说「这不是为了钱」。好吧，我确实有点在乎钱，但也许你说得对。瑞典人的心态是自我实现：做你享受的事，在某种意义上是最高形式的声望——可以是音乐、诗歌，什么都行，几乎不重要。那有什么是你不享受的？说实话，我讨厌融资。还行吧——遇到好人的时候我喜欢，但挺有挑战的：你得对很多人说不。这是个好烦恼。「哦，你得对很多人说『不』。」哦 Erik，我知道，抱歉，那太残忍了。「哦 Harry，你只能投一千万，不是五千万。」你从来没做过 deck 吧？我从没做过融资演示文稿。「是啊。每——」

[360s - 480s] Every round we’ve done so far, it’s I I like to get to know people well in advance of of raising money, so so every round we’ve done. I mean, most recently we announced the D round yesterday with Accel. I’ve been speaking to Rich and Gonzalo at Accel for I think at least a year, and at the at some point, you know, we’re like, actually, you know what? We could take some money. Like, what are the terms? Like, let’s discuss terms. And then it happened very quickly because we already had that relationship. How do you think about price optimization? Is it a game of getting the highest price? No, I don’t think so. I I think maybe in later stages it becomes a little bit more of a commodity, like the money itself. But but I think in the early days, like like I mean, for several reasons, like one is that there’s a misaligned incentive. Like VCs always want to put in way more money than a company needs, and and and that you know you just take on a lot of dilution. It’s not good as a shareholder. But the other thing is like you actually want to hold back. I think and and it’s often better to raise a lot less money at a lower valuation and then keep the VCs hungry for the next round. Can I ask you? Is this not the generation though where actually capital is the moat and you need so much more money than you think? You said that VCs want to give you more money than the company needs. I thought AI was finally the output exhaust for venture dollars that was needed. I think there are times in the market where you are capital constrained. We have not been capital constrained for most of the time of the company. We’ve been extremely capital efficient. We burn only a few million a month. I mean, only like you know, it’s a lot of money. I respect for the money. You burn a few million a month. No, I mean up until like last year. Now we have to spend a lot of money on capacity. How much do you burn a month now? A lot. I mean, we we actually are profitable on a P and L basis, but like cash flow is quite negative because we have to spend a lot of money on on on getting GPUs and building out capacity, and so that is a different period in you know in the company and how we think about capital. Capital is I don’t want to call it a moat because you know like I feel like people you know throw around that word word like to mean like things that you know moat to me has like a very strict definition. But but you know so.

【译】 ——我们做过的每一轮都是这样：我喜欢在融资之前很久就去认识人。最近这轮——我们昨天和 Accel（转写作 Excel）宣布了 D 轮——我和 Accel 的 Rich、Gonzalo 已经聊了至少一年。到某个时刻我们说：其实，我们可以拿一些钱。条款是什么？来谈条款吧。因为有这层关系，事情很快就成了。你怎么看价格优化？是把价格抬到最高的游戏吗？我不这么认为。后期阶段，钱本身可能有点商品化了。但早期不是，原因有几个：一是激励错位——VC 总想投得远多于公司需要的，你因此吃下大量稀释，作为股东这不好。另一件事是你其实想留有余地：通常更好的做法是以更低的估值融更少的钱，让 VC 对下一轮保持饥饿。我能问一下吗：这一代难道不是「资本即护城河」、你需要的钱远比想象中多的一代吗？你说 VC 想给的钱比公司需要的多——我以为 AI 终于是风投资金真正的出口了。市场上确实存在资本受限的时期，但我们公司大部分时间不缺资本，我们资本效率极高，每月只烧几百万美元——当然这也是很多钱，我对钱保持尊重。「每月几百万」？不，我是说直到去年。现在我们必须在算力上花很多钱。现在每月烧多少？很多。其实我们按损益表算是盈利的，但现金流非常负，因为要花很多钱买 GPU、建容量。这是公司的一个新时期，我们对资本的思考方式也变了。资本……我不想叫它护城河，人们把这个词用得太滥了，护城河对我来说有非常严格的定义。但，你知道，所以。

[480s - 600s] Certainly helps. Capital helps at this point. What’s the moat definition for you? I mean, moat to me is like you go read like Seven Powers or like one of those things. I love Seven Powers. It’s a good book, but but I think you know then people read that and they get delusional, like you know they’re like, yeah, we’re starting this company, it has all these moats. You never have moats in the early days, like you know, like Google could do anything they wanted to. Google could, yeah, but it’s also like, you know, like what are the seven? I don’t even remember. It’s like cornering resources, like market power, like whatever. It’s like none of those things happen until you’re like ten years in, you know. I there’s an old blog post I always come back to by Ben Thompson. It talks about Snap and the like the gingerbread man strategy. Do you know this thing? And it’s like the whole point is like Snap doesn’t have a moat. They just run faster. Like they’re the gingerbread man, and I actually always come back to it. I think that’s so true. Like for almost every startup these days, it’s like you just have to like move faster than everyone else. I actually think like the three things that matter today when I invest is like one, team. Are they top one percent team? Yeah. Two, speed of product decision making. Very specifically, you look at like Instinct today. The speed of feature shipping is really very impressive, and they’re very. And then three is brand. Yeah. And I think those are the three determinants to me of defensibility moat, but it’s a kind of package. Yeah, yeah. I mean, again, I don’t know if I would, but but I kind of agree your point. Like I think brand. I don’t want to call it a moat, but it’s it’s actually I think the value of brand has gone up a lot. Can I be a dick? Yeah, please. Do you think Modal have done well with brand? I think pretty well. I think we have. It’s coupled with the developer experience and the product and the website and the visual identity and all these things. We just did an event in in San Francisco, and I thought a lot about that, sort of how we package it, the event and the visual look and feel. But like, do you think you have a bigger brand than Baseten and Fireworks? I think we have that. I think we have like really really good vibes. Like I’m I’m terminally online. Like people tweet like way more about Modal. I think there’s like a little bit more of a community. I think there’s like much better vibes around us. Like, but like I am paranoid about those companies. Like I am paranoid.

【译】 资本当然有帮助，资本在这个阶段有帮助。「你对护城河的定义是什么？」对我来说，护城河就是你去读《Seven Powers》之类的书。我爱《Seven Powers》，是本好书。但人们读完会陷入幻觉：「我们创业公司拥有这么多护城河。」早期你根本不会有护城河——那种年代 Google 想干什么都能干。「Google 什么都能做。」对，但还有，七种力量是什么来着？角资源、市场力量之类的——那些东西要到你第十年才会出现。我总会想起 Ben Thompson 的一篇旧文，讲 Snap 和「姜饼人策略」。你知道这个吗？核心观点是：Snap 没有护城河，它只是跑得更快——它就是姜饼人。我总回想这篇。我认为今天几乎每家创业公司都是这样：你只需要比其他所有人移动得更快。我今天投资时看重的三样东西：一，团队——是不是前 1% 的团队；二，产品决策的速度——非常具体地，你看今天的 Instinct，功能交付速度真让人印象深刻；三，品牌。对我来说这三样决定了防御力，但它们是一整套。「对，品牌——」品牌，我不想叫它护城河，但品牌的价值确实涨了很多。「我能毒舌一句吗？」「请便。」「你觉得 Modal 品牌做得好吗？」我觉得做得挺好，我认为我们有。它和开发者体验、产品、网站、视觉识别这一切耦合在一起。我们刚在旧金山办了一场活动，我在怎么包装它、活动的视觉观感上想了很多。「但你觉得你们的品牌比 Baseten（转写作 Base Ten）和 Fireworks 大吗？」我认为是的。我们有真的很好的氛围。我重度泡在网上，人们发推聊 Modal 远多于别人，社区感更强，围绕我们的氛围好得多。但我对那些公司非常警惕，非常警惕。

[600s - 720s] By competition, of course, I track it. Like I, of course, like I, I see everything that gets tweeted and about our competitors, and you know, it’s. But this is important. Like, get everyone wants to be a part of the community. Everyone, you know, what I think is like that. Like people misunderstand is I love TypeSafe with Jev yeah because like everyone wants to clone Jev. No one wants to get up on stage and wear a fucking pink fur coat and just like talk about the Jevolution in an unhinged way. That is what brand to me is like. It’s like that guy is like he’s playing the game on a whole different level, and I think it’s just like so like true about like companies today. Like you have to think about that. What so can you sum up why? Because he’s authentic and different. I think so. I think you just have to create a personality. You kind of have to like figure out like what is the personality of your company and just be consistent and create like almost like a sense of movement, a sense of like community. Like this is something cool, and and people want to use you because you’re one of the cool products out there. Well, like, have you tried Modal? It’s pretty cool. What What has Modal done that Baseten and Fireworks haven’t to build that? I think we started with the developer experience, so that was, and we had a very like serious or like very clear sort of PLG product market fit, product led growth, in the sense that people came bottoms up; they just sign up on a product. I think the consumer background of coming from Spotify is very important in how I thought about building what’s now a B two B company selling tech effectively. You know, it’s like a very boring category, but I think you know, like having personality and and having you know a little bit of an edge is a big part of how you create that sort of movement and that sort of like sense of excitement and the feeling that you’re different. And I wanted to start the moment you go to our website and the moment you try the product and you just feel that it’s different. The magic moment. Can I ask you? We’ve spoken about the three different providers there. Inference is a market has many different speculators commenting on it. When you look at the next five years, how does the inference market play out? First of all, I think inference is going to be an enormously large.

【译】 竞争嘛，我当然追踪。我看到竞争对手被提到的每一条推文，这很重要——人人都想成为社区的一部分。我觉得大家误解的一点是——我喜欢 TypeSafe 搭配 Jev（转写作 Jav）这套打法，因为人人都想克隆 Jev，但没有人愿意穿上该死的粉色皮草大衣上台、用 unhinged 的方式去讲 Jevolution。在我看来那才是品牌：那个人在完全不同的层面上玩这个游戏。我觉得这对今天的公司来说千真万确，你必须去想这件事。「能总结一下为什么吗？」因为真实和不同。我认为是。你必须塑造一个人格，你必须想清楚你的公司人格是什么，然后保持一致，制造出一种运动感、一种社区感——「这是个很酷的东西」，人们用你是因为你是那些酷产品之一。「你试过 Modal 吗？挺酷的。」「Modal 做了什么 Baseten 和 Fireworks 没做的？」我们从开发者体验起步，有非常严肃或者说非常清晰的产品驱动增长（PLG）：人们自下而上地进来，直接在产品上注册。我来自 Spotify 的消费产品背景，这对我现在经营一家本质上是卖技术的 ToB 公司影响很大。这是个很无聊的品类，但有人格、有一点棱角，是制造那种运动感、兴奋感、「你不一样」感觉的重要部分。我想让你从打开我们网站的那一刻、试用产品的那一刻，就感觉到它不一样。魔法时刻。「我能问一下吗——我们聊过那三家提供商。推理市场有很多评论者在投机式发言。未来五年，推理市场会怎么演化？」首先，我认为推理会变成一个巨大的——

[720s - 840s] Market, it’s like the new cloud, basically, and but I don’t think inference in itself is maybe like the only thing you need. One of the things that I feel very strongly about is like it has to come with a whole set of different, like a lot of other products. It has to come with post-training sandboxes, agentic stuff, all kinds of you know other types of inference, like low latency inference, batch inference. So, so building that platform, and by the way, we’ve been working on this for the last five years. I think that to me is you know not a moat, but it’s a very strong. It’s a very hard product to build. And why? Why do you need to have the breadth of that product set? Otherwise, you’re just in a compression layer. Yeah, because otherwise you’re just selling commodity product. I think, and like if you’re just selling like tokens for dollars, there’s no switching cost. You know, people can just switch to different provider, right? It’s then it’s you just it’s just a race to the bottom. But if you couple it with all these other things around it, sandboxes and post-training and and and and reinforcement learning and all these things, like you you can actually create a a whole set of tools. Like no one wants to build a you know buy a single trick pony. People want to buy a portfolio of different things. Like they want to buy you know whole suite of tools that basically lets them build amazing AI applications on top of it, right? And and that to me, and the in the infrastructure underneath is a lot of differentiation, a lot of you know staying power in the market. It’s very hard to copy, like the infrastructure required to build all of these products. What does no one see about the inference market, or you don’t see commented on much that is very obvious to you? One thing that people don’t see about the inference market is I actually think in the long term there is so much gross margin opportunity in the space because you can build all these things around it, all the sandboxes, all the agentic stuff, workflows, post-training, reinforcement learning. If you do all those things, there’s actually so much opportunity to make money. Is it a market that has software margins? Because we’re looking at margins now in the twenty-five to forty percent range. That’s kind of.

【译】 ——市场，基本上就是新云。但我不认为推理本身就是你唯一需要的东西。我非常坚信的一点是：它必须伴随一整套不同的产品——post-training 沙箱、agentic 的东西、各种其他类型的推理，低延迟推理、批量推理。建设这个平台——顺便说，我们已经为此做了五年——在我看来不是护城河，但它是一个非常强的、非常难造的产品。「为什么你需要那么宽的产品面？」「否则你就只是一个压缩层。」对，因为否则你卖的就是商品化产品。如果你只是按美元卖 token，就没有切换成本，人们随时可以换供应商，那就是逐底竞争。但如果把它和周边这一切耦合起来——沙箱、post-training、强化学习——你就能真正造出一整套工具。没人想买一个单一功能的产品，人们要买的是一个组合、一整套让他们能在上面造出惊艳 AI 应用的工具。而对我来说，这下面的基础设施里有大量差异化、大量在市场上立足的能力，要复制这套基础设施非常难。「关于推理市场，有什么是没人看到的、或者很少被评论、但对你来说非常显而易见的？」人们没看到的一点是：长期来看，这个领域有巨大的毛利机会，因为你可以在它周围建所有这些东西——沙箱、agentic、工作流、post-training、强化学习。如果你把这些都做了，真的有非常多赚钱的机会。「这是一个有软件级毛利的市场吗？因为我们现在看的毛利在 25% 到 40% 的区间。差不多是——」

[840s - 960s] Traditional players, I don’t know if that’s yours, but that’s kind of what inference providers are generally. Do they go to seventy as you own more and more of the stack? I think basically almost no company today has software margins anymore. Like what you used to see in like Silicon Valley in the you know ten years ago with this like ninety percent margins plus, I think that’s basically gone. And the reason is actually almost like topological. It’s like you have this like stack of vendors, right? Like today, you have applications, you have models, you have infrastructure, you have cloud, you have Nvidia, and so when a customer pays a dollar at the top of the stack, it has to flow through a lot of different layers. And at every layer, there’s going to be you know twenty percent, thirty percent gross margins. So I think there’s a lot of margin opportunity, and there’s a lot of opportunity to to to make a lot of money. I mean, the space is so big, but you’re also splitting it across a lot of layers in the stack. So, I mean, you know, like the ninety percent software margin businesses are gone. I I think Modal has an opportunity long term to get to forty, fifty, sixty percent margins, but we’re never going to get to software margins. What’s a wild prediction you have on the future of the inference market that you think many people would disagree with? I think one prediction about the inference market is going to split into a lot of different things because the market got so big in the last. I don’t think anyone expected me included. I don’t think anyone expected it to be you know so big. It’s like ten x or a hundred x bigger than I ever thought, and that to me means there’s going to be so much room for specialization, low latency, high latency, like various types of fine tuning and post training. I think there’s going to be some wacky new technologies that let you do it with way less resources, new hardware, like all kinds of different things. I’m really sorry. Does that not go against what you just said about the need to own it all and have that breadth of product? If it’s so much larger that you need real specialization, specialization by product. I think it’s going to fragment into a lot of different products. Now we can build a lot of those products and host it all under Modal, and that to me is very exciting because because a lot of different you know there’s going to be a lot of different products that companies are going to need. For one product, they’re going to need low latency inference. For one product, it’s going they’re going to need batching inference. For one product, they’re

【译】 ——传统玩家，我不知道那是不是你的情况，但一般推理提供商差不多就是这个水平。「随着你拥有技术栈越来越多，会涨到 70% 吗？」我认为今天基本上没有公司还有「软件毛利」了。十年前你在硅谷看到的那种 90% 以上的毛利，基本已经消失了。而原因几乎是拓扑学的：你有一摞供应商——今天你有应用层、模型层、基础设施层、云、英伟达。当客户在栈顶付一美元时，这一美元要流过很多不同的层。而在每一层，都会留下 20%、30% 的毛利。所以毛利机会很大，赚钱的机会很多，市场空间巨大，但你要把它分给栈里的很多层。所以，90% 毛利的软件生意没有了。我认为 Modal 长期有机会做到 40%、50%、60% 的毛利，但我们永远不会达到软件级毛利。「关于推理市场的未来，你有什么疯狂的、很多人会反对的预测？」一个预测是：推理市场会分裂成很多不同的东西，因为过去这段时间市场变得太大了。没人预料到，包括我自己。它比我曾经想象的大 10 倍、100 倍。而那意味着专门化的空间巨大：低延迟、高延迟、各种类型的微调和 post-training。我认为会出现一些古怪的新技术，让你用少得多的资源做推理，还有新硬件、各种不同的东西。「抱歉——这不是和你刚才说的『需要拥有一切、产品要有宽度』矛盾吗？如果市场大到需要真正的专门化……」是按产品专门化。我认为它会碎裂成很多不同的产品，而我们可以造出其中很多、全部托管在 Modal 之下。这让我非常兴奋，因为公司将需要很多不同的产品：一个产品会需要低延迟推理；一个产品会需要批量推理；对一个产品，他们——

[960s - 1080s] They need sandboxes, which we have for one product. They’re going to do pre-training their own models. You know, maybe it’s a you know certain specialized you know audio models or whatever. Like we do a lot of stuff besides LM inference, and and so that to me, that creates a lot of opportunity to basically kind of rethink what a cloud looks like in the era of AI and and kind of deconstruct the cloud and rebuild it for this new era of software. Can I ask you? Earlier, you said about building capacity with GPUs. Yeah. What’s it like doing that today? Is it as crazy as reports suggest? So it is pretty crazy. I spent some time in March at GTC, Nvidia’s big conference in San Jose. I would describe that as like an open air drug market, but GPUs instead of drugs. You’re running around, you know, working, talking to 500 different cloud vendors. You’re like, do you have GPUs? Like, please get and and like you know, in the end, you come out like you do a lot of deals, and and and that’s something we’ve had to develop the muscle. I think part of the value prop of modal is like our customers don’t have to think about that. We do the dirty work of doing all the GPU deals under the hood. We work with 25 different neo clouds, or you know, right now we control like you know 0.1 percent or whatever. I don’t even know the number of like the world’s compute, but like my like you know crazy like you know how I think about the the long term goal is like as we control more and more of the world’s compute, you can aggregate that capacity and aggregate demand on the other side and manage that capacity and demand extremely efficiently. So that to me is like something we actually think we can do. We can fill a big gap. What does that allow you to do? Being that aggregation layer. Well, so we built a lot of our infrastructure ourselves. Like we we don’t use Kubernetes and Docker and things like that. We built our own container files. We built our own file system. We built our own container runtime. We built our own scheduler. We built our own basically how we managed the the capacity, and that that enables us to to basically treat like all of the world’s compute as one big pool of capacity. And so if you if you think about the fact that there’s you know we already have many thousand paying customers, and and over time it just grows and grows. Each of

【译】 ——需要沙箱，某一个产品要沙箱，我们有；某一个产品要预训练他们自己的模型，也许是某种专门的音频模型之类。我们做很多 LLM 推理之外的事情。对我来说，这创造了大量机会去重新思考 AI 时代的云长什么样，把云解构、为软件的新时代重建它。「我能问一下吗——你之前说用 GPU 建容量。今天做这件事是什么体验？有报道说的那么疯狂吗？」确实相当疯狂。三月份我在 GTC 待了一阵——英伟达在圣何塞的大会。我会把它形容成露天毒品市场，只不过货是 GPU。你跑来跑去，和 500 家不同的云厂商谈：「你有 GPU 吗？求你了……」最后你谈成一堆交易。这是我们不得不练出来的肌肉。我认为 Modal 价值主张的一部分就是：我们的客户不用去想这些，我们在幕后做所有 GPU 交易的脏活。我们与 25 家不同的 neo-cloud 合作。现在我们控制的大概是世界算力的 0.1% 还是什么，我都不知道确切数字。但我对这个长期目标的疯狂想法是：随着我们控制越来越多的世界算力，你可以把这些容量聚合起来、在另一边聚合需求，然后极其高效地管理这个供给与需求。这是我们真心认为自己能做到的，我们能填上一个巨大的缺口。「作为那个聚合层，这让你能做到什么？」我们的大量基础设施是自己建的：我们不用 Kubernetes 和 Docker 这些东西。我们自建容器文件，自建文件系统，自建容器运行时，自建调度器，自建我们管理容量的整套东西。这让我们能够把全世界的算力当作一个大容量池来对待。我们已经有很多数千级别的付费客户，而且随着时间只会越来越多。每一个——

[1080s - 1200s] Those customers have like very sort of variable requirements, or like their their their workloads kind of vary over the day, and they have maybe bursts and and and in each individually, that they’re like very sort of unpredictable. But if you pull all of those customers together, you end up with a very sort of smooth overall demand, and that is something we can manage extremely efficiently. And so, one of the value props of modal, besides like we we can scale up with you, is also you don’t even have to do capacity planning. Just come to us; we give you a hundred GPUs. I think it’s so amazing going back to TypeSafe, for instance. Like they, we we we didn’t even we had talked to them, I think, a long time ago before they launched. But you know, they launched overnight on modal, and we we weren’t even aware of it. To just like scale up, and we gave them a ton of GPUs, and it just happened automatically. So so that magic, I think, is is something that is people really value with modal. It’s like we can give them a lot of GPUs on very short notice, and we can handle that sort of variable. Like if you go to us, you don’t have to lock up, you know, like a rectangular block of GPUs over three years. You have to know the exact number. Like it’s it’s pay as you go. You scale up and down. You pay per GPU second. How do how do you do GPU buying in terms of forecasting then? Because you don’t want to have an insane excess of supply just sitting on your books. Totally, yeah. The the GPU lead times right now is about three to four months. So so we basically do a straight line or exponential line extrapolation, and we look at an aggregate across all our customers. We have a very diversified set of customers. You can actually look at the demand in aggregate pretty well and predict based on monthly growth rates, which have been about thirty forty percent month to month for the last twelve months. And you just project that you know three to four months into the future, you end up with like two to three x. You go out and you know buy that capacity and make sure to ramp up. And and sometimes can you just buy that capacity? You can, yeah. I mean, I think I think t here is a little bit of like a misconception that like we’re out of capacity. I mean, Nvidia’s like shipping like an insane amount of GPUs like every month. I don’t know how many hundred thousand Blackwells are getting installed. Prices are up for sure. You have to pay more.

【译】 ——客户的需求都各自非常波动，负载一天里起起伏伏，可能有突发，单个客户完全不可预测。但如果你把所有这些客户放在一起，你得到的是一条非常平滑的总需求曲线，而这是我们能够极高效管理的东西。所以 Modal 的价值主张之一，除了「我们能陪你扩容」，还有：你甚至不用做容量规划。来找我们就行，我们给你一百张 GPU。回到 TypeSafe 特别有意思——很久以前他们发布之前我们聊过。但他们是一夜之间在 Modal 上线的，我们事先根本不知道。就直接扩起来了，我们给了他们一大堆 GPU，全自动发生。这种魔法，我认为是人们真正看重 Modal 的地方：我们能在极短时间内给你大量 GPU，能处理那种波动。如果你用我们，你不用锁定一块矩形的三年期 GPU、不用精确知道数量。按用付费，随扩随缩，按 GPU 秒计费。「那你们怎么做 GPU 采购的需求预测？因为你肯定不想账上趴着疯狂过剩的供给。」完全正确。GPU 交期现在大约是三到四个月。我们基本做直线或指数外推，看全体客户的聚合——我们的客户构成很多元，你其实可以相当好地看聚合需求，按月增长率来预测，过去十二个月月环比大约 30% 到 40%。把它外推三到四个月，你就得到两到三倍。你出去买那些容量，确保爬坡到位。「有时候能直接买到那种产能吗？」能，可以。我觉得存在一点误解，以为我们没卡了。英伟达每月出货的 GPU 数量疯狂——不知道每个月装了多少十万块级别的 Blackwell（转写作 black walls）。价格肯定涨了，要拿到同样数量的 GPU，你得付更多钱。

[1200s - 1320s] To get the same amount of GPUs, and I don’t want to say it’s like an efficient market, but if you’re willing to pay more than other people, you can get a lot of GPUs. Is there a price at which it doesn’t make sense for you? Of course there is, but at what point? I don’t think so. I think in fact, like we’re actually more long than we’re short GPUs because we, you know, we we can charge a higher price to our customers. Like there’s effectively a market price, and you know, it’s like different dynamics because we lock up longer-term reservations and rent them out on slightly shorter duration. But you know, if if the underlying price goes up, people are also willing to pay more. Totally get you. What happens to the price over time? Yeah, H100s were meant to depreciate and go down. Idea. Yeah, I mean, and like, I mean, actually, like it’s funny. Well, dude, if you have no idea, what am I meant to have? Did you see Jensen yesterday? He was in London actually. Someone asked him this question, and he’s like, “Well, obviously, like demand is going to catch up with supply, or sorry, supply is going to catch up with demand eventually, and then we’ll see what happens with pricing. It might happen in a year. It might happen in five years. He said exactly the same thing I always tell people: like hardware prices, like it’s like gravity. Hardware prices always go down in the very long run. Right now, we’re in a very, you know, there’s a supply shortage, and you know that means that the prices have gone up. They may continue to go up for another year, two years, five years. I don’t know, but eventually it will come down. How early are we? Because like you know, I run Instinct for for everything. I have over a hundred automations on Instinct doing things that are just absolutely wild, from buying concert tickets to planning shows to helping with like medication for my mother. The amount of compute I am using is Jevons’ paradox times a hundred. Yeah. If that is equivalent for actually many people in society. Yeah. Because it’s accessible within WhatsApp and what. How early are we though? I think we’re very early. I think you’re saying the same thing. I mean, yeah, Jevons’ paradox, whatever. Like, I think there’s just so much latent demand for these things, and you know, I think you know, eventually we’re going to start to run up against these capital constraints, right? Like, there’s interest rates are going up. Like, we just can’t build data.

【译】 要拿到同样数量的 GPU——我不想说这是个有效市场，但只要你愿意比别人付更多，你就能拿到很多 GPU。「有没有某个价格高到让你觉得不划算？」当然有。「在哪个点？」我觉得没有。实际上，我们在 GPU 上更多是多头而不是空头，因为我们可以向客户收更高的价格——存在一个市场价格，动态是这样：我们锁定更长期的预订，以略短的时长租出去。如果底层价格上涨，人们也愿意付更多。「完全明白。价格随时间会怎样？H100 本来是该贬值下跌的。」「是啊，我——有意思。老兄，你都没数，我能有什么数？你昨天看 Jensen 了吗？」他其实在伦敦。有人问了他这个问题，他说：「显然，供给终将赶上需求——抱歉，供给终会赶上需求，到那时再看价格怎么走。可能一年，可能五年。」他说的和我一直告诉大家的完全一样：硬件价格像重力，超长期看总是下降。现在我们处于非常严重的供给短缺，这意味着价格涨上去了。可能还会再涨一年、两年、五年，我不知道，但最终会降下来。「那我们到底有多早期？因为我什么都用 Instinct，我在 Instinct 上跑着一百多个自动化，干的事都疯得不行：买演唱会门票、规划节目、帮我母亲管理用药。我消耗的算力是杰文斯悖论（转写作 Javon’s paradox）的 100 倍。如果社会上很多人也等价地这样用——因为它在 WhatsApp 里就能用——那我们到底有多早期？」我觉得非常早期。你在说和我一样的话。「是啊，杰文斯什么的。」我认为这些东西的潜在需求太大了。而最终，我们会开始撞上资本约束：利率在上升，我们没法把数据中心建得够快。也许这是好事，我不知道——这算是对过度建设的一种保险。但我也不知道我们会不会过度建设这些东西，因为需求实在太大了。「对，我认为未来五年你不可能过度建设。」大概不会。不，不会。我认为可能发生的是人们意识到——Instinct 是个好例子：据称他们还不收费、还没有收入，所以是个好例子，但 Instinct 显然未来能赚大钱。市场上有大量公司还没开始赚钱，却在用大量 GPU。他们最终必须开始赚钱，而那个滞后可能要一阵子，我们可能会撞上那个问题。

[1320s - 1440s] It’s fast enough. Maybe that’s good. I don’t know. Like I think in in a way, like you know, it’s a guarantee against like overbuild. Which, but I also don’t know if we can overbuild these things because there’s just so much demand for it. Yeah, yeah. I don’t think you can overbuild in the next five years. Probably not. No, no. I think I think what could happen is like people realize, you know, Instinct is a good example. Like allegedly, they don’t you know charge or they don’t make revenue yet. So I think it’s a good example. But like clearly, Instinct can make a ton of money in the future. But so you know, that’s a dynamic in the market. There’s a lot of companies that are not yet making money yet, and they’re using a lot of GPUs. They just have to eventually start to make money. But like that that lag could take a while, right? And and that we might run up against that. You mentioned the specialization that comes within the inference market, and we we spoke about Nvidia there. We’re seeing more and more specialization within the chip market. Yeah. Is Nvidia’s business today stronger than ever, or actually weaker than ever? As you see the fragmentation of the market into very specialized players from Cerebras to Etched to Olex to Fractal that weren’t there before. I I think you’re trying to phrase it as like one versus the other. My my view is like I think Erik Vishio is actually it’s funny like same thing like all of these things are going to grow. Like I think Nvidia is going to keep growing. They’re going to keep shipping amazing chips. I think there’s also going to be like Cerebras and all these Etched and all these other ones like shipping specialized chips, and they’re also going to grow. So I think there’s going to be markets for all of it. So why Cerebras down thirty percent? I don’t know. I don’t follow really the stock market. You don’t follow the stock market? No, I’m in like a very passive. You know, buy the index mostly because of my own time. Like I don’t feel like I have the energy to follow the stock market. To what extent is CUDA the moat that everyone says it is for Nvidia? I actually I think it’s a little overstated. I don’t think CUDA itself is the moat. CUDA is actually kind of annoying to work with. It’s like a pretty low-level programming model. But I do think all the Nvidia the ecosystem is like I think you know all the kernel optimizations and and all the.

【译】 你提到推理市场里的专门化，我们也聊到了英伟达。我们看到芯片市场里越来越多的专门化。「英伟达今天的生意是史上更强还是史上更弱？」随着市场碎裂成非常专门的玩家——从 Cerebras 到 Etched（转写作 Etch）到 Olex 到 Fractal——这些以前都不存在。我觉得你想把它表述成一个对另一个。我的看法是——我认为 Eric Vishio（转写疑作，人名未确认）说得对，有趣的是：所有这些东西都会增长。英伟达会继续增长，继续出货惊艳的芯片；同时也会有 Cerebras、Etched 和所有这些其他玩家出货专门芯片，他们也会增长。所以我认为所有这些东西都会有市场。「那为什么 Cerebras 跌了 30%？」我不知道，我真的不太跟股市。「你不跟股市？」不跟，我非常被动，基本买指数，主要是没那个时间精力。「CUDA 在多大程度上是大家说的英伟达护城河？」我其实觉得它有点被高估了。我认为 CUDA 本身不是护城河。CUDA 用起来其实挺烦人的，是个相当低层的编程模型。但我确实认为英伟达的整个生态——所有的 kernel 优化，以及所有这些——

[1440s - 1560s] Stuff that you know, every every like Torch always supports Nvidia, you know, much better than than anything else, and and and the open source engines like vLLM and SGLang because of that they always supported much better. So like everything is Nvidia first today, and that’s probably going to continue. It’s like funny for the last like fifteen years I’ve been saying, you know, like next year is the you know it’s the year of non Nvidia accelerators. It never keeps happening. I’m always wrong. Nvidia has just you know shown to to be incredibly good at having that staying power and just dominating the market. We’ll see. I mean, there’s like Trainium, there’s TPUs, there’s AMD, there’s a lot of other stuff that does look quite good and they’re catching up, but you know I still think they’re quite behind. To what extent are the neo clouds a friend because they’re partnered with you in a lot of cases? I’m sure as you mentioned, I think 25 you said, or a foe? You know, I had the founder of Nebius on, and he was like, we’re coming for the inference market. Like we need to own full stack, and I don’t mean that against him. I totally understand it, but are they friends or are they they foes? I feel like everyone in startup land is like a friend and a foe at the same time. They have this like weird sort of you know relationship with everyone. I I think some of the top neo clouds, Nebius included, they are trying to move up the stack. I’m personally somewhat skeptical. I think it’s like very hard to be good at like your core business and also trying to move into different market at the same time. Why would it be hard for them? Like, what’s the bare case on them being able to move into the inference layer efficiently? I mean, I’m biased, but like we spent the last five years building you know an incredibly complex infrastructure stack. I think it’s very very hard to replicate what we built. What’s the hardest thing? Just all the code and you know everything around the file system, the container runtime, the scheduler, like building the system that can you know place like millions of sandboxes every minute. You know, scale up and down many many thousands of GPUs. You know, power hundreds of thousands of tasks running concurrently. These are these are very hard engineering problems. Like I mean, I’m I’m an infra geek. Like I love that shit. It’s so much fun to work on it. But like these are hard problems.

【译】 ——比如 PyTorch 对英伟达的支持总是远好于其他任何硬件，而开源引擎 vLLM（转写作 VLM）和 SGLang 也因此对英伟达的支持好得多。所以今天一切都是英伟达优先，这可能会持续下去。有意思的是，过去十五年我一直在说「明年是非英伟达加速器之年」，从来没兑现过，我一直错。英伟达就是展现了那种惊人的立足能力，统治着市场。再看看吧——有 Trainium（转写作 Tranium）、TPU、AMD，很多别的东西看起来相当不错，他们在追赶，但我仍然认为落后相当多。「那 neo-cloud 算朋友还是敌人？」他们很多情况下是你的合作方——你说了，我想是 25 家。我记得 Nebius 的创始人上过节目，他说「我们要拿下推理市场，我们要全栈自建」——我不是针对他，我完全理解，但他们到底是友是敌？创业圈里的每个人，感觉同时是朋友又是对手，跟所有人都是这种奇怪的关系。一些头部 neo-cloud，包括 Nebius 在内，正在尝试向技术栈上层走。我个人有点怀疑。我认为很难既把你的核心业务做好、同时又进入一个不同的市场。「为什么对他们来说很难？他们高效进入推理层的空头论点是什么？」我有偏见，但我们花了过去五年建一套极其复杂的基础设施栈。我认为要复制我们建的东西非常非常难。「最难的是什么？」所有那些代码，以及文件系统、容器运行时、调度器周围的一切——建一个能每分钟放置上百万个沙箱的系统，上下伸缩几千张 GPU，驱动几十万个并发运行的任务。这些都是非常难的工程问题。我是基础设施极客，我爱死这些东西了，做起来太好玩。但这些是真难的问题。

[1560s - 1680s] Is Nebius a better business than CoreWeave? I actually don’t know. Like, I don’t know. I think they’re both like they’ve executed extremely well, both of them. I have a tremendous respect for what they’ve done. I think at the at the end of the day, like what the the the thing that I have the most respect for is actually more in the financial engineering side. They’re very good at that kind of stuff. Why do you say that? Because I think a lot of the constraint for building a neo cloud very quickly is on the financial side. You need to be very good at you go lock in like a long term customer contract, then you go raise a bunch of money against it, then you go buy the chips, then you deliver the chips, and then you pay off the debt, like you know, over the next few years. Doing that sort of dance with the financials, like, is kind of an interesting game. And CoreWeave has done it exceptionally well. It’s a capital game. It’s a financial game. When you move a layer beneath them, we spoke about Jev, who obviously worked with Modal. You, we see incredible movement from the open source community, which has been exciting. Yeah. But then you also see Haiku and Anthropic’s newest models being incredibly performant and incredibly cheap. What’s the benefit of open models if frontier models are almost as cheap and as efficient? I think open source is always going to be. There’s always going to be great models coming out. Let’s also remember, like the open source the the open source models, they’re all moving into licensing models. So, like, they’re getting better at capturing the value of those models, and I think that means they’re going to have more funding to train better models. And I really hope there’s going to be European and US models because I don’t think it’s a good situation that we have so many Chinese models. I mean, they’re obviously I have tremendous respect for all the Chinese labs. They’re extremely good at what they’re doing, but I think it’s good for the world if there’s you know more power spread out. But but anyway, going back to the question, I do think that open source there was there’s always going to be people who want to fine tune, want to have more control over their models, and and sort of tweak the parameters like go low latency, high throughput, like whatever it is, and and open source is going to play a big role in that.

【译】 「Nebius 是比 CoreWeave 更好的生意吗？」我真不知道。我觉得他们两家都执行得极其出色，我对他们做成的事怀有巨大的敬意。归根结底，我最尊敬的其实更多在财务工程那一面。他们非常擅长这种东西。「为什么这么说？」因为快速建一个 neo-cloud 的很多约束在财务侧。你必须非常擅长这个游戏：你先锁定一个长期客户合同，然后据此融一大笔钱，然后去买芯片，然后交付芯片，然后在未来几年偿还债务。做这种财务上的舞蹈是个挺有意思的游戏，而 CoreWeave 把它玩得极其漂亮。这是一场资本游戏，一场财务游戏。当你往下移一层看他们。我们聊过 Jev（转写作 Java）——它显然和 Modal 合作过。我们看到开源社区令人兴奋的大动作。「但另一边，你也看到 Haiku 和 Anthropic 的新模型表现极强又极其便宜。如果前沿模型几乎一样便宜高效，开放模型的好处是什么？」我认为开源永远都会有好模型出来。还要记住：开源模型们都在转向授权收费（licensing）模式。所以他们越来越擅长捕获这些模型的价值，而我认为这意味着他们会有更多资金去训练更好的模型。我真心希望会有欧洲和美国的模型，因为我不认为中国模型那么多是个好局面。我对所有中国实验室当然怀有极大的敬意，他们做的事情极其出色，但我认为如果权力更分散，对世界更好。但回到问题本身：我确实认为开源——总会有人想微调、想更多控制自己的模型、调整参数——要低延迟、高吞吐，不管什么——而开源会在其中扮演重要角色。

[1680s - 1800s] Does it Windows versus Linux? With I’m sorry, with the greatest respect, do you think Europe will produce a competitive open source model? Yeah, I mean, I’m I love Europe. I grew up in Europe. I grew up in Sweden, but you know, I do think that like Europe does struggle with you know accumulating large amounts of capital. I think both of those things are things that Europe struggles with, and and that is something that you know China has done really well. They’ve accumulated massive amounts of human capitals, human capital, and and U.S. obviously has very you know very good at like raising enormous amounts of money and like training very large models. I I don’t know if Europe has the either of those. Does it matter if Nvidia owns the open source model layer in the U.S. with first Poolside and now reflection? I I think Nvidia is very smart to focus on the ecosystem, building up an ecosystem, and I think that includes like propping up a lot of open source models. And I think they probably saw an opportunity to pick up open source models. Like these these companies are very resource intensive, right? Very capex intensive, and and and I think it’s smart of Nvidia to like buy some of those and sponsor them effectively, in order to create like American built models. By the way, I want to take back like one thing I said. Europe does have an insane amount of human capital. I do think what Europe is very bad at is like aggregating them and putting all of them at the same company. That is something I think U.S. is very good at. And I think inspiring a mindset of ambition, yes, yeah, huge huge scale. I think to the point that you said earlier, which is like, oh, I didn’t have the belief in myself. I respectfully think if you’re in the U.S., you might have been surrounded by influences that told you to start a company earlier. Yeah, yeah. I mean, it’s almost like too much on the other side to that point. But oh, in all four, you’re encouraged to start. Yeah, exactly. Drop out. Kindergarten wasn’t worth it. Exactly. I totally agree with you. I do want to talk about you. We said about models and Jev in particular. Are we seeing like the unbundling of models on?

【译】 「这是 Windows 对 Linux 的局面吗？恕我直言——你觉得欧洲会造出一个有竞争力的开源模型吗？」是啊，我热爱欧洲，我在欧洲长大，在瑞典长大。但我确实认为欧洲在积累大额资本方面很挣扎。这两样东西欧洲都难，而中国做得非常好——他们积累了大量的人力资本；美国显然极擅长融巨额资金、训练超大模型。我不确定欧洲两样里有哪一样。「如果英伟达通过先 Poolside（转写作 pool side）后 Reflection，拥有美国的开源模型层，这要紧吗？」我认为英伟达聚焦生态、建起一个生态非常聪明，而这包括扶持很多开源模型。我认为他们可能看到了捡起开源模型的机会：这些公司资源密集、资本开支密集，英伟达买下其中一些、实际上赞助它们，来打造「美国制造」的模型，我觉得很聪明。顺便，我想收回刚才说的一句话：欧洲有惊人数量的人力资本。我确实认为欧洲非常不擅长的是把这些人聚合起来、放进同一家公司——

[1800s - 1920s] Like per function basis, where Jev is obviously very good for its certain tasks, I think so, and I think that’s a function of what I talked about earlier. Where like the market for inference is so large, and and that you know when a market is like you know billion dollars, like you’re not going to have too many sub markets, but when a market is starting to get ten billion, hundred billion, whatever, like you’re going to have a lot of different sub markets, and and Jev does classification and things like that, regression type things, very very well, and and there’s large enough market for that. Okay, welcome. It’s your first day as a partner at Twenty VC Fund, and we’re in the investment committee, and very kindly we have the opportunity to invest all in Jev. We’re all in on Jev’s new all in all in. What’s the bull case for Jev being worth a hundred billion? I told you before, but like I think the brand and the community and the movement around you know this just this sort of everything they’re doing is is so incredible, and that to me is like a massive massive asset in this new age. That’s a hundred billion dollar company. I think there is an opportunity to create a hundred billion dollar company with a lot of specialized models in inference. You would have done it though at eight point five billion. I would have done it eight point five billion. I think there’s a pretty big outcome here. Could be an exit. Could be you know like a public company. Who knows? But but I do think there’s a lot of upside here. Dude, I’d love to see him ringing the bell on the New York Stock Exchange. Yeah, do you know what I mean? Yeah, fantastic guy. Love that guy. Is he great? I mean, you know, his charisma, his like, you have to create like a personality like around your brand. You have to create this like visual identity. It’s obviously not the visual identity I would have picked. But I think I think then I hear you. But I think then you should be more loud. If I’m just honest, Erik. Modal. No, we yeah, but you. Yeah, I mean that’s possible, right? Like I I don’t think there’s like a personality lad cult from modal from your perspective. I think that’s a fair point of criticism. I think I could be out there even more. I am tweeting. I try to tweet. I try blogging. I’ve been kind of bad at that recently. No one gives a shit about blogging. Yeah, maybe right. Maybe I just need to like go more unhinged on Twitter. It’s like fan related marketing, basically. Like in the early days of modal, like that was a big part of our growth, right? Like me.

【译】 ——这是美国非常擅长的事。还有激发一种野心心态。「对。」说到你之前说的「我对自己没信心」——恕我直言，如果你在美国，你周围的影响可能会让你更早创业。「对，而在另一头几乎又过头了。」「在硅谷，人人被鼓励创业。」「对，辍学，幼儿园都不值得上。」完全同意，我百分百同意。我想聊聊你，聊聊模型，特别是 Jev。我们是不是在看到模型按功能、按单一用途被非捆绑？Jev 显然在某些特定任务上非常好。我认为是，而这是我之前讲过的那件事的函数：推理市场太大了。当一个市场是十亿美元时，你不会有太多子市场；但当市场开始到百亿、千亿美元，你会有很多不同的子市场。而 Jev 把分类、还有回归类的事情做得非常好、非常好，而那有足够大的市场。好，欢迎，这是你在 20VC 基金当合伙人的第一天，我们坐在投委会里，而我们非常幸运有机会投资——全部押注 Jev。我们全押 Jev 的新……全押、全押。「Jev 值一千亿美元的看多论点是什么？」我之前就跟你说过：品牌、社区、围绕这一切的运动，在我看来是这个新时代里巨大的、巨大的资产。那是一家千亿美元公司。「专门化模型加推理，有机会造出千亿美元公司。」「但你在 85 亿美元的时候就会投了。」「我 85 亿就会投。」我认为这里有相当大的结局：可能是退出，可能是一家上市公司，谁知道呢。但确实有很大上行空间。老兄，我很想看到他去纽交所敲钟。是啊，你懂我意思吧。太棒的家伙了，我爱那个人。「他很棒对吧？」他的人格魅力……你必须围绕品牌塑造一个人格，必须创造这种视觉识别。那显然不是我会选的视觉识别，但我听到了。「那我觉得你应该更大声，老实说，Erik。Modal。」「不，我们……对，但你们……」是，有可能。我不觉得从你的视角看 Modal 有「人格领袖崇拜」。「我觉得这是个公正的批评。」我可以更抛头露面一点。我在发推，试着写博客，最近有点懈怠。「没人在乎博客了。」是啊，也许吧。也许我该在推特上更放飞（unhinged）一点。这基本就是创始人营销。在 Modal 早期，这是我们增长很大的一部分，比如我——

[1920s - 2040s] Just like tweeting all the time about what we’re building, and I think that’s really important. If I mean, every every company has to find its own sort of you know go to market strategy. But in the early days, like we spent a lot of time just talking to people on Twitter, just like hitting people up, try the product. What do you think? Half of them ended up angel investing. You know, they’re now supporters of Modal, and and you know, you have to lean into what you’re good at. And in the early days, I did a lot of shit posting. Do you know, for every follower that followed me before fifty thousand, I would send them a personal like thank you message. That’s amazing. You typed it out yourself every night from one till two thirty a.m. That’s amazing. And I’d say like, I see that you’re from Sweden. I love the Grand Hotel. Let me know when you’re in London, and I would send like fifty to a hundred a night. That’s dedication. That’s amazing. No, but it was the customer acquisition flywheel. Yeah, like they would go into the office and they’d stand, be like, “Wow, you know this guy. He’d sent me a message, and like, you should listen to a show. Yeah, it’s smart. I love what you’ve been doing in terms of building up this like media brand. It’s it’s fantastic. Neil Mater has a great thing, which is like, how can I tenx the customer experience? And actually, like, let’s let’s do it. Fuck it, let’s play a game. How do I tenx the current Modal customer experience? I think the biggest thing we need is we need way more products in in Modal. Like, we want to create a basically comprehensive experience of anything you need to do in order to build an AI application should be a product inside Modal. The other thing we can do is hire way more FDEs and way more sales reps. I will say we’ve been behind on that, but we should hire like way more people that reach out to you personally and say, “Hey, I saw you get started in Modal. I saw you’re working in the bio space. How can we help you? Here’s some bio examples. You know, protein folding, whatever. Right? Like, we work with a bunch of other similar companies. You seem skeptical. Well, no, I don’t at all. I’m actually just thinking about like, I like the product ambition, but I go as a European who goes, “Wow, we’re in the B two B space, so I think it’s like you know, but I do worry about being stretched too thin. Yeah, to your point on like you know, why can’t Nebius do what you do so efficiently because you’re so

【译】 ——就是不停地发推讲我们在建什么，我认为这真的很重要。每家公司都必须找到自己的 go-to-market 策略。但在早期，我们花了大量时间就是在 Twitter 上跟人聊、私信他们：「试试产品，你觉得怎么样？」其中一半人后来成了天使投资人，现在是 Modal 的支持者。你必须发挥自己的长处。早期我发了大量搞怪帖（shitposting）。还有——在粉丝数到五万之前，每一个关注我的人，我都会发一条亲手写的感谢私信。「太惊人了。你每天晚上自己打字，凌晨一点到两点半？太惊人了。」我还会写：「我看到你来自瑞典，我爱 Grand Hotel，来伦敦时告诉我。」我每晚发 50 到 100 条。「这是奉献精神，太惊人了。」不，这就是获客飞轮。他们会走进办公室说：「哇，你认识这人吗？他给我发了条私信，你该听听他的节目。」很聪明。我喜欢你在打造这个媒体品牌上做的事，非常棒。Neil Mater（转写疑作）有个很好的说法：「我怎么把客户体验乘以十？」那来吧——管他呢，我们玩个游戏：怎么把 Modal 当前的客户体验乘以十？我认为最大的事情是：Modal 需要多得多的产品。我们想创造一种全面的体验——你建 AI 应用所需做的一切，都应该是 Modal 里的一个产品。另一件我们能做的是雇多得多的 FDE（转写作 FDS）和多得多的销售。我要说我们在这方面落后了，但我们该雇更多人主动联系客户：「嘿，我看到你在 Modal 上起步了，看到你在做生物领域。我们能怎么帮你？这里有些生物方面的例子——蛋白质折叠什么的。我们服务很多类似的公司。」「你好像有点怀疑。」「不，完全没有。我其实只是在想——我喜欢这个产品野心，但作为一个欧洲人我会想：『哇，我们在 ToB 领域』……但我确实担心被摊得太薄。」「对，呼应你说的——为什么 Nebius（转写作 Nabius）做不到你们这么高效，因为你们这么——」

[2040s - 2160s] So good, and you spent so much time. As you expand into all the specialities, there are players in each of them who’ve been doing that for years. Where I’m like, I don’t think that’s true. I don’t think they’re very good. A lot of them, but but I mean, I sort of agree with your higher level premise. Like, if there’s anything that’s like super critical and never goes away, it’s it’s prioritization and resource allocation, right? Like, you can’t boil the water and try to do everything at the same time. You need to be extremely ruthless. And and having a multi product strategy, I think, is a very hard balance. Like, when do you branch out into a second product or a third product? Right? Like, it took us four years to see success with a second product, which was sandboxes. Four years to see success with your second product. So we we launched sandboxes five years to a billion in revenue. Yeah, so we we launched sandboxes in 2023. Didn’t see a lot of success in the early days. I mean, there was like some success, you know, it wasn’t horrible, but you know, it was a relatively insignificant part of our revenue up until summer of 2025, when models, I think, like Sonnet 4.5. One of those, like, they just got really good at generating code and building whole apps, and then it’s it just exploded. I mean, sandboxes today is a 300 million run rate business. It’s about a third of our revenue, and and so that’s been kind of a crazy journey. And now we’re thinking about you know what’s the third and the fourth. We just launched a training product. I’m super excited about that. Working on batch products. Working on reinforcement learning. We’re working on all kinds of other stuff. And so that evolution into becoming a platform, it’s actually there. Like if you go back to the very early decks of modal, you know, I put together some just for my own, you know, thinking. It was all there. It’s like we we had this like market map and all these, you know, it ended up being very different products than we thought. But it’s so important for me to create like you know to focus on a category and focus on a set of solutions, not just like be a single single widget vendor. What do you know about the sandboxes market today that you wish you’d known when you started and released the first product? It it actually is like one of those products. I think one of the benefits of being an infrastructure, but also one of the drawbacks in a weird way, is like.

【译】 ——强，而且你们投入了这么多时间。当你扩张进所有这些专业领域时，每个领域里都有做了很多年的玩家。嗯，我不觉得那成立。他们中很多并不怎么样。但我大体上同意你更高层的前提：如果有什么东西是超级关键、永远存在的，那就是优先级排序和资源分配。你不能把所有水同时烧开、试图什么都做。你必须极端无情。而多产品策略是一个非常难的平衡：你什么时候分支去做第二个、第三个产品？我们第二个产品（沙箱）用了四年才看到成功。「四年才看到第二个产品的成功。」「我们发布沙箱……五年到十亿收入。」对，我们 2023 年上线了沙箱，早期没看到太大成功——有一些成功，不算差，但直到 2025 年夏天之前，它在收入里都是相对不起眼的部分。那时模型——我想是 Sonnet 4.5（转写作 sawn 4.5）那批——突然在生成代码、构建整个应用上变得非常强，然后就爆炸了。沙箱今天是一个 3 亿美元 run rate 的业务，约占我们收入的三分之一。这一直是段相当疯狂的旅程。现在我们在想第三个、第四个产品是什么。我们刚上线了一个训练产品，我超级兴奋。在做批量产品。在做强化学习。在做各种其他的东西。而这种演化成平台的事——其实一直就在那里。如果你回看 Modal 最早期的 deck——我自己拼的那些，只是为了自己思考——全都在里面：我们有那张市场地图和所有这些东西，最终的产品形态和想的不一样，但对我来说非常重要的是：聚焦一个品类、聚焦一组解决方案，而不是只做一个单一零件的供应商。「关于今天的沙箱市场，有什么是你希望自己在发布第一个产品时就知道的？」它其实是那种产品之一。作为基础设施的好处之一，同时也是某种意义上的坏处，就是——

[2160s - 2280s] You kind of just have to trust your gut on like what are like important primitives, and just build like amazing building blocks. I think we’re in the space of like providing building blocks. We try to look at like what what do engineers need, and and not like listen too carefully because like they’ll ask for like all kinds of random stuff. But kind of distill it down into like what are some of the building blocks. And we put out a serverless function product in in 2021. It saw success in 2023 with with Gen AI inference. We put out a sandbox product in 2023 because we saw a lot of people using LLMs to generate code. And then I saw success in in 2025, like I said. So a lot of like how we think about building products has always been kind of like let’s just put out like good building blocks, and then it exploded. And and but like to to answer your question, it was reinforcement learning, it was vibe coding, background agents, and maybe some other stuff too in the future. And those were things we actually didn’t anticipate. So so I think I think I think infrastructure is often like that. Like we didn’t anticipate Stable Diffusion when we started building our serverless function product. It just happened to be the killer app. What’s your takeaway from that? The only thing is that matters is you’ve got to stay in the game long enough for your timing to be right. Sometimes I think that’s right. I mean, we we definitely been extremely lucky in certain ways with timing. Like we put out products like right before we started the company 2021, and then you know Stable Diffusion and ChatGPT came out 2022. We had really good timing with that. Right, just created so much demand, and we’re very well positioned. But some of it is also like with infrastructure. Like again, like I think it’s about like building products that you know you anticipate will be useful in the future, and and from your own experience, are things that you always wanted to see. I think you have to kind of trust your gut in that. Dude, you said that about resource allocation and that being very very important as you think about this breadth of products. Where did you allocate resources? Where with the benefit of hindsight, you wish you hadn’t, and how do you reflect on that? I think we should have allocated more resources towards LLM inference like earlier. I think we’re a little bit on on the the sidelines for a while. We didn’t.

【译】 你其实只能凭直觉去判断什么是重要的原语，然后把积木做得惊艳。我们做的就是提供积木的生意。我们试着看工程师需要什么，但又不能听得太认真，因为他们会要各种随机的东西——而是把它蒸馏成：什么是那几个基本积木。我们在 2021 年发布了 serverless 函数产品，它在 2023 年随着 GenAI 推理见到成功。我们在 2023 年发布了沙箱产品，因为我们看到很多人在用 LLM 生成代码。然后在 2025 年见到成功，像我说的。所以我们建产品的思路一直是：把好的积木放出去，然后它爆了。而回答你的问题——是强化学习、vibe coding、后台 agent，未来也许还有别的——这些是我们当时没预料到的。所以我认为基础设施经常就是这样的。我们做 serverless 函数产品时也没预料到 Stable Diffusion，它恰好成了那个杀手级应用。「你从这中间的结论是什么？」唯一要紧的是：你得在游戏里待得够久，让时机来对你。有时候我觉得是这样。我们确实在某些方面运气极好：我们在 2021 年、公司刚起步前后发布了这些产品，然后 2022 年 Stable Diffusion 和 ChatGPT（转写作 Chat GPT）出来了。我们的时机真的很好，这创造了巨量需求，而我们占位极好。但另一部分也是基础设施本身的特性：建那些你预判未来会有用、而且从你自身经验出发一直想看到的产品。我觉得你必须相信自己的直觉。

[2280s - 2400s] Realized how big open source inference would be, and we thought it would be kind of a race to the bottom, just a commodity space. And then DeepSeek came out, and we’re like, “Fuck! Like, we got to move very fast.” And so we staffed up a team immediately. You know, went from like you know one or two people working on this to like thirty people twelve months later, and and now we’re one of the top three providers. Jason Lemkin, dear friend of mine, who I do a show with every Thursday, he always says that now we’ve reached peak. Actually, open source LLM inference usage, we’re going to see frontier providers bluntly provide incredibly cheap products that they are now. U.S. enterprises don’t want to use Chinese models. That will continue, and so actually, with those two core prevailing winds, frontier will actually research, and we will see a reduction slash plateauing of the OS LLM inference model. I think that’s a little bit of a fixed pie size mentality. Like, I, I think you know, with with all due respect, I do think both spaces will grow five x, ten x, twenty x over the next twelve months and and beyond. Like, and so like it’s hard to argue like you know which one is going to be bigger than the other, but I think both will grow like in you know orders of magnitude over the next few years. Five x ten x twenty x over the next twelve months. Yeah, for sure. You just hit a billion in revenue. Yeah. Well done. Yeah. Does that feel pretty cool? It feels pretty cool. Yeah. I mean, I, I, it’s you know, like I, I don’t know, like I’m just like you know Swedish guy that you know grew up in a random. For me, it’s like these numbers are absurd, but like, yeah, like it feels great. It’s so much money, Erik. You’re so Joe. I wasn’t expecting you to be so like likable and cool. I don’t know. Like, it’s like a lot of zeros. Like, how many zeros are in a billion? It’s it’s a lot. Your wealth manager will tell you when you look at that. Where will you be in twelve months? I don’t know. What I can say is, we’ve been growing. We grew fifteen x in the last twelve months, so you can extrapolate that. I mean, I don’t know. Like at some point, we start. You know, like markets just get very big and they start to slow down a little bit, but.

【译】 「老兄，你说了资源分配、说它在你思考产品宽度时非常重要。你把资源分配到了哪里？事后看，你希望当初没分配到哪里？」我认为我们应该更早地把更多资源投向 LLM 推理。我们有一阵子是在场边观望的。我们没意识到开源推理会有这么大，我们当时以为那会是逐底竞争、纯商品化的空间。然后 DeepSeek 出来了，我们整个人都：「卧槽，我们得动得非常快。」所以我们立刻组了团队，从只有一两个人做这个，到十二个月后大约三十人，而现在我们是前三名的提供商之一。Jason Lemkin（转写作 Jason Lamkin），我的挚友，我每周四和他一起做节目，他总说现在我们已经见顶了——实际上他说的是：开源 LLM 推理的使用量已达峰值，我们会看到前沿厂商 bluntly 地提供他们现在这种极其便宜的产品；美国企业不想用中国模型，这会持续。所以在这两股主导风向下，前沿会真的胜出，我们会看到开源 LLM 推理的下滑或者说平台化。我觉得这有点「固定饼大小」的心态。恕我直言，我确实认为未来十二个月以及之后，两个空间都会增长 5 倍、10 倍、20 倍。所以很难争论谁会比谁更大，但我认为未来几年两者都会以数量级增长。「未来十二个月 5 倍 10 倍 20 倍。」「对，肯定的。」「你刚到 10 亿收入。干得好。感觉爽吗？」感觉挺爽的。是啊，我……我不知道，我只是一个在随机地方长大的瑞典人，对我来说这些数字荒谬得很。但感觉确实很棒。「这可是好多钱，Erik。你太……我没料到你这么讨人喜欢、这么酷。」我不知道，就是好多零。「十亿里有几个零来着？」很多。「你的财富管理师看报表时会告诉你的。十二个月后你们会在哪？」我不知道。我能说的是：我们一直在增长，过去十二个月增长了 15 倍，你可以据此外推。我不知道——到某个时候，市场会变得非常大、开始放缓一点。但——

[2400s - 2520s] I with with the growth rate right now, like the growth rates are not slowing down. Like we’re growing thirty, forty percent every month. Four billion. I mean, I definitely think we should. I hope to be way beyond that. With the greatest of respect, fifteen billion then feels like quite a cheap price. Like if I’m an ambassador, so this is now my ballpark. Like if if you think you’re going to be at four billion, say dropping, I don’t know, a billion two to the bottom line, yeah, then they actually only paid what fourteen x end of year next year bottom line. Three x top line. Yeah, it’s kind of cheap. That feels pretty cheap. I think it’s cheap. I mean, like I wouldn’t do this if I didn’t think we could grow and become like a generational company. I think there’s so much upside. I think this is one of the largest markets opportunities that’s ever existed. I think that the demand for compute, the demand for AI applications, it’s just gonna it’s gonna be enormous, and they’re all gonna need a place to run. And I hope a lot of them are gonna run on Modal. What revenue line do you not have today, or is tiny that you think will be very significant in three years’ time, five years’ time? That is a good question. I think training is is something we just launched. I think that could be very large. I hope that could get to 100 million run rate in the next six to 12 months. Post training, similarly, working on batch jobs. I’m very excited also about a lot of like sort of traditionals like data warehousing stuff, databases, workflow orchestration, feature stores like vector database. I mean, there’s just so many parts of this, and they all kind of fit together very nicely. And that that to me is like what I’m so excited about. It’s like once you start building in this space, there’s so many natural adjacencies. Like this is like a problem with me. It’s like I get excited about. So I have like no lack of ideas. Like at modal, like we have so many ideas of things we want to go. CI, I would love to get into continuous integration. That’s like a very interesting space. Agent harnesses. Maybe we should train models. Maybe you know, this is five thousand directions. You know, and and that’s like my ADD brain. Like wants to run in all of those directions. You have to be obviously ruthless with prioritization and only take on.

【译】 ——以现在的增速，增长率并没有慢下来。我们每月增长 30%、40%。「四十亿？」我的确认为我们应该……我希望远超那个数字。「恕我直言，那 150 亿的估值就显得相当便宜了。让我来 ballpark 一下：如果你认为你们会到 40 亿收入，假设落 12 亿到净利上，那他们实际上只付了明年年底利润的 14 倍、收入的 3 倍。这感觉太便宜了。」「我认为是便宜的。我做这件事的前提，就是我们能成长成一家 generational company。上行空间太大了。我认为这是有史以来最大的市场机会之一：对算力的需求、对 AI 应用的需求，会大到吓人，而它们全都需要一个地方去跑。我希望其中很多都跑在 Modal 上。」你今天没有、或者很小、但认为三五年后会非常显著的收入线是什么？好问题。训练是我们刚上线的，我认为它可以非常大，我希望它在未来六到十二个月能到 1 亿美元 run rate。post-training 类似，我们在做批量任务。我还对很多传统东西感到兴奋，比如数据仓库、数据库、工作流编排、特征存储、向量数据库。这个领域有这么多部分，而它们全都咬合得很好。这正是我兴奋的地方：一旦你开始在这个空间里建设，就有这么多自然的相邻扩张。这其实也是我的一个毛病，我对什么都兴奋：CI——我很想进持续集成，非常有意思的空间；agent harness；也许我们该训练模型——五千个方向。这就是我的 ADD 大脑，想冲进所有方向。你显然必须无情地排优先级、一次只——

[2520s - 2640s] One or two things at the right time, but but but but there’s so many different directions we can go. It’s amazing that your wife calls you for decisions. She must get a very incoherent answer. I know, I know. What is this? Is this? I just wanted some curtains, darling. Yeah, you know, with the curtains, I just like you pick one of them. We’ve seen Simpsons and like Rainer Wolfcastle becomes the president, and they’re like, which one do you want to pick? You know, one to seven. He’s like, I was born to lead, not to read. Number four. Yeah, I I totally get that. Can we’ve seen this real shift? I think in in emphasis, but also in resource allocation from pre-training to post-training. How do you feel about that? And do you think that is the new normal? I I I don’t know. I I think post-training works because there is a lot of tasks that people do in real life that have fairly finite sort of feedback mechanisms. So coding is like a great example, right? Like you can do coding with reinforcement learning pretty well because you know you can ever you can run code in sandboxes, and a lot of labs are running modal sandboxes for that exact purpose. I start to look at cybersecurity is another area. Math, obviously, with lean, I mean, we’ve seen some fantastic. I think it’s kind of incredible all these like OpenAI proofs the other day. So law, maybe to some extent, you can you know do like eval. So I I think I think post-training right now like there’s like such an incredible focus on it because you you you have all these tasks that traditionally only humans can do, but are actually well within reach for for agents and and for LMs. But but but like I I don’t know if it extends beyond that. Like I don’t know. Like there’s a lot of tasks out there that like have way longer feedback loops that I don’t know if like post-training quite yet are is within reach. A lot of people ask me like, what’s going to happen with software engineers in the long term? And I think you know we’re probably going to automate everything that that an agent can do within 24 hours or a week or something like that. But I kind of feel like you know engineers like they they work on a longer timescale and they have feedback loops that are longer, so.

【译】 ——在正确的时间做一两件事。但我们可以去的方向太多了。「有意思，你妻子打电话让你做决定时，一定得到一个非常不连贯的回答。」「我知道，我知道。」「这是什么？窗帘的事？」「我就是想要个窗帘，亲爱的。对，窗帘的事，你就挑一个就行了。」我们看过《辛普森》，雷纳·沃尔夫卡斯尔（Rainer Wolfcastle，转写作 Raymond）当了总统，他们说：你想选几号？一到七。他说：「我生来是领导人的，不是来阅读的。四号。」我完全懂。我们能聊聊这个真实的转移吗——无论是重视程度还是资源分配，从预训练转向后训练？你怎么看？你觉得这是新常态吗？我不知道。我认为 post-training 之所以有效，是因为人们在现实生活里做的很多任务有相当有限的反馈机制。写代码就是绝佳例子：你可以用强化学习把编码做得很好，因为你可以在沙箱里运行代码，而很多实验室正是为了这个目的在跑 Modal 的沙箱。网络安全是我开始关注的另一个领域。数学显然有 Lean——我们前几天看到 OpenAI 那些惊人的形式化证明。法律，某种程度上可以靠 eval。所以我认为 post-training 现在有这种难以置信的聚焦度，因为这些传统上只有人类能做的任务，实际上已经在 agent 和 LLM 的能力范围之内。但它是否延伸到那之外，我不知道。外面有很多任务的反馈回路要长得多，我不确定 post-training 是否已经够得着。

[2640s - 2760s] I have a feeling like reinforcement learning. We’re going to start to run into some limitations around that. But I’ve been wrong for so long, so maybe I’ll be wrong and just keeps going. And eventually, we can do reinforcement learning on things that take six months for a human. I don’t know. What does the day look like in three years for an engineer? I mean, I’m extremely bullish on software engineering. I’ve been coding for thirty years, and I’ve seen so many different changes with like programming languages and the cloud, and you know, so many different frameworks. And it’s you know, every single time we make programmers more efficient, it just increases demand for software. There’s so many parts of the world that desperately need software, and so I’m very, very bullish on software and the need for software engineering. And I think one of the biggest misconceptions from the outside is that developer is about like translating from English to code. Like that has never been in you know the job of software engineers. It’s like to understand the requirements in the first place and be kind of an interface between the humans you’re building for humans at the end of the day and like trying to figure out like what does that even mean in terms of processes and workflows and and and that continuously moves up in the stack. You know, back in the sixties, we used to write assembly and punch cards, and then you moved to COBOL, and then to C plus plus, and then to Python, and and now we’re using agents. And that’s just like one and like it’s definitely like the biggest thing that I’ve seen in the last thirty years. But I do think we have many, many analogs and many parallels throughout the the history of of software and computers. To me, that makes me feel like you know, software engineers are going to be fine for the next ten, twenty years. How much money do you spend on coding agents a month at Modal? I think a half a million or something like that. How many engineers? One hundred and fifty. Dude, that’s so that’s so worryingly low. You want to should be token maxing more. There should be ten x that. Yeah, maybe I don’t know. Brandon said, “McCool, that they spend three x their human salaries on inference.” Yeah, yeah, yeah. We’re not quite there yet, but I will say like our software engineers are extremely agent-pilled, including myself. I mean, I write code with agents like all the time. But I think like the big question, if you’re looking at like an Anthropic IPO and my.

【译】 我有一种感觉：强化学习会开始撞上一些限制。但我错了太多次了，所以也许我会继续错下去、一切照旧向前，最终我们能对人类要花六个月的事情做强化学习。谁知道呢。「三年后一个工程师的一天是什么样的？」我对软件工程极度看多。我写了三十年代码，见过太多变化：编程语言、云、各种框架。而每一次我们让程序员更高效，它都只是扩大了软件的需求。世界上有那么多地方急需软件，所以我对软件、对软件工程的需求非常非常看多。我认为外界最大的误解是：开发者是做「英语翻译成代码」的。那从来都不是软件工程师的工作。工作是首先理解需求本身，在你最终为之建设的人群之间充当接口，弄清楚这在流程和工作流上到底意味着什么——而这持续地在技术栈里上移。六十年代我们写汇编和打孔卡，然后到 COBOL，到 C++，到 Python，现在我们在用 agent。这只是又一次。而且这绝对是我三十年里见过的最大变化，但历史上确实有很多类似物、很多先例。这让我觉得：软件工程师未来十年、二十年都会好好的。「你们在 Modal 每月花多少钱买编码 agent？」大概五十万美元吧。「多少工程师？」一百五十个。「老兄，这低得让人担心。你们应该更狠地烧 token，应该是这个的十倍。」是吗，也许吧，我不知道。「Brandon 说（转写疑作）他们花三倍于人力薪资的钱在推理上。」对对对，我们还没到那一步。但我要说，我们的软件工程师全员深度 agent 化（agent-pilled，转写作 agent pills），包括我自己。我一直在用 agent 写代码。不过我认为，如果你看 Anthropic 的 IPO——

[2760s - 2880s] I’m not doing a Bloomberg and like pulling you into be like, what do you think of the anthropic IPO? I hate those ones. Yeah, you know where it’s like, yeah, it’s great. It’s no one in AI is like, oh, not a fan. Yeah, yeah, yeah. Everyone is against it. Go, but I do think a big, the motherload, so to speak, is Claude Code. It’s 60 billion dollars of revenue. It largely kind of dictates the two trillion dollar price. And you look at Mark Benioff, who says that they spend five percent of engineers’ salaries, 300 million, on Claude Code today. Now, if that comes way down as efficiencies improve, you name it, that’s a challenge. Or will that go up as it just eats more and more of engineers’ workflows? Yeah, it’s just going to eat more. I mean, like again, like we’ve seen this so many times. Like you make something more efficient, it makes something more powerful. It just unlocks like five new use cases, and then demand just goes up in the end. Like that’s that’s just the history of like the last 50 years of computers and software. But like I wouldn’t be here if I wasn’t such an optimist. That’s why I started a company. You know, like I’m just like the biggest like optimist that there is in the world about software and computers and like I just love this shit. And I think it’s just going to keep growing. The data market is an interesting one. We have so many players that scaled so fast. Yeah, it’s incredible. If I was 20 and I just wanted to make a lot of money, I would go out and start an environment company and just sell data to Frontier Labs. There is so much money to be made in that space. It’s kind of incredible. Let’s go a layer deeper. What would you do next? You specialize on a specific market. I mean, the coding space is where you see the most demand right now. Like, if you can sell like code environments to to the labs, like you can make a lot of money. Like, is a big space in that. Still, still, yeah, yeah. I meet startups all the time. I angel invested in a bunch of that. I mean, they’re incredible businesses. I I don’t think what I will say like my one piece of advice is like you make a lot of money today, but like is this a venture fund? Do you want to get on the venture treadmill? I’m not sure. Like I think these are companies.

【译】 「我不会像彭博那样拉着你问『你怎么看 Anthropic IPO』。」「我讨厌那种问题。」「就是那种『挺好的』。AI 圈里没有人会说『不看好』。」都对，但真正的母矿（motherlode）是 Claude Code。它有 600 亿美元收入，基本决定了那 2 万亿的估值。你看 Marc Benioff 说他们今天花了工程师薪资的 5%——3 亿美元——在 Claude Code 上。如果随着效率提升这块大幅下降，那是个挑战；还是说它会涨，因为它吃掉工程师越来越多的工作流？它只会吃得更多。再说一次，这种事我们见过太多次了：你把某个东西变得更高效，它就变得更强大，解锁五个新用例，最终需求就是会上涨。这就是过去 50 年计算机和软件的历史。但如果我不是这么乐观的人，我今天就不会坐在这里，我也不会创业。我是全世界对软件和计算机最乐观的人，我就是爱死这些东西了，而它会继续增长。数据市场也很有意思——那么多玩家规模暴涨。「不可思议。如果我现在 20 岁、只想赚大钱，我会出去开一家环境公司（environment company），把数据卖给前沿实验室。那个领域能赚的钱多到不可思议。我们再往深一层：你接下来会做什么？你会专注一个特定市场。」编码空间是现在需求最大的地方。如果你能把代码环境卖给实验室，你能赚很多钱。那里面有很大的空间。现在还是。「还是，是。我天天遇到这类创业公司，我天使投了一批。它们是不可思议的生意。我要说的一点建议是：你今天能赚很多钱，但这是一门风投生意吗？你想上风投的跑步机吗？我不确定。这些公司——」

[2880s - 3000s] They’re going to make a lot of money today. I don’t know if they’re going to make a lot of money in very long run, but like, who cares? Like, you know, take the money. It’s like great money. Like, it’s a fantastic business today, and that’s all that matters. I think there’s there’s data marketplaces, and then there’s data and services. And when you look at the leading providers, they started as the data acquirers and then just sell you the data, and that’s kind of it. And I think they’ve really gone full lifecycle now to deployment, implementation, evaluation, monitoring, with a real enterprise handholding ability similar to the FDEs, where they will really be that partner to ensure full implementation. Yeah. That’s how you expand margin and have some form of business in the long term. For sure, yeah, and I think it’s going to be a tough business in the very long run, but like today, you can have extremely good cash flow selling these datasets. Can I when we think about that kind of implementation, working with partners so closely, you said and Lin has said too from Fireworks about the importance of owning your own intelligence. Can you unpack how this looks because specialized models for companies in this way? Kind of jury’s still out. Yeah. We have Harvey who says that it’s fantastic and you know it’s been amazing, and then you have Lagora who say ridiculous. The underlying models move so fast; it’s the wrong strategy. Yeah, for some stuff, it’s good to own your intelligence. For other stuff, is not like I don’t think it matters. Like, like we have a lot of customers who run proprietary models on Modal too, like and and their customers, you know, they use a proprietary model. Like, I don’t I don’t think it makes sense for everyone to own their own intelligence. Why does it make sense for you to versus not? I I think for certain use cases, you can take an open source model and you can fine tune it, and you can get really good performance. And the nice thing is you know what the fine tuned model is based on. It’s not going to change. Like they’re not going to take away the model from you. It’s nice to have that value, like that consistency and and the ability to like continue to like tweak it without without you know the underlying model changing, which has happened in some cases with the Frontier Labs. That’s like you know that’s nice. Like you know I support that, but like I you know I think there’s many other things that matter to.

【译】 ——今天会赚很多钱。我不知道它们超长期是不是还能赚很多，但管他呢——把钱赚了。这是很棒的钱，今天是一门极好的生意，这才是全部。我认为存在数据市场（data marketplaces），然后存在「数据+服务」（data and services）。看那些头部提供商：他们起步是做数据采集的，然后就是把数据卖给你，仅此而已。而他们现在已经走完整个生命周期——到部署、实施、评估、监控，具备真正的企业级陪跑能力，和 FDE 类似，真正成为确保完整落地的伙伴。这就是你扩张毛利、并在长期拥有某种生意的方式。「肯定。」是啊，我认为超长期这是门艰难的生意，但今天卖这些数据集可以有极好的现金流。问到与伙伴紧密合作这件事——你说过，Fireworks 的 Lin 也说过「拥有自己的智能」的重要性。你能展开讲讲吗，公司用这种方式做专门化模型？陪审团还没定论。我们有 Harvey 说这棒极了、非常惊艳；然后你有 Lagora（转写疑作）说荒谬——底层模型变得太快，这是错误的策略。有些东西，拥有自己的智能是好的；另一些东西，我觉得无所谓。我们有很多客户在 Modal 上跑专有模型，他们的终端用户用的就是那个专有模型。我不认为「人人都该拥有自己的智能」成立。「那对什么样的公司成立、什么样的不成立？」某些用例里，你可以拿一个开源模型微调，得到非常好的性能。而好处是你知道这个微调模型建立在什么之上。它不会变。没人会把模型从你手里拿走。拥有这种一致性、能持续调参而底座不变的能力是很好的——而底座改变这种事，在前沿实验室身上已经发生过几次了。这很好，我支持。但还有很多别的东西同样要紧。

[3000s - 3120s] To people, you know, owning your intelligence is like just one thing out of many. Like, I think latency matters. I think cost matter, performance. I think you know hackability. I think you know being you know having a platform of a lot of adjacent services. Like these things matter too. So owning your own intelligence, like yeah, I’m supportive, but you know, there’s many other things. Isn’t that kind of going against your business? No, I don’t think so. I, I, we. I love the honesty, but like, dude, it would be in your favor to say yes. Every company is going to need it, and we’re the partner to help you. I kind of fucking teed you up there, and you. Well, I mean, not not really. I mean, if you think about TypeSafe with Jev, right? Like, you know, it’s a highly proprietary model. Like, I want people to use Jev. I mean, it runs on Modal. Like, like so, we have many customers like that themselves are proprietary model like that are hosting on Modal that are serving their customers, and I’m very, very in favor of that too. I think the beautiful model is always that it’s like a general purpose platform, and we support so many different things. That that to me is like you know the beauty is like you know we can if you want to use open source models like that’s great like we’ll help you and there’s many customers that build incredible AI applications on top of open source models fine tuning them reinforcement learning have a bunch of you know adjacent sandbox use cases or training whatever but there’s a lot of companies that have built highly successful businesses you know Suno Runway TypeSafe and many others like running their own proprietary model on Modal like and and their customers don’t care and and that that that I think is like you know puts us in a really good spot. I want to ask two more questions before I do a quick fire. Okay. Modal routing, as the dear Americans say, is that a defensible business? What do you say in the UK? Routing, routing, routing. I love that. Well, I mean, we are actually the original language. America just took us and bastardized it. But yes, my question to you is that a defensible segment of this value chain, or is that a commoditized product that everyone just builds themselves? I think it’s possible to build a successful business in model routing. I think you can do that.

【译】 对人们来说，「拥有自己的智能」只是众多因素之一。延迟要紧，成本要紧，性能要紧，可改造性（hackability）要紧，拥有一个带很多相邻服务的平台要紧。这些东西同样重要。所以拥有自己的智能——我支持，但还有那么多别的东西。「这不是和你的生意相反吗？」「不，我不这么认为。」「我爱这份诚实，但老兄，说『是』对你有利啊：每家公司都会需要它，而我们就是帮你的伙伴。我都把球给你喂到这了，你却……」嗯，其实不是。你想想 TypeSafe 和 Jev：那是个高度专有的模型，我希望人们用 Jev——而它就跑在 Modal 上。我们有很多这样的客户：他们自己有专有模型，托管在 Modal 上、服务他们自己的客户，我也非常支持那样。我认为最美的模型始终是：一个通用平台，我们支持这么多不同的东西。美就美在：如果你想用开源模型，太好了，我们会帮你——有非常多客户在开源模型之上、通过微调、强化学习和一堆相邻的沙箱用例或训练，建起了不可思议的 AI 应用；但也有一大批公司建成了高度成功的生意——Suno、Runway、TypeSafe 和很多别家——把他们自己的专有模型跑在 Modal 上，而他们的客户根本不在乎。这让我们的位置非常好。快问快答前我再问两个问题。「模型路由（routing）——用美国人的说法——是一门可防御的生意吗？你们在英国怎么说？Routing（发音不同）。我爱这个。我们才是英语的原产地，美国人拿过去糟蹋了。」但我的问题是：它是这条价值链里可防御的一段，还是人人都能自建的商品化产品？我认为在模型路由里做成一门成功的生意是可能的，你可以做到。

[3120s - 3240s] Today, as a standalone product, I wonder if long-term it has to be part of a more comprehensive platform. At the end of the day, I think what what these companies want they want evals, they want you know ways to fine tune, they want you know more control over performance, like latency, throughput, like all these things. And I I think it’s like one of those things that kind of makes sense today. Like these are very good businesses. I invested in OpenRouter personally, and they had a great outcome. I actually think they could have done even better on a standalone basis. But but I think in the very long term, it feels like one of those categories. Like you know, and but that’s my view on everything in AI. It should really be a part of you know larger, more comprehensive set of tools. Congratulations, you’re on the OpenRouter board, and the acquisition offer comes in. Please make the case for why they would have reject. Yeah, please make the case. I think I think OpenRouter, like the growth rate and the revenue, like this is and and the brand, frankly, like they have a really strong brand. I understand. I mean, there’s a lot of like there’s 500 other competitors, but but but but the the space is growing so fast, and there’s so much revenue. I would have told them to keep going. Final one, I think a really telling question is what do you believe that does not go in the favor of your business? And so, for example, I think that seed investing today is kind of dead and bullshit. This idea of like two on ten and we’ll slowly build for 18 months. Well, it’s kind of bad. Like it’s kind of bad if you want to make money, right? But it’s great if you want to develop a relationship with engineers, or I mean, with founders. Sure, but that goes against what I do. I’m I’m traditional. You’re you’re against a relationship with founders. No, no, no. I’m saying like I think the death of seed investing, the two on ten and build for 18 months, is dead. Yeah, no, I understand what you’re saying. So, what would yours be? I I don’t know. I I mean, I think like one one thought I have is like obviously at some point this AI stuff is going so fast, and we’re here, and we’re going to go here, and then there’s going to be pullback, you know, 20 percent, 30 percent in the stock market. It was going to.

【译】 今天，作为一个独立产品——我怀疑长期看它必须成为一个更全面平台的一部分。归根结底，这些公司要的是 evals，是微调的方式，是对性能更多的控制——延迟、吞吐，所有这些。它是那种「今天说得通」的东西之一，这些是很好的生意。我个人投了 OpenRouter，他们有个很好的结局。我其实认为独立发展他们本可以做得更好。但在超长期，它感觉是那种（会被吸收的）品类之一——我对 AI 里的一切都是这个看法：它真的应该属于一套更大、更全面的工具。「恭喜，你在 OpenRouter 董事会里，收购要约来了。请论证一下他们为什么该拒绝。对，你来论证。」我认为 OpenRouter 的增速和收入——坦率地说还有品牌，他们有非常强的品牌。我理解有一堆 500 个竞争对手，但这个空间增长太快了，收入体量太大了。我会告诉他们继续干。最后一题，我认为非常能说明问题：「你相信什么对你自己的生意不利的事？」比如我认为今天的种子投资某种意义上已经死了、是扯淡——那种「200 万投 1000 万估值、然后慢慢建设 18 个月」的想法。如果你想赚钱，这挺糟的；但如果你想和创始人建立关系，很棒。「当然，但那和我做的事相反，我是传统派。」「你是反对和创始人搞好关系？」「不不不，我是说种子投资的死亡——200 万投 1000 万、慢慢建 18 个月——已经死了。」「对，我懂你的意思。那你的呢？」我不知道。一个想法是：显然在某个时点，这个 AI 的事情跑得这么快，我们现在在这里，会冲到那里，然后股市会出现回调——20%、30%。它总会——

[3240s - 3360s] Freak out and like people are going to say it’s the end of the world. I’m almost certain it’s going to happen, but like I’m so bullish because then we’re going to keep going. And like you know, like so so that that would be like one thing that I’m like you know there’s going to be a pullback. I’m okay with that because I think the long-term opportunity is so fucking big. It’s like so large. Totally agree. Do you take secondaries? Very little. Why? I I mean I personally don’t need a lot of money. I you know I I’m pretty happy with life and I have enough. So as a personally, I I actually supported for my employees though. I think it’s it’s quite important actually as a founder to to that you make the equity worth as much as possible. And one of the way to do that and to tell future candidates that no, this isn’t just paper money. This is real money. Is to have a precedent of doing employees secondaries. And you can point to you know last round we let employees sell this much or whatever and you know up to ten percent. Like I think you have to have guardrails. Like you don’t want employees to sell too much. You want them to still be very you know have a lot of skin in the game and and be very incentivized to to make the company successful. But I think it’s a it’s a way to like you know like create a very liquid hard currency out of the equity. I totally agree. And as you think about a talent marketplaces, absolutely more competitive. It’s very extremely competitive out there. Yeah, is it nuts? It’s nuts. It’s nuts. I mean, you know, like the the offers we see for we’ve lost new grads going to you know some of the labs for like like talking you know twenty plus million. I mean, these are exceptional people. Exceptional. What makes an exceptional exceptional new grad? You know, like just just fucking sharp people. You know, these are like I know it when I see it. Like they’re just you know they’re just extremely hungry, hardworking. And where do they go?

【译】 ——吓坏所有人，人们会说世界末日到了。我几乎百分百确定这会发生。但我如此看多，因为之后我们会继续向前。所以那就是一件事：会有回调，我接受，因为长期机会大得要命，就是那么大。「完全同意。你们做二级出售吗？」非常少。「为什么？」我个人不需要很多钱。我对生活挺满意的，已经够了。就个人而言……不过我实际上支持员工做。我认为作为创始人，把股权价值做上去其实相当重要。方法之一，也是告诉未来的候选人「这不只是纸面财富，这是真钱」的办法，就是留下做员工二级出售的先例。你可以指着说：上一轮我们让员工卖了这么多，上限 10% 之类。你必须设护栏：你不想让员工卖太多，你要让他们仍然有足够的 skin in the game、有足够激励把公司做成功。但我认为这是把股权变成一种高流动性硬通货的方式。「我完全同意。」再看人才市场——绝对更激烈了，极度激烈。「疯了吗？」疯了。疯了。我们看到的 offer——我们流失的新毕业生去了某些实验室，2000 万美元以上的包。这些人出类拔萃。出类拔萃。「什么造就一个出类拔萃的新毕业生？」就是他妈的极其锋利的人。我看到就认识：就是极度饥饿、极度勤奋的人。他们都去哪？

[3360s - 3480s] I mean, they go to some of the big labs. Is that it? They go to OpenAI, yeah, the big labs. Yeah, they pay a lot. But is that it? Is that where the talent comes? There’s like a few other ones, but but I mean, mainly those people are are known for paying very a lot of money. And you won’t pay that much. We pay a lot. We can go very aggressive. We want to be out there paying 95th percentile for in the market. And and there’s like a few companies that I know, like they’re probably going to pay a little bit more. I think we have way more upside in the equity. And going back to what I said about secondaries, I think it’s important that for for that reason also, like to create like a real tangible value and the currency in the equity that makes us, you know, our offers seem very very strong. Plus the fact that we have a lot of upside. I mean, you you said it right. Like our stock seems cheap. Like I kind of agree. There’s a lot of upside in our stock, so that enables us to not necessarily have to match with the labs, but I think you know create a compelling sales pitch for why our comp package is like basically better than the labs, and we’re willing to go very far to get the best talent. Okay, we’re going to do a quick fire round because otherwise I could talk to you all day. You have a great angel portfolio. OpenRouter, you mentioned lovable before. You can put all of your money into only one. I love turbopuffer. I know Simon really well. Danish guy. So you have this Scandinavian connection. Just a fantastic entrepreneur. I just love the business. How big can that business be? That’s a good question. I mean, it is like somewhat specialized. They’re doing vector databases. I I think it could be a very large category. If I ran that company, I would have a 500 different ideas of how to expand into many adjacent things around databases, data warehouses, other types of query querying system. Like there’s, it could be a large category. What have you changed your mind on most in the last 12 months? Probably the magnitude of the sandbox space. I think it’s kind of insane. Like I look at the revenue and it just keeps going up, and I’m like, well, clearly has to stop right now. And I think I’m like doing some sort of reverse psychology on the revenue because it just keeps going up. And like, what is going on? It just keeps growing and growing. And like, I’ve been wrong so many times. Like, you know,

【译】 他们去一些大实验室。「就这些吗？」他们去 OpenAI，对，那些大实验室。他们付很多钱。「但就这样吗？人才就是从那来的？」还有另外几家，但主要是那些以付天价著称的公司。「而你们付不到那么多。」我们付很多，我们可以非常激进。我们要做市场上付 95 分位的公司。我知道有几家公司可能会付得再多一点，但我认为我们在股权上的上行空间大得多。回到我说的二级出售——也是为了这个原因：把股权做成真实有形的价值和通货，让我们的 offer 显得非常非常强。加上我们有巨大上行空间——你说得对，我们的股票看起来便宜，我也这么觉得。我们股票有大量上行空间，这让我们不必在现金上硬拼实验室，而是能讲一个有说服力的故事：为什么我们的薪酬包基本好过实验室。而且为了顶级人才，我们愿意走得非常远。「好，快问快答，不然我能跟你聊一整天。你有一流的天使组合：OpenRouter，还有你之前提到的 Lovable。只能押一家、全押。」我爱 turbopuffer（转写作 TurboPuffer）。我和 Simon 很熟，丹麦人，所以你们有这层斯堪的纳维亚连线。他就是个 fantastic 的创业者，我就是爱这门生意。「这门生意能做多大？」好问题。它有点专门化，做的是向量数据库。我认为那可以成为一个非常大的品类。如果我经营那家公司，我会有 500 个扩张进相邻领域的点子：数据库周边、数据仓库、其他类型的查询系统。它可以是个大品类。「过去 12 个月你在什么事上改观最大？」大概是沙箱市场的量级。我觉得有点疯了：我看着那条收入曲线一直涨，我心想「现在它总该停了」。我觉得我在对收入做某种逆向心理，因为它就是一直涨。怎么回事？它就一直长、一直长。我已经错了太多次——

[3480s - 3600s] In a good way, like you know, and and and so it’s incredible how large that space has gotten. What do you think is the most overheated category within AI today? Oh boy, I don’t know. I don’t think anything is overheated. I will say, like one of the things that I’m a little nervous about is companies that are dependent on continual infusion of equity capital. I think I would not invest in them, and and there’s a lot of them that I I would be nervous about anything where like if there’s a venture capital pullback, they are you know they will struggle to raise more money. I think it’s important you raise the the every round as if it’s your last round, and there’s like a real path towards commercialization and revenue. You got two kids. Two kids. Biggest lesson on being a good dad and being a good CEO. I mean, sometimes I feel like it’s kind of similar. Like, you mean like in terms of like how to balance it, or do you mean in terms of like being a good dad? What does it take to be a good dad? Setting them up for success in life, and I always like think about like the right amount to push the kids. I’m not a tiger dad, but like I think a little bit you need to push the kids a little bit to like get them exposure. I don’t want to create you know in the future a situation where my kids come back to me and say, Hey, Dad, I wish you’d made me do X or I wish you’d made me do Y. But I also don’t want to create a situation where my my kids come back to me when they’re 20 and they say, Hey, Dad, why did you make us do X or why did you make us do Y? So a little bit of pushing. I make them play piano. I don’t know why, but like I think it’s like good way to like learn something that I hope you know when they’re 20, 30, they will actually appreciate. They don’t like it that much, but I think stuff like that you got to push them a little bit, but not too much. My my parents were very lax, and you know I discovered math and and programming, and they were like Erik, why are you doing so much math? Like you should go out and play like the other kids. So in a way, I actually wish they would have pushed me even harder, but I pushed myself so hard anyway, so it doesn’t matter. So so that like right amount of. Why did you push yourself so hard? You’re sorry, you’re.

【译】 ——以一种好的方式——那个空间变得这么大，太不可思议了。「你觉得今天 AI 里最过热的品类是什么？」天哪，我不知道。我不觉得任何东西过热。我要说的倒是：让我有点紧张的一类公司，是那些依赖持续股权资本输血的公司。我不会投它们，而这样的公司很多。任何「一旦风投回撤就会挣扎着融不到钱」的东西，我都会紧张。重要的是：每一轮融资都当作最后一轮来融，并且有一条通往商业化和收入的真实路径。「你有两个孩子。」两个。「当好爸爸和当好 CEO，最大的心得是什么？」有时候我觉得这俩挺像的。你是说怎么平衡，还是怎么当个好爸爸？「当好爸爸需要什么？」为他们人生的成功打好基础。我总在想推孩子的「合适量」。我不是虎爸，但我觉得需要推一点，让他们获得接触。我不想在未来造出一种局面：孩子回来对我说「爸，我真希望你当年逼我学 X、逼我学 Y」；但我也不想造出另一种局面：他们 20 岁时回来问「爸，你当年为什么逼我们做 X、做 Y」。所以要推一点。我让他们弹钢琴。我也不知道为什么，但我觉得那是学一样东西的好方式，我希望他们 20 岁、30 岁时会真心感激。他们现在不太喜欢，但这种事就是要推一点，但不能太多。我父母非常放养，我自己发现了数学和编程，他们说「Erik，你搞这么多数学干嘛？出去和别的孩子玩去」。所以某种意义上，我其实希望他们当年推我更狠一点——不过反正我把自己推得够狠了，所以无所谓。就是那个「合适量」……「你为什么把自己推得这么狠？你是说——抱歉，你是……」

[3600s - 3720s] Actually, from a nation where you are told that it’s like bad to stand out. I’m forgetting that word, but I had a Swedish girlfriend. Yeah, where it’s like it’s bad to to stand out. Yeah, don’t think you’re special. Yeah, why? But I don’t know if that’s true. Like, I think there’s a lot of misconceptions to Scandinavian culture. I do think there’s a lot of like respect for people who follow their dream and like do things and like whether, like I said, you know, it’s music or writing books or you know, making furniture or art, like or being an entrepreneur. Like, I actually think there’s a lot of respect for. I think this like you know self actualization and following your dreams is like kind of highly regarded. And I think you know it’s like like that’s the flip side of having a very lax culture. Is like you kind of pursue people to follow their dreams. And I don’t know. My dream was like doing math and programming. I love that shit, so I did that. Unless you’re from Norway, where a tax structure exists that basically makes it impossible to start a company. Yeah, I mean, Norway has some issues. It also barely works. Yeah, this is very true. But we do have a sovereign wealth fund. Yeah, so they are rich. Thank God for that. Maybe it’s like the first post-economic society. I don’t know. Literally, in certain, in multiple meanings of the word. Final one for you. What are you most excited for? It can be anything. Like I look at drug discovery for kind of chronic conditions. My mom actually drug discovery is like one of the applications I think is so incredibly cool. Like I can’t wait to take AI generated drugs. You know, I’m like old and my skin is getting bad, and like I don’t know, like you know, every time I drink a beer, I feel like horrible the day after. Like I don’t know, give me all the AI generated drugs I can take. Like I want to live long. I want to live a healthy life. You know, and I think do you think that is actually a realistic future? I think so, but but I think with the FDA and other things, it’s going to take a while. But I think there’s so much incredible progress today in like drug in AI aided drug design. I think there’s so many cool stuff. So many things. Dude, I’ve so enjoyed this show in a way that I I honestly wasn’t expecting you to be this fun, open, charismatic. You’ve been amazing, so thank you so much. Thank you so much.

【译】 「你其实来自一个大家告诉你『出头不好』的国家。我想不起那个词了，但我有过一个瑞典女友——就是那种『出人头地不好』『别觉得自己特别』。为什么？」我不确定那是不是真的。我认为大家对斯堪的纳维亚文化有很多误解。我确实认为，那里对追随梦想、动手做事的人有很高的尊重——不管是音乐、写书、做家具、艺术，还是当创业者。我认为有很多这样的尊重。自我实现、追随自己的梦想，是被高度推崇的。而这正是一种非常宽松的文化的另一面：你放手让人们去追梦。我的梦想就是做数学和编程，我爱死它了，所以我做了。「除非你来自挪威——那里的税制基本让你没法开公司。」挪威是有些问题。「它也几乎不能运转。」「这倒很对。但我们有主权财富基金，所以他们很富有。」「谢天谢地。也许那是第一个『后经济社会』——在这个词的多个意义上，都是字面意义的。」「最后一题：你最期待什么？什么都可以。」我看的是慢性病的药物发现。我妈其实……药物发现是我认为酷得不可思议的应用之一。我等不及要吃 AI 生成的药了。我老了，皮肤在变差，每次喝啤酒第二天都难受得要命。不知道——把所有 AI 生成的药都给我吧。我想活得更久，想过健康的生活。「你觉得那真是一个现实的未来吗？」我认为是，但考虑到 FDA 和其他因素，会需要一阵子。而今天 AI 辅助药物设计已经有太多不可思议的进展，太多酷的东西了。老兄，我太享受这期节目了——说实话我没想到你会这么有趣、这么开放、这么有魅力。你太棒了，非常感谢。非常感谢。

[3720s - 3732s] Much for having you. This is a lot of fun.

【译】 （续上句）邀请我来。这期聊得太开心了。