Anthropic CFO Sees Revenue Reach $30 Billion as Compute Spend Top $100 Billion

Anthropic CFO预计营收将达300亿美元，算力支出超千亿美元 将绝大部分算力重新投入内部AI研发，相比出售容量获取即时客户收入，能更快加速前沿智能的发展。 Krishna Rao详细拆解了Anthropic如何将算力作为其核心战略命脉进行管理，以应对年化运行率收入（run-rate ARR）从90亿美元跃升至300亿美元的指数级业务增长。他概述了涵盖Trainium、TPU和GPU的多云多硬件平台战略，以及超过1000亿美元、规模达数吉瓦的长期基础设施承诺。讨论还探讨了缩放定律为何依然成立、企业对前沿智能的回报为何会引发杰文斯悖论，以及Claude Code等内部工具如何在编写Anthropic超过90%代码库的同时重塑财务等内部运营流程。 算力是 Anthropic 的核心生命线，占据了 CFO 约 30% 至 40% 的日常精力；公司在 AWS Trainium、Google TPU 与 NVIDIA GPU 三大硬件平台间实现了完全的算力互换与调度，保障全天候动态优化。 (5:17) 公司坚持前沿模型研发的算力保底底线，通过将顶尖模型深度用于内部加速推动递归自进化；目前 Anthropic 内部超过 90% 的代码均由 Claude Code 编写。 (15:00) 企业级客户对真实任务与 Agent 工作流的需求带来了极高的前沿智能回报，推动 Anthropic 的年化运行收入从年初的 90 亿美元跃升至第一季度末的 300 亿美元，且全美财富 10 强企业中已有 9 家采用其服务。 (12:38) Anthropic 的长期算力布局涵盖超过 1000 亿美元的“千层蛋糕”式基础设施承诺，包括自 2027 年起由 Google 和博通提供的 5GW TPU，以及来自亚马逊高达 5GW 的 AWS Trainium 算力部署。 (23:17)

Anthropic CFO Krishna Rao argues that managing compute as a flexible, unified portfolio is essential to sustaining frontier AI capabilities, which continuously unlock high-value enterprise demand despite severe operational and capital risks. Compute Procurement and the Cone of Uncertainty (0:00) Managing compute accounts for 30% to 40% of Rao's time as CFO, driven by binary capital risks: overcommitting to infrastructure can threaten solvency, while under-procuring prevents the company from serving clients or advancing frontier models. Because power and data centers require long lead times, Anthropic models potential growth across a wide 1- to 2-year 'cone of uncertainty.' To support this runway, the company secured a multi-year 'layer cake' of infrastructure, including a $100+ billion commitment spanning up to 5 gigawatts of Google and Broadcom TPUs starting in 2027 and up to 5 gigawatts of AWS Trainium capacity, complemented by opportunistic short-term deals like SpaceX's Colossus site in Memphis. Hardware Fungibility and Dynamic Resource Allocation (5:17) Unlike peers tied to single ecosystems, Anthropic operates across AWS Trainium, Google TPUs, and NVIDIA GPUs, deploying custom compilers to shift workloads across architectures from the bare metal up. Compute is managed as an integrated capital envelope rather than standard SaaS marginal cost, dynamically reallocated between model training, internal R&D, and enterprise serving. The company enforces an inviolable compute floor for frontier research, prioritizing capability gains even when customer capacity is constrained. Recursive Improvement and Internal Leverage (15:00) Anthropic foregoes near-term revenue by directing massive compute internally to create compounding development loops. Over 90% of the firm's internal code is generated by Claude Code, which also authors significant portions of its own infrastructure, enabling Anthropic to release 30 product updates in a single month. Similar agentic workflows in finance automate 90% to 95% of Monthly Financial Reviews and draft statutory statements across global entities. Rao describes this dynamic as a 'Jevons paradox for labor,' where elevating worker productivity expands operational ambition rather than reducing overall headcounts. Commercial Trajectory and Enterprise ROI (12:38) Rao reports that Anthropic scaled run-rate revenue from $9 billion at the start of the year to $30 billion by the end of Q1, supported by annualized net dollar retention exceeding 500% and adoption by nine of the Fortune 10. The company rejected traditional assumptions about deceleration, arguing that enterprise clients consistently migrate to cutting-edge models because real-world execution ROI far exceeds token pricing. Structural price reductions on tiers like Opus expanded aggregate consumption by enabling previously cost-prohibitive workloads, while core safety research—such as neural network interpretability—acts as a direct enterprise sales driver. Deployment Thresholds and Downside Pre-Mortem (55:38) Surges in capability have forced structured adjustments to product rollouts. When Mythos identified 250 security vulnerabilities in a codebase where prior models found 22, Anthropic instituted a phased release limited to defensive patching to mitigate dual-use offensive exploitation. In a strategic pre-mortem, Rao identifies three primary vulnerabilities that could pull growth toward the lower end of their planning cone: organizational adoption friction across legacy enterprises, unexpected plateaus in empirical scaling laws, and fierce competition at the technological frontier. While Anthropic's exponential revenue and infrastructure expansion validate the thesis of outsized returns on frontier intelligence, long-term stability hinges on scaling laws persisting uninterrupted and enterprise adoption keeping pace with aggressive multi-gigawatt capital commitments.

Anthropic首席财务官Krishna Rao指出，算力是前沿人工智能实验室的核心生命线；通过在三大主流芯片平台间维持高弹性调度、主动牺牲短期外部商业化收益以保障内部研发飞轮，并依托对齐安全研究获取企业级客户的深度信任，公司正以非线性方式拓展商业与技术边界。 算力治理与跨芯片弹性布局 (0:00) Rao透露，其日常约30%至40%的时间直接投入于算力战略与采购管理。由于大型电力与算力基础设施建设周期极长，前沿实验室面临着极为严峻的双向风险：采购过剩可能导致资本链断裂，而采购不足则可能导致技术掉队或无法承载客户需求。为了在Rao所称的“不确定性锥体”中保有调整余地，Anthropic构建了跨AWS Trainium（由Annapurna Labs支持）、Google TPU以及NVIDIA GPU三大硬件平台的完全可替代性架构，通过自研编译器直接在底层裸机上优化利用率，实现工作负载在不同芯片代际间的自由迁移。 在基础设施规划上，Anthropic采用了结合长短期的“千层饼”架构。一方面，公司通过近期合作（例如利用SpaceX与xAI在孟菲斯设立的Colossus设施）灵活补充即时算力；另一方面，公司锁定了总额超过1000亿美元的长期资本承诺，其中包括自2027年起从Google与Broadcom获取最高达5吉瓦的TPU算力，以及从亚马逊获取最高达5吉瓦的Trainium算力，以此平抑长期供给风险。 递归自我改进与内部算力优先策略 (15:00) 在日常算力分配的零和博弈中，Anthropic严格划定了前沿模型研发的非谈判算力底线。Rao强调，若完全切断公司内部对先锋模型的算力占用，转而将其投入外部商业化服务，短期内可立即释放“数十亿美元的收入”；但Anthropic坚持战略性地限制即时收入，全力保障内部研究算力，旨在通过高人才密度与顶尖AI工具的结合加速递归式自我改进。 目前，Anthropic内部超过90%的代码由Claude Code编写，且该工具自身底层基础设施的大量代码也由其自主生成；依托这种高密度杠杆，工程团队曾创下单月推出30项独立产品与功能更新的纪录。在非工程领域，财务团队通过内部AntStats平台构建了70多个定制化技能，使月度财务回顾的文档生成自动化率达到90%至95%，分析耗时从数小时压缩至30分钟，并且全球法人的法定财务报表初稿也已由Claude全自动起草。 商业化飞轮、定价弹性与杰文斯悖论 (32:42) Rao主张跳出传统SaaS将算力简单计入边际履约成本（COGS）的分析框架，转而将算力视为涵盖推理、内部工具与未来模型训练的统一资产组合。由于算力供给依赖提前约12个月的硬件供应链铺设，其扩容并非线性跟随短期业务波动，Anthropic的营收扩张是在预先承诺的固定基础设施投资之上释放的。据Rao介绍，公司年化运行营收从年初的90亿美元快速攀升至第一季度末的300亿美元，年化净收入留存率（NDR）维持在500%以上，且已为《财富》前10强企业中的9家提供服务。 在定价与需求弹性方面，降价策略展现了显著的“杰文斯悖论”。早先降低Opus层级价格后，客户不再勉强将复杂任务降级委托给Sonnet，单价降低所释放的潜在需求远超传统预期，直接引发了消耗总量的爆发式增长；后续Anthropic推出Opus 4.6等升级模型时维持基准价格不变，使企业客户在锁定预算确定性的同时获得了性能跃升。 安全驱动的商业壁垒与高密度组织文化 (45:00) 在产业生态中，Anthropic倾向于扮演类似AWS的基础设施平台角色，克制垂直业务扩张，优先通过API与金融、安全等垂直领域的行业龙头合作。Rao认为，公司在“可解释性研究”（充当神经网络的MRI）与对齐科学上的前瞻性投入，已从单纯的安全使命转化为关键的商业护城河，让交付高度敏感核心数据的顶尖企业建立了技术信任。针对高危能力的双重用途风险，Anthropic在推出网络安全漏洞发掘能力激增至250个（上一代仅22个）的Mythos模型时，首次确立了以防御性修补为核心的分阶段受控部署模式。 在组织治理方面，Anthropic维持“美国优先”的地缘政治合作立场，并以反官僚和反山头主义为核心。招聘过程中，文化面试拥有一票否决权，技术能力出众但不具备协作谦逊度的候选人会被直接淘汰。这种文化使得公司在竞对高薪挖角中仅流失2人，7位联合创始人与绝大多数早期员工至今全员在岗，每两周由CEO面对全员解答未经筛选的提问，确保战略意图高度透明。 下行风险预演与生命科学终极图景 (70:00) 在针对业务增长的“事前面前”（Pre-mortem）复盘中，Rao归纳了可能阻碍公司实现预期算力产出的三大潜在风险：大型企业在组织架构与工作流习惯上的消化吸收阻力、底层模型缩放定律（Scaling Laws）理论上可能遭遇的平台期，以及前沿智能赛道极高资本强度下的多方竞争。尽管当前经验数据表明缩放定律仍未见顶，但外部落地摩擦可能导致商业化增速与模型能力脱钩。 在更长期的技术愿景上，Rao认为前沿智能最有意义的应用场景在于生命科学与医疗健康。除了自动化处理临床研究报告与合规文书，若前沿模型能够直接指导复杂分子结构优化与基础生物实验，将实验室有效通量提升10倍至100倍，人类发现重疾靶向疗法的周期将被大幅压缩，释放巨大的社会与经济回报。 Rao所展示的核心论点在于，前沿AI竞争已非单纯的算法较量，而是演变为集百亿美元级芯片调度、算力内部再投资飞轮、企业信任架构与高密度组织文化于一身的复合工程。然而，尽管Anthropic依靠非线性扩张突破了传统软件增长模型，这种资本密集型路径仍高度依赖缩放定律的持续有效性，并面临着传统企业工作流改造滞后以及高昂前沿竞赛的多重考验。

## Background 本期内容由 Colossus 旗下的知名播客节目《Invest Like the Best》（第 472 期）于 2026 年 5 月 13 日在 RSS 及各大播客平台发布。节目主要采访了 Anthropic 首席财务官 Krishna Rao，围绕其如何采购与调配被视作公司命脉的算力资产展开，探讨了多芯片调度、不确定性锥体模型、前沿智能的超高回报率以及公司向 300 亿美元 ARR 扩张的底层逻辑。 ## TL;DR Anthropic CFO Krishna Rao 在对谈中深入拆解了支撑大模型实验室指数级狂飙的算力与财务底盘。他详细阐述了公司如何在 Trainium、TPU 与 GPU 三大异构芯片间实现动态可替代编排，并坚守模型前沿研发算力的刚性底线；凭借前沿智能解锁的多维落地能力，Anthropic 实现了 ARR 从 90 亿美元到 300 亿美元的爆发式跨越，内部更借助 Claude Code 实现了超过 90% 代码自写的递归自我改进。此外，他还分享了将算力总支出回报率（Return on Compute）作为核心北极星指标的全新财务范式，并就劳动力市场的杰文斯悖论、极度务实的内部文化以及未来由虚拟协作者重塑 40 万亿美元知识工作的愿景展开了深度剖析。 ## Atmosphere & style 整场对话洋溢着理性克制却极具冲击力的高密度智识张力，主持人 Patrick O'Shaughnessy 与 Krishna Rao 之间的问答精准而富有穿透力，毫无空泛的行业公关辞令。听众能真切感受到一位身处历史级科技浪潮核心的掌舵人，在面对以百亿美元计的物理资本承诺与不可逆的指数级复合增长时，那种将第一性原理、冷峻风险推演与深邃技术乐观主义融于一体的从容定力。从解构算力千层蛋糕的严谨逻辑，到结尾追忆兄长无私让渡教育资源时的温情流露，整期节目在磅礴的商业硬核推演中交织着深厚的人文关怀。 ## Bulletin - `12:38` Anthropic 凭借前沿模型能力的跨越与企业落地，年化运行营收（ARR）在短短数月内从 90 亿美元急剧攀升至 300 亿美元，净金额留存率（NDR）突破 500%。 - `23:17` 公司已锁定了联合 Google/博通的 5 吉瓦 TPU 及亚马逊的 5 吉瓦 Trainium 算力协议，未来资本承诺总额超过 1000 亿美元。 - `16:32` Anthropic 内部全面实现自举式研发，目前超过 90% 的工程代码由自研工具 Claude Code 编写，且底层代码也由其自身迭代构建。 - `49:02` 算力资产具备极致可替代性（Fungibility），白天作为高并发外部推理支撑营收，夜间与低谷则无缝切回研发训练模型，打破了传统 SaaS 的研发与营业成本边界。 ## Key takeaways - `02:29` 算力采购具有苛刻的双向风险：买太多会面临破产，买太少则会失去前沿领先地位，CFO 日常需将 30% 到 40% 的精力倾注于算力调度与决策。 - `05:17` Anthropic 建立在 Amazon Trainium、Google TPU 与 NVIDIA GPU 三大异构芯片平台的完全可替代性之上，通过自研编译器和裸机调度压榨极致效能。 - `09:08` 面对指数级增长的“不确定性锥体”，Anthropic 设立绝对刚性底线：无论外部客户需求多旺盛，前沿模型研发的算力配额绝不缩减。 - `10:00` 打破模型“跑车能耗悖论”，Opus 4 到 4.7 的代际演进不仅提升了能力，还实现了 Token 处理效率的成倍优化与强化学习沙盒推理的提速。 - `12:38` 企业级市场对前沿智能具有极端的高溢价倾向，评估标准已从通用单维跑分转向长视程复杂任务规划、工具调用及计算机使用能力。 - `18:46` 坚持“人才密度胜于人才规模”，顶尖科学家的价值不仅未被 AI 取代，反而通过前沿工具获得了开拓科学未知的指数级杠杆。 - `20:00` 从内部预训练损失曲线和实证数据来看，缩放定律（Scaling Laws）在内部演进中完全没有放缓的迹象。 - `27:53` 战略定位坚持横向平台而非全面收割垂类应用，类似早期 AWS 构建使能层，自研 Claude Code 等工具旨在牵引生态和验证模型能力。 - `32:42` Opus 4 的主动降价触发了典型的“杰文斯悖论”，单价降低反而激活了高价值企业负载的暴增，促使客户算力总消费非线性提升。 - `38:32` 财务团队全面“吃狗粮”，通过 Claude 沉淀 70 多个财务专属 Skill，自动生成 95% 成熟度的月度财务审查初稿，高年资税务主管成为 Token 消耗冠军。 - `45:00` 早期对 AI 安全与可解释性（神经网络 MRI）的纯科研投入形成了双重飞轮，不仅反哺前沿模型研发上限，更成为拿下《财富》前 10 强中 9 家的核心信任护城河。 - `55:38` 高危通用模型 Mythos 在网络安全测试中检出漏洞数量较前代跃升逾 10 倍（22 个到 250 个），促使 Anthropic 确立分阶段受控安全发布的行业模板。 - `58:51` 组织内部奉行“对手极其能打，成功绝非必然”的谦逊文化，7 位联创至今零流失，在全行业高薪挖角潮中仅损失 2 人。 - `01:03:48` 下一代前沿边界瞄准价值 40 万亿美元的知识工作市场，通过结合全局上下文、长期记忆和工具链的“虚拟协作者”颠覆传统软件研发模式。 - `01:10:00` 未来可能导致增长落入不确定性锥下界的三大风险：企业组织流程变革阻力、缩放定律潜在撞墙以及高烈度前沿竞争中的执行掉队。 ## Detailed walkthrough ### `02:29` 算力是 Anthropic 的命脉：生死攸关的采购与“不确定性锥体” 在访谈开篇，主持人 Patrick O'Shaughnessy 强调了 Krishna Rao 所处职位的特殊性——作为 Anthropic 的首席财务官，他身处当今全球最具爆发力的技术浪潮中心，每天都要面对海量算力的采购与调配。Krishna 直言，算力绝不仅仅是一项常规资本支出，而是整个公司的命脉，“它是构筑一切的底层画布（the canvas on which everything else gets built）”。因此，决定采购多少算力，是 Anthropic 内部最具深远影响、也是最艰难的战略决策。 算力采购面临着极度苛刻的双向风险：如果采购过多，巨大的资本开销会让公司直接面临倒闭风险；如果采购过少，不仅无法满足爆炸式增长的客户需求，更会直接从前沿模型竞争的第一梯队中掉队。现实世界的物理限制更放大了这种挑战——算力基础设施不是下个星期就能随叫随到的消费品，企业根本不可能今天决定要 1 个吉瓦（GW）的电力与算力，下周就能顺利交付。这要求团队必须建立极其严苛的前瞻性规划体系。Krishna 透露，即便到了今天，他个人依然要将 30% 到 40% 的精力完全倾注在算力相关事务上。 为了在指数级增长与长期交付周期之间取得平衡，Anthropic 采用自下而上的需求预测模型，并将其置于所谓的“不确定性锥体（cone of uncertainty）”中进行持续评估。公司在算力交易结构中预留了极高的商业灵活性，同时在技术使用层面也极度强调柔性，以确保在业务呈指数级增长时，每一份算力资产都能被榨取到极致效率。 - **核心矛盾**：“买太多算力会破产，买太少算力则无法服务客户且丧失前沿地位。” - **物理交付壁垒**：吉瓦级算力集群的规划与建设周期极长，必须提前数年进行布局。 - **管理层精力分配**：作为 CFO，Krishna Rao 日常 30% 至 40% 的时间专门用于算力采购与资源调度决策。 --- ### `05:17` 三大可替代芯片平台与全栈编排体系 当被问及算力战略中的“灵活性”究竟意味着什么时，Krishna 解释称，Anthropic 最大的技术护城河之一，在于能够在三大异构芯片平台之间实现高度“可替代（fungible）”的自由调度。目前，Anthropic 同时深度使用亚马逊的 Trainium 芯片、谷歌的 TPU 以及英伟达的 GPU。无论是前沿模型的预训练与后训练、内部员工的提效工具，还是外部客户的推理调用，这三大芯片平台都能在各类工作负载之间无缝承接。 这种架构灵活性并非一朝一夕之功，而是 Anthropic 团队历时数年重金研发的结果。Krishna 回忆道，早在团队开始大规模采用第三代 TPU（TPU v3）时，外界许多人甚至认为他们“疯了”，因为当时整个行业几乎完全绑定在英伟达的 GPU 与 CUDA 生态之中。然而，Anthropic 坚持深入硬件底层，不仅直接构建了自己的编译器，还从裸机（bare metal）层面自研调度编排层，甚至与亚马逊的 Annapurna Labs 团队展开极其密切的联合研发，直接反向影响芯片的硬件演进路线图。 这种自下而上的全栈控制力，让 Anthropic 能够根据不同代际芯片的硬件特性，将其精准匹配到最适合的内部负载上。其最终结果是：Anthropic 内部投入的每一美元算力成本，都能比行业其他前沿实验室产生更高的效能乘数。 ```mermaid flowchart TD accTitle: Anthropic 算力架构与资源流向 accDescr: 展示三种异构芯片平台经过 Anthropic 编排与自研编译器后，灵活供给三大核心业务负载。 subgraph Hardware["底层算力供给（三大可替代硬件平台）"] A["Amazon Trainium"] B["Google TPU"] C["NVIDIA GPU"] end subgraph Engine["Anthropic 核心基础设施层"] D["自研编译器与裸机优化"] E["跨集群动态编排层"] end subgraph Workloads["上层工作负载"] F["前沿模型研发（设定不可突破的底线）"] G["内部员工赋能（Claude Code / 研发加速）"] H["外部客户推理服务（商业化变现）"] end A --> D B --> D C --> D D --> E E --> F E --> G E --> H ``` - **三大硬件支柱**：Amazon Trainium、Google TPU、NVIDIA GPU。 - **历史前瞻选择**：在行业普遍迷信 GPU 的时代，从早期（如 TPU v3）就开始探索非 GPU 路线，构建多芯片跨平台支持能力。 - **软硬件协同突破**：自研编译器与编排层，与 Annapurna Labs 深度共创芯片路线图，直达裸机层优化。 --- ### `07:31` 指数级思考与“不确定性锥体”的应对策略 随后，Krishna 详细拆解了“不确定性锥体（cone of uncertainty）”这一核心认知框架。他指出，人类天生的思维模式是线性和渐进式的，习惯以固定的绝对值增量去推演未来；然而 AI 实验室的增长曲线是指数级的。在指数级复利模型下，周增长率或月增长率哪怕产生极为微小的波动，在经过一到两年的时间复利放大后，都会导向落差极其悬殊的终局结果。 Krishna 坦言，自己在加入 Anthropic 的两年时间里，被迫打破的最重要认知习惯就是“线性思维”，强迫自己全面转向“指数级思维”。为了应对这种极宽的结果分布范围，Anthropic 会建立一至两年的多情景推演模型，在不确定性锥体的上沿（极端乐观高增长）与下沿（相对保守平缓）之间设定不同的锚点，并以此倒推当前的采购与资源规划。 在此框架下，内部算力的分配逻辑展现出极为反直觉的权衡。Krishna 举例说明：如果 Anthropic 今天为了追求短期财务表现，剥夺内部员工使用自家模型的算力配额，把这些算力全部用于外部商业化服务，公司立刻就能多创造数十亿美元的当期收入。但从不确定性锥体的长远视角来看，Anthropic 选择坚决向内部技术人员倾斜算力。因为借助算力赋能研发人员（例如使用 Claude Code 自动化编程），能够加速模型迭代并找出更多的算力效率乘数，从而推动公司向锥体最顶端的高潜结果演进。 - **思维范式转移**：从人类本能的线性增量推演，全面转向指数级复合增长推演。 - **长短期利益取舍**：内部研发占用的算力若直接对外销售，价值可达数十亿美元，但公司宁愿放弃当期收益以换取内部研发速度的几何级提升。 - **规划方法论**：跨度 1 至 2 年的情景建模，基于锥体上下沿推导容错空间，依靠算力调优弥补预测与现实的偏差。 --- ### `09:08` 算力分配的博弈：研发底线、内部赋能与客户需求 随着业务规模扩大，Anthropic 不仅要开会决定买多少算力，更要频繁召开例会讨论算力在三大主要去向之间的即时分配：前沿模型研发（Model Development）、内部员工生产力（Internal Use）、以及外部客户服务（Customer Demand）。Patrick 提出，外人可能会主观猜测这三者是对称的“各占三分之一”分配模式，但实际的决策权衡远比想象中动态和激烈。 Krishna 强调，这些分配讨论之所以能够高效进行，首先依赖于 Anthropic 极具协作性、非零和博弈的企业文化，内部不存在争夺算力领地的“诸侯封建割据（fiefdoms）”。在具体的分配准则上，Anthropic 确立了一条不可动摇的红线：**模型研发所分配的算力拥有绝对的保底下限（floor）**。哪怕外部客户需求极度旺盛导致交付承压，或者不得不采取某些非常规手段来压缩服务，公司也绝不会削减用于研发下一代前沿模型的核心算力。 Anthropic 坚信，“前沿智能的投资回报率（returns to frontier intelligence）正在变得越来越高”，尤其是在高价值的企业级软件与复杂工作流场景中，只有真正处于绝对第一梯队的前沿模型，才能解锁最丰厚的商业利润。在研发底线获得绝对保护之后，剩余算力则由各个团队带着具体的预期 ROI 进行公开、坦诚的讨论，在赋能内部研发加速与承接外部商业变现之间动态流动。 - **三大核心算力池**： 1. 前沿模型研发（探索智能边界与架构迭代）； 2. 内部生产力工具（提效工程师研发、寻找工程优化乘数）； 3. 客户需求交付（API 与企业级产品推理）。 - **底线原则**：前沿模型研发算力享有不可突破的刚性保底，优先度高于短期客户服务承载。 - **战略底层逻辑**：前沿智能在企业级市场具备极高的收益倾斜，落后半步便会导致产品溢价大幅缩水，因此持续领跑是最根本的商业壁垒。 ### `10:00` 算力效率的飞跃与研发的正向飞轮 关于算力效率的度量，外界常常陷入一种直觉误区，认为大模型迭代就像汽车升级：从普通轿车升级到高性能跑车，性能虽然提升了，但油耗也必然随之暴涨——没人会为了省油去买超级跑车。然而在 Anthropic 的实际迭代中，这一类比完全被打破了。模型在能力发生阶跃的同时，处理效率（以更低的消耗高效处理 Token）同样实现了成倍的提升。从 Opus 4 逐步迈向 4.5、4.6 以及 4.7 的每一个版本升级，都伴随着 Token 处理效率的乘数级跨越。 这种效率的提升不仅惠及外部客户，也在内部引发了巨大的飞轮效应。在模型进行强化学习（Reinforcement Learning, RL）时，其本质就是在带有奖励函数的沙盒环境中进行大规模推理；一旦模型的基础推理效率提升，整个 RL 训练与对齐过程的效率也会成倍加速。Anthropic 借此实现了双赢：对外，新模型赋予客户更强能力且推理成本显著降低；对内，在重大版本发布的间隙，团队持续动态部署各类中间效率优化。驱动这一良性循环的核心正是研究团队——前沿模型能力的研发、算力效率的挖掘、客户服务推理以及内部利用未发布的最强模型加速研发，彼此深度咬合为一个严密的整体。 ```mermaid flowchart TD accTitle: Anthropic 算力研发正向飞轮架构 accDescr: 展现前沿模型、推理效率、内部强化学习与客户业务之间的循环互促关系 subgraph R_and_D ["前沿研发与效率引擎"] M1["新一代模型能力突破\n(Opus 4 / 4.5 / 4.6 / 4.7)"] EFF["Token 处理效率乘数级提升\n(跨代与代际间动态优化)"] RL["内部强化学习 (RL)\n沙盒环境高效推理"] end subgraph Flywheel ["内部杠杆与外部落地"] INTERNAL["内部工具与研发加速\n(使用未发布前沿模型)"] CLIENT["企业级客户解锁全新用例\n(推理成本与吞吐优化)"] end M1 --> EFF EFF --> RL RL --> INTERNAL INTERNAL --> M1 EFF --> CLIENT CLIENT -->|"收入反哺算力投资"| M1 ``` - **跑车悖论打破**：模型能力的代际跃升并不以牺牲能耗效率为代价，相反，Opus 4 到 4.7 每一代都带来了处理 Token 效率的成倍优化。 - **强化学习效率互锁**：由于 RL 本质是带奖励机制的沙盒推理，基础推理效率的突破直接大幅降低了后续模型训练与探索的算力门槛。 - **全天候动态优化**：除了大版本的阶跃式更新，团队在各版本生命周期之间持续动态推送能效补丁与算法改进。 --- ### `12:38` 前沿智能的超高回报：从 90 亿到 300 亿美元 ARR 市场中曾有一种流行观点：企业只需落后前沿半步，使用半年前的模型即可享受低廉成本，并逐步追平技术差距。但现实却截然相反，前沿智能（Frontier Intelligence）的经济回报率极其惊人。在真实商业世界中，智能并非一个简单的单维 IQ 分数（例如从 110 分涨到 125 分），公开基准测试评测榜单也早已出现饱和现象。Anthropic 衡量智能的唯一标准是客户反馈的真实世界落地能力——包括长链路复杂任务（Long-horizon tasks）的执行能力、工具调用能力（Tool use）、计算机使用能力（Computer Use），以及以极快速度完成复杂 Agent 任务的交付能力。 用一个生动的职场案例来说明：两位员工能力看似相当，但完成同一项任务，一人需要一周，另一人仅需一天；在持续运转的业务中，后者的生产力输出就是前者的七倍。前沿智能带来的质变直接为企业客户打开了过去无法实现的全新潜在市场空间（TAM）。Anthropic 自身的惊人增长正是这一逻辑的最佳注脚：公司在当年开局时的年化运行收入（Run-rate revenue, ARR）约为 90 亿美元，而在短短一个季度结束时，ARR 已经飙升至 300 亿美元。这种指数级爆发是纯粹由前沿模型智能跨越及其上层产品推动的。在企业端，客户见证了新模型解锁的全新生产力，立刻大规模增加对最新模型 Token 的预算投入，形成了“越前沿、越创造 ROI、越追加投资”的强劲正向循环。 - **智能的多维本质**：摒弃单一的跑分或智商维度，企业级智能核心看重长视程复杂决策、环境交互与工具操控能力。 - **业绩指数级飞跃**：仅凭前沿模型能力的解锁与配套产品落地，Anthropic 在短短 4 个月内将 ARR 从 90 亿美元推升至 300 亿美元。 - **ToB 市场的极端追求**：与对前沿模型不敏感的部分 C 端场景不同，B 端企业对能力边界极为敏锐，率先从编程场景蔓延至全业务流程，并疯狂为更高阶的智能埋单。 --- ### `16:32` 递归自我改进：Claude Code 书写自身与全速产品迭代 科技界一直探讨的“递归自我改进”（Recursive Self-improvement）如今已在 Anthropic 内部成为日常工程现实，这一趋势宛如科幻小说走入现实。外界猜测 Scaling Laws（缩放定律）是否见顶，但 Anthropic 近期的模型研发（如内部项目 Mythos）证明缩放定律依然极其强劲有效。更震撼的现实在于内部研发方式的彻底重构：目前 Anthropic 内部超过 90% 的代码已经完全由 Claude Code 编写，甚至 Claude Code 自身的大量底层代码也是由 Claude Code 自行编写完成。 这也解释了为何 Anthropic 愿意在内部算力调度中顶着“放弃对外商业化变现收入”的机会成本，大量划分顶尖算力供内部模型使用——因为前沿模型正在全力协助研究员打造下一代模型。在这个逻辑下，Anthropic 内部并不将模型割裂地划分为“开源”或“闭源”，而是清晰地划分为“前沿模型”与“非前沿模型”。只有身处最前沿的系统才能真正捕捉到最大经济价值并为客户带来实质性的 ROI。这种技术杠杆甚至直接反映在惊人的发布节奏上：仅在 1 月份单月，Anthropic 就密集推出了 30 项独立的产品与功能更新，这种极致的产品爆发力正是顶尖工程团队搭配自身前沿模型协同开发的结果。 - **代码自主书写比例**：Anthropic 内部超过 90% 的工程代码由 Claude Code 撰写，实现了自举式的代码构建与迭代。 - **内部算力分配逻辑**：宁可牺牲短期对外商业化租金，也要保障内部研发算力，核心在于模型自我加速对下一代模型产出的决定性价值。 - **产品推新速度**：在内部 AI 杠杆加持下，团队创下单月（1 月）连续交付 30 项产品与功能上线的惊人纪录。 --- ### `18:46` 人才密度胜于人才规模：AI 时代研究者的终极杠杆 当模型开始自己编写代码并参与下一代设计时，是否意味着人类研究员和顶尖工程人才终将被算力边缘化、甚至完全被取代？Krishna Rao 给出了极为坚定的否定回答。他强调，Anthropic 的核心底色始终是一家前沿科学研究实验室，这一上游探索引擎决定了后续所有商业化与产品化的走向。现阶段的前沿突破并非完全由模型自主生成，而是由人类科学家主导实验设计、定义优先级，并探索未知的科学边界。 在这个过程中，模型的角色不是取代人类，而是赋予顶尖人才不可思议的放大杠杆。Anthropic 内部崇尚的核心信条是“人才密度远胜于人才规模”（Talent density beats talent mass）。公司不需要庞大的人力团队去堆砌基础执行，而是致力于构建全球密度最高、水平最顶尖的 AI 研究员与推理工程专家群体。当这群最高认知密度的人才配备上最强大的前沿模型作为工具时，他们开拓全新领域和进行深度探索的能力被指数级放大，这构成了 Anthropic 最具防御力且无法轻易被复制的核心竞争力。 - **实验室定位不可动摇**：无论商业规模达到何种量级，驱动一切的上游核心依然是探索前沿极限的科学实验团队。 - **人才密度法则**：AI 时代不再需要堆砌平庸的人力规模，顶尖技术专家的密度结合前沿算力工具，才是爆发最高生产力的“制胜组合”。 - **人机协同的真实形态**：模型负责大幅消除繁琐的底层实现阻力，而人类专家则聚焦于方向设定、科学假设构建与未知前沿的发现。 ### `20:00` 缩放定律的内部视角：未见放缓的科学实证 关于缩放定律（Scaling Laws）是否见顶的讨论，外界目前的共识趋向于将其拆解为多个独立演进的环节——预训练（pre-training）、后训练（post-training）以及推理阶段计算（reasoning），普遍认为只有当所有环节同时受阻，技术才会真正撞墙。当被问及 Anthropic 内部如何看待这套逻辑时，Krishna Rao 阐明了他们基于模型全生命周期的实证方法。在预训练阶段，研究团队会紧密追踪训练中的损失曲线（loss curves），将其与上一代模型实时对照，从而预判能力的跃迁；在强化学习（RL）阶段亦是如此。 更为关键的输入则来自于真实世界的反馈。当客户上手模型后，他们所遭遇的卡点与痛点，会直接反哺并转化为 R&D 团队的下一代训练靶标。Krishna Rao 特别强调，Anthropic 在企业级端绝不使用客户数据进行训练，在专业消费者端也仅限于用户主动选择（opt-in）的情况；这种反馈更多体现为客户主动沟通的需求——例如模型在某个特定推理节点卡住，阻碍了客户构建下一代产品。Anthropic 对这些客户的建议通常非常干脆：“直接按照你的终态去设计产品，因为我们在研发端一定会随着时间把这项能力补齐。”尽管缩放定律开山论文的多位核心作者正是 Anthropic 的联合创始人，内部团队却始终是一群极其严苛的怀疑论者。他们严格恪守科学方法，不断用最新实验挑战过去的假设，但从内部所有模型快照及内外部对照指标来看，缩放定律在 Anthropic 内部“完全没有看到放缓的迹象”。 - **研发评估闭环**：通过预训练损失曲线预判基模能力，通过后训练强化学习优化推理表现，通过客户真实卡点反哺研发靶标。 - **数据隐私界限**：企业端数据绝不用于训练；消费端与专业端仅在用户显式授权（opt-in）时参与优化。 - **科学质疑精神**：创始团队虽为缩放定律奠基者，但内部崇尚不断证伪与推翻旧假设的严谨文化。 --- ### `21:54` 指数级思维：打破季度预测与以代码跃迁为参照系 面对持续演进的缩放定律，人类大脑习惯于线性思考的弱点成为了商业决策的最大障碍。Krishna Rao 坦言，理解复合增长率已非易事，而要去推演“模型能力的指数级暴增”所带来的连带影响，更是对认知维度的极限考验。在日常工作与业务管理中，他坚决抛弃传统的静态点估计（point estimate），转而全面依托场景化分析（scenarios）。更为关键的是，组织必须建立“极低的先验更新门槛”——如果一个月前成立的前提在今天被证伪，就必须毫不犹豫地推倒原有的业务与财务模型。 传统企业“每季度做一次预测、等待三个月后的董事会再行复盘”的节奏，在当下的前沿大模型行业已彻底失效。随着模型能力的阶跃，总可寻址市场（TAM）的边界每隔几周就会被重新定义。Krishna Rao 指出，这一飞轮效应在代码领域展现得最为淋漓尽致：从 Claude 3.5 Sonnet 开始，编程能力出现了令人惊叹的单点跳跃，紧随其后的便是爆发式的客户采用率、使用时长和营收激增。由于事先极难通过线性思维预测这种爆发，如今财务与战略团队将代码能力的突破作为“先行参照系（analog）”，去镜像和推演经济体系中其他垂类板块即将迎来的跃迁。 ```mermaid flowchart TD accTitle: 动态情景预测飞轮 accDescr: 展示从模型能力指数级突破到快速更新财务与业务场景预测的闭环机制 subgraph Capability ["模型能力突破"] A["缩放定律推动能力跃迁<br/>(例如: Claude 3.5 Sonnet 代码能力)"] end subgraph Market ["市场响应"] B["客户极速采用与推理用量激增"] C["TAM 范围重构与营收爆发"] end subgraph Decision ["动态财务与战略响应"] D["打破季度静态预测<br/>实施极低门槛的先验更新"] E["以代码爆发为参照系<br/>建立多情景动态推演模型"] end A --> B B --> C C --> D D --> E E -.->|"指引算力扩张与资源配置"| A ``` --- ### `23:17` 算力的“千层蛋糕”：从近场应急到千亿美元长约 为了喂饱算力需求，Anthropic 在全球范围内极具创造性地发掘供给。在访谈前数分钟，市场刚刚获悉他们与 xAI 达成合作，接入其位于田纳西州孟菲斯的 Colossus 算力中心（访谈中提及 SpaceX 孟菲斯设施这一同属马斯克旗下的算力网络）。Krishna Rao 表示，此类合作能迅速支撑消费者端与专业消费者端的扩张，但它本质上属于 Anthropic 随时搜寻的“近场算力（near compute）”。随着公司算力总盘子的急速膨胀，这类现货性质的近场算力在整体版图中的占比正变得越来越小。 在评估任何一笔算力时，团队都会综合考量五个核心维度：**价格**、**合同期限**、**物理空间位置**、**芯片架构类型**以及**内部运行效率**。这套严格的经济回报评估模型，不仅用于近场算力，更被运用于主导超长周期的重磅长约。就在上个月，Anthropic 签署了两笔震撼业界的超级大单： 1. 联合 Google 与博通（Broadcom），锁定了从 2027 年开始交付的 **5 吉瓦（GW）TPU** 算力； 2. 与亚马逊（Amazon）敲定了高达 **5 吉瓦（GW）Trainium** 算力协议。 这两笔交易构成了总额超过 **1000 亿美元** 的长远资本承诺。Krishna Rao 将 Anthropic 的算力架构形容为一层层叠加的“千层蛋糕（layer cake）”：不同交付批次、不同芯片架构、不同代际的硬件在时间线上错落入场，而财务与算力团队必须动态评估每一层蛋糕的性价比演进，确保其在进场时刻能够无缝承接内部研发或外部商业化的特定诉求。 --- ### `25:00` 颗粒度到芯片级的权衡与极端算力消化能力 当深入到微观的算力调度时，企业必须在多重变量之间做出极度精细的权衡。在 TPU（跨越 v5e、v6、v7 代际）、Trainium（Trainium2、Trainium3）以及各代 GPU 这三大芯片平台之间，性价比（每 token 成本）与吞吐速度（吞吐量与时延）各自处在不同曲线上。低成本高吞吐关乎推理经济性，而极高的单 token 速度则能解锁以往不可想象的交互新场景。算力团队会协同全公司各业务单元，将需求细致拆解到极小颗粒度： - 例如强化学习（RL）对 CPU 和特定辅助计算的需求比例是多少？ - 最先进、速度最快的旗舰芯片应当留存多少用于前沿预训练，多少直接切入高价值商业客户的推理池？ 面对主持人提出的极端假设——“如果明天凭空向 Anthropic 空投 2 倍甚至 10 倍的混合算力，公司能否瞬间吞吐？”——Krishna Rao 给出了极其自信的肯定答复。在一两年前，由于异构芯片平台在底层软件栈、驱动适配与操作习性上存在巨大差异，快速消化异构算力极具挑战。然而今天，Anthropic 已经攻克了这一技术壁垒，构建了高度可替代的异构运行环境。即便算力规模成倍空投，公司不仅依然受限于三大方向（前沿训练、内部前沿模型开发、商业客户需求）的强劲胃口，而且能够在极短时间内将几乎任何形态的计算资源瞬间拉起并高效满载，“这正是 Anthropic 当前真正的核心竞争壁垒之一”。 --- ### `27:53` 平台与应用之争：像早期 AWS 那样构建使能层 在 AI 产业中，底层模型公司究竟是该坚守纯粹的使能平台，还是下场收割垂直应用（例如构建类似 Claude 设计工具去对抗 Figma），是业界最具争议的战略命题。Krishna Rao 旗帜鲜明地指出，**Anthropic 绝大部分的精力与业务终局都在于做“平台（Platform）”**。他们坚信平台层虽能捕获巨大价值，但构建在平台之上的客户群必将创造并累积更庞大的生态价值，这与早期的亚马逊 AWS 如出一辙。从底层的模型原始调用，到 Prompt Caching、虚拟环境交互、Claude Agents SDK 以及托管智能体（Managed Agents），这些能力本质上都是赋能外部公司将前沿智能嵌入自身产品的“访问矢量（vectors）”。 然而，Anthropic 也会在特定场景下自研应用，但这必须满足明确的前提逻辑： 1. **预演并牵引模型能力的发展方向**：以 Claude Code 为例，以往市面上的编程工具大多以传统的“开发者思维为主导（developer-led）”，而 Claude Code 是一个完全以“大模型能力为主导（Claude-led）”的范式。在一年前立项时模型能力尚无法完美支撑，但 Anthropic 预见到了模型演进的轨迹，通过自研应用提前下注并最终验证了这一路径。 2. **为整个生态系统打造标杆示范**：通过推出面向特定垂直场景的解决方案（如 Claude for Financial Services 金融服务版、Claude for Life Sciences 生命科学版或网络安全方案），展示底层平台的组合潜力，激发其他企业跟进与模仿。 最关键的是，Anthropic 自身的应用研发团队与外部开发者“完全运行在同一套底层平台之上”。这种公平的竞技场环境，确保了平台绝不会为了偏袒自有应用而牺牲生态伙伴的利益。 ### `30:00` 横向平台定位与化解生态伙伴的“恐惧感” Krishna Rao 明确界定了 Anthropic 的生态定位：公司的核心战略是“横向平台”（mostly horizontal）。虽然 Anthropic 偶尔会做纵向探索（垂直应用），但那仅仅发生在团队确信自身拥有独到见解、能够产生增量价值，或旨在向全行业直观展示平台潜力的场景下。他强调，底层智能大部分的经济价值终将归属于构建在平台之上的客户。Anthropic 的重心在于打造最强的基座模型，并提供让智能能够在客户业务内生根发芽的工具与产品。 针对外界“开发者与客户是否会对 Anthropic 掌控核心智能感到恐惧”的疑问（类比于当年 AWS 对传统 IT 与托管服务商带来的震慑），Krishna 坦承这种敬畏与焦虑是现实存在的。由于前沿模型的能力迭代速度完全打破了历史周期——以往需要 5 年、10 年甚至 20 年的技术浪潮演进，如今被压缩在短短几个月内完成——模型的突破往往连 Anthropic 自身都会感到意外。当新模型或产品发布时，客户自然会产生强烈的震撼与危机感。 为了平衡这种生态焦虑，Anthropic 采取了极度“以合作为导向”（partner-oriented）的做法： - 设立抢先体验计划（Early Access Programs），与核心客户深度共创并倾听他们对功能维度的具体诉求。 - 在安全、电子签名、金融服务等垂直领域发布产品时，均采用联合生态伙伴的协同方式，而非封闭式替代。 - 鼓励客户站在前沿技术的最前线，利用 Anthropic 提供的工具反哺自身业务，从而将恐惧转化为生产力上的加速。 ### `32:42` 稳定的定价哲学与 Opus 降价引爆的“杰文斯悖论” 随着推理算力供不应求、高端芯片（如 H100）租金曲线呈现“微笑曲线”（两头高企），外界常好奇为何头部 AI 实验室不通过激进提价来筛选需求。Krishna 回顾道，Anthropic 成立刚刚五年出头，从收到第一美元商业收入至今仅满三年，而真正意义上的前沿模型（Frontier Model）也是在 2024 年 3 月才正式成型。在这样极度压缩的时间刻度下，Anthropic 的全系模型定价（Haiku、Sonnet、Opus，以及后来的 Mythos）一直保持惊人的稳定性。 唯一一次重大的价格调整发生在发布 Opus 4 时调低了 Opus 家族的单价。其背后的核心动因是团队敏锐地捕捉到了“产能与需求的不匹配”：原本适用顶级 Opus 解决的高难业务，往往被客户迫于成本强行塞给 Sonnet；而 Anthropic 在系统效率优化上取得了重大突破，具备了低成本规模化供给的能力。降价之后，市场迅速呈现出经典的“杰文斯悖论”（Jevons Paradox）——由于定价触达了让企业 ROI 爆发的甜蜜点，客户的实际消耗量出现爆发式非线性增长，整体算力消费不降反升。 ```mermaid flowchart TD accTitle: Pricing Strategy and Jevons Paradox in Opus accDescr: 降价策略如何通过系统工程效率引发杰文斯悖论并促进长线采用 subgraph Efficiency["底层系统工程"] A["推理服务效率大幅优化"] --> B["边际服务成本显著降低"] end subgraph PricingAction["战略定价调整"] B --> C["Opus 4 主动调低价格"] C --> D["匹配企业客户 ROI 甜蜜点"] end subgraph JevonsEffect["杰文斯悖论释放"] D --> E["业务负载不再降级迁就 Sonnet"] E --> F["Opus Token 消耗量超预期剧增"] F --> G["深度嵌入工作流: 升级至 Opus 4.6 维持平价"] end ``` 这种定价机制带来的后续价值是极其稳固的锁定效应：客户将 Opus 深度嵌入其生产工作流后，后续即便升级到性能更强大的 Opus 4.6，Anthropic 依旧维持原价不变，给企业带来确定性的成本预期。 ### `35:26` 算力回报率（Return on Compute）与打破 SaaS 边际成本思维 在衡量商业毛利与商业模式时，Krishna 旗帜鲜明地反驳了用传统软件业（SaaS）的边际变量成本模式来看待前沿大模型公司的做法。传统财务逻辑总试图测算“为单个客户推理提供服务的增量边际成本”，但现实中算力是以巨大包络（Compute Envelope）的形式、按照 12 个月甚至更早前锁定的供给曲线分阶段上线的，无法按需即时弹性扩缩。 因此，Anthropic 在财务顶层设计中引入的核心北极星指标是**算力总支出回报率（Return on Compute Spend Writ Large）**。在这个框架下，所有算力分配都被视为在不同时间维度上对公司营业收入的直接支撑： | 算力分配去向 | 时间兑现周期 | 商业与战略收益价值 | |:--- |:--- |:--- | | **客户推理服务（Inference）** | 即时 / 当期（Today） | 直接兑现当期账面营收与现金流 | | **前沿模型研发（Model Dev）** | 中期（如 6 个月后） | 解锁全新模型能力，开拓全新潜在市场空间（TAM） | | **内部工具与应用加速（Internal Ops）** | 持续演进 | 赋能内部生产力，加速新产品（如 Claude Code）落地 | 在整个算力包络受限的情况下，第一季度营收能够以爆发式曲线增长，依靠的绝非短期内堆叠新芯片，而是通过算法及基础设施的优化，从固定的算力底池中榨取更高的系统回报。 ### `37:22` 超越采购：三大云厂商与芯片硬件的深度联合设计 作为业内唯一一家横跨三大公有云（AWS、Google Cloud、Microsoft Azure）并通吃三大 AI 芯片架构（AWS Trainium、Google TPU、NVIDIA/Broadcom GPU）的前沿实验室，Anthropic 与算力硬件供应商之间的关系早已超越了常规的商业采购或机架租赁。 Krishna 强调这种共生体系是全方位、深层次的： - **深度协同研发**：以 AWS 为例，Anthropic 的工程团队与亚马逊旗下的 Annapurna Labs 团队深度嵌合，双方在底层硬件架构与编译器优化层面高度协作，全力优化 Trainium 芯片的计算效率。 - **协同容量规划**：所有算力上线并非简单下订单，而是与各大超大规模云服务商（Hyperscalers）共同进行多季度的算力容量落地演练与系统化规划。 - **双向分发网络**：三大云平台既是算力供应源，又是极其强大的客户分发渠道；与此同时，Anthropic 自身的一方平台（First-party platform）也保持着强劲的独立直销增长。 ### `38:32` 财务团队的“吃狗粮”实践：从 Claude Code 到 70+ 个专属 Skill 在 Anthropic 内部，最前沿的 AI 工具并不是实验室里的玩具，而是被全面注入到支持部门的日常运转中。约一年前，Anthropic 财务团队就开始日常使用 Claude Code。Krishna 回忆，最初大家只是抱着尝试“氛围编码”（vibe coding）的心态，但很快发现 Claude Code 的智能推理本质上可以充当一个全能的“数字同事”（Digital Coworker），这一早期内部实践甚至直接反哺了后续代理型产品“Coworker”的概念诞生。 如今，Anthropic 财务部的日常运作已经实现了高度的 Agent 化与自动化： - **法定财务报表生成**：公司旗下所有法定实体（Legal Entities）的法定财务报表已全面交由 Claude 自动化编制生成，最后仅由人类审计与财务人员负责终审复核。 - **实时数据分析平台 AntStats**：过去需要财务分析师耗费大量时间手工筛查海量数据、总结业务动因并撰写日常分析备忘录的流程，现已实现自动化解析与报告推送。 - **公共财务技能库（Skills Repository）**：团队沉淀并维护了一套专属于财务领域的 Claude 技能库，目前内部已有超过 70 个开箱即用的专业技能，供全员按需调用以驱动复杂财务分析。 ### `40:00` 财务团队的“超级用户”文化与 MFR 实践 在 Anthropic 内部，财务团队深度融入了 Claude 的原生工作流。团队基于模型构建了一套专门的月度财务审查（MFR, Monthly Financial Review）技能，能够自动生成成熟度达 90% 到 95% 的初稿。这项变革彻底重塑了财务例会的讨论焦点：团队不再把大把时间浪费在核对数字、单纯充当“播报天气预报”的角色上，而是转由 Claude 深入挖掘数据波动背后的核心驱动因素（Drivers），直接分析“事情为何发生”以及“对战略有何启示”。以往需要耗费数小时才能整理出的每周营收动因或算力利用率报表，现在压缩到 30 分钟内即可完成，业务领导层也能以更快的节奏获得洞察支撑。 Krishna Rao 透露，他的管理看板上甚至维护着一个内部 Token 消耗排行榜。虽然公司不会基于 Token 使用量进行绩效考核，但排行榜呈现出一个极具启发性的现象：团队中最顶尖的“Token 大户”并非刚入职、带着周末编程习惯的 22 岁年轻技术工程师，而是那些经验丰富的资深财务主管。其中位居榜首的是税务主管，他专注于构建自动化税务政策引擎，将税务团队中大量繁重复杂的合规流程彻底智能化。Krishna 强调：“如果我们自己都不能成为这款产品的超级用户、不去主动探寻它的能力边界，又凭什么指望客户能做到这一点？” - **MFR 生成效率**：初稿自动化完成度达 90%~95%，业务报告产出时间从数小时骤降至 30 分钟。 - **分析维度的转变**：从机械式的事后汇报升级为探究归因与制定业务行动方案（"What do we do? What are the implications?"）。 - **资深员工的驱动**：税务主管等高年资领导者成为内部 Token 消耗冠军，主导高价值垂直任务的深度自动化。 --- ### `41:32` 人类被 AI 支配的隐忧，与劳动力领域的“杰文斯悖论” 针对主持人提出的隐忧——随着 AI 成为更优异的统筹者与全局优化器，人类是否在一步步沦为“听从算法指令行事”的附庸，甚至带有一丝反乌托邦色彩，Krishna 给出了完全不同的乐观解读。他借用了经济学中的“杰文斯悖论”（Jevons Paradox）来阐释劳动力市场的演进：当高素质人才因 AI 赋能而实现生产力飞跃时，企业面临的工作需求不但没有枯竭，反而在无限膨胀，Anthropic 因而招聘了更多优秀人才。 在传统企业运作中，财务人员的大部分精力都被困在月末结账或跨表格核对数字（Tie out a number）等琐碎泥潭中。而当 Claude 接管了这些低效摩擦后，团队得以将认知带宽转移到高维度的思考上：如何将资金进行战略性再投资？如何更动态地调配稀缺的算力与资本资源？在 Krishna 看来，AI 是生产力的强效催化剂，不仅没有弱化人类的能动性，反而在整个组织甚至各行各业的技能爬坡期内，成倍释放了优质人力的产出上限。 --- ### `43:08` 融资历程：从 FTX 阴影、DeepSeek 冲击到指数级认知重塑 回顾资本运作历程，Krishna 在两年前 Anthropic 进行 Series D 轮融资时加盟。那次融资过程充满波折：前沿模型刚刚推向市场，又恰逢 FTX 破产清算其持有的 Anthropic 股份，二级市场抛压严重。彼时外部投资人充斥着误解与质疑：为何非要投入巨资研发前沿模型？追逐模型智能的商业回报究竟在哪里？甚至有人认为“AI 安全使命”与“打造一家商业巨头”本质上相互背叛；还有投资人死搬传统企业软件的框架，质疑 Anthropic 的销售团队规模为何如此之小。 到了 2024 年底进行 Series E 轮融资时，公司的年化运行营收（Run Rate Revenue）已逼近 10 亿美元大关。然而首批交割当天，DeepSeek 的发布消息突然席卷市场，引发了投资者对 AI 底层逻辑与估值模型的剧烈震荡与重估。即便 Anthropic 如期完成了巨额融资，许多机构在审视预测模型时依然难以置信，断言打破“物理定律”的超高速增长难以为继，甚至搬出云计算数十年缓慢渗透本地部署（On-prem）的历史，论证企业级 AI 的落地周期必将极为漫长。 ```mermaid flowchart TD accTitle: Anthropic 融资与关键里程碑演进 accDescr: 展示从 Series D 到 Series E 期间 Anthropic 面临的外部挑战与认知打破过程。 subgraph D_Round["Series D 阶段（约两年前）"] D1["FTX 清算股权抛压"] --> D2["投资人质疑：安全与商业矛盾？"] D3["前沿模型初现，商业模式待验"] --> D2 end subgraph Growth["突破拐点"] G1["ARR 快速攀升至 10 亿美元"] end subgraph E_Round["Series E 阶段（2024 年底）"] E1["交割首日突遇 DeepSeek 消息冲击"] --> E2["市场重构 AI 估值逻辑与确定性"] E3["投资人质疑：大数定律与企业采纳慢"] --> E2 end D_Round --> Growth --> E_Round ``` --- ### `45:00` 安全研究的双重飞轮与企业客户的信任护城河 面对外部的普遍犹疑，Anthropic 凭借实打实的商业表现印证了核心论点：追求前沿智能具有极高的经济回报，驱动增长的核心引擎是“模型驱动增长”（Model-led growth），并由产品体验、渠道分发与 Go-to-Market 团队协同赋能。更深层的战略红利在于，Anthropic 最初出于使命驱动而重注的 AI 安全研究，在商业化端产生了意想不到的技术外溢与飞轮效应。 这种双重飞轮具体体现在研发与客户信任两个维度： 1. **技术研发反哺**：Anthropic 开创了可解释性研究（Interpretability），相当于为神经网络注入了“核磁共振（MRI）”扫描能力，能清晰透视模型内部的运算机理；同时开创了对齐科学（Alignment science），确保模型严格遵循指令而不偏移。实践证明，能够看清模型内部机理的团队，反而在构建更强前沿模型时拥有更高的工程上限。 2. **企业级信任壁垒**：如今《财富》前 10 强（Fortune 10）企业中已有 9 家采用 Claude。企业将极其敏感的客户信息、核心业务数据、员工协作乃至客户交互工作流托管给 Claude 平台，对系统可靠性与企业可信度有着苛刻的要求。Anthropic 在安全、可解释性与对齐领域的底蕴，直接转化为了最坚固的企业级护城河。 Krishna 补充道，自他加入公司以来已完成了数百亿美元规模的庞大资本储备（包括来自 Amazon 与 Google 的后续资金注入承诺）。在资本高度密集型的前提下，Anthropic 的业务运营保持着极高的资金利用效率；公司大规模储备弹药的动机，绝非为了填补日常亏损，而是为了在面对激烈的“不确定性圆锥”（Cone of Uncertainty）时，拥有充足的底牌抢占战略先机。 --- ### `47:31` 从线性思维到指数级跃迁：Dario 的超前预测力 谈及对 10 倍速增长的心理预期，Krishna 坦言自己最初经历了一场剧烈的认知洗礼。在他刚加入时，公司的 ARR 约为 2.5 亿美元，内部目标是冲击 10 亿美元。当时作为传统财务背景出身的 Krishna，脱口而出的第一个问题却是线性思维式的：“太棒了，我们打算在哪一年实现？” 事实证明，联合创始人兼 CEO Dario Amodei 对营收跃升的预判能力始终比财务团队更加精准敏锐。传统商业逻辑所笃信的物理定律、大数定律以及对企业采购周期的陈旧认知，在强大的多层指数级（Exponential）底盘面前被逐一击碎。Krishna 的核心认知也彻底完成了范式转换：在财务预测上依然保持底层的审慎与多情景推演，但在战略视野上彻底摒弃线性增量逻辑，全然拥抱并押注指数级增长带来的颠覆性可能。 - **线性与指数的碰撞**：面对 10 亿美元目标，传统直觉往往按年度做增量规划，而底层技术裂变直接推动了跨越式达成。 - **复合指数效应**：表面上的营收指数级爆发，背后是由模型能力跃升、推理成本优化与场景渗透等多重底层指数曲线共同托举的结果。 --- ### `49:02` 投资人最难理解的概念：算力的极致可互换性（Fungibility） 当被问及当前阶段与投资人沟通时最难以解释的痛点是什么，Krishna 明确指出：依然是**算力作为核心资产的可互换性（Fungibility）范式**。大部分投资人依然习惯性地把算力看作是某段时期内的普通可变成本，或者是固定死板的硬件开销，无法理解其在企业内部的动态流转价值。 在传统软件公司或制造业工厂中，成本结构泾渭分明——研发（R&D）部门的工程师薪酬属于研发费用，你不可能在当天下午把他们突然变成产品制造费用或直接营业成本（COGS）。然而在 Anthropic 的架构下，算力芯片具备极致的弹性： - **白天的峰值时段**：同一批计算芯片全力支持高并发推理（Inference），响应企业客户的实际业务请求，转化为商业收入； - **下午或夜间低谷**：算力瞬间无缝切入研发模式，直接用于前沿新模型的训练开发（Model Development），转化为研发资产。 正是这种打破传统 R&D 与 COGS 静态壁垒的极致可互换性，决定了算力回报率（Return on Compute）在 Anthropic 的商业闭环中占据着无可替代的枢纽地位，也构成了它区别于传统软件企业的本质特征。 ### `50:00` 投资人视角下的算力 ROI 与客户真实价值 如果让 Krishna Rao 换位成为顶尖投资机构的投资人去考察各家前沿大模型实验室，他会直击最核心的三大疑问。首先是全口径算力的 ROI（投资回报率）：各家模型公司正在进行史上规模前所未有的巨额资本投入，这些算力究竟是如何被有效利用的？当下能看到怎样的回报？长期回报又将在何时、以何种形态释放？其次，也是客户层面的真实商业闭环：企业客户究竟只是在试水 Demo，还是已经进入了具有规模经济效益的实质性部署阶段？ 就 Anthropic 自身的业务数据而言，客户价值正在迅速兑现。Anthropic 的年化净金额留存率（Net Dollar Retention Rate）已经超过 500%，财富 10 强企业中有 9 家在使用 Anthropic 的产品与服务。Krishna 分享了一个极具现场感的生动细节：就在前往本次播客录制现场的 20 分钟 Uber 车程中，他在手机上直接签署了两份千万美元级别的企业大单。全球最成熟、要求最苛刻的大型企业和顶尖初创公司在拥有众多选择的情况下，依然坚定选择了 Anthropic。第三个核心考量则是长期的算力供应格局：上游算力供应商同时也是潜在的销售商甚至内部使用者，随着时间推移，该如何平衡算力采购来源，保持灵活度并与头部算力巨头紧密合作。 - **三大投资人审视维度**：全口径算力 ROI 与资本回收形态、企业客户的真实采购闭环与落地规模、未来算力供应链的灵活性与平衡策略。 - **关键财务指标**：年化净金额留存率（NDR）突破 500%，财富 10 强企业中 9 家为其实际客户。 - **签约速度与规模**：在赶往录音室的 20 分钟车程内连签两笔千万美元级别的长期客户承诺（Commitment）。 --- ### `52:15` 破除公众对 AI 的疏离与恐惧感 面对“大众对 AI 概念的好感度甚至低于美国国会”这一惊人现象，Krishna 剖析了科技界与普通大众之间的脱节。过去无论是工业革命、互联网还是云计算，社会变革都有较长的适应周期；而 AI 技术将几十年乃至上百年的技术进步压缩在短短几个月内发生，人类习惯于线性思维的本能很难适应这种指数级的狂飙，从而容易引发内心的疏离与抗拒。要改变这种公共认知，AI 从业者必须在展现宏大愿景的同时保持严谨与坦诚。 一方面，科技界需要更具体地描绘技术带来的福祉，正如 Dario Amodei 在长文《Machines of Love and Grace》中所阐述的：利用 AI 彻底重塑药物研发流程，攻克主流甚至罕见疾病；提升医疗交付水准并让全球资源匮乏的发展中地区普遍受益，确保经济成果惠及全人类。另一方面，必须摒弃“盲目乐观主义”，不能只报喜不报忧。如果只宣传美好未来而对风险闭口不谈，公众自然不会建立信任。行业必须诚实评估可能遇到的颠簸与阵痛，携手商业机构与政府共同摸索解决方案。 ```mermaid flowchart TD accTitle: AI 公共沟通与信任重建机制 accDescr: 展示了如何通过愿景落地与风险诚实披露双轮驱动，建立社会与公众对 AI 技术的信任。 subgraph Problem["公众认知困境"] A["指数级变革压缩在数月内"] --> B["大众线性思维产生脱节与恐惧"] end subgraph Solution["双轨重建公众信任"] C["具象化展现技术红利<br/>(医疗、罕见病攻克、普惠全球)"] D["坦诚披露风险与颠簸<br/>(非盲目乐观、兼顾政府与产业协同)"] end Problem --> Solution Solution --> E["建立真正长期的公众信任"] ``` --- ### `55:38` Mythos 的震撼登场与分阶段安全发布 当被问及此前让众多资深观察者感到“有些后背发凉”的模型 Mythos 时，Krishna 揭秘了这款高危高能模型背后的内部决策考量。很多人误以为 Mythos 仅仅是一款专精的网络安全模型，但实际上它在诸多维度上都具备极其强悍的通用智能，只是在网络安全（Cyber）能力上出现了极其惊人的尖峰突破。在内部测试中，针对同一个开源代码库，先前的顶尖模型只扫出了 22 处安全漏洞，而 Mythos 一口气揪出了 250 处漏洞。 正是这种令人敬畏的强大能力，促使 Anthropic 放弃了直接公开全量上线的做法，而是确立了一套分阶段发布的全新模板。团队并没有因噎废食地雪藏这项突破，而是选择将其优先开放给一个逐步扩大的受控群体，聚焦于防守型代码修复与系统防护，确保该技术发挥防御性价值而非被滥用于攻击。Mythos 成为了前沿大模型在安全可控与能力释放之间取得平衡的典范样本。 - **安全能力质的飞跃**：在同一开源代码库测试中，前代模型检出 22 处漏洞，Mythos 检出 250 处漏洞。 - **发布机制创新**：确立分阶段（Phased approach）的发布模板，从受控测试逐步扩展，坚决将尖端黑客攻防能力引导至防御与补丁修复方向。 --- ### `57:31` 政府监管与“美国优先”的技术担当 随着 Anthropic 的体量与影响力跃升至国家关键基础设施级别，公司不可避免地直面政府监管与地缘政策的审视。面对诸如大模型上线前是否需要政府前置审批、以及国防等敏感场景的互动，Krishna 强调 Anthropic 一直将政府沟通与合规治理置于战略高度。公司明确秉持“美国优先（America first）”的价值导向，致力于让前沿技术服务于美国本土及其全球民主盟友的战略优势。 在 Krishna 看来，监管不仅不是创新的绊脚石，反而是确保技术安全落地的必要规则。关键在于如何在高速创新与严谨的责任框架之间取得平衡。大模型的技术外溢效应深远，科技实验室无法单方面解决所有社会层面的挑战，必须与政府管理机构保持透明、务实的常态化沟通机制。 --- ### `58:51` 极度务实与谦逊的 Anthropic 内部文化 谈及如何向父母解释 Anthropic 的文化内核，Krishna 指出，公司最神奇的地方在于 7 位联合创始人在纸面上看似庞杂难以协调的架构，在实际运转中却极为顺畅和谐，为整个组织注入了谦逊与纯粹的基因。Anthropic 拥有极其严苛的“文化面试”环节，这不是走走形式的打勾打叉，而是拥有绝对的一票否决权：即便应聘者在业务与技术能力上是面试官见过最聪明的人，但如果无法通过文化维度的检验，也绝不录用。 这种文化集中体现为“极度协作”与“极度谦逊”。公司内部绝不容忍拉帮结派建立各自的“独立王国（Fiefdoms）”，也杜绝争抢功劳与排挤他人的“锋利手肘（Sharp elbows）”。Anthropic 员工的笔记本电脑上都贴着一张标志性的贴纸，上面印着一句座右铭：**“我们的对手极其能打，成功绝非必然（Our competitors are incredibly capable and success is far from guaranteed）”**。因此，每当团队取得重大里程碑或突破时，办公室里从不会有人撒彩带开香槟庆祝，大家的标准反应永远是平静地问一句：“下一个攻坚目标是什么？”这种高度对齐使命、淡化个人英雄主义的理性氛围，构成了 Anthropic 区别于硅谷其他巨头的最鲜明特质。 ### `01:00:00` 深度辩论、高度透明与人才护城河 在谈及 Anthropic 的内部文化时，克里希纳·拉奥（Krishna Rao）强调，这种文化深植于公司的方方面面。团队内部保持着极其严谨的辩论、求知的开放性以及智识上的诚实（intellectual honesty）。员工们不会隐藏观点，而是开诚布公地各抒己见，并展开富有建设性的对话，最终确定前行路线。一旦达成决定，全员便会产生真正的思想与行动统一（alignment）。以算力分配为例，即便大家对资源应当投向何处持有不同意见，彼此也会理性探讨哪里的回报率最高；而一旦做出决策，所有人都会全力以赴，绝不会出现事后猜忌、推诿、办公室政治或划地为牢的“领地意识”（fiefdoms）。 这种文化的另一大支柱是极高透明度。首席执行官达里奥·阿莫代伊（Dario Amodei）每两周就会站到全员大会前。他通常会提前写好一份简明备忘录，分享三到四个主题，随后直接接受全员的公开提问。这些问题绝不是走过场的“软球”（softballs），也不是提前安排好的托儿，而是大家心中最真实、最尖锐的疑虑。达里奥会竭尽全力坦率回应。这个全员会虽然不是决策机构，却为全体员工提供了一个了解领导层思维方式的窗口。这种文化让团队凝聚力极强：Anthropic 的 7 位联合创始人至今无一人离职，前 20 至 30 名早期核心员工绝大多数也仍然留在公司。 当 Meta 等巨头开出令人咋舌的天价薪酬包、从各大前沿大模型实验室疯狂挖人时，其他实验室流失了数十名顶尖技术专家，而 Anthropic 仅损失了两人。针对研究员为何愿意留下，克里希纳指出，其核心在于“人才密度（talent density）远重于人才数量（talent mass）”。研究人员希望最大化自身的影响力，渴望在一个协作、透明、无需为争取资源而勾心斗角的求真环境中工作。更重要的是，他们被公司的使命深深吸引：以负责任的方式开发能够改变世界的颠覆性技术。Anthropic 倡导“向上竞争（race to the top）”，希望通过树立高标准的行业榜样，让全行业都能更好、更安全地推进 AI 发展，这一愿景对顶级人才具有无法抗拒的号召力。 - **顶尖人才留存数据**：在行业面临巨头大规模高薪挖角的狂潮下，其他同行实验室流失了数十人，Anthropic 仅流失了 2 人；初创团队的 7 位联合创始人全员在岗，前 20-30 名员工绝大部分依然在职。 - **治理与沟通机制**：达里奥每两周撰写备忘录并召开全员会，无预设、无筛选地直面员工最尖锐的真实提问。 - **人才哲学**：坚信人才密度胜过人才体量，团队成员追求自身影响力的最大化，反感政治内耗与黑箱决策。 --- ### `01:03:48` 下一个前沿边界：价值 40 万亿美元知识工作的“虚拟协作者” 展望下一阶段的 AI 前沿，克里希纳认为其突破口不仅限于基础模型能力的演进，更在于重构整个经济体中知识工作的生产力。全球每年的知识工作总价值高达约 40 万亿美元，Anthropic 将未来的核心产品形态定义为**“虚拟协作者”（Virtual Collaborator）**。这种协作者远不仅是一个通用意义上聪明的模型，而是与企业特定业务场景深度共生、具备多维智能的综合体。 虚拟协作者的核心在于具备企业内部的全局上下文，能无缝调用组织自研或外部采购的所有工具链，拥有长期记忆，并且能从人类以及它自身过往的错误中持续学习。它不再只是帮人完成单点的小任务，而是能在极长的时间跨度内独立推进一个完整的业务构想。在编码领域，Claude Code 已经率先展现出这一范式的雏形，而新推出的协同工具 Cowork 的采用速度在同一时间基准下甚至超越了 Claude Code。在 Anthropic 内部，产品研发方式已经被彻底颠覆：不再由 1 位产品经理带着 2 名工程师耗费 3 个月才发布一个版本，而是由遍布全公司的庞大 Agent 舰队承接各项任务，每天都在高频迭代上线，这也意味着“每位员工都在某种意义上变成了管理者”。 ```mermaid flowchart TD accTitle: Virtual Collaborator Architecture accDescr: Diagram illustrating how context, memory, and tools combine to transform 40 trillion dollars of knowledge work into a fleet of agents shipping daily. subgraph Inputs["核心能力要素"] A["企业上下文<br/>(Context)"] B["自研与外部工具链<br/>(Tool Use)"] C["长期记忆与反思<br/>(Memory & Reflection)"] end subgraph Agent["虚拟协作者 (Virtual Collaborator)"] D["多维场景智能<br/>长周期推进复杂任务与构想"] end subgraph Applications["生产力重塑 (40万亿美元市场)"] E["Claude Code<br/>(自动化代码构建)"] F["Cowork 协同套件<br/>(极速渗透业务流)"] G["Agent 舰队日常迭代<br/>(人人成为管理者)"] end Inputs --> Agent Agent --> Applications ``` - **市场空间**：全球知识工作年产值约 40 万亿美元，虚拟协作者旨在通过多维智能与全新交互形态全面加速这一流程。 - **虚拟协作者的四大特质**：理解企业级专属上下文、调用原生/采购的各类工具、具备自省纠错与长期记忆、支持长周期复杂构思与任务落地。 - **产品增长对比**：Cowork 在发布同期的增长速度已经超越了极速普及的 Claude Code。 - **内部研发范式变革**：告别“1名 PM + 2名工程师耗时3个月”的传统软件迭代周期，转由 Agent 集群驱动每日持续交付。 --- ### `01:05:40` 与指数级跃迁同行：CFO 的认知重塑与第一性原理 伴随 Anthropic 超常规的爆发式扩张，克里希纳回顾了自身如何在心理与管理能力上应对前所未有的指数级演进。他坦言这极其艰难，唯有彻底打破既有成见（priors），回归第一性原理思考。他特别提到了 2024 年初自己刚加入公司前与首席算力官汤姆·布朗（Tom Brown）的一次深谈。当时汤姆作为面试官之一带他散步，一路将他送回了家。在漫长的散步中，汤姆向他描绘了未来算力基础设施的宏大蓝图以及模型将在极短时间内达成的能力飞跃。 克里希纳回忆道，汤姆描述的世界在当时听起来简直如同科幻小说。回到家后，他震撼地对妻子说：“这太疯狂了。如果他说的哪怕只有 10% 变成现实，它都将打破我和绝大多数人过去所见过的一切思维范式。”事实证明，汤姆当时所预言的技术跃迁绝大多数都已成为眼前的现实。面对业务如此巨大的暴露面，克里希纳将自己的管理方式定义为寻找平等的“合伙人”而非居高临下的下属。凭借过去在黑石集团（Blackstone）私募股权投资部门严苛的颗粒度训练，以及在疫情期间协助 Airbnb 完成在 7 周内收入暴跌 70% 时的紧急融资历练，他学会了在没有历史先例、无模板可依的极端动态环境中保持清晰与敏锐。 在个人生活层面，面对高强度工作带来的巨大吞噬，克里希纳坚持每周至少给自己留出一小段独处的安静时刻——无论是在深夜的车里还是片刻的宁静中——去由衷地感恩与惊叹：“能在这个时代的关键节点，与这样一群杰出的人一起在这样一家公司攻克这个问题，这是何等不可思议的机遇。”他认为，正是这种敬畏之心与汤姆·布朗等技术先驱身上展现出的深刻乐观主义，成为了支撑他不断拓展认知边界的力量源泉。 - **早期的思想震荡**：2024 年初与汤姆·布朗的散步深谈奠定了认知基础；汤姆勾勒的超大规模算力规划与模型演进速度曾看似科幻，如今已基本成真。 - **合伙人式领导风格**：不把团队视作下级执行者，而是鼓励白板推演与直接反对意见的平级合伙人；克里希纳指出，在如此庞大的业务复杂度面前，管理者既不能始终漂浮在 50,000 英尺的高空，也无法独自事无巨细地下沉到 500 英尺，必须依靠顶级合伙人共同分担。 - **历史极端案例的映射**：曾参与主导 Airbnb 疫情爆发期面临 70% 营收骤降时的关键融资，积累了在缺乏模板和先例的未知领域中独立研判的定力。 - **心态调节机制**：每周固定留出几分钟静默反思，保持对当下参与历史级科技变革的敬畏感与欣赏之心。 ### `01:10:00` 不确定性锥的下行风险与潜在瓶颈 在 Anthropic 内部，团队经常用“兼抱光明与阴影（holding light and shade）”来形容审视未来的视角。尽管过去公司始终运行在“不确定性锥（cone of uncertainty）”的高端预测轨道上，但主持人提出了一项事前面检（premortem）式的设问：如果一年后回看，发现实际所需的算力远低于预期，究竟是哪些因素导致轨迹跌落至不确定性锥的下界？ Krishna Rao 拆解了可能导致发展减速的三大核心瓶颈： 首先，模型能力的提升目前远快于企业应用落地速度，二者处于“应用追赶模型”的状态。真实世界中，大型组织内的员工受制于长期固化的工作流、工具生态与操作惯性，“变革是极其艰难的（change is hard）”。一旦新技术在企业内部的渗透与扩散撞上阻力墙或发生放缓，营收增长的斜率就会直接受到冲击。 其次，是底层扩展定律（scaling laws）的持续性问题。尽管目前 Anthropic 完全没有看到扩展定律失效的迹象，且深信这一技术演进曲线，但在前沿科学探索中断言百分之百的确定性是极不理性的，模型能力阶段性见顶依然是不可忽视的潜在风险。 最后，是前沿竞争地位的不确定性。Anthropic 目前处在前沿位置并正在定义智能体 AI（agentic AI）的演进方向，但前沿地位绝非高枕无忧。这是一个极其残酷的竞争市场，保持领先需要持续压注巨量技术、算力以及渠道资源，任何执行上的偏差都可能导致掉队。 ```mermaid flowchart TD accTitle: 影响算力需求与增长的潜在瓶颈 accDescr: 展示可能导致 Anthropic 掉入不确定性锥下界的三个核心风险点。 subgraph Risks ["可能跌入锥体下界的风险要素"] R1["企业工作流扩散受阻<br/>(组织惯性与变革成本高)"] R2["扩展定律撞墙<br/>(模型能力增长遭遇瓶颈)"] R3["前沿竞争失守<br/>(技术、算力与商业化掉队)"] end Risks --> Impact["算力需求不及预期 / 收入增长放缓"] ``` --- ### `01:11:32` 前沿智能的最美愿景：生命科学与药物研发的质变 跳出下行风险，当谈及身处见证未来的前沿席位上最令人振奋的前景时，Krishna 毫不犹豫地指向了生物技术与医疗健康领域。他描绘了一个正在逼近的现实：未来人们若被确诊某种目前无法治愈的绝症，在患者的有生之年里，AI 能够以远超以往的速度找到解药，从而免于死亡。 目前的落地主要集中在提升药物研发中后期的流转效率，例如自动处理海量合规文书、生成临床研究报告等行政与文档密集型环节，Claude 的方案已在显著压缩这些流程的耗时。但真正令 Krishna 感到心潮澎湃的，是技术向药物发现与早期研发源头的深层渗透： 分子与蛋白质的内部结构极端复杂，微观维度的极细微变动就会对宏观药理结果产生颠覆性影响。人类科研人员受制于体能与算力很难穷尽探索，而前沿 AI 恰恰在这个复杂高维空间中具备降维打击式的优势。一旦实验室的吞吐量因 AI 的介入提升 10 倍乃至 100 倍，科研人员便能以几何级倍数并发推进实验，更低成本、更快产出高精度结果。这种突破将不再局限于少数特定罕见病，而是全面辐射整个人类疾病谱系，彻底重塑人类生存与互动的根本方式。 - **行政与临床文书加速**：通过自动化临床试验报告及监管文书处理，大幅缩减新药进入市场的周期。 - **分子与蛋白质机理研究**：针对细微改变即可导致巨大结果差异的高维生化结构，AI 展现出远超传统人工试错的预测精度。 - **研发吞吐量数量级跃升**：推动湿实验与干实验结合的研发通量提升 10x 到 100x，实现多管线快速验证。 --- ### `01:13:30` 最善意之举：兄长为我留下的“选择权” 访谈进入节目的标志性收尾环节，主持人询问 Krishna 这辈子他人为他做过最友善的一件事是什么。Krishna 分享了一段此前深埋多年、甚至需要他后来主动追问才从兄长口中得知的往事。 他的哥哥比他大五岁半。当年全家生活在加利福尼亚州，家境是典型的稳健中产阶级。那是大约二十五到三十年前，大学的奖学金与助学金体系远不如当下这般优厚。哥哥在升学时展现出顶尖的才华，拿到了所有申请名校的录取通知，且立志毕业后攻读医学院。然而，出乎意料的是，哥哥最终选择留在加州就读本州的州立大学。 直到很多年后 Krishna 才得知真相：哥哥当时做这个决定的核心考量之一，是希望把家庭的经济资源沉淀下来，留给当时只有十二三岁、距离读大学还有六年的弟弟。"很大一部分原因是为了确保我未来能拥有选择的权利。" 兄长在自己的人生关键节点选择退让，只为给弟弟一个完全不被经济条件绑架、可以自由挑选理想大学的未来。这份跨越时光与无私牺牲的亲情，成为 Krishna 至今仍深深铭记在心的最善之举。 ## Notable quotes > `02:29` 它是构筑一切的底层画布（the canvas on which everything else gets built） — Krishna Rao > `02:29` 买太多算力会破产，买太少算力则无法服务客户且丧失前沿地位。 — Krishna Rao > `09:08` 前沿智能的投资回报率（returns to frontier intelligence）正在变得越来越高 — Krishna Rao > `18:46` 人才密度远胜于人才规模（Talent density beats talent mass） — Krishna Rao > `20:00` 直接按照你的终态去设计产品，因为我们在研发端一定会随着时间把这项能力补齐。 — Krishna Rao > `40:00` 如果我们自己都不能成为这款产品的超级用户、不去主动探寻它的能力边界，又凭什么指望客户能做到这一点？ — Krishna Rao > `47:31` 太棒了，我们打算在哪一年实现？ — Krishna Rao > `58:51` 我们的对手极其能打，成功绝非必然（Our competitors are incredibly capable and success is far from guaranteed） — Krishna Rao > `58:51` 下一个攻坚目标是什么？ — Krishna Rao > `01:03:48` 每位员工都在某种意义上变成了管理者 — Krishna Rao > `01:05:40` 这太疯狂了。如果他说的哪怕只有 10% 变成现实，它都将打破我和绝大多数人过去所见过的一切思维范式。 — Krishna Rao > `01:05:40` 能在这个时代的关键节点，与这样一群杰出的人一起在这样一家公司攻克这个问题，这是何等不可思议的机遇。 — Krishna Rao > `01:10:00` 变革是极其艰难的（change is hard） — Krishna Rao > `01:13:30` 很大一部分原因是为了确保我未来能拥有选择的权利。 — Krishna Rao ## Mentioned Krishna Rao, Patrick O'Shaughnessy, Dario Amodei, Tom Brown, Anthropic, Claude, Claude Code, Claude 3.5 Sonnet, Opus 4, Opus 4.5, Opus 4.6, Opus 4.7, Haiku, Mythos, Cowork, AntStats, Trainium, TPU, GPU, Amazon, AWS, Annapurna Labs, Google, Google Cloud, Broadcom, NVIDIA, Microsoft Azure, xAI, Colossus, DeepSeek, FTX, Meta, Blackstone, Airbnb, Machines of Love and Grace, Invest Like the Best ## Video description My guest today is Krishna Rao, the CFO of Anthropic. The center of our conversation is how he navigates the decision around procuring and allocating compute, which he describes as the canvas on which everything else gets built. We talk about what he calls the cone of uncertainty, the three chip platforms Anthropic uses fungibly across Trainium, TPUs, and GPUs, and the daily meetings they run to allocate compute between model development, internal use, and serving customer demand. He explains why the returns to frontier intelligence keep getting higher, especially in enterprise, and how Anthropic thinks about the line between platform and application and why they choose to build their own products like Claude Code. Krishna has such a unique seat watching one of the fastest growing businesses in history, and he is generous in sharing what he has learned since joining the company two years ago. For the full show notes, transcript, and links to mentioned content, check out the episode page ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠here⁠⁠⁠⁠⁠.  ----- Become a Colossus member to get our quarterly print magazine and private audio experience, including exclusive profiles and early access to select episodes. Subscribe at ⁠colossus.com/subscribe⁠. ----- ⁠Ramp’s⁠ mission is to help companies manage their spend in a way that reduces expenses and frees up time for teams to work on more valuable projects. Go to⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ ⁠ramp.com/invest⁠⁠ to sign up for free and get a $250 welcome bonus. ----- Trusted by thousands of businesses, ⁠Vanta⁠ continuously monitors your security posture and streamlines audits so you can win enterprise deals and build customer trust without the traditional overhead. Invest Like the Best listeners get a special offer of $1,000 off Vanta when you go to ⁠vanta.com/invest⁠.  ----- WorkOS⁠ is the infrastructure B2B and AI-native companies use to sell to enterprise. It covers everything enterprise security requires: SSO, SCIM, RBAC, Audit Logs, AI governance, and more. Trusted by 2,000+ fast-growing companies, including OpenAI, Anthropic, Cursor, and Vercel. ----- Rogo is the AI platform for finance. They're building agents for Wall Street that are trained to understand how bankers and investors actually do work: from diligence and modeling, to turning analysis into deliverables. To learn more, visit rogo.ai/invest. ----- ⁠Ridgeline⁠ has built a complete, real-time, modern operating system for investment managers. It handles trading, portfolio management, compliance, customer reporting, and much more through an all-in-one real-time cloud platform. Visit⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ridgelineapps.com⁠. ----- Editing and post-production work for this episode was provided by The Podcast Consultant (⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://thepodcastconsultant.com⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠). Timestamps: (00:00:00) Welcome to Invest Like The Best (00:02:29) Episode Intro: Krishna Rao (00:03:14) Compute as Anthropic's Lifeblood (00:05:17) Three Fungible Chip Platforms (00:07:31) The Cone of Uncertainty (00:09:08) Competing Ways to Allocate Compute (00:10:36) What Drives Compute Efficiency (00:12:38) Why Frontier Returns Are So High (00:16:32) How Claude Code Writes Its Own Code (00:18:46) Will Talent Become Obsolete? (00:20:07) How Scaling Laws Are Holding (00:21:54) Exponential Thinking (00:23:17) The Layer Cake of Compute (00:26:36) How Anthropic Deploys New Compute (00:27:53) Platform v. Application Layer (00:32:42) Why Model Pricing Has Stayed Stable (00:35:26) Measuring Return on Compute (00:37:22) Working With Chip Providers (00:38:32) How Anthropic's Finance Team Uses Claude (00:41:32) The Jevons Paradox for Labor (00:43:08) Anthropic's Fundraising & Growth Journey (00:47:31) The Exponential Revenue Curve (00:49:02) The Hardest Thing to Explain to Investors (00:52:15) AI's Public Perception Problem (00:55:38) Mythos (00:57:31) Relationship With Government (00:58:51) Inside Anthropic's Culture (01:03:48) The Next Frontier: Virtual Collaborators (01:06:22) How Leaders Scale With a Business (01:10:55) The Biggest Risks to Continued Progress (01:12:09) What Krishna is Excited About  (01:13:45) The Kindest Thing [Source](https://traffic.megaphone.fm/CLS3629415437.mp3?updated=1778621552)

[2:29] Krishna Rao: 它是构筑一切的底层画布（the canvas on which everything else gets built） [2:29] Krishna Rao: 买太多算力会破产，买太少算力则无法服务客户且丧失前沿地位。 [9:08] Krishna Rao: 前沿智能的投资回报率（returns to frontier intelligence）正在变得越来越高 [18:46] Krishna Rao: 人才密度远胜于人才规模（Talent density beats talent mass） [20:00] Krishna Rao: 直接按照你的终态去设计产品，因为我们在研发端一定会随着时间把这项能力补齐。 [40:00] Krishna Rao: 如果我们自己都不能成为这款产品的超级用户、不去主动探寻它的能力边界，又凭什么指望客户能做到这一点？ [47:31] Krishna Rao: 太棒了，我们打算在哪一年实现？ [58:51] Krishna Rao: 我们的对手极其能打，成功绝非必然（Our competitors are incredibly capable and success is far from guaranteed） [58:51] Krishna Rao: 下一个攻坚目标是什么？ [63:48] Krishna Rao: 每位员工都在某种意义上变成了管理者 [65:40] Krishna Rao: 这太疯狂了。如果他说的哪怕只有 10% 变成现实，它都将打破我和绝大多数人过去所见过的一切思维范式。 [65:40] Krishna Rao: 能在这个时代的关键节点，与这样一群杰出的人一起在这样一家公司攻克这个问题，这是何等不可思议的机遇。 [70:00] Krishna Rao: 变革是极其艰难的（change is hard） [73:30] Krishna Rao: 很大一部分原因是为了确保我未来能拥有选择的权利。

[9:08] Krishna Rao: There's a level of compute for model development that we will not go below. Even if it means it's harder to serve customers... we want to continue to make that long-term investment in developing the best models because we think the returns to frontier intelligence are extremely high. [18:46] Krishna Rao: We don't really think about models as closed or open. We think of them as frontier or not. The ones that are at the frontier clearly are capturing this economic value, driving meaningful ROI for customers. [32:42] Krishna Rao: Changing the pricing for Opus, you see this Jevons paradox. We lowered the price of it, but the consumption went up way, way more than what you would have expected. [38:32] Krishna Rao: What used to take hours to produce a weekly report for what's driving revenue or what's driving our compute utilization now comes down to 30 minutes. And then we can spend our time on the actual strategic implications of the business. [45:00] Krishna Rao: Consistently, Dario has been a much better predictor of the revenue than I have. [55:38] Krishna Rao: This was the first model that we decided to release in a different way... focus on how it can actually be used positively in a defensive way as opposed to in an offensive way. [60:00] Krishna Rao: When Meta and others were out with these huge packages for some of the technical talent across the large language labs, I think we lost two people and other labs lost dozens. [65:00] Krishna Rao: He walked me all the way home. I remember I came in and I told my wife, I was like, this is going to be wild. If even 10% of that is true, this is going to bend all paradigms of not just things I've seen, but what most people have seen. [73:48] Krishna Rao: A big part of his decision, a big factor in his decision, was giving me the opportunity to have choice. I didn't know that 12-year-old me or 13-year-old me would have never understood. But many years later, I think that's something that was incredibly kind and is something that I still kind of hold with me today.

[0:00] Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all run on Ramp. Mine does too, and so should yours. Learn more at ramp.com slash invest. Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, context, and standards. Send Felix an email like, take these comments and turn them for me, or update my tracker with the context of these emails, or run the ability to pay math on this buyer, and Felix sends back finished PowerPoint decks, Excel models, and sourced research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai slash Felix. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Hello and welcome everyone I Patrick O and this is Invest Like the Best This show is an open exploration of markets ideas stories and strategies that will help you better invest both your time and your money If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus along with all of our podcasts at colossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum, All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Positive Sum may maintain positions in the securities discussed in this podcast. To learn more, visit PSUM.VC. My guest today is Krishna Rao, the CFO of Anthropic. The center of our conversation is how he navigates the decision around procuring and allocating compute, which he describes as the canvas on which everything else gets built. We talk about what he calls the cone of uncertainty, the three-chip platforms Anthropic uses fungibly across Tranium, TPUs, and GPUs, and the daily meetings they run to allocate compute between model development internal use and serving customer demand He explains where the returns to frontier intelligence keep getting higher especially in enterprise and how Anthropic thinks about the line between platform and application and why they choose to build their own products like CloudCode Krishna has such a unique seat watching one of the fastest growing businesses in history, and he is generous in sharing what he has already learned since joining the company two years ago. Please enjoy my conversation with Krishna Rao. Krista, I have been so excited for this conversation because you get to see from the inside one of the most interesting businesses in world history at maybe the most interesting time in world history, at least if you're a technologist or care about technology. One of the things that fascinates me most is this question of compute that you have to deal with all day, every day. It's a key part of what you do. It's a key part of what these companies are doing. And there's just this whole revolution happening. I'd love you to just start by explaining what it's like to have to deal with that. I understand at one point you were having like a daily meeting about how to allocate to compute and to who and why. Bring us into that part of your life, because I think it's right at the cutting edge of what's going on. The compute that we procure is the lifeblood of our business. It is the most important thing in the company. It's like the canvas on which everything else gets built. And so the decisions we make and how much compute to buy are some of the most consequential and hardest decisions to make in the entire company Think of it this way If you buy too much compute you go out of business If you buy too little compute you can serve your customers and you not at the frontier Same thing. So we talk a lot about this cone of uncertainty, but the idea of just these purchases that have these real world implications, you can't just go out and buy a gigawatt of compute and have it delivered next week. You have to really think ahead to plan for this. And so we really take a very disciplined approach to how we think about it. So we look bottoms up, we model what we think demand will be. Obviously, we sometimes get that wrong. We think about the compute we need to stay at the frontier. And we really look ahead and try to estimate that. And then as we go out and do these deals to procure compute, flexibility is really important to us. And so we build that flexibility into the deals themselves. We build that flexibility into how we use the compute as well. Because the way in which we bridge from a position we are today to where we want to go when the business is growing exponentially is to use that compute as efficiently as possible. I would say I spend 30 or 40% of my time on compute even today. What does flexibility mean in that example? It means a couple of different things. Number one, we use three different chip platforms. So we are customers of Amazon's Tranium chip, Google's CPUs and NVIDIA's GPUs. [5:00] use these chips fungibly. If you think about the compute we buy, we're using it for model development. We're using it internally to speed up our own product and model development. We're also using it obviously to serve customers. Across those three chip platforms, we're using compute for all of those internal and external uses. And that flexibility took us a long time to be able to do that. We've invested in that over multiple years to be what I believe the most efficient users of compute amongst any of the frontier labs. And that's not something that just happened overnight. When we started using TPUs, I think it was third generation TPUs was the first one we used. At scale, people thought, oh, well, you're crazy. Everyone's using GPUs. Why aren't you using GPUs? And we've invested very heavily to be able to use that compute incredibly flexibly. And then we look across the different generations of those chip platforms and use each generation for the best workload internally. And so we really built this orchestration layer that gives us that flexibility to use all different types of compute. And in doing so, we also are able to get the most value out of it. Am I thinking about this in the right way that something like CUDA that has been a part of NVIDIA story for a long time now that allows you to do a lot with the underlying actual hardware that you want to sort of eke your way into being as close to the bare metal as possible That part of this flexibility and being able to control as many of the variables as you can Is that the journey that you been on That part of the journey for sure, but it's also been actually pretty collaborative. So we work really closely with the Annapurna Labs team at Amazon to help to influence the roadmap of these chips, because we believe what we're doing is really stressing the limits of what these chips are capable of. And that means that like a dollar of compute inside our organization goes further than I think it does anywhere else. Importantly, we basically want to utilize each chip to its best purpose within the company. So that does mean that we're building our own compilers. We're really building things from the chip level up in order to have that customization and that flexibility to use it internally the way we think will generate the most ROI. Can you explain this cone of uncertainty thing? I want to ask about all the component parts of this, but that feels like a really key starting point or overall frame for how to think about both sourcing and then the uses of compute. Can you just explain what that concept is? When you're building and growing a business exponentially, really small movements in monthly or weekly growth rates result in compounding very very different outcomes And so as we thinking ahead even with our revenue growth it really hard to predict this business And it really hard I think humans mostly think linearly and you think incrementally I been at the company for two years That something that a paradigm I had to break for myself to stop just thinking linearly and think on this exponential. When you're on this exponential, again, the range of outcomes starts to be really, really wide. We look at a range of scenarios and we look at different points in that cone of uncertainty over a one to two year period. And then we kind of work backwards from that. And what we want to do is we want to be at a place where we can obviously still be at the frontier. That's the most important thing to be able to serve customers and then to be able to have enough internal compute to accelerate our employees. It's interesting if we were to say to our employees, you can't use our models anymore. We could serve billions of dollars of revenue with that compute that we allocate to employees internally. But we want to take a long term view and a long term perspective on that cone of uncertainty because we want to range towards the top end of these outcomes. but we have to plan for that. And as we go, that's how we think about buying compute in a disciplined way. The most important thing is what happens if you are at one point in the cone of uncertainty, but you've only bought compute for a different point. That's where this compute efficiency is something that really has helped us out. Can you bring us into the room for the conversations around the trade between those I so interested by those three buckets of training research internal use broadly speaking and then serving customer demand Naively you might think okay it a third a third a third allocation or something How much does that range around? What are the trade-offs? What is that discussion like on an ongoing basis? In addition to meeting about compute procurement, we meet a lot about compute allocation. I think what's important is it starts with a place where our culture is one that's incredibly collaborative. And that informs how this conversation happens. So there's not fiefdoms. It's done in a very collaborative, not a zero-sum way. There's a level of compute for model development that we will not go below. Even if it means it's harder to serve customers, or we have to do kind of unnatural things when it comes to that, we want to continue to make that long-term investment in developing the best models because we think the returns to frontier intelligence are extremely high. And it's extremely high, especially in enterprise. That kind of puts a floor on the compute that's allocated to model development. And then as we think about the internal use of compute, it really helps us to speed up that model development and to speed up finding those compute efficiency multipliers that really get us more from each dollar of compute. So when we're talking about it, each team is kind of representing what they would do with that compute. And then we have a really open and frank discussion about how we think about ROI and because we can allocate that. [10:00] compute so dynamically, we can make changes. We can make adjustments in that on a relatively short time horizon. The efficiency thing is so interesting to me. I'm curious if you have a sense of how much more efficient you are versus your own internal benchmarks from a year ago or something or versus others that you have some sense of how efficient they are. How do you measure what efficiency means? There's a couple of different ways I would think about it. From a model perspective, I think the analogy people have when these new models come out is They're kind of like cars. You had a sedan before, and then you might have the higher inversion of that sedan and you're moving up the chain. And I think that is true in terms of model intelligence. The place that analogy kind of breaks down a little bit is people think, OK, I'm going from the sedan to the sports car. We get much less fuel efficiency. I'm not going to buy the sports car for the gas mileage. In our case, we actually see huge improvements in capability, but also model efficiency. And so if you look at going from Opus 4 to 4.5, 4.6, and now 4.7, they're not equal. But each one of those leaps to a new model has a multiplier in terms of how much more efficient it is at processing tokens effectively. And that just doesn't serve customers. That also helps us internally as well Because if you think about if we doing reinforcement learning on the model it basically inference within a sandbox with a reward function And so if the model is better at more efficient inference that RL is more efficient as well We able to do this kind of win-win where the customer is getting more capability when we release a new model. And then we're able to serve that model sometimes, again, a multiple more efficient than the prior generation. And then when we're in between generations, we're dynamically deploying efficiency improvements in between these more step function model changes. It is always getting more efficient over time. And what fuels that is the research team. So if you think about it, all these things are very connected. These various tasks and workloads that we have internally all fit together in this way of doing R&D for model capabilities, for compute efficiency, for serving customers, and then having internal workloads that can be sped up by using the best models, sometimes models that we haven't released. You said something really important before, which is the returns to being at the frontier are really high. Can you just explain that in as much detail as you can? It sounds obvious when you say it, but there's certainly been some camps I can use the six month old model and it's a fraction of the cost and I just use that and that be catching up all the time And that just like hasn been the case The second Opus 4 comes out even me as a consumer the thing you do is you switch it on or GPT 5 comes out you switch on the new one right away Like I want the best Talk about the returns to being on the frontier and why it so high I think it's a couple of things. It's every time we have a new model, there's a set of capabilities that are different. People tend to think about model intelligence as IQ. It's a single number. OK, this model was at 110 and then it goes to 125. We think of it differently. Intelligence for us is multidimensional. It's not just a score. In fact, we find that, yes, everyone publishes their model benchmark cards. We find that a lot of those benchmarks are saturated. We publish it too. But what our measurement is what the customers tell us. What is the real world capability of this model? As we've released better and better models, what we've seen is it's not just the outright intelligence. It's also the ability to do long horizon tasks. It's the ability to use tools or computer use. It's the ability to do things for agentic tasks that have specific value even faster. In some sense, if you have two employees and they're maybe both equally capable, someone takes a week to do an assignment, someone does it in a day. Well, that second person, if they're continuing to do that, can be seven times better. They might be equally capable at something maybe just take longer times to do it All of those factor in to then how customers experience it And what we found very consistently is by releasing new models the TAM is unlocked in a unique way More TAM gets unlocked, more use cases are possible. And a good illustration of that is this last four months that we've had at the company, right? We started the year with about $9 billion of run rate revenue, and we ended the quarter with the $30 billion of run rate revenue. That kind of a change is really enabled by these model intelligence leaps and then the products that we build around them. And so that's what I mean by the returns to frontier intelligence are really high. I think that's unique to enterprise. Then consumers, sometimes you don't see that as readily, that consumers really are pushing the limits of what the models can do. Whereas in enterprise, like our customers are always now, it started with coding, but it's really expanded beyond that very meaningfully. But each model generation gives you the chance to do more with it, to do it better, to do it more efficiently. And customers see that, and then they invest really heavily in more tokens with the newer models. And we just have seen that cycle play out again and again. And that's a core thesis of our business, that especially in enterprise, the returns to frontier intelligence are not slowing down. Vanta automates security and compliance for over 16,000 fast-paceds. [15:00] moving companies like Ramp, Cursor, and Harvey, keeping them audit-ready around the clock. It's the number one agentic trust platform, and it now helps companies like yours watch for the risks that show up between audits, across your vendors, your AI tools, and your whole environment. Every new tool your team signs up for, every vendor that turns on AI features, is an opportunity for something to go wrong, and most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24-7 GRC engineer in the background, finding issues, drafting fixes for you and cutting vendor assessment time by up to 50%. Whether you're a fast-growing startup or a global enterprise, Vanta helps you earn and prove trust. Invest like the best listeners get a special offer for $1,000 off at vanta.com slash invest. Ridgeline is the first end-to-end system of record with embedded AI for investment management firms running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared to anything else in investment management software, which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline unified platform If you serious about your firm AI strategy Ridgeline should be part of that conversation You can request a demo at ridgeline The things that push that frontier, this is like a sci-fi story or something from books I was reading when I was growing up. It seems as though in the major labs, we've reached this point, someone on your team said it recently, of recursive self-improvement where the models themselves are building and doing a lot of the research to do the next generation of improvement. If I think about the frontier that you're pushing and OpenAI is pushing and compare that to the open source models, that maybe the gap will widen as a result of you getting there first to this like recursive thing. How do you think about that? Tell us how we should think about this idea of recursive self-improvement in the models themselves, because it seems like getting there first is incredibly important because then you just can continue to separate yourself versus those that haven't reached it yet. We do see progress accelerating. I can't speak for other companies, but for us, the scaling laws are alive and well, and we're seeing that with releases more recently like Mythos. But right now within the company, 90 plus percent of our code is actually written by Cloud Code A lot of Cloud Code code is written by Cloud Code And so you think of this as why do we allocate compute internally Why would we forego revenue for it It because the models themselves are helping us to build that next generation of models In addition to this capability leap that you would have from the scaling laws, talent is really important. And that talent with the best models can really accelerate the development of the capabilities. And we're really seeing that. We don't really think about models as closed or open. We think of them as frontier or not. The ones that are at the frontier clearly are capturing this economic value, driving meaningful ROI for customers. We are just investing behind that thesis. And that means both compute, but it also means talent to use that compute and use our own models to really accelerate the development. And that's been something we've been doing for a long time. It's not a new phenomenon, but we're seeing the fruits of that in everything we've done. The other piece of it is to, it's not just the models, it's the products that get built on top of them. We had 30 different product and feature releases in January. The pace of that is accelerated as well. And that's enabled in part by utilizing the models with the talent that we have to accelerate ways to access this underlying intelligence. That's kind of our theory of the case on the product side How do you think about this weird world where you mentioned the talent and the leverage and they not writing code themselves and CloudCode writing its own code It seems like the last step of that would be you don even need the talent to tell the thing what to do It just figures out what to do on its own And that's like the ultimate thing then just runs and is only constrained by computer or something. Am I being too crazy about that? Or is that future possible, do you think? The core of our company is still a research lab. I think it's maybe not as well understood. Maybe it's getting more understood from the outside. But we're doing experiments. We are doing things that push the limits of what our models can do. And that research and that engine is upstream of everything else that we've talked about. That is enabled by the models today. It's not entirely done by the models. Over time, we think that the models will get better. They'll be more helpful in that process. But having the best talent to set the direction and to help effectively guide that process, not just the priorities, but some of the new areas of discovery, it just actually makes that research talent even better. I think of it as accentuating and accelerating the talent that we already have. We talk a lot about how talent density beats talent mass. And I think that's true here. Like we want the densest collection of AI research talent and inference engineering talent. And that enabled with the best models is a really winning combination. [20:00] What are scaling laws talked about internally? The sort of consensus has been you've got different components of them. You've got pre-training, you've got post-training, you've got reasoning, and that all of these are kind of moving at different paces. And to hit a true wall, they would all need to fall down. That's sort of like how the world is now conceptualizing scaling laws. How are they talked about internally? How do you think about them? We look at models at various points in their development. So we can see during a pre-training run? How does this model compare to a prior model that we did on these kind of loss curves? And that gives us a sense for model capability. You can do the same thing as you think about RL, probably as importantly as when customers get their hands on it. What are they seeing? Where are they identifying pain points? And those pain points become like training targets for us. We don't train on customer data in the enterprise side, on the prosumer side. It's only if you opt But customers tell us things like, hey, I wish the model were better at this, or I had this particular place where it got stuck and I could build this other product, but the capability needs to be further than that. What we usually tell them is build your product for that because we're going to, on the R&D side, improve that over time. And so there is this connected loop but internally we always looking at different models that are being trained different snapshots that we have and comparing them internally and to a lesser extent externally against our own measure and then ultimately how our customers view them as well And it feels like there's just no slowdown in the scaling laws themselves. Is that a fair characterization? For us, that's a fair characterization. Obviously, a bunch of the authors of the scaling laws papers are amongst our founders. Notwithstanding that, we can be a bit of a skeptical bunch. We hold ourselves to a really high standard. It's very kind of scientific method and people are constantly challenging previously held assumptions. But from what we see, the scaling laws are not slowing down. So if that's true, you said before, it's hard for humans to be exponential in their thinking and not linear. That continues to be true for however many more turns of the crank here. How do you do that thing of not thinking linear and thinking exponential yourself in your job and for the business? The implications are really hard to reason through. Exponential growth rate is one thing, but exponential growth of capability. I don't even know how to get my head around it. How do you get your head around it? We think about the world as scenarios. It's very hard to have a point estimate in this business. And then having a very low bar for updating your current priors, basically, or your current perspective, because it could be the case that something a month ago was true that just not true today And that breaks your model And you have to go back and update it And so this old like well we forecast once a quarter and we revisit this in three months at the next board meeting That doesn work for our business It's so dynamic that we have to always think about, oh, our models couldn't do this before and they could do this now. What does that mean for the TAM? We've seen this in coding first, starting with around Sonnet 3536. we started to see this really remarkable jump in capability, which was then followed by adoption and usage and revenue. It was a little hard to predict that, but now we can use coding as an analog for a lot of what's happening elsewhere in the economy and elsewhere in our business. We kind of look at pattern recognition in our own business to try to predict what's going to happen in the future. Literally 15 minutes before you got here, the news came out about your partnership with XAI and the Tennessee facility. But it makes me curious about how you are canvassing the world, that is an opportunity you decided to do. I'm sure there's a universe of things that you've explored. What is the strategy for trying to get more in creative ways? Bring us a little bit more into that. We announced a partnership with SpaceX for their Colossus facility in Memphis. We're really excited about that. It's going to allow us to continue to expand, especially on the consumer and prosumer side But that just one example of us just as you said looking for near compute wherever we can get it As the compute base grows that near compute becomes a smaller and smaller fraction of what available and what out there We look at it as, can we deploy that compute that's available productively? Sometimes the answer is yes and sometimes it's no. But if we can, then we look at the economic return on it based on what its price, what duration we have it for, where it's located, what type of compute it is and how efficiently we can run it. So we have kind of a process to assess. And that same process, by the way, we use to assess longer term deals as well. So last month, we signed a five gigawatt deal with Google and with Broadcom for TPUs starting in 2027. We also signed a deal with Amazon for Tranium for up to five gigawatts as well. It was an over $100 billion commitment. And a lot of that computer is actually already landing and will land in the rest of this year into next year. And so if you think about it, it's a bit of this layer cake of compute that's starting at different times with different capabilities. And we're very dynamically comparing that compute. It's price performance over time. That's really, really important to us when it lands and what we think we can do with it internally in the business. And so there's so many different variables you have to optimize for around what compute it is at what cost and over what time horizon. [25:00] But we have a pretty dynamic way of looking at near-term compute and then medium to long-term compute. But the things we're assessing are largely the same. What is different is just the time horizon. What about the trade-off use of price per performance? Trade-off between like cost per token or something throughput and speed. Both are important, of course. From the customer perspective, they care about both. Speed probably unlocks some capability and use cases that are really interesting that we don't know about yet as these things get faster. Can you talk a little bit about that trade-off in compute as you're assessing it? As we look across three different chip platforms, we also have multiple generations of chips within it. It could be TPU, V5E, and V6, and V7, and TrainiM2, TrainiM3. All of them are at different places on the price performance curve. And then we importantly look at how we will utilize it. Price performance is important because of efficiency. Speed is also important for certain use cases as well. So we look at the compute down to a very granular level in terms of what it can deliver for us and when. And that's something that we do. Again, our compute team leads that, but we closely collaborate across the business to say, like, where do we need this compute and for what? OK we might need CPUs for RL OK how much of that do we need and where and what the nature of it We might need this more leading edge compute and we going to deploy it for our best and fastest models or for training them From our perspective it customer demand but it also really down quite granular in terms of what is each chip best for and then what will we have when. I'm always so curious by the metabolism of, in this case, Anthropic for new compute, how fast you could take if I airdropped on you twice the compute that you have tomorrow? Like, would you consume that? And how fast would you consume that? If I ever dropped 10 times the compute on top of you, how fast would you consume it? Can you calibrate us on that? It feels like demand's unlimited between these three uses, training, internal customer demand. Everyone's saying the same thing. Shortages everywhere. Memory stocks, mooning. Is it that extreme that like if you 2x or 5x or 10x the amount available to you tomorrow, you just like more or less instantly consume it? This goes back to like how we use it and the fungibility of it. So the answer is we're constrained across those use cases. A year or two ago, it would have been harder to consume, especially like a heterogeneous compute drop in your example, really quickly because these chip platforms are different and they are different. Some are harder to operate. Some of them have idiosyncrasies in terms of how we use it. I would say today that getting a bunch more compute I think it would be deployed very rapidly across those different use cases We probably have the same allocation or calibration that we do with compute today, but it's become a lot easier for us to spin up very quickly and deploy almost any type of compute. And that's something, again, we think is a real advantage. One of the interesting tensions and trade-offs that I'm fascinated to hear how you think through is between the platform approach where I build my business on top of Claude and it powers my thing versus you doing the thing that I wanted to build. This is like the classic Claude design versus Figma or something like this. How do you think about the right balance of how deep into the application layer you should go versus just being a pure enabling layer of we're going to provide the reasoning engine and the intelligence and world go forth and build whatever you want, pay us through the API or whatever. That seems like a fascinating internal discussion and tension to some degree. The way I would think about it is most of what we're building is platform. We think that there's so many examples of where a platform can accrue a lot of value, but the customers who are building on that platform actually accrue even more value. We think that's what we're setting up for today. It's maybe akin to the early days of AWS. If you think about the cloud platform and all the tools and services that are now built into it because it not just the raw model access it is prompt caching and the ability to use virtual machines and cloud code being called within there or dispatch on and on The cloud agents SDK, managed agents. All of these are effectively, I think of as vectors to access that model intelligence for other companies to build into their own products. That's most of what we're focused on and really most of where we think the business is going from where we are today. We will also build our own applications on that same platform where a couple of things are true. Number one, if we feel like we have a vision into where the models are going and we can kind of demonstrate that and create customer value in that, that might be something like CloudCode. We are able to say, actually, a lot of what's out there in the market was developer led. CloudCode is a platform that's Cloud led. And we think the models can't quite do that today when it was launched, you know, a little over a year ago. We think they'll get there and they have. One is kind of building ahead to model capabilities. The second is thinking about ways to like demonstrate value for the ecosystem that others might emulate. If you think about Cloud for Financial Services or Cloud for Life Sciences or even something like Cloud Security, these are ways in which we've composed the platform. Again, we're building on the same platform as our customers. That creates a level playing field. We also think that there's so much value that's... [30:00] going to accrue in these areas that our customers can win and we can win as well, which is why you've seen as we've launched some of these products, we've done them in a collaborative partnership oriented way, whether that be on the security side or sign or financial services, we've partnered with the ecosystem. So I think of our strategy as mostly horizontal. We'll build vertical where we think we have some value to add or a perspective that's useful or a way to demonstrate to the market, how we think about our platform adding value, a lot of the value is going to accrue to the customers that are building on top of it. Our goal is build the best models and then build the products and tools and services that allow that intelligence to proliferate within customers. How much do you care that it's just a reality that people are scared of you? There's a sense that because you control the most essential piece of these new applications, the underlying intelligence reasoning engine, that may totally be true, maybe already is true, that more of the value is accruing on top of the Anthropic platform than is being captured by it. But nonetheless, it's still scary to imagine. And I guess maybe you could say something similar about cloud and AWS or something like that But how much do you think and care about the fact that some of your would customers or existing customers are in fact scared of you as a competitor Part of what is hard in this business is it changing so quickly So the model capabilities sometimes even surprise us And so when we release models or products on top of that there is an element of what's happened in prior waves over the course of five years, 10 years, 20 years is happening in months now. When we release things, people are also surprised by it in some ways in the same way that we were surprised by it. But I think fundamentally, what we are trying to do is be very partner oriented towards the ecosystem. And that means that we have early access programs. We work very closely with customers. We listen to them about what capabilities they want. That doesn't mean that the things we release are sometimes not moments where you're like, wow, that's way more powerful than I thought it would be. Or I didn't realize the models would be able to do that this quickly. Part of that is a reality of where we are in this cycle, in this development of intelligence. Part of it is also like we want to make those capabilities really accessible. And that should accrue a lot of value to customers as well. And customers that are front footed on that and adopt and also ones that are building and using the tools that we offer on our platform, we think we can actually accelerate them. And so I think some of it a reality of frontier model development but our approach to it is probably a little different and more partner oriented You said before going nine to 30 in the first quarter the pace is so insane which makes me wonder about pricing The dynamic of how to price tokens or use of the system is so fascinating to me because I think a lot of people a year ago would say price is going to constantly fall. But actually what's happening is it's going up in many cases. This is true at different levels, whether it be the mythos pricing that is quite high because it's so powerful, the cost of an H100, the rental price of a cost of an H100 looks like a smile curve. I'm very curious why if everyone is compute constrained, why everyone doesn't just raise prices a lot to try to find what the right equilibrium is. And so I'd love you to just riff on pricing, how you think about it, what the tradeoffs are, why not raise prices a lot. The company is only a little over five years old. This past March was the third anniversary of the first dollar of revenue into the business. And we only had a frontier model for real for the first time in March of 2024. So the timescale of these things, it's an important backdrop as I think about it. Our pricing has been relatively stable across Haiku, Sonnet, and Opus. And now Mythos is obviously newer We made very few pricing changes The biggest pricing change we made was to bring down the price of the Opus family when we launched Opus 4 If we thought about why did we do that? It's really because we found that Opus class models were underutilized relative to their capability. People were trying to often fit an Opus problem into a sonnet. We were able to serve that very efficiently from our perspective, but actually bring down the price, which made it more accessible to customers. We want our customers to generate a lot of value from it. And they're generating a ton of ROI from our models today. We want that to just continue because our goal is to proliferate this throughout the ecosystem. We think we're in the very, very early innings on all of these use cases. The best way to do that is get this intelligence in the hands of as many businesses from startups to digitally native businesses to the largest companies in the world, some of that means that you have to make it in a price point that's accessible and that allows them to get a lot of value from it. Changing the pricing for Opus, you see this Jevons paradox. We lowered the price of it, but the consumption went up way, way more than what you would have expected. And so because we hit that sweet spot for customers, they were able to use it a lot more. We had the efficiency to be able to serve it to customers at scale. And then they were able to build a lot of the cost. [35:00] that into their workload such that when we released Opus 4.6, it's a model improvement, they can slot it in, we didn't change the price. So we think pricing stability is important. And we also think that pricing to get that value and to see that kind of Jevons paradox happen is really important. The other component of this is margins and how you think about margins as a business. Again, because this is so unbelievably capital intensive to build these frontier labs, you've got the levers we talked about, which is efficiency, price, both those things relate to margin, given how much capital you need, why not just say we want to have a healthy margin and set the price accordingly. And maybe that price can come down if efficiency is better or whatever. And I'm just curious how you think about margins as it relates to pricing in the business. We think about what is the return on our compute spend writ large. So that is all of the different workloads that we've talked about, whether it's serving customers. If you think of all of those are kind of in support of revenue over different timescales. And if I serve inference, it's in support of revenue today. If I do model development, it might help for a capability that unlocks TAM that drives revenue six months from now and everything in between. If I do internal acceleration to launch a new product all of these things are in support of that Our returns on that compute expense today are robust We think of it as what is the return on that full envelope of compute We feel really good about where we are from that perspective And we're balancing delivering value to customers with also seeing a really, really strong return on that compute ourselves. If you think about when revenue grows, as we mentioned, kind of in Q1, it's not like we onboarded a bunch of new compute in that time period. We talked about compute comes based on a ramp that might've been determined 12 months ago. This idea of a variable cost that's like on the incremental to serve a customer, it doesn't really fit our business. It tries to maybe fit our business into like a software paradigm. That's not the case. In actuality, compute is supporting all of these activities. We're really generating a robust return on that compute. And that's our measuring stick. I think it's something where we think of the compute envelope that we have as the thing that is able to govern how much we're able to drive revenue, both over the short term and the long term. If you're this great customer of the compute providers, what does that group need to do to be a great provider to you to help you drive that return? We're fortunate in that we have really great partners in Amazon in Google in Microsoft but also with Broadcom and NVIDIA as well Our ecosystem we are the only model that on all three clouds Today we the only model that has large language lab that's using all three of these chip platforms. These collaborations are much deeper than just procurement. I think that's something that's often lost. If you think about our relationship with Amazon, our teams are deeply embedded with the Annapurna Labs team. We are good users of Tranium. We've spent a lot of time and energy and worked closely with the team and internally. And that's something where we plan capacity together. We do that with the other hyperscalers as well. If you think about the three clouds, they're great distribution engines for us too. We have a really, really robust first-party business as well. But these are multifaceted partnerships, whether it be on developing the chips themselves, landing that capacity, serving it, and then ultimately distributing it to the customers. Each one of them has that element to it. They're different and each of them has relative strengths, but we've been able to work really well across those clouds and then obviously across Broadcom and NVIDIA as well. I'm thinking about your function like the finance team. I picturing this like ROI on compute thing on different horizons with all these complex variables which makes me wonder how do you use these powerful tools yourself internally to run your group and the business What is the deployment of Cloud Code and Cloud in general on the finance team at Anthropic This is really interesting because we were using CloudCode about a year ago. And I started asking people, is everyone just kind of taking up vibe coding or what is it? And we started to use CloudCode as almost like an assistant, a digital coworker, not just for coding tasks. And that actually was early in what eventually became co-worker. That was kind of an extension of Cloud Code to say what it's done for agentic software development, it should do for all of knowledge work. But then we started to productionize that. And I'm actually really proud. We spend a lot of time with our product team too. They kind of see how we use it and get input and feedback from that. But today, all of our legal entities, we can produce the statutory financial statements using Cloud. And yes, the human checks it, but all of those financial statements are produced with Cloud. We also have a more real-time platform called AntStats. And it used to take a lot of time to sift through all the data, get to conclusions, write a memo about it, or publish a regular report on what's happening over the course of the day, what's driving it. We now have a library of skills for Claude that are specific to finance. Last I checked, there were over 70 of them that everyone can kind of access through this common repository. [40:00] And on top of that, we built an MFR, a monthly financial review skill, and it can produce a monthly financial review. It's 90 to 95% ready. And then all of our discussion becomes about what do we do? What are the implications? Not what exactly happened, because Claude is not just reporting the weather. It's also helping to think about drivers. Why did the number change in the way it did? And that gives you tremendous insight into the business, both in terms of this MFR that we do, but also on a daily basis. And so what used to take hours to produce a weekly report for what's driving revenue or what's driving our compute utilization now comes down to 30 minutes. And then we can spend our time on the actual strategic implications of the business. We can also get it in the hands of business leaders much more quickly. The insight engine is a lot faster. I have a dashboard. I look at token usage. Leaderboard. Yeah, we don't compensate people on it. No one's trying to token max for that. But it's really interesting because some of the most senior people within the finance team are actually the biggest users of tokens. So it is not just the 22-year-old who joined and has a coding background and was doing that on the weekends and brought it to work. It also people using the tools to change how they working I think our number one user is our head of tax and he focused on tax policy engines and automating large parts of the workloads that are happening within the team. So I love seeing that. I tell people, if we're not super users of this, if we're not pushing the limits of it, how can you expect customers to do that? Just as a human, does it freak you out at all? I've heard so many examples like this. It starts to feel like we just start doing the stuff that AI tells us to do in the sales example or the calendar or whatever. And maybe that's great. Maybe it's just such a better coordinator and wide thinker and optimizer than we ever could be that we should do what it tells us to do. But it feels ever so slightly dystopian to me that that reality is coming quickly. I've had examples of it too. It feels kind of cool, like it's helpful. But at the same time, if I really close my eyes, I'm like, oh, I'm just like doing what it tells me versus me telling it what to do. It's a really interesting just human dynamic that I'm curious for your take on. I maybe have a slightly different view on it. in that I think we've been able to hire great people at the company, but it has made even those incredibly talented people so much more productive. There's a little bit of this, I think of it again, like Jevin's paradox, but for labor, which is that we have people who become incredibly more productive. We hired a lot more people because of that because there no shortage of work to do And now with the assistance of Claude people are spending less time in that MFR trying to reconcile some number but they actually thinking oh how do we reinvest this in the business How do we think about dynamically allocating resources Whereas before I'm working to tie out a number or I'm in that accounting example, taking a long time to close the books. So I actually think of it as even more optimistically that it is an accelerant to our productivity. And that actually means that we can get a lot more done and that even as we grow the team, those people are more productive as well as they come up the curve on how to use Claude within our company. And I think that's starting to be true across many companies as well. I'd love to talk about investors and capital formation. Of course, you've had to raise tons and tons of capital. At the same time, it seems as though if I just squint my eyes and think about the multiple on current revenue, it's not that crazy in terms of where you're raising money. I'm so curious for you to teach us about what it's been like to interact with investors, how you've seen their understanding of the company evolve and mature? Where do you think the investors as a group understand it now? Where are their misunderstandings about Anthropic? Tell us that side of your life. And I don't know what percentage it is of your job, but it's an important part. So I joined the company about two years ago. We were closing our Series D at the time That was not a straightforward fundraising The company really only had a frontier model in the middle of that fundraising Towards the tail end of it the FTX transaction was happening, which liquidating a bunch of Anthropik shares. That was kind of the starting point. And at that point, the questions were around, why do you need to have a frontier model? What's the returns to this? They're also around our mission and how we approach things. People said, well, AI safety and building a really big business, aren't those things at odds? And there were also a lot of other misconceptions on your Salesforce is really small. Don't you need to scale it like all these enterprise software companies? There was just a bit of a paradigm around trying to fit us into a particular mold that it existed before. Over time, it's evolved. At the end of 2024, we raised the Series E. The business had scaled to close to a billion dollars of run rate revenue. But the day of our first close was the day the DeepSeek news came out. We got the close done, but certainly a ton of volatility is people then said, wait a minute, should I just totally re-underwrite how I think about AI in total? That was a series E. Obviously, we brought on great investors across all of these, but people still had some of those questions, but they looked at our forecast and they thought, I get it. You've grown to a billion dollars of running revenue so quickly, but there's no way you're going to be able to keep up. Yeah, that's just not possible. There's laws of fixation. [45:00] physics, you're in enterprise, which is great, but the adoption is going to take so much longer. I mean, look at how long it took with cloud and how many people are still on-prem. The business continued to prove out the thesis that the return to frontier intelligence is really high. We are really focused on what's really happened is model-led growth, enabled byproducts in our go-to-market team and our distribution. What they also saw was that this thesis of, hey, it's really important to build this transformative technology, but to do it in the right way and do it responsibly, that had this really interesting interlink with our business that most people didn't really understand or really believe, which was that we invest in research, not just in model development, but also in AI safety research. We've pioneered interpretability, which is think of it as like an MRI for the model to see inside the neural network, how it works. We pioneered alignment science, which is you want the model to do, what do you tell it to? And how often does it do that? And how often does it stray from that? Those things are important for our mission. And that's why we did them. They had these downstream effects where it turns out if you can look inside the model, you're better at building them And then the last linkage if you selling to enterprises we now sell to nine of the fortune 10 All of those enterprises are entrusting us with customer information their data they interacting with their employees sometimes even interacting with their customers as well. Those are the most sensitive workloads. More and more of these businesses are running on Claude and our Claude platform. When you have this investment that we've made and will continue to make in safety, interpretability, alignment, that actually enures the benefit of the enterprise customers as well. And all of our customers, because they want that, if they're going to entrust us with all that access and all that data and the ability to work in the most sensitive workflows within their company, they want a company that they can trust. That's not why we invested in it, but it did have this downstream effect that we've really seen prove out again and again to be a company that is both at the frontier, but one that is investing in safety and that you can trust. We've raised $75 billion since I joined the company. We have another $50 billion that'll come in into the future from the Amazon and Google deals that we, that we in class fund. And so that's a tremendous amount of capital, but it's a capital intensive business. And we need this capital to support that growth. It all goes to the fact that the business is running very efficiently. And so the reason we raise this capital is more because of that cone of uncertainty than it is to fund losses in the business How is your own perception of this perspective for a 10X growth of the business The first time that happened did you personally believe that it was possible Did that seem absurd? Now that it's becoming consistent, maybe it's becoming more commonplace to you or something, but what was your own view staring at this cone about the odds of hitting a 10X type of growth so many years in a row? When I joined the business, it had about 250 million of run rate revenue and the plan was to get to a billion. And I said, great, in what year? That was linear thinking. Consistently, Dario has been a much better predictor of the revenue than I have. I think we're going to close the gap over time as we get better at forecasting and understanding the business. But definitely the first time I saw it, you have all these arguments about the laws of physics and law of large numbers and this can't. Where is the revenue coming from? And how can it be added this quickly? And how can customers move this quickly? And is this even possible in enterprise? All of those things start to get broken down over time as you see how the business works internally and you see how the adoption curves and the exponentials that are happening, again, we have the exponential that's happening on revenue, but that's underlies it or these many other exponentials that support that, you start to see and believe in that more. Now, that doesn't mean we not disciplined and thoughtful about the forecast and how we think about the range of scenarios But it does mean that my thinking has shifted a lot more from linear and incremental towards leaning into this exponential and really believing in its potential and how this is just different than how other businesses have evolved. As you've talked to investors at every stage, I'm sure every stage, every round that you've raised, there's something that is like the most common or hardest thing to explain to investors or that they're struggling the most to understand to get their heads around. What is that today? I think it is this paradigm of how compute is used, thinking of it as not just something that is a variable cost over some time period, but really this resource that's so fungibly utilized. We run workloads on one day in the morning on a chip for inference, and in the afternoon or evening, we use it for model development. That paradigm does not exist in a company like a software company or a factory. You can't repurpose if you have a bunch of people doing R&D and that's your R&D expense. They can't go and become COGS and vice versa in most traditional companies. Here, you really have that fungibility that's possible. And I think that's where the return on compute is so important. And I think people are beginning to understand that. [50:00] there's still a tendency towards treating it like, oh, I have to separate these two costs, when in actuality, they're very self-reinforcing. And that flexibility is actually what helps to drive revenue short-term and long-term. If I was to force you out of your role and into an investor seat at a great big investing firm, and then I said, your job is to go grill these companies and invest in the best ones, what questions would you be asking of the labs or companies that are building models to really get at the heart of the points of uncertainty, of skepticism, of things that might not make these the best businesses of all time. Curious from that angle, how you would approach it. First, what is the ROI on compute all up? How are you utilizing it? What return are you seeing today? And how is that coming over time? That's certainly a question I get asked a lot by investors, but I think it's a really important question. These are the massive, unprecedented investments that companies like us are making. What's the return that you're generating on that? And when does it come and what is the shape of it? So I think that's one. A second one is how do your customers see ROI? I get asked this question a lot as well Like are people just using this for testing Are they actually deploying this at meaningful scale For our business we seeing that in spades Our net dollar retention rate is over 500 on an annualized basis Nine out of the Fortune 10, these are real customers making significant buying decisions. They aren't pilots anymore. Exactly. On the way here, I was in an Uber and I signed two double-digit million-dollar commits in the car ride, which was like 20 minutes. From that perspective, we're seeing it. We're now being judged by some of the biggest companies in the world. the most sophisticated buyers and startups. They have choice in the market and they're choosing us. But I think one question I get a lot is, or I would ask from the investor seat, the skeptical investor seat, is how are your customers getting returned from this? Maybe a third one is how do you think about compute in the future and where does it come from? Because obviously some of the places that we buy compute from, they sell the compute to others. They may use the compute internally. What is the balance of that over time? So your philosophy there is be involved with great players and have flexibility. That's right. There's this crazy stat about AI, just the generic concept being less popular than Congress amongst like the general populace. And it's like kind of funny when you first hear it, but when you really think about it, you're like, this is kind of fucked. We need to solve this problem It doesn seem like the general world that isn in technology doesn live in the Bay Area or New York does not yet feel or understand why this is good for them just as measured by their opinion of it. What do you think we need to do as an industry about that problem? If we think about the transformation that's happening, there's been other transformative waves before, all the way back to the industrial revolution, the internet, cloud, et cetera. One of the things that's different about AI is it's all happening so quickly. You can have years or decades of progress that are being compressed into months. Going back to humans thinking in terms of exponentials versus linear, that can be jarring. We are very optimistic generally about the potential for this technology. We as an industry can continue to do a better job of articulating, you know, Dario wrote this essay, Machines of Love and Grace. It's all about the potential for this technology to transform the way that we live, whether that be in drug development and curing diseases that are more mainstream, but also rare diseases, the ability to accelerate that biological progress. Number two, in healthcare and how healthcare is delivered to raise our standard of living in the developing world and in places where resources are not as plentiful how do we actually make sure that the economic gains are also accruing to not just a small number of people in the world but across the world I think that all of those things are part of the promise and the potential of AI We could probably do a better job of painting that picture. And we want to show more tangible results for that over time. I think that is coming. And that's one of the things I'm most optimistic about. I think on the other side, though, we do want to articulate the risks. I don't think we should just tell everyone everything's going to be great. There are likely to be bumps on the road. And so I think people generally gravitate towards more honest and balanced assessments. If I feel like somebody is just telling me all the good news and none of the bad news, then I'm like, OK, do I really trust this perspective? That's where there's a need for balance. Look, these are some of the things that happen when change is compressed over a short amount of time. How do we work across commercial and government to actually come up with some of the solution sets? So I think it's about a clear articulation of the opportunities. It's about thinking about what those solutions may be. And that's not any one company that can come up with it. We don't have this blueprint that's going to solve everything. But to at least have that dialogue about some of the risks and downsides and what we can do to address it. And then I think it's being transparent about both of those things when we talk about it. I do think that over the long term, the opportunity is going to be significantly higher and greater than... [55:00] and some of the risks and the downsides that will happen. But that doesn't mean it's going to be perfectly smooth on the curve. The release of Mythos was such an interesting moment. It was the first time many people, friends of mine that are careful watchers of this stuff, said something like, this one kind of makes me scared. So it relates back to the safety question. It's also the first example of you coming out and saying, we want to make sure this isn't used for bad. And it's maybe the first one that we are worried could be used for bad. I'm curious what that discussion was like internally before the world heard about it, the decision-making process around it. And just using that as an example to talk about the things that do scare you as we continue to advance and the scaling laws continue to hold. One of the things about Mythos is that people may be misconstrued as just a cyber model. It is a incredibly capable model across many, many different dimensions. What we found was that cyber in particular was a place where it spiked. This was the first model that we decided to release in a different way. The way in which we did that, again, is consistent with our mission, our principles. We wanted to do it in that way. We have this phased approach to it because we think that when a model is this capable, and again, cyber is the thing that people focused on but there are other things as well We think it can be used in a positive way to patch code bases You seen these examples where we had an open source code base that you know a prior model found 22 security vulnerabilities and Mythos then found 250 That is kind of scary, but that informed the way in which we released it. So we didn't say we're never going to release it. We said, let's do it in a phased way. Let's do it to, you know, a group that will expand over time where we can focus on this one cyber capability and how it can actually be used positively in a defensive way as opposed to in an offensive way. And we think that's a template that could be used for the future. But because of this one particular area, we wanted to be cognizant of that in like how we released it. You're so big now that you run into everything and everyone. And one example of this is the government just a couple of days ago said maybe there'd be this new system where you have to sort of pre-approve the release of a new model with the government before it was released to the public. Obviously, you had the crazy experience of the Department of War, which I'm really curious what that was like as you went through it. Now everyone cares about this company and this technology and a couple other companies that are building it. How do you navigate that stuff? Some of it is just, I guess, beyond your control, but I'm sure you're trying to work with people as best you can. Maybe talk about those two examples of the government now as a very relevant partner, player overseer et cetera We prioritize having a strong relationship on this because we do think that regulation has a role to play We been pretty vocal about that We are very America first in our approach. We want the technology to support the U.S. as well as democratic countries around the world. And that's one of the reasons why we've been working closely with the administration. I do think that there's a balance, right? You want to be able to have innovation happen really quickly and have that not be slowed down. But you also want to have this responsibility framework for how these things are deployed, because we've long said that this technology has implications and we should have an honest conversation about them. And that includes with the government. And so I think the mythos process is a good example of that. Can you teach us a bit more about how you would describe their cultural tenets to your parents or something like this? What feels like it really drives most of the culture? I'm especially curious about the writing. You hear often that Dario publishes these long essays every so often externally. My understanding is he does that way more frequently. And there's a lot of writing culture internally. I'm trying to get a sense of what the culture is like to be in and what makes it the most distinctive from other companies, maybe that you've worked at or from other companies that are trying to do the same thing What your sense of the differences and the distinctiveness The culture is a real unique aspect of Anthropic and it is something that we do talk about externally but it different when you're in there living it. I can tell you a little bit about some of my observations. First of all, there's seven co-founders that shouldn't work on paper, but it really does in practice. And I think they've really set the example for the culture and the things that really matter to the company. We take culture extremely seriously. We do a culture interview And it's not some pro forma thing we do just to check a box. It is a real part of the evaluation process. So somebody could be flying colors on everything else. And really, really the smartest person you've met in this role, we won't hire them if they don't pass the cultural bar. And the way I would describe it, I like that frame. How would you describe it to your parents is it's one incredibly collaborative. And this means that we don't really tolerate fiefdoms or the sharp elbows or the I need to take credit for this. It's incredibly humble. We have this sticker on their laptops internally. Our competitors are incredibly capable and success is far from guaranteed. And I think that's part of how the company operates. If we reach a milestone and something good happens, there's not confetti on the floor. It's like, what's next? And I think it's just that focus on the mission and the alignment [60:00] that is imbued throughout the culture of the company. The other thing I would say is there's rigorous debate and intellectual openness and intellectual honesty that happens where people question things. People will express a point of view, but then there's dialogue around it that's productive. And then we'll decide on a path forward. And then after that happens, there's real alignment. So in something like compute allocation, we were talking about before, people might have different perspectives on how to allocate that compute. They will engage in a thoughtful discussion about where the returns are the highest or the best. And when they do that, you know, and we come to a decision, then there's alignment on it. There's not second guessing. There's not politics or fiefdom. The other piece of it is it's remarkably transparent. Dario gets up in front of the company every two weeks, usually writes a short document, and he talks about usually three or four topics and then takes open questions from the company. These are not softballs. They're not planted questions. They're just real questions that are on people's mind. And he answers them the best that he can. It's not a decision making forum, but it is a way for the company to get a window into how leadership is thinking, how he's thinking. And there debate and dialogue in that That is something that people really value It is a transparent culture It is one where all seven of the co are still at the company The vast majority of the first 20 to 30 employees are still at the company. The culture underpins the reason why we've been able to attract and retain some of the best talent in the industry, because we don't always pay people the most. We have very competitive compensation packages. But when Meta and others were out with these huge packages for some of the technical talent across the large language labs, I think we lost two people and other labs lost dozens. What parts of the business and the culture specifically for researchers, why do you think that is true? It really is underpinned by the culture. That's not just something we feel. It's like empirically, when you talk to people, it's I want to have the most impact possible. I want to work in a place where, again, this idea of talent density mattering more than talent mass. And I want to work in a place that is actually collaborative versus I have to like fight for this one thing. And I feel like it wasn't discussed and debated in the right way, or there wasn't transparency around how a decision was made. I think that actually really matters because most of our team just wants to do really, really good work. And they're attracted to the company for the mission The idea of having an impact on a company like ours that is trying to develop this transformative technology but to do it in a responsible way I think that really matters to people not just on the research team, but across the company. And that is a real advantage for us. And it's not something that we take lightly. We have this concept of a race to the top. We don't always have all the right answers. We don't always do everything perfectly. But we want others to look at some of the things we do and maybe emulate some pieces of that and actually have the technology be developed in a better way across the industry. I think people are also really attracted to that as well. Again, not that we have all the answers, but that we can be a part of contributing and leading to how this can go well for humanity. If we now think forward, as you're having conversations with people internally, what does the frontier feel like? I don't just mean the model frontier. I mean, the next couple of rolls of the dice here in building AI in general, everyone is kind of wise to like, these things are powerful, everyone's using them. What feels to you like the frontier from the inside? I think it's this idea. And again, it's because we're focused on enterprise and because, you know, we're really trying to change the productivity of knowledge work that's done in the economy I think it is towards this vision or this goal of like a virtual collaborator And so think of this as something that has context within your organization that can use all of the tools that are specific to you whether they be homegrown tools or tools that you purchase, that has memory and the ability to effectively learn from mistakes you've made, but also mistakes that it's made over time. The ability to work over a very long time horizon on not just a task, but an actual idea. What that means for us is the model capability has to continue to grow to support that. The products we build on top of it can unlock this virtual collaborator that we think can really accelerate knowledge work. There's something like $40 trillion of knowledge work done annually in the world. We think that the productivity and the acceleration of it can happen, but you have to get it in the right form factor. This is where intelligence is not just a single dimension, but the virtual collaborator kind of combines many of those things. Something that's not just generically smart, but it's smart for your use cases. What we're seeing in coding is something that we expect to see elsewhere. For us, Cloud Code has led the way on that as well as much of the business that we have great customers that are pushing the coding frontier as well. But then you also see something like Cowork come along and start to unlock that. Cowork is growing faster than... [65:00] And cloud code was if you index them to the same point in time. That's kind of remarkable because developers are really fast adopters of this technology. But I think it's because the model capabilities and the products are pushing towards this notion of a virtual collaborator, where even our product development today is not done by like one product manager with two engineers shipping something over three months. It's shipping daily. And there's a fleet of agents that are working across the company on a specific task. So everyone kind of becomes a manager. I think the implications of that and the productivity gain that can come from that when it's in the right form factor, we're very, very early in that. But the potential for it is incredible. I'm curious how you've had to personally evolve to be able to stay doing this. You hear a lot about these stories. The executives have to scale with the company. They're all still getting new executives. The business that you were at prior to this was a great business, but it was a tiny, tiny fraction of the scale. You, like everyone, is in this new unprecedented thing. You talked about the example of getting out of linear into a more exponential type thinking. That's one example of what I mean. But how have you managed it personally? What have you had to do? What's been the most painful? How do you manage your own ability to scale with this thing that scaling faster than what we seen before It really hard I think the important thing is to think in first principles Everyone has priors when they come to something new. Thinking in first principles and having intellectual openness. I spent a lot of time with Tom Brown, our chief compute officer. He was actually one of the first people to interview me at the company. And I remember we went on a walk and he started to tell me about his vision for the future of the company. And this is in 2024, early 2024. And I'll be honest, it sounded crazy. He walked me all the way home. I remember I came in and I told my wife, I was like, this is going to be wild. If even 10% of that is true, this is going to bend all paradigms of not just things I've seen, but what most people have seen. And it turns out that a lot of what Tom said during that walk has come to fruition. I remember that as like an early formative thing. This walk I took with him and coming home and being like, holy shit, this is going to be totally different and new, really incredible experience, but also really challenging. And that's what it's been. The other piece of this is just hiring great people. I try to hire people and I tell people during the interview process I like I not really hiring you as like a direct report of mine I hiring you as a partner and I want you to treat it as a partnership which means that there might be things that you and I disagree on I want to hear that I want to whiteboard it I want to understand we hired people from some of the best companies in the world They come to this from a different perspective. They might come to it from a hyperscaler or a large software company or from financial services. In another lifetime, you know, I worked at Blackstone in the private equity group. That training is really valuable and thinking about things at a granular level and not losing that. I'm not somebody who is comfortable at 50,000 feet. That's just not me. But you can't be at 500 feet at everything in this business. There's too much surface area. And so having people who can be partners in that is really, really critical. And I found we've been able to attract really great people. They've pushed my thinking quite a bit. I think the last piece is to think about how the business evolves over time and where there might be moments or analogs to things that have happened in the past. I helped lead the financing that Airbnb did in the middle of the pandemic. Very different situation, right? The business lost 70% of its revenue in seven weeks, and O'Brien did a show with you. That was a harrowing time, but it was also a time kind of without precedent where you had to think about things with a clear perspective when it was rapidly changing and there was not a good template or pattern to match And then on a personal level look it is hard to balance everything family and friends. And this job takes a big bite out of all that. But what I do try to do maybe once a week is in a quiet moment, just think, wow, this is really cool. It's an incredible opportunity to work with this group of people on this problem at this company at this moment in time. Maybe it's in a car ride, maybe it's late at night or something like that. Having that recognition and that appreciation is really important. So I try to bring that to at least my life for a couple minutes once a week. What did Tom tell you on the walk that sounded most crazy? We talked a lot about the scale of the compute infrastructure, what models could do in a short amount of time. I think he described a world that I would have said is kind of sci-fi. A lot of what we're experiencing here and now have really roots in that conversation. And so there's even more things that he talked about that are probably beyond where we are today. But I think the commonality of it was that everything is going to happen much quicker than we think, and that both the implications, but also the capabilities of that can change. And then he also had like a really incredible optimism about [70:00] The future that I think we talk about internally, holding light and shade, that's one of the things we say. I came from that conversation with just a bunch of questions, but also a sense of positivity about what could happen in the future. It seems like we spend most of our time talking about because it's been the reality that we exist at the high end of that cone. What can you imagine that would cause that to change to the low end of that cone? If we were to do like some sort of premortem on a year from now, we're like, wow, actually we didn't need nearly as much compute as we thought or something like that. What can you imagine that would shift us meaningfully in that cone? The capability and the use cases are playing catch up to where the model is. We are talking about humans in large organizations with a set of tools and practices and things that they've been doing for a really long time. Change is hard. To the extent that that diffusion hits a wall or slows down or something like that, that could affect the rate of change in terms of revenue growth. The scaling laws slowing down or not holding, we don't see that, but can't say that with 100% certainty. I think that would be silly. We certainly believe in the trajectory, but the model capabilities leveling off would be another thing Third is how we think about being at the frontier Today we at the frontier I think we defining the frontier of agentic AI We need to stay there It a competitive market and we going to continue to invest in the technology and the compute and the go that required to be there but that's not guaranteed either. To finish on an optimistic note, if those are the things that could cause us to go to the low end, what are you most excited about? You have a privileged seat. You sort of get to literally see the future because it's happening inside the business before those outside the business see it. With that perspective, and in that seed, what are you most excited about? I really think that the biotechnology and healthcare outcomes that can come from this technology are the things that I'm most optimistic about it. We may live in a world where you're diagnosed with a disease that is not curable, but in your lifetime, that cure can be found much more rapidly and you might not die of that disease. A lot of what we're doing today is helping to speed up the drug development process. A lot of the paperwork and clinical studies reports and things like that that are needed to be done. AI and our solutions in particular are helping to rapidly accelerate that. I'm really most optimistic and excited about when it goes further back into drug development and drug discovery because humans are capable of research. But if you think about these molecules and proteins they so complex and such small changes have such big implications for the outcomes AI is perfect for that What can happen when the lab throughput goes up 10x or 100x and we can run that many more experiments, probably get better results faster. And that can be something that helps people around the world. And it doesn't have to be limited to a small set of diseases or disorders that can really go much further down the chain. And so I think that has the potential to greatly alter the way that we live and the way that we interact. And that's really exciting to me. Sure. Hope you're right. It sure seems like we're on that trajectory and it's quite a future to imagine. This is so much fun. I feel like we covered so many interesting aspects of the business. When I do these, I ask the same traditional closing question. What is the kindest thing that anyone's ever done for you? I have a brother who's five and a half years older than me. We lived in California when he went to college and he got into everywhere he applied to and he was going to go to medical school after that. And I didn't know any of this at the time. So he ended up going to college in state. He did exceptionally well. Kind of years later that I kind of had to pull this out of him. But in deciding where to go to college, you know, we were solidly middle class. This was like 25 30 years ago The financial aid packages weren as robust as they are today A big factor in his decision I found out was wanting to give me the opportunity to go wherever I wanted That was six years out, and who knows how I would turn out. A big part of his decision, a big factor in his decision, was giving me the opportunity to have choice. I didn't know that 12-year-old me or 13-year-old me would have never understood. But many years later, I think that's something that was incredibly kind and is something that I still kind of hold with me today. I've done this like 600 times or something. I've never heard an answer of that type. That's awesome. Amazing. Love to meet your brother. Krishnan, thanks so much for doing this with me. Thanks for having me, Patrick. Really enjoyed it. If you enjoyed this episode, visit Colossus.com. You'll find every episode of this podcast complete with hand edited transcripts. You can also subscribe to Colossus, our quarterly print, digital and private audio publication featuring in-depth profiles of the founders, investors, and companies that we admire most. Learn more at Colossus.com slash subscribe. [75:00] You know how small advantages compound over time? That's true in investing and just as true in how you run your company. Your spending system is your capital allocation strategy. Ramp makes it smarter by default. Better data, better decisions, better economics over time. See how at ramp.com slash invest. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com slash invest. The best AI and software companies from open AI to cursor to perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit RidgelineApps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process rogo does it's an ai platform built specifically for wall street connected to your data understanding your process and producing real outputs check them out at rogo.ai invest