Stripe Automates 30% of Code as AI Signups Surge 50%

AI入驻激增50%，Stripe实现30%代码自动化 AI驱动的软件开发对互联网经济的拓展速度超过了传统工具的淘汰速度，促使Stripe扩大内部编程智能体规模并在全球拓展稳定币通道。 Stripe总裁Will Gaybrick详细阐述了公司如何在处理超2万亿美元交易量的同时积极拥抱AI生产力，将内部编程工作流程转向自动化AI智能体（“Minions”），这些智能体目前生成了Stripe 30%的拉取请求（pull request）。Stripe并未利用AI效率来裁减人员，而是赋能技术人才“构建一切”，并观察到新的AI初创公司与SaaS群体变现速度更快，营收同比增幅极为显著。讨论还勾勒了智能体商业的前景——即AI智能体发现API、通过HTTP 402协议支付微额款项，并利用Tempo等稳定币基础设施——最后介绍了Stripe如何通过持续的产品模拟来保持工程品味。 Stripe将AI生产力工具视为大幅提升产出和打造更多产品的机遇，而非单纯用于削减成本或精简工程团队规模。 (0:00) Stripe已从单纯的支付处理商演变为涵盖25至30个品牌产品的多产品金融基础设施平台。 (1:04) AI初创企业的迅速崛起带来了新的财务挑战，例如算力密集型免费试用的滥用，以及覆盖100多个国家和地区的即时全球税务与合规需求。 (1:04) Stripe的核心战略依赖于获取早期初创公司——它们扮演着高标准的“煤矿里的金丝雀”——并在其成长为企业巨头的过程中提供服务。 (5:05) AI能力正推动企业创立呈现前所未有的加速态势，Stripe的2026年入驻客群在同期产生的收入比2025年客群高出50%。 (7:58)

Stripe President Will Gaybrick argues that artificial intelligence is accelerating software creation and business formation rather than commoditizing the tech industry. Consequently, Stripe is orienting its strategy around maximizing developer agency, building infrastructure for autonomous agentic commerce, and treating compute tokens and stablecoins as fundamental economic primitives. AI-Driven Software Expansion and Cohort Growth (7:58) Rejecting predictions that AI automation will diminish software demand, Gaybrick cites internal Stripe cohort data showing unprecedented acceleration in business creation. First-half new business signups grew 50% year-over-year, while the 2026 customer cohort generated 50% more revenue than the 2025 cohort at the same developmental stage—which had already outpaced the 2024 cohort by 70%. Furthermore, new SaaS platform signups on Stripe grew 103% year-over-year. Gaybrick attributes this trajectory to lower baseline software production costs combined with an expanded frontier of technical feasibility. Rather than reducing the total addressable software market, agent-assisted development allows lean startups to build complex products that were previously impossible and adopt multiple financial infrastructure products earlier in their life cycles. Engineering Leverage and Injection Molding Workflows (14:25) Instead of using AI efficiency to downsize engineering headcount, Stripe focuses on turning senior engineers into high-agency builders who oversee end-to-end execution. Stripe's internal agent framework, Stripe Minions, executes one-shot code generation, runs tests, and submits ready-to-review pull requests. Minion output scaled from 1,200 pull requests per week to 7,000 per week, representing 30% of all pull requests merged across the company. Similarly, an internal AI knowledge tool named Kai achieved 83% weekly active usage, lifting sales productivity by 20% and prompting Stripe to hire more sales representatives via Jevons' paradox. Gaybrick compares modern engineering to industrial injection molding: senior developers build standardized architectural templates and repository markdown context files, enabling AI agents to synthesize integration code in parallel. This methodology allowed Stripe Tax to expand from US-only filing to global filing across diverse jurisdictions in one-third of the original development time, while flatter team structures allow single engineers to coordinate up to 16 agents simultaneously. Treating Compute Tokens as Economic Units (43:10) With Stripe processing more than $2 trillion in total payment volume, Gaybrick observes that AI model tokens are functionally operating as a primary medium of economic exchange. Unlike traditional software with negligible marginal serving costs, AI inference incurs direct compute expenses per call. This dynamic introduces severe financial-style abuse vectors, such as model distillation farming; for example, Cursor discovered that roughly one in six free-trial signups was abusive, prompting Stripe to deploy foundation-model defense pipelines that now help ElevenLabs block over 2,000 fraudulent signups daily. Because token consumption directly drives unit economics, Gaybrick contends that financial platforms must treat token tracking as a core operational mandate. Companies increasingly require tooling to manage token usage across two distinct vectors: tracking internal developer usage as operating expenses (OpEx) and monitoring customer-facing model routing as cost of goods sold (COGS) to preserve gross margins. Autonomous Agent Commerce and Stablecoin Settlement (28:00) Gaybrick anticipates that digital commerce will transition from human-driven, recurring monthly software subscriptions to ephemeral micro-consumption executed by autonomous agents. To remove transaction friction for machine-to-machine interactions, Stripe co-developed Tempo and the Machine Payments Protocol using HTTP 402 payment specifications, alongside the Link Agent Wallet and Link CLI, which let agents programmatically transact using Stripe Link's base of 400 million consumer profiles. To facilitate lightweight, high-frequency transactions across borders, Stripe relies on stablecoins as a neutral global settlement layer, expanding its operational footprint from 60 countries under fiat rails to 150 countries. Addressing the limitations of general-purpose blockchains—specifically speculative network congestion, unpredictable gas fees, and public ledger transparency—Stripe is developing Tempo as an enterprise-grade payment rail optimized for privacy and fee stability, with early partners including DoorDash. Preserving Product Craft Through Simulation (51:06) As AI agents handle an expanding share of routine coding, Gaybrick argues that maintaining design quality and product craft requires concrete operational mechanisms rather than top-down slogans. While acknowledging the skeptical view that invoking "taste" is merely human justification against automation, Stripe views product elegance as critical for user retention and engineering morale. To institutionalize quality across sprawling product suites, Stripe uses flight-simulator-style environments. These test setups strip out personally identifiable information, randomize growth metrics, and generate authentic merchant friction such as automated chargebacks and dispute spikes. Engineering Managers are required to run these simulations firsthand, exposing them directly to user-facing friction and allowing them to allocate upcoming engineering sprints specifically toward refining product UX. Gaybrick's thesis assumes that AI productivity gains will permanently expand software market demand rather than commoditize software pricing power. While Stripe's internal cohort data demonstrates accelerating startup formation and product adoption, the long-term viability of the agentic economy depends on industry-wide standardization around machine payment protocols, effective mitigation of token fraud, and the enterprise adoption of stablecoin infrastructure over legacy financial rails.

Stripe产品与业务总裁Will Gaybrick指出，AI不仅未导致软件商品化或萎缩，反而大幅降低了构建成本并推动了新一轮软件爆发；随着处理规模突破2万亿美元，Stripe正通过内部代码智能体、稳定币网络以及面向机器的微交易基础设施，将业务重心由传统支付扩展为全周期的金融与代币生态。 AI推动软件爆发与Stripe队列扩张 (7:58) 针对AI将导致垂直SaaS或软件行业走向终结的市场论点，Gaybrick依据Stripe平台数据提出了相反的判断。Stripe的2026年上半年注册量同比增长50%，其中新SaaS平台签约量相比2025年队列增长103%；按同期水平衡量，2026年新客户队列产生的收入比2025年队列高出50%，而2025年队列亦比2024年高出70%。Stripe Billing的增长速度甚至超过了整体网络交易规模，显示出软件原生业务模式的集中涌现。 Gaybrick将这一加速归因于软件边界的扩张与代码构建成本的急剧下降。如同Suno或Higgsfield这类依靠生成式能力的新产品在数年前根本无法成立，如今更低的开发门槛正直接激发现代企业的复合需求。以ElevenLabs为例，其已在平台上接入14款Stripe独立产品；初创团队正以远快于传统SaaS周期的节奏，在成立早期便整合计费、税务、合规与反欺诈等多维系统。 敏捷工程演进：从代码生成到组织瓶颈 (14:25) 在内部运作方面，Stripe选择借助AI将资深工程师转变为具备高度自主权的“内部创始人”，而非以此裁员削减运营支出。其内部构建的单次执行智能体系统“Stripe Minions”允许工程师提交提示词后自主完成构建、运行CI/CD管道、执行单元测试并提交审查就绪的PR；截至录制时，Minions每周生成约7,000个PR，占Stripe全公司合并代码请求总量的30%，高于年初的1,200个。在Stripe Projects的开发中，一名资深工程师通过终端同时调度16个AI智能体，协同极少数成员在数周内完成上线。 开发提速使瓶颈从代码编写转移至制度与后台系统，如内部审批、定价页面更新与销售协同。为应对内部知识阻滞，Stripe由两名工程师用时约六个月开发了内部AI知识工具Kai，在全公司取得83%的周活与约60%的日活，使销售团队生产力提升20%。Gaybrick指出，Stripe将工程效率提升视作“注塑成型”，团队专注于搭建代码规范与Markdown上下文作为“模具”，使Stripe Tax等复杂合规工具的全球报税支持开发周期缩短至原先仅支持美国本土时的三分之一。 机器商务、HTTP 402与稳定币清算网络 (28:00) 随着自主智能体接管越来越多的日常任务与B2B采购，传统由人类填写表单、绑定按月订阅的模式面临摩擦。Gaybrick认为，智能体在网络中像蜂鸟一样频繁抓取数据、调用瞬时存储或执行短时算力，促使互联网必须从固定的9.99美元月度订阅演进为动态的、经授权限额内的即时微额消费。为此，Stripe参与推出了机器支付协议（Machine Payments Protocol）与Tempo，利用HTTP 402状态码为机器服务声明付款请求与自主结算提供统一协议，并基于覆盖超4亿用户的Stripe Link推出了面向代理的Link Agent Wallet与Link CLI。 传统卡组织的结算成本使微额支付长期难以经济落地，而稳定币提供了低成本、无国界的清算层。Gaybrick对比指出，印度本土即时支付UPI使5美元以下交易占比达86%，而美国信用卡网络该比例仅为个位数。Stripe通过将稳定币原生整合进Stripe Treasury，把跨境支付覆盖范围从法币的约60个国家拓展到约150个国家。以美墨走廊的汇款服务Felix为例，其完全基于稳定币轨道运营，仅用数年时间即夺取了该走廊5%至10%的汇款份额。针对公链存在的投机交易拥堵、动态Gas费剧烈波动及隐私泄露缺陷，Stripe打造了专为支付设计的Tempo区块链，当前已获得DoorDash等企业合作伙伴的测试采用。 代币经济学的安全防线与产品品味仿真 (43:10) 在模型调用计费普及的背景下，Gaybrick提出“代币（Token）正在成为新的美元”。平台（如Cursor、Replit）消耗AI代币作为经济媒介，但也引来了类似于传统金融网络的多账户攻击、模型蒸馏与试用期滥用。以往微不足道的边际服务器成本在AI调用下转变为真实算力支出，例如Cursor早期面临高达约六分之一的试用注册滥用。Stripe在周末搭建了基于基础模型、全网嵌入与推理层的防御流水线，目前帮助ElevenLabs等客户每日拦截超过2,000次试用滥用，并将代币使用纳入开发OpEx与产品销货成本（COGS）的双重监控体系。 针对AI时代如何维持产品高“品味”与质量标准，Gaybrick指出管理层除了自上而下反复强调质量之外，关键机制在于环境仿真。借鉴航空飞行模拟器及NVIDIA在芯片流片前的仿真机制，Stripe为工程主管创建了剔除个人身份信息（PII）、模拟随机增长曲线、退款与争议波动的真实商户模拟账号。由于工程经理掌控冲刺周期排期，通过直接承受模拟运营中的断点摩擦，团队得以主动将工程周期分配给消除UX瑕疵，而非仅由指标驱动交付。 Gaybrick的讨论勾勒出AI推动下的软件供应扩张与新型微额基础设施蓝图，但在机器自主支付完全爆发前，仍面临通用公链性能限制、跨国合规摩擦以及机器与人类交接验证等过渡性瓶颈。

## Background 本期视频由知名风险投资机构 a16z 于 2026 年 8 月 20 日发布在 YouTube 频道。内容由 a16z 普通合伙人 David George 对话 Stripe 产品与业务总裁 Will Gaybrick，探讨 AI 正在如何颠覆 Stripe 的产品研发机制、团队组织阵型，以及驱动底层互联网商业、智能体电商和稳定币金融基础设施的全面重构。 ## TL;DR Stripe 正从最初的单一支付处理器跃迁为承载企业全生命周期收入的金融基础设施平台，并在 AI 浪潮下坚定推行以软件扩张突破瓶颈的“建造一切”策略。通过内部自动化编程 Agent“Stripe Minions”，公司每周自动生成数千个 PR，极大释放了开发吞吐量，并重塑出扁平且高能动性的组织模式。面对未来，Will 深入拆解了机器对机器交易的智能体电商（Agentic Commerce）、消灭传统结账页面的瞬态微消费模式，以及基于稳定币构建全球无国界资金流转网络的时代机遇。同时，面对代码生产力井喷，Stripe 借助高仿真沙箱环境和自上而下的匠心文化，打破虚幻概念将高标准产品品味深度规模化。 ## Atmosphere & style 整场对话沉浸在顶级科技机构之间极高信息密度的前沿研讨氛围中，充满理性思辨与工程师文化的硬核极客感。主持人 David George 提问精准且深谙技术投资脉络，Will Gaybrick 的回答则从容笃定、金句频出，既有对行业流行“内卷式降本”的犀利调侃与哲学反思，又有拆解底层架构时的扎实与坦诚。两人在探讨“Token 即新货币”与“品味本质”时默契交锋、笑声不断，让听众不仅能真切触摸到硅谷头部公司重塑组织与代码生产的滚烫现实，更感受到一种极其强烈的、坚定“做多未来”的野心与澎湃张力。 ## Bulletin - [14:26](https://www.youtube.com/watch?v=P5iICDVn5gc&t=866s) Stripe 内部自动化编码智能体“Stripe Minions”单周生成的 PR 数量已从 1,200 个暴增至约 7,000 个，占全公司当周总 PR 量的近 30%，研发瓶颈全面从写代码转移至中后台系统流程。 - [12:23](https://www.youtube.com/watch?v=P5iICDVn5gc&t=743s) 面对 AI 带来的生产力跃升，Stripe 拒绝以裁员降本为主的“做空未来”策略，反而践行“建造一切（build everything）”的进攻姿态，利用杰文斯悖论全面扩产与加速新业务落地。 - [02:30](https://www.youtube.com/watch?v=P5iICDVn5gc&t=150s) 由于 AI 推理具有实际边际成本，免费试用滥用成为新型恶意黑产；Stripe 借助全网嵌入与基础大模型构建归因风控防御，帮助 ElevenLabs 实现单日拦截 2,000 名试用滥用者。 - [28:02](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1682s) 智能体商业（Agentic Commerce）的瓶颈在于缺乏机器对机器的原生原语，Stripe 正在联合推进“机器支付协议（MPP）”唤醒 HTTP 402，并推动结账页面彻底消亡和基于稳定币的瞬态微支付复兴。 - [38:32](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2312s) 稳定币正演变为全球资金流转的“谢林点”，Stripe Treasury 接入稳定币后将覆盖范围从约 60 个国家/地区直接外扩至约 150 个，并与合作伙伴联合打造面向支付优化的专有区块链 Tempo。 ## Key takeaways - [01:04](https://www.youtube.com/watch?v=P5iICDVn5gc&t=64s) Stripe 已完成核心定位逆转，从“提供附加特性的支付工具”跃迁为覆盖收入与现金流全生命周期的金融基础设施平台，平均每家 AI 初创企业深度集成其 11 款产品。 - [02:30](https://www.youtube.com/watch?v=P5iICDVn5gc&t=150s) 传统软件的免费试用成本可忽略不计，但 AI 算力推理具有硬性成本，这让免费试用黑产与模型蒸馏成为严峻威胁，风控防御必须依托全网数据网络与归因大模型进行精准拦截。 - [05:04](https://www.youtube.com/watch?v=P5iICDVn5gc&t=304s) 初创企业不仅是未来的商业巨头，更是要求最严苛、反馈最敏锐的“金丝雀”，Stripe 坚守“赢得所有初创公司，然后再次赢得它们”的商业飞轮。 - [07:55](https://www.youtube.com/watch?v=P5iICDVn5gc&t=475s) 随着大模型拓展商业可行性版图以及智能体大幅降低软件构建门槛，Stripe 平台新增企业与订阅计费客户呈现井喷，反驳了“AI 导致软件同质化贬值”的虚无论调。 - [09:46](https://www.youtube.com/watch?v=P5iICDVn5gc&t=586s) 单人资深工程师产能匹敌以往两个团队，传统管理机制已“无处可抄”，企业必须转向赋能“内部创始人”的高能动性（High Agency）模式并大幅压缩中层协调链条。 - [22:42](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1362s) 将 AI 提效用于裁员和缩减 Opex 是在“做空自己的未来潜能”，唯有像二战后“注塑成型”技术革命那样重构范式并全力加速研发交付，才是真正的“做多未来”。 - [19:41](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1181s) 当生产要素效率提升时总需求反而爆发，这就是“杰文斯悖论”；内部 AI 知识工具 Kai 提升人效后，Stripe 不仅没有缩减销售编制，反而因回报周期缩短而进一步加大招聘。 - [22:42](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1362s) 软件工程演变为由人类定义模具（如规范指南与 Markdown 模版）、智能体负责高频注塑产出并进行自动化代码审查的新型工业化协作形态。 - [29:10](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1750s) 当前基于浏览器自动化去模拟人类填表的智能体电商属于“拟物化过渡方案”，未来的终局是机器原生直付与传统面向人类的结账页面彻底消失。 - [30:54](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1854s) AI 智能体在长任务中像蜂鸟一样穿梭于互联网，迫切需要告别固定月度订阅，全面转向预算控制下的瞬态、一次性微消费 API 与极速清算。 - [34:22](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2062s) 稳定币是解决跨主权边界低成本微支付的终极武器，相比被高额固定刷卡费压制的传统信用卡，它在底层成本与结算效率上具备代际优势。 - [41:35](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2495s) 通用公链由于投机交易容易出现拥堵与 Gas 费暴涨，无法支撑严肃商业支付，打造原生重视隐私保护与交易费可预测性的专有区块链 Tempo 具有高度必要性。 - [43:10](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2590s) Token 正在成为 AI 时代的通用货币，针对 Token 的套现与滥用呈现出传统金融黑产特征；企业需要像管理美元一样管理作为开发 Opex 和模型能效的核心 Token 资产。 - [48:06](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2886s) 构建企业软件的核心壁垒在于让业务抽象贯穿公司底层运转，单纯靠模型手写代码无法替代长期沉淀的收入技术栈与合规抽象能力。 - [49:00](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2940s) 所谓的“品味（Taste）”不应成为人类面对 AI 时的防御性借口，在组织规模化扩张中，高品味必须依靠管理层不厌其烦的宣讲与高拟真的业务沙箱仿真系统来落地维系。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=P5iICDVn5gc&t=0s) 开场预告：AI 重构软件工程与商业范式 播客开篇由一段极具冲击力的高光对话切入：在 AI 时代，企业若想以更快的速度交付更多产品，就必须在组织内部建立起“像创始人一样的自主权（founder-like agency）”。如今，得益于智能体编程（agentic coding）的赋能，单个工程师所能产出的成果，已经足以匹敌两年前两个工程团队的工作量。面对这种效率红利，业界出现了两条截然不同的路径：很多公司将智能体带来的效率提升单纯视作缩减成本结构的手段；而 Stripe 的哲学截然相反，Will Gaybrick 带着些许调侃却无比严肃地提出——去“构建一切（build everything）”。 伴随着代码合并量的激增，Stripe 内部系统承受着空前压力，为此他们推出了名为“Stripe Minions”的自动化代码编写智能体，将代码合并效率推向全新维度。与此同时，在商业基础设施底层，Stripe 年处理交易量已突破 2 万亿美元。从底层模型调用到未来无需收银结账页面的智能体电商，AI 正在彻底重塑软件被创建、被使用以及被购买的整个链条。 ### [01:04](https://www.youtube.com/watch?v=P5iICDVn5gc&t=64s) 今日 Stripe：从支付工具跃迁为金融基础设施平台 a16z 合伙人 David George 将话题引向 Stripe 当前的发展全貌。回溯 a16z 早期投资 Stripe 时，它本质上还是一个简化线上收款的支付处理器（payment processor）。而如今，平均每家 AI 公司要使用 11 种不同的 Stripe 产品；在近期的 Stripe Sessions 大会上，团队更是创下了单次发布 288 项独立产品/功能的惊人纪录。 ```mermaid flowchart TD accTitle: Stripe 核心价值定位的跃迁与产品矩阵 accDescr: 展示 Stripe 从最初单一的支付网关演化为覆盖企业全生命周期收入与现金流的金融基础设施平台架构。 subgraph Core["核心战略定位"] direction TB Friction["减少业务摩擦 (Reduce Friction)"] Agency["增强商业自主权 (Increase Agency)"] end subgraph Products["核心产品矩阵 (25-30 款主力产品)"] direction LR P1["Payments & Billing<br/>(支付与订阅计费)"] P2["Stripe Connect<br/>(平台与分润网络)"] P3["Radar & Tax<br/>(欺诈防御与税务合规)"] end subgraph Value["赋能客户增长"] V1["极速响应商业模式演进"] V2["一键合规扩张至全球市场"] V3["全周期掌控收入与现金流"] end Core --> Products Products --> Value ``` 针对这种惊人的迭代速度，Will Gaybrick 阐释了 Stripe 内部价值主张的“倒置”：Stripe 已经从“一家提供附加功能的支付公司”，演变为“一个专注于金融基础设施的多产品平台”。其核心框架始终围绕两件事展开：“消除摩擦（reducing friction）”与“赋予自主权（increasing agency）”，帮助客户更灵活地调整商业模式、开拓跨国市场，并在所有涉及收入与现金流的环节跑得更快。目前，Stripe 拥有 25 到 30 款标杆级产品（如 Payments、Billing 订阅账单、Connect 平台分润、Radar 欺诈风控、Tax 税务等），以及支撑其下的数千项细分特性。 - **平均采用率**：平均每家 AI 初创企业深度集成 11 款 Stripe 产品。 - **发布密度**：近期 Stripe Sessions 集中推出了 288 项新产品与功能更新。 - **主力矩阵**：拥有 25 至 30 款核心品牌产品，覆盖全球数以万计的底层业务能力。 ### [02:30](https://www.youtube.com/watch?v=P5iICDVn5gc&t=150s) AI 时代的防御战：破解免费试用滥用与全球落地摩擦 在“消除摩擦”层面，Will 举了一个极具时代特征的痛点：免费试用滥用（free trial abuse）。在生成式 AI 爆发前，由于传统软件拥有极高毛利，羊毛党消耗的计算资源成本微不足道（de minimis）；但 AI 应用的每次推理调用都有实打实的算力成本。AI 代码编辑器 Cursor 是 Stripe 客户中最早遭遇此问题的先锋之一。当时平台上大约每 6 个试用用户中就有 1 个属于恶意注册，这些黑产用户通过不断注册新账号薅取算力额度，甚至用于模型蒸馏（model distillation）。 Stripe 团队迅速与 Cursor 团队并肩“扎进掩体”，仅用一个周末就基于自身的基础模型与全网嵌入向量（embeddings）搭建了一套防御流水线，并在其上叠加了推理层，能够明确给出“该账号被判定为试用滥用者，是因为……”的具体归因信号。语音生成巨头 ElevenLabs 最近向 Stripe 反馈，他们借助这套风控信号，每天能够精准拦截约 2,000 名试用滥用者。这种基于网络效应的风控能力，不仅止住了算力“失血”，更避免了误伤正常付费转化。 而在“赋予自主权”层面，针对数字原生 AI 企业跨国扩张极其繁琐的税务难题，Stripe 推出了托管支付服务（Stripe Managed Payments）。以 SEO 巨头 Ahrefs 为例，在其欧美主要本土市场，他们保留自己的登记销售方（Seller of Record）身份；而在哈萨克斯坦等长尾或亚太新兴市场，Stripe 则直接充当登记商户（Merchant of Record），代为承担税费计算、申报与缴纳，使企业能够合规无缝地覆盖全球 100 多个国家与地区。 ```mermaid flowchart LR accTitle: 基于 Stripe 全网信号的 AI 免费试用滥用防御流水线 accDescr: 展示用户注册请求如何经过 Stripe 全网网络嵌入、基础模型与推理层，最终阻断算力滥用并输出归因。 User["用户尝试注册试用<br/>(Cursor / ElevenLabs)"] --> Net["Stripe 全网网络信号<br/>& 嵌入向量数据库"] Net --> Model["Stripe 基础风控大模型"] Model --> Reason["推理归因层<br/>(输出具体判定证据)"] Reason --> Decision{"判定性质"} Decision -->|正常用户| Pass["放行试用 & 转化付费"] Decision -->|滥用者| Block["精准拦截 (如拦截模型蒸馏)"] ``` - **滥用比例**：早期 AI 工具的免费试用群体中，一度有高达六分之一（约 16.7%）属于恶意滥用账户。 - **实战拦截效果**：ElevenLabs 借助该功能实现单日拦截 2,000 起免费试用薅羊毛行为。 - **全球合规覆盖**：通过托管支付与商户代理服务，帮助出海企业一站式打通 100 多个司法管辖区的税务合规。 ### [05:04](https://www.youtube.com/watch?v=P5iICDVn5gc&t=304s) 客户策略：“赢得所有初创公司，然后再次赢得它们” 谈及客户群体，David 指出 Stripe 起步于服务早期初创企业，曾伴随 YC 孵化期的 DoorDash 和 Instacart 一同壮大，如今则在不断向大型企业渗透。Will 引用了 Clerk 创始人 Colin 在 X 上的精辟概括来阐明其战略：“Stripe 的策略毫不掩饰，那就是——赢得所有的初创公司，然后再次赢得它们（win all the startups and then win them again）。” 这背后的商业逻辑在于：初创企业最具野心，往往会成长为未来的商业巨头，同时也是洞察行业演进的“金丝雀”。更微妙的一点在于，初创公司反而是所有客户中“要求最严苛、标准最高”的群体。Will 分享了一个耐人寻味的细节：在查看 Stripe 报表功能的客户满意度（CSAT）时，大企业客户普遍赞不绝口，夸赞其数据比以往使用的传统老牌软件出色得多；而初创团队却直接在反馈中毫不留情地痛骂：“你们的报表简直就是垃圾，赶紧给我修好，我都快被逼疯了！”正是初创企业的敏锐与苛刻，倒逼 Stripe 保持极高的产品进化节奏。 目前，Stripe 正在逼近服务全美财富 500 强（Fortune 500）企业半壁江山的里程碑，客户已涵盖亚马逊、微软等巨头。尽管大企业的销售周期更长、合同签订后的落地交付更为重型，但底层逻辑与服务初创企业完全一致：紧跟用户需求，用实打实的核心业务指标证明自身卓越的价值。 - **战略核心**：“赢得所有初创公司，然后再次赢得它们”的双重飞轮。 - **市场渗透度**：财富 500 强企业的合作比例已接近 50%。 - **用户画像差异**：传统大企业受制于对旧系统的低预期而容易满足，敏捷型初创企业反而具有最严苛的产品敏锐度与纠错推力。 ### [07:55](https://www.youtube.com/watch?v=P5iICDVn5gc&t=475s) AI 队列大爆发：软件创建门槛瓦解与新企业潮涌 在探讨业务增长时，David 列举了一组惊人的公开数据：Stripe 在当年前半年的新注册用户数同比激增了 50%；2026 年新入驻队列企业的收入中位数比 2025 年同期队列高出 50%，而 2025 年队列又比 2024 年高出 70%。这种加速式的陡峭爬升，折射出整个软件行业的底层巨变。 Will 分析认为，这一爆炸性增长由两大驱动力共同促成： 第一，**商业可行性版图的大幅外扩**。大模型能力让许多在几年前根本不可能落地的商业构想变为现实。例如四年前根本无法做出像 AI 音乐生成平台 Suno 或 AI 视频生成平台 Higgsfield 这样水准的产品，AI 带来了全新的产品形态与商业切入点。 第二，**软件构建成本的雪崩式下降**。由于智能体辅助编程的普及，如今开发一款产品所需雇用的工程师数量大幅减少。门槛的降低直接带来了全新软件公司创立潮的井喷。这一趋势直观地反映在数据上：Stripe Billing（订阅计费）的使用量同比增速显著超越了 Stripe 大盘的整体增速，因为新增入驻的企业正压倒性地向新型软件公司集中。 ### [09:46](https://www.youtube.com/watch?v=P5iICDVn5gc&t=586s) 借鉴商业巨头的管理机制与“无处可抄”的全新现实 在谈及 Stripe 究竟如何维持如此惊人的产品交付速度时，Will Gaybrick 指出，虽然“对用户的极度专注”永远是公司的第一准则，但“速度”紧随其后。长久以来，Stripe 一直将自身视为一个让优秀制度沉淀的容器，不断向那些经历过历史考验的商业巨头“偷师学艺”：Stripe 改编并采用了类似 Google 的 OKR 目标管理机制；销售团队的组织方式在很大程度上借鉴了微软；内部推行类似苹果的 DRI（Directly Responsible Individual，直接责任人）文化，并沿用了苹果式严苛的产品质量标准与评审流程；而 Will Gaybrick 自己日常用来管理业务运转的机制，更是直接借鉴了前波音民机及福特汽车 CEO 艾伦·穆拉利（Alan Mulally）的管理方法论。通过汇聚这些顶级机制，Stripe 旨在建立一个生命力超越任何单个个体的长青机构。 然而发展到当下这个节点，Will Gaybrick 感叹他们进入了一个“再也没有榜样可以抄袭”的新阶段。当一名资深工程师凭借 AI 赋能就能完成两年前两个工程团队的工作量时，现有的大企业管理模板全部失效了。Stripe 得出的心智模型是：它必须在内部成为一个为“内部创始人”提供支撑的赋能平台，正如它一直以来对外为创业者提供商业基座一样。事实上，创始人文化一直在 Stripe 内部极具生命力： - 被收购的计费基础设施公司 Metronome 联合创始人兼 CEO Scott Woody，如今在 Stripe 掌管着整个 Metronome 以及 Stripe 庞大的 Billing 计费业务。 - 钱包基础设施公司 Privy 与稳定币编排平台 Bridge 被收购后，Bridge 的联合创始人 Zach Abrams 负责领导 OpenUSD 与 Bridge 业务，Henry 负责领导绝大部分加密业务，Privy 的联合创始人 Aayush 则主导着加密工程团队。 - Lemon Squeezy 创始人 JR Farr 目前挑起大梁，领导 Stripe Treasury 团队。 ### [12:23](https://www.youtube.com/watch?v=P5iICDVn5gc&t=743s) 降本裁员还是“建造一切”：以扩张突破组织瓶颈 面对高级工程师生产力呈指数级爆发的局面，企业通常面临两种截然不同的战略抉择。许多硅谷大厂和初创团队将 Agent 带来的效率提升视作压缩成本结构的契机，通过裁员与削减运营支出（Opex）来让组织变小。但 Will Gaybrick 明确表示，Stripe 采取了截然相反的路径——“开个半开玩笑的说法，我们的策略是去‘建造一切’（build everything）”。在他看来，优化成本结构的终极手段不是收缩，而是以更高的速度扩张并实现业务增长；Stripe 手头多年来积压的用户需求堆积如山，工程师能力的大跃进恰恰是消化这些长期未满足需求的绝佳历史机遇。 要让团队真正抓住这一红利、交付更多产品，企业就必须在内部释放出媲美创始人般的自主能动性（founder-like agency）。Will Gaybrick 直言，目前制约 Stripe 研发推进的瓶颈早已不在于代码编写本身，而在于下游的各类“后台与中台系统”。因为前端合并的代码量相比去年呈现井喷，直接对全链路带来了极限施压： - 新功能如何及时同步并录入内部销售人员的系统？ - 定价页面如何快速更新与调整？ - 当销售人员甚至来不及接受系统培训时，新产品该如何有效推向市场？ 从产品创意的萌芽、用户反馈的归集，到产品最终交付到用户手中，整条关键路径上的阻碍必须被逐一清除，必须让每一位创造者像电影导演或创始人那样大展拳脚，而不被中心化流程的泥潭所束缚。 ```mermaid flowchart TD accTitle: Stripe Code Velocity Bottleneck Flow accDescr: A diagram showing how increased code output from AI agents shifts organizational bottlenecks to back-office systems. subgraph Development["研发提速层"] A["资深工程师 + AI Agent"] --> B["代码合并量激增 (PRs 暴涨)"] end subgraph Friction["中后台承压与瓶颈"] B --> C["销售系统集成滞后"] B --> D["定价页面更新跟不上"] B --> E["销售团队培训脱节"] end subgraph Solution["Stripe 应对举措"] C & D & E --> F["全流程关键路径优化"] F --> G["赋予内部人员创始人级能动性"] end ``` ### [14:26](https://www.youtube.com/watch?v=P5iICDVn5gc&t=866s) Stripe Minions 与一人调度 16 个 Agent 的组织重构 为了彻底释放开发端生产力，Stripe 内部开发者效能团队（DevProd）打造了名为“Stripe Minions”的自动化编码 Agent 工具。与让人反复推敲计划、反复多轮对话的辅助插件不同，Minions 被设计为高度自动化的“One-shot”（一次性生成）机制：工程师只需要输入一条 Prompt 描述需求，Minions 就会自动编写代码、触发 CI/CD 流程、完成全套测试，最后直接提交 Pull Request（PR）等待人类进行最终审查。Stripe 内部将“Minions 贡献的 PR 绝对数量与占比”作为最高优先级的工程效能指标之一。数据表现尤为惊人： - 在当年一、二月份公司内部发布相关博客时，Minions 每周生成的 PR 数量约为 1,200 个。 - 到了本期播客录制的前一周，Minions 单周产出的 PR 数量已经跃升至约 7,000 个，占 Stripe 全公司当周总 PR 量的近 30%。 在组织阵型上，由于工程师往往集工程师、产品经理、设计师三重角色于一身，Stripe 顺应这一趋势将组织进一步极度扁平化，大幅削减了过去负责“工作编排与协调”的管理中层。以最新推出的 Stripe Projects 为例，该项目核心仅由一名 PM 与一名核心资深工程师（Alexander）在短短数周内主导搭成（另有少数工程师提供阶段性协助）。核心开发者 Alexander 在工作中直接通过屏幕同时编排调度 16 个 Agent 协同推进。针对团队规模的讨论，Will Gaybrick 认为当前的趋势并非单纯缩减编制，而是“杠杆放大”：管理跨度或许保持不变，但一个 8 人的基层团队，在 AI 的加持下已能够以以前 3 倍的吞吐量同时推进 3 项重大任务。 ### [18:45](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1125s) 对内死磕后台优化 vs 对外打造客户产品 David George 提到了行业当前的核心争鸣：市场上充斥着大量“内向型”叙事，即认为企业应当利用 AI 重新梳理内部流程，把一切行为结构化为 Markdown 文件、在后台穷尽一切手段做内部降本增效；而 Stripe 的选择则是大步向外，将 AI 的核心力量直接投射到构建全新、面向客户的外部产品上。对此，Will Gaybrick 剖析了两种思路背后的底层哲学，指出现阶段虽然利用 Agent 提高人效、用更少的人做同样的事确实有其商业合理性，但单纯沉迷于内部成本优化的策略往往会丧失对未来更大市场机会的捕捉力，Stripe 将继续坚定贯彻以产品创新为核心的进攻姿态。 ### [19:41](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1181s) 杰文斯悖论与内部工具 Kai：工程师越高效，反而越需要更多人 Stripe 内部对 AI 生产力的思考始终围绕着“如何让员工更具生产力”展开，而非单纯为了缩减人力。Will Gaybrick 分享了一个典型范例：大约六个月前，仅由两名工程师构建的内部知识型 AI 工具“Kai”，如今在 Stripe 内部已达到 83% 的周活跃度（WAU）和约 60% 的日活跃度（DAU）。该工具直接让销售团队的人均生产力提升了 20%。但 Stripe 的管理层并未因此认为“可以少雇销售”，反而一致判定“我们需要招聘更多的销售人员”——因为销售团队的投资回报周期（Payback Period）大幅缩短，单位投入的产出效率变得极其诱人。同样，运营团队的日常工作体验也大幅改善，重复繁重的体力活减少，工作满意度显著上升。 这种现象正是科技圈常谈的“杰文斯悖论”（Jevons Paradox）：当某种要素的使用效率提升、成本下降时，市场对其总需求并不会减少，反而会呈现爆发式增长。Will Gaybrick 指出，软件工程师的现状也是如此。尽管外界曾充斥着“AI 会消灭程序员”的悲观论调，但 Stripe 看到的现实完全相反：软件的创造速度正在以不可思议的节奏飙升。以垂直 SaaS 平台为例，即便九个月前市场曾唱衰 SaaS 平台的前景，但 Stripe 平台上 2026 年新成立的 SaaS 平台群组规模，相较于 2025 年激增了 103%。这意味着越来越多的垂类行业正在被新建的 SaaS 软件覆盖，软件构建的速度正变得越来越快。 对于初入职场的应届生与年轻工程师而言，Stripe 正演变为一个鼓励高能动性（High Agency）的内部孵化器。公司负责修筑平整的跑道与底层基础设施（Paved Paths），让工程师可以毫无阻碍地探索贴近用户痛点的创新项目，在现有业务的临近领域（Adjacencies）不断叠加复利。例如，与资金管理（Treasury）紧密相邻的“支出管理”（Spend Management）功能原本排在两年后的产品路线图上，但一名高行动力的工程师主动接手立项并全权推进，快速将其实现。Will 打趣提到，自己甚至为该项目提交了最初的代码，随后项目便交到了“更具能力的人手中”。 - **内部知识工具 Kai**：2 人团队耗时约 6 个月搭建，全员 WAU 达 83%，DAU 达 60%，推动销售人效提升 20%。 - **SaaS 平台爆发**：Stripe 平台上 2026 年的新增 SaaS 平台客群规模较 2025 年同比大涨 103%。 - **资源扩展原则**：短期内“用现有资源做更多的事”（Doing more with what we have），长期看则是“用更多资源做成倍的事”（Doing more with more）。 --- ### [22:42](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1362s) 软件工程的“注塑成型”时刻：做多未来 vs 做空未来 谈及如何在削减成本与开发新产品之间做权衡，David George 指出，创始人之所以成为后期风投最核心的“资产类别”，正是因为他们能够不断带领企业开辟下一代增长曲线。Will Gaybrick 赞同这一观点，并表示管理层虽然适度施加“资源回压”（Resourcing Backpressure）有助于逼出更高阶的工具利用效率，但本质上绝不能落入缩减成本的下行螺旋。用户对产品的需求是无限的，这驱动着 Stripe 必须以更快的速度交付更多功能。 一个直观的对比是财税产品 Stripe Tax。当年为了跑通美国各个州完全自动化报税，Stripe 团队耗费了大量时间去攻关错综复杂的地方法规与提报流程；而随后面对全球各个国家千差万别的报税诉求时，在复杂度成倍提升的情况下，团队仅用了当初构建美国系统三分之一的时间，便交付了全球报税功能。又如近期推出的 Stripe Treasury，面对银行业务解构的浪潮，Stripe 并不追求成为一家实体银行，而是致力于打造全球资金流动的门户底座，支持多国企业持有数十种法定货币。如今各类产品的研发交付周期，相比两年前被一次又一次地大幅压缩。 ```mermaid flowchart LR accTitle: AI-Driven Injection Molding Code Paradigm accDescr: Diagram illustrating the parallel between industrial injection molding and agentic software engineering. subgraph Industrial["传统工业：注塑成型 (1940s)"] direction TB M1["金属模具与精密铸件"] M2["往复螺杆均匀熔融塑料"] M3["高频批量冲压成型"] M1 --> M2 --> M3 end subgraph Software["现代软件：智能体工程 (Agentic Coding)"] direction TB S1["模版与 Markdown 指南 (AGENTS.md)"] S2["AI 智能体基于规范自动产出代码"] S3["智能体深度自动化代码审查 (Code Review)"] S1 --> S2 --> S3 end ``` Will 提出了一个新颖的隐喻：当下的 AI 驱动型软件工程就像二战后的“注塑成型”（Injection Molding）技术革命。当年 James Watson Hendry 发明螺杆式注塑机后，塑料加热熔化与模具注塑实现了极致的标准化与规模化，进而引发了日用消费品与玩具行业的爆炸式增长。今天的软件开发如出一辙：工程师的工作重心转变为打造“模具”与“范式”——例如在代码仓库中定义规范严密的 Markdown 指南（如 `AGENTS.md`）和金融机构集成接口规范，随后将大量重复性编码交由 AI 智能体执行。人类工程师甚至只需在末端参与 Code Review，而智能体本身在代码审查上的细致程度也已超越过往的人工排查。 针对这种生产方式的转变，David 与 Will 达成了共识：聚焦于优化成本结构本质上是在“做空自己的未来潜能”（Going short on your future potential），这就像一家企业选择退还现金而非再投资；而通过 AI 工具把软件生产力推向极致、加速构建新产品，才是真正的“做多未来”（Going long on your future potential）。 --- ### [28:02](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1682s) Agentic Commerce（智能体商业）：缺失的底层原语 随着话题由内部工程效率转向商业生态演进，David George 提到了行业当前高度关注但概念仍显模糊的前沿领域——AI 智能体在互联网上自主参与商业活动（Agentic Commerce）。关于智能体商业的现状、阻碍与演进路径，Will Gaybrick 直言，目前智能体商业尚未迎来属于它的“Claude 3.5 Sonnet”或“GPT-4”式的寒武纪大爆发时刻。现实中尚未出现被高频验证、广泛复制的杀手级原生使用场景。 造成这一停滞的核心症结在于商业底层“积木块”的缺失，即缺乏支撑机器之间互相交易的原生原语（Missing Primitives）。当前的互联网商业架构是完全围绕人类交互（如信用卡表单、重定向结账页面、图形用户界面）而设计的，无法直接兼容自主运行的 AI Agent。 为此，Stripe 正与开源社区及 Tempo 等合作伙伴联合推进基础协议的制定，例如确立全新的“机器支付协议”（Machine Payments Protocol，简称 MPP）。该协议旨在唤醒 HTTP 协议早年预留的状态码——“402 Payment Required”（需要付款）。当一个 AI 智能体向线上某个服务请求获取某张图片、某段数据或受版权保护的内容时，目标服务器可以直接返回 402 响应，并在机器可读的协议载体中精确标明所需支付的金额、货币形式以及机器结算接口，从而为无人工干预的机器对机器（M2M）商业闭环奠定基石。 ### [29:10](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1750s) 机器互购与结账页面的消亡 在智能体电商（Agentic Commerce）的早期探索中，核心挑战在于构建机器与机器之间交易的“原生原语”。Will Gaybrick 与 David George 探讨了智能体时代的结账究竟应该长什么样：当前许多团队尝试使用浏览器自动化（Browser Automation）让 Agent 像人类一样逐个去爬取、填写真实的表单，但这只是典型的“拟物化（Skeuomorphic）过度方案”，未来必然会出现属于机器的“原生版本”。 一个不可逆转的趋势是——面向人类的传统结账页面（Checkout Pages）本身将会彻底消失。目前 Stripe Link 已经积累了约 4 亿用户，Shopify 的 Shop Pay 也有庞大的用户基数，用户在结账时已经能极速完成。但 Will 提出，未来甚至连商品详情页（PDP）跳转到结账页这一步都不需要存在，用户只需说一句“买了”，系统便能静默完成交易。这是消除结账页面的确定性演进，而智能体电商将把这种摩擦力降为零。 ```mermaid flowchart LR accTitle: 结账交互的演进历程 accDescr: 展示从人类手动填表到拟物化自动化，再到原生智能体直付的演变。 subgraph Phase1["拟物化与过渡阶段"] A["传统人类结账<br/>跳转页面/手动填表"] --> B["浏览器自动化<br/>Agent 模拟人爬取填表"] end subgraph Phase2["原生智能体时代"] B --> C["Link Agent Wallet / CLI<br/>提取凭证由机器直付"] C --> D["零结账页面<br/>机器原生指令直接扣款"] end ``` - **Stripe Link 规模**：拥有约 4 亿注册用户，成为跨商户一键结算的底座。 - **Link Agent Wallet**：Stripe 推出了针对智能体的钱包与 Link CLI，允许 Agent 在人类设定的安全权限与授权规则内，自动读取并使用 Link 支付凭证完成扣款。 --- ### [30:54](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1854s) B2B 智能体采购与一次性微消费需求 Will 强调，他们目前对智能体电商最兴奋的方向实际上是 B2B 领域。以 Stripe 推出的 Stripe Projects 为例，表面上它是一个搭建和脚手架应用的工具，但其底层最核心的突破在于“让 Agent 具备自主配置与采买 B2B 服务的原语”。例如，代码 Agent 可以在完全不打开 Vercel 官网界面的情况下，为项目自动购买并开通 Vercel 托管；同理，它也能自主采购 Browserbase 等开发者基础设施。David George 补充指出，这也深刻影响了 a16z 近期的开发者工具投资逻辑——评估一款工具的标准正变成“未来 Agent 采买工具时，是否会首选这个”。 随着复杂长任务的普及，传统按月扣费的订阅模式（Subscription）开始成为阻碍 Agent 执行的巨大包袱。Will 分享了他用 Claude Code 为侄女写生日歌的经历：如果为了制作一首歌，必须分别在 Suno、Higgsfield 等服务上按月花 9.99 美元或 10.99 美元订阅账号，无论是人还是 Agent 都会被劝退。David 也提到自己 10 岁的儿子正在用 AI 工具创作说唱歌曲并上传到 Spotify。他们一致认为，互联网急需一套“瞬态、一次性微消费 API（Ephemeral/Micro-consumption APIs）”，让用户只需给 Agent 设定预算（例如“给你 15 美元额度，去发现并调用所需的服务来完成任务”），Agent 就能自由调度整个网络的计算与存储资源。 - **实测案例**：Stripe 团队曾现场演示由 Agent 驱动 Browserbase 会话，自主打开 espn.com 查阅数据并填报 NCAA 对阵预测竞猜榜（Bracket），效果堪比甚至超越人类。 - **模式变革**：由强制绑卡的月度订阅制（$9.99/月），走向轻量级、无需建档的单次微额消费接口。 --- ### [34:22](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2062s) 微支付的复兴与稳定币的底层契合 微支付（Micropayments）自互联网诞生之初就被反复提及，却始终未能真正成立。Will 剖析了其历史困境：在过去的内容消费中，单篇文章或零散内容的微支付总是被两头夹击——要么被高 ARPU 的会员订阅制通吃，要么被零门槛的广告变现碾压，中间的微支付根本找不到健康的商业模式。然而在 Agent 经济下，逻辑发生了根本逆转。 未成年与成年智能体就像“在互联网间穿梭的蜂鸟”，它们需要在一个站点抓取一点数据、在另一处做一点瞬态存储、再在第三方调用一次算力模型。人类不可能为蜂鸟途经的每一个站点预先注册账号并绑定信用卡，微交易因此成为了 Agent 经济能够运转的刚性前提。而解决微支付落地成本与跨境结算障碍的终极武器，正是稳定币（Stablecoins）。尽管目前稳定币对普通人类而言仍存在钱包与合规等链上门槛，但对于只看执行逻辑的 Agent 而言，在后台用美元直接结算稳定币、使用公用余额完成极速清算，没有任何心理或操作负担。 ```mermaid flowchart TD accTitle: Agent 蜂鸟经济与微支付调用模型 accDescr: 展示用户设定预算后，Agent 穿梭于各微消费服务之间并以稳定币完成即时清算。 User["用户<br/>设定总预算（如 $15）"] --> Agent["AI 智能体 (蜂鸟模式)"] subgraph MicroServices["瞬态微消费生态"] Agent --> S1["数据抓取 (单次调用 $0.05)"] Agent --> S2["临时存储 (瞬态保留 $0.02)"] Agent --> S3["模型推理/生成 (微计算 $0.20)"] end subgraph Settlement["底层结算"] S1 & S2 & S3 --> Stables["稳定币极速微结算通道<br/>(零摩擦/免建账号)"] end ``` --- ### [37:28](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2248s) 全球资金流转平台：稳定币与国家级支付网对比 针对稳定币在未来五年的市场定位，Will 从纯粹的基础设施视角给出断言：稳定币是人类目前所拥有的“更优越的资金流转平台（A better platform for moving money）”。相比传统银行卡组织和复杂的跨境清算电汇网络，稳定币在结算效率与交易成本上具备代际优势。 作为对比，部分由主权国家统一推动的基础设施同样展现了极低成本微支付的力量： - **印度 UPI**：国家主导的即时支付系统极其繁荣，其交易中单笔金额低于 5 美元的交易占比高达约 86%。 - **美国信用卡网络**：与印度形成鲜明对比，美国信用卡交易中单笔低于 5 美元的笔数占比仅为个位数百分比（Single digits%），微额交易长期被高额的固定刷卡手续费抑制。 - **巴西 Pix**：同样作为国家级低成本、快速清算的支付网络，在巴西本土呈现出爆发式增长。 Will 认为，主权国家虽然能在单一境内市场打造类似 UPI 或 Pix 这样快速廉价的系统，但在跨越主权边界的全球互联场景下，稳定币正是填补这一空白、支撑全球智能体商业与微支付运转的底层基石。 ### [38:32](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2312s) 稳定币与全球经济的“谢林点”：突破法定货币的地域壁垒 全球经济需要一个能够让所有人达成共识的“谢林点”（Schelling point），而加密货币支付轨道从根本上解决了传统跨国资金流动中的政治壁垒与摩擦。David George 指出，如果整个世界都能统一采用稳定币，全球金融体系将运行得更快、成本更低。Will Gaybrick 对此完全认同，但他也直言这往往属于“理应如此、却未必能自然实现”的难题，因此 Stripe 正全力推动这一变革。Stripe 已将稳定币作为底层原生资产接入 Stripe Treasury，使得企业不仅能持有美元、欧元或英镑余额，也能直接持有稳定币余额。 对于 Stripe 而言，稳定币的核心机遇在于“更高性能、更低成本、更具全球化”。在传统的法定货币网络中，Stripe 支持的国家和地区大约在 60 个左右；而通过稳定币，Stripe 的覆盖范围迅速扩大到约 150 个国家和地区。这种能力让全球各地的创业者得以直接参与软件经济与 AI 浪潮——无论是在泰国只有两名工程师的微型初创团队，还是在巴西的开发者，都可以与全球最前沿的 AI 公司无缝结算。David 赞许了 Stripe 的策略，认为企业不会一刀切地放弃法币转向加密货币，而 Stripe 兼顾法币与稳定币的混合架构，恰好契合了企业逐步过渡的市场节奏。 Will 以专注于美墨汇款走廊的金融科技公司 Felix（前身为 Felix Pago）为例，展现了稳定币底层设施的爆发力。美墨汇款走廊是全球规模最大的汇款通道，过去由西联汇款（Western Union）或 Wise 等传统及现代金融机构耗费多年深耕，而 Felix 基于稳定币构建网络，仅仅几年时间就在该走廊占据了 5% 到 10% 的汇款份额，这种颠覆性的效率在传统法币世界几乎是不可想象的。 - **法币与稳定币覆盖对比**：Stripe 法定货币覆盖约 60 个国家/地区，而稳定币支持拓展至约 150 个国家/地区。 - **汇款案例落地**：Felix 凭借稳定币轨道，在短短数年内拿下了全球最大汇款走廊（美墨走廊）5%~10% 的市场份额。 - **无国界的 AI 创业**：两名身处泰国的工程师能够与美国或巴西的团队一样，平权接入全球资金网络。 ```mermaid flowchart TD accTitle: Stripe 稳定币与全球资金流动架构 accDescr: 展示 Stripe 如何通过原生整合稳定币，将全球覆盖范围从 60 国法币扩展到 150 国 subgraph FiatRail["传统法币网络 (Fiat Rail)"] F1["支持约 60 个国家/地区"] F2["跨境银行结算 / 政治与中继摩擦"] F3["传统汇款公司 (耗时长/费用高)"] end subgraph StripeCore["Stripe Treasury 统一底座"] T1["原生多币种账户：USD / EUR / GBP / 稳定币"] T2["渐进式迁移：无需全面抛弃法币"] end subgraph StablecoinRail["稳定币支付网络 (Stablecoin Rail)"] S1["全球谢林点：覆盖约 150 个国家/地区"] S2["即时清算、低费率、无国界"] S3["新兴案例：Felix 在美墨汇款走廊占 5%-10%"] end FiatRail --> StripeCore StablecoinRail --> StripeCore StripeCore --> AI["全球创业者 (如泰国 AI 初创团队)"] ``` --- ### [41:35](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2495s) 专为支付而生的区块链：揭秘 Tempo 项目 谈及备受瞩目的区块链项目 Tempo 时，Will 调侃自己要再次扮演“底层技术书呆子”（infrastructure nerd）。他坦言，市场上已经存在许多优秀的通用公链，如果脱离支付场景，世界根本不需要另一条全新区块链。然而，现有公链大多偏向加密资产交易与金融投机，当市场发生剧烈交易波动时，网络性能往往严重恶化，Gas 费用飙升。对于实际的商业支付而言，交易费用的可预测性和网络的确定性是刚性要求。 正因如此，打造一条专门面向支付优化的区块链具有极其清晰的商业逻辑。Tempo 在设计上将隐私性作为一等公民（First-class primitive）来构建，以解决公开链账本容易被逆向推导商业机密的缺陷；同时保障极高的吞吐量永远不被牺牲，并根除浮动交易费导致的成本暴涨，确保资金移动始终保持高效、稳定与低廉。目前 Tempo 仍处于早期构建阶段，但已经获得了强劲的发展势头，不仅吸引了 DoorDash 等知名企业的合作，也在逐步确立其作为 Stripe 默认但非唯一底层区块链的地位。 - **隐私保护作为首要原语**：防止企业链上交易数据与商业机密被外界轻易逆向分析。 - **杜绝 Gas 费暴涨**：剥离投机交易对底层性能的干扰，确保任何网络拥堵环境下支付手续费平稳可预测。 - **商业伙伴支持**：与 DoorDash 等头部企业展开合作，打造大规模真实商业场景下的结算底层。 --- ### [43:10](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2590s) Token 即新货币：AI 时代的算力资产管理与软件大爆发 David 将话题引向当前庞大的“Token 经济”（Token Economy），指出 Stripe 年交易量已突破 2 万亿美元，而 Token 经济或将成为科技史上规模最大的产业之一。Will 笑着表示，自己一直非常喜欢“Token”这个词，因为“它听起来就像是货币的一种近似物”，而在实际应用中，Token 正在真正演变为通用货币。随着通用编程平台（如 Cursor、Replit）的兴起，用户可以在上面消耗 Token 完成几乎所有计算与开发任务。随之而来的是极其精密的攻击手段——黑客通过免费试用漏洞、批量注册多账户来恶意套取 Token，其套路与 Stripe 过去防范盗刷信用卡和资金欺诈如出一辙。Token 与美元之间的界限正在彻底模糊。 面对这种趋势，Will 强调 Stripe 将自身角色定义为“不仅帮客户安全合规地移动和存储美元，更有责任在 Token 领域提供同样的保护”。在未来的长周期智能体任务（Agentic tasks）中，企业过去可能需要雇佣多人完成结账或审计，并以美元支付薪资；而今后这些任务将由 AI Agent 辅助或完全自动化执行，企业转而以 Token 为计算单元支付服务成本。Stripe 的使命是让“Token 与美元之间的流转，就像美元与欧元之间的兑换一样无缝且安全”。谈及在这一生态中的定位，Will 澄清 Stripe 内部从不傲慢地谈论所谓的商业“权利”（rights），而是将其视为必须反复赢得的“使命”（mandates）。 在企业内部管理上，Token 消费正在迅速分化为两大维度：其一是作为运营支出（Opex）的开发端消耗，即研发构建产品的成本；其二是作为产品功效（Product efficacy）的核心要素，即企业需要动态衡量自身产品的 Token 消耗产出比，并实时在不同模型之间调度编排以实现最高效率。关于“AI 是否会导致软件严重同质化与贬值（Commoditized）”的争论，Will 指出虽然理论上有此可能，但现实数据展现的却是截然相反的景象：全球软件创建正迎来指数级爆发，企业商业化变现的速度远超以往。Stripe 的 2026 届客户群规模比 2025 届扩大了 50% 且增速更快，而 2025 届又比 2024 届扩大了 70%，更多软件正在被以前所未有的速度创建并投入市场。 - **Token 与法币边界消融**：Token 滥用（多账号白嫖、免费试用攻击）呈现出与传统金融盗窃高度一致的黑产特征。 - **Token 管理的两大支柱**： 1. **运营支出端（Opex）**：开发工具消耗管理与预算控制。 2. **产品功效端（Product Efficacy）**：作为模型编排层，评估并调度各模型间的性能与成本表现。 - **客户群爆发式增长**：Stripe 2026 届签约客户规模较 2025 届增长 50%，2025 届较 2024 届增长 70%，有力反驳了“软件价值走向消亡”的虚无论断。 ### [48:06](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2886s) 超越代码的软件价值：贯穿全栈的收入基础设施 在探讨企业对金融工具的采纳时，Will Gaybrick 分享了一个具体案例：AI 语音生成独角兽 ElevenLabs 目前在生产环境中深度使用了多达 14 款 Stripe 产品。主持人 David George 随即开玩笑打趣道，考虑到 ElevenLabs 创始人兼首席执行官 Mati Staniszewski 凡事都喜欢以数字 “11” 为倍数，原本觉得用 11 款最契合品牌，但既然已经达到了 14 款，“那你们干脆直接瞄准 22 款去推吧”。这种深度绑定印证了当今快速崛起的 AI 公司对成熟商业基建的极度渴望。 Will 借此深入阐述了企业软件的本质：软件开发远不止于“编写和交付代码”本身。构建一套系统的真正难点，在于如何让底层的业务抽象与业务逻辑深度穿透并贯穿整家公司的日常运转。如果一家企业完全选择在内部从零自研这些复杂的金融与计费模块，一旦经历人员流动，原本掌握核心专长的人员离职，内部系统就会迅速缺乏维护并开始退化。Stripe 的核心价值正在于承揽整个“收入技术栈”（Revenue Stack）的繁杂底层，帮助企业高效管理现金、梳理收入、加快关账进程，并在全球化扩张中免去后顾之忧。在当下这个软件爆发的周期，这一市场机遇比以往任何时候都要巨大。 - **多产品深度渗透**：ElevenLabs 单一客户即接入了 14 款不同的 Stripe 产品，体现出 AI 新兴企业对一体化金融基础设施的高度依赖。 - **自研系统的退化陷阱**：内部自建系统在遭遇团队更迭（Turnover）时往往面临知识流失与架构退化，而标准化外部基建能提供长期的专业抽象与持续迭代。 --- ### [49:00](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2940s) “品味”的反思：AI 时代人类价值的借口与真实追求 播客接近尾声时，话题转向了科技圈近期极其火热的一个概念——“品味”（Taste）。David 指出，尽管 Stripe 从本质上讲是一家底层金融基础设施公司，但在整个硅谷和开发者群体中，Stripe 出色的产品品味与设计审美是公认的标签。在越来越多的代码与工作流被 AI 接管、公司规模不断扩大的当下，Stripe 究竟如何守住并规模化这种品味？ 听到这个问题，Will 忍不住笑出声来。他坦言最近“Taste”这个词在行业里被讨论得过于频繁，甚至带着一种微妙的防御心理。从愤世嫉俗的视角来看，“品味”更像是人类在面对日渐强大的 AI 时用来为自身未来存在价值进行辩护的借口——“当程序员突然发现自己的手指写代码不再像以前那样不可替代时，大家便开始紧抓‘品味’这种虚无缥缈的概念不放，坚信‘模型永远不可能拥有品味’”。但 Will 直言，这种关于 AI 永远没有品味的断言大概率并不成立，David 也对这一观点表示认同。 --- ### [50:01](https://www.youtube.com/watch?v=P5iICDVn5gc&t=3001s) 规模化品味的核心法宝：反复宣讲与高拟真业务仿真系统 抛开对概念的哲学辩论，谈及 Stripe 究竟如何落地产品品味，Will 强调“产品质量”（Product Quality）是深植于 Stripe 文化与团队身份认同的核心。这背后源于两大朴素的动力：其一是极度聚焦的用户视角，创业本身就是一场充满艰辛的跋涉，如果开发者能用上被打磨得惊艳、顺手且极其趁手的工具，整个旅途的挫败感就会大大降低；其二是纯粹的内部成就感，对于研发团队而言，做出一款连自己看了都忍不住赞叹“哇，这做出来太漂亮了”的产品，其带来的乐趣远非那些东拼西凑、勉强能跑的蹩脚软件所能比拟。 关于如何在千人规模的大组织中将这种品味规模化，Will 总结了两个硬核方法论： 1. **自上而下不厌其烦地宣讲**：将品质提升上升为全员信条，管理层必须像初创公司反复抓战略那样，反反复复、随时随地强调高质量的要求。 2. **构建高拟真环境并强制亲身体验产品**：Will 借用了两个经典工业领域的类比——航空业在飞机试飞前依赖极其严苛的飞行模拟；英伟达（Nvidia）在发展史上真正实现“逃逸速度”（Escape Velocity）的关键节点，就是他们不再需要被动等待晶圆厂（Fab）把物理芯片造出来才能测试，而是通过高精度的芯片性能仿真系统大幅缩短研发周期。 受此启发，Stripe 在“产品使用模拟”（Simulation of Product Usage）上投入了巨大的工程资源。研发团队不需要拿真实的生产数据冒险，而是可以通过工具指定一个典型账户，系统即刻生成一个完全脱敏（剔除所有 PII 个人身份信息、随机化增长率与规模）但行为特征极度贴近真实世界的虚拟商户沙箱。在这个模拟环境里，工程师会身临其境地遭遇该客户真实面对的业务波动、季节性起伏，甚至会实时收到虚拟生成的客户纠纷（Disputes）和退款请求（Refunds）。Stripe 赋予工程经理（EM）极大的主导权，当 EM 在亲自使用中感受到“这里的体验太糟糕了”，他们便有权调配资源，将接下来的整个 Sprint 周期全部投入到“品质跃升”（Elevating Quality）的专项打磨中。 ```mermaid flowchart TD accTitle: Stripe 规模化品味与品质的闭环流程 accDescr: 展示 Stripe 如何通过文化宣讲、业务仿真、痛点体验及资源倾斜来维护高水准产品品味的流程图。 subgraph Culture ["1. 文化与意识驱动"] A["自上而下宣讲<br/>强调品质信条"] --> B["用户同理心与<br/>自豪感激发"] end subgraph Simulation ["2. 工业级业务仿真系统"] C["参考航空/英伟达仿真逻辑<br/>摆脱实物/生产环境依赖"] --> D["生成脱敏沙箱账户<br/>模拟真实增长/季节性波动"] D --> E["动态注入真实日常痛点<br/>如虚拟拒付纠纷、退款潮"] end subgraph Execution ["3. 体验驱动与资源闭环"] F["工程团队/EM<br/>身临其境试用产品"] --> G{"发现粗糙体验与缺陷?"} G -- "是" --> H["EM 调动资源决策<br/>将下个 Sprint 锁定为品质优化"] H --> I["持续打磨与品味规模化"] end B --> F E --> F ``` --- ### [53:02](https://www.youtube.com/watch?v=P5iICDVn5gc&t=3182s) 结语：在 Stripe 内部孕育“下一个 Stripe” 对话进入尾声，David 总结了这次交流的核心亮点，盛赞 Stripe 在 AI 时代以极具开拓性的姿态重塑软件构建模式，并走在智能体商业（Agentic Commerce）与稳定币金融前沿。在对未来的探索中，Stripe 内部始终恪守着一个充满创业野心的愿景指标：“我们如何确保在 Stripe 内部孕育出‘下一个 Stripe’？”（"How do we assure that we build the next Stripe inside of Stripe?"）。Will 对此深表认同，并在对未来互联网商业无限可能性的期待中结束了本次对话。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=P5iICDVn5gc&t=0s) 在 AI 时代，企业若想以更快的速度交付更多产品，就必须在组织内部建立起“像创始人一样的自主权（founder-like agency）”。 — Will Gaybrick > [05:04](https://www.youtube.com/watch?v=P5iICDVn5gc&t=304s) Stripe 的策略毫不掩饰，那就是——赢得所有的初创公司，然后再次赢得它们（win all the startups and then win them again）。 — Colin / Will Gaybrick > [12:23](https://www.youtube.com/watch?v=P5iICDVn5gc&t=743s) 开个半开玩笑的说法，我们的策略是去“建造一切”（build everything）。优化成本结构的终极手段不是收缩，而是以更高的速度扩张并实现业务增长。 — Will Gaybrick > [19:41](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1181s) 短期内“用现有资源做更多的事”（Doing more with what we have），长期看则是“用更多资源做成倍的事”（Doing more with more）。 — Will Gaybrick > [22:42](https://www.youtube.com/watch?v=P5iICDVn5gc&t=1362s) 聚焦于优化成本结构本质上是在“做空自己的未来潜能”（Going short on your future potential）……而通过 AI 工具把软件生产力推向极致、加速构建新产品，才是真正的“做多未来”（Going long on your future potential）。 — Will Gaybrick > [37:28](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2248s) 稳定币是人类目前所拥有的“更优越的资金流转平台（A better platform for moving money）”。 — Will Gaybrick > [43:10](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2590s) Token 听起来就像是货币的一种近似物……让 Token 与美元之间的流转，就像美元与欧元之间的兑换一样无缝且安全。 — Will Gaybrick > [49:00](https://www.youtube.com/watch?v=P5iICDVn5gc&t=2940s) 当程序员突然发现自己的手指写代码不再像以前那样不可替代时，大家便开始紧抓“品味”这种虚无缥缈的概念不放，坚信“模型永远不可能拥有品味”。 — Will Gaybrick > [53:02](https://www.youtube.com/watch?v=P5iICDVn5gc&t=3182s) 我们如何确保在 Stripe 内部孕育出“下一个 Stripe”？（How do we assure that we build the next Stripe inside of Stripe?） — David George ## Mentioned Will Gaybrick, David George, Stripe, a16z, Stripe Minions, Stripe Projects, Stripe Payments, Stripe Billing, Stripe Connect, Stripe Radar, Stripe Tax, Stripe Managed Payments, Stripe Treasury, Stripe Link, Kai, Cursor, ElevenLabs, Ahrefs, Colin (Clerk), DoorDash, Instacart, Amazon, Microsoft, Google, Alan Mulally, Scott Woody, Metronome, Zach Abrams, Bridge, Privy, JR Farr, Lemon Squeezy, Suno, Higgsfield, Tempo, MPP (Machine Payments Protocol), HTTP 402, Shopify (Shop Pay), Vercel, Browserbase, Claude Code, UPI, Pix, Felix, Mati Staniszewski, Nvidia ## Video description a16z General Partner David George is joined by Will Gaybrick, President of Product & Business at Stripe, to discuss how AI is changing the way Stripe builds products, organizes teams, and thinks about the future of internet commerce. Stripe has evolved from a payments company into a multi-product financial infrastructure platform, while a new generation of AI companies is growing and monetizing faster than previous software cohorts. Will explains why Stripe sees AI productivity as an opportunity to build more rather than simply cut costs, including how its internal coding agents now generate thousands of pull requests each week. They discuss creating founder-like agency inside large companies, building smaller and flatter teams, and why Stripe believes dramatically more software will be created as the cost of building continues to fall. They also look ahead to agentic commerce, why checkout pages could disappear, the potential return of micropayments, stablecoins as infrastructure for a global economy, and a future where AI agents increasingly buy software and services from other machines. Timestamps: 00:00 - Intro 01:00 - What Is Stripe Today 08:00 - The AI Cohort Explosion: 50% More Signups & Software Creation Booming 14:25 - Stripe Minions: From 1,200 to 7,000 PRs a Week 19:24 - Build Everything: Going Long vs Going Short on Your Future 28:00 - Agentic Commerce: The Missing Primitives 37:28 - Stablecoins & the Global Money Movement Platform 49:00 - Scaling Taste at Stripe Resources: Follow Will Gaybrick on LinkedIn: https://www.linkedin.com/in/william-gaybrick-5730347/ Follow Will on X: https://x.com/gaybrick Follow David George on X: https://x.com/DavidGeorge83 Follow Stripe on X: https://x.com/stripe Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=P5iICDVn5gc)

[0:00] Will Gaybrick: 在 AI 时代，企业若想以更快的速度交付更多产品，就必须在组织内部建立起“像创始人一样的自主权（founder-like agency）”。 [5:04] Colin / Will Gaybrick: Stripe 的策略毫不掩饰，那就是——赢得所有的初创公司，然后再次赢得它们（win all the startups and then win them again）。 [12:23] Will Gaybrick: 开个半开玩笑的说法，我们的策略是去“建造一切”（build everything）。优化成本结构的终极手段不是收缩，而是以更高的速度扩张并实现业务增长。 [19:41] Will Gaybrick: 短期内“用现有资源做更多的事”（Doing more with what we have），长期看则是“用更多资源做成倍的事”（Doing more with more）。 [22:42] Will Gaybrick: 聚焦于优化成本结构本质上是在“做空自己的未来潜能”（Going short on your future potential）……而通过 AI 工具把软件生产力推向极致、加速构建新产品，才是真正的“做多未来”（Going long on your future potential）。 [37:28] Will Gaybrick: 稳定币是人类目前所拥有的“更优越的资金流转平台（A better platform for moving money）”。 [43:10] Will Gaybrick: Token 听起来就像是货币的一种近似物……让 Token 与美元之间的流转，就像美元与欧元之间的兑换一样无缝且安全。 [49:00] Will Gaybrick: 当程序员突然发现自己的手指写代码不再像以前那样不可替代时，大家便开始紧抓“品味”这种虚无缥缈的概念不放，坚信“模型永远不可能拥有品味”。 [53:02] David George: 我们如何确保在 Stripe 内部孕育出“下一个 Stripe”？（How do we assure that we build the next Stripe inside of Stripe?）

[5:05] Will Gaybrick: Stripe's strategy is unabashedly to win all the startups and then win them again. [18:45] Will Gaybrick: Going long on your future looks like building everything, and going short on your future looks like cutting headcount. [37:20] Will Gaybrick: If we all like went to sleep tonight and woke up tomorrow and we all held stables, it would just be a better global economy with less friction and less cost. [43:10] David George: Tokens are the new dollars. [46:10] Will Gaybrick: We talk about mandates. We feel we have a mandate to help users with this, and so that's why we're really leaning into it. [51:06] Will Gaybrick: In aviation, simulation is so important... You can't just say, 'I hope the jet works.' You have to kind of prove to yourself that it does before you ever fly it. [51:46] Will Gaybrick: How do we ensure that we build the next Stripe inside of Stripe?

[0:00] If you want to ship more and build [0:02] faster, you have to create founder-like [0:04] agency inside your company. A single [0:06] engineer can do what two teams of [0:08] engineers could do two years ago. [0:10] Suddenly the market opportunity [0:12] landscape is just much broader and you [0:13] can do more with software. How do you [0:15] guys ship so much product? Companies [0:17] have skewed in that direction where [0:19] they've seen this new agentic efficiency [0:23] and power as a way to optimize cost [0:25] structure. our belief I'm being a little [0:27] cheeky but build everything the Opus 45 [0:30] Cambrian explosion moment we haven't had [0:32] that in agentic commerce today and I [0:34] think there's a few reasons for that one [0:36] is we are merging so much more code than [0:38] last year and it is stressing every [0:40] system so we created something called [0:42] strike minions you're not going to [0:43] iterate you're not go into planning mode [0:44] you're just going to say this is what I [0:45] want go do it that's kind of where the [0:47] world is going you guys are I think over [0:50] two trillion of volume now what do you [0:52] think is sort of the future state of how [0:54] we all access models and tokens [0:56] >> and very bullish on [1:00] >> I want to start with just a state of [1:03] stripe. [1:04] >> Yes. [1:04] >> So what is stripe today and you know [1:06] when we originally invested it you know [1:07] it's a payment processor but now I think [1:10] the average AI company uses 11 different [1:13] stripe products and at sessions [1:16] >> I believe you had 288 distinct [1:19] >> product launches. So I want to talk [1:21] about velocity also but just start with [1:22] just what is the state of stripe today? [1:24] So internally we think about Stripe as [1:27] having you know inverted our value [1:29] proposition from being a payments [1:31] company with sort of add-ons to now [1:33] being this multi-product platform where [1:36] everything sort of focuses on financial [1:38] infrastructure helping you grow by [1:41] reducing the friction and increasing the [1:43] agency you know to um be more agile with [1:47] your business model to uh operate in [1:50] more countries uh and just go faster [1:54] when it comes to everything that touches [1:55] revenue and cash. So you know just [1:59] practically speaking went from payments [2:01] to then you know billing subscriptions [2:04] and invoicing uh connect if you're a [2:07] platform or marketplace radar for [2:09] mitigating fraud you know radar now for [2:11] doing many more things than that tax and [2:14] just I think we we don't actually count [2:16] but somewhere in and around 2530 [2:19] products that are sort of headlining [2:21] branded products and then of course you [2:22] know hundreds and thousands of features [2:24] below that but again the framework work [2:27] we think about is really reducing [2:28] frictions and increasing agency. So a [2:31] good example last year uh we saw a lot [2:34] of users for the first time experiencing [2:35] free trial abuse and this wasn't really [2:38] an issue pre- AI because [2:41] >> you know most stripe users are software [2:43] companies margins [2:46] burden wasn't there yet. [2:47] >> Yeah. So they're wasting a little [2:48] compute but but it's it's it's it's [2:50] negligible. It's minimous but now [2:53] software has a cost structure and so [2:54] actually I think cursor was the first [2:57] user that we [2:57] >> I was going to say it was our portfolio [3:00] where we exactly experienced it and uh [3:03] you know I I guess internet users can be [3:04] crafty but yes [3:06] >> yes and and I think it was something [3:08] like one in six users of free trials [3:12] were abusive [3:14] um [3:15] >> and so you know you're just throwing [3:17] money at these you know users who are [3:19] just signing up for another account, [3:20] another account, another account, you [3:22] know, maybe even doing model [3:23] distillation and and and things like [3:24] that. And so we sort of, you know, got [3:26] in the bunker with them and just stood [3:29] up in a weekend a pipeline where we were [3:31] able to, you know, use our foundation [3:33] model, look across the entire Stripe [3:35] network, use our embeddings, and then, [3:37] you know, after that put a reasoning [3:38] layer on top of so you could sort of say [3:40] we think this is a free trial abuser [3:42] because [3:43] um and point to those signals. And I I I [3:47] I think it's today, I think it's 11 labs [3:51] recently told us that they're blocking [3:53] 2,000 free trial abusers per day using [3:56] stripe signals. [3:56] >> Incredible. [3:57] >> So, you just think about how much money [3:58] you'd be, you know, burning if you were [4:01] um you know, well, how much money you [4:03] wouldn't how how much revenue you [4:04] wouldn't be creating if you weren't [4:06] giving these trials, you know, how false [4:08] positives and the false Yeah, exactly. [4:09] Exactly. [4:10] >> And then on the increasing agency side, [4:12] >> it's still just so hard to go global. [4:15] you know, you think about um having to [4:17] register, calculate, remmit taxes in so [4:20] many geographies [4:21] >> and this new crop of AI companies, you [4:23] know, they're digital goods companies [4:25] and they want to just, you know, go [4:26] global very fast. And so I guess, you [4:30] know, AEFs is a good example. um they're [4:33] using uh this product strike managed [4:34] payments where in their home markets you [4:37] know the US let's say they have an [4:39] entity in Europe they are the seller of [4:41] record but you know in longtail markets [4:44] you know the Kazakhstans and you know [4:46] all over um APAC uh where they you know [4:50] haven't sort of entities you know stripe [4:52] stands in as the actual merchant of [4:54] record we handle all tax calculation tax [4:57] remittance and everything like that and [4:58] so this just lets them cover you know [5:00] 100 plus geographies [5:02] and compliantly. [5:03] >> Yeah, that's amazing. Yeah. And if you [5:05] go back to the to the fraud example, you [5:07] know, if you identify abusive actors, [5:10] you know, there's actually a network [5:10] effect in that in that business, right? [5:12] Which is um which is super powerful. Um [5:15] and obviously we appreciate it as uh as [5:18] as you know, investors and many of the [5:20] companies who benefited from it. Um so [5:22] if you look at the product set today [5:25] >> you know obviously the origin of stripe [5:28] was startups and so you can very easily [5:32] stand up payments you know sell stuff [5:34] online [5:35] >> and then famously part of the the beauty [5:38] of Stripe's growth and business model is [5:40] that you captured companies like Door [5:42] Dash and Instacart when they were in YC [5:45] >> and then you grew with them along the [5:46] way. How do you think about the product [5:49] strategy as it relates to serving [5:52] startups versus serving large [5:53] enterprises? [5:55] >> So, I think it was actually um Colin at [5:59] Clerk who who summarized our strategy [6:01] nicely on on X recently where he said [6:04] stripe strategy is unabashed. I'm [6:06] paraphrasing his words, but it's win all [6:08] the startups and then win them again. [6:09] >> Yes. Um and I think you know to some [6:11] extent that's just a good business model [6:14] because startups are um very ambitious. [6:18] They typically grow into the biggest [6:19] companies of tomorrow. They're sort of [6:21] canaries for what the next opportunity [6:23] is. So on another subtle reason to win [6:27] all the startups and win them again is [6:28] they actually have the highest standards [6:30] of all of our customers. uh you know I I [6:33] looked recently at uh the seesap for [6:36] stripes reporting and for our very very [6:38] large enterprise users like your [6:40] reporting is great like we love the data [6:42] you go it's way better than anything we [6:43] get from any of our other [6:44] >> because they're accustomed to the [6:44] incumbents. Yeah, exactly. And then you [6:46] you look at it for startups like your [6:48] reporting is garbage. You have got to [6:49] fix this. This is driving me crazy. And [6:52] so there's um this persistent sense that [6:55] startups just make us better by being [6:57] the fastest, by being most demanding and [6:59] so on. But of course once we start [7:01] working with startups, we want to work [7:02] with them forever. And so this pulls us [7:04] up market and you know forces us to [7:07] become you know as sort of surprisingly [7:10] great. That's the standard we try to [7:11] hold ourselves to for enterprises as for [7:13] startups. And now we're working with I [7:16] don't know what the exact percentage is [7:17] today but it's not quite half but we're [7:20] getting there. Uh Fortune 500 companies [7:22] you know we work of course with the [7:24] Amazons of the world um Microsofts and [7:27] you know at the enterprise it's sort of [7:29] a different sales cycle. It's a [7:31] different uh you know post sale [7:32] activation motion. It's a lot more [7:35] required you know after you've already [7:37] signed the contract whereas startups [7:39] sort of go live like that. But [7:40] fundamentally it's you know the same [7:42] thing. You just stay close to the user. [7:43] you hear their their needs. Um you show [7:46] that you're sort of provably better when [7:48] it comes to the metrics they care about [7:50] and um and yeah, so it's really both at [7:52] this stage. [7:52] >> Yeah. Win them and then win them again. [7:54] I love that. That's a great that's a [7:55] great dynamic. Um I want to shift gears. [7:58] So, um I think you guys publicly said [8:00] that uh first half signups grew 50% [8:03] >> year-over-year [8:05] >> and that the median 26 cohort is [8:09] generating 50% more revenue than the [8:12] comparable 25 cohort. So, um and then I [8:17] think the 25 cohort was generating 70% [8:20] more revenue than the comparable four [8:22] cohort. [8:23] >> Um so, what has changed? [8:24] >> Yeah. Well, I think two things. one is [8:28] um [8:30] AI is giving rise to so much opportunity [8:34] for new business creation. There's just [8:36] things you couldn't do before that you [8:38] can do now. [8:39] >> Uh you couldn't build a sunno four years [8:42] ago. Um or you could maybe build a much [8:44] worse four years ago. [8:46] >> Uh you couldn't build a Higsfield 4 [8:48] years ago. You know, there's um there's [8:50] just new things that you can do. And so [8:52] suddenly the um market opportunity [8:56] landscape is just much broader and you [8:57] can do more with software. Uh on the [9:00] other side the sort of cost of doing [9:02] more software has decreased a lot [9:04] because you just need many fewer [9:05] engineers to build the things that you [9:07] want to build because of you know [9:09] agentic coding. And so we're just seeing [9:12] this explosion in new software creation. [9:15] You know, when you look at the [9:16] year-on-year increase in usage of Stripe [9:19] billing, it's actually a lot higher than [9:22] usage of Stripe overall because it [9:24] disproportionately skews towards people [9:26] creating software companies. [9:27] >> Yeah. Yeah. Oh, that's fantastic. That's [9:29] awesome to hear. Um I want to shift now [9:32] to [9:33] >> um some of the products that we talked [9:35] about, right? So, you have this [9:38] proliferation of new products to serve [9:40] all the needs of companies online. Um [9:44] the high level question is like how do [9:46] you do that? Like how do you guys ship [9:48] so much product? [9:49] >> Yeah. [9:52] >> Well speed has always been you know near [9:56] the top maybe maybe user relentless [9:59] focus on user is probably the top part. [10:02] Yeah. Exactly. [10:03] >> But speed is is right up there. Um [10:07] and it's interesting you know we're I [10:08] think a lot about institutional [10:10] progress. Um, and you know, we we study [10:14] the great companies that have come [10:16] before us and now are, you know, beside [10:18] us. Um, you know, the Amazons are still [10:20] around, of course, but you look at [10:22] Stripe and you'd say, okay, well, we [10:24] have sort of an adapted adapted version [10:26] of of how we set goals. Um, that's quite [10:30] similar to Google's OKRs. Um, you know, [10:33] our sales team is organized pretty [10:35] similarly to Microsoft's. Um, you know, [10:38] we have a DRRi culture similar to Apples [10:41] and we, you know, have a product quality [10:43] standard and how we do our product [10:45] reviews is similar to theirs. Um, you [10:47] know, a lot of the day-to-day mechanisms [10:49] that I use to sort of run the business, [10:51] I've basically stolen from Alan Mali, [10:54] the former CEO of [10:56] >> of Ford um, and Boeing Aviation. And so [10:59] we study these companies and we try to [11:01] bring it all into stripe and we think [11:04] about just how do we create an enduring [11:05] institution that can sort of outlast any [11:08] of us because it's you know a great [11:10] container for entrepreneurship [11:13] and then you sort of reach today and [11:16] >> there's just no one to copy like what do [11:18] you do when you know a single engineer [11:21] can do what you know two teams of [11:23] engineers could do two years ago and you [11:27] know sort of the theory of mind right [11:29] now is that Stripe needs to become more [11:32] than it ever has been before a platform [11:35] for founders internally and it's always [11:37] been a platform for founders externally. [11:40] >> Yeah, of course the customers. Yeah, [11:41] >> exactly. And founders have always done [11:42] incredibly well at Stripe. You know, we [11:44] um actually it was an adjacent company [11:46] um Metronome we acquired Metronome last [11:48] year and Scott Woody the CEO is thriving [11:52] at Stripe. He's leading [11:53] >> leading all of you know Metronome and [11:56] billing at this point. Uh we acquired um [11:59] Privy uh which is a wallets [12:01] infrastructure company and Bridge uh [12:03] over the last couple of years and you [12:04] know Henry is now leading most of [12:06] crypto. Zach is leading OpenUSD and [12:08] bridge of course um [12:11] uh AA from from privy is leading a lot [12:13] of engineering for crypto. Uh JR Far [12:16] from Lemon Squeezy is leading Stripe [12:18] Treasury at this point. And so just [12:20] founders have always done very well at [12:21] Stripe. [12:23] But now you have this very interesting [12:24] moment where [12:26] senior engineers are just so powerful. [12:29] >> Yes. [12:30] >> And so you say, well, there's a few [12:32] different ways our or could be shaped [12:34] going forward. It could be that we just [12:36] have many fewer engineers and maybe [12:38] that's what what the world would look [12:40] like. And I think a lot of companies [12:42] have skewed in that direction where [12:45] they've seen this new, you know, agentic [12:49] uh efficiency and power as a way to [12:51] optimize cost structure. Uh you saw some [12:54] layoffs. You saw, you know, companies [12:56] shrinking opex. [12:58] And our belief is that it's just an [13:01] opportunity to I'm being a little [13:02] cheeky, but build everything. [13:04] >> Yes. [13:05] >> And you know, the best way to optimize [13:08] your cost structure is to grow more. [13:10] >> Yes. you know, we have reams of user [13:12] asks going back many years that are [13:14] unmet. And so we just want to get [13:15] through them all faster. [13:17] >> And so while we have, you know, more and [13:20] more productivity from those senior [13:22] engineers, if you want to ship more and [13:25] build faster, you have to create [13:27] founder-like agency inside your company. [13:30] >> And so the main things that are holding [13:32] back our progress today are actually [13:34] sort of back office things. You know, we [13:37] are merging so much more code than last [13:39] year. And it is stressing every system. [13:42] It is stressing how do we get things [13:43] into our seller systems? How do we get [13:44] things onto our pricing page? How do we [13:46] actually bring things to market when we [13:48] can't train sellers on them fast enough? [13:50] And so we are really trying at this [13:53] stage to optimize every single phase of [13:57] we think of it as sort of the critical [13:59] path in Stripe from you know the ideas [14:02] behind product development the user asks [14:04] all the way to products being in users's [14:05] hands. And so, you know, to your [14:07] question of how do you go faster, it's [14:09] really two things. It's one just how do [14:10] you create that agency for everyone at [14:12] Stripe? How do you let everyone at [14:14] Stripe be an otur like a founder, you [14:16] know, a a a creative and not have them [14:19] held back by the morass of, you know, [14:22] centralized processes [14:24] overhead. Yeah. Exactly. [14:26] >> Exactly. And on the other side then it's [14:28] how do you you know with your de uh your [14:30] dev prod team just make the tools better [14:32] and better and better. Um, so we created [14:35] something called Stripe Minions, which [14:36] we've blogged about a little bit. And, [14:39] you know, of course, we've got tons of [14:41] developer tooling to help users iterate [14:43] with agents, but minions, uh, we think [14:46] of as, you know, the most like one of [14:49] our most important metrics internally is [14:52] how many PRs and what percentage of our [14:54] PRs are created by minions. And the [14:56] reason for that is that minions are one [14:58] shot. [14:59] >> Yeah. So you you give it a prompt and [15:02] you know it's going to build it and then [15:04] it's going to you know go through CI/CD [15:07] and all of testing and then you're going [15:08] to review it. So you're not going to [15:10] iterate not go into planning mode. [15:12] You're just going to say this is what I [15:13] want go do it. [15:15] >> And so we think you know that's kind of [15:16] where the world is going. You know it's [15:18] the the oneshot or the glorified real [15:21] loop where you're just letting the agent [15:23] be super super powerful. And so we [15:26] blogged about Minions, I think in [15:29] January, February, and they were doing [15:30] 1,200, you know, PRs per week. And last [15:34] week, I think 7,000 PRs came from [15:37] minions. Wow. [15:38] >> Yeah. And and about 30% of our PRs in [15:41] that week came from Minions. So, [15:42] >> you know, that sort of developer tooling [15:45] and then the internal processes just get [15:48] out of the way and make the tools [15:50] amazing. And that's how we're going to [15:51] go fast. [15:52] >> That's amazing. In terms of empowering [15:55] senior engineers who are now more [15:57] powerful than ever, have you made any [15:59] dramatic changes to how you organize the [16:03] company internally? So, obviously that [16:05] the minions example is incredible and I [16:07] assume the 30% is going to go to a very [16:08] high percentage within a year or [16:10] something. Um, but some of the things [16:12] that we've heard from other companies is [16:14] taking your most powerful engineers, [16:16] putting them into business units or [16:18] closer to the customers. Um but then you [16:21] have sort of coordination problems that [16:22] come from that. So like how have you [16:24] guys approached that empowerment? [16:26] >> Yeah. So engineers have always been you [16:29] know some of the most important product [16:31] leaders at Stripe. You know we build for [16:32] engineers for technical users. Of course [16:34] we built for many personas at this point [16:36] but engineers have always been sort of [16:38] the hero you know ICP um uh for Stripe. [16:42] And so because of that engineers have [16:44] always been some combination of [16:46] engineer, PM, designer. [16:48] >> Yeah. Uh, and so I think we're just [16:50] leaning even more into that. You know, [16:51] we're building tools and platforms so [16:54] that engineers can sort of do a lot of [16:56] the the front-end development themselves [16:58] and um and front-end design themselves. [17:01] Um, [17:03] probably, you know, just the the [17:06] shortest answer to your question is [17:08] flatter. [17:08] >> Yeah, flatter. Smaller, I assume smaller [17:10] teams. [17:11] >> Smaller teams, flatter. uh you had you [17:13] know a lot of layers that were sort of [17:15] orchestrating work and very valuably so [17:18] in most cases. Exactly. But you just [17:21] don't need that anymore. I was talking [17:22] to a one of our most senior engineers [17:24] recently. In fact [17:25] >> the guy who um uh who built Stripe [17:28] Projects because Stripe Projects was [17:29] >> you know basically a PM and one of a [17:33] very senior engineer and a few weeks and [17:35] and a couple of other engineers who who [17:37] um jumped in as well. So I want to give [17:39] them due credit but most of the PRs came [17:41] from one guy. He's got Alexander now. [17:43] He's just got a screen. He's [17:44] orchestrating 16 agents and he's going a [17:47] whole lot faster. [17:48] >> Yeah, that's amazing. Do you have a [17:49] standard uh team size for projects like [17:52] this? Is it like we've gone around it's [17:54] like okay one you know one there's no [17:57] social dynamic and so maybe that doesn't [17:59] work and you know four is probably the [18:00] max size and we've heard everything in [18:02] between. You have a view on it? [18:04] >> Not really. I mean we still have I mean [18:08] you kind of want um in a flatter or you [18:12] both want wider teams right because that [18:14] means fewer layers of management [18:17] >> but then you know you also sort of [18:20] >> believe you can have fewer people [18:23] because each individual is is more [18:25] empowered these days and so I think [18:27] those forces are offsetting to some [18:29] extent today and in fact it's it's more [18:31] like [18:32] >> you know if you have a a linemanaged [18:34] team of eight people, they're just doing [18:36] 3x more. [18:37] >> Yeah, exactly. [18:37] >> So, you might have the same manager to [18:39] the manager to IC ratio, but the team is [18:42] doing three things rather than one. [18:43] >> Yeah. Yeah, it makes total sense. So, [18:45] one of the things that you and I have [18:47] talked about is um and you touched on [18:50] it, but I'd like to expand on a little [18:51] more is this idea that you can do all [18:55] these new things with AI. And so, where [18:58] do you direct the efforts? And so you [19:00] you guys have very clearly directed the [19:02] efforts toward the front-facing stuff. [19:04] So building new products for customers. [19:06] Um and lots of what is discussed in the [19:09] market is backend facing stuff. So [19:11] optimizing your call structure as you [19:12] said everything is a markdown file and [19:15] you know let's let's map out everything [19:16] that everyone does in the organization [19:18] and optimize it to death. Um talk about [19:20] your philosophy for why you want to go [19:23] for the former and not the latter. [19:24] >> Yeah. So I think there was an an early [19:26] narrative and it still you know lingers [19:29] for sure um and there's some truth to it [19:32] um of just we have agents now we should [19:35] be more efficient um we should be able [19:37] to do more with less um and and the main [19:41] thing I actually think about on this [19:42] front is making our people more [19:45] productive uh we have a tool internally [19:48] called Kai which is our sort of [19:50] knowledge AI tool and uh we built it I [19:54] don't know maybe 6 months ago, two [19:56] people built it. Um, now 83% daily [20:00] activives. Uh, or sorry, 83% weekly [20:03] activives, about 60% daily activives at [20:05] Stripe. Uh, seller productivity has [20:07] increased by 20%. Uh, but we're not [20:10] saying, "Wow, we need fewer sellers." [20:11] We're saying, "So, we need a lot more [20:13] sellers." [laughter] [20:14] You know, it's just, uh, you know, [20:15] >> the idea of payback of our sellers just [20:17] got way better. We should have way more [20:18] of them. [20:19] >> Exactly. And then on the flip side, you [20:21] know, operational teams are just feeling [20:23] much more productive. Their day-to-day [20:25] life is is more enjoyable. They're using [20:27] better tools. They're having to do less, [20:29] you know, manual work. Um, and so, you [20:32] know, it's almost trit in tech circles [20:34] now, but the Javon's paradox, you know, [20:36] assets becomes more productive. You [20:38] don't want less, you want more. [20:39] >> Yeah. Exactly. [20:40] >> Uh, and it is very much the case for us [20:42] when it comes to engineers. You know, [20:43] our engineers are just magnificently [20:45] more productive. And so I mentioned [20:48] earlier just making Stripe a platform [20:51] for founders like internal platform for [20:53] founders you know going forward you [20:56] could say well because engineers are so [20:58] productive there'll just be fewer [21:00] engineers in the world and I just really [21:02] don't believe that and that's not what [21:04] we're seeing at all you know you see you [21:06] know that much more software creation we [21:09] you know there was a narrative um 9 [21:11] months ago about you know SAS platforms [21:14] and how there you know vertical [21:15] platforms you know um how they were all [21:17] going to struggle. Uh our um new SAS [21:21] platform cohort is 103% larger than uh [21:25] in 2026 than it was in 2025. So people [21:27] are just saying well there's so many [21:28] spaces that need a SAS platform. I'm [21:30] building it faster and faster and [21:31] faster. And so I think there will be [21:34] more software engineers in the world. Uh [21:38] we will need few of them fewer of them [21:40] to do the things we're already doing, [21:42] >> right? And so Stripe then becomes for [21:46] new grads and early career engineers [21:48] more of an incubator of come in figure [21:51] out a high agency project that can help [21:54] our users. Let's get out of your way. [21:56] Let's create the pave paths. Let's make [21:58] you make it easy for you to go faster [22:00] and we should be able to just build more [22:03] and more and more and you know not just [22:05] random things but adjacencies that sort [22:08] of compound what we're doing today. Um, [22:12] so spend management is adjacent to [22:14] Treasury and that might have been [22:16] roadmapped for two years from now. [22:18] Instead, one engineer has picked it up [22:20] as a project and they're just building [22:21] it themselves. Um, actually I wrote the [22:23] first commit for it, but they're now [22:26] they're now running with it. [22:27] >> Two more capable hands now. [22:28] >> Exactly. More capable hands. Um, so [22:30] yeah, we we um we believe the big [22:33] opportunity and the one that we're [22:35] leaning into is uh doing more with what [22:38] we have and over time doing more with [22:40] more. Yeah, I totally agree with that. [22:42] And um I wrote a piece recently about [22:46] founders being the asset class, like the [22:48] the representation of our market of late [22:50] stage venture, which you guys are [22:52] probably at the forefront of. Um [22:53] founders are the asset class because [22:55] founders are going to be the ones who [22:56] enable the companies to find the next [22:58] product areas. [22:58] >> And so this idea that you have to make [23:01] trade-offs today around do we try and [23:04] build new products for revenue or cut [23:06] costs? Y [23:07] >> I think for companies like Stripe that [23:08] are run by you guys There's probably [23:10] always going to be that next opportunity [23:12] to create additional revenue [23:14] opportunities, you know, find the new [23:16] customer products that are a big hit and [23:18] that is never going to go away. You're [23:19] always going to have the chance to build [23:20] more. [23:20] >> Yes, absolutely. I do think it's a good [23:23] moment for company leaders to create [23:25] some resourcing back pressure because [23:27] that back pressure drives more efficient [23:30] and more expensive use of new tools. Um, [23:32] [snorts] [23:33] but I also imagine that in the years [23:35] ahead, we'll have a budgeting process [23:37] that looks not that dissimilar for what [23:40] it's been in the past of we're doing so [23:42] much with what we have. We can do even [23:44] more if we continue to expand [23:45] resourcing. Um, but in this moment, I [23:48] think it is a good time to say, well, [23:50] let's really do a lot with what we have [23:52] today. Um, because you still see [23:54] increases in productivity, you know, [23:57] from what you what you already have. [23:58] Beyond that, you know, from from a users [24:00] first standpoint, just there's so much [24:02] more they want from us. Um, you know, [24:05] we've heard from from a lot of users um [24:07] that for Stripe Tax, they want global [24:10] filing and it took us a long time to get [24:13] to, you know, US filing on Stripe Tax, [24:16] meaning that, you know, every state in [24:17] the US will just automatically file for [24:19] you. Um, it's hard. It's a hard problem [24:21] because you are dealing with um so many [24:24] different jurisdictions that accept [24:25] filing in different ways [snorts] and so [24:26] it took us a long time to get there. Um [24:29] global filing was the next task. You [24:30] know, we just don't want to have to [24:31] think about this in any market. And we [24:34] built that in about a third of the time [24:37] that it took us to get to US filing. [24:39] >> Wow. With much greater complexity. [24:40] >> Exactly. Much greater complexity. And so [24:42] it's just like you an infinite list of [24:44] of asks and opportunities like that. And [24:46] so if you're just relentlessly focused [24:47] on users, you're going to focus on on [24:50] that. Uh we um we launched recently a [24:54] product called Stripe Treasury. Um and [24:57] it's an interesting moment in sort of [24:59] fintech right now because I think there [25:02] is sort of an unbundling of what banking [25:05] means. And Stripe, you know, is not a [25:07] bank, does not aspire to be a bank, but [25:10] does aspire to provide your banking [25:11] portal and to help you with global money [25:13] movement. This is something that we've [25:15] we've we've been asked for uh for a very [25:17] long time. And so we built Stride [25:19] Treasury to allow our users to hold [25:21] funds in dozens of currencies, you know, [25:23] across many countries and, you know, [25:26] just build what the banking portal [25:28] should be able to do. [25:30] And so, you know, we we've, you know, [25:32] focused on bringing, you know, this [25:36] experience to users uh extremely [25:39] quickly. Um, and what we're finding in [25:43] every case is that the timelines that we [25:46] thought, you know, were were the the [25:49] best doable 2 years ago, uh, are now [25:52] being compressed again and again and [25:54] again. And it's because [25:57] the direction that we're finding, um, [26:00] from an agent engineering standpoint is [26:03] kind of akin to injection molding. You [26:06] know, it's it's it's a bit reductionist. [26:07] Yeah. Uh but when you look at sort of [26:09] the growth in consumer package goods and [26:10] toys and things like that, a lot of it [26:12] just exploded postwar and it was because [26:15] we got really good in the 40s uh at [26:19] melting plastic uh you know it's the [26:20] inject injection molding screw that sort [26:22] of you know melts plastic very evenly [26:25] and you know sort of injects it into uh [26:28] you know a metal casting and that's sort [26:31] of how we're doing code production at [26:33] this point. you know, we're creating the [26:34] molds and the templates and you [26:36] mentioned markdown files like those are [26:37] extremely important. You need your [26:38] agents files in every repo [26:40] >> and as long as you create those patterns [26:43] and you sort of understand the pattern [26:45] of how you integrate with financial [26:46] institutions, you can hand off so much [26:49] of the work to agents and they just get [26:50] it done and all of the time you end up [26:53] spending is sort of on the back end at [26:55] code review. Um, but even then agents [26:58] are doing code review even better than [26:59] humans did before. So just again [27:02] compressing timelines, it just feels [27:04] like a much bigger opportunity than [27:05] optimizing cost structure. [27:06] >> Yeah, I think you may be the first [27:08] person to compare uh modern AI based [27:11] software engineering to injection [27:13] molding. So this is this is great. We're [27:14] covering your ground. And I've actually [27:15] been to an injection molding plant [27:17] before. [27:18] >> It's pretty amazing. [27:18] >> It's pretty amazing, especially when you [27:20] think about the fact that it was [27:21] probably 70 years ago that it was [27:22] created. Uh [27:24] >> James Watson Henry, I think the guy I've [27:27] been in Pennsylvania. [27:28] >> Yeah, there you go. That's uh that's [27:30] great. Um, you know, if you, you know, [27:32] as an investor in Stripe and a large [27:34] large investor in Stripe, um, another [27:36] way that I think about what you just [27:38] described is the opportunity to optimize [27:40] cost is finite, right? You said long [27:42] verse short. I really like. [27:43] >> Yeah. Yeah. Yeah. Yeah. No, it's [27:45] optimizing cost structure is going [27:46] short. Uh, sort of your own future [27:48] potential and just building faster, [27:51] building more is going long, you know, [27:52] your future potential. [27:53] >> It's like the capital allocation [27:54] equivalent of uh like returning cash. [27:57] Yes. As opposed to reinvesting cash. [27:59] Yeah. Yeah. I uh I got it. Okay. Um I [28:02] want to shift over to a new topic which [28:04] you guys are on the cutting edge of and [28:06] which gets a lot of play but almost in [28:09] like a nebula sense. I' I'd like to go a [28:10] little bit deeper on it with you. Um [28:12] which is sort of agents engaging in [28:15] commerce on the internet. So [28:17] >> what is the state of where we are today? [28:19] What are the bottlenecks and how do you [28:21] sort of view a framework for how that [28:23] could be done in the future? [28:24] >> Yeah. [28:26] So we are, you know, we talked about the [28:29] Opus 45 Cambrian explosion moment. We [28:32] haven't had that in a commerce today. Um [28:35] like there isn't a um just there aren't [28:38] a ton of canonical use cases that you [28:40] just see repeated over and over and over [28:42] again. Um and I think there's a few [28:44] reasons for that. One is we're missing [28:46] primitives, [28:47] >> right? [28:47] >> Uh and just that's something we're very [28:49] focused on. you know, we we uh we [28:51] created with Tempo the machine payments [28:53] protocol so that uh services can [28:56] indicate, you know, what needs to be [28:59] paid and how you can pay for it. So, if [29:02] you're selling, you know, an image or a [29:04] piece of content online, you can request [29:07] it and get a 402 response back. Um, and [29:10] it just says like here's how you buy me. [29:12] So, that's a primitive like we think [29:13] machines will want to buy from from [29:16] other machines. Um, and there's really a [29:18] question of like what what should [29:19] checkout look like for agents? And it's [29:23] still sort of an open question. You [29:24] know, browser automation is getting [29:25] better. So maybe agents should just be, [29:26] you know, crawling through checkout [29:28] forms and them out. I think there's [29:30] still a lot of open questions. It's like [29:31] the Yeah, that that's the skuorphic [29:33] version. There's going to be a native [29:35] version. Yeah. [29:35] >> Yeah. I I think that's right. I think [29:36] that's right. Um, so we're we're sort of [29:39] at the missing primitives phase and [29:41] we're sort of figuring that out. And [29:43] then there's a sociological phase of [29:45] just in what ways [29:47] will agentic commerce be better than [29:50] non-agentic commerce. [gasps] [29:52] I think there's some very non [29:54] speculative ways in which it'll be [29:56] better like just checkout pages [29:58] shouldn't exist for humans you know with [30:00] with stripe link and you know shop with [30:02] shop pay and [30:04] it's very easy to get through a checkout [30:06] page these days but should you even need [30:08] to go to one or should you just be able [30:10] to say buy it on a you know product [30:13] display page I think so I think just [30:16] check out pages will go away at the best [30:18] I think they'll go away and so that's [30:20] like a sort of nonspecul ulative but [30:22] maybe less ambitious form of agent [30:24] commerce. [30:25] >> I actually think that's interesting [30:26] because you they will go away for human [30:28] users too. This is not just like only [30:30] Yes. Okay. So, stripe link like you have [30:32] like 400 million users on strip link and [30:34] then shop pay probably has some [30:35] comparably large number and then you can [30:37] you can use those and just automate the [30:38] whole process. [30:39] >> Exactly. Yeah. We launched recently the [30:40] link agent wallet so that there's now [30:42] just a link CLI so an agent can just you [30:46] know sweep up link credentials and go [30:48] use them um with you know humans in the [30:50] loop to say how they can use them um the [30:54] I think one of the places we are the [30:55] most excited about gentic commerce [30:57] actually is B2B so we launched straight [31:00] projects straight projects is you know a [31:03] way to scaffold apps you know that's [31:05] sort of the the narrow definition of it [31:07] uh but the most exciting thing that [31:09] Stripe projects is is a way to provision [31:11] B2B services agentically. [31:13] >> Yep. [31:13] >> So you can go um adop an agent can go [31:16] adoptic Versell uh for hosting and they [31:19] can do it without you needing to go to [31:21] versell.com and do anything. [31:23] >> This is this has actually informed our [31:24] investment thesis in a few investments [31:26] recently that are developer tools and we [31:28] say okay assume that the agents are [31:30] going to be the shoppers in the future. [31:32] Exactly. Like is this the one that the [31:33] agents will want to pick? [31:34] >> Yes. Exactly. Exactly. Um and that's a [31:37] very good framework. um you know, an [31:39] agent can go adopt browserbase. Uh and [31:42] we actually had a really great demo uh a [31:44] while back at Stripe where we um [31:47] >> uh we used uh browserbase uh live to uh [31:51] with via an agent to fill out an NCAA [31:53] bracket. [31:55] >> That's [laughter] cool. That's great. [31:56] >> Which I thought was awesome because I [31:57] have three brothers and you know, we all [31:59] love sports, but I'm always sort of too [32:00] busy to engage in the family pools. And [32:03] so it was so cool to just watch the [32:04] agent, you know, pick up um a [32:06] browserbased session and, you know, fire [32:08] up espn.com and do a bit of research and [32:11] just like fill it out. [32:12] >> How'd it do? [32:13] >> Uh I don't actually know. I don't think [32:15] we ever sort of ran the evals on it, but [32:16] uh [32:17] >> I suspect probably as good as human. [32:19] >> It's probably better. Yeah, exactly. [32:20] Probably better. Um [32:22] >> and so I I I do think agents sort of [32:24] adopting B2B or sort of B2C services [32:28] that are more utilitarian uh will be [32:31] very popular. And so today again as a [32:33] primitive we're just trying to make it [32:35] easier for agents to adopt services. Um [32:40] but I do think the sort of sociology [32:41] will continue to evolve. Um you know the [32:45] the longunning tasks uh are not sort of [32:49] uh I don't know that we've sort of [32:51] figured out exactly how to do them on [32:53] the consumer side. [32:54] >> Yeah. It's not there yet. my um uh I was [32:58] uh trying the other day I was using uh [33:00] cloud code to try to uh compose a song [33:03] for my niece's birthday. [33:04] >> Oh [33:05] >> yeah, it was pretty cool. Um and you [33:08] know in the past without a pseudo I just [33:10] like definitely wouldn't have done that. [33:11] >> Um with ao I might do that. Uh you might [33:14] do it with a Higsfield as well. We're [33:16] talking about them earlier [33:17] >> but I'm not sure that I want to go [33:19] create a $10.99 or $9.99 a month account [33:22] with them. I just want the sort of micro [33:24] consumption for that. And so, you know, [33:27] along with the actual commerce [33:29] primitives, I think there's these micro [33:31] consumption APIs that need to start to [33:33] exist. Um, and browser rates we're [33:35] talking about, they're sort of leaning [33:36] into it. [33:37] >> But I'm very bullish on, you know, all [33:40] these services standing up, [33:41] >> you know, sort of not necessarily [33:43] anonymous, but just ephemeral or [33:45] one-time consumption. And I think that [33:47] will really unlock agent commerce too [33:49] because then you won't have to say well [33:50] I need to pay these guys 10 bucks a [33:52] month or these guys 10 bucks a month [33:53] these accounts and all this stuff just [33:56] use the service and pay [33:57] >> agent do this discover the services your [33:58] budget is $15 go [34:01] >> my uh my 10-year-old has been creating [34:03] rap songs so uh so I'm very familiar [34:05] with all of these services and uh and by [34:08] the way you know the the quality is [34:09] actually pretty good and he's you know [34:11] he posts them on Spotify [34:12] >> uh but uh but the execution maybe is not [34:15] quite as good The lyrics are very good [34:17] from AI and the music itself is pretty [34:19] good. So it's yeah I would welcome that. [34:22] Um so the the critical thing I think or [34:24] one of the critical things that you just [34:26] said is this concept of micro payments [34:28] or microtransactions. So this has been [34:30] something that has been talked about [34:32] >> probably since the advent of the [34:33] internet uh as an opportunity like why [34:35] do you think now it could work? [34:38] >> Well agents really increase human [34:41] agency. I guess it almost sounds um uh [34:43] almost sounds redundant, but you know, [34:46] talking about these um these [34:47] applications, uh composing a song for [34:49] your niece, filling out your NCAA [34:51] bracket, whatever else, it's really hard [34:53] to do without agency. They take a lot of [34:55] time. You're creating all these [34:56] different accounts. [34:58] And so, I'm just very bullish on [35:00] services leaning into saying you can use [35:03] me ephemerally, you can use me um in a [35:06] very um you sort of lightweight way. Um, [35:11] and I think to make that work, you're [35:14] just going to need to support [35:16] microtransactions. Um, [35:19] I think the case against [35:20] microtransactions has always been, well, [35:23] you're trying to consume content and if [35:26] you're trying to sell an article, you're [35:29] always going to be squeezed between the [35:30] subscriber business model and the free [35:32] business model. [35:32] >> Yeah. Like you you want to capture the [35:34] excess as part of the subscriber. [35:35] Exactly. Like that's the business model. [35:37] or you do it with ads and just like [35:39] they're one or the other is better than [35:41] what you're doing. [35:42] >> And I think that was probably true in [35:44] the past. But as you give agents more [35:47] complex tasks, you know, you want them [35:50] to be these little sort of hummingbirds [35:51] like going around the internet just [35:52] slurping up a little data here, you [35:54] know, pulling it over here, you know, [35:56] some very transient ephemeral storage, [35:59] doing a little compute over here. And [36:01] you don't want the individual human to [36:03] have to think about what they're using. [36:06] you want them in that these services are [36:08] secure, right? [36:11] But you don't want them to sort of [36:12] create accounts everywhere. So I think [36:14] microtransactions will just be necessary [36:17] for that economy to exist, the sort of [36:20] agentic economy. [36:21] >> And then on the flip side, they're now [36:23] eminently possible because of stable [36:25] coins. Um, [snorts] [36:26] >> and stable coins are today relatively [36:30] unorganic. So if you're a human, you [36:32] know, you have to go jump through a [36:34] bunch of hoops. But if you just give an [36:36] agent a budget, they can easily, you [36:38] know, take dollars, move into stables [36:41] or, you know, just use a share a stored [36:43] balance um and find a way to check out [36:47] and they don't they don't mind the sort [36:48] of back and forth stables. Yeah. [36:50] >> Yeah. Yeah. I totally agree with that. [36:52] And look, we've we've seen that [36:54] transition that's happened in a lot of [36:56] software business models today, right? [36:58] like the predominant business model of [37:00] selling subscriptions and seeds and you [37:02] know we've now seen a shift to [37:03] consumption and I think you know there's [37:05] probably analogies where um you know the [37:08] consumers are better off because they [37:09] can get get access to more stuff and [37:11] then the businesses can access more [37:13] people which they otherwise couldn't [37:14] access with a subscription service. So [37:16] my hope is that that [37:17] >> you know that does come into play. Um so [37:20] on the topic of stable coins so we've [37:22] covered sort of how you guys build [37:24] software how you move so fast uh in an [37:26] AI development world you know we've [37:28] talked about the idea of commerce you [37:30] know aentic commerce um talk about [37:32] stable coins like you said there's a lot [37:34] of friction in the process today but [37:35] just state state of the stable coin [37:37] market and what you see the opportunity [37:39] to be for the next call it five years [37:42] [sighs and gasps] [37:43] >> so I'm an infrastructure nerd so I [37:45] always will go there first and stable [37:48] coins just a better platform for moving [37:50] money than exists otherwise. Um [37:52] >> we strongly agree. [37:53] >> Yeah. Um and better for a couple [37:55] reasons. You know there are certain [37:57] countries uh that have you know rolled [37:59] out really good payment systems. Most of [38:01] them are nationalized. You know UPI in [38:04] in India. Um actually it's interesting [38:07] when you look at these economies uh the [38:09] percentage of sub$5 payments in India I [38:14] believe is on UPI is I think something [38:17] like 86% whereas in the US cards I think [38:20] it's [38:21] >> single digits%. So, you know, there's [38:23] some good, cheap, fast payment schemes [38:27] that are nationalized. Picks, another [38:29] one in Brazil, which has grown [38:30] meteorically, [snorts] [38:32] >> but you need a shelling point for the [38:34] global economy, right? [38:36] >> Just what can we all agree on? And [38:38] that's where crypto rails solve this [38:40] political problem where just this [38:42] platform works everywhere. [38:45] >> And so, if we just all used it, the [38:47] world, the global financial system would [38:49] work better, be faster, be cheaper, and [38:51] so on. Um, so if we all like went to [38:53] sleep tonight and woke up tomorrow and [38:54] we all held stables, it would just be a [38:56] better global [38:57] >> less friction and less cost in the [38:59] economy. [38:59] >> Exactly. Um, now that might be one of [39:02] those, uh, should work this way, may [39:05] never, you know, work this way problems [39:08] and so we are sort of resolutely focused [39:10] on trying to change that. Um, we talked [39:13] earlier about stripe treasury and, you [39:16] know, we decided to make stable coins [39:18] native to stripe treasury. So you can [39:19] just hold a balance in stables just like [39:22] you hold a balance in USD or EUR or GBP [39:24] or anything like that. [snorts] [39:26] So um we believe that the opportunity [39:29] for stables is just faster um money [39:32] movement, cheaper money movement and [39:35] more global money movement. Um, right [39:38] now you can be a Stripe user in I can't [39:42] remember what the exact number is, but [39:43] around 60 countries in fiat, but you can [39:46] be a Stripe user in stable coins in I [39:49] think about 150 countries. So just [39:51] bringing more people into the [39:55] uh online economy uh in a way that [39:58] allows them to you know transact with [40:00] you know the AI companies and you know [40:02] as the as the economy the software [40:04] economy gets more and more global [40:06] bringing more people into it is more and [40:08] more valuable. You can build a powerful [40:11] AI company with two engineers in [40:13] Thailand uh just as well as you can do [40:15] it in the US or as well as you can do it [40:17] in Brazil. So you know for stables for [40:19] us it is higher performance more global. [40:21] >> Yeah and I like the position that you [40:23] guys are in as it relates to stable [40:24] coins just because again talk about [40:26] meeting the market where it is like [40:29] people don't want to cut over full [40:31] wholesale and just you know drop all [40:33] their fiat and move over to stable [40:34] coins. So I think [40:35] >> the relationships that you have and the [40:37] sort of comprehensive offering that you [40:38] have allows for adoption to happen at [40:40] the pace that the enterprises want to [40:42] adoption adopt at or startups. [40:44] >> Yeah. Yeah. And all of this works, you [40:47] know, when you really lean into it. Uh [40:49] Felix Bogar, Felix, I think, is as as [40:51] they're called now. Um which started as [40:54] a remittance company between um the US [40:56] and and Mexico and now, you know, has [40:59] other corridors as well. Um they built [41:01] it on stables and uh a few years in, [41:05] they're now between 5 and 10% of [41:09] remittances along that corridor, which [41:11] is the largest remittance corridor in [41:13] the world. [41:14] >> It's amazing. And so, you know, you [41:15] think about how long it takes companies [41:17] to um, you know, to to move fiat money [41:21] efficiently. You know, how long it took [41:23] initially for the wisest of the world to [41:24] do this, let alone the Western unions [41:26] before it. And, you know, these are [41:27] great companies, but to get to 5 to 10% [41:30] in just a few years is amazing. [41:32] >> That's remarkable. Yeah. Um, I'd love to [41:35] have you talk a little bit about tempo. [41:36] Yeah. So, um, you know, obviously it's a [41:39] big important project that you guys are [41:40] are in the center of. [41:42] um talk about, you know, where that is [41:43] and and what the aspirations for it are. [41:46] >> Yeah, I think this um I I'll just play [41:49] the infrastructure nerd card again for a [41:51] second. Um there are great blockchains, [41:54] you know, um and I think there are, you [41:58] know, there was no obvious need for [42:01] another blockchain outside of payments, [42:03] >> right? [42:04] >> And there's a whole lot of reasons why a [42:07] payments specific blockchain can work [42:10] really well. you know privacy for one [42:12] because you know blockchains are [42:13] generally public and you sort of reverse [42:15] engineer you know what's happening on on [42:17] on blockchains so I want to make sure [42:18] that privacy is a first class primitive [42:21] uh making sure that throughput is never [42:23] sacrificed um blockchains are [42:27] disproportionately used for trading [42:29] >> and so when you see you know massive [42:31] trading events happen you'll typically [42:33] see you know performance degrade a lot [42:35] transaction fees like finding a way to [42:37] create a a blockchain where the [42:39] transaction fee is never going to spike [42:41] because you know other blockchains have [42:43] >> capability is important. Yeah. [42:44] >> Exactly. And and and they have you know [42:46] floating sort of gas fees and so the the [42:50] project there is just saying how do we [42:52] move money you know as efficiently [42:54] consistently and cheaply as possible. [42:57] Still in the early stages of building [42:59] out getting a lot of great traction you [43:00] working with companies like um Door Dash [43:03] you know making it sort of the default [43:04] but not only blockchain in Stripe and uh [43:07] feeling very optimistic about it. That's [43:09] great. That's awesome. [43:10] >> Um, [43:11] >> last topic I would love to cover with [43:13] you is, um, just the sort of token [43:16] economy. [43:17] >> Um, and so, you know, it sounds a little [43:19] buzzwordy, but you guys are, you know, I [43:21] think over two trillion of volume now. [43:25] >> Um, and I think, you know, you and I [43:27] would probably share the belief that the [43:29] token economy is going to be one of the [43:30] biggest things we've ever seen. Um, so, [43:33] you know, what do you think is kind of [43:34] the state of play for accessing tokens? [43:38] you know, obviously there's a lot of [43:39] work that gets done in first party [43:40] applications today. Um, but you know, [43:43] what do you think is sort of the future [43:44] state of how we all access models and [43:46] tokens? [43:47] >> Yeah. [43:48] >> Well, I love I've always loved the word [43:50] token. Um, and the reason uh is that it [43:54] just sounds like an approximation of [43:55] money. [43:56] >> Yeah. [43:56] >> And [laughter] increasing that's just [43:58] what it is, right? We're seeing, you [43:59] know, we talked earlier about, you know, [44:00] free trial abuse, multi-accounting, [44:03] and the attacks that we're seeing [44:06] against users who have sort of general [44:09] whose platforms are general purpose [44:12] token consumers. You know, you can do [44:14] just about anything with tokens on [44:15] cursor, right? Um, you do just about [44:17] anything with tokens on on replet. [44:20] >> Uh, so the attacks against these users [44:22] are very sophistic sophisticated. you [44:24] know, they're they're very reminiscent [44:25] of what we see, [44:27] >> you know, in terms of people trying to [44:28] steal money from Stripe users. [snorts] [44:31] >> And so there's this blurring that you [44:34] see between tokens and dollars. [44:36] >> Yeah. [44:37] >> And so for us, we think about [44:39] >> wanting to help users move money, store [44:41] money, send money safely, compliantly, [44:44] and so on. [44:45] >> And the uh we now feel this mandate to [44:49] do the same thing on tokens, right? [44:51] to um protect our users in the same way [44:55] and you know over time I think this is [44:57] only going to happen more right you know [44:59] we talk about [45:00] >> some of these longunning agentic tasks [45:03] and a lot of them will replace you know [45:05] services and you'll just be able to say [45:07] you know close my books and I used to [45:09] pay you know a human in tokens for that [45:12] now I or many humans now I pay fewer [45:14] humans or a human or sorry in dollars [45:16] for that now I pay fewer humans or a [45:17] human tokens [45:18] >> in tokens right um because they're [45:20] augment rented by by agents. So, we just [45:24] want to make sure that moving between [45:27] tokens and dollars is as seamless as [45:29] moving and safe as moving between, you [45:31] know, dollars and euros. [45:33] >> We're at the beginning of this journey, [45:35] but we think it's going to be a big part [45:36] of the future of Stripe. [45:38] >> Yeah, that's awesome. Yeah, it's uh it's [45:40] it's super exciting and and look, [45:42] personally speaking, I think you guys [45:44] >> have a real right to play a role there [45:46] uh because of the relationship that you [45:48] have with so many so many companies [45:49] since inception all the way through to [45:50] enterprise as you said. [45:51] >> Yeah. And we we use it's funny um we're [45:53] we're always a little bit careful [45:54] internally about talking about about [45:56] rights and so we talk about mandates. [45:58] >> That's good. Yeah. We feel [45:59] >> win and then win it again. Exactly. We [46:01] have a mandate to help users with this [46:04] and [46:05] >> uh and so that's why we're really [46:06] leaning into it. You know, there's uh [46:08] we're also focusing even more on spend [46:10] management for our customers. You know, [46:12] they're using Stripe Treasury. They're [46:13] thinking a lot about how much they're [46:14] spending and how to, you know, uh reduce [46:17] their spending. How their spending ties [46:18] through to revenue and you know, a lot [46:21] of token usage is actually in product. [46:24] >> Yeah. [46:25] >> So, you see sort of two different types [46:26] of token usage. One is, you know, to [46:28] build things and this is sort of the [46:30] opex management side of of of token [46:34] management. And then there's the product [46:36] efficacy side right of of token [46:39] management is how good is my token [46:42] oriented or sort of token driven product [46:45] you know how should I shift between [46:47] different models to make it the most um [46:50] effective and so we're really thinking [46:52] about how do we help users on both of [46:53] these [46:54] >> yeah for sure I mean certainly in the [46:55] latter this sort of you know uh position [46:58] as an orchestration layer that helps you [47:01] effectively get the most out of tokens [47:03] and [47:04] >> you also maintain some form of control I [47:06] think has like never been more top of- [47:08] mind than it is right now. [47:09] >> Yes, absolutely. Yeah. And it's it's [47:11] interesting because um there's a [47:14] hypothesis that you know software will [47:17] be severely commoditized. Um, and it's [47:21] plausible. You know, I think we're we're [47:22] at the u let's say we're either in the [47:25] singularity or creeping towards the [47:26] singularity and it's it's very hard to [47:29] um to estimate what a future looks like [47:31] where you know models are sort of [47:33] recursively you know generating models [47:35] and so on and so on. Um but we are [47:39] seeing the exact opposite right now [47:41] where you know software creation is [47:43] exploding. [47:44] >> Yes, [47:44] >> customers are monetizing faster than [47:47] ever. You know, we talked about our 2026 [47:48] cohort growing 50% faster than our or [47:52] 50% larger and growing faster than 25 [47:54] cohort. That one being 70% larger and [47:56] growing faster than the 24 cohort. [47:59] >> And so, you know, more software being [48:01] created. It's being adopted faster than [48:03] ever before. [48:04] >> And then more Stripe products are being [48:06] adopted than ever before. You know, uh [48:08] 11 Labs is using 14 uh stripe products. [48:12] Sort of wish it were 11 in a way, but [48:13] actually I'm I'm glad it's 14. [48:15] >> Just make it 22. He got really close [48:17] with Matty and he does everything in [48:18] denominations of 11. So you got to shoot [48:20] for 22. [48:20] >> Exactly. And you know there's a lot more [48:23] to software than just building right. [48:24] There's the expertise about how [48:26] abstractions should thread through your [48:29] entire business and you sort of keeping [48:31] those up to date and you have you know [48:33] turnover inside of your company and so [48:34] you build a system internally and then [48:36] you know it starts to sort of degrade [48:38] and that person leaves the company and [48:40] the expertise goes away. And so we think [48:42] about just how can we help companies [48:44] across their entire revenue stack, you [48:48] know, manage cash, manage revenue, close [48:50] the books faster, grow faster, globalize [48:52] faster, and that opportunity, I think, [48:55] is just larger than it was ever ever was [48:57] before. [48:57] >> Yeah. Couldn't couldn't agree more. [48:58] >> Yeah. Um, so in closing, uh, I wanted to [49:02] just get your take, and this is I don't [49:04] want to get too deep into, you know, the [49:06] role of humans versus AI, but one of the [49:09] things that Stripe has always been [49:11] universally recognized for and [49:14] appreciated for, I think, is your taste, [49:16] right? And it sounds funny because [49:17] you're an infrastructure company, but I [49:19] think it's it's generally agreed upon. [49:21] >> Um, how do you continue to uphold that [49:24] at such a large scale and with so much [49:26] of the work that you're doing now being [49:27] done by AI? [49:29] >> [laughter] [49:29] >> Mhm. [49:30] >> Yeah. It's interesting how much the word [49:32] taste is being used right now. [49:34] [laughter] [49:35] >> And it's sort of hard sometimes to, you [49:37] know, the the cynical view is that it's [49:39] how we're all justifying our future [49:41] value. [49:42] >> Yeah. That no one actually has taste and [49:43] it's just we're all reproducing [49:45] reproducing machines of things that we [49:47] had previously perceived. [49:48] >> Yes. Yes. suddenly my fingers are no [49:50] longer as useful, but this nebulous [49:53] notion of taste is um and and and models [49:56] will never have taste and so on, which [49:57] is probably just not true. Um [49:59] >> yeah, I agree I agree with you on that. [50:01] >> But uh to your um you know, more [50:04] approximate to your question, you know, [50:07] product quality is really really core to [50:10] our culture and to our identity. And [50:14] it's for a few reasons. Like one is just [50:16] we want users to have amazing tools that [50:19] are carefully crafted that you know can [50:22] make their um make them go faster you [50:24] know enjoy uh you know the the um the [50:28] journey of building a company that much [50:30] more you know it's not an easy journey [50:32] >> but if you have you know tools that are [50:34] surprisingly great it feels that much [50:36] better. So it's it's a big user lens. Um [50:39] >> another dimension of it is it's just [50:41] more fun. you know, you look at what you [50:44] created with your team and you're like, [50:46] "Wow, that looks amazing." Versus, you [50:48] know, you kind of janked this thing [50:50] together and it barely works and, you [50:52] know, it just isn't as fun, you know, to [50:54] to to, you know, go through the slog, [50:56] the long the long nights, the long [50:59] weekends and so on to build it. Um, [51:02] and [51:04] I think the single thing I would say uh [51:06] for scaling cases, right? Well, actually [51:08] two things. One is I think it has to be [51:09] a tops down. You just have to say it% [51:11] over and over and over again. You know, [51:13] it's like it's like the the the the [51:15] company strategy of when the startups [51:16] win them again. It's like talk about [51:18] quality and talk about it again. [51:20] >> And then the other thing is just use the [51:22] product. [51:23] >> Yes. [51:23] >> And make it easy to use the product. You [51:26] know, in in you know, uh aviation, you [51:28] know, simulation is is so important. [51:31] Like you can't uh you can't just say [51:33] like I hope the jet works. You have to [51:34] kind of prove to yourself that it does [51:35] before you ever fly it. um or Nvidia, [51:38] you know, when they hit really hit [51:38] escape velocity was when they, you know, [51:41] went from needing to wait for the fab to [51:42] give them the chips to try the chips to [51:44] actually simulate [51:45] >> chip performance. And so we actually [51:47] invest a lot in simulating the usage of [51:50] products. [51:50] >> Yeah. Wow. [51:51] >> And so you can point to an account and [51:53] say, you know, give me something that [51:56] looks like that. [51:57] >> Yeah. But make it you know completely [52:00] PII you know free you know randomize the [52:03] growth rates so I don't actually know [52:05] how big they are but just you know the [52:07] day-to-day problems that they face you [52:10] know the ups and downs of their business [52:11] the seasonality you know give me a sense [52:13] of of that and then you know make it [52:16] feel live you know make it so that I'm [52:18] feel like I'm receiving disputes from [52:20] customers and refunds are happening and [52:22] of course these aren't real disputes [52:24] they're not real refunds but just give [52:26] me the experience that user is having. [52:28] So I can really live in it and and use [52:30] that with my team. And so I think, you [52:32] know, you have the the mandate to use [52:34] the product and we really ask our EMs to [52:38] lead this because they're the ones who [52:39] control the resources and they're the [52:40] ones who can sort of say, "Wow, this [52:42] does not feel good. It needs to get [52:43] fixed. So we're going to, you know, uh [52:45] our next sprint is going to be elevating [52:47] quality." And the other side, it's how [52:48] do you give people the tools to do it [52:50] more easily. Um so scaling taste for us [52:52] is you know culture and it is the you [52:56] know shival lifts and then it is the [52:59] daily rhythms of just using the products [53:01] and stepping into these issues. [53:02] >> That's awesome. That's so great. Uh will [53:05] this is so fun uh to talk to you about [53:07] all these topics. you know, obviously [53:08] covering uh how you're building so much [53:11] product inside Stripe, which is which is [53:13] super unique uh and right on the cutting [53:15] edge, I think, um and the future of how [53:18] we engage in agent commerce and and [53:20] stable coins. Um you know, one of the [53:22] things that I think you guys have said [53:23] is, you know, how do we assure that we [53:25] build the next Stripe inside of Stripe? [53:27] Yes. And uh I love that in closing and [53:30] um I really appreciate the time. [53:32] >> Yeah. Thank you. Great to be here.