Zipline Drone Delivery Costs Drop to $12 to Undercut Ground Transport

Zipline无人机配送成本降至12美元 低于地面运输 随着无人机网络向日均100万单规模扩张，自主飞行的全成本单体经济效益已低于人工汽车配送。 Zipline的自主飞行里程已超过1.4亿英里且保持零安全事故，证明了飞行器物理硬件仅占自主物流网络的15%，而85%依赖于辅助基础设施与空中交通融合。从在卢旺达开展挽救生命的医疗物资配送、将产妇死亡率降低51%起步，该公司通过双冗余安全架构、专有声学设计以及持续的垂直整合，克服了严峻的监管与工程难题。如今，在自主空中配送正式比传统地面运输更便宜、更快速且更安全的拐点推动下，Zipline正向日均100万次配送的目标迈进。最终，创始人呼吁重振雄心勃勃的实体基础设施和美国硬件制造业，以变革全球物流。 Zipline在卢旺达早期的实践表明，无人机硬件仅占自主物流系统复杂度的15%，其余85%由库存与订单管理、空域监管整合及空中交通管理等配套基础设施构成。 (0:00) Zipline网络已从最初单一医院扩展至8个国家的5,000多家医疗机构，通过即时血液配送将当地产妇死亡率降低51%，并达成了与美国国务院价值5.5亿美元的商业外交合作。 (6:49) 在实现超1.4亿英里自主飞行且保持零安全事故的记录下，Zipline利用消费级智能手机供应链组件搭建双冗余飞行计算机与独立仲裁架构，以极低成本实现客机级安全冗余。 (11:15) 系统采用在100米高空悬停结合系缆投递舱的设计，由舱载Nvidia GPU及感知系统主动沿X/Y轴机动规避地面障碍物，在保障投递精度的同时将地面噪音控制在微风拂叶水平。 (20:25) 随着配送目标从累计100万次迈向每日100万次，Zipline将监控模式升级为单名“舰队指挥官”管控100架无人机，并通过自主研发AI空管与全栈垂直整合应对庞大的低空飞行流量。 (27:21)

Zipline co-founders Keller Rinaudo Cliffton and Eric Watson argue that autonomous aerial logistics cannot scale through off-the-shelf drone hardware alone, but requires end-to-end vertical integration where auxiliary ground infrastructure and software represent 85% of total system complexity. By pairing aerospace failover architectures with low-cost consumer electronics, Zipline has flown over 140 million incident-free commercial miles while driving per-flight costs below conventional road delivery. Vertical Integration and Low-Cost Aerospace Architecture (13:55) Zipline’s leadership explains that physical drone hardware accounts for only roughly 15% of autonomous logistics complexity, with the remaining 85% encompassing inventory management, automated docking, regulatory integration, and airspace control systems. After early experiments with commercial quadcopters and defense contractor components resulted in repeated crashes, the company transitioned to complete vertical integration. Zipline now designs 700 unique components and 43 major sub-assemblies in-house—including custom electric motors, battery management systems, and flight controllers—subjecting them to Highly Accelerated Lifetime Testing across thermal extremes from -25°C to 49°C (-13°F to 120°F). To achieve commercial airliner safety standards without multimillion-dollar aerospace pricing, Zipline adapts components from high-volume smartphone supply chains. The aircraft architecture deploys dual-redundant flight computers operating simultaneously alongside a lightweight third arbiter computer that actively directs actuator failover if the primary node falters. Combined with a culture influenced by SpaceX's engineering approach of aggressively deleting unneeded mechanical parts, full-stack control enables rapid iteration between factory assembly and airfield testing within single-day cycles. Acoustic Isolation and Real-World Environmental Navigation (20:25) Accumulating 140 million autonomous commercial miles with zero safety incidents exposed Zipline to unpredictable physical edge cases unrepresented in software simulations. Atmospheric disruptions from solar flares and coronal mass ejections altered radio frequency speeds between satellites and ground stations, corrupting standard GNSS reception. Zipline addressed this by hardening receiver software for centimeter-level GNSS calculation during ionospheric delays and adding secondary, non-GNSS redundancy systems to guide aircraft safely during satellite signal degradation. To deliver packages safely in suburban neighborhoods, Zipline developed a two-tiered tethered delivery architecture. The primary fixed-wing aircraft hovers at roughly 100 meters (328 feet), keeping high-thrust propeller acoustics to the volume of rustling tree leaves for bystanders below. An autonomous delivery pod descends via tether, utilizing an onboard Nvidia GPU and computer vision to identify micro-obstacles like outdoor furniture or pets, actively maneuvering along its X and Y axes to place packages within an 18-inch radius on unmapped ground terrain. Fleet Commander Ratios and Decentralized Airspace Management (27:21) Achieving commercial economic viability required overturning early regulatory mandates that enforced a 1:1 ratio between human pilots and autonomous drones. As Zipline demonstrated verified flight safety over millions of miles, regulators incrementally widened this threshold to 1:3, 1:20, and eventually 1:100. Formally redesignated as "fleet commanders" in FAA filings, single human operators now supervise constellations of 100 autonomous drones mid-flight across extensive regions like Dallas-Fort Worth, intervening solely on high-level operational exceptions while cloud systems handle dynamic routing and automated hub docking. As Zipline scales toward an engineering goal of 1 million deliveries per day, the company projects its flight volume will reach 40 to 80 times the daily flight count of all conventional U.S. airlines combined. Because existing air traffic control relies on aging manual conventions and faces severe controller retirements, Zipline designed an autonomous airspace management platform. Drones continuously broadcast telemetry and negotiate deconfliction trajectories in real time, establishing open protocols intended to support shared, decentralized commercial skies. Cost Curve Inversion Against Ground Automobile Logistics (49:21) When launching blood deliveries in Rwanda in 2016, Zipline charged $30 per flight while incurring fully burdened costs of $300. Through eight years of hardware consolidation and automated continuous 24/7 flight ops, Zipline reduced fully burdened unit costs to $12 for long-range international flights. Keller Rinaudo Cliffton reports that per-delivery expenses have now crossed below the cost of human automobile delivery, eliminating ground vehicular congestion, carbon emissions, and customer tipping fees while fulfilling orders in under five minutes. This cost inversion is accelerating commercial demand. Having taken roughly a decade to achieve its first 1 million cumulative lifetime deliveries, Zipline is now receiving requests from individual commercial partners seeking 1 million deliveries per day. Based on consumer ordering habits observed in suburban Texas hubs, Cliffton projects that substituting 4,000-pound passenger cars with instant aerial delivery could expand the current 5.5 billion annual U.S. delivery runs into an addressable domestic market of 55 billion automated drops per year. While Zipline aims to replace ground road transport with autonomous aerial logistics, expanding daily operations from 5,000 flights to 1 million daily deliveries will test whether decentralized air traffic protocols can safely share low-altitude airspace at unprecedented densities. Additionally, scaling factory production to 20,000 aircraft annually will challenge Zipline's fully integrated manufacturing and automated depot maintenance to eliminate component defects across tens of thousands of active vehicles.

Zipline联合创始人Keller Rinaudo Cliffton与系统工程负责人Eric Watson阐述了该公司如何通过全栈垂直整合、航空级冗余架构和底层运营创新，建立起累计自主飞行超1.4亿英里且零事故的商用无人机物流网络。随着单次交付成本从最初的300美元降至12美元，该系统正在迎来跨越传统汽车公路配送的经济学与运力拐点。 场景破局与全栈垂直整合的系统工程 (0:00) Zipline于2014年转向自动化物流，但早期在美国因超视距飞行（BVLOS）缺乏监管路径而难以推进，促使其在2016年转向卢旺达开展血液应急运送。这一落地经验让团队认识到，无人机硬件本身仅占系统复杂度的15%，其余85%在于订单管理、空管调度以及对接国家民航与医疗体系的地面外围基础设施。针对非洲复杂环境的验证不仅实现了产品市场契合，宾夕法尼亚大学的同行评审研究亦显示，该服务使服务区域的产妇死亡率降低了51%。 为了兼顾极高可靠性与低于地面交通的成本结构，Zipline摒弃了拼凑通用组件的做法，转而执行全栈垂直整合。其飞行器包含700个自研专用零部件与43个主要子组件，覆盖高推重比定制电机与电池管理系统。在工程文化上，团队践行第一性原理与部件精简原则，例如将一代机型（Platform 1）机身长达1米的可展开尾钩完全剔除，简化为2厘米静态结构，将减速捕获的机械动作转移至地面回收机器人，从而大幅减轻空中机械负载与故障风险。 消费级供应链下的航空级双冗余安全体系 (11:15) 在保持1.4亿英里零事故记录的背后，Zipline的设计哲学建立在“假定所有部件终将失效”之上。商业客机如Boeing 777依赖昂贵的航电备份，而Zipline通过智能手机供应链的低成本消费级元器件，构建了双主动飞行计算机系统。两台计算机同时计算飞行轨迹并处理传感器数据，由独立的轻量仲裁计算节点监控运行状态并在主节点异常时实现毫秒级切换，以此在极低物料成本下达到航空级容错能力。 面对高频实际飞行中的极端物理边界，Zipline针对太阳黑子爆发引发的电离层扰动，自研了抗延迟的高精度GNSS算法及非卫星辅助导航层，防止无线电信号畸变破坏降落精度。此外，所有硬件必须接受高加速寿命试验（HALT），在-25°C至49°C的温控室、风洞及高湿高盐环境中验证耐受极限。红杉资本合伙人Alfred Lin曾提出对标要求，推动团队将安全目标设定为Waymo自动驾驶汽车安全水平的两倍（即人类驾驶安全性的约20倍）。 双层系留静音投递与1比100网络化运管 (20:25) 为进入高密度城郊配送场景，Zipline研发了悬停与系留两层架构。主机在约100米（328英尺）高空悬停，将推进器的高分贝噪音阻隔在高空，地面噪音仅相当于树叶沙沙声。系留释放的智能投递舱搭载Nvidia GPU及计算机视觉传感器，下落过程中动态检测露台桌椅、宠物或杂物等障碍物，利用X/Y轴微调推进器修正落点，在无需预先测绘坐标的情况下实现半径18英寸（约45厘米）以内的精密着陆。 空域运维模式经历了根本性重组。早期受FAA监管要求，地面必须维持1名操作员监控1架飞机的1:1比例，造成单位经济效益低下。随着自主飞行系统安全性得到验证，监管机构逐步放宽至1:3、1:20直至当前的1:100。操作员被正式更名为“舰队指挥官”（Fleet Commander），转为基于云端网络处理极少数异常情况，日常起降、空中路径规划与自动化机库对接均由算法自主完成。 空管重构、规模跃迁与单均成本拐点 (31:52) 传统美国空中交通管制（ATC）依赖20世纪50年代的人工雷达与纸质追踪卡，加之管制员队伍老龄化严重（50%超过45岁，20%濒临退休），无法承载密集的低空物流。美国最大商业航空公司的日航班量峰值约为5000架次，而Zipline正从日均5000次配送向年底3万次、最终日均100万次推进，其低空飞行量将达到传统全美商业客机总和的40至80倍。为此，Zipline研发了基于自主广播与协同避障算法的去中心化空管软件，推动建立开放的空中防撞通信标准。 运力的指数级扩张带动了全成本的断崖式下降。2016年在卢旺达首飞时，单次交付的完全分摊成本高达300美元（当时向客户收费30美元）；经过数轮软硬件精简与规模摊薄，该成本历经120、75、40、28、18美元，目前国际长距离航线已降至12美元。Keller指出，完全自主空中物流的单位经济成本正正式跨过传统燃油车外卖配送的临界点。根据达拉斯商用试点数据，即时配送的便利性将大幅激发消费频次，美国市场潜在容量预计可从年均55亿次人工配送扩展至550亿次自主航空投递。 Zipline的发展验证了深科技硬件与全栈制造在改造传统物流中的可行性，但随着业务规模由累计百万次冲刺至日均百万次，企业仍需在超高频维护容错、去中心化自主空管协议的国际标准化，以及各国低空空域适航监管政策的全面放开之间持续寻求平衡。

## Background 本期内容由红杉资本（Sequoia Capital）于 2026 年 7 月 7 日发布于 YouTube 平台。视频由红杉合伙人 Alfred Lin 与 Pat Grady 主持，对话 Zipline 联合创始人 Keller Rinaudo Cliffton 与系统工程及安全负责人 Eric Watson，深入剖析这家全球最大商业自主飞行系统如何以双飞控容错等全栈工程实现 1.4 亿英里零事故飞行，并揭示无人机本体仅占系统 15% 复杂度背后的万亿级物流重构之路。 ## TL;DR 节目全面梳理了 Zipline 从 2016 年因美国监管死局远赴卢旺达空运血液、将当地产后大出血孕产妇死亡率降低 51% 起步，逐步演进为全球自主物流巨头的历程。嘉宾详细阐述了硬件仅占 15%、全栈自研与空域调度系统占 85% 的工程架构，包括抵御太阳耀斑干扰的双飞控仲裁冗余，以及搭载独立 GPU、实现 100 米高空精准悬停投送的投递舱（Delivery Pod）黑科技。随着人机监控比从 1:1 跃升至 1:100 的“舰队指挥官”协同模式，单次交付成本从初期的 300 美元断崖式降至 12 美元，正式全面击穿地面传统燃油汽车配送的成本线。面对未来全美 550 亿次即时配送的爆炸性需求，Zipline 践行质疑需求与极致删减零件的第一性原理，正以日均百万单为目标重塑低空物流基础设施。 ## Atmosphere & style 整场对话兼具深度的硬核工程硬度与创业自省的克制幽默，既有投资人与创始人之间直面早期“十倍亏损”与严苛对账的坦率交锋，也浸透着年轻工程师扎根非洲前线、通宵排障的硬派极客情怀。叙事节奏张弛有度，从早期如同“黑暗骑士”般坐了十年硬件冷板凳、在极端温差与破坏性测试中硬磕细节的苍凉磨砺，自然过渡到打破传统空管瓶颈与地面拥堵的宏大构想。听众能真切感受到一种褪去概念炒作浮躁后，由上千万英里飞行数据和挽救鲜活生命所淬炼出的沉静、自信与偏执。 ## Bulletin - [06:49](https://www.youtube.com/watch?v=6bGxm8gX41o&t=409s) Zipline 累计完成 1.4 亿英里自主商业飞行且保持“零安全事故”，覆盖 8 国超 5,000 家医疗机构，并在覆盖区域使产后大出血等导致的孕产妇死亡率骤降 51%。 - [04:09](https://www.youtube.com/watch?v=6bGxm8gX41o&t=249s) 无人机硬件本体仅占整体系统复杂度的 15%，其余 85% 为仓储冷链管理、空域监管直连、预测性维护及全自动充电机巢等深层基础设施。 - [26:57](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1617s) 人机协同监管比例已从早期的 1:1 严苛监管破冰跃升至 1:100，监管层正式采纳“舰队指挥官（Fleet Commander）”定义，标志着人机协同进入高维战略监控阶段。 - [51:01](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3061s) 长航程单次配送成本已从初期的 300 美元大幅压降至 12 美元，全口径摊薄成本全面击穿传统燃油车人工配送线，迎来自动化物流成本倒挂的重大拐点。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=6bGxm8gX41o&t=0s) 真实的“产品市场匹配（PMF）”始于客户对服务时长的刚需抱怨：卢旺达医生反问“为何生病不分昼夜而服务只开 12 小时”，促使 Zipline 第一年便升级为 24/7/365 全天候运行。 - [02:28](https://www.youtube.com/watch?v=6bGxm8gX41o&t=148s) 以出海破除监管死局：在美国严禁超视距（BVLOS）飞行的初期，Zipline 转向以生死救援为刚需的卢旺达，换取政府特批空域，在真实极限场景中跑通闭环。 - [04:09](https://www.youtube.com/watch?v=6bGxm8gX41o&t=249s) 硬件仅占整体解决方案的 15%，真正壁垒在于涵盖仓储冷链、民航空域直连、医疗物资处方流转与机队预测维护的 85% 幕后系统工程。 - [08:24](https://www.youtube.com/watch?v=6bGxm8gX41o&t=504s) 商业外交取代传统援助：凭借低成本与高可靠性，Zipline 与美国国务院达成 5.5 亿美元商业外交合作，推动发展中国家将未来基建建立在先进机器人与 AI 技术之上。 - [11:15](https://www.youtube.com/watch?v=6bGxm8gX41o&t=675s) 自研双飞控与微型仲裁器容错架构，即便在单粒子翻转、接触不良或主飞控突发故障时也能实现无缝接管，用低成本消费电子供应链实现航空级高安全冗余。 - [17:20](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1040s) 坚持破坏性测试哲学与极端天候飞行：通过 -25°C 至 49°C 极限温差考验，将系统安全目标设定为超越传统汽车 20 倍以上（即比 Waymo 再安全 2 倍）。 - [20:25](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1225s) 母机悬停 100 米高空避免了地面旋翼噪音和安全隐患，搭载独立英伟达 GPU 与自主推进器的投递舱在下降中自主识别障碍，实现分米级平稳投送。 - [26:57](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1617s) 人机监控比例跃迁至 1:100，监控人员正式转型为战略层面的“舰队指挥官”，系统通过云端大脑自主调度高密度机群实现全网运力自平衡。 - [28:22](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1702s) 迈向日均 100 万单使百万分之一的罕见故障沦为每日必现常态，倒逼物流系统向无固定基地部署、新机自飞出厂、故障机自飞返修的超高度自动化演进。 - [32:53](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1973s) 打破依赖 1950 年代人工纸笔的老旧空管体系，推动由机载算法自主广播与分布式协同避障的低空交通管制标准，以承载超民航数十倍的高密空域架次。 - [36:08](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2168s) 外购零部件的频繁炸机迫使团队走向全栈垂直自研，底层涵盖电机控制器、BMS、GNC 算法与空管软件，打造出能直接击败燃油车成本的汽车级可靠硬件。 - [41:11](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2471s) 严格践行第一性原理与部件删减原则，如将沉重的 1 米可展开机载机械尾钩砍至 2 厘米固定微钩，将机械抓捕复杂度转移至地面智能抓捕臂。 - [46:47](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2807s) 即时配送将迎来 10 倍暴增的“Uber 效应”：用 2 吨汽车送轻包裹难以为继，极速、廉价且免小费的无人机物流有望将全美即时配送规模推高至 550 亿单。 - [52:17](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3137s) 在全行业追逐轻资产 SaaS 的狂热时代坐了十年冷板凳，Zipline 坚守重实体制造与底层硬科技工程，证明了实体硬件对未来文明基础设施的不可替代性。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=6bGxm8gX41o&t=0s) 卢旺达前线的启示：生病不挑时间与产品市场匹配 创业初期的探索往往充满了戏剧性的反差。系统工程与安全负责人埃里克·沃森（Eric Watson）回忆起当年在卢旺达一线的经历：当时他还是一个满怀抱负、初出茅庐的年轻工程师，怀着对飞行器技术的极大热情，深入一线去走访他们所服务的医生和化验室技术人员。埃里克原本满心期待对方会大聊特聊无人机硬件多么惊艳、飞行控制多么先进，于是主动询问对方使用体验如何、有什么反馈。不料对方给出的核心反馈彻底颠覆了他的工程师视角：“人们生病是不分昼夜的，是一年 365 天、一天 24 小时随时会发生的，为什么你们一天只营业 12 个小时？” 这一反问让团队醍醐灌顶——尤其考虑到 Zipline 当时运送的是抢救生命的血液制品。尽管早期的无人机产品本身还谈不上完美、各种毛病层出不穷，但这种看似挑刺的反馈恰恰证明他们找到了极其强烈的“产品市场匹配”（Product-Market Fit, PMF）。客户最迫切的抱怨不是技术不好，而是现有的服务时间远远无法满足刚需，他们渴望获得更多、更持久的服务。这种来自滩头阵地市场的真实痛点，直接促使 Zipline 在第一年内便迅速将服务升级为全年无休的全天候运营。 ### [01:03](https://www.youtube.com/watch?v=6bGxm8gX41o&t=63s) 拒绝“无人机公司”标签：隐于幕后的即时传送网络 红杉资本（Sequoia Capital）的合伙人阿尔弗雷德·林（Alfred Lin）与帕特·格雷迪（Pat Grady）在开场时指出，Zipline 如今已经毫无争议地成长为全球最大的商业化自主飞行系统提供商。然而，联合创始人凯勒·里纳乌多·克利夫顿（Keller Rinaudo Cliffton）却明确表示，自己其实非常反感外界将 Zipline 简单定义为一家“无人机公司”。凯勒强调，Zipline 骨子里是一家以客户为中心的企业，而真实用户根本不在乎无人机本身。“我们所有的客户完全不在乎幕后运转的是什么技术”，他们的愿景从一开始就是为地球构建一套自动化物流系统，去逼近“瞬间传送”（Teleportation）的终极体验。 对于今天使用 Zipline 服务的普通消费者而言，最直观的体验不过是：在手机上下载并打开 App，挑选琳琅满目的品牌或心仪的餐厅美食，轻轻点击下单，5 分钟后货物便精准空投到眼前。技术应当彻底退居幕后，变成一种无感而神奇的可靠基础设施，而不是拿来炫耀的机械玩具。 ```mermaid flowchart LR accTitle: 用户视角的无感即时配送链路 accDescr: 展示从用户端下单到无人机幕后系统交付物资的精简链路 subgraph FrontEnd["用户端体验"] A["打开 App 下单<br/>(餐饮 / 医疗物资)"] end subgraph BackEnd["幕后自主系统 (85% 复杂度)"] B["自动化仓储与温控"] C["空域协调与自主航路"] D["双备份容错飞行器"] end subgraph Delivery["终端交付"] E["5分钟内精准降落/投送<br/>(近似瞬间传送)"] end A --> B B --> C C --> D D --> E ``` ### [02:28](https://www.youtube.com/watch?v=6bGxm8gX41o&t=148s) 逆风启航：规避监管死局的全球破局之路 回溯 Zipline 的诞生历程，其发展节点几乎完全走在现代人工智能（AI）、大语言模型（LLM）以及通用机器人风潮的最前端。公司最早成立于 2011 年，并在 2014 年完成关键业务转向。在那个硬件与机器人被资本市场视作“最愚蠢创业方向”的年代，23、24 岁的凯勒与刚加入团队的埃里克四处奔走融资，描绘着一个“速度提升 10 倍、成本降低一半、零碳排放”的全球自主物流愿景。然而，几乎所有硅谷投资人在听完推介后都会抛出同一个致命问题：“这在美国不是违法的吗？” 当时的美国监管法规极其严苛，甚至完全禁止超视距飞行（Beyond Visual Line of Sight, BVLOS），商业无人机根本没有合法升空的空间。面对“我们不投非法生意”的闭门羹，Zipline 被逼上了一条看似极其冒险、实则柳暗花明的出海之路：既然本土不合法，那就去世界上最需要这项技术的地方验证价值。2016 年，Zipline 选择在卢旺达正式启动运营，专注于直接向医院和基层医疗机构空运紧急输血用血。正因为这是一个事关生死的极限救援场景，其创造的社会价值如此巨大，以至于卢旺达政府愿意全力配合，甚至专门修改现有空域监管框架给予特批，让商业自主飞行真正合法落地。 ### [04:09](https://www.youtube.com/watch?v=6bGxm8gX41o&t=249s) 硬件仅占 15%：首年惨败砸碎的极客幻想 2016 年正式在卢旺达投入运营时，年轻的团队曾盲目笃信自己研发的无人机飞机是行业顶尖之作，甚至觉得只要造出酷炫的硬件就能万事大吉。然而现实却迎头痛击——起步阶段几乎演变成一场彻头彻尾的灾难。当时 Zipline 雄心勃勃地与当地政府签订了覆盖 21 家医院的供货合同，但在最初的长达 9 个月时间里，系统实际能够稳定服务的竟然只有区区 1 家医院。 埃里克回忆起那段日子，几乎处于长期的睡眠剥夺状态。由于卢旺达与美国之间存在巨大时差，驻守美国后方的团队常常在午夜被刺耳的警报电话叫醒——大洋彼岸的卢旺达配送中心刚一启动，系统往往全面瘫痪，什么都无法正常工作。团队陷入了无休止的通宵排障和连轴转的周末加班。这场惨痛教训让团队彻底清醒：在整体解决方案中，无人机硬件本身的复杂度仅仅占了 15%，剩下的 85% 是一座深不可测的技术与运营冰山。要实现全年无休、有如魔法般的可靠物流，Zipline 不得不从零研发整套辅助系统： - 极其复杂的仓储与冷链库存管理系统； - 与国家民航局（Civil Aviation Authority）直连的空域管控与飞行集成系统； - 与国家公共卫生体系深度嵌套的处方与物资下发链路； - 全自动化的预测性维护与机队健康诊断软件； - 高并发的实时订单调度与需求管理平台。 ### [06:49](https://www.youtube.com/watch?v=6bGxm8gX41o&t=409s) 1.4 亿英里与拯救生命：从单一据点到商业外交支柱 渡过早期的生死考验后，Zipline 迅速兑现了全天候运转的承诺，在第一年内就实现了 24/7/365 全天候无休运行。如今每逢圣诞节，凯勒依然保持着给全球所有配送中心逐一致电慰问的传统。从最初艰难支撑的 1 家医院，到 20 家、500 家，如今该网络已经覆盖全球 8 个国家的超过 5,000 家医院与卫生机构，成为当今地球上规模最大的商业自主运行系统。 为了让外界直观感受这一体量，埃里克打了一个生动的比方：Zipline 累积的商业自主飞行里程刚刚突破 1.4 亿英里（约合 2.25 亿公里），这相当于将全美境内的每一条公路反反复复完整开上超过 30 遍。伴随这一天文数字飞行里程而来的，是极其扎实的社会效益。宾夕法尼亚大学（University of Pennsylvania）近期发布的一项独立学术研究显示，在 Zipline 覆盖的服务区域内，因产后大出血等紧急情况导致的孕产妇死亡率骤降了 51%，挽救了整整一半原本可能逝去的母亲。根据合作伙伴的综合评估，系统目前每年直接挽救的生命在 10,000 到 12,000 人之间。 这一不可替代的生命线网络也促成了其国际地位的质变。随着美国国际开发署（USAID）部分职能的调整，美国政府急需寻找一种更高效、更具创新力的对外合作抓手。前不久的 12 月，Zipline 与美国国务院正式达成了一项高达 5.5 亿美元的重大商业外交合作伙伴关系，旨在借助其低成本、高可靠性的无人自主网络，向更多新兴市场国家规模化复制这一挽救生命的空中医疗动脉。 ### [08:24](https://www.youtube.com/watch?v=6bGxm8gX41o&t=504s) 商业外交与发展中国家的新基础设施 美国在国际发展策略上正在经历一场重大转变，被称为“商业外交”（commercial diplomacy）。过去，许多受援国只能依靠非政府组织（NGO）提供的低质量免费援助，但这些国家已经对这种模式感到厌倦。免费援助往往容易催生依赖性，反而阻碍了当地经济的自我发展。这些发展中国家真正渴望的是高薪工作、创业机会和先进技术。 因此，美国战略的核心在于推动整个发展中国家将未来建立在美国的 AI 和机器人技术之上。通过激励这些国家采用高科技基础设施，不仅能在当地拯救生命并节省资金，还能让美国在未来几十年的制造业和先进机器人领域持续巩固其全球领导地位。 - 传统免费援助容易造成受援国被动依赖，抑制本土经济增长。 - 发展中国家迫切需要的是高薪就业机会、技术转让与创业生态。 - 商业外交战略致力于将美国的 AI 与机器人技术打造为全球发展中国家的新型基础设施底座。 --- ### [09:35](https://www.youtube.com/watch?v=6bGxm8gX41o&t=575s) 真实物理世界带来的突发挑战：太阳天气与 GNSS 干扰 当被问及在真实世界中运营全球最大的自主系统近十年所积累的意外经验时，团队指出，普通工程师坐在电脑屏幕前根本无法预见现实环境的极端复杂性。许多人会以为最主要的飞行障碍是刮风下雨等常规气象，但实践中遇到过最奇特的问题之一竟然是“太阳天气”（solar weather）。 太阳耀斑爆发会释放高能辐射，直接影响地球的电离层。这种大气层扰动会导致来自 GPS 卫星的射频（RF）信号传播速度发生异常波动，从而引发导航系统的精度退化和定位挑战。在系统研发之初，团队从未设想过需要专门攻关天体物理带来的影响。为此，Zipline 在硬件和算法上进行了深度改造：一方面重新设计了导航与 GNSS 系统，使其即便在剧烈的太阳耀斑干扰下仍能保持厘米级的定位精度；另一方面在 GNSS 之外引入了多层互补冗余机制，确保在极端信号恶化时无人机依然能安全飞行。 ```mermaid flowchart TD accTitle: 太阳耀斑对无人机导航的影响及应对冗余 accDescr: 展示太阳耀斑干扰电离层信号后，Zipline的双重应对策略 A["太阳耀斑爆发 (Solar Flares)"] --> B["电离层扰动与辐射增强"] B --> C["GPS/GNSS 射频信号传输延迟或异常"] subgraph Defense ["Zipline 容错导航方案"] C --> D["高鲁棒性 GNSS/导航算法<br/>(保持厘米级定位精度)"] C --> E["非 GNSS 冗余导航系统<br/>(无信号或极度衰减时接管)"] end D --> F["持续稳定自主飞行"] E --> F ``` --- ### [11:15](https://www.youtube.com/watch?v=6bGxm8gX41o&t=675s) 双飞控系统与仲裁器容错架构 在系统安全领域，Zipline 坚守一条基本法则：“必须假设系统的任何部件都可能出现故障或短暂异常。”无论是软件偶发缺陷、连接器接触不良，还是宇宙辐射导致的单粒子翻转（bit flip），系统都必须具备自我恢复能力。 为了彻底消除单点故障，无人机搭载了双飞行计算机（Flight Computer）。两台飞控在飞行期间同时接收所有传感器数据并实时计算向执行机构（电机、控制舵面）下发的指令，彼此都认为自己处于飞行控制状态。系统引入了一个微型独立的“仲裁器”（Arbiter）计算机来监控两台飞控的健康指标，并负责向全机其他节点广播究竟听从哪一台的指令。当被追问“如果仲裁器自身坏了怎么办”时，团队解释了极其干脆的降级逻辑：一旦仲裁器失效，当时处于控制地位的主飞控将直接维持现状继续执行飞行任务。 就在访谈前几周，这套容错架构在实战中发挥了关键作用：一架无人机在向客户精准投递包裹后，主飞控突发故障，系统瞬间无缝切换到备用飞控，飞机自主安全返航降落，全程未发生任何事故。这种贯穿到底的高可靠性设计，正是支撑其累计飞行 1.4 亿英里、完成 250 万次配送且保持“零安全事故”的核心基石。 ```mermaid flowchart TD accTitle: Zipline 双飞控与仲裁器容错架构 accDescr: 两台飞行计算机同时接收传感器并计算输出，由仲裁器裁决控制权，并具备仲裁器自失效降级保护 subgraph Inputs ["感知层"] S["全机传感器集群"] end subgraph Compute ["计算与仲裁层"] FC1["主飞行计算机 (Primary)"] FC2["备用飞行计算机 (Secondary)"] ARB["仲裁器 (Arbiter 监控芯片)"] end subgraph Actuators ["执行层"] ACT["电机与舵面执行机构"] end S --> FC1 S --> FC2 FC1 -.->|"状态心跳"| ARB FC2 -.->|"状态心跳"| ARB ARB -->|"广播有效指令控制权"| ACT FC1 -->|"控制指令"| ACT FC2 -->|"控制指令"| ACT FC1 -.->|"仲裁器失效时<br/>主飞控默认锁定接管"| ACT ``` - **航空级安全与消费级成本**：波音 777 同样采用双飞控架构，但单套系统成本高达数百万美元。Zipline 的突破在于借鉴传统航天的安全标准，并利用智能手机供应链级零部件将其成本压缩至数十到数百美元，以 1/100 的成本实现了接近传统航天的安全性，研发迭代速度提升了 100 倍。 - **实战零事故记录**：累计自主飞行 1.4 亿英里，完成 250 万次商业交付，始终保持零安全事故。 --- ### [14:05](https://www.youtube.com/watch?v=6bGxm8gX41o&t=845s) 全栈自研垂直整合与极限硬件测试 无人机本体实际上仅占整体解决方案的 15%，其余 85% 涵盖庞大的地面、空域协调及系统工程网络。Zipline 并非纯软件公司，而是一家深植于物理世界的 AI 机器人企业。其硬件可靠性依赖极其严苛的组件级与集成系统级测试，其中包括高度加速寿命试验（HALT测试）：电动机等关键部件被置于暴雨、极热、高湿、腐蚀性盐雾及紫外线照射等极端恶劣环境中持续满负荷运转，以验证其耐受极限。 之所以采取这种近乎偏执的测试手段，是因为 Zipline 具备极高的垂直整合度。传统汽车或航天企业在被问及零件可靠性时，通常直接依赖供应商提供的数据报告，但 Zipline “自己就是供应商”。从底层硬件到整机，全机多达 700 种核心零部件均为团队自主从零设计，包括： - 飞行控制计算机 - 动力分配板与电机控制器 - 动力电池组及电池管理系统（BMS） - 配备英伟达（NVIDIA）GPU 飞控的微型投递舱（Pod，用于在悬停时垂直降下完成精准入户交付） - 自动机库对接站（Docking Station）的配套电子电气系统 - 定制高推重比无刷电机（市面现有商用电机均无法满足其严格的推重比指标） 全机包含 43 个核心大型子装配总成，全在自建产线上完成装配。整机组装后还会被送入振动台、风洞和高低温交变热舱中进行系统级极限测试，确保从底层硅片、执行器到上层机电系统均拥有端到端的绝对掌控力。 ### [16:33](https://www.youtube.com/watch?v=6bGxm8gX41o&t=993s) 破坏性测试哲学与极端温差考验 Eric Watson 深入探讨了 Zipline 的工程测试哲学：测试绝不仅是为了证明无人机“足够好”或完成测试清单，而是为了主动将设备推向极限，直至其发生故障。工程团队的核心诉求是彻底搞清楚“这套系统究竟会怎么坏掉”，从而判断损坏模式是在预料之内的可控范围，还是需要彻底返工改进。他们从不满足于“跑完整轮测试没出故障就万事大吉”，而是秉持“把东西测到彻底报废，看看极限究竟在哪里”的破坏性测试思维。 Alfred Lin 随后追问极限测试的具体指标，Eric 与 Keller Rinaudo Cliffton 列举了无人机所经受的严苛温度测试：高至 49°C（非常炎热），低至 -25°C（极度严寒）。当被质疑是否真有必要在 49°C 飞行时，Keller 指出，夏季的美国凤凰城（Phoenix）气温就能达到这一极端水平；而 -25°C 则对应美国北部的冬日严寒。 - **测试温度范围**：从 -25°C 的极寒环境到 49°C 的酷暑高温。 - **目标飞行地域**：夏季的美国西南部沙漠城市（如凤凰城）以及冬季的美国北部高纬度地区。 --- ### [17:20](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1040s) 绝不向恶劣天气妥协：始于救命使命的安全权衡 对话转向了一场深刻的哲学讨论：面对太热、太冷或暴雨等极端天气，为什么不干脆选择停飞以降低风险？Keller 解释道，Zipline 之所以没有采取这种妥协方案，根本原因在于其最初切入的核心业务是“拯救生命”。如果对客户承诺“只要艳阳高照，我们就能挽救你和挚爱的生命”，这在逻辑上根本站不住脚。在早期的运营中（尤其是与卢旺达民航局紧密合作期间），团队甘愿承担巨大的运营风险，因为只要有一条生命悬于一线，无人机就必须逆风起飞。 事实证明，这种冒险几乎每一次都成功挽救了患者的生命。在极少数系统不可控的情况下，无人机配备了作为最后安全冗余手段的副翼降落伞（Parawing），可将飞行器平稳、缓慢地带回地面。在创立之初，这种降落伞迫降相对常见，但随着工程演进，如今已变得极其罕见。 Keller 还透露了一个董事会上的趣闻：Zipline 最初设定的安全标准是“比传统汽车安全 10 倍”，但投资人 Alfred Lin 在最近一次董事会上直言不讳地提出批评，认为这简直是无稽之谈，因为汽车本身就是落后过时的技术标准；Alfred 提出，“真正的目标应该是比 Waymo 还要安全两倍”。如今，这已成为 Zipline 的核心目标。Waymo 的自动驾驶系统目前大约比人类驾驶安全 10 到 12 倍，这意味着 Zipline 的目标是将安全性提升到比普通汽车安全 20 倍以上的水准。 ```mermaid flowchart TD accTitle: Zipline 与地面交通安全基准演进对比 accDescr: 展示从传统人类驾驶到 Waymo 自动驾驶，再到 Zipline 设定的无人机安全目标的技术飞跃。 subgraph Ground["地面交通安全基准"] A["传统人类驾驶汽车<br>(行业基准：每 1.4 亿英里产生 600 起事故)"] B["Waymo 自动驾驶<br>(较人类安全约 10-12 倍)"] end subgraph Air["空中自主投递安全目标"] C["Zipline 早期目标<br>(比普通汽车安全 10 倍)"] D["Zipline 现行目标<br>(比 Waymo 再安全 2 倍 / 达到汽车安全的 20 倍以上)"] end A -->|"自动驾驶技术升级"| B A -.->|"最初设定"| C B -->|"Alfred Lin 敦促更高标准"| D ``` - **飞行里程与安全记录**：Zipline 累计已完成 1.4 亿英里（140M miles）商业自主飞行，保持“零安全事故（Zero Incidents）”。 - **地面交通事故对照数据**：若驾车行驶 1.4 亿英里，平均会发生约 600 起交通事故、导致 100 人受伤，并造成 2 至 6 人死亡。 - **耐用性里程碑**：通过破坏性测试打造的机身极为皮实，商业机队中已有单架无人机累计自主飞行里程超过了 100 万英里。 --- ### [20:25](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1225s) 100米高空悬停与自主投递舱（Delivery Pod）黑科技 主持人 Pat Grady 对 Zipline 惊人的投递精度表达了赞叹：无人机在高空悬停，即使面对强风暴雨，也能将包裹精准降落在一个半径仅 18 英寸的狭小圆圈内。Eric 纠正道，无人机母机并不是在 100 英尺高度，而是在更高、难度更大的 100 米（约 330 英尺）高空悬停。 为了实现分米甚至厘米级的落点精度，Zipline 采用了分层控制与实时环境感知的复合架构。母机利用实时动态差分定位技术（RTK GNSS）获得厘米级自身定位，并逆风调整机身偏移以抗衡空中风力。然而，室外环境如后院咖啡桌并没有预设的绝对 GPS 坐标。因此，母机在高空释放悬吊的自主投递舱（Delivery Pod），由投递舱搭载的机载计算系统在下降过程中独立完成末端感知与位姿微调。 投递舱配备了独立的英伟达（NVIDIA）GPU，运行专有的 AI 自主感知算法堆栈，即便在漆黑的夜晚也能精确探测地面状况。例如，如果桌上摆满了饮料杯，投递舱会自动识别并避开桌面，改将包裹稳妥放置在桌边的平整地面上。投递舱不仅能看懂环境，还能在 X 轴和 Y 轴方向自主施加推力以调整自身水平位置。当包裹平稳触地、舱门开启并放下货物后，投递舱迅速收拢并被母机绞盘收回，整机随即返航。 ```mermaid flowchart TD accTitle: Zipline 平台式母机与自主投递舱协同运作架构 accDescr: 展示母机高空抗风悬停与下方搭载独立 GPU 的投递舱自主识别地面的两层协作流程。 subgraph Mothership["空中母机 (100米高空悬停)"] M1["RTK GNSS 厘米级自身定位"] M2["顶风偏航调节<br>抵抗高空乱流"] M3["绞盘缆绳下放系统"] end subgraph DeliveryPod["自主投递舱 (Delivery Pod)"] P1["独立 NVIDIA GPU<br>自主运行 AI 感知堆栈"] P2["主动 X/Y 轴平面推进控制"] P3["实时障碍物识别<br>(避开杯子、杂物，识别最优落点)"] P4["平稳触地开门交付 -> 收起升空返回"] end M1 --> M2 --> M3 M3 -->|"缆绳下放"| P1 P1 --> P3 --> P2 --> P4 ``` --- ### [22:46](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1366s) 极致降噪与极致安全的双重飞跃 Keller 强调，将动力母机保留在 100 米高空、仅降下轻量化投递舱的设计带来了两大颠覆性优势： 1. **极致的声学体验（Acoustics）**：绝大多数无人机的高频旋翼噪音令人难以忍受。Zipline 组建了一支由空气动力学、气动声学和飞行控制专家组成的庞大研发团队，对机身每个构件进行气动声学优化，目标是让其噪音微弱得犹如“微风轻拂树叶”。此外，由于真正产生大推力噪音的母机停留在 100 米高空，地面上的人几乎听不到动静。 2. **极高的接触安全性（Human Safety）**：高速旋转的桨叶和重型机身绝不靠近地面人员。唯一接近用户、儿童和宠物的，只有那个从天而降的“外形呆萌、完全拟人化且非常安全的保丽龙（Styrofoam）投递小舱”。 在进入美国达拉斯（Dallas）等本土城市展开常规外卖与包裹配送之前，Zipline 自 2016 年在卢旺达启动首个商业运营项目以来，已在海外深耕测试了大约 6 至 8 年时间，积累了无与伦比的现实世界飞行数据与工程鲁棒性。 ### [24:37](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1477s) 进军美国监管与从单一航机到网络基础设施的跃迁 新一代的到家配送服务（Home Delivery Service）目前正处于令人惊叹的超大规模扩张阶段，但这套系统其实在去年一月才正式启动。从 2014 年起步算起，这项技术的打磨实际上耗费了近九年的光阴。谈及进入美国市场是否一帆风顺时，团队坦言，他们与美国联邦航空管理局（FAA）及其他监管机构进行实质性深入沟通始于 2020 年左右，绝非到了 2025 年就能直接平稳落地。回顾 2016 年在卢旺达起步时，这类无人机配送在法律上甚至没有合规路径、处于“某种违法”的灰色地带。进入美国后，双方建立起一种紧密的伙伴关系，共同聚焦于一个目标：“安全且高效的空域融合（Safe and efficient airspace integration）”。Zipline 带着在多个国家成功运营的丰富实战经验与技术理念，与监管机构共同探讨未来的监管范式，历时数年逐步磨合统一认识，最终跑通了合规路径。 要管理如此庞大的飞行机队，Zipline 必须开发一整套全闭环的软件与控制系统。正如 Keller 所强调的，飞行器硬件本身仅占整个系统复杂度的 15%，Zipline 真正搭建的是一套底层“基础设施层”，以实现对商品的即时访问。这绝非单纯靠单架飞机从 A 点飞到 B 点，而是依赖覆盖区域的充电机巢网络、数百架分布式协同的飞机，以及云端自主调度系统。云端大脑必须实时洞察需求潮汐与运力分布，精准预判即将到来的晚餐高峰，权衡各区域风力与气象条件，自主实现全网运力自平衡。虽然航机全程自主飞行、无需人类介入，但系统仍会在遇到恶劣突发大风或异常时智能警示人类介入。 ```mermaid flowchart TD accTitle: Zipline 自动化云端基础设施与运力调度网络 accDescr: 展示 Zipline 如何通过云端大脑连接需求、充电网络、自主机群与人类舰队指挥官 subgraph Inputs["实时感知与输入层"] A["订单需求预测<br/>(如晚餐高峰潮汐)"] B["微气象监测<br/>(风速/恶劣天气预警)"] C["全网充电站与机巢状态"] end subgraph CloudBrain["云端自主调度中枢 (基础设施层)"] D["全局运力自平衡算法"] E["航线动态规划与避障"] F["异常感知与安全升级触发器"] end subgraph Operation["执行与监管层"] G["自主无人机群 (100+ 巡航中)"] H["舰队指挥官 (1:100 战略监控)"] end Inputs --> CloudBrain D --> G E --> G F -->|遇异常告警| H H -.->|战略干预| CloudBrain ``` ### [26:57](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1617s) 从“飞手”到“舰队指挥官”：1:100 的人机协同飞跃 在最初进入美国领空运营时，监管层给出的许可非常严苛，要求必须实行“1 对 1”的人员配比，即一名被称作远程当班机长（Remote Pilot in Command）的专业人员坐在办公室屏幕前，专门监视单单一架自主飞行的飞机。Keller 笑着回忆，尽管人类实际干预并发送指令的概率极低极低，但这种“一人看一机”的模式对单笔经济模型（Unit Economics）来说简直是灾难。随着 Zipline 在运营中反复证明系统的极致可靠性，监管比例迅速被改写并破冰：从 1:1 走向 1:3、1:6、1:20、1:40，如今已经达到一名操作员监控 100 架飞机的规模，未来还将迈向更高数量级。 伴随这一规模跃迁，Zipline 决定彻底弃用“飞手（Pilot）”这个容易引起误解的词汇。受科幻小说《安德的游戏》（Ender's Game）的启发，他们将这支专业监控团队正式命名为“舰队指挥官（Fleet Commanders）”。这一术语甚至被正式写入了 FAA 的官方适航监管文档中，明文规定“舰队指挥官应当执行如下规程”。这展现了技术赋能人类的本质转变：机器人接管了底层的琐碎操作与重体力劳动，而人类则被解放出来，站在更高的战略层级上指挥和管理庞大的机器人机群。 - **人机监控比例跃迁轨迹**：1:1 $\rightarrow$ 1:3 $\rightarrow$ 1:6 $\rightarrow$ 1:20 $\rightarrow$ 1:40 $\rightarrow$ 1:100（当前已落地标准）。 - **行业术语更迭**：摒弃传统“Remote Pilot”，确立官方监管认可的“Fleet Commander”职位定义。 - **人机关系定位**：操作员不再“操控飞行”，而是转型为高维度的系统战略运维与异常监督者。 ### [28:22](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1702s) 奔向单日百万单：重构制造、维护与全天候自主法则 谈及超大规模扩张所带来的全新挑战，团队坦言整个系统的要求发生了质的蜕变。从 2014 年起步研发，Zipline 耗费了整整十年时间（直到 2024 年末至 2025 年初），才在公司历史中累积完成了前 100 万次递送。然而在不久的将来，Zipline 将迎来“单日完成 100 万次递送”的拐点。这是一个令人敬畏的体量变化：过去所谓的“百万分之一小概率故障”，在单日百万单的现实下，将变成“每天必然发生一次的常态”。原有的解决模式将被彻底打破，系统的严苛度被推升到全新维度，维护成本、机队故障容错和流程效率都必须进行根本性重构。 自主系统展现出了传统物流难以企及的优势：它们天然渴望且能够做到 7x24 小时全天候运转。人类卡车物流受制于工作时长和排班，而自主机群既可以在下午 2 点准时送餐，也能在凌晨 2 点或 3 点悄无声息地将包裹投递在后院或门廊，等待用户早晨 6 点起床后取用，且能实现 5 分钟极速送达。未来的运维流程也被完全自主化：飞机即将实现直接从南旧金山（South San Francisco）的制造工厂起飞，自主飞入商业运营网络，类似特斯拉汽车自主驶出工厂交付车主。一旦飞机需要预防性维护，它会自动飞向维护枢纽（Depot），维修人员快速处理后，飞机再度自主起飞返岗。在大都市圈内，飞机没有固定的“家”，而是完全由系统根据需求动态调配。Eric Watson 指出，这正是埃隆·马斯克所说的“制造机器的机器”，Zipline 正在重点投资能够自动化制造与运转整套系统的底层软硬件。 - **交付体量指数跃升**：从“耗时 10 年累积 100 万次”转向“单日递送 100 万次”。 - **极速响应与全时服务**：支持 5 分钟极速达，24/7 全天候无休运转，重构隔夜及黎明即时配送。 - **自主巡航闭环**：新出厂飞机直接从工厂飞入商业航网；遇维保需求自飞回基地，维修完毕自飞回前线。 - **无固定基地部署**：全城运力动态打散与重组，航机无固定停泊节点，完全随需求潮汐流动。 ### [31:52](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1912s) 颠覆空中交通管制：比全美航司大 80 倍的空域革命 在宏观航空运输尺度上，Zipline 正以惊人的速度重塑空域格局。目前全美最大的商业客运航空公司每天大约执行 5,000 个架次航班，而 Zipline 在下个月的单日飞行架次就将超越这一历史记录。当单日配送量迈向 100 万单大关时，Zipline 在美国商业空域每天运行的飞行班次，将达到全美所有民航客货运航司总和的 40 倍到 80 倍之多。 尽管无人机与传统民航客机属于完全不同的航空器级别，所解决的也是截然不同的问题，但对空中交通管制（ATC）而言，空域划分与安全性要求是一致的。这意味着整个空中交通管制体系必须迎来一场深刻的大变革。Keller 指出，大众往往对地面电动车和自动驾驶汽车充满热情，但地面始终受限于狭窄、古老且拥挤的公路路网；随着交通革命深入，未来空中的自动驾驶载具数量将是地面车辆的 10 倍以上。天空是一片极其广袤的三维立体空间，一个由无人自治系统驱动的空域新时代正在全面到来。 ### [32:53](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1973s) 重构空中交通管制：告别1950年代的纸笔系统 如果人类能够更高效地利用低空空域，就能将地面空间归还给人类——让居民社区更安静、更安全，减少碳排放与道路拥堵。然而，要实现这一愿景，美国乃至全球的空中交通管制（ATC）逻辑必须经历根本性的变革。现行的空管系统成型于1950年代，本质上依赖纸笔、索引卡以及人类调度员用肉眼紧盯窗外或雷达屏幕的落后方式；而面对未来海量自主飞行器的涌入，空管系统必须从底层以 AI 与全自主运行为核心进行全新设计。 这一转变无法靠单一企业单打独斗，必须依靠全行业的协同创新。Zipline 认为未来空域管理的核心逻辑是“飞行器之间的直接对齐与沟通”——无人机在空域中自主广播自身位置，遇到航线冲突时由机载算法自主做出避让决策（例如“你向上爬升，我向下避让”）。目前 Zipline 的飞行器已经具备这种分布式协同规避能力，并正在与其他进入空域的自主无人机企业联手打通协议，推动系统互联互通。同时，Zipline 也在与民航监管机构及标准组织紧密合作，将这些创新实践沉淀为通用行业标准。 在开拓国际市场时，由于多国缺乏适用于高频自主无人机的现成监管平台，Zipline 不得不从零编写了一整套空管与监控软件交付给当地民航局使用。美国本土有数家市值超百亿美元的上市公司专门开发空管软件，而 Zipline 内部为了跑通业务闭环，实际上相当于“在公司内部被迫孵化了一家百亿级的空管软件公司”。当前的传统空管体系正面临极为严峻的人力与安全危机：华盛顿特区发生过军机撞击直升机的险情，机场跑道与滑行道上的飞机剐蹭也屡见不鲜。其背后症结在于严重的劳动力断层——全美50%的空中交通管制员年龄超过45岁，20%即将退休，且几乎没有新鲜血液愿意进入这一行业。劳动力枯竭叠加大规模自主飞行器入网的现实，倒逼整个空域管理体系必须加速拥抱 AI 与自动化。 - **空管现状危机**：全美空管人员超半数在45岁以上，五分之一面临退休，行业面临巨大的人才断档。 - **协同规避机制**：低空飞行器必须支持 V2V（机对机）自主通信与冲突自主检测化解，而非事事依赖地面人工调度。 - **内部隐藏的“独角兽”业务**：Zipline 独立研发的民航空管监控软件，其对标的传统公开市场软件公司市值通常超过100亿美元。 ```mermaid flowchart TD accTitle: 空中交通管制模式演进对比 accDescr: 对比传统1950年代依赖人工纸笔的中心化空管与面向未来的AI分布式自主避障空管模式 subgraph Legacy["传统 1950s 空管模式"] L1["人工管制员 (面临老龄化与用工荒)"] --> L2["纸质便签与地面雷达监控"] L2 --> L3["单向无线电语音指挥"] L3 --> L4["容量极低 / 极易出现人为疏漏"] end subgraph Modern["自主空域协同模式"] M1["无人机机群自主广播位姿"] <--> M2["机载算法自动解算空域冲突"] M2 --> M3["分布式动态避让 (如一机爬升/一机下降)"] M3 --> M4["海量无人机高密安全并发"] end ``` --- ### [36:08](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2168s) 被迫走向极致垂直整合：从踩坑外包到自研底层 从硬件零件制造、整机组装、核心飞控代码，到日常仓储配送与航线运营，Zipline 展现出了极其罕见的全栈高度垂直整合结构。面对投资人关于为何不直接采购市面上成熟软硬件零部件的提问，联合创始人 Keller 坦言，这绝不是一开始就蓄意选择的“宏大蓝图”，而是在现实撞墙后的无奈抉择。他办公桌上方挂着一面自嘲的标语：“我们选择做这件事不是因为它是容易的，而是因为我们当初以为它很容易”（"We do this not because it is easy but because we thought that it would be easy"）。 几乎所有顶尖硬件企业的演进史都经历过试图走捷径的阶段。就像特斯拉早期制造初代 Roadster 时，曾寄希望于使用莲花（Lotus）Elise 的现成底盘外购电池组进行组装，最终却导致车辆成本奇高且故障频发，从而倒逼出 Model S 的全自研路线。Zipline 最初同样尝试向各种供应商采购货架产品（COTS）或付费外包设计，但随之而来的是测试场上接二连三的惨烈炸机。市面上现存的无人机硬件存在严重的两极分化：要么是大疆制造的消费级塑料多旋翼，其可靠性无法满足商业物流生命周期的苛刻要求；要么是美军耗资2000万美元、时速仅100英里的捕食者级别军用无人机。在民用高频物流领域，没有任何现成设备的安全性、可靠性与单次边际成本能够直接对标并击败在地面上跑了上百年的汽车与摩托车。 为了打破这种“外购件昂贵却极其不可靠”的死循环，团队在极度绝望中开启了一场彻底的置换手术：先是拔掉不可靠的现成配件，从零自研电机控制器；接着推倒现成模块，从零自研 GPS 与导航制导系统（GNC）。这种被迫的全栈自研，最终让 Zipline 打造出一套完全贴合其业务需求、具备“汽车级可靠性”（automotive-grade）且具备极佳成本竞争力的整套自主飞行闭环系统。 - **早期外包阵痛**：外购零部件和外包研发导致测试场频繁坠机，成本高昂且故障难以溯源排查。 - **跨界竞争基准**：无人机配送的真正竞争对手并非其他无人机，而是经历了上百年工业打磨、极度廉价且成熟的汽车与摩托车运力。 - **底层自研范围**：电机控制器、GPS 模块、核心导航定位系统、飞行控制算法全链路自制。 --- ### [39:07](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2347s) “沙丁鱼罐头”式协同：物理距离与零壁垒研发 全栈自研带来的最核心壁垒在于研发迭代的敏捷速度与系统级集成的紧密性。在 Zipline 的研发总部，所有的研发力量被高度压缩在极具密度的物理空间内：固件工程师、机械工程师、自主飞控工程师、云端基础设施团队、气动声学专家、制导导航与控制（GNC）专家、系统工程人员以及制造团队全部肩并肩挤在一个像“沙丁鱼罐头”一样紧凑的办公空间里。 这种极致的物理贴近构建起了无法被传统航企复制的反馈闭环。Zipline 的制造工厂距离总部研发大楼仅有 3 分钟车程，而试飞测试场也仅有一小段车程。工程师在工位上写完代码或完成硬件改版，转头就能去产线看装配，驱车几分钟就能到试飞场实地观察飞行表现。一旦系统出现故障，根本不存在传统航空巨头那般冗长且充满政治阻力的推诿链路。如 Keller 所对比的，波音 787 客机如果出现电池故障，各级供应链与主机厂之间往往需要通过打官司、扯皮两年才能查明到底是谁的责任，技术细节被层层采购合同和外包黑盒所掩盖；而在 Zipline，当飞行器发生意外时，系统工程师可以直接走到责任工程师的桌前说：“今晚你和我一起通宵把这个问题彻底干掉。” 系统工程与安全负责人 Eric Watson 进一步阐述，让背景截然不同的跨学科人才紧密合作其实非常简单，其前提就是打破部门墙，以真实的问题为导向。因为没有外部供应商所谓的“代码是专有知识产权（IP），不能提供源代码”这一类黑盒壁垒，任何工程师都能完全看清最底层的故障诊断逻辑。所有团队坐在同一张桌子上，朝着统一度量衡的目标并肩排障，从而在底层工程上彻底抹平了跨学科协作的沟通鸿沟。 ### [41:11](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2471s) 跨职能敏捷协作与第一性原理：质疑一切需求并勇于删减部件 在解决高度复杂的硬件与软件系统问题时，Eric Watson 指出最有效的方式就是毫无阻碍的扁平沟通。当某项测试未通过时，工程师直接走到相关人员的工位旁拉出代码：“‘测试失败了，跟我讲讲这部分系统是怎么运作的。’‘你做了这个假设？我设计的时候可不是这么想的。’走，我们一起查到底。”当制造团队、运营团队与工程团队紧密无缝地聚在一起时，看似棘手的工程难题攻关就会变得极其迅速高效。 Keller Rinaudo Cliffton 补充道，在打造融合了 AI、机器人技术、硬件和软件的复合系统时，必须领会一条近乎宗教般深刻的真理——“既定需求通常都很愚蠢”。Zipline 践行了 SpaceX 著名的“马斯克五步工作法”（Elon's Algorithm）首要法则：“质疑每一项需求”（Question every requirement）。无论是哪个部门提出来的指标，只要向下深挖两层，往往就会暴露出其不合理之处。伴随质疑需求的另一大原则是“果断删减部件”（Delete the part）。Keller 直言，飞机上最可靠的零部件，就是根本不在飞机上的零部件，因为它在上一代设计中就已经被彻底删除了，一个不存在的零件永远不会发生故障。就像 SpaceX 的猛禽发动机从 Raptor 1、Raptor 2 到 Raptor 3 一路极致精简管线一样，参观 Zipline 工厂的人也能从初代平台的 10 个硬件迭代版本，再到 EV1、EV2 直至最新的 EV3 飞机上，看到这种大刀阔斧的“删除”哲学。然而，删减零件是一项巨大的心理挑战和“勇气的表现”，因为一旦系统因缺少某零件而失效，删除它的人就会显得像个傻瓜；唯有对物理底层逻辑和系统极限性能拥有绝对自信，以及坚持全栈一体化垂直自研，团队才敢将多余的结构彻底剔除。 Eric 以初代平台（Platform 1）的回收机制为例分享了一个经典故事：早期的飞机降落依托一根长达 1 米的可展开式尾钩，飞机进场时放下长钩钩住回收线完成减速，结构沉重且机械构件极其复杂。工程团队萌生了一个基于第一性原理的设想——为什么不把复杂性转移到地面系统，让地面回收装置像航母甲板一样，用带驱动机构的机械手臂主动向上捕获飞机？这样机身就能变得极其轻盈且故障率大幅降低。然而在下一代飞机即将投产前夕，地面主动捕捉系统其实还没调试成功。面对下周就要开工造飞机的节点，团队围坐在一间会议室里孤注一掷：“是继续装 1 米长的尾钩，还是直接把它砍掉、换成机尾只有 2 厘米的固定微型钩，赌我们能搞定地面系统？”最终大家一致拍板：“删掉它！”随后的几周里，包括 Eric 在内的全员疯狂熬夜加班赶工，而在首批新飞机飞回的那天，地面机械手臂精准完成了空中捕捉与着陆。 ```mermaid flowchart TD accTitle: 复杂性转移架构演进 accDescr: 展示 Zipline 飞机回收系统如何将机械复杂度由机载部件转移至地面智能机器人系统。 subgraph Early["早期方案（复杂机载机构）"] A["1米长可展开机械尾钩"] --> B["机构沉重 / 易故障 / 增加机载能耗"] end subgraph Modern["优化后方案（第一性原理简化）"] C["2厘米极简固定式微型挂钩"] D["地面智能机械抓捕臂"] C --> E["机身极致轻量化与高可靠性"] D --> E end Early -.-> Modern ``` - **第一性原理与部件删减**：“飞机上最可靠的部件，就是被你从设计中彻底删掉的那个零件，因为它永远不会坏。” - **尾钩迭代数据**：初代飞机原采用 1 米长（约合 3.3 英尺）可折叠投放尾钩，在激进的技术路线重塑中被砍至仅有 2 厘米长的固定钩，将复杂机构转移至地面机器人。 - **全栈整合的优势**：唯有掌握从软件、飞控、机体制造到地面回收站的端到端全栈控制权，团队才有底气去跨层级打破需求并删减硬件。 --- ### [44:45](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2685s) 运送人类的畅想与从 100 万到爆发式增长的市场奇点 面对投资人兼董事会成员半开玩笑的提问——“既然这套系统安全、可靠、饱经实战检验，未来是否会从运送救命药品和芝士汉堡走向运送人类、彻底终结地表街道的拥堵暴政”，Keller 展现出了极具节制的战略清醒。他深知一旦松口可能就会引发“再向投资人募集 10 亿美元”的扩张冲动。Keller 坦言，单是将无人机配送网络打造成如同互联网传输数据般高效的“全球下一代物理基础设施”，就足以诞生一家全球规模最大的超级企业。面对这一万亿美元级的物流蓝海，团队必须对未来执行层面的极其残酷的难度保持“谦逊与偏执（humble and paranoid）”。 随后，Keller 披露了一个令人震惊的需求拐点：不久前 Zipline 设定的远景目标是实现“全网日均配送 100 万单”，并以此为基准规划了一座年产 20,000 架飞机的制造工厂。但在过去的短短几个月里，市场格局发生了翻天覆地的剧变——已经有多家大型合作伙伴各自要求直接向 Zipline 采购“每天 100 万单”的运力额度。“我们原先的整体经营规划指标，如今竟然变成了单一客户的采购计量单位（Our operating plan has now become our unit of sale）。”这种市场需求的爆发性跃迁，彻底打破了现有的产能模型，也意味着背后维系日均数十万乃至百万次飞行所必需的调度算法、空域避障及软硬件全栈系统，正在承受指数级上升的超高复杂度考验。 --- ### [46:47](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2807s) 550 亿次即时配送的未来：Uber 效应与全自动空中物流的必然性 探讨即时物流赛道时，Keller 揭示了现有地面配送体系的荒谬现状：全美每年由人工完成的即时即刻配送订单高达 55 亿单，而人类完成这些几磅重包裹配送所采用的工具，竟然是一辆重达 4,000 磅（近 2 吨）的燃油内燃机汽车。这种模式不仅耗时长达 30 分钟、45 分钟甚至一小时，伴随着巨大的碳排放和交通事故隐患，甚至在调查中，有超过 50% 的外卖骑手承认过去一个月内曾偷吃过客户的食物。 更具颠覆性的是市场潜在体量的重新估算。Zipline 在达拉斯（Dallas）都市区实现大规模常态化运营后，正计划在未来 4 个月内进军全美另外 4 大都市圈。通过深度剖析达拉斯消费者的真实复购行为，Keller 发现：如果将达拉斯用户展现出的高频即时配送习惯推演至全美，美国的即时配送真实需求根本不是 55 亿单，而是会直接暴增 10 倍，达到惊人的 **550 亿单**。 这种爆炸式增长与当年 Uber 颠覆出租车市场的逻辑如出一辙：在 Uber 创立初期，很多分析师认为就算它抢下旧金山 33% 的出租车份额，也不过是一家市值 15 亿美元的小公司；但他们忽视了一旦某种服务变得更加便捷、低廉且体验卓越，它将彻底重塑并无限扩张整个基本盘——今天的 Uber 已经是当年出租车市场总盘子的 10 倍以上。在实际运营中，Zipline 用户的习惯迅速从每月点一两次变成了日常每天点单： - Keller 提到，他最近遇到一位住在养老社区的 80 岁老奶奶，在过去一年内通过 Zipline 下单了 350 次。 - 老年公寓和退休中心正在成为无人机配送的巨大需求中心；老人完全不像外界刻板印象中那样排斥新技术，他们整天使用 iPhone，极其熟练地使用 Face ID 与 Apple Pay 完成下单并等待物资精准降落。 面对 550 亿单的未来需求，Keller 强调，全美既没有足够的人手去充当外卖司机，地表道路也不可能承受让汽车数量翻倍带来的致命拥堵。唯一的出路就是通过低空全自动无人飞行器将物流转移至空中，从而把道路还给人类的正常生活与城市繁荣。目前 Zipline 的全网每日自主飞行量已逼近 **5,000 架次**，并正以一条陡峭的增长曲线向着年底的目标高速冲刺。 ### [49:33](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2973s) 从日均三万次到百万次规模的飞跃 Zipline 目前的日常飞行量已达到每天约 30,000 架次，而团队眼前的首要目标，是以极快速度向日均 100 万次自主飞行的体量冲刺。Keller Rinaudo Cliffton 指出，公司管理层目前正在规划与执行的每一项供应链决策、产能爬坡计划以及制造工艺改进，不仅是为了支撑每日 100 万次配送的目标，更是为了在穿透这个里程碑后继续加速扩张。 - **规模化目标**：从当前的日均 30,000 架次自主飞行，全力推进至日均 100 万架次甚至更高量级。 - **底层驱动**：现阶段所有产能规划、全球供应链搭建与制造架构均围绕指数级放量的承载力展开。 --- ### [49:52](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2992s) 硬件公司的残酷现实与惊人的降本曲线 在探讨硬件创业与单位经济模型（Unit Economics）时，Keller 分享了一条发自肺腑的建议。作为早期天使投资人，他经常遇到意气风发的硬件创业者宣称“这个系统的制造成本只需要 X”。Keller 的标准回应往往是：“做好心理准备，当你真正做出来时，它一定会花掉 10X。” 这并非泼冷水，而是基于 Zipline 踩过的血泪教训。 2016 年 Zipline 在卢旺达签署第一份商业合同时，约定按每次配送 30 美元的价格提供 80 至 100 英里的血液运输服务，这在当时被认为与传统急救运输成本相当，团队也天真地以为成本就在 30 美元左右。然而系统上线交付的那一刻，实际单次运营成本高达 300 美元！面对整整十倍的成本亏损，红杉投资人 Alfred Lin 表现得“出奇地冷静（surprisingly chill）”。正是从那一刻起，团队开启了长达数年、死磕工程细节的降本攻坚战：成本从最初的 300 美元逐年压低至 120 美元、75 美元、40 美元、28 美元、18 美元，直至如今海外长航程技术实现了单次仅 12 美元的水平。 ```mermaid flowchart LR accTitle: Zipline 交付单次成本下降曲线 accDescr: 展现从 2016 年到现今海外长航程系统的成本骤降路径，从 300 美元压降至 12 美元。 subgraph 初期危机["初期亏损阶段"] A["2016年商业上线<br/>实际单次成本: $300"] --> B["阶段攻坚<br/>成本: $120"] end subgraph 规模化降本["工程优化与规模效应"] B --> C["优化至 $75"] C --> D["优化至 $40"] D --> E["优化至 $28"] end subgraph 拐点降临["击穿汽车成本"] E --> F["降至 $18"] F --> G["现阶段长航程<br/>降至 $12 / 次"] end ``` - **给硬件创始人的忠告**：若预估成本为 X，务必提前按 10X 的成本压力推演商业模式能否存活。 - **卢旺达初期落差**：合同签约价为 30 美元/次，实际初期营运成本高达 300 美元/次。 - **降本路径**：$300 → $120 → $75 → $40 → $28 → $18 → $12。 --- ### [51:01](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3061s) 历史性拐点：无人机成本全面击穿传统燃油车 Keller 指出，一场悄无声息但意义深远的革命正在当下发生：Zipline 系统的全口径摊薄单位经济成本（fully burdened unit economics）正全面降至传统汽车配送成本之下。虽然外界和主流媒体尚未大肆报道，但在物流领域用机器人替代人工配送已具备更优的成本效益，这个拐点将在未来几个月内重塑人们的日常生活方式。 这种转变对社会具有多维度的红利：不仅能大幅削减环境排放，使社区告别车流拥堵、噪音与安全隐患，更能让终端用户享受更加快速、确定性高且便宜的服务。更有趣的一点在于消费者的真实反馈——在所有好评中，呼声最高的一条竟然是“没有小费（no tip!!）”。消费者极度喜爱这种免除心理负罪感、费用清晰透明的确定性服务体验。 - **成本倒挂拐点**：自主飞行物流的全成本正式跌破地面汽车派送的成本线。 - **多方系统性共赢**：减少交通事故与碳排放、降低社区噪音、大幅压缩端到端配送时长。 - **用户第一痛点反馈**：全流程无人化配送免去了繁琐的小费文化，让账单明晰透明。 --- ### [52:17](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3137s) 坐了十年冷板凳的“黑暗骑士”与实体制造的回归 Alfred Lin 调侃原以为 Keller 总结的经验会是“耗费的时间远超想象”，Keller 坦言“时间跨度确实也漫长得多”。他特别提到了红杉合伙人 Shaun Maguire 撰写的一份极具分量的内部备忘录。该备忘录深刻剖析了为什么硬件制造是推动人类未来几十年发展最具影响力的赛道、为什么硬件项目极难起步与融资，以及风险投资机构究竟该如何下注。 回顾过去十年，Zipline 几乎一直扮演着创投圈的“害群之马（black sheep）”——在全行业追逐高毛利纯软件 SaaS、移动端 App 与轻资产互联网红利的狂热年代，重资产硬件与机器人公司饱受冷落。Keller 引用了电影《蝙蝠侠：黑暗骑士崛起》中反派贝恩（Bane）的经典台词来打趣：“‘你们只是借用了黑暗，而我生于黑暗之中（You adopted the darkness. I was born in it.）’。我们在造机器人这件事还不酷的时候，就已经在泥潭里硬生生干了十年。” 在对谈尾声，Keller 发出了对重建实体基础设施能力的呼吁。他以近期旧金山金门大桥加装防自杀防护网为例：仅仅安装这张防护网的耗资，竟然超过了祖辈当年建造整座金门大桥的总成本。现代社会依然极度依赖祖辈遗留下来的老化基础设施，如果希望给子孙后代留下一个科幻小说中所描绘的光明未来，社会各界必须重拾对制造业、复杂供应链与实体硬科技的敬畏与耐受力。 > “在造机器人这件事成为潮流之前，我们已经在黑暗中硬生生干了整整十年。” > —— Keller Rinaudo Cliffton - **红杉内部备忘录**：深刻论证了重型实体硬件对人类文明演进的必要性，以及资本支持硬件攻坚的底层逻辑。 - **金门大桥的鲜明反差**：新建防自杀防护网工程的总耗资，居然超越了当年建设整座金门大桥的历史成本。 - **呼唤科幻未来的勇气**：重拾实体工程制造能力与复杂供应链组织能力，坚信并动手构建曾被许诺过的科幻未来。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=6bGxm8gX41o&t=0s) 人们生病是不分昼夜的，是一年 365 天、一天 24 小时随时会发生的，为什么你们一天只营业 12 个小时？ — 卢旺达医生与化验室技术人员 > [01:03](https://www.youtube.com/watch?v=6bGxm8gX41o&t=63s) 我们所有的客户完全不在乎幕后运转的是什么技术 — Keller Rinaudo Cliffton > [11:15](https://www.youtube.com/watch?v=6bGxm8gX41o&t=675s) 必须假设系统的任何部件都可能出现故障或短暂异常。 — Eric Watson > [17:20](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1040s) 真正的目标应该是比 Waymo 还要安全两倍 — Alfred Lin > [24:37](https://www.youtube.com/watch?v=6bGxm8gX41o&t=1477s) 安全且高效的空域融合 — Keller Rinaudo Cliffton > [36:08](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2168s) We do this not because it is easy but because we thought that it would be easy — Keller Rinaudo Cliffton > [39:07](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2347s) 今晚你和我一起通宵把这个问题彻底干掉。 — Eric Watson > [41:11](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2471s) 飞机上最可靠的零部件，就是根本不在飞机上的零部件，因为它在上一代设计中就已经被彻底删除了，一个不存在的零件永远不会发生故障。 — Keller Rinaudo Cliffton > [44:45](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2685s) Our operating plan has now become our unit of sale — Keller Rinaudo Cliffton > [49:52](https://www.youtube.com/watch?v=6bGxm8gX41o&t=2992s) 做好心理准备，当你真正做出来时，它一定会花掉 10X。 — Keller Rinaudo Cliffton > [52:17](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3137s) You adopted the darkness. I was born in it. — Keller Rinaudo Cliffton > [52:17](https://www.youtube.com/watch?v=6bGxm8gX41o&t=3137s) 在造机器人这件事成为潮流之前，我们已经在黑暗中硬生生干了整整十年。 — Keller Rinaudo Cliffton ## Mentioned Zipline, Sequoia Capital, Keller Rinaudo Cliffton, Eric Watson, Alfred Lin, Pat Grady, Shaun Maguire, Elon Musk, SpaceX, Tesla, Waymo, NVIDIA, FAA, USAID, 美国国务院, 宾夕法尼亚大学, 卢旺达民航局, Uber, Lotus Elise, Boeing 777, Boeing 787, 安德的游戏 (Ender's Game), 蝙蝠侠：黑暗骑士崛起 (The Dark Knight Rises), 金门大桥, Platform 1, Delivery Pod, EV1, EV2, EV3, Raptor ## Video description The largest commercial autonomous system on earth isn't a robotaxi fleet — it's Zipline, which has flown 140 million autonomous miles with zero safety incidents. Co-founder Keller Rinaudo Cliffton and Eric Watson, who leads systems engineering and safety, explain why the drone itself is only 15% of the solution. The rest spans inventory management, air traffic integration, and engineering systems such as a dual flight computer failover protocol that recently saved a delivery mid-flight. They trace Zipline's path from launching blood delivery in Rwanda in 2016 (when drone delivery was illegal in the US) to a 51% reduction in maternal mortality in that country, a $550 million commercial diplomacy partnership with the State Department, and a cost curve that fell from $300 per delivery to $12. Zipline is now racing toward a million deliveries a day, and a quiet inflection point when autonomous delivery becomes cheaper than sending a car. Hosted by Alfred Lin and Pat Grady, Sequoia Capital 00:00 Introduction 02:28 Early Vision and Regulation 04:09 Rwanda Launch Hard Lessons 06:49 Scaling to 24/7 Impact 09:35 Real World Ops Surprises 11:15 Safety Redundancy Failover 20:24 Precision Delivery Pod Tech 25:34 Building the Drone Network 26:51 Fleet Commanders Explained 28:22 Scaling to a Million a Day 29:51 Autonomy Enables 24 7 Ops 31:52 Reinventing Air Traffic Control 36:08 Why Zipline Is Vertical 41:40 First Principles Delete Parts 44:45 Market Explosion and Closing Thoughts [Source](https://www.youtube.com/watch?v=6bGxm8gX41o)

[0:00] 卢旺达医生与化验室技术人员: 人们生病是不分昼夜的，是一年 365 天、一天 24 小时随时会发生的，为什么你们一天只营业 12 个小时？ [1:03] Keller Rinaudo Cliffton: 我们所有的客户完全不在乎幕后运转的是什么技术 [11:15] Eric Watson: 必须假设系统的任何部件都可能出现故障或短暂异常。 [17:20] Alfred Lin: 真正的目标应该是比 Waymo 还要安全两倍 [24:37] Keller Rinaudo Cliffton: 安全且高效的空域融合 [36:08] Keller Rinaudo Cliffton: We do this not because it is easy but because we thought that it would be easy [39:07] Eric Watson: 今晚你和我一起通宵把这个问题彻底干掉。 [41:11] Keller Rinaudo Cliffton: 飞机上最可靠的零部件，就是根本不在飞机上的零部件，因为它在上一代设计中就已经被彻底删除了，一个不存在的零件永远不会发生故障。 [44:45] Keller Rinaudo Cliffton: Our operating plan has now become our unit of sale [49:52] Keller Rinaudo Cliffton: 做好心理准备，当你真正做出来时，它一定会花掉 10X。 [52:17] Keller Rinaudo Cliffton: You adopted the darkness. I was born in it. [52:17] Keller Rinaudo Cliffton: 在造机器人这件事成为潮流之前，我们已经在黑暗中硬生生干了整整十年。

[11:15] Two heads are better than one... We actually had one of these events happen a couple weeks ago. After a delivery, we delivered the package to the customer, and then we had a hiccup on the main flight computer, and we switched over to the backup, the aircraft flew itself home, landed, everything was totally fine. [31:52] Keller Rinaudo Cliffton: There are going to be 10 times as many autonomous vehicles in the air as there are using these teeny archaic constrained things that we call roads... You can give Earth back to humans. You can make neighborhoods quieter, safer, less pollution, less traffic. [36:08] Keller Rinaudo Cliffton: I have this flag over my desk that says: 'We do this not because it is easy, but because we thought that it would be easy.'... You never do it intentionally. Maybe you just slowly freak out and through desperation realize like, 'Wow, we got to tear all this out, and we got to build it all from scratch.' [49:21] It cost $300 a delivery and Alfred was surprisingly chill about it... It's now $12... The fully burdened unit economics of these systems is just now in the process of falling below the cost of using cars to deliver things. [52:17] We built a robotics company for 10 years before building a robotics company was a cool thing to do... We got to get good at building stuff again. And we got to get good at building not just vehicles and hardware, we got to get good at building infrastructure.

[0:00] I remember being in Rwanda early days [0:02] and going out and meeting with some of [0:03] the doctors and lab techs that we were [0:05] serving and asking for them like, you [0:06] know, how's it going? What what do you [0:08] think? What's your feedback? Here I am [0:10] kind of you know, up and coming you [0:12] know, learning engineer thinking they're [0:13] going to say something about the drone [0:15] or some of these things and the main [0:17] piece of feedback that I received was [0:19] people get sick 24/7. Why are you guys [0:22] only open 12 hours a day? [0:23] >> Mhm. [0:24] >> Right? [0:24] >> Especially when you're delivering [0:25] life-saving blood. [0:26] >> Yeah, exactly. And so that was a really [0:28] key insight for me where it's like, man, [0:31] we have found product market fit in a [0:33] market where yeah, our our you know, our [0:34] product wasn't great yet, but it was [0:36] solving a real need. [0:38] Uh and so having that that really [0:39] beachhead market where there's a real [0:40] problem being solved and when your [0:41] customer is telling you that their main [0:43] feedback is they want [music] more of [0:44] your service, it's like that's a good [0:46] sign. [0:55] >> [music] [1:03] >> Welcome Keller and Eric um to the show. [1:06] You guys have been working at Zipline [1:08] for a long period of time. Keller is the [1:10] co-founder and Eric, you are in charge [1:12] of systems engineering and safety and we [1:15] got lots of things to talk about in this [1:17] whole world of drones, drone systems and [1:20] how you guys started in this hardware [1:22] space before LLMs even started. So we [1:25] have lots of questions. [1:26] >> Awesome. But you don't like you don't [1:28] like Zipline being described as a drone [1:30] company even though you're the you're [1:32] probably the largest autonomous drone [1:34] company in the world right now. [1:36] >> I mean, you know, we we've always wanted [1:39] to be an extremely customer obsessed [1:41] company and the reality is none of our [1:42] customers care at all about drones. You [1:45] know, like we our our goal was always to [1:47] build an automated logistics system for [1:49] Earth and to approximate teleportation. [1:52] And all the customers who are like [1:54] living on Zipline today, they really [1:56] don't care how they don't care about the [1:58] technology operating behind the curtain. [2:00] What they care about is their ability to [2:01] like download an app, open it up, uh [2:05] you know, see a huge number of different [2:08] brands and amazing, you know, [2:09] restaurants that they want to shop with [2:10] and then click a button and have it [2:12] delivered to them 5 minutes later. So, [2:14] we've always really tried to focus on [2:15] the experience rather on rather than on [2:17] like the specific technology. [2:19] >> Well, this show is about technology. [2:22] >> [laughter] [2:22] >> We're excited about that. [2:23] >> This podcast is about technology. What [2:25] is the underlying technology behind [2:27] Zipline? You started in 2011. [2:30] You pivoted in 2014. [2:33] This is way before anything related to [2:36] >> AI or robotics. [2:37] >> AI, robotics, foundation models, [2:39] anything related to that. [2:41] >> Yeah. [2:41] >> But you've kind of you were before all [2:43] of that and you're riding the wave of [2:45] all of the things that have come [2:46] afterward as well. [2:47] >> Yeah, this was when like starting a [2:49] robotics company was the dumbest thing [2:51] you could possibly do. Like why you [2:53] know, you're talking to an investor in [2:54] the time about [2:56] I mean, it wasn't easy and it was [2:57] particularly hard cuz so many of those [2:59] conversations, you know, I mean [3:01] I was 23, 24. Eric joined the company [3:03] around that time and we were starting to [3:06] describe this vision of autonomous [3:08] logistics system for Earth that would be [3:10] 10 times as fast, half the cost, zero [3:12] emission. [3:14] Um [3:15] you know, one of the biggest problems [3:16] when we're trying to raise money for [3:17] that vision was investors would say, [3:18] "Isn't this illegal in the US?" In fact, [3:20] I think that that's the question you [3:21] asked me when we started talking about [3:22] this. [3:23] >> We weren't We weren't allowed to fly [3:25] beyond uh visual line of sight. [3:27] >> weren't allowed to fly at all, really, [3:29] but like yeah, and and so the answer was [3:31] yes, it's illegal. And then most [3:32] investors would be like, "Well, we don't [3:33] invest in illegal things. So, like we're [3:35] not [laughter] going to invest." But you [3:36] know, for weird reasons, this is what [3:38] basically took Zipline down this path of [3:41] like well, if it's illegal in the US, [3:42] then we can launch in other parts of the [3:43] world where you know, the the value of [3:45] the service would be extremely high. [3:47] Zipline decided to launch in Rwanda in [3:50] in [3:50] delivering blood transfusions directly [3:52] to hospitals and primary care [3:53] facilities. [3:54] Um this enabled us to have a use case [3:57] was so powerful that a government would [4:00] work very closely with us to make it [4:02] happen. [4:03] >> Make it legal. [4:03] >> And to make it legal [4:05] um or at least make an exemption to [4:06] their existing kind of regulatory [4:08] framework. [4:09] And um you know, and then the other [4:11] thing, you know, when it comes to how to [4:13] think about Zipline as a company, you [4:14] know, when we launched in 2016, we were [4:16] like, we have this really cool drone. We [4:18] put all this work into designing these [4:19] really cool aircraft that you know, and [4:21] it's it has all these great fundamental [4:23] features. And when we launched, it was a [4:24] total disaster. [4:26] Um because the reality what we what we [4:28] learned in that first year is for the [4:30] first eight we we'd signed a contract to [4:31] sign 21 to serve 21 hospitals. [4:34] And we served one hospital for the first [4:36] 9 months. And Eric in particular, like [4:39] how much did you sleep during those [4:40] nights? I mean [4:41] >> a lot of time in Rwanda and didn't sleep [4:43] a lot. [4:43] >> when you're in the US, like we'd get [4:45] woken up at like midnight cuz that's [4:46] when the distribution center was turning [4:47] on and it would everything would be [4:48] broken. Nothing was working. And it was [4:50] totally desperate, constant all-nighters [4:52] and working through the weekends because [4:54] we'd made this big error which was that [4:56] thinking that like the cool vehicle was [4:58] the majority of the solution. [5:00] What we learned during that first year [5:02] is that the drone is 15% of the [5:04] complexity of the solution. Like Zipline [5:06] had [5:07] >> drone, the hardware of it. [5:08] >> Yeah. Is only 15%. to build so many [5:10] auxiliary software systems, [5:13] maintenance systems, uh it how do we [5:15] hold the inventory and do inventory [5:17] management, how do we integrate with a [5:19] national civil aviation authority which [5:21] we'll talk more about, how do we [5:22] integrate with a national healthcare [5:23] system, how do we do ordering and demand [5:25] management. You know, we had to build [5:27] out all of these other parts of the [5:28] overall logistics system. So this is the [5:30] reason I think a lot of people might [5:32] look at Zipline be like, wow, it's a [5:33] cool drone company. They build a cool [5:34] aircraft. The reality is the aircraft is [5:36] like 15% of the solution that's required [5:39] to build something that just feels like [5:41] magical, reliable, teleportation 24/7, [5:45] 365 to the [5:47] now, you know, hundreds of millions of [5:48] people who depend on the service. [5:50] >> Speaking of 24/7, I remember being in [5:52] Rwanda early days and going out and [5:54] meeting with some of the doctors and lab [5:55] techs that we were serving and asking [5:58] them like, you know, how's it going? [5:59] What what do you think? What's your [6:00] feedback? Just really being customer [6:02] obsessed and and wanting to optimize the [6:04] product. And, you know, here I am kind [6:06] of you know, up-and-coming you know, [6:09] learning engineer thinking they're going [6:10] to say something about the drone or some [6:12] of these things. And the main piece of [6:14] feedback that I received was people get [6:16] sick 24/7. Why are you guys only open 12 [6:19] hours a day? [6:20] >> Mhm. [6:20] >> Right? And so [6:22] >> Especially when you're delivering [6:23] life-saving blood. [6:24] >> Yeah, exactly. And so that was our, you [6:25] know, you got to start somewhere, right? [6:26] So we started being open 12 hours a day [6:28] and trying to expand and and grow from [6:30] there. And so that was a really key [6:32] insight for me where it's like, man, [6:34] we found product-market fit in a market [6:36] where yeah, our our, you know, our [6:37] product wasn't great yet, but it was [6:39] solving a real need. Uh and so having [6:41] that that really beachhead market where [6:43] there's a real problem being solved and [6:45] when your customer is telling you that [6:46] their main feedback is they want more of [6:47] your service, [6:48] it's like that's a good sign. [6:49] >> Yeah, we were 24/7 within the first [6:52] year. [6:52] >> 24/7. [6:53] >> 24/7 365. I mean, on Christmas day I [6:55] usually call all of our different [6:57] distribution centers to like thank them [6:58] and check in with them. So like there is [7:00] no day when these facilities don't [7:03] depend on, you know, we went from [7:05] serving one to 20 to 500 now to 5,000 [7:09] hospitals and health facilities across [7:10] the world across eight countries that [7:12] are served by the system. It's become [7:13] the largest commercial autonomous system [7:15] on Earth. [7:15] >> Can you size for us the largest [7:18] system on Earth? [7:19] >> We just crossed 140 million commercial [7:20] autonomous miles, which I mean, how many [7:22] times is that I think that that's like [7:23] to the sun and back or [7:25] >> One of the one of the things that I like [7:26] is every road in the United States, [7:28] there's a lot of roads in the United [7:29] States, driving on every single road [7:31] more than 30 times. [7:32] >> Wow. [7:33] >> It's a good stat. [7:34] >> It's a lot. That's a lot just to put in [7:35] perspective. [7:36] >> You know, seeing the impact that that [7:37] system is now having in across all these [7:40] eight countries, I mean, University of [7:42] Pennsylvania just published a study [7:44] showing a 51% reduction in maternal [7:46] mortality [7:48] thanks to Zipline. So, half as many moms [7:50] losing their lives in childbirth. Um we [7:52] have you know, across all the different [7:54] use cases that Zipline serves, some of [7:56] our partners estimate that we're saving [7:58] between 10 and 12,000 lives a year. And [8:00] that impact is growing exponentially as [8:02] we're now expanding, especially as a [8:03] result of this new partnership we have [8:04] with the US State Department. [8:06] >> What is that partnership with the State [8:07] Department? [8:07] >> December, we announced a $550 million [8:10] partnership with the US State Department [8:12] to expand the impact of Zipline's [8:14] life-saving service across a lot of the [8:16] countries where we're already operating. [8:17] So, with USAID being shut down, the US [8:20] was really seeking like new ways of [8:23] engaging in these countries and helping [8:24] save lives in these countries, but they [8:26] wanted to do it in a way that would [8:28] um that that would accelerate the [8:29] economies of these countries and help [8:31] the US economically. And so, the new [8:32] strategy they're calling commercial [8:34] diplomacy. The idea is that we want all [8:36] of the developing world should be built [8:38] on top of US AI and robotics technology. [8:41] We should be going and, you know, [8:43] economically helping. We should be [8:44] bringing the best that the US has to [8:46] offer. Interesting thing is when you [8:48] talk to these countries about what they [8:49] want, they'll tell you they are sick of, [8:52] you know, low-quality aid provided by [8:54] NGOs for free because these services [8:56] engender dependence and prevent economic [8:58] growth in the countries. What they want [9:00] is high-paying jobs, [9:02] entrepreneurship, technology. [9:04] >> Mhm. [9:04] >> And so, the US is [9:06] is, you know, going through a big [9:07] strategic shift where it's like, well, [9:08] we have that. We have those things. [9:10] Like, so let's let's basically go out [9:12] and incentivize these countries to adopt [9:15] that kind of infrastructure, make sure [9:16] that, [9:17] you know, as these countries are [9:18] accelerating, they're they're doing it [9:20] using US [9:22] robotics and and AI technology. And this [9:24] is something that will be great for [9:26] those countries. It saves lives. It [9:27] saves them money, but it also means that [9:28] it will it will [9:30] make it possible for the US to secure [9:32] our lead in manufacturing and robotics [9:34] of the decades to come. [9:35] >> I'm curious about, you know, you guys [9:37] because you now run the largest [9:39] autonomous system in the world and you [9:42] you launched it 10 years ago at this [9:43] point. So, you've been in production for [9:45] 10 years. You've learned a lot of stuff [9:48] that your average engineer sitting [9:50] behind a computer screen has no idea [9:51] they're going to run into when they try [9:53] to deploy AI into the real world. And [9:56] so, I I'm curious what some of those [9:58] lessons learned are. And maybe one way [10:00] to ask the question is what popped up [10:02] over the last 10 years that you never [10:04] would have guessed you needed to be good [10:06] at when you first started launching [10:08] these systems in 2016. [10:09] >> Yeah. [10:10] >> You know, we started off delivering [10:12] life-saving products, right? And our [10:13] customers need [10:15] need life-saving products all the time [10:17] in all weather conditions. And you would [10:19] think it's, you know, wind, these [10:20] things, but one of the weirdest things [10:22] is actually solar weather. [10:24] So, there's solar flares that happen on [10:26] the sun. So, they're basically big [10:28] explosions that send radiation to the [10:30] earth. They can mess with the ionosphere [10:32] and that can cause basically the RF [10:34] signals coming from GPS satellites to be [10:36] faster or slower than you expect. And [10:38] that can lead to degradation and [10:40] challenges in navigation systems. And [10:43] so, here's one example that, we know, [10:44] when we were starting off, we didn't [10:46] think that this was going to be [10:47] something we have to figure out. Uh but [10:48] we actually have, uh you know, gone [10:50] pretty deep in this space and and really [10:51] it's two things. One is [10:53] designing our navigation system and our [10:56] our GNSS systems to be robust to these [10:58] conditions to ensure that we can still [11:00] know where, you know, where our aircraft [11:02] are with centimeter-level precision in [11:04] those conditions and those challenging [11:05] solar flare times, as well as designing [11:08] the system to have redundancy beyond [11:11] GNSS such that if things get really bad, [11:13] we can still safely operate. [11:15] >> Eric, you're in charge of safety. Tell [11:17] us about [11:18] uh what you've learned about safety [11:19] today and and specifically about the [11:22] compute failover system that you have. [11:25] >> Yeah. Yeah, absolutely. I mean, there's [11:26] so many things that we've learned over [11:27] the last decade of operating, you know, [11:29] the the system in the real world. Um one [11:31] of the things that that we're proud of [11:33] is [11:34] uh how we we've developed to to your [11:35] point um compute failover. So, there's a [11:37] flight computer, flies the aircraft, [11:39] lots of sensors come into this uh into [11:42] this computer, and that basically does a [11:44] lot of math and sends commands to [11:46] actuators, right? So, motors, control [11:48] services, these things. So, this is the [11:49] brain that flies the aircraft, right? [11:52] Um, [11:53] there, you know, one of the things that [11:54] we've learned is you need to assume that [11:57] any part of the system can have a fault, [11:59] can have a hiccup, something can go [12:00] wrong. And that's how you really design [12:02] something to be robust, reliable, and [12:04] safe. So, what do we do if this flight [12:07] computer has a challenge? Could be a [12:09] software challenge, it could be a [12:10] connector challenge, could be these [12:11] different things. [12:11] >> Bit flip due to solar radiation. [12:13] >> All kinds of things, right? And so, what [12:15] we've done is we have two flight [12:16] computers. And both of these flight [12:18] computers think that they're flying the [12:20] aircraft at any given point in time. [12:22] They're all receiving all the [12:23] information from the sensors, they're [12:24] all sending in commands to the [12:25] actuators, and there's uh like a kind of [12:27] a third arbiter, a little computer, that [12:30] is monitoring the health of those two [12:32] and telling everyone every other node on [12:34] the on the aircraft who to listen to, [12:36] who's actually in charge. [12:37] >> What if the arbiter fails? [12:38] >> Yeah, if the arbiter fails, then the the [12:40] primary computer that was flying just [12:42] keeps flying. Right? So, one's in [12:44] charge, and if the thing that's [12:45] monitoring itself fails, then now we [12:47] say, "Okay, like, you know, now we're [12:49] just going to keep flying on the thing [12:50] that was good, and we're going to keep [12:52] flying the mission." Um, so [12:54] >> Two heads are better than one. [12:56] >> [laughter] [12:57] >> Yeah, it's something we're really proud [12:58] of. Um, we had actually had one of these [13:00] events happen [13:01] uh a couple weeks ago, yeah, where we [13:03] had after a delivery, we delivered the [13:04] package to the customer, and then we had [13:06] a a hiccup on the main flight computer, [13:08] and we switched over to the backup, the [13:10] aircraft flew itself home, landed, [13:12] everything was totally fine. So, just, [13:14] you know, designing the systems to be [13:16] robust and reliable through and through [13:18] is is, you know, how you get the 2.5 [13:19] million deliveries and 140 million miles [13:22] flown with no safety incidents. [13:23] >> And a lot of what Zip line's doing, it's [13:24] not like, "Oh, this is totally [13:26] revolutionary. No one has ever thought [13:27] about having a secondary flight [13:28] computer. That's how Boeing 777 works. [13:31] But the cost of a flight computer on a [13:34] Boeing 777 is in the millions of [13:35] dollars. And so a lot of what Zipline's [13:37] having to do is take a lot of the best [13:40] ideas that you can see in aerospace [13:42] safety [13:43] >> best practices for aerospace companies. [13:45] >> And you and then you got to figure out [13:47] how to build that using components [13:49] coming out of the smartphone supply [13:51] chain. [13:51] >> Yeah. [13:51] >> do it for you know tens of dollars or [13:53] hundreds of dollars. You can achieve [13:55] similar levels of safety to traditional [13:58] aerospace, but you can move 100 times as [14:01] fast at 1/100 of the cost. [14:03] >> Yeah. [14:03] >> So you you mentioned that the aircraft [14:05] is only 15%. [14:07] Describe the other 85% in like layers [14:10] and maybe go down deep in some of your [14:13] systems that are really really [14:15] sophisticated like your like I know this [14:17] because of what being a board member [14:19] like the detect and avoid systems. [14:21] >> So you know how do we test why do we [14:23] test? Really um the way I think about it [14:25] is there [14:27] maybe first of all we're we're not a [14:28] software company, right? We're a real [14:30] world AI robotics company. And so [14:33] there's electromechanical systems out [14:34] out in the real world. So there's [14:36] hardware test aspects, there's software [14:38] test aspects, and there's the integrated [14:39] system test aspects. We have a lot of [14:42] different environments that we test a [14:43] lot of different approaches. I'll I'll [14:44] name a few of them. You know on the [14:46] hardware side we do a lot of component [14:48] level testing [14:49] halt testing highly accelerated lifetime [14:51] testing where we're taking components [14:53] maybe it's a motor these kinds of things [14:54] and we're putting them through you know [14:56] through hell, right? We're putting them [14:57] through all kinds of challenging [14:58] conditions making it rain making it hot [15:00] making it human making it corrosive all [15:02] these things while we're exercising you [15:03] know while we're spinning the motor [15:04] while we're moving things all of the [15:05] things, right? UV like you name [15:07] >> And just to give a context for scale and [15:10] there are 700 unique components on the [15:12] aircraft designed from scratch by [15:14] Zipline. We are designing not just the [15:16] flight computer from scratch, the power [15:18] distribution board, the motor [15:19] controllers, the battery, the battery [15:22] management system, the you know, the pod [15:24] is the smaller robot that we're using to [15:26] actually make deliveries to people's [15:27] homes. There's an entire Nvidia Nvidia [15:29] GPU-powered flight computer on the pod. [15:32] We're building the electronics that go [15:34] into the docking station where the Zip [15:36] is flying in and out of. [15:37] Uh we [15:39] you all of that, even the electric motor [15:41] being designed from scratch by Zipline [15:42] because we need a you know, a a [15:45] thrust-to-weight ratio that is not [15:46] available in off-the-shelf electric [15:47] motors. So, you have to design something [15:48] from scratch. So, you know, 700 unique [15:50] components, 43 major sub-assemblies on [15:52] the aircraft, all then coming together [15:54] on the manufacturing line that you both [15:56] have gotten to visit. Uh [15:58] and getting assembled into one overall [16:00] aircraft. But anyway, that's the So, for [16:01] So, for each of those components [16:03] >> They're all right. Going through this [16:04] type of testing. And what you know, [16:05] thinking about other industries, often [16:06] times when I talk to people from maybe [16:08] automotive or aerospace some of these [16:10] like, "Hey, how do you think about [16:12] reliability challenges?" And a common [16:13] answer is like, "Well, I ask the [16:14] supplier what the reliability of the [16:15] part is." [16:16] >> Yeah. [16:16] >> And I'm like, "Okay, cool. Like, what if [16:18] we're the supplier, you know?" So, [16:19] anyway, so we're that vertical [16:20] integration where we have component [16:22] testing on the ground, we have system [16:23] testing on the ground where we're taking [16:25] full aircraft for you know, and as well [16:27] as other parts of the of the system and [16:29] putting them through vibration tables, [16:31] wind tunnels, thermal chambers you can [16:33] walk into, like all of these things um [16:36] in order to understand is how is this [16:38] going to break, right? More than just is [16:40] it good enough? Like, we want to know [16:41] how it's going to break. And then we can [16:43] understand, okay, cool. Like, let's make [16:44] it better. Or maybe it's like, "Oh, [16:46] that's not too worrisome. Like, great." [16:48] You know, we we didn't break any of the [16:49] ways we were worried about. It broke in [16:50] that way. Fantastic. So, like, we don't [16:51] just want to say, "We ran the test [16:53] campaign and nothing failed. We're [16:54] done." It's like, "No, no, let's take [16:55] this thing to failure, right? Let's see [16:56] where the limits are." [16:57] >> 49° C, [16:59] which is very hot. [17:00] >> Hot. Cold. [17:01] >> -25° [17:02] C, which is very cold. [17:04] >> all the things, yeah. [17:05] >> You don't fly anywhere at 49° C. [17:08] >> You do? That's what We wouldn't test at [17:10] 49 if we're not worried about it. Where [17:11] are you flying at 49? [17:12] >> I think Phoenix during the summer. [17:13] >> Phoenix during the summer, yeah. So, [17:15] >> And then where's -25? [17:17] >> So, [17:17] >> northern parts northern parts of the [17:18] United States. Actually curious what you [17:20] guys think about that like [17:21] I could imagine a different version of [17:23] the world where you guys are like hey [17:25] look [17:26] if it's too hot we're just not going to [17:27] fly. [17:28] >> Totally. [17:28] >> And if it's too cold we're just not [17:30] going to fly. And if it's raining too [17:31] hard we're just not going to fly. [17:33] >> Yeah. [17:33] >> And and and like there are trade-offs to [17:35] be made you know and obviously your [17:36] customers would prefer that you fly at [17:38] all times. But how do you think about [17:39] those trade-offs? [17:40] >> The easiest way to think about the [17:41] trade-off was because of the use cases [17:43] that Zipline started with. [17:44] >> That's right. [17:44] >> Which was basically saving lives. [17:47] >> Yeah you can count on us with your life [17:49] and the lives of your loved ones as long [17:51] as the sun is shining. [17:52] >> [laughter] [17:54] >> It's not super compelling you know like [17:58] >> [clears throat] [17:58] >> Zipline would fly and in fact I mean for [18:00] the first you know for the first [18:03] couple years we took a lot of risk. I [18:05] mean we would basically fly. We were [18:07] like look if it's a life saving [18:09] delivery happening and there's someone [18:11] whose life is on the line we're going to [18:13] go for it. And we had a civilian [18:14] aviation authority that was you know [18:16] generally a great partner with us on [18:17] that front. We took a lot of risk we [18:19] learned a lot and you know almost always [18:22] it worked out in favor of like we saved [18:24] the person's life. And you know the [18:25] worst thing that could happen was you [18:27] know we had a parawing which is the kind [18:28] of like Zipline's safety mechanism of [18:30] last resort is we can pull a parachute [18:32] on the aircraft and bring it gently to [18:33] the ground. [18:34] >> How often does that [18:34] >> learned a lot. It happened very often in [18:37] the first few years like yeah very very [18:39] rare today. I mean to put it into [18:41] perspective you know our original goal [18:43] was to be 10 times safer than cars. [18:45] Actually Alfred was the one put pushing [18:47] in our last board meeting he's like [18:48] that's a BS goal we need to be two times [18:50] safer than Waymo. And so Eric literally [18:52] went and reset the goal and now we're [18:54] like the Zipline's target for the end of [18:56] this year is to be two times safer than [18:57] Waymo. He's like cars that's like [18:59] archaic technology. [19:00] >> Waymo I think is about 10X right? [19:02] >> They're about 10X safer than 10 12. [19:04] >> So our goal is to be 2X safer than [19:05] Waymo. [19:06] >> Yeah. [19:06] >> And [19:07] >> the right comparison you're flying. You [19:09] have to be safe in the in in the air not [19:11] safe on the ground. [19:12] >> we're substituting something that's [19:14] typically going in cars, so it's like [19:15] debatable. But um you know, suffice it [19:18] to say I mean we we now have 140 million [19:20] commercial autonomous miles and zero [19:21] safety incidents. [19:22] >> Zero. [19:23] >> If you were to drive 140 million miles, [19:25] you would have 600 accidents, 100 [19:28] injuries, and somewhere between two and [19:29] six fatalities depending on what country [19:31] you're talking about. And um [19:33] you know, this is why it's you know, we [19:35] really pride ourselves in like picking [19:36] the right use cases. It's like it's life [19:37] saving and it really makes a lot of [19:39] sense to go do it. And also we're going [19:40] to be by God we're going to be as safe [19:42] as humanly possible from an engineering [19:44] and testing and validation perspective. [19:45] We really take that um [19:47] that's a that's a deep part of the DNA [19:48] of the company. [19:49] One last point, you know, what is the [19:51] outcome of all of that testing that Eric [19:53] is talking about? The outcome of all [19:55] that testing is we have individual [19:57] aircraft in the commercial fleet that [19:59] have flown more than a million [20:01] commercial autonomous miles. And so I [20:02] think people, you know, that's just from [20:04] an intuition perspective, a lot of [20:06] people look at this and they're like, [20:07] "Wow, it kind of seems maybe exquisite [20:09] or fragile, probably very sensitive to [20:11] like [20:12] extreme conditions or weather." I mean, [20:14] you know, raise your hand if you have a [20:16] car that has a million miles on it. [20:19] >> It's pretty impressive. [20:19] >> These systems are already like way more [20:21] rugged and durable and robust than [20:23] people necessarily think. [20:25] >> ask you about the precision? [20:26] Like one of the things that blew my mind [20:28] when I saw some of the I haven't had a [20:29] chance to experience in person, you [20:31] know, a delivery. [20:32] >> Patrick. [20:33] >> I got to I got to go experience it. But [20:35] just but just in watching the videos, [20:37] the drone's 100 ft up [20:39] and it drops the package, it lowers the [20:42] package to a I don't know, a circle [20:44] that's got a 18-in radius or whatever it [20:46] is, right? Like how do you guys achieve [20:48] such precision even when it's windy, [20:51] even when it's raining? Yeah. How do you [20:53] How do you pull that off? [20:54] >> First of all, the aircraft's about 100 m [20:55] up. [20:56] >> 100 m up. Okay, yeah, there you go. [20:58] >> it harder. Um [20:59] and uh you know, there multiple layers [21:01] of there's the the delivery pod that [21:04] comes down, right? So the delivery pod [21:06] comes down. That's really the delivery [21:08] and pick up like precision part of it, [21:10] right? So the the drone is hovering [21:12] above. [21:13] Um it it you know, it knows where the [21:14] target is. Maybe let's say it's it's [21:16] this coffee table, for example, if there [21:17] wasn't a roof above us. Um so it's this [21:19] coffee table, and so the aircraft is [21:21] going to hover above, but it actually [21:23] needs to consider what the wind [21:24] conditions are, right? So if the wind's [21:26] blowing in one direction, then the [21:27] aircraft's going to kind of be shifted, [21:29] you know, upwind, right? So it's going [21:30] to shift in the direction to help um [21:32] with that with those wind conditions. [21:34] And then it's going to lower that [21:35] delivery pod down. Um as Keller [21:37] mentioned, we we do uh take advantage of [21:41] GNSS, so real-time kinematic GNSS. That [21:44] gives you centimeter level um confidence [21:46] of of where you are. But the thing is, [21:48] we don't know the GPS coordinates of [21:51] this table, right? It's not like someone [21:53] came in and surveyed the middle of the [21:54] table and sent us the coordinates, [21:55] right? No one wants to do that. So what [21:57] we have to do is we kind of that we use [21:59] that to kind of get close, right? We're [22:00] like, "Okay, here's the backyard. Here's [22:02] where we kind of know things that the [22:03] things roughly are." And then the job of [22:05] this delivery pod is to be lowered down, [22:08] you know, fight the wind conditions, [22:09] fight these different things, and be [22:10] able to use its onboard perception [22:13] autonomy systems to identify where's the [22:15] best place for me to leave the package, [22:17] right? Like if there's a little table [22:18] and there's a whole bunch of drinks, [22:20] probably I shouldn't, you know, try and [22:22] drop down on top of these drinks and [22:23] make a mess. Maybe I should go to the [22:24] ground right next to the table, right? [22:26] And so it has these autonomous, you [22:28] know, onboard real-time compute uh to be [22:31] able to identify what am I looking at? [22:33] What am I seeing? And how can I find the [22:35] best place to leave the package? And [22:37] then come down, touch the ground, opens [22:39] its doors, gets retracted back up, and [22:41] there you go, the package is left on the [22:42] ground, and the the delivery pod comes [22:44] back up, stows, and the aircraft flies [22:46] back home. [22:46] >> A couple big advantages, I mean, just to [22:48] be specific. So that pod, it not only [22:50] has its own Nvidia GPU running its own [22:52] AI autonomy stack, so it [22:54] >> Yeah, and survey and like know exactly [22:56] where it's delivering, even at night. [22:58] >> Yeah. [22:59] >> But um it's it's also controlling its [23:01] own position [23:02] >> That's right. [23:03] >> in the X and Y axis, so it can [23:04] >> Yeah. it it can not just know, but then [23:06] move. Um and the advantage of that [23:08] architecture, which you can probably [23:10] guess, but there are two huge advantages [23:12] of doing it in this way. One is it's [23:14] quiet. People have this perception of I [23:16] mean, first of all, most drones are [23:18] really freaking annoying. Like the sound [23:20] is just It's basically the most grating [23:22] annoying sound that you could possibly [23:23] subject a human to. And so, you know, [23:26] like we Zipline has a big team of [23:29] aerodynamics, aeroacoustics, and [23:31] controls experts. Every part of the [23:33] vehicle is designed with sound in mind [23:36] for the vehicle to be as quiet as [23:37] humanly possible. We want it to be no [23:39] louder than the sound of like gentle [23:42] leaves moving in trees. [23:44] Um and for the when the pod is [23:46] delivering, we're keeping the main [23:47] aircraft 100 m in the air. So, it's like [23:50] the thing that is creating noise is [23:52] really far away. That's also a huge [23:54] benefit from a safety perspective. [23:56] Because the only thing that is coming [23:57] anywhere close to you, your family, your [24:00] pets, your kids, is something that is [24:02] super cute and safe. And it's you know, [24:03] it's really like a Styrofoam kind of [24:05] like a cute anthropomorphic Styrofoam [24:08] >> tub [24:08] >> tub. [24:09] >> [laughter] [24:10] >> Yeah. How long was the [24:13] technology tested outside the United [24:14] States before it came to the United [24:16] States? And was the path to getting into [24:18] the US now that you're flying in Dallas [24:20] and delivering packages there? [24:22] >> We spent 8 years, I think, right? About [24:25] about about 8 years. I mean, I depending [24:26] on how you measure it, maybe like six [24:28] six to eight years. And then it was I [24:30] mean, we launched in Rwanda in 2016, our [24:33] commercial service. And we really [24:34] launched the this kind of next [24:37] generation home delivery service, the [24:38] thing that's now like in sort of insane [24:40] hyper scaling mode. That only launched [24:42] January of last year. So, depending on [24:44] how you measure it, you could even say [24:45] it was like almost 9 years. [24:47] >> And then when you got to the US, was it [24:49] just smooth sailing? What What was the [24:51] sort of regulatory path that you had to [24:52] go through? [24:53] >> Yeah, I mean, we really started, I would [24:56] say, you know, like meaningfully [24:57] engaging with US, what you know, FAA and [25:00] other um regulators in the US around [25:03] 2020 or so. [25:05] Um so, it doesn't mean we didn't show up [25:06] in 2025 and everything was smooth [25:07] sailing. It was really a partnership of [25:09] working through [25:10] um as you mentioned in kind of 2016, all [25:12] of this stuff was there was no pathways. [25:14] It's kind of illegal as we as we joked [25:15] earlier. And so, [25:17] um yeah, so we really was a partnership [25:19] to identify, "Hey, you know, [25:21] we have a shared goals, right? Our [25:23] shared goals are safe and efficient [25:26] airspace integration. And so, while we [25:29] have uh experience doing that [25:30] successfully in different countries, uh [25:33] we can bring some of that experience and [25:34] we have opinions on how this should [25:35] work, the regulators had opinions on [25:37] maybe how they thought it should work. [25:38] And so, it was a partnership over the [25:39] course of a couple years to identify [25:41] what those paths look like and how we [25:43] could kind of converge and align before [25:45] we were able to execute on that. [25:46] >> And you had to show your ability to [25:48] manage all these aircrafts that are [25:49] flying. So, you wrote systems, you built [25:51] systems. [25:52] >> Yeah, I think it's a huge part of, you [25:53] know, Keller's mentioning that the drone [25:55] is only a part of the of the, you know, [25:56] of the overall system, the overall [25:58] complexity. What we're really building [26:00] is an infrastructure layer, right? We're [26:02] building an infrastructure layer that [26:03] can enable instant access to products. [26:06] And you don't do that with one aircraft [26:08] flying from one place to another place. [26:09] You do that with a network of um [26:12] charging locations, hundreds of aircraft [26:14] uh spread across an area that uh with [26:17] the autonomous systems in the in the [26:19] cloud that can understand where am I [26:21] having demand, where where do I have [26:23] supply, where do I have aircraft, um [26:24] what's coming up is about to be the [26:26] dinner rush, uh what's the weather at [26:28] these different locations, how can I [26:29] kind of self-balance these things, as [26:31] well as how do I efficiently uh [26:35] pull like pull in people when needed, [26:37] right? So, these these aircraft are [26:38] autonomous, they're operating, they [26:40] don't require human intervention through [26:42] these flights, uh but there are times in [26:44] which it makes sense to alert a person [26:46] that, "Hey, maybe there's there's an [26:47] issue here or the weather is a little [26:49] bit uh the wind is is in this area, [26:51] right?" So, there are humans, you know, [26:52] trained aviation professionals that are [26:54] monitoring our like our network, I would [26:57] call it. [26:57] >> They're fleet commanders. [26:58] >> Fleet commanders, that's right. [27:00] >> We used to call them pilots, [27:01] >> [laughter] [27:02] >> you know, cuz when we originally [27:03] launched in the US, the first regulatory [27:05] permission we got was to fly one to one. [27:07] So, that meant that we had one pilot [27:09] sitting in an [27:10] >> Remote pilot in command. [27:12] >> who was sitting in an office basically [27:13] just observing an aircraft do its thing. [27:16] And again, you know, it's exceedingly [27:18] rare that a human should ever have [27:20] >> to issue any kind of a command to a [27:21] vehicle, but we would have one human [27:23] watching one aircraft. Not great for [27:24] unit economics. [27:25] Uh [27:26] but as Zipline proved out these systems, [27:29] we went from one to one to one to three, [27:31] one to six, one to 20, one to 40, and [27:34] now operating one to 100, and have plans [27:37] to go well beyond that. [27:38] >> Well, one fleet fleet commander now. [27:40] >> So, yeah, we technically changed the [27:41] name cuz I think pilot's confusing. So, [27:43] we're inspired by Ender's Game. We now [27:44] call these this group of this team of [27:46] people at Zipline, we call them fleet [27:48] commanders. And I'd actually says that [27:50] in the FAA documentation, we say fleet [27:52] commander shall do the following. And [27:54] yeah, they are overseeing a group of 100 [27:57] aircraft. And to me, this is like the [27:58] exciting cool thing about technology cuz [28:00] people think about like, well, you know, [28:04] what about um you know, how humans used [28:06] to solve this problem? It's like it's [28:07] not, you know, the [28:09] it it it it's cool how robots enable [28:11] humans to like up level, right? Like the [28:13] human is still getting to like [28:14] strategically manage the system. It's [28:16] just the human is now maintaining and [28:18] commanding robots rather than like doing [28:20] the actual work [28:21] herself. [28:22] >> Now that you guys are kind of in hyper [28:23] scale mode, you solved so many problems [28:25] in the last 10 or 15 years. What new [28:27] problems are you running into? [28:29] >> Yeah, I mean, what I would say like [28:30] thematically, I I mentioned earlier that [28:33] as you know, getting to 2.5 million [28:34] deliveries, the you know, the it only [28:37] happens every couple of years. It's like [28:39] kind of a one in a million chances. Um [28:41] these things start to matter, right? [28:43] We're we're on the path towards a [28:44] million deliveries every day. And if you [28:46] have a one in a million situation, it's [28:48] going to happen every single day, right? [28:49] >> Yeah, can I just just just to really [28:50] make that clear. So, [28:52] it took Zipline from 2014 [28:56] when we started building the original [28:57] version of the technology to 2020 [29:01] 4 to do our first million deliveries. [29:02] Was it the end of 2024? Maybe it was [29:04] even early 2025 actually that we did we [29:06] had done a million deliveries in the [29:08] cumulative history of the company. [29:10] >> Yeah. [29:11] >> So, it's a almost a decade. Maybe say [29:13] about a decade to do a million [29:14] deliveries. Zipline is now in the very [29:17] near future [29:18] going to be doing a million deliveries a [29:19] day. [29:20] And so, that is definitely [29:23] humbling. It's like, wow, okay. [29:25] Everything about the way we've been [29:28] solving the problem is going to break. [29:29] The bar goes way way up. [29:31] >> Yeah. [29:31] >> And I mean, you know, one specific [29:33] example, you know, maintenance becomes [29:36] really hard. Like, you know, the scale [29:38] of the problems, the number of vehicles [29:39] that you're managing in the fleet, the [29:41] cost of a screw-up or if some if a [29:44] certain process is operating in very [29:45] inefficient ways becomes extremely high. [29:48] And so, there's just high degree of [29:50] criticality for all these systems. One [29:52] interesting point though, you know, [29:53] there are a lot of ways that these [29:54] systems operate that I think people [29:56] don't yet appreciate the advantages of [29:58] autonomy. One good example is that like [30:00] the system wants to operate 24/7. It [30:02] does operate 24/7. So, I think people [30:04] are used to like logistics is generally [30:06] being like, well, here are the hours [30:07] when humans are driving trucks. [30:10] That's not how these systems operate. [30:12] They want to operate 24/7. They can be [30:14] fully utilized. They they can be as [30:16] happily delivering at 2:00 a.m. and 3:00 [30:18] a.m. delivering something so it's like [30:19] ready for you on your doorstep or in [30:21] your backyard when you wake up at 6:00 [30:22] a.m. before you go to work as they are [30:24] delivering at 2:00 p.m. Um, they can [30:26] deliver in 5 minutes. They are they are [30:28] available 100% of the time. We are soon [30:30] going to be flying vehicles straight out [30:31] of our factory in South San Francisco [30:33] into commercial operation. If you've [30:35] seen, you know, Tesla Model uh 3's and [30:37] Model Y's delivering themselves to [30:39] customers, [30:40] Zipline aircraft will fly straight from [30:41] the factory into operation. It's huge [30:43] advantage from a maintenance perspective [30:45] that as soon as a vehicle needs to go [30:47] through some kind of proactive [30:48] maintenance, it will fly itself to the [30:49] maintenance depot. So, the human can [30:51] then quickly make, you know, [30:53] do whatever process necessary and then [30:55] the vehicle flies itself back into [30:56] operations. We can also dynamically [30:58] assign capacity in a metro [31:00] based on what the system is seeing. [31:02] There's no like set home for a vehicle. [31:04] It can go to wherever it's needed. [31:06] >> Yeah, I think to to to your question [31:08] about, you know, getting to a million a [31:09] day and what are the new challenges, I [31:11] think, you know, Keller hit on some of [31:12] them. To the previous thought about the [31:14] drone is only 15% of the problem. [31:16] Really, it's the way that we currently [31:19] manufacture aircraft, maintain aircraft, [31:21] support all these things, you know, [31:22] troubleshoot problems. Like, the way [31:24] that we do it today isn't going to work [31:26] when we're at a million deliveries a [31:28] day. And so, there's still like, okay, [31:30] we need better tools, we need better [31:31] software systems, we need better [31:32] processes, we need better, you know, all [31:34] these things. So, it's like, um, you [31:35] know, Elon talks about designing the [31:37] machine that builds the machine. And so, [31:40] you know, this is really one of the [31:42] things that I see Zipline tackling over [31:44] the coming couple years is we're going [31:46] to be investing much more in the [31:48] machines that build and run the [31:51] machines. [31:52] >> Mhm. [31:52] >> I mean, from a scale perspective, I [31:54] think the largest airline in the US is [31:56] doing about 5,000 flights a day. [31:57] >> Yeah. [31:58] >> Zipline is going to surpass that in the [32:00] next month. And when we get to a million [32:03] deliveries a day, Zipline will be doing [32:05] like somewhere between, I don't know, 40 [32:08] and 80 times as many flights in the US [32:11] in commercial airspace as all other [32:13] airlines combined. [32:14] >> Yeah. [32:15] >> And so, it's obviously a different class [32:17] it's completely different class of [32:18] aircraft, it's a totally different kind [32:19] of problem, but the reality is when you [32:21] look at air traffic control, [32:23] they don't make a distinction. And so, [32:24] there's there's also when you talk about [32:26] all the, you know, auxiliary systems [32:28] that have to be built, there is a huge [32:29] transformation that's going to have to [32:31] happen in air traffic control as we [32:33] start to realize that, you know, people [32:35] are really excited about electrification [32:36] of vehicles, People are excited about [32:38] autonomous vehicles. The reality is [32:40] what as those transformations occur, [32:42] there are going to be 10 times as many [32:44] autonomous vehicles in the air as there [32:45] are using these teeny archaic [32:48] constrained things that we call roads. [32:50] And so like the sky is a big place. It [32:53] makes sense to utilize it. You can give [32:55] Earth back to humans. You can make [32:57] neighborhoods quieter, safer, less [32:59] pollution, less traffic. You know, you [33:01] can make huge improvements to Earth if [33:03] we can more effectively utilize the sky. [33:05] This is going to require huge [33:08] transformation of how we think about air [33:10] traffic control in the US. And it means [33:12] that [33:13] we need to design it with AI and [33:15] autonomy in mind rather than the way it [33:17] was designed which was in 1950 using, [33:19] you know, pencils and paper and note [33:21] cards and like a human looking out [33:22] trying to watch the airplane. [33:23] >> Are you helping the FAA to design that? [33:26] >> It's really yeah, I mean what needs to [33:27] happen is [33:29] like collaborative innovation is one way [33:31] to put it, right? It's like if one [33:33] company solving this problem for [33:34] themselves is not going to solve the [33:36] problem for the industry. And so we are [33:38] heavily involved in I mean first of all, [33:40] what a key part of the solution we [33:42] believe is aircraft should be talking to [33:45] each other. They should be telling each [33:46] other where they are. They should be [33:48] automatically detecting that hey, [33:50] there's a conflict on the horizon here. [33:51] And so therefore we're going to you [33:53] know, you go up, I go down, right? These [33:54] kinds of these kinds of things. Um and [33:56] our aircraft do that. And we're working [33:59] with other [34:00] kind of you know, other new entrants [34:02] into the airspace with autonomous [34:03] aircraft and autonomous drones to do the [34:05] same thing. To make sure that our [34:07] systems can talk to their systems and we [34:08] can all collaborate to make sure it's [34:10] efficient and safe use usage of the [34:12] airspace. We're also to your point [34:14] Alfred working with regulators, working [34:16] with standards bodies to take some of [34:17] these best practice and innovations that [34:19] that we and others have developed and [34:21] try and make them, you know, broadly [34:23] accepted and utilized so that we can all [34:25] collaborate and we can all, you know, [34:27] safely and efficiently use the airspace. [34:28] >> Cuz you guys are have developed a very [34:30] sophisticated [34:31] >> we because we were when we were [34:32] launching in all these other countries [34:34] like we had to build something from [34:35] scratch and so we built the thing from [34:36] scratch. We provided all this software [34:38] to the civil aviation authorities so [34:39] that they could use it to monitor this [34:41] entirely new class of autonomous [34:43] vehicles in the airspace. Interestingly, [34:45] you know, there are multiple public [34:47] companies in the United States that [34:48] build air traffic control software that [34:50] are worth more than 10 billion dollars, [34:52] right? And so it's like I often look at [34:54] that I mean I think there are many [34:55] companies inside Zip line that are [34:56] likely [34:57] it's like oh that's like a public [34:58] company inside Zip line that is having [35:00] to get built from scratch. We're [35:01] building it because every part of the [35:03] ecosystem we sort of had to build from [35:05] scratch to enable the overall technology [35:07] to flourish. Um [35:09] you know, air traffic control is an [35:10] interesting like the more you learn the [35:12] more disturbing it is. I mean we're [35:14] starting to see the impact you know, you [35:16] read about like you know, a plane [35:18] crashing into a helicopter in DC a few [35:20] months ago. You read about like two [35:21] planes colliding on um I think on a on a [35:25] taxiway in an airport. I don't remember [35:27] where that was a month ago. You're like [35:28] wow, why are all these accidents [35:30] happening? Turns out like 50% of air [35:32] traffic controllers are over the age of [35:33] 45, 20% are are about to retire and [35:38] nobody is going into air traffic control [35:40] as a career path right now in the US. [35:42] And so there's actually a huge labor [35:44] crisis [35:45] uh around these kinds of jobs and and [35:48] so you you you have pressure coming from [35:50] different angles for like transformation [35:51] is required. That we cannot use a system [35:54] that was designed for airspace for [35:56] airspace in the 1950s. The labor isn't [35:58] available to do it even if we wanted to [36:01] and also there is this like giant influx [36:03] of new technology AI and autonomous [36:05] vehicles that are going to require us to [36:06] transform how these systems work. [36:08] >> So you're a hardware company and a [36:09] software company. You build you design [36:11] your own [36:12] >> operations, manufacturing [36:14] >> You design your own parts. You build [36:16] your own aircraft. You write your own [36:18] software. You You do your own [36:20] operations. This look pretty vertically [36:23] integrated company. Talk about the [36:25] benefits of like complete vertical [36:27] integration versus buying component [36:29] parts or [36:31] buying component software and putting it [36:32] all together. [36:33] >> And how you get people who come from [36:35] such different disciplines and domains [36:38] to see eye to eye and work together [36:40] collaboratively. [36:41] >> Yeah. I mean, I think that [36:42] interestingly, you know, this is [36:45] doing it is such an incredible pain in [36:47] the butt that you would never [laughter] [36:49] do it. Like if you, you know, I have [36:50] this flag over my desk that says we do [36:53] this not because it is easy but because [36:55] we thought that it would be easy. [36:57] >> [laughter] [36:57] >> And this is definitely like the [36:58] definition of Zipline, you know, and [37:00] it's such a pain in the butt actually [37:02] that it's almost if you look at the [37:03] history of all these hardware companies, [37:04] they all tried to not do it first. Like [37:07] you can look at the Roadster, right? [37:08] They're like we're going to use a Lotus [37:09] Elise chassis, we're going to buy the [37:11] battery pack from a secondary supplier, [37:13] and we're just going to put the two [37:14] together and it's going to be awesome. [37:15] You know, kind of Roadster lost a lot of [37:19] money and wasn't very reliable, right? [37:20] But like it was it was an important part [37:22] of getting to the Model S. Zipline, when [37:24] we started, you know, Eric knows well, [37:26] we were like buying everything from [37:27] suppliers, we were like, you know, [37:29] paying people to design different parts [37:31] of the system for us or trying to buy [37:33] off-the-shelf stuff, and we crashed [37:35] airplanes. I mean, at test sites, and we [37:37] just we crashed and we crashed and we [37:38] realized, wow, this stuff is like super [37:40] expensive and it's also totally [37:41] unreliable. And so, you know, you know, [37:44] part by part, you're like, all right, [37:45] well, like rip that out. We'll design [37:47] the motor controller from scratch. Okay, [37:48] rip that out. We're going to have to [37:49] describe, you know, design the GPS [37:51] module from scratch. Navigation system. [37:53] So, you know, part by part, you sort of [37:55] like rip it out. And I think there's a [37:57] fundamental realization, probably [37:58] similar to the realization that happened [38:00] that made the Model S possible, is like, [38:01] hey, if we want to build a really great [38:04] specific product in this totally new [38:06] area of technology, we're going to have [38:08] to design every single one of these [38:09] components from scratch to meet the [38:10] specific requirements of this new area. [38:13] You know, you might think, oh, like [38:14] drones, I mean, there are already lots [38:16] of drones cuz DJI makes, you know, [38:18] plastic quadcopters and they make [38:20] millions of them. And like the US buys [38:22] $20 predator aircraft that can fly 100 [38:25] mph. The reality is actually both of [38:27] these systems are very unreliable, and [38:29] nothing is in a level of like [38:31] reliability and safety at unit economics [38:36] that would work for this new industry [38:38] that Zipline was trying to kind of like [38:39] pioneer. And so, we realized we had to [38:41] go build like an automotive-grade [38:43] solution. It has to be super reliable, [38:45] and it has to be extremely [38:47] cost-effective, cuz you're competing [38:48] against cars and motorcycles, which are [38:50] actually really cost-effective. I mean, [38:52] we've had 100 years to make them [38:53] reliable and cheap. So, [38:55] you never do it uh I think [38:58] intentionally. Maybe you just like [38:59] slowly freak out and through desperation [39:02] realize like, "Wow, we got to tear all [39:03] this out, and we got to build it [39:04] all from scratch." The advantage of [39:07] doing it from scratch is like is speed [39:10] and integration. And so, you know, our [39:12] offices, you guys know cuz you've been, [39:14] but like when you visit Zipline's [39:15] offices, I mean, we are all like [39:17] absolutely packed into like, you know, [39:20] sardines into this small building where [39:22] you have firmware engineers sitting next [39:24] to mechanical engineers sitting next to [39:26] autonomy engineers sitting next to, you [39:27] know, cloud infra sitting next to um [39:30] aero-aero acoustics, guidance navigation [39:33] controls, systems engineering, [39:35] manufacturing and everything all [39:37] everybody in one place. And then our [39:38] factory is a 3-minute drive away. And [39:41] so, our team is like on the factory [39:44] floor working, seeing parts get [39:46] integrated into the overall system, and [39:48] then we have our test sites, which are [39:50] just a short drive away. So, you can go [39:51] to the test sites, watch the vehicles [39:53] flying, observe how the system is [39:55] performing. Like, combining all these [39:57] things together means that, you know, [39:59] stuff is always breaking. Stuff's always [40:01] going wrong. As Erica described, the [40:03] advantage is when the thing goes wrong, [40:05] we can basically go straight to the [40:07] person's desk and be like, "You and I [40:08] are pulling an all-nighter tonight." [40:09] >> Mhm. [40:10] >> Whereas if you're Boeing and something's [40:12] going wrong with the battery on the 787, [40:16] you're like going and suing a supplier [40:19] and taking, you know, 2 years to try to [40:21] figure out whose fault it is and like [40:22] it's three layers deep in the rats nest, [40:25] you know, [40:25] cluster of [40:27] like how these procurement deals and [40:29] supply chains work for for aerospace. [40:31] It's why it's so broken. [40:32] >> Yeah, I think back to your your kind of [40:34] question there about getting these [40:36] different discipline folks to work [40:37] together. Honestly, I think it's quite [40:38] easy. [40:39] It's easy [40:41] when you have set up the way that Keller [40:43] just mentioned, right? Like first of [40:45] all, everyone's rowing in the same [40:46] direction. We all have the same goals. [40:47] And when you can ground it in reality [40:49] and it's tangible, then we're all just [40:51] here to solve the same problems, right? [40:52] >> Yeah. [40:52] >> So, we actually with a vertical [40:54] integration, with having a very diverse [40:56] team, we actually cut through a lot of [40:58] the stuff, right? A lot of the things [41:00] that happen where oh, you know, that [41:01] engineer won't tell me what the actual [41:03] source code does cuz they said it's IP [41:05] and so we don't actually know what the [41:07] fault detection looks like and you don't [41:09] have any of that. You just like [41:09] literally go walk over, sit next to the [41:11] person's desk and be like, "Hey, I we [41:13] failed that test. Tell me about how this [41:15] part of the system works." [41:15] >> Mhm. [41:16] >> Oh, cool. Pull up the code. Great. Let's [41:17] look through it. Oh, interesting. You're [41:19] making that assumption? That's not how I [41:20] designed it, right? Cool. Let's get to [41:22] the bottom of it, right? And so this [41:23] idea of just rapid collaboration where [41:26] you just, you know, the manufacturing [41:27] team, the operations team, the [41:29] engineering team are all just like [41:30] really together is the way to solve [41:32] these problems and I I found that it's [41:35] actually not that hard, right? When you [41:36] have that those ingredients, it actually [41:38] makes it, you know, it makes it pretty [41:39] fast and efficient. [41:40] >> And, you know, too, I mean, [41:42] Eric's saying that it really makes you [41:43] realize when you build these like [41:45] complex [41:46] you know, AI and robotic systems that [41:48] combine hardware and software, you [41:49] really appreciate like the deep [41:51] religious truth of how [41:54] dumb requirements usually are. [41:56] >> Mhm. [41:56] >> Question every requirement, which is, [41:58] you know, the number one part of like [42:00] Elon's algorithm they talk about at [42:01] SpaceX. Like question every requirement [42:03] is like is is like so profoundly and [42:06] deeply true. You must have every team [42:09] question every requirement. The [42:10] requirement is always stupid. When you [42:12] And you're like, well, you know, it's [42:13] supposed that you go to this team, that [42:15] team. You like Often you have to dig [42:16] like two levels deep to realize like [42:17] this is But um [42:19] questioning every requirement is a [42:21] fundamental part of like getting through [42:22] this. And then, you know, the other [42:24] thing is um delete the part. The most [42:26] reliable part on an aircraft is the part [42:29] that is not on the aircraft at all [42:30] because you deleted it in the last [42:32] design. That part will never fail. [42:34] And [42:35] you know, you take a lot of inspiration [42:36] from looking at like the Raptor 1, [42:38] Raptor 2, Raptor 3. I'm sure you've seen [42:40] you know, those engines next to each [42:41] other. Actually, a lot of people who [42:42] come to the factory now and get to see [42:44] like the EV3 aircraft. You can see the [42:46] EV2 aircraft, the EV1 aircraft, plus [42:47] like the 10 different hardware versions [42:49] that we built on the initial on the on [42:51] the first version of Zipline's [42:52] technology. Um you would just delete, [42:55] delete, delete. Like, you know, there's [42:57] a huge amount of [42:59] it's really hard to delete things. It's [43:01] an act of courage. [43:02] No one wants to delete the thing. You [43:04] look like an idiot if you delete the [43:06] thing and then like the system can't [43:08] perform or doesn't work because you [43:09] deleted the thing, but like, you know, [43:11] true confidence in like the physics and [43:14] the performance of the system enables [43:15] you to start deleting things. It's a big [43:17] advantage of having like full integrate [43:20] full stack integrated control of all of [43:22] these systems. It makes it possible to [43:24] question every requirement. It makes it [43:25] possible to delete parts. [43:27] >> Yeah, I think first principles thinking [43:28] is a huge part of that. I remember the [43:31] um platform one aircraft, early days it [43:33] had a deployable tail hook cuz how it [43:36] landed. It had this big hook, it's like [43:37] a meter long, that would come down from [43:39] the aircraft and make had a line that [43:40] would catch that and slow the airplane [43:41] down. It's this kind of complicated [43:43] contraption. And we had this idea that [43:46] we should be able to move that [43:47] complexity to the ground systems and [43:50] have the recovery system, the landing [43:51] system, more like an aircraft carrier, [43:53] like crab the airplane, right? We can we [43:55] can you know, put the actuation on the [43:57] basically a robot that goes up and grabs [43:58] the airplane. And we're like, man, [44:00] that's going to make the aircraft so [44:01] much simpler, so much lighter, so much [44:02] more reliable. [44:03] Um [44:04] we didn't have it working yet. And it [44:06] was time to to build that next [44:07] generation of the aircraft and we're [44:10] like, so do we building these next week? [44:12] Do we build them with the meter long [44:13] tail hook or do we delete the tail hook [44:15] and put the 2 cm long tail hook on the [44:17] back and bet that we can get this thing [44:19] working. We got a room like [44:21] delete it, right? Like let's do this [44:22] thing. And so we're like from first [44:24] principles it should work. We can make [44:26] it work. We haven't done it yet, but we [44:27] can do it. And the next couple weeks [44:28] looked like [44:30] myself included a lot of people pulling [44:31] a lot of a lot of late nights getting [44:34] that thing working and sure enough those [44:35] first aircraft came and we caught them [44:37] and landed them. So it's a lot of [44:38] courage [44:39] but that like really being grounded in [44:41] first principles thinking with a tight [44:42] integrated team is is how you do that. [44:45] >> Is there a version of the future in [44:46] which instead of delivering life-saving [44:48] medicine and cheeseburgers [44:51] you're delivering human beings? [44:52] >> Uh-oh. [laughter] [44:54] >> And the board member that's control [44:55] their costs. [44:57] >> [laughter] [44:59] >> I mean, you know, safe, reliable, battle [45:02] tested. I don't know. Seems like [45:04] >> Seems like if we're going to liberate [45:06] ourselves from the tyranny of streets [45:07] it's a pretty decent solution. [45:09] >> Gosh, I I I think I agree with you. [45:11] Um I think that [45:14] >> He's going to come to you and ask for [45:15] another billion dollars. [45:16] >> [laughter] [45:18] >> Um I think you know, [45:21] a couple thoughts. Like one is that I [45:23] think Alfred knows I'm measured in the [45:25] way I answer that question because to be [45:27] clear like [45:29] you know, building a new infrastructure [45:31] layer for the planet that can deliver [45:32] packages as efficiently as the internet [45:34] moves information is going to be one of [45:36] the biggest companies on earth. Like [45:38] it's a huge opportunity and we [45:39] definitely want to stay like humble and [45:43] paranoid about how super hard that's [45:44] going to be. The level of execution for [45:46] us to scale the way we want to scale [45:47] over the next couple years. [45:49] And [45:50] you know, to put into perspective I I [45:52] described this goal of getting to a [45:53] million deliveries a day in the in the [45:55] very near future. [45:56] We now have many partners who are at [45:59] each asking to buy a million deliveries [46:01] a day of capacity from Zipline [46:03] in the last few months. [46:04] And so our operating plan has now become [46:07] our unit of sale. That's a pretty crazy [46:10] realization and it's leading us, you [46:12] know, we had originally built then we [46:13] we'd sized the entire factory to build [46:16] uh 20,000 aircraft a year. That was what [46:18] about what was required for a million [46:19] deliveries a day. Like [46:21] all of this is kind of being thrown [46:22] we're realizing the market is way [46:24] bigger. And one one thing, you know, [46:27] when you look at this, you know, totally [46:29] hyperbolic curve that I think I showed [46:31] you only a few months ago of like, you [46:32] know, the what our total flights have [46:35] total daily flight volumes have done [46:36] over the last [46:38] 16 months, the level of complexity of [46:40] all the different systems that are [46:41] required to basically like stay on that [46:43] track [46:45] um is quite high. Yeah. But you know, [46:47] there are 5 and 1/2 billion instant [46:50] deliveries being done [46:52] by humans [46:53] in the United States [46:55] every year. [46:56] And that's what you know, we're we're [46:57] using a 4,000 lb gas combustion vehicle. [47:00] >> It's like [47:01] >> Yeah, half an hour to an hour. [47:02] >> Exactly. It's good marketing that it's [47:03] called instant, but yeah, exactly. And [47:06] uh you know, 30 minutes, 45 minutes, an [47:07] hour, you know, [47:09] uh significant percentage of the drivers [47:11] report eating some of the food that [47:12] they've delivered in the last month, [47:13] like more than 50%. Um there are [47:16] significant safety, you know, concerns [47:18] associated with these kinds of delivery. [47:20] But 5 and 1/2 billion instant [47:22] deliveries, what we're realizing when [47:23] you look you know, Zipline is now at [47:24] massive scale in Dallas and we're we're [47:26] now launching four more metros in the [47:28] next 4 months. Um [47:30] when you just look at Dallas, if you [47:32] were to extend the buying behavior that [47:34] we're observing from Zipline customers [47:36] in Dallas to the rest of the United [47:37] States, there would be 55 billion [47:39] instant deliveries happening, not five. [47:41] >> Wow. [47:42] >> 55 billion. [47:43] >> Yeah. There's a huge market expansion. I [47:45] think it's similar to how people looked [47:47] at Uber when they were launching in San [47:49] Francisco and they're like, oh, even if [47:51] Uber gets to be 33% of the taxi market [47:53] in San Francisco, it's only going to be [47:54] a $15 billion company. And obviously [47:56] what they missed is like Uber's now 10 [47:58] times the size of the taxi market. You [48:00] know, like if you make something [48:02] more convenient [48:04] and less expensive and a better product [48:06] experience, [48:07] people are going to consume a lot more [48:09] of it. We are clearly seeing customer [48:11] behavior where customers order every day [48:12] rather than a couple times a month. Um I [48:14] mean, I met a grandma the other day [48:16] who's ordered 350 times from Zipline in [48:17] the last year. [48:18] >> [snorts] [48:19] >> She's 80 years old. [48:19] >> Amazing. [48:20] >> Um actually nursing homes are like big [48:23] Zipline like they're like big demand [48:25] centers for Zipline. It makes sense. [48:26] Like and actually it's funny old people [48:28] as like maybe [48:31] being, you know, not capable of using [48:32] technology. They're all like living on [48:34] their iPhones, you know, like they [48:36] they're probably doom scrolling actually [48:37] which is maybe not a good thing, but [48:38] like um they are very comfortable using [48:41] like, you know, Apple ID, [48:43] um Apple Pay or Face ID, Apple Pay and [48:46] just ordering and having it delivered [48:47] directly to them. Um so, there are [48:50] definitely not enough humans in the [48:52] United States to do 55 billion [48:55] deliveries. [48:55] >> Yeah. [48:56] >> The only way we're going to be able to [48:58] serve this kind of demand is with [49:00] automated systems and there's definitely [49:01] not enough roads. And when you look at, [49:03] you know, traffic in most of our major [49:05] cities, you're like, "Oh, can we just [49:06] like maybe double the number of cars on [49:08] the road so that we can do way more [49:09] deliveries?" [49:10] It obviously doesn't work. We actually [49:12] need to be taking cars off the roads if [49:14] we want to like enable human growth and [49:16] flourishing. And so, I think it, you [49:18] know, this this change is it is [49:20] inevitable. [49:21] >> So, how many flights are you doing a day [49:24] now and how many will you do in a month? [49:26] >> Zipline is now doing almost 5,000 [49:28] flights a day um and you know, we're [49:31] anticipating exiting this year at about [49:33] 30,000 flights uh a day. And our goal is [49:36] to get to a million flights a day as [49:38] fast as humanly possible which we we [49:39] expect to achieve in the very near [49:40] future. Like all of the supply chain [49:43] manufacturing and capacity decisions [49:44] we're making right now are designed not [49:47] just to get us to a million deliveries a [49:48] day, but also um [49:50] accelerate past that. [49:52] >> The things that are interesting to think [49:53] about on the unit economics front is [49:54] like whenever we meet hardware companies [49:56] and you always talk to them about like [49:57] how much do you think the system is [49:58] going to cost? And they're always like [50:00] it's going to cost X and you're like [50:02] cool, it's going to cost 10 X just so [50:03] you know. Like when you build it, it's [50:05] going to cost 10 X. [50:05] >> That's your advice to founders. [50:06] >> That's my advice to founders is like for [50:08] hardware companies like if I cuz you [50:09] know, I'm I'm like try to you know, be a [50:11] good seed investor and pay it forward [50:13] and stuff and like you're always meeting [50:14] these founders and they're always like [50:15] it's going to cost this much. I'm like [50:16] cool, just like [50:18] assume it's going to cost 10 times that [50:19] and like does it work and what would you [50:20] do if it cost 10 times that? And I we're [50:22] speaking from experience. Like when we [50:23] launched our system in 2016, we were [50:26] charging $30 a delivery to deliver a [50:28] blood transfusion over 80 to 100 miles [50:31] and that was like cost comparable. And [50:34] so we that's what we signed the contract [50:36] for and we thought that we're going to [50:37] launch a system that cost about $30 a [50:38] delivery. How much do you think it cost [50:40] when we launched? [50:41] >> $300. [50:41] >> Yeah, [50:42] >> [laughter] [50:42] >> it cost $300 a delivery and Alfred was [50:44] surprisingly chill about it. [50:46] Um and you know, we were like all right, [50:48] we got work to do. And so you know, the [50:49] next year we got it to like 120 and then [50:51] the next year we got it to 75, then the [50:53] next year we got it to 40, and then we [50:54] got it to 28, then we got it to 18. It's [50:57] now 12 [50:58] for the kind of you know, the the long [51:00] range technology that we operate outside [51:01] the US. Right now what's happening this [51:03] summer is the the fully burdened unit [51:06] economics of these systems is [51:08] just now in the process of falling below [51:11] the cost of using cars to deliver [51:13] things. And I think it's it is a cool [51:14] moment that I think most people don't [51:16] really realize. It's happening quietly. [51:18] Like you're not reading about this in [51:19] the New York Times or whatever, but um [51:21] you know, [51:23] I think that this this thing is [51:24] happening [51:26] in the next month or two that is going [51:28] to have a big impact on the world and [51:29] how the world looks and and how people [51:31] how people most people normal people [51:32] even live their lives because it is now [51:34] more cost effective to use a robot in [51:37] logistics [51:38] than it is to use a human. And that's [51:40] really good news for the environment. [51:42] It's really good news for neighborhoods [51:44] that are going to get quieter and safer [51:45] and less traffic, less pollution. And [51:47] it's really good news for customers cuz [51:49] you can get things way faster and more [51:51] reliably. And and and and [51:53] for less expensive. You know, our [51:55] customers love like there are obviously [51:57] so many cool things about the system [51:59] that, you know, you can talk about and [52:00] and that you see customers taking [52:01] advantage of, but like no tip [52:03] {exclamation point} {exclamation point} [52:05] [laughter] {exclamation point} is like a [52:06] big, you know, that's probably the [52:07] number one comment. I think that um [52:10] customers love [52:12] not having to feel guilty and being able [52:14] to just have a system that they know how [52:15] much it's going to cost. [52:17] >> Well, thank you, Keller and Eric, for [52:19] being here with us. I thought you were [52:21] going to say it it takes longer than you [52:23] thought, not 10x more than it costs. [52:25] But, anyway. [52:27] >> Yeah. [52:27] >> That's a great great way to end. [52:29] >> It does also take a lot longer. I mean, [52:30] I think, you know, [52:32] the memo that Sean wrote here at Sequoia [52:35] a few years ago, I think, is like is [52:37] deeply true. [52:38] I don't know if he'll ever publish that [52:39] publicly or if it'll be allowed, but I, [52:41] you know, I do think, you know, suffice [52:42] it to say, there's an internal Sequoia [52:44] memo that has had a big impact on me [52:46] talking about A, why hardware companies [52:49] are going to be some of the most [52:50] impactful companies for humanity's [52:52] progress over the coming decades, and B, [52:55] why it's super hard to get those [52:57] companies off the ground and fundraise [52:59] for them, and C, like how, you know, [53:01] investors should think about funding [53:02] those kinds of companies. It's [53:03] interesting like when you look at the [53:04] world today to see [53:06] I mean, it wow, how fast the world [53:08] changes. Cuz think about it, we spent 10 [53:10] years being the freaking black sheep. [53:13] Like, hardware company, no, thank you. [53:14] Like, like, [laughter] let's invest in, [53:16] you know, SaaS, let's invest in margins. [53:18] Like, this this is where the whole [53:19] future was and like, you know, iPhone [53:21] apps, blah blah blah. So, [53:23] I don't know. I guess I feel like um [53:25] >> Now you're cool. [53:25] >> Bane. Remember Bane and Batman? What [53:27] does he say? [53:28] Like, you adopted the darkness. I was [53:30] born in it. Like, man, we built a [53:32] robotics company for 10 years before [53:33] building a robotics company was a cool [53:34] thing to do. [53:36] But, um you know, I do think that [53:37] especially important for like US [53:39] competitiveness and just for our ability [53:42] to like build the future that we'd be [53:44] really proud to hand to our kids and to [53:46] our grandkids, and to build the sci-fi [53:48] version of the future that we were all [53:50] promised, like we got to get good at [53:51] building stuff again. And we got to get [53:53] good at building not just you know, uh [53:56] vehicles and hardware, we got to get [53:57] good at building infrastructure. Like [53:59] we're depending on the crumbling [54:01] infrastructure that our grandparents [54:02] built for us. [54:04] I read the other day [54:05] the you know, we just installed these [54:07] like anti-suicide nets on the Golden [54:08] Gate Bridge. Have you guys read about [54:09] that project? It cost more to install [54:12] those nets than it cost our grandparents [54:14] to build that bridge. [54:17] >> I believe it. [54:17] >> So anyway, we get really excited just [54:19] like we think the future like promising [54:21] future is like we should be able to [54:22] build infrastructure you know, we got to [54:24] we have to be interested in it um [54:26] and I think people have to have the [54:28] stomach for it. [54:29] And we have to learn how to manufacture [54:31] and run complex supply chains again. And [54:33] we have to be you know, bold and like [54:35] believe in sci-fi versions of the future [54:37] if we're going to build them. [54:39] >> Awesome. Let's end it at at that. [54:41] Believe in the sci-fi future. [54:43] >> Yeah. [54:44] >> Thank you guys for being with us. [54:45] >> Thank you. [54:46] >> Thanks for inviting us. [54:50] >> [music] [55:03] [music] [55:12] [music]