Zipline Cuts Drone Delivery Cost to $12 to Undercut Car Transport

Zipline将无人机配送成本降至12美元，低于汽车运输 随着Zipline朝着每日100万次飞行规模迈进，自主空中配送正变得比人工驾驶更便宜。 Zipline领导人Keller Rinaudo Cliffton和Eric Watson回顾了公司从2016年在卢旺达开创远程血液配送，到运营全球最大商业自主物流系统的历程。他们指出实体无人机仅占其整体解决方案的15%，并重点介绍了冗余双飞控计算机故障转移架构、700个定制设计的零部件、严苛的硬件极限压力测试以及AI驱动的空中交通管制。对话还探讨了监管审批如何从1对1的飞手比例过渡到单个机队指挥官监管100架飞行器，同时单次配送成本从300美元骤降至12美元，低于传统人工汽车配送。 Zipline 在卢旺达启动血液配送时验证了产品市场契合点（PMF），医护人员对 24/7 全天候服务的迫切需求推动其迅速转型为不间断运营的生命线物流体系。 (0:00) 物理无人机仅占自主物流系统的 15%，其余 85% 依赖辅助软件、调度和地面运营；Zipline 已实现 1.4 亿英里零事故自主飞行，使服务地区的孕产妇死亡率降低了 51%。 (5:00) 为了以消费级供应链成本实现波音级的安全性，Zipline 采用了双飞行计算机加仲裁节点的冗余架构，并自主设计了超 700 种定制零部件以经受严苛的破坏性极限测试。 (10:00) 随着 FAA 监管的推进，Zipline 的监控人机比从最初的 1:1 扩大至 1:100，单个“舰队指挥官”已能同时监管 100 架自主飞行器的空中网络。 (24:30) 其单次配送成本已从早期的 300 美元大幅降至 12 美元，迎来历史性拐点：自主无人机配送的完全成本历史上首次低于传统汽车的人工配送。 (49:58)

Zipline argues that autonomous aerial logistics has transitioned from high-risk medical relief into a scalable infrastructure system capable of undercutting human-driven transport on cost and safety. Achieving this required abandoning off-the-shelf components in favor of vertical hardware integration, fail-safe redundant engineering, and automated airspace management. Operational Lessons and the 15% Rule (5:00) Regulatory restrictions on beyond-visual-line-of-sight flight in the United States initially forced Zipline to launch in Rwanda in 2016, delivering blood supplies to hospitals. The company's initial contract covered 21 hospitals, but early operational and technical hurdles limited service to a single hospital for the first nine months. These challenges revealed that the aircraft hardware represents only 15% of the overall solution, while 85% consists of supporting software, maintenance protocols, national aviation coordination, and round-the-clock logistics. In-House Engineering and Redundant Aircraft Architecture (10:00) After discovering that third-party consumer and military drone components were either too fragile or economically unviable, Zipline vertically integrated its manufacturing, engineering 700 custom components across 43 sub-assemblies. The aircraft relies on dual flight computers that calculate trajectory simultaneously, regulated by an isolated arbiter node that switches control if the primary computer experiences faults mid-flight. To survive environmental extremes ranging from -25°C to 49°C and solar flares that disrupt standard satellite signals, the company relies on custom navigation systems and stress-testing to failure. Following principles of part deletion, engineers also replaced a complex one-meter mechanical tail hook with a two-centimeter stub by transferring capture mechanisms to robotic ground systems. Safety Benchmarks and Increasing Operator Leverage (18:30) Across 140 million commercial autonomous miles, Zipline reports zero safety incidents—contrasting with the roughly 600 accidents typical of passenger automobiles over identical distances. Prompted by Sequoia partner Alfred Lin, the company shifted its benchmark from matching human driving safety to targeting a record twice as safe as Waymo. Precision delivery is executed from a 100-meter hover, where an active-maneuvering foam payload pod powered by an NVIDIA GPU scans the terrain and lands packages within an 18-inch target. These demonstrated safety margins enabled the Federal Aviation Administration to relax staffing ratios from one pilot per drone in 2020 to one Fleet Commander overseeing up to 100 autonomous aircraft. Cost Trajectory and Air Traffic Modernization (30:00) Zipline's fully burdened long-range delivery cost declined from $300 at launch in 2016 to $12 per delivery, dropping below the operating cost of human-driven ground transport. While the network took a decade to log its first million cumulative deliveries, the firm is preparing to handle one million deliveries per day, projecting a potential US instant delivery market of 55 billion annual orders based on early Dallas ordering patterns. Leadership asserts that sustaining this volume—amounting to 40 to 80 times the current daily volume of all US commercial airlines combined—will require replacing legacy, human-constrained air traffic control with automated, direct aircraft-to-aircraft conflict resolution. While Zipline has validated its safety architecture and reduced per-delivery costs across 140 million miles, reaching planetary scale depends on completely overhauling legacy air traffic control and proving that consumer demand for instant aerial delivery outside initial metropolitan test beds matches its projected 55-billion-order market.

Zipline创始人Keller Rinaudo Cliffton与工程团队指出，自主无人机配送的核心并非单纯制造飞行器硬件，而是构建涵盖自研软硬件、航天级冗余架构与空域协同的端到端自动化物流基础设施。在累计商业飞行1.4亿英里且保持零安全事故的记录下，该系统单次配送成本已跌破传统人力汽车，并推动着空域管理体系向AI自主控制转型。 早期验证：从无人机转向全栈物流系统 (0:00) Zipline在卢旺达提供血液应急配送初期发现，一线医护人员的核心诉求并非无人机硬件参数，而是打破12小时轮班限制的24/7全天候即时服务。创始团队据此提出“15%法则”：物理飞行器仅占整体解决方案的15%，其余85%依赖辅助软件、地勤维护、库存调度、医疗协同与国家航空整合。聚焦整体系统可靠性后，公司首年内便实现全天候运行，服务范围从最初9个月仅能维持1家医院，快速扩展至跨越8个国家的5000多家医疗机构。 根据宾夕法尼亚大学的研究数据，Zipline在服务区域内帮助产妇死亡率降低了51%，合作方估算其每年挽救1万至1.2万人的生命。截至目前，该系统累计完成了1.4亿英里商业自主飞行且保持零事故，里程相当于环绕全美所有公路30余次。此外，公司与美国国务院建立了5.5亿美元的商业外交合作，通过输出自动化基础设施替代传统的依赖型非政府组织（NGO）物资援助。 全栈垂直整合：低成本实现航天级冗余 (10:00) 由于消费级四轴无人机故障频发，而数千万美元的军用无人机又缺乏商业物流可行性，Zipline选择全面垂直整合，自主设计了700个独特零部件与43个主要子组件。自研范畴覆盖高推重比定制电机、电池管理系统以及搭载NVIDIA GPU的智能配送舱。为了检验物理极限，所有组件均需在-25°C至49°C的极端环境仓及高加速寿命测试（HALT）中经历震动与风洞破坏性试验，贯彻“测试直至失效”的工程准则。 在航电安全上，Zipline借鉴波音777架构开发了双主飞行计算机系统，两台计算机并行处理传感器数据，由独立的仲裁节点监控健康状态并分配控制权。近期一次投递任务中，主计算机在卸货后突发微小异常，系统瞬时在空中切换至备用计算机并安全返航。同时，团队严格践行“质疑需求并剔除部件”的第一性原理，例如将早期平台长达1米的可展开尾钩直接缩减为2厘米短桩，将拦阻复杂性转移至地面机器人机构，大幅消除了机载机械故障源。 厘米级空投与监管突破：重塑人机杠杆 (20:00) 为确保在居民区安静且安全地投递，Zipline主机保持在100米高空悬停，将旋翼噪声降至接近树叶沙沙声。实际卸货由轻质泡沫配送舱通过系留缆绳下降完成：系统先借助RTK GNSS实现厘米级空间定位，随后配送舱机载NVIDIA GPU与感知算法在下落中主动在XY轴抵抗侧风，避开桌椅或杂物等地面障碍，在触地释放包裹后迅速收回。红杉资本合伙人Alfred Lin曾提出安全基准不应对比人类驾驶汽车，而应达到Waymo自动驾驶汽车的两倍安全性，工程团队随即将该挑战确立为核心安全指标。 在空域准入方面，Zipline自2020年起与美国联邦航空局（FAA）深度合作。地面人员的监管配比逐步从最初的1人监控1架，跃升至1:3、1:20直至目前的1:100，操作员角色也正式更名为“舰队指挥官”（Fleet Commander）。指挥官仅在系统发出恶劣天气或空域冲突警报时介入处理，实现了单人运营杠杆的跃升。 成本拐点与空域重构：迈向百万级自动化 (30:00) 根据Keller Rinaudo Cliffton提供的数据，卢旺达长途平台的单次全口径配送成本已从2016年上线时的300美元，经多代优化逐步降至120美元、75美元直至当前的12美元，历史上首次实现了自动化飞行配送成本低于传统汽车人力配送，并消除了小费等隐性支出。Keller表示，目前全美依赖燃油车的人工即时配送年订单量约为55亿单，而参考达拉斯运营区中老年社区等高频日常订购习惯，若该模式推向全美，潜在即时配送需求将达到550亿单。 目前Zipline日均飞行近5000架次，年底预计突破3万架次，长期目标为每日执行100万次配送，该体量将达全美所有民航航班总和的40至80倍。面对美国50%航空管制员超过45岁且20%临近退休的人员断层危机，1950年代成形的传统人工空管系统已无法承载未来密度。Zipline正协助监管机构推动基于机对机直接通信与AI动态规避的自动化空管协议，同时规划无人机直接从旧金山南部制造工厂起飞交付运营，打造全天候自主运转的立体运输走廊。 Zipline验证了全栈硬件垂直整合与自动化规模效应在实体物理世界的落地路径，其单次配送成本与安全数据确立了空中物流替代地面人力车队的商业可行性。然而，当配送规模从单日万次迈向百万次大关时，罕见的百万分之一极端异常将转化为日常必现的工程挑战，其长远扩张不仅考验底层供应链的自研韧性，更有赖于国家航空监管体系能否彻底打破半个世纪前的人工空管架构，真正接受由算法原生的分布式空域调度协议。

## Background 该内容由红杉资本（Sequoia Capital）合伙人 Alfred Lin 与 Pat Grady 于 2026 年 7 月 7 日在 RSS 与网络播客平台发布。节目简介重点介绍了全球最大商用自主运行系统 Zipline，围绕其累计完成 1.4 亿英里自主飞行且零安全事故的系统级工程、从卢旺达医疗血液急救起步的演进历程，以及配送成本从 300 美元降至 12 美元并全面跌破汽车配送门槛的技术商业拐点展开深入探讨。 ## TL;DR Zipline 联合创始人 Keller 与系统工程负责人 Eric 指出，无人机机体仅占整体复杂度的 15%，其余 85% 则是涵盖空管对接、仓储调度与极致冗余的复杂软硬件基础设施。面对早期美国本土严苛的超视距飞行监管限制，团队逆向出海卢旺达切入急救刚需，在将当地产妇死亡率降低 51% 的同时，攻克了太阳耀斑干扰、双飞控故障转移与百米悬停边缘计算微调投递等核心技术闭环。通过推行“1对100”的舰队指挥官监管协议以及全栈自研的去官僚化垂直整合模式，Zipline 彻底颠覆了传统航空外包研发的低效泥潭。伴随单次配送成本从 300 美元大幅优化至 12 美元，全包无人机物流经济模型首次低于传统汽车配送，驱动公司向日均百万单的全自主物流网络加速迈进。 ## Atmosphere & style 整场对话洋溢着硬核工程的严谨务实与底层硬件创新的真实痛感，毫无空洞的概念包装，全是一线踩坑淬炼出的血泪实战经验。红杉资本主持人以董事会成员与思考伙伴的敏锐视角切入，在直面董事会关于“比 Waymo 安全两倍”的激烈交锋与长尾故障挑战中，展现出高密度的思维激荡。叙事从卢旺达至暗时刻的通宵修障一路铺陈至十载坚守硬件“黑羊”的终迎破晓，既有理工团队对物理定律和第一性原理的敬畏，又散发着摆脱纯软件依赖、亲手重塑物理世界造物能力的磅礴野心与感染力。 ## Bulletin - `07:25` Zipline 累计商用自主飞行里程突破 1.4 亿英里并保持零安全事故，独立医疗评估显示其服务使进驻区域产妇死亡率直降 51%，年均挽救超 1 万条生命。 - `07:25` 商业外交协议达成：美国国务院与 Zipline 签署 5.5 亿美元合作，以输出高自主度软硬件及本土造血机制替代传统低效无偿援助。 - `50:00` 单次配送全包成本实现断崖式下降，从 2016 年卢旺达初期的 300 美元一路压降至目前的 12 美元，正式跌穿传统汽车送货成本底线。 - `13:30` 全栈垂直整合推进极致制造：机身拥有 700 个自研零件与 43 个子总成，在 FAA 监管框架下实现了 1 名“舰队指挥官”协同调度 100 架自主飞机的运力体系。 ## Key takeaways - `00:00` 拒绝被定义为“无人机公司”，Zipline 的终极定位是构建接近“瞬间移动”体验的端到端全球自动化物流基础设施。 - `02:30` 在早期遭遇美国本土严苛禁飞监管（BVLOS）后，团队转向卢旺达以国家级急救血液配送切入，依靠刚需撬动政策豁免与商业闭环。 - `04:15` 在全套端到端物流交付方案中，飞行器硬件本身仅占 15% 复杂度，剩余 85% 涵盖空管直连、医疗对接、仓储库存及动态调度等全栈基建系统。 - `06:05` 客户反馈推动运营机制重塑，Zipline 迅速从日间 12 小时作业演变为 24/7/365 全天候无休响应，找到了极度强烈的 PMF。 - `10:00` 面对太阳耀斑和高空恶劣工况，Zipline 采用由轻量仲裁器管理的主备双飞行计算机架构，以消费级元器件成本达成传统航空级冗余安全。 - `13:30` 全机包含 700 个自研独立零件，并推行彻底测到结构损坏的高加速寿命测试（HALT），温区覆盖 -25°C 至 49°C 的严苛工况。 - `17:00` 确立“比 Waymo 还要安全两倍”的年终安全考核基准，在 1.4 亿英里自主运行中交出完全零伤亡、零责任事故的成绩单。 - `20:00` 新一代平台采用母机在 100 米高空悬停、内嵌 NVIDIA GPU 算力的轻量投递舱垂直下放的架构，兼顾了厘米级精准投放、极致静音与地面人身安全。 - `26:15` 空域管控角色实现跨越式升级，远程调度员从 1:1 单机监管跃升至 1:100 并被 FAA 官方认证为“舰队指挥官”，极大重塑了单位经济模型。 - `28:30` 系统从耗时约十年完成首个 100 万单，全面跨入“单日交付 100 万单”的超大规模扩张期，使得百万分之一概率的长尾故障成为每日必现的常态挑战。 - `35:10` 全美空管人员正面临 50% 超过 45 岁的人力断层危机，Zipline 倡导利用机对机（V2V/D2D）自主协同规避来重构 1950 年代遗留的人工空管体系。 - `36:55` 因外购零部件频频炸机而被迫走向全面垂直整合，南旧金山办公室将多学科工程团队挤在同处，3 分钟直连工厂与试飞场形成极速物理反馈回路。 - `42:08` 奉行第一性原理与马斯克式的“删减部件”信条，早期平台果断删去 1 米长的沉重尾钩，将复杂度转嫁给地面主动机械抓取系统以极致减轻机载负荷。 - `44:48` 即时配送将迎来类似当年 Uber 的“诱发需求”大扩容，全美市场有望从每年 55 亿次跃升至 550 亿次，无人机成为破解司机短缺与地面拥堵的唯一途径。 - `50:00` 硬件创业必须警惕“10倍成本法则”，Zipline 上线初期单次配送全包成本曾高达 300 美元，历经多代技术与工艺迭代才打磨至 12 美元。 - `51:40` 全包无人机物流成本首次全面低于传统汽车送货，且“告别小费”的透明定价体验彻底消除了终端消费者的社交焦虑与心理负担。 - `52:45` 反思现代社会造物能力的退化，Zipline 以十年“黑羊”坚守呼吁重拾物理世界基础设施建设与高端制造实力，将科幻未来照进现实。 ## Detailed walkthrough ### `00:00` 不做“无人机公司”：追求近乎瞬移的自动化物流体验 红杉资本的合伙人 Alfred Lin 与 Pat Grady 开启对话时提到，Zipline 早在 2011 年成立、2014 年转型，远在大语言模型（LLM）和现代机器人基础模型大爆发之前就已深耕硬件自主系统领域。尽管 Zipline 如今已是全球规模最大的商用自主无人机系统，联合创始人 Keller Rinaudo Cliffton 却明确表示，团队非常抗拒被称为一家“无人机公司”。他强调，Zipline 始终坚持极度的“以客户为中心”（customer-obsessed），而现实是客户根本不在乎无人机本身。 Zipline 真正的终极愿景是为地球打造一套全自动化的物流体系，无限逼近“瞬间移动”（approximate teleportation）。对于如今依赖 Zipline 服务的终端用户来说，幕后运行着何种高科技并不重要，他们真正在意的是体验本身：打开手机 App，浏览琳琅满目的品牌和喜爱的餐厅，点击下单，五分钟后物品便平稳送达。Keller 认为，一旦将重心完全放在用户体验而非某种特定载具上，技术的定义与边界就会彻底改变。 - **成立背景**：公司始于 2011 年，2014 年完成战略转型，全面切入自主硬件与物流系统研发。 - **核心定位**：非单纯的硬件或无人机制造商，而是提供近乎瞬间移动体验的全球自动化物流网络。 ### `02:30` 绕道卢旺达：从“在美国属于违法”到拯救生命的切入点 回溯到 2014 年前后，Keller 坦言在那个时期创立一家机器人公司几乎被视作“最愚蠢的决定”。当时他和系统工程与安全负责人 Eric Watson 都只有二十三四岁，拿着一套“速度快 10 倍、成本减半、零排放”的全球自主物流构想四处向投资人推介，却遭遇了巨大的制度阻碍。投资人抛出的第一个致命问题往往是：“这在美国不是违法的吗？” 当时的监管法规严禁超视距飞行（BVLOS），甚至在多数场景下无人机商业飞行根本不被允许。面对投资人“既然违法我们绝不投资”的集体闭门羹，Zipline 被迫做出逆向抉择：既然美国无法落地，那就去全世界最迫切需要这项服务、价值能被无限放大的地方。2016 年，Zipline 远赴卢旺达开展血液及急救制品的跨空配送。拯救生命的刚性诉求，促成了当地政府以前所未有的深度与初创团队紧密合作，甚至特批豁免并联合制定了全新的监管法规，为无人机生命线破例放行。 ```mermaid flowchart TD accTitle: Zipline 早期出海破局路径 accDescr: 展示 Zipline 在美国面临法规壁垒时转型进入卢旺达的决策流程 subgraph US_Barrier ["美国本土困境 (2014)"] A["构想自动化物流系统"] --> B{"监管法规限制"} B -->|"禁止超视距飞行 (BVLOS)"| C["被投资人判定为违法项目"] end subgraph Rwanda_Pivot ["破局切入：卢旺达 (2016)"] C --> D["寻找超高价值刚需场景"] D --> E["国家级急救血液与医疗物资配送"] E --> F["政府深度协同与监管特批豁免"] F --> G["成功跑通商业与生命救援闭环"] end ``` ### `04:15` 硬件只占 15%：首年惨败与真正的全栈物流系统 2016 年在卢旺达落地伊始，团队满心以为只要造出一台设计精良、功能强悍的酷炫飞行器，问题就迎刃而解了。然而现实却迎头痛击——第一年的运营堪称“彻底的灾难”。Zipline 当时与卢旺达官方签订了服务 21 家医院的合同，但在前整整九个月里，他们拼尽全力也只能勉强维持其中 1 家医院的运转。 负责系统工程与安全的 Eric Watson 对那段至暗时刻记忆犹新。当时他在卢旺达连轴转，几乎没有合眼的机会；而即使回到美国，每到半夜零点，卢旺达的配送中心一开始运营，各种系统接连崩溃报警，团队只能在深夜惊醒，通宵达旦、牺牲周末排查修障。这段惨痛教训让团队彻底清醒：在整个商业交付方案中，无人机硬件本身仅仅占到 15% 的复杂度，剩下的 85% 全是看不见的系统级工程挑战。 ```mermaid flowchart LR accTitle: 自动化物流系统的复杂度构成 accDescr: 区分无人机硬件与外围关键运营软件系统的占比 subgraph Solution ["端到端自动化物流体系"] direction TB subgraph Hardware ["硬件载具 (15%)"] H1["无人机机体设计"] H2["机载飞控与动力"] end subgraph Infrastructure ["全流程基建与系统 (85%)"] S1["库存管理与仓储调配"] S2["国家民航局空管系统直连"] S3["国家医疗体系深度集成"] S4["全天候维护检修系统"] S5["订单调度与需求预测管理"] end end ``` - **初始履约落差**：签约承诺覆盖 21 所医疗机构，前 9 个月仅能艰难维持 1 所机构运作。 - **系统核心组成**：除去 15% 的机体硬件，其余 85% 涵盖了精密温控与库存管理、自动化维护检修、民航空管对接接入、国家级医疗下单网络及动态运力需求管理。 ### `06:05` 从 12 小时到 24/7/365：寻找真正的产品市场契合度（PMF） 在卢旺达服务初期，Eric 作为一线工程师常常深入各家医院和化验室走访医护人员，满怀期待地询问他们对这套高科技无人机有什么改进建议。然而，医生与化验员们给出的反馈让他大吃一惊。对方根本不关心飞行器飞得如何，而是直截了当地反问：“病人生病是不分昼夜的，为什么你们每天只营业 12 个小时？” 这一灵魂质问瞬间点醒了团队：哪怕当时的原型系统千疮百孔、故障频发，它却真真切切解决了一个人命关天的核心痛点。当客户最核心的诉求是“希望得到你们更多、更长时段的服务”时，团队意识到他们已经找到了极度强烈的“产品市场契合度”（PMF）。团队随即咬牙攻坚，在第一年内就强行将系统运营能力升级为 24/7 全天候无休。如今，即便在圣诞节当天，Keller 也会致电全球每一个配送中心慰问坚守岗位的员工，整个网络已全面进入 365 天无间断运行模式。 - **规模跨越**：服务节点从最初挣扎维系的 1 家，逐步突破至 20 家、500 家，如今已常态化覆盖全球 8 个国家的 5000 余所医院与基层医疗机构。 - **全时运作**：打破传统物流受限于光照和气候的壁垒，实现全年全天候 24/7/365 紧急生命物资响应。 ### `07:25` 1.4 亿英里与 5.5 亿美元：商业外交与全球最大自主运行系统 历经十载打磨，Zipline 已悄然构筑起全球规模最大的商用自主运行系统。其累计商用自主飞行里程刚刚跨越 1.4 亿英里（约合 2.25 亿公里）。为了形象化这一天文数字，Keller 打了个比方：这相当于把全美所有道路来回开上 30 遍以上，或者往返太阳一整趟。宾夕法尼亚大学（UPenn）发布的一项独立医学评估显示，在 Zipline 进驻的区域内，产妇分娩死亡率断崖式下降了 51%；综合各方合作伙伴测算，这套网络每年能直接拯救 1 万至 1.2 万条生命。 伴随这一巨大社会效益，Zipline 迎来了地缘与商业维度的重大转折：与美国国务院签署了价值 5.5 亿美元的合作协议。在美国国际开发署（USAID）架构调整的背景下，美国政府转向推行被称为“商业外交”（commercial diplomacy）的新战略。长期以来，发展中国家对传统 NGO 提供的无偿低效救济深恶痛绝，因为这种模式极易滋生技术依赖并扼杀本土经济活力；他们真正渴求的是高薪就业机会、高精尖科技基础设施与自主创业生态。通过这一战略，美国将顶尖的 AI 与自主机器人基础设施直接引入新兴市场，在挽救生命的同时，加速了当地经济造血能力，也从战略层面锁定了美国在高端机器人制造与自动化领域的长期主导地位。 - **运行体量**：商业自主运行里程突破 1.4 亿英里，安全零责任事故，远超任何现存自动驾驶出租车（Robotaxi）车队。 - **医疗成效**：宾夕法尼亚大学研究证实产妇死亡率直降 51%，年均挽救生命 10,000 至 12,000 人。 - **战略合作**：获得美国国务院 5.5 亿美元支持，以“商业外交”替代传统低效无偿援助，全面向海外输出高自主度软硬件体系。 ### `10:00` 意料之外的挑战：太阳耀斑与双飞控故障转移系统 在回顾过去十年的技术历程时，主持人和嘉宾探讨了一个核心问题：2016 年刚启动无人机配送时，有哪些系统挑战是当时完全始料未及的？Zipline 的初衷是运送救命物资，医疗客户在任何极端天气条件下都需要物资保障。常人首先想到的是狂风骤雨等气象考验，但团队遇到的最怪异障碍之一却是“空间天气”——太阳耀斑（Solar Flares）。太阳表面的巨大爆炸会向地球喷射大量高能辐射，直接扰乱电离层，导致 GPS 卫星的射频信号出现不可预测的快慢偏差，进而使机载导航系统严重降级。为此，Zipline 深入研究了该领域，一方面增强自身导航与 GNSS 系统的鲁棒性，确保在太阳耀斑爆发时仍能维持厘米级定位精度；另一方面在 GNSS 之外引入多重冗余，确保在卫星信号极度恶劣时依然能绝对安全地飞行。 负责系统工程与安全的 Eric Watson 详细拆解了飞机内部极为硬核的计算故障转移机制（Compute Failover）。飞机大脑——飞行计算机负责整合传感器海量数据并向电机、舵面等执行器下达指令。在现实环境中，任何部件都可能出故障，包括软件偶发异常、接插件接触不良，甚至是宇宙射线导致的内存比特翻转（Bit Flip）。Zipline 的解决策略是配备两台完全对等的飞行计算机：两台计算机在任何时刻都“自认为正在掌控全局”，同时接收全部传感器数据并持续计算输出控制指令。系统中间设有一个轻量的第三方仲裁器（Arbiter），时刻监控两台飞控的健康状况，向机上所有执行节点指示当前听从谁的指挥。 ```mermaid flowchart TD accTitle: 双飞行计算机冗余与仲裁架构 accDescr: 展示主备两台飞行计算机如何同时接收传感器数据，并由仲裁器决定执行控制权。 subgraph Sensors["传感器输入"] S["全机传感器集群<br/>(GNSS / IMU / 空速管等)"] end subgraph DualCompute["双飞控计算核心"] FC1["主飞行计算机 (FC 1)<br/>『实时解算并输出指令』"] FC2["备用飞行计算机 (FC 2)<br/>『实时解算并待命接管』"] ARB["仲裁微处理器 (Arbiter)<br/>『持续监测两台飞控健康度』"] end subgraph Actuators["执行机构与底盘"] ACT["电机与舵面执行器"] end S --> FC1 S --> FC2 FC1 -.->|"状态心跳"| ARB FC2 -.->|"状态心跳"| ARB ARB -->|"指派控制权"| ACT FC1 ==>|"执行指令 (默认)"| ACT FC2 ==>|"执行指令 (故障切换)"| ACT ``` 当被问及“如果仲裁器本身坏了怎么办”时，Eric 解释道，系统设计遵循最简默认路径：一旦仲裁器宕机，当前正在飞行的主计算机将直接保持主导地位，继续执行飞行任务。就在几周前，这套系统刚在实战中立下大功：一架无人机在刚完成用户交付后，主飞行计算机突然出现小故障，仲裁系统瞬间无缝切断至备用飞控，飞机随后完全自主返航并平稳降落。正是依靠这种深入骨髓的冗余设计，Zipline 创造了 250 万次交付、1.4 亿英里自主飞行且零安全事故的行业纪录。联合创始人 Keller Rinaudo Cliffton 补充指出，双飞控架构在理念上并非 Zipline 独创，波音 777 同样采用类似机制，但波音航电设备成本高达数百万美元；Zipline 真正颠覆的地方在于，充分利用智能手机供应链的高性价比成熟元器件，以数十至数百美元的硬件成本达成传统航空级的安全冗余，研发迭代速度快了 100 倍，成本却只有传统方式的百分之一。 ### `13:30` 700 个自研零部件与极端环境下的破坏性测试 关于“无人机本身只占整体解决方案的 15%”这一论断，团队深入剖析了背后极其复杂的全栈硬件验证体系。Eric 强调，Zipline 绝不是一家纯软件公司，而是一家扎根物理世界的“真实世界 AI 机器人企业”。在硬件层面，团队推行严苛的高加速寿命测试（HALT）。面对电机等关键运动部件，工程团队会将其推入极端严酷的环境——在高温、高湿、强腐蚀盐雾以及强烈紫外线交织的舱室内，让电机以极限工况高速旋转并往复受力，榨取其性能极限。 Keller 透露了令人惊叹的垂直整合深度：整架无人机上拥有 700 个由 Zipline 从零自研的独立部件，并组成 43 个核心子总成。从机载主控、配电板、电机控制器，到电池模块及其电池管理系统（BMS），无一不是自主开发。甚至用于精准落锁投递家庭后院的下潜小机器人（Delivery Pod），内部都搭载了自研的高性能 NVIDIA GPU 飞行计算机；连用于自动起降充电的地面扩展坞站（Docking Station）电子设备与高推重比定制无刷电机也是从零定制，因为市售消费级电机根本无法满足严苛的功耗与推重比需求。 在与传统汽车或航空业工程师交流时，对方衡量可靠性往往是“直接向供应商索取测试报告”，而 Zipline 的角色本身就是顶层供应商。因此，地面验证不仅覆盖元器件，更延伸至整机级别：完整的无人机被装载进巨型振动台、风洞以及可供人员进出的大型步入式高低温湿热交变箱中。测试哲学的核心不是“证明它能正常运转”，而是“彻底测到它发生结构或电气破坏”。团队必须摸清部件极限在哪里、会以何种模式断裂或失效，从而准确评估风险或做针对性结构优化。其测试温度跨度极其惊人，高至 49°C，低至 -25°C。针对现场主持人的好奇，Keller 坦承这完全源自实际运营场景：美国凤凰城的酷暑夏季机体地表温度轻松逼近 49°C，而美国北部高纬度地区的凛冬作业环境则低至 -25°C。 - 核心硬件规模：整机包含 700 个自主研发的独立零件与 43 个主要子总成。 - 边缘算力配置：末端交付舱（Delivery Pod）内嵌专门由 NVIDIA GPU 驱动的自主飞控系统。 - 极端测试温区：覆盖 49°C（如亚利桑那州凤凰城夏季高温）到 -25°C（如美国北部严寒冬季）。 ### `17:00` 生命重于一切：从伞降容错到“Waymo 两倍安全”的新标准 面对极端天气，商业公司通常最容易做出的妥协是“天气太差就停飞”。主持人 Pat Grady 提问 Zipline 如何在运营安全性与出勤率之间做取舍。Keller 直言，这种毫不妥协的工程定力源于公司的起点业务——紧急医疗生命救援。“如果有人生命垂危，你总不能告诉对方‘只要外面阳光明媚，你就可以托付生命给我们’，这在商业和道义上都毫无说服力。”成立之初的前几年，Zipline 在民航监管机构的密切协同下承担了极大的探索风险。当时团队抱定信念：只要能拯救产后大出血或严重外伤患者的生命，就必须迎难而上。即使在极端偶发情况下飞机遭遇不可逆故障，系统底层的终极安全防线——降落伞系统（Paraland）也会瞬间弹出，让飞机极其轻柔地降落地面。早期伞降事件偶有发生，但换来的是挽救了无数生命并沉淀下极为宝贵的系统边界数据；而在如今的成熟体系下，伞降机制已极罕被触发。 在安全基准讨论中，红杉资本合伙人、Zipline 董事会成员 Alfred Lin 回忆了上一届董事会上的激烈交锋。当时管理团队汇报称目标是“比传统人类驾驶汽车安全 10 倍”，Alfred 现场驳斥道：“这是个毫无意义的目标（BS goal），汽车是极度原始陈旧的技术。Waymo 的自动驾驶已经做到比人类汽车安全约 10 倍，我们必须将目标定为‘比 Waymo 还要安全两倍’！” ```mermaid flowchart LR accTitle: 安全目标演进逻辑 accDescr: 对比人类驾驶、Waymo自动驾驶与Zipline设定的终极安全冗余层级。 CAR["传统人类汽车驾驶<br/>基准事故率"] -->|"Waymo 设定目标<br/>约 10x 更安全"| WAYMO["Waymo 自动驾驶汽车"] WAYMO -->|"Alfred 董事会要求<br/>再提升 2x 安全冗余"| ZIP["Zipline 年底终极安全目标<br/>(2x Safer than Waymo)"] ``` Eric 随即将该指标作为安全团队的年终考核标准。尽管现场围绕“低空飞行器与地面自动驾驶汽车是否具备可比性”展开了探讨（有人认为空中没有行人和复杂路况，但 Keller 强调无人机本身替代的就是原本穿梭在公路上的送货车辆），但不可辩驳的事实是：Zipline 商业自主航行 1.4 亿英里，至今保持完全零伤亡、零安全事故。作为对比，在地面交通中，普通汽车行驶 1.4 亿英里平均会导致 600 起碰撞车祸、100 人受伤以及 2 至 6 人死亡。正是这种在极端场景中被千锤百炼打磨出的系统韧性，让商业机队中的单机经受住了极高的飞行寿命考验。 - 关键安全战绩：商业自主飞行里程突破 1.4 亿英里，实现 0 安全责任事故。 - 地面事故率参照：汽车行驶 1.4 亿英里统计学上对应约 600 次车祸事故、100 次人员受伤及 2-6 起致死事故。 - 核心演进目标：正式将安全基线从“比人类驾驶汽车安全 10 倍”提升至“比 Waymo 还要安全 2 倍”。 ### `20:00` 百米高空悬停与下放舱：厘米级精准投放的技术架构 许多人直觉上会认为无人机配送系统精致而脆弱，在面对恶劣天气或极端环境时往往难以应对。但 Keller Rinaudo Cliffton 指出，Zipline 的无人机已经在商用自主飞行中累积了数百万英里，其耐用性、坚固度与鲁棒性远超公众预期——“如果你的汽车跑了一百万英里，那绝对令人刮目相看，而我们的系统早已在这个量级上稳定运行”。当主持人 Pat Grady 惊叹于无人机如何在风雨天气中将包裹精准放置在极小半径（约 18 英寸）的目标圆圈内时，Eric Watson 修正了一个常见的误区：无人机并不是在 100 英尺的高度作业，而是在距地面约 100 米（约 300 英尺）的高空悬停。高度提升让操控难度呈几何级数增加。 面对风向变化，主机在悬停时会自主向上风向偏移，利用微调抵消气流干扰，随后垂直下放独立的递送舱（Delivery Pod）。虽然系统利用实时动态载波相位差分技术（RTK-GNSS）实现了厘米级自身定位，但用户绝不可能提前测量自家后院某张具体咖啡桌的 GPS 坐标。RTK 只能将递送舱引导到目标区域的大致范围（例如后院上空），真正的微调与安全判断则完全依靠递送舱自身携带的感知与边缘计算单元。递送舱配备了专门的 NVIDIA GPU，运行着独立的 AI 自主飞行算法栈，甚至在黑夜中也能自主绘制地形。如果感知系统发现下方桌子上摆满了饮料，它会智能规避打翻饮料的风险，主动选择降落在桌子旁的平整空地上。 ```mermaid flowchart TD accTitle: Zipline 平台2代高空精准投递架构流程 accDescr: 展示从100米高空悬停、感知环境、下放具备独立动力的投递舱到精准触地回收的完整步骤。 subgraph 母机协同与粗定位 ["100 米高空：母机悬停"] A["RTK-GNSS 厘米级定位导航"] --> B["逆风偏移补偿气流"] B --> C["匀速下放轻量级投递舱"] end subgraph 边缘感知与机动 ["近地端：投递舱自主决策"] C --> D["NVIDIA GPU 运行视觉感知系统"] D --> E{"识别落点障碍物\n(如：桌上饮料、宠物)"} E -- 存在障碍 --> F["X/Y 轴微推进器机动平移"] E -- 落点安全 --> G["精准接触目标地面"] F --> G end subgraph 完成交付 ["回收与返航"] G --> H["底舱自动开启释放包裹"] H --> I["投递舱高速收卷回母机仓位"] I --> J["母机自主规划航线返航"] end ``` 不仅如此，递送舱还能通过微型推进器在 X 轴与 Y 轴上主动控制自身位置。这种“母机高空悬停 + 边缘智控子舱下放”的架构带来了两大核心优势： - **极致静音**：传统无人机刺耳的高频噪音对人类耳朵而言极具折磨性。Zipline 组建了规模庞大的空气动力学、气动声学与飞行控制专家团队，目标是将飞行噪音降至“微风拂过树叶般的沙沙声”。将产噪最大的主机悬停在 100 米高空，彻底拉开了噪声源与居民的物理距离。 - **极致安全**：真正接近用户、家庭、宠物和儿童的，只有那个质感类似轻质泡沫塑料材质的下放舱，“外观小巧且人畜无害”，从物理层面上杜绝了旋翼近距离伤害的隐患。 --- ### `24:45` 八年磨一剑：从非洲血库配送到美国本土监管破冰 在将新一代家庭配送网络引入达拉斯等美国核心城市之前，Zipline 经历了长达数年的海外验证与漫长的监管对接。从 2016 年在卢旺达开启全球首个商用血液无人机配送服务算起，整个技术在海外打磨了六到八年；而目前处于“疯狂超大规模扩展模式”（Hyperscaling Mode）的新一代家庭配送服务，实际上直到前一年（2025 年初）才正式商用。综合算来，整体技术的演进与铺垫已耗时近九年。 进入美国本土绝非一蹴而就的“顺风航行”。Zipline 从 2020 年左右开始与美国联邦航空管理局（FAA）等监管机构进行深度接洽。在 2016 年前后，商用无人机长距离自动配送在法律层面几乎没有任何准入路径。两者的合作建立在共同的目标之上——安全、高效地融入国家空域（Airspace Integration）。Zipline 将在非洲多个国家积累的数百万英里实战经验带入谈判，监管机构亦有其对空域管控的严格审查标准，双方历经数年磋商与技术对齐，才最终敲定了一套合法合规、可复现的监管执行框架。 --- ### `26:15` 从 1:1 到 1:100：自主基础设施与“舰队指挥官”的诞生 Eric Watson 强调，正如 Keller 所言，无人机硬件本身仅占系统整体复杂度的 15%，Zipline 真正构建的是一套提供即时商品获取能力的基础设施层（Infrastructure Layer）。支撑系统运作的并非从 A 点单向飞往 B 点的孤立飞机，而是一个由高密度充电桩网络、数百架分布式协同无人机以及云端自主调度大脑组成的多维生态。云端调度系统能够实时计算需求与供给分布、预测晚餐高峰时段的配送潮、评估各微气候网格内的风速气象，并进行运力自平衡。 在全自动运行的体系中，人类的角色被重新定义。系统虽然不需要人工操纵飞机，但会利用算法在出现风速骤升或偶发异常时主动呼叫专业航空人员介入监督。这一运营角色的进化轨迹直接决定了单位经济模型（Unit Economics）的成败： | 演进阶段 | 监管模式 / 控制比例 | 职责定义与经济效应 | |:--- |:--- |:--- | | **初期探索阶段** | **1 对 1 (1:1)** | 每位远程机长在办公室盯守单架飞机，极少干预，但单位经济模型极差 | | **规模化爬坡期** | **1:3 → 1:6 → 1:20 → 1:40** | 伴随系统冗余度与安全认证完善，单人监管规模成倍提升 | | **当前常态化** | **1 对 100 (1:100)** | 成为 FAA 正式审批的标准操作协议，单人全局把控百架机队 | | **未来目标** | **1 对 100+** | 向更高量级的自主机队集群调度迈进 | 这一职位的命名也经历了语义更迭。由于“飞行员”（Pilot）极易让人误以为有人在拿手柄实时操纵，团队从科幻小说《安德的游戏》（Ender's Game）中汲取灵感，将其正式更名为“舰队指挥官”（Fleet Commander），该称号甚至被正式写入了 FAA 的官方监管合规文件中。Keller 充满激情地表示，这正是技术最迷人的地方：机器人并非在剥夺人类的工作，而是在“赋能人类向上跃升”——专业人员不再去低头做基础苦力，而是站在战略层面上统筹协调上百台自主机器人。 --- ### `28:30` 奔向单日百万单：百万分之一极端案例的常态化挑战 随着业务进入爆炸性扩展阶段，系统的工程挑战完全转向了极端尺度下的长尾问题。Eric 透露，当历史累计交付量突破 250 万次后，“百万分之一概率”的罕见边缘工况（Edge Cases）开始频频现身。 Keller 对此给出了极具冲击力的对比数据：Zipline 从 2014 年起步研发底层技术，整整耗费了大约十年时间（直到 2024 年底或 2025 年初），才在全公司历史上累计完成了第一个 100 万单递送。然而在当下与不久的将来，Zipline 即将实现**每天完成 100 万单**的常态化递送。这种指数级的跨越带来了极大的敬畏感： - **长尾故障常态化**：过去那些“几年才遇一次”的罕见偶发事件，在日均百万单的体量下，将变成**每天必发生一次**的确定性事件。 - **旧模式的全面失效**：过去的测试标准、调试手段与解决思路在此等吞吐量下几乎会全面瓦解，所有系统指标的门槛都被迫大幅抬高。 - **超大规模运维（Maintenance）瓶颈**：随着机队在役硬件数量的激增，任何微小的流程低效或失误都会被放大成灾难性的维护成本，系统的每一个模块都必须具备极端的容错与自愈能力。 尽管规模扩张带来了沉重的工程压力，但高度自主化架构在规模效应下也正在展现出许多传统物流系统未曾被认识到的巨大底层优势。 ### `30:00` 24/7全天候运转与自主“自我交付”系统 Keller Rinaudo Cliffton 指出，大众往往习惯了传统物流由卡车司机工作轮班所决定的运营时间表，但自主飞行系统天然渴望且能够做到“24/7全天候运转”。无人机机群在凌晨两三点进行配送与在下午两点配送同样游刃有余。这种全天候能力让系统可以在用户早晨6点起床上班前，就将物品悄无声息地送达门廊或后院，平均单程配送仅需5分钟，并保持100%的可用性。 更具颠覆性的是机群的自主调度机制。Zipline 即将实现让飞行器直接从位于南旧金山（South San Francisco）的制造工厂起飞，自主飞入商业运营空域，正如特斯拉 Model 3 和 Model Y 尝试自主交付给客户一样。这种能力在维护层面带来了巨大优势：一旦某架飞行器触发预防性维护需求，它会自行飞往维护中心；地勤人员完成必要检查与维护后，飞机再自行飞回业务运营区域。此外，系统能根据全城各区域的实时需求动态调配运力，飞机没有固定的“老巢”，哪里需要就飞往哪里。 - 配送时效与可用性：全天候 24/7 待命，单次配送耗时约 5 分钟，可用率达 100%。 - 自主运维闭环：飞行器从南旧金山工厂自主飞入运营空域，需要维护时自动飞往维护中心，完工后自行返岗。 - 动态运力网络：全城范围内无固定停机位绑定，算法根据实时订单密度动态迁移运力。 ### `31:40` 日行百万单的挑战：造就“制造机器的机器” 谈到迈向“每天100万次配送”的全新挑战时，Eric Watson 重申了核心观点：飞行器本身仅占整个系统工程难度的15%。Zipline 目前制造飞机、维护机群、支持运营和排查故障的整套模式，在每天100万单的超大规模下将完全无法维系。这要求团队必须在更强大的工具链、软件架构和底层工艺上进行质的跃升。 正如 Elon Musk 经常强调的“制造那台造就机器的机器”（designing the machine that builds the machine），Eric 认为 Zipline 未来数年的核心任务正是重金投资于用于制造和运转机群的基础设施。从业务规模对比来看，目前全美最大的民航商业航空公司每天飞行约 5,000 个架次，而 Zipline 在下个月的飞行量就将超越这一数字。当达到每天100万单时，Zipline 在美国商业空域的每日飞行量将是“全美所有传统商业航空公司总和的40到80倍”。 ```mermaid flowchart TD accTitle: Zipline 从工厂到空域的规模化运营架构 accDescr: 展示飞行器从工厂出厂、自主飞往商业运营空域、遭遇冲突自主协商避让，到触发维护自主返回机库的完整闭环。 subgraph FactoryOps["智能制造与自我部署"] A["南旧金山制造工厂"] -->|"出厂自主起飞"| B["商业运营网络"] end subgraph AirspaceCoord["空域实时协同"] B --> C{"AI / V2V 空中防撞监测"} C -->|"检测潜在航线冲突"| D["自主协商航线：一机上升 / 一机下降"] D --> B end subgraph MaintenanceLoop["自主闭环维护"] B -->|"触发预防性维护阈值"| E["自主飞往维护中心"] E -->|"地勤快速检修"| B end ``` ### `33:10` 重塑空中交通管制：空域协同与自研ATC系统的商业价值 面对体量悬殊的新型飞行器涌入空域，现有空中交通管制（ATC）系统并未做好分类管理的准备。Keller 强调，天空是一片无比宽广的物理空间，而人类目前却拥挤在狭窄、古老且严重受限的陆地道路上。随着电气化与自主载具浪潮的发展，天空中运行的自主飞行器数量未来将是地面车辆的10倍。更有效地利用天空，可以把地面空间重新还给人类，带来更安静、安全、少污染和不拥堵的社区环境。然而，美国现行的空管体系仍停留在1950年代的设计逻辑中——依赖纸笔、卡片和管制员肉眼观察窗外飞机，根本无法支撑未来的自主飞行密度。 针对 Alfred Lin 关于 Zipline 是否在协助美国联邦航空管理局（FAA）进行体系设计的提问，Eric 指出这是一种“协同创新”（collaborative innovation）。单一企业不可能孤立解决全行业的问题。Zipline 坚信解决方案的核心在于飞行器与飞行器之间的直接协同（V2V/D2D 通信）：飞行器应自主向邻近机群广播自身坐标，并在发现潜在航迹冲突时自动协商规避方案（例如一架爬升、一架下降）。Zipline 正联合其他新兴自主飞行器厂商、监管机构和行业标准组织，将这些已在实践中得到验证的最佳方案推行为通用行业标准。 Keller 补充道，由于早年在非洲多国落地时当地根本没有适用的现代化空管监控工具，Zipline 被迫完全从零自研了整套空管软件，并直接交付给当地民航局使用。他半开玩笑地指出，美国有多家市值超过100亿美元的上市公司专门为空管开发软件，而“Zipline 内部实际上孕育着好几家潜在的百亿美元级独立上市公司”，空管软件只是其中之一。为了让整体技术生态运转起来，他们几乎把所有外围模块都亲手重造了一遍。 ### `35:10` 传统空管体系的人力危机与技术代差 当前航空系统的脆弱性正不断暴露在公众视野中。从华盛顿特区飞机与直升机相撞的险情，到各机场滑行道上频繁发生的飞机擦碰事故，传统空管系统承受的压力已濒临极限。Keller 揭示了隐藏在这些事故背后的冷酷现实：全美空管人员中约有 50% 年龄超过 45 岁，其中 20% 已临近退休年龄；更糟糕的是，当下几乎没有年轻人愿意将空中交通管制选作终身职业。 这场突如其来的劳动力断层，使美国航空业陷入了系统性的用工危机。即便管理机构想沿用1950年代确立的人工空管模式，现实中也根本没有足够的人手可用。一边是即将彻底空心化的老龄化专业劳动力，另一边是汹涌而至的 AI、自动化技术和海量自主飞行器，双重夹击正倒逼整个空管体系进行根本性的底层范式重构。 - 空管劳动力结构：50% 的空管员年龄超过 45 岁，20% 即将达到法定退休年龄。 - 职业断层：年轻一代几乎无人进入空管行业，人力供给完全无法匹配机队增长。 - 事故警钟：从华盛顿特区的空中险情到机场滑行道相撞，旧有依靠人眼和纸卡的调度模式已不堪重负。 ### `36:55` 绝望中走向全面垂直整合：从炸机中诞生的车规级自研体系 当被问及 Zipline 为什么从硬件设计、制造组装、底层软件到现场运营全链路高度垂直整合时，Keller 直言，没有任何一家硬件公司一开始就心甘情愿做垂直整合，因为“这过程简直痛苦得要命（an incredible pain in the butt）”。Keller 透露自己在工位上方挂了一面旗子，上面写着一句调侃肯尼迪演讲的名言：“我们做这事不是因为它们容易，而是因为我们当初以为它容易（We do this not because it is easy, but because we thought that it would be easy）”。 几乎所有成功的硬件公司最初都试图走捷径：就像第一代特斯拉 Roadster 试图直接套用莲花跑车（Lotus Elise）的底盘并外购电池包，结果既不可靠又巨额亏损，但它是通往 Model S 的必经之路。Zipline 早期同样选择外购现成货架零部件（COTS），出资雇佣外部供应商开发组件。然而，测试机在靶场一次接一次地坠毁（“We crashed and we crashed”）。团队最终痛定思痛：这些外购零件不仅极其昂贵，而且稳定性奇差。 面对现实，团队只能绝望地像剥洋葱一样把系统全部拆掉重写：拔掉不可靠的电机控制器，从零自研；拆掉通用的 GPS 导航模块，从零自研。虽然大疆（DJI）每年生产数百万台塑料四轴飞行器，美国军方也拥有造价高达 2000 万美元、能飞 100 英里的“捕食者”（Predator）无人机，但这两种极端路线在商业配送所要求的“高可靠性、高安全性与极致单位经济模型”面前全都不堪一击。面对早已拥有百年降本增效积淀的汽车和摩托车，Zipline 必须制造出“车规级（automotive grade）”、成本极低且高可靠的全新硬件体系。“你绝不是预谋要做垂直整合的，你只是在一次次抓狂和绝望中意识到：我们必须把所有这些垃圾全扔掉，一切从头自研。” ```mermaid flowchart LR accTitle: Zipline 垂直整合核心反馈流 accDescr: 展现研发工位、生产车间与测试试飞场地之间的物理极速闭环流程。 subgraph Headquarters["同处一室的研发团队"] A["固件与自研芯片"] B["机械与航气工程"] C["GNC / 自主导航算法"] end subgraph PhysicalLoop["物理极速验证闭环"] D["生产制造工厂 (车程3分钟)"] E["外场测试基地 (短途车程)"] end Headquarters -->|"原型快速试产"| D D -->|"整机装配上天"| E E -->|"真实炸机与性能数据"| Headquarters ``` ### `38:50` “沙丁鱼罐头”式的紧密协同：软硬件全链路的物理闭环 一旦硬着头皮完成了全自研，垂直整合带来的终极红利便体现为两点：“迭代速度与系统深度融合”。Keller 描述了 Zipline 极具特色的办公生态：走进公司的南旧金山办公室，所有工程团队全部像“沙丁鱼”一样紧密挤在同一栋大楼里。 办公区没有任何部门壁垒：固件工程师（firmware）紧挨着机械工程师（mechanical），旁边坐着自主算法工程师（autonomy），再旁边是云端基础架构（cloud infra）、气动学（aero）、声学工程（aeroacoustics）、制导导航与控制（GNC）、系统工程（systems engineering）以及制造部门的同事。更关键的是地理布局上的极致紧凑——Zipline 的制造工厂距离这栋办公楼只有 3 分钟车程，而测试试飞基地也仅有很短的车程。 工程师们可以在办公室写完代码或完成图纸，3分钟后直接站在工厂车间里紧盯零件总装过程，随后驱车前往试飞场亲眼观察飞机冲上云霄、测试性能表现并收集数据。这种跨越机械、电气、算法到实体测试的极速物理反馈回路，构成了 Zipline 难以被外包组装型竞争对手复制的技术护城河。 ### `40:00` 垂直整合与去官僚化协作：比传统航空工业快得多的排错模式 在复杂的工程系统中，故障几乎每天都在发生。Keller Rinaudo Cliffton 与 Eric Watson 指出，Zipline 能够快速迭代的核心秘诀在于极致的组织扁平化与全栈垂直整合。当系统出现严重缺陷时，工程师可以直接起身走到同事桌前说：“今晚你和我一起通宵把这事搞定。”这种协作模式与传统航空巨头（如波音）形成鲜明反差。以波音 787 电池问题为例，传统航司往往深陷三层以上的采购合同与外包供应链泥潭，光是打官司、厘清具体是哪家供应商的责任就要耗费整整两年。 Eric 强调，要让不同学科背景的人无缝协作其实非常简单，关键在于目标一致且“立足于真实的工程实体”。在很多大型企业里，工程师之间互相推诿，甚至以“源码是团队知识产权”为由拒绝告知具体故障检测逻辑。但在 Zipline，制造、运营与工程团队坐在一起：一旦测试失败，大家立刻当面拉出代码逐行排查，当场发现最初的系统假设差异，并迅速解决根因。 ```mermaid flowchart TD accTitle: Zipline 协同模式与传统航空外包模式对比 accDescr: 对比 Zipline 垂直整合下工程、运营与制造团队的面对面即时排错，与传统航空深层外包、多层级供应链下的诉讼与推诿。 subgraph ZIP["Zipline 垂直整合模式"] Z1["测试失败 / 硬件故障"] --> Z2["直接走到同事工位"] Z2 --> Z3["制造、运营与软件团队当面拉取代码"] Z3 --> Z4["当晚通宵闭环解决问题"] end subgraph BOE["传统航空分包模式（如波音 787 电池事故）"] B1["出现故障"] --> B2["受制于 3 层以上外包采购合同"] B2 --> B3["代码与硬件细节被视为商业机密互不透明"] B3 --> B4["耗费 2 年诉讼供应商与推诿责任"] end ``` ### `42:08` 第一性原理与“删减部件”的勇气：拦阻钩的工程赌注 Keller 提到，构建融合软硬件的高复杂度 AI 与机器人系统时，必须遵循一条核心信条：质疑每一条需求（"Question every requirement"）。这一理念与 SpaceX 埃隆·马斯克推崇的算法如出一辙——需求往往是荒谬的，工程师必须深入挖掘两层以上去验证其必要性。与此相辅相成的是“删减部件”（Delete the part）：一架飞机上最可靠、绝不会发生故障的零件，就是那个被你彻底删掉的零件。从 SpaceX Raptor 1、Raptor 2 到 Raptor 3 引擎的极简演进，再到 Zipline 从早期的 10 个硬件迭代版本，直到 EV1、EV2 与最新的 EV3 机型，贯穿始终的主线就是持续不断地“删除、删除、再删除”。Keller 直言，删掉部件是一项极具勇气的行为，因为一旦删除后系统无法运转，工程师会显得极度愚蠢；但唯有对物理规律和系统表现抱有真正的信心，才能迈出这一步。 Eric 分享了一个关于早期平台一号（Platform 1）着陆系统的经典案例： - 最初设计采用了一根长达一米的可展开尾钩（Tail hook），飞机降落时放下大钩勾住地面阻拦索减速，结构沉重且极其复杂。 - 团队基于第一性原理推演：能否将机械复杂度完全转移到地面系统，让地面机器人像航空母舰舰载机回收系统一样，主动伸出机械爪把飞机抓下来？如此一来，机载尾钩就能从一米缩减到仅有两厘米。 - 在即将投产新一代飞机的关键节点，地面抓取系统尚未调试成功。团队围在会议室里做出抉择：是继续沿用一米的大长钩，还是把尾钩删减至两厘米并在数周内逼自己搞定地面系统？ - 最终团队拍板删掉大钩子，全员连续通宵数周攻坚。当第一架搭载两厘米短钩的飞机飞抵时，地面系统精准捕获降落，成功将飞机重量大幅减轻，并大幅提升了整体可靠性。 ### `44:48` 市场容量的真正拐点：从 50 亿到 550 亿次即时配送的大扩张 当被主持人问及“未来是否会运送人类乘客”时，Keller 谨慎而坚定地表示，将地面交通从“道路的暴政”中解放出来确实是诱人的未来，但 Zipline 眼下保持着极度的谦逊与危机意识，首要目标是构建起一个像互联网传输信息一样高效传输物理包裹的全球新基础设施层。 当下即时配送市场正在迎来颠覆性扩容： - 美国的现状是每年约有 55 亿次由人类司机驾驶 4000 磅重燃油车完成的“即时配送”，耗时普遍在 30 至 60 分钟。这种模式不仅低效、不环保，还充斥着安全隐患与服务质量问题（超过 50% 的外卖骑手承认曾偷吃过顾客的食物）。 - Zipline 在达拉斯的大规模商业运营展示出惊人的消费习惯变迁：配送变得极其快捷、便宜且体验绝佳后，顾客的订购频次从原本的每月几次跃升为每天下单。一位 80 岁的养老院老奶奶在一年内通过 Face ID 和 Apple Pay 下单了 350 次 Zipline 配送。 - 若将达拉斯展现出的消费行为外推至全美，市场规模不是 55 亿次，而是高达 550 亿次。这与当年 Uber 进驻旧金山如出一辙——投资人最初以为 Uber 顶多占领出租车市场的 33%，却未曾料到其通过提升便利性直接将市场整体容量扩大了 10 倍。 Keller 指出，全美根本没有足够的卡车司机来承接 550 亿次配送，大城市的道路网络更不可能承受汽车数量翻倍带来的拥堵灾难。唯一能满足这一海量需求的可行路径，唯有高度自动化的无人机配送系统。 - 当前日飞行量：Zipline 目前每天飞行近 5,000 架次。 - 年底目标：预计在今年年底突破每天 30,000 架次。 - 扩张速度：计划在未来 4 个月内再落地 4 个全新大都市圈。 - 终极产能目标：原计划年产 20,000 架飞机以支撑每天 100 万次配送；如今已有数个合作伙伴各自要求每天购买 100 万次配送运力，“每天 100 万次”已经从最初的公司经营目标演变成了单笔合同的销售单位，促使工厂与供应链全面向更高数量级扩产。 ### `50:00` 硬件创业的“10倍成本法则”与降至12美元的成本曲线 当谈及给硬件创始人的建议时，Keller Rinaudo Cliffton 分享了一条在血泪教训中摸索出的铁律：当你预估某项硬件成本是 X 时，请直接假设它需要 10X。在做天使投资和回馈创业生态时，他经常遇到声称成本只要极少数额的团队，他总会反问对方：“假设实际成本是它的10倍，商业模型还能跑通吗？如果真变成10倍，你该怎么应对？”这并非危言耸听，而是 Zipline 在 2016 年上线时的真实经历。 当时 Zipline 签署合同为卢旺达运送单程 80 到 100 英里的输血急救包，定价为每单 30 美元，这与传统地面运输成本相当，团队也天真地以为成本差不多就是 30 美元。结果系统正式投入运营时，单次交付的实际成本高达 300 美元——整整翻了10倍。幸运的是，投资人 Alfred Lin 表现得异常淡定，团队也随即沉下心开始长达数年的降本攻坚战。随后几年里，单次配送成本从 300 美元一路猛降至 120 美元、75 美元、40 美元、28 美元、18 美元，直至今天在美国境外运行的远程物流网络中，单次配送成本已经降到了 12 美元。 ```mermaid flowchart TD accTitle: Zipline 单次配送成本演进历程 accDescr: 从2016年上线时的300美元一路优化至目前的12美元的成本曲线 subgraph Phase1["初创期与高昂探索成本"] A["2016上线成本: $300<br/>(原定目标价: $30)"] --> B["次年优化: $120"] end subgraph Phase2["规模化与工艺迭代"] B --> C["第三年: $75"] C --> D["第四年: $40"] end subgraph Phase3["成熟运营与拐点临近"] D --> E["降至: $28"] E --> F["降至: $18"] F --> G["现阶段远程网络: $12"] end ``` - **初始落差**：合同约定配送价为 30 美元/单，但系统在 2016 年投入使用时的全包成本高达 300 美元/单。 - **降本轨迹**：300 美元 → 120 美元 → 75 美元 → 40 美元 → 28 美元 → 18 美元 → 12 美元（海外长距离航线）。 - **资本耐性**：面对上线初期 10 倍的成本差距，红杉资本合伙人 Alfred Lin 当时表现得“出奇从容”，为技术与供应链成熟争取了宝贵的时间窗口。 --- ### `51:40` 拐点降临：机器人配送成本全面低于汽车与“告别小费” Keller 指出，一场鲜为人知但影响深远的剧变正在发生：全包计费（Fully Burdened）体系下的自主无人机配送单位经济模型，正在这个夏天悄悄跌穿用汽车送货的成本底线。这一历史性的拐点虽然没有登上《纽约时报》等主流媒体的头条，却标志着机器人进入物流领域的成本正式低于人类司机。它将彻底重构人类社区的运转面貌，不仅能让街道变得更安静、更安全、告别尾气拥堵，也让终端配送变得更快、更廉价。 在用户体验方面，除了惊人的配送时效与稳定性，消费者反馈中最热烈、最直观的一条评价竟然是“不用付小费！！”（No tip, exclamation point!）。在如今小费文化盛行且带来普遍道德绑架与心理内疚感的背景下，无人机配送不仅成本透明可预期，还彻底免除了顾客在门前额外给小费的社交焦虑。 - **成本十字交叉点**：全包无人机物流成本首次全面低于传统燃油车/人力配送。 - **环境与社会效益**：减少道路交通伤亡、削减城市尾气碳排放、降低社区噪音干扰。 - **最受欢迎的用户反馈**：告别小费（“No tip!!!”），透明定价彻底消除了消费者的心理负担。 --- ### `52:45` 十年“黑羊”的坚守：重塑造物能力与科幻未来的信念 在投资人打趣硬件不仅费钱、耗时也远超预期时，Keller 提到了红杉合伙人 Shaun 撰写的一篇内部备忘录。该备忘录深刻阐述了为何硬件科技企业将在未来几十年成为推动人类文明进步的中流砥柱，同时也道尽了硬件创业起步之艰难与募资之辛酸。回顾过去十年，Zipline 就像硅谷投资圈里的一只“黑羊”（Black Sheep）——在人人都追逐高毛利纯软件 SaaS、移动 App 和轻资产代码的黄金年代，几乎没有投资人愿意多看硬件公司一眼。 Keller 借用《蝙蝠侠：黑暗骑士崛起》中反派贝恩（Bane）的经典台词打趣道：“你们只是适应了黑暗，而我是在黑暗中出生的（You adopted the darkness, I was born in it）。”早在硬件和机器人重新变成风口之前，Zipline 就已经在无人问津的冷板凳上死磕了十年。 面向未来，Keller 呼吁重新建立对实体世界的建设能力。当今社会过于依赖祖辈留下的老化基础设施，他举了一个极具讽刺意味的例子：旧金山金门大桥最近安装防自杀防护网的工程造价，竟然比当年祖辈从零建造整座金门大桥的总成本还要高昂。要维护国家竞争力并亲手兑现童年时那些令人心潮澎湃的“科幻未来”，整个产业必须找回强大的制造魄力，重新精通复杂的实体供应链，勇敢去建造属于下一代人的全新物理基础设施。 - **贝恩式创业自嘲**：“你们只是适应了黑暗，而我是在黑暗中出生的”——以此形容 Zipline 在 SaaS 狂热时代坚守硬件与机器人领域十年的孤独行径。 - **基建成本反思**：金门大桥加装防自杀网的费用甚至超过当年修桥的历史总投入，印证了现代社会重型实体工程与造物能力的退化。 - **未来愿景**：摆脱纯数字软件的单一依赖，下定决心重新掌握复杂制造与硬件工程链条，将许诺给子孙后代的科幻图景变为现实。 ## Notable quotes > `02:30` 这在美国不是违法的吗？ — 投资人 > `06:05` 病人生病是不分昼夜的，为什么你们每天只营业 12 个小时？ — 卢旺达医护人员 > `17:00` 这是个毫无意义的目标，汽车是极度原始陈旧的技术。Waymo 的自动驾驶已经做到比人类汽车安全约 10 倍，我们必须将目标定为‘比 Waymo 还要安全两倍’！ — Alfred Lin > `20:00` 如果你的汽车跑了一百万英里，那绝对令人刮目相看，而我们的系统早已在这个量级上稳定运行 — Keller Rinaudo Cliffton > `26:15` 机器人并非在剥夺人类的工作，而是在‘赋能人类向上跃升’——专业人员不再去低头做基础苦力，而是站在战略层面上统筹协调上百台自主机器人。 — Keller Rinaudo Cliffton > `36:55` We do this not because it is easy, but because we thought that it would be easy — Keller Rinaudo Cliffton > `36:55` 你绝不是预谋要做垂直整合的，你只是在一次次抓狂和绝望中意识到：我们必须把所有这些垃圾全扔掉，一切从头自研。 — Keller Rinaudo Cliffton > `40:00` 今晚你和我一起通宵把这事搞定。 — Keller Rinaudo Cliffton > `42:08` Question every requirement — Keller Rinaudo Cliffton > `42:08` 一架飞机上最可靠、绝不会发生故障的零件，就是那个被你彻底删掉的零件。 — Keller Rinaudo Cliffton > `50:00` 当你预估某项硬件成本是 X 时，请直接假设它需要 10X。假设实际成本是它的10倍，商业模型还能跑通吗？如果真变成10倍，你该怎么应对？ — Keller Rinaudo Cliffton > `51:40` No tip, exclamation point! — Keller Rinaudo Cliffton > `52:45` You adopted the darkness, I was born in it — Keller Rinaudo Cliffton ## Mentioned Zipline, Keller Rinaudo Cliffton, Eric Watson, Alfred Lin, Pat Grady, 红杉资本 (Sequoia Capital), 卢旺达, 美国国务院 (U.S. Department of State), 美国国际开发署 (USAID), 宾夕法尼亚大学 (UPenn), 美国联邦航空管理局 (FAA), Waymo, 波音 (Boeing), 波音 777, 波音 787, NVIDIA, NVIDIA GPU, SpaceX, Elon Musk, Raptor 引擎, Tesla Model 3, Tesla Model Y, Tesla Roadster, Lotus Elise, DJI (大疆创新), Predator 无人机, Uber, Apple Pay, Face ID, 《安德的游戏》(Ender's Game), 《蝙蝠侠：黑暗骑士崛起》, 贝恩 (Bane), 金门大桥 ## 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 [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI2621449336.mp3)

[2:30] 投资人: 这在美国不是违法的吗？ [6:05] 卢旺达医护人员: 病人生病是不分昼夜的，为什么你们每天只营业 12 个小时？ [17:00] Alfred Lin: 这是个毫无意义的目标，汽车是极度原始陈旧的技术。Waymo 的自动驾驶已经做到比人类汽车安全约 10 倍，我们必须将目标定为‘比 Waymo 还要安全两倍’！ [20:00] Keller Rinaudo Cliffton: 如果你的汽车跑了一百万英里，那绝对令人刮目相看，而我们的系统早已在这个量级上稳定运行 [26:15] Keller Rinaudo Cliffton: 机器人并非在剥夺人类的工作，而是在‘赋能人类向上跃升’——专业人员不再去低头做基础苦力，而是站在战略层面上统筹协调上百台自主机器人。 [36:55] Keller Rinaudo Cliffton: We do this not because it is easy, but because we thought that it would be easy [36:55] Keller Rinaudo Cliffton: 你绝不是预谋要做垂直整合的，你只是在一次次抓狂和绝望中意识到：我们必须把所有这些垃圾全扔掉，一切从头自研。 [40:00] Keller Rinaudo Cliffton: 今晚你和我一起通宵把这事搞定。 [42:08] Keller Rinaudo Cliffton: Question every requirement [42:08] Keller Rinaudo Cliffton: 一架飞机上最可靠、绝不会发生故障的零件，就是那个被你彻底删掉的零件。 [50:00] Keller Rinaudo Cliffton: 当你预估某项硬件成本是 X 时，请直接假设它需要 10X。假设实际成本是它的10倍，商业模型还能跑通吗？如果真变成10倍，你该怎么应对？ [51:40] Keller Rinaudo Cliffton: No tip, exclamation point! [52:45] Keller Rinaudo Cliffton: You adopted the darkness, I was born in it

[0:00] people get sick 24/7. Why are you guys only open 12 hours a day? [5:00] All the customers who are like living on Zipline today, they really don't care about the technology operating behind the curtain... they care about their ability to download an app... click a button and have it delivered to them five minutes later. [15:00] We don't just want to say, 'We ran the test campaign and nothing failed. We're done.' It's like, no, no, let's take this thing to failure, right? Let's see where the limits are. [20:00] People have this perception... most drones are really freaking annoying. Every part of the vehicle is designed with sound in mind to be as quiet as gentle leaves moving in trees. [30:00] If you've seen Tesla Model 3s and Model Ys delivering themselves to customers, Zipline aircraft will fly straight from the factory into operation. [30:00] 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. [35:00] Keller Rinaudo Cliffton: We do this not because it is easy, but because we thought that it would be easy. [35:00] You might think, 'Oh, like drones. I mean, there are already lots of drones because DJI makes plastic quadcopters... and the U.S. buys $20 million Predator aircraft.' The reality is actually both of these systems are very unreliable... [40:00] The most reliable part on an aircraft is the part that is not on the aircraft at all because you deleted it in the last design. [40:00] So do we build these next week with the meter-long tail hook? Or do we delete the tail hook and put the two-centimeter-long tail hook on the back and bet that we can get this thing working? We got in a room, we're like, 'Delete it.' [49:58] No tip!!! [52:16] You merely adopted the dark; I was born in it

[0:00] I remember being in Rwanda early days and going out and meeting with some of the doctors and lab techs that we were serving and asking for them, like, you know, how's it going? What do you think? What's your feedback? Here I am kind of, you know, up and coming, you know, learning engineer thinking they're going to say something about the drone or some of these things. And the main piece of feedback that I received was people get sick 24 seven. Why are you guys only open 12 hours a day? Right. Especially when you're delivering life saving blood. Yeah, exactly. And so that was a really key insight for me where it's like, man, we have found product market fit in a market where, yeah, our, you know, our product wasn't great yet, but it was solving a real need. And so having that really beachhead market where there's a real problem being solved. And when your customer is telling you that their main feedback is they want more of your service, it's like, that's a good sign. Welcome, Keller and Eric, to the show. You guys have been working at Zipline for a long period of time. Keller is the co and Eric you are in charge of systems engineering and safety And we got lots of things to talk about in this whole world of drones drone systems and how you guys started in this hardware space before LLMs even started So we have lots of questions. Awesome. But you don't like Zipline being described as a drone company, even though you're probably the largest autonomous drone company in the world right now. I mean, you know, we we've always wanted to be an extremely customer obsessed company. And the reality is none of our customers care at all about drones. You know, like we our goal was always to build an automated logistic system for Earth and to approximate teleportation. And all the customers who are like living on Zipline today, they really don't care how they don't care about the technology operating behind the curtain. And what they care about is their ability to download an app, open it up, see a huge number of different brands and amazing restaurants that they want to shop with, and then click a button and have it delivered to them five minutes later. So we've always really tried to focus on the experience rather than on the specific technology. Well, this show is about technology. We're excited about that. This podcast is about technology What is the underlying technology behind ZipLine You started in 2011 You pivoted in 2014 This is way before anything related to AI robotics foundation models anything related to that But you were before all of that and you're riding the wave of all of the things that have come afterward as well. Yeah, this was when starting a robotics company was the dumbest thing you could possibly do? Like why, you know, you're talking to an investor in that time about, I mean, it wasn't easy and it was particularly hard because so many of those conversations, you know, I mean, I was 23, 24. Eric joined the company around that time and we were starting to describe this vision of an autonomous logistic system for Earth that would be 10 times as fast, half the cost, zero emission. You know, one of the biggest problems when we're trying to raise money for that vision was investors would say, isn't this illegal in the US? In fact, I think that that's a question you asked me when we started talking about this. We weren't allowed to fly beyond visual light of sight. I mean, we weren't allowed to fly at all, really. But like, yeah. And so the answer was, yes, it's illegal. And then most investors would be like, well, we don't invest in illegal things. So we're not going to invest. But for weird reasons this is what basically took Zipline down this path of like well if it illegal in the US then we can launch in other parts of the world where the value of the service would be extremely high Zipline decided to launch in Rwanda in 2016 delivering blood transfusions directly to hospitals and primary care facilities This enabled us to have a use case that was so powerful that a government would work very closely with us to make it happen. Make it legal. And to make it legal, or at least make an exemption to their existing kind of regulatory framework. And, you know, and then the other thing, you know, when it comes to how to think about Zipline as a company, you know, when we launched in 2016, we were like, we have this really cool drone, we put all this work into designing this really cool aircraft that, you know, and it's, it has all these great fundamental features. And when we launched, it was a total disaster. Because the reality what we what we learned in that first year, for the first eight, we'd signed a contract to sign 21 to serve 21 hospitals. And we serve one hospital for the first nine months. And Eric in particular, how much did you sleep during those? I spent a lot of time in Rwanda and didn't sleep a lot. And then when you're in the US, we'd get woken up at midnight because that's when the distribution center was turning on and everything would be broken. Nothing was working. It was totally desperate, constant all-nighters and working through the weekends because we made this big error, which was the thinking that the cool vehicle was the majority of the solution. [5:00] What we learned during that first year is that the drone is 15% of the complexity of the solution. The physical drone, the hardware of it is only 15%. We had to build so many auxiliary software systems, maintenance systems. How do we hold the inventory and do inventory management? How do we integrate with a national civil aviation authority, which we'll talk more about? How do we integrate with a national healthcare system? How do we do ordering and demand management? You know, we had to build out all of these other parts of the overall logistics system. So this is the reason I think a lot of people might look at Zipo and be like, wow, it's a cool drone company where they build a cool aircraft. The reality is the aircraft is like 15% of the solution that's required to build something that just feels like magical, reliable teleportation 24-7, 365 to the now, you know, hundreds of millions of people who depend on the service. Speaking of 24-7, I remember being in Rwanda early days and going out and meeting with some of the doctors and lab techs that we were serving and asking for them, like, you know, how's it going? What do you think? What's your feedback? Just really being customer obsessed and wanting to optimize the product. And you know here I am kind of you know up and coming you know learning engineer thinking they going to say something about the drone or some of these things And the main piece of feedback that I received was people get sick 24 Why are you guys only open 12 hours a day Especially when you're delivering life-saving blood. Yeah, exactly. And so you got to start somewhere. So we started being open 12 hours a day and trying to expand and grow from there. And so that was a really key insight for me where it's like, man, we have found product market fit in a market where our product wasn't great yet, but it was solving a real need. And so having that really beachhead market where there's a real problem being solved. And when your customer is telling you that their main feedback is they want more of your service, it's like, that's a good sign. Yeah, we were 24-7 within the first year. So we went 24-7. We're now 24-7, 365. I mean, on Christmas Day, I usually call all of our different distribution centers to thank them and check in with them. So there is no day when And these facilities don't depend on, you know, we went from serving one to 20 to 500 now to 5,000 hospitals and health facilities across the world, across eight countries that are served by the system. It's become the largest commercial autonomous system on earth. Can you size that for us? The largest system on earth? Just crossed 140 million commercial autonomous miles, which I mean, how many times? I think that that's like to the sun and back. One of the things that I like is every road in the United States there a lot of roads in the United States driving on every single road more than 30 times Wow That good said That a lot That a lot just to put it in perspective You know, seeing the impact that that system is now having across all these eight countries. I mean, University of Pennsylvania just published a study showing a 51% reduction in maternal mortality thanks to Zipline. So half as many moms losing their lives in childbirth. We have, you know, across all the different use cases Zipline serves, some of our partners estimate that we're saving between 10 and 12,000 lives a year. And that impact is growing exponentially as we're now expanding, especially as a result of this new partnership we have with the U.S. State Department. What is that partnership with the State Department? In December, we announced a $550 million partnership with the U.S. State Department to expand the impact of Zipline's life-saving service across a lot of the countries where we're already operating. So with USAID being shut down, the U.S. was really seeking new ways of engaging in these countries and helping save lives in these countries. But they wanted to do it in a way that would accelerate the economies of these countries and help the U.S. economically. And so the new strategy they're calling commercial diplomacy. The idea is that we want all of the developing world should be built on top of U AI and robotics technology We should be going and economically helping We should be bringing the best that the U has to offer The interesting thing is when you talk to these countries about what they want they tell you they are sick of low aid provided by NGOs for free because these services engender dependence and prevent economic growth in the countries. What they want is high paying jobs, entrepreneurship, technology. And so the U.S. is going through a big strategic shift where it's like, well, we have that. We have those things. So let's basically go out and incentivize these countries to adopt that kind of infrastructure, make sure that as these countries are accelerating, they're doing it using US robotics and AI technology. And this is something that will be great for those countries. It saves lives, it saves them money, but it also means that it will make it possible for the US to secure our lead in manufacturing and robotics over the decade to come. I'm curious about, you know, you guys, because you now run the largest autonomous system in the world and you launched it 10 years ago at this point. So you've been in production for 10 years. You've learned a lot of stuff that your average engineer sitting behind a computer screen has no idea they're going to run into when they try to deploy AI into the real world. And so I'm curious what some of those lessons learned are and maybe what's [10:00] One way to ask the question is, what popped up over the last 10 years that you never would have guessed you needed to be good at when you first started launching these systems in 2016? Yeah. We started off delivering lifesaving products, right? And our customers need lifesaving products all the time in all weather conditions. And you would think it's wind, these things. But one of the weirdest things is actually solar weather. So there's solar flares that happen on the sun. So they're basically big explosions that send radiation to the Earth. They can mess with the ionosphere and that can cause basically the RF signals coming from GPS satellites to be faster, slower than you expect. And that can lead to degradation and challenges in navigation systems. And so here's one example that when we were starting off, we didn't think that this was going to be something we have to figure out. But we actually have gone pretty deep in this space. And really, it's two things. One is designing our navigation system and our GNSS systems to be robust to these conditions to ensure that we can still know where aircraft are with centimeter level precision in those conditions, in those challenging solar flare times as well as designing the system to have redundancy beyond GNSS such that if things get really bad we can still safely operate Eric you in charge of safety Tell us about what you learned about safety today and specifically about the compute failover system that you have. Yeah, yeah, absolutely. I mean, there's so many things that we've learned over the last decade of operating the system in the real world. One of the things that we're proud of is how we've developed, to your point, compute failover. So there's a flight computer, flies the aircraft, Lots of sensors come into this computer and that basically does a lot of math and sends commands to actuators, right? So motors, control services, these things. So this is the brain that flies the aircraft, right? One of the things that we've learned is you need to assume that any part of the system can have a fault, can have a hiccup. Something can go wrong. And that's how you really design something to be robust, reliable, and safe. So what do we do if this flight computer has a challenge? It could be a software challenge. It could be a connector challenge. It could be these different things. Bit flip due to solar radiation. All kinds of things, right? And so what we've done is we have two flight computers. And both of these flight computers think that they're flying the aircraft at any given point in time. They all are receiving all the information from the sensors They all sending commands to the actuators And there like a kind of a third arbiter a little computer that is monitoring the health of those two and telling everyone every other node on the on the aircraft who to listen to who actually in charge What if the arbiter fails? If the arbiter fails, then the primary computer that was flying just keeps flying. Right. So one's in charge. And if the thing that's monitoring its health fails, then now we say, OK, like, you know, now we're just going to keep flying on the thing that was good and we're going to keep flying the mission. um so two heads have been one yeah something we're really proud of um we had actually had one of these events happen uh a couple weeks ago yeah where we had 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 aircraft flew itself home landed everything was totally fine so just you know designing the systems to be robust and reliable through and through is is you know how you get two and a half million deliveries and 140 million miles flown with no safety incidents. And a lot of what Zipline is doing, it's not like, oh, this is totally revolutionary. No one has ever thought about having a secondary flight computer. That's how Boeing 777 works. But the cost of a flight computer on a Boeing 777 is in the millions of dollars. And so a lot of what Zipline is having to do is take a lot of the best ideas that you can see in aerospace safety best practices best practices And then you got to figure out how to build that using components coming out of the smartphone supply chain You can do it for tens of dollars or hundreds of dollars. You can achieve similar levels of safety to traditional aerospace, but you can move a hundred times as fast at one one hundredth of the cost. So you mentioned that the aircraft is only 15%. Describe the other 85% in layers and maybe go down deep in some of your systems that are really, really sophisticated. Like I know this because of being a board member, like the detect and avoid systems. So, you know, how do we test? Why do we test? Really, the way I think about it is there, maybe first of all, we're not a software company, right? We're a real world AI robotics company. And so there's electromechanical systems out in the real world. So there's hardware test aspects, there's software test aspects, and there's the integrated system test aspects. We have a lot of different environments that we test, a lot of different approaches. I'll name a few of them. on the hardware side, we do a lot of component level testing, halt testing, highly accelerated lifetime testing, where we're taking components, maybe it's a motor, these kinds of things, and we're putting them through hell, right? We're putting them through all kinds of challenging conditions, [15:00] hot, making it humid, making it corrosive, all of these things while we're exercising, you know, while we're spinning the motor, while we're moving things, all of the things, right? UV, like you name it. And just to give a context for a scale, I mean, there are 700 unique components on the aircraft designed from scratch by Zipline. We are designing not just the flight computer from scratch, the power distribution board, the motor controllers, the battery, the battery management system. The, you know, the pod is the smaller robot that we're using to actually make deliveries to people's homes. There's an entire NVIDIA GPU powered flight computer on the pod. We're building the electronics that go into the docking station where the zip is flying in and out of. All of that, even the electric motor being designed from scratch by Zipline because we need a thrust to weight ratio that is not available in off the shelf electric motors. So you have to design something from scratch. So 700 unique components, 43 major sub-assemblies on the aircraft, all then and coming together on the manufacturing line that you both have gotten to visit and getting assembled into one overall aircraft. But anyway, that's the scope. So for each of those components. Yeah, right, going through this type of testing. And thinking about other industries, oftentimes when I talk to people from maybe automotive or aerospace, some of these I like hey how do you think about reliability challenges And a common answer is like well I asked the supplier what the reliability of the part is Yeah I like OK cool Like what if we the supplier You know so anyway so we that vertical integration where we have component testing on the ground. We have system testing on the ground where we're taking full aircraft, you know, as well as other parts of the system and putting them through vibration tables, wind tunnels, thermal chambers that you can walk into, like all of these things in order to understand how is this going to break, right? More than just, is it good enough? Like, we want to know how it's going to break. And then we can understand, okay, cool. Like, let's make it better. Or maybe it's like, oh, that's not too worrisome. Like, great. You know, we didn't break any of the ways we're worried about. It broke in that way. Fantastic. So, like, we don't just want to say, we ran the test campaign and nothing failed. We're done. It's like, no, no, let's take this thing to failure, right? Let's see where the limits are. 49 degrees Celsius, which is very hot. Hot, cold. Down to negative 25 degrees Celsius, which is very cold. All the things. You don't fly anywhere at 49 degrees. We do. You do? We wouldn't test at 49 if we're not worried about that. Where are you flying at 49? I think Phoenix during the summer. Phoenix during the summer, yeah. And then where's minus 25? It's in northern parts of the United States. Actually, can I ask you guys how you think about that? I could imagine a different version of the world where you guys are like hey look if it too hot we just not going to fly Totally And if it too cold we just not going to fly And if it raining too hard we just not going to fly Yeah And like there are tradeoffs to be made you know and obviously your customers would prefer that you fly at all times. That's right. But how do you think about those tradeoffs? The easiest way to think about the tradeoff was because of the use cases that Zipline started with. That's right. That makes sense. Which was basically like saving lives. Yeah. You can count on us with your life and the lives of your loved ones as long as the sun is shining. Yeah. it's not super compelling you know you develop the capability because you had to for the initial use case zip line would fly and in fact i mean for the first you know for the first um couple years we took a lot of risk i mean we would basically fly we were like look if it's a life-saving delivery happening and there's someone whose life is on the line yeah we're gonna go for it and we had a civilization authority that was you know generally great partner with us on that front. We took a lot of risk. We learned a lot. And, you know, almost always it worked out in favor of like we saved the person's life. And, you know, the worst thing that could happen was we had a Paraland, which is the kind of like zipline safety mechanism of last resort is we can pull a parachute on the aircraft and bring it gently to the ground. But we learned a lot. It happened very often in the first few years Like yeah very very rare today I mean to put it into perspective you know our original goal was to be 10 times safer than cars Actually Alfred was the one pushing in our last board meeting He like that a BS goal We need to be two times safer than Waymo And so Eric literally went and reset the goal. The Zipline's target for the end of this year is to be two times safer than Waymo. He's like, cars, that's like archaic technology. Waymo, I think, is about 10x, right? 10x cars. So our goal is to be 2x safer than Waymo. It's not the right comparison. You're flying. You have to be safe in the air, not safe on the ground. I think it depends. We're substituting something that's typically going in cars, so it's debatable. But suffice it to say, we now have 140 million commercial autonomous miles and zero safety incidents. Zero. If you were to drive 140 million miles, you would have 600 accidents, 100 injuries, and somewhere between two and six fatalities, depending on what country you're talking about. And this is why we really pride ourselves on picking the right use cases. It's like it's lifesaving and it really makes a lot of sense to go do it. And also, we're going to be by God, we're going to be as safe as humanly possible from an engineering and testing and validation perspective. We really take that. That's it. That's a deep part of the DNA of the company. One last point. You know, what is the outcome of all of that testing that Eric is talking about? The outcome of all that testing is we have individual aircraft in the commercial fleet that have flown. [20:00] more than a million commercial autonomous miles. And so I think people, you know, that's just from an intuition perspective. A lot of people look at this and they're like, wow, it kind of seems maybe exquisite or fragile, probably very sensitive to like extreme conditions or weather. I mean, you know, raise your hand if you have a car that has a million miles on it. It's pretty impressive. These systems are already like way more rugged and durable and robust than people necessarily think. Can I ask you about the precision? Like one of the things that blew my mine when I saw some of the, I haven't had a chance to experience in person, you know, a delivery. Gotta come Pat. I know. I gotta go experience it. But just, but just in watching the videos, the drone's a hundred feet up and it drops the package. It lowers the package to a, I don't know, a circle that's got a 18 inch radius or whatever it is. Right. Like how do you guys achieve such precision even when it's windy, even when it's raining? Yeah. How do you pull that off? First of all, the aircraft's about a hundred meters up. A hundred meters up. Okay. Yeah. Yeah, there you go. It makes it harder. And, you know, the multiple layers, there's the delivery pod that comes down. So the delivery pod comes down. That really the delivery and pickup like precision part of it So the drone is hovering above It knows where the target is Maybe let say it this coffee table for example there wasn a roof above us This is this coffee table And so the aircraft is going to hover above, but it actually needs to consider what the wind conditions are. Yeah. Right. So if the wind's blowing in one direction, then the aircraft's going to kind of be shifted upwind. Right. It's going to shift in the direction to help with those wind conditions. And it's going to lower that delivery pod down. as keller mentioned we we do uh take advantage of gnss so real-time kinematic gnss that gives you centimeter level um confidence of where you are but the thing is we don't know the gps coordinates of this table right it's not like someone came and surveyed the middle of the table and sent us the coordinates right no one wants to do that so what we have to do is we kind of that we use that to kind of get close right we're like okay here's the backyard here's where we kind of know things things roughly are. And then the job of this delivery pod is to be lowered down, you know, fight the wind conditions, fight these different things and be able to use its onboard perception autonomy systems to identify where's the best place for me to leave the package, right? Like if there's a little table and there's a whole bunch of drinks, probably I shouldn't, you know, try and drop down on top of these drinks and make a mess Maybe I should go to the ground right next to the table right And so it has these autonomous you know onboard real compute to be able to identify what am I looking at what am I seeing and how can I find the best place to leave the package, and then come down, touch the ground, opens its doors, gets retracted back up, and there you go, the package is left on the ground, and the delivery pod comes back up, stows, and the aircraft flies back home. A couple big advantages, I mean, just to be specific, so that pod, it not only has its own NVIDIA GPU running its own AI autonomy stack, So it's a survey and like know exactly where it's delivering even at night. Yeah. But it's also controlling its own position. That's right. In the X and Y axis. So it can not just know but then move. And the advantage of that architecture, which you can probably guess, but there are two huge advantages of doing it in this way. One is it's quiet. People have this perception of, I mean, first of all, most drones are really freaking annoying. Like the sound is just it's basically the most grating, annoying sound that you could possibly subject a human to. And so, you know, like we, Zipline has a big team of aerodynamics, aeroacoustics and controls experts. Every part of the vehicle is designed with sound in mind for the vehicle to be as quiet as humanly possible. We want it to be no louder than the sound of like gentle leaves moving in trees And when the pod is delivering we keeping the main aircraft 100 meters in the air Yeah it like the thing that is creating noise is really far away. That's also a huge benefit from a safety perspective. Because the only thing that is coming anywhere close to you, your family, your pets, your kids is something that is super cute and safe. And it's really like a styrofoam, kind of like a cute anthropomorphic styrofoam tub. tub. How long was the technology tested outside the United States before you came to the United States? And what's the path to getting into the U.S. now that you're flying in Dallas and delivering packages there? We spent eight years, I think, right? About eight years. I mean, depending how you measure it, maybe like six to eight years. And then it was, I mean, we launched in Rwanda in 2016, our commercial service. And we really launched this kind of next generation home delivery service, the thing that's now like in sort of insane hyperscaling mode that only launched January of last year. So depending on how you measured it, you could even say it was like almost nine years. And then when you got to the U.S., was it just smooth sailing? What was the sort of regulatory path that you had to go through? Yeah, I mean, we really started, I would say, you know, like meaningfully engaging with U.S., you know, FA. [25:00] and other regulators in the US around 2020 or so. So it doesn't mean we didn't show up in 2025 and everything was smooth sailing. It was really a partnership of working through, as you mentioned in kind of 2016, all of this stuff was, there was no pathway. It was kind of illegal as we joked earlier. And so, yeah, so it really was a partnership to identify, hey, you know, we have shared goals, right? Our shared goals are safe and efficient airspace integration. And so while we have experience doing that successfully in different countries, we can bring some of that experience in. We have opinions on how this should work. The regulators had opinions on maybe how they thought it should work. And so it was a partnership over the course of a couple of years to identify what those paths look like and how we could kind of converge and align before we were able to execute on that. And you had to show your ability to manage all these aircrafts that are flying. So you wrote systems, you built systems. Yeah, I think it's a huge part of, you know, Keller's mentioning that the drone is only a part of the, of the, you know, of the overall system, the overall complexity. What we're really building is an infrastructure layer, right? We're building an infrastructure layer that can enable instant access to products. And you don't do that with one aircraft flying from one place to another place. You do that with a network of charging locations hundreds of aircraft spread across an area that with the autonomous systems in the cloud that can understand where am I having demand Where do I have supply? Where do I have aircraft? What's coming up is about to be the dinner rush. What's the weather at these different locations? How can I kind of self-balance these things? As well as how do I efficiently pull in people when needed, right? So these aircraft are autonomous. They're operating. They don't require human intervention through these flights. But there are times in which it makes sense to alert a person that, hey, maybe there's an issue here, or the weather is a little bit, the wind is climbing in this area, right? So there are humans, trained aviation professionals that are monitoring our network, I would call it. They're fleet commanders. Fleet commanders. That's right. We used to call them pilots, because when we originally launched in the US, the first regulatory permission we got was to fly one-to-one. So that meant that we had one pilot sitting in an office, remote pilot in command, who was sitting in an office basically just observing an aircraft do its thing. And again, it's exceedingly rare that a human should ever have to issue any kind of a command to a vehicle, but we would have one human watching one aircraft. Not great for unit economics but as Zipline proved out these systems we went from one to one to 1 to 3 1 to 6 1 to 20 1 to 40 We now operating 1 to 100 and have plans to go well beyond that One fleet commander now. So yeah, we technically changed the name because I think pilot's confusing. So we're inspired by Ender's game. We now call this group of this team of people at Zipline, we call them fleet commanders. And it actually says that in the FAA documentation. We say fleet commanders shall do the following. and yeah they are overseeing a group of 100 aircraft and to me this is like the exciting cool thing about technology because people think about like well you know what about um you know how humans used to solve this problem it's like it's not you know the it it's cool how robots enable humans to like up level right like the human is still getting to like strategically manage the system it's just the human is now maintaining and commanding robots rather than like doing the actual work herself. Now that you guys are kind of in hyper scale mode, you've solved so many problems in the last 10 or 15 years. What new problems are you running into? Yeah. I mean, what I would say like thematically, I mentioned earlier that, you know, getting to two and a half million deliveries, the, you know, the, oh, it only happens every couple of years. It like kind of a one in a million chances Yeah These things start to matter right We we on the path towards a million deliveries every day And if you have a one in a million situation it going to happen every single day Yeah Can I just to really make that clear So it took Zipline from 2014 when we started building the original version of the technology, to 2024 to do our first million deliveries. Was it the end of 2024? Maybe it was even early 2025, actually, that we had done a million deliveries in the cumulative history of the company. So it was almost a decade, maybe say about a decade to do a million deliveries. Zipline is now in the very near future going to be doing a million deliveries a day. And so that is definitely humbling. It's like, wow, okay, everything about the way we've been solving the problem is going to break. The bar goes way, way up. And I mean, one specific example, maintenance becomes really hard. The scale of the problems, the number of vehicles that you're managing in the fleet, the cost of a screw up or if a certain process is operating in very inefficient ways becomes extremely high. And so there's just high degree of criticality for all these systems. One interesting point though, there are a lot of ways that these systems operate that I think people don't yet appreciate the advantages of autonomy. One good example, [30:00] is that the system wants to operate 24-7. It does operate 24-7. So I think people are used to logistics as generally being like, well, here are the hours when humans are driving trucks. That's not how these systems operate. They want to operate 24-7. They can be fully utilized. They can be as happily delivering at 2 a.m. and 3 a.m., delivering something so it's ready for you on your doorstep or in your backyard when you wake up at 6 a.m. before you go to work, as they are delivering at 2 p.m. They can deliver in five minutes. They are available 100% of the time. We are soon going to be flying vehicles straight out of our factory in South San Francisco into commercial operation. If you've seen Tesla Model 3s and Model Ys delivering themselves to customers, ZipLine aircraft will fly straight from the factory into operation. It's a huge advantage from a maintenance perspective that as soon as a vehicle needs to go through some kind of proactive maintenance, it will fly itself to the maintenance depot. So the human can then quickly do whatever process necessary, and then the vehicle flies itself back into operations. We can also dynamically assign capacity in a metro based on what the system is seeing. There's no like set home for a vehicle. It can go to wherever it's needed. Yeah I think to your question about you know getting to a million a day and what are the new challenges I think you know Keller hit on some of them to the previous thought about the drone is only 15 of the problem Really it the way that we currently manufacture aircraft maintain aircraft, support all these things, you know, troubleshoot problems. Like the way that we do it today isn't going to work when we're at a million deliveries a day. And so there's like, okay, we need better tools. We need better software systems. We need better processes. We need better, you know, all these things. So it's like, you know, Elon talks about designing the machine that builds the machine. And so, you know, this is really one of the things that I see Zipline tackling over the coming couple of years is we are going to be investing much more in the machines that build and run the machines. I mean, from a scale perspective, I think the largest airline in the U.S. is doing about 5000 flights a day. Yeah. Zipline is going to surpass that in the next month. And when we get to a million deliveries a day, ZipLine will be doing like somewhere between, I don't know, 40 and 80 times as many flights in the US in commercial airspace as all other airlines combined. And so it's obviously a completely different class of aircraft. It's a totally different kind of problem. But the reality is when you look at air traffic control, they don't make a distinction. And so there also when you talk about all the auxiliary systems that have to be built there is a huge transformation that going to have to happen in air traffic control as we start to realize that people are really excited about electrification of vehicles People are excited about autonomous vehicles. The reality is, as those transformations occur, 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. And so, like, the sky is a big place. It makes sense to utilize it. You can give Earth back to humans. You can make neighborhoods quieter, safer, less pollution, less traffic. You can make huge improvements to Earth if we can more effectively utilize the sky. This is going to require huge transformation of how we think about air traffic control in the US. And it means that we need to design it with AI and autonomy in mind rather than the way it was designed, which was in 1950 using pencils and paper and note cards and like a human looking out trying to watch the airplane. Are you helping the FAA design that? It's really, yeah, I mean, what needs to happen is like collaborative innovation is one way to put it, right? It's like that one company solving this problem for themselves is not going to solve the problem for the industry. And so we are heavily involved in I mean first of all a key part of the solution we believe is aircraft should be talking to each other They should be telling each other where they are They should be automatically detecting that hey there a conflict on the horizon here And so therefore, we're going to, you know, you go up, I go down, right? These kinds of these kinds of things. And our aircraft do that. And we're working with other kind of, you know, other new entrants into the airspace, autonomous aircraft and autonomous drones to do the same thing, to make sure that our systems can talk to their systems and we can all collaborate to make sure it's efficient and safe usage of the airspace. We're also, to your point, Alfred, working with regulators, working with standards bodies to take some of these best practice and innovations that we and others have developed and try and make them broadly accepted and utilized so that we can all collaborate and we can all safely and efficiently use the airspace. Because you guys have developed a really sophisticated detector of the right system. Yeah, because when we were launching in all these other countries, we had to build something from scratch. And so we built the thing from scratch. We provided all this software to the Civil Aviation Authority so that they could use it to monitor this entirely new class of autonomous vehicles in the airspace. Interestingly, there are multiple public companies in the United States that build air traffic control software that are worth more than $10 billion, right? I often look at that. I think there are many companies inside Zipline that are likely, it's like, oh, that's a public company inside Zipline. [35:00] having to get built from scratch. We're building it because every part of the ecosystem we sort of had to build from scratch to enable the overall technology to flourish. You know, air traffic control is an interesting, like the more you learn, the more disturbing it is. I mean, we're starting to see the impact. You know, you read about like, you know, a plane crashing into a helicopter in DC a few months ago. You read about like two planes colliding on, I think on a taxiway in an airport. I don't remember where that was a month ago. You're like, wow, why are all these accidents happening? Turns out like 50% of air traffic controllers are over the age of 45. 20% are about to retire. And nobody is going into air traffic control as a career path right now in the US. And so there's actually a huge labor crisis around these kinds of jobs. And so you have pressure coming from different angles for like transformation is required. that we cannot use a system that was designed for airspace in the 1950s. The labor isn't available to do it even if we wanted to. And also there is this giant influx of new technology, AI, and autonomous vehicles that are going to require us to transform how these systems work. So you a hardware company and a software company You design your own questions Operations manufacturing You design your own parts you build the own aircraft you write your own software you do your own operations It's a pretty vertically integrated company. Talk about the benefits of complete vertical integration versus buying component parts or buying component software and putting it all together. And how you get people who come from such different disciplines and domains to see eye to eye and work together collaboratively. Yeah. I mean, I think that interestingly, you know, this is doing it is such an incredible pain in the butt that you would never do it. Like if you, you know, 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. And this is definitely like the definition of Zipline, you know, and it's such a pain in the butt actually that it's almost if you look at the history of all these hardware companies, they all try to not do it first. You can look at the Roadster, right? They're like, we're going to use a Lotus Elise chassis. We're going to buy the battery pack from a secondary supplier. And we're just going to put the two together and it's going to be awesome. You know, kind of Roadster lost a lot of money and wasn't very reliable. Right. But like it was an important part of getting to the Model S. Zipline when we started you know Eric knows well we were like buying everything from suppliers We were like you know paying people to design different parts of the system for us or trying to buy off the shelf stuff And we crashed airplanes I mean at test sites And we just crashed and we crashed and we realized, wow, this stuff is like super expensive and it's also totally unreliable. And so, you know, part by part, you're like, all right, well, like rip that out. We'll design the motor controller from scratch. Okay, rip that out. We're gonna have to describe, you know, design the GPS module from scratch, navigation system. So, you know, part by part, you sort of like rip it out. And I think there's a fundamental realization, probably similar to the realization that happened that made the Model S possible is like, hey, if we want to build a really great specific product in this totally new area of technology, we're going to have to design every single one of these components from scratch to meet the specific requirements of this new area. You know, you might think, oh, like drones. I mean, there are already lots of drones because DJI makes, you know, plastic quadcopters and they make millions of them. and like the US buys $20 million Predator aircraft that can fly a hundred miles. The reality is actually both of these systems are very unreliable and nothing is in a level of like reliability and safety at unit economics that would work for this new industry that Zipline was trying to kind of like pioneer And so we realized we had to go build like an automotive grade solution It has to be super reliable and it has to be extremely cost effective because you competing against cars and motorcycles which are actually really cost effective And we've got 100 years to make them reliable and cheap. So you never do it, I think, intentionally. Maybe just like slowly freak out and through desperation realize like, wow, we've got to tear all this shit out and we've got to build it all from scratch. The advantage of doing it from scratch is speed and integration. And so our offices, you guys know because you've been, but when you visit Zipline's offices, we are all absolutely packed into sardines, into this small building where you have firmware engineers sitting next to, mechanical engineers sitting next to, autonomy engineers sitting next to, cloud infra sitting next to, aero, aero acoustics, guidance navigation controls, systems engineering. manufacturing, everything, everybody in one place. And then our factory is a three minute drive away. And so our team is like on the factory floor, working, seeing parts get integrated into the overall system. And then we have our test sites, which are just a short drive away. So you can go to the test sites, watch the vehicles flying, observe how the system is performing. Like combining all these things together means that, [40:00] Stuff is always breaking. Stuff's always going wrong. As Erica described, the advantage is when the thing goes wrong, we can basically go straight to the person's desk and be like, you and I are pulling an all-nighter tonight. Whereas if you're Boeing and something's going wrong with the battery on the 787, you're like going and suing a supplier and taking two years to try to figure out whose fault it is. And it's three layers deep in the rat's nest cluster of how these procurement deals and supply chains work for aerospace is why it's so broken. Yeah, I think, Pat, to your kind of question there about getting these different disciplined folks to work together. Yeah. I honestly think it's quite easy. It's easy when you have set up the way that Keller just mentioned, right? Like, first of all, everyone's rowing in the same direction. We all have the same goals. And when you can ground it in reality and it's tangible, then we're all just here to solve the same problems, right? So we actually, with the vertical integration, with having a very diverse team, we actually cut through a lot of the stuff, right? A lot of the things that happen where, oh, you know, that engineer won't tell me what the actual source code does because they said it's IP and so we don't actually know what the fault detection looks like and you don't have any of that You just like literally go walk over sit next to the person desk and be like hey we failed that test Tell me about how this part of the system works Oh cool Pull up the code Great Let look through it Oh interesting You making that assumption That's not how I designed it. Right. Cool. Let's get to the bottom of it. Right. And so this idea of just rapid collaboration where you're just, you know, the manufacturing team, the operations team, the engineering team are all just like really together is the way to solve these problems. And I have found that it's actually not that hard. Right. When you have those ingredients, it actually makes it, you know, makes it pretty fast. inefficient and you know to i mean eric saying that it really makes you realize when you build these like complex you know ai and robotic systems that combine hardware and software you really appreciate like the deep religious truth of how dumb requirements usually are question every requirement which is you know the number one part of like elon's algorithm they talk about at spacex like question every requirement is like this is like so profoundly and deeply true you must have every team question every requirement the requirement is always stupid when you and you're like well you know it's but that you go to this team that team you like often you have to dig like two levels deep to realize like this but um questioning every requirement is a fundamental part of like getting through this and then you know the other thing is um delete the part the most reliable part on an aircraft is the part that is not on the aircraft at all because you deleted it in the last design that part will never fail and you know you take a lot of inspiration from looking at like the raptor one raptor two raptor three i'm sure you've seen you know those engines to each other and actually a lot of people who come to the factory now and get to see like the ev3 aircraft you can see the ev2 aircraft the ev1 aircraft plus like the 10 different hardware versions that we built on the initial on the on the first version of ziplines technology, you would just delete, delete, delete. There's a huge amount of, it's really hard to delete things. It's an act of courage. No one wants to delete the thing. You look like an idiot if you delete the thing and then the system can't perform or doesn't work because you deleted the thing. But true confidence in the physics and the performance of the system enables you to start deleting things. It's a big advantage of having full stack integrated control of all of these systems. It makes it possible to question every requirement. It makes it possible to delete parts. Yeah, I think first principles thinking is a huge part of that. I remember the Platform One aircraft, early days it had a deployable tail hook is how it landed. So it had this big hooks like a meter long that would come down from the aircraft and we had a line that would catch that and slow the airplane down It this kind of complicated contraption And we had this idea that we should be able to move that complexity to the ground systems and have the recovery system the landing system more like an aircraft carrier like crab the airplane, right? We can put the actuation on basically a robot that goes up and grabs the airplane. And we're like, man, that's going to make the aircraft so much simpler, so much lighter, so much more reliable. We didn't have it working yet. And it was time to build that next generation of the aircraft. And we're like, so do we're building these next week, do we build them with the meter long tail hook? Or do we delete the tail hook and put the two centimeter long tail hook on the back and bet that we can get this thing working? We got in a room, we're like, delete it, right? Like, let's do this thing. And so like, from first principles, it should work, we can make it work. We haven't done it yet. But we can do it. And the next couple weeks looked like myself included a lot of people pulling a lot of a lot of late nights, getting that thing working. And sure enough, those first aircraft came and we caught them and landed them. So it's a lot of courage, but that like really being grounded in first principles thinking with a tight integrated team is how you do that. Is there a version of the future in which instead of delivering life-saving medicine and cheeseburgers, you're delivering human beings? Uh-oh. Hey, I'm the board member that has to control their costs. I mean... Yeah. [45:00] You know, safe, reliable, battle tested. I don't know. It seems like it seems like if we're going to liberate ourselves from the tyranny of streets, it's a pretty decent solution. Gosh, I think I agree with you. I think that, you know, come to you and ask for another billion dollars. I think, you know, a couple of thoughts. Like one is that I think Alfred knows I'm measured in the way I answer that question, because to be clear, like, you know, building a new infrastructure layer for the planet that can deliver packages as efficiently as the Internet moves information is going to be one of the biggest companies on Earth. Like it's a huge opportunity. And we definitely want to stay like humble and paranoid about how super hard that's going to be. the level of execution for us to scale the way we want to scale over the next couple of years. And, you know, to put into perspective, I described this goal of getting to a million deliveries a day in the very near future. We now have many partners who are each asking to buy a million deliveries a day of capacity from Zipline in the last few months. And so our operating plan has now become our unit of sale That a pretty crazy realization And it leading us you know We had originally built we had sized the entire entire factory to build uh 20 000 aircraft a year that was what about what was required for a million deliveries a day like all of this is kind of being thrown we realizing the market is way bigger and one one thing you know when you look at this you know totally hyperbolic curve that i think i showed you only a few months ago of like you know the what our total flights have you know total daily flight volumes have done over the last 16 months, the level of complexity of all the different systems that are required to basically like stay on that track is quite high. But, you know, there are five and a half billion instant deliveries being done by humans in the United States every year. And that's, you know, we're using a 4,000 pound gas combustion vehicle. It's not really instant. It's like half an hour to an hour. It's good marketing that it's called instant. But yeah, exactly. And, you know, 30 minutes, 45 minutes, an hour, you know, a significant percentage of the driver's report eating some of the food that they've delivered in the last month, like more than 50 percent. There are significant safety concerns associated with these kinds of delivery. But five and a half billion instant deliveries. what we realizing when you look you know Zipline is now at massive scale in Dallas and we now launching four more metros in the next four months When you just look at Dallas if you were to extend the buying behavior that we observing from Zipline customers in Dallas to the rest of the United States, there would be 55 billion instant deliveries happening, not five. 55 billion. Yeah. There's a huge market expansion. I think it's similar to how people looked at Uber when they were launching in San Francisco. And they're like, oh, even if Uber gets to be 33% of the taxi market, it's only going to be a $15 billion company. And obviously what they missed is like Uber is now 10 times the size of the taxi market. You know, like if you make something more convenient and less expensive and a better product experience, people are going to consume a lot more of it. We are clearly seeing customer behavior where customers order every day rather than a couple of times a month. I mean, I met a grandma the other day who's ordered 350 times from Zipline in the last year. She's 80 years old. Amazing. Actually, nursing homes are like big Zipline. Like they're like big demand centers for zipline it's probably pretty fun if you're in a nursing home it makes sense like and actually it's funny people perceive i think old people as like maybe being you know not capable of using technology they're all like living on their iphones you know like they they're probably doom scrolling actually which is maybe not a good thing but like um they are very comfortable using like you know apple id um apple pay or Face ID Apple Pay and just ordering and having it delivered directly to them So there are definitely not enough humans in the United States to do 55 billion deliveries Yeah. The only way we're going to be able to serve this kind of demand is with automated systems. And there's definitely not enough roads. And when you look at traffic in most of our major cities, you're like, oh, can we just maybe double the number of cars on the road so that we can do way more deliveries? It obviously doesn't work. We actually need to be taking cars off the roads if we want to enable human growth and flourishing. And so I think this change is inevitable. So how many flights are you doing a day now? And how many will you do in a month? Saplen is now doing almost 5,000 flights a day. and we're anticipating exiting this year at above 30,000 flights a day. And our goal is to get to a million flights a day as fast as humanly possible, which we expect to achieve in the very near future. Like all of the supply chain manufacturing capacity decisions we're making right now are designed not just to get us to a million deliveries a day, but also accelerate past that. The things that are interesting to think about on the unit economics front is like whenever we meet hardware companies and you always talk to them about like, how much do you think the system is going to cost? and they're always [50:00] like it's going to cost X. And you're like, cool, it's going to cost 10X, just so you know. Like when you build it, it's going to cost 10X. That's your advice to founders. That's my advice to founders is like for hardware companies, like if, because, you know, I'm like, try to, you know, be a good seed investor and pay it forward and stuff. And like, you're always meeting these founders and always like, it's going to cost this much. I'm like, cool. It's just like, assume it's going to cost 10 times that. Like, does it work? And what would you do if it costs 10 times that? And we're speaking from experience. Like when we launched our system in 2016, we were charging $30 of delivery to deliver a blood transfusion over 80 to 100 miles. And that was like cost comparable. And so that's what we signed the contract for. And we thought that we were going to launch a system that cost about $30 of delivery. How much do you think it cost when we launched? $300. Yeah, it cost $300 of delivery. And Alfred was surprisingly chill about it. And we were like, all right, we got work to do. And so the next year we got it to like $120. And then the next year we got it to $75. Then the next year we got it to $40. Then we got it to 28. Then we got it to 18. It's now 12 for the long range technology that we operate outside the US. Right now what happening this summer is 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 It is a cool moment that I think most people don really realize It happening quietly Like you not reading about this in the New York Times or whatever But I think that this thing is happening in the next month or two that is gonna have a big impact on the world and how the world looks and how most normal people even live their lives because it is now more cost-effective to use a robot in logistics than it is to use a human. And that's really good news for the environment. It's really good news for neighborhoods that are going to get quieter and safer and less traffic, less pollution. And it's really good news for customers because you can get things way faster and more reliably and for less expensive. You know, our customers love like there are obviously so many cool things about the system that you can talk about and that you see customers taking advantage of. But like no tip, exclamation point, exclamation point, exclamation point is like a big, you know, that's probably the number one comment. I think that customers love not having to feel guilty and being able to just have a system that they know how much it's going to cost. Well, thank you, Keller and Eric, for being here with us. I thought you were going to say it takes longer than you thought not 10x more than it costs But anyway that a great way to end It does also take a lot longer I mean I think the memo that Sean wrote here at Sequoia a few years ago I think is deeply true I don know if he'll ever publish that publicly or if it'll be allowed, but I do think, suffice it to say, there is an internal Sequoia memo that has had a big impact on me talking about A, why hardware companies are going to be some of the most impactful companies for humanity's progress over the coming decades and B, why it's super hard to get those companies off the ground and fundraise for them and see how investors should think about funding those kinds of companies. It's interesting when you look at the world today to see, I mean, wow, how fast the world changes. Because think about it, we spent 10 years being the freaking black sheep, like a hardware company, no thank you. Let's invest in SaaS, let's invest in margins. This is where the whole future was and like, you know, iPhone apps, blah, blah, blah. So I don't know, I guess I feel like, you know, Bane, remember Bane and Batman? What does he say? Like you adopted the darkness, I was born in it. Like, man, we built a robotics company for 10 years before building a robotics company was a cool thing to do. But, you know, I do think that especially important for like US competitiveness and just for our ability to like build the future that we be really proud to hand to our kids and to our grandkids and to build the sci version of the future that we were all promised. 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. We're depending on the crumbling infrastructure that our grandparents built for us. I read the other day, we just installed these like anti-suicide nets on the golden gate bridge if you guys read about that project it cost more to install those nets than it cost our grandparents to build that bridge i believe it so anyway we get really excited just like we think the future like promising future is like we should be able to build infrastructure you know we got it we have to be interested in it um and we i think people have to have the stomach for it and we have to learn how to manufacture and run complex supply chains again. And we have to be bold and believe in sci-fi versions of the future if we're going to build them. Awesome. Let's end it at that. Believe in the sci-fi future. Thank you guys for being with us. Thank you. Thank you for inviting us. [55:00] Thank you.