全球电动车产业里,找不到第二家赚钱的公司。
马斯克(Elon Musk)在特斯拉 2024 年第三季度财报电话会开场时给出的判断是:放眼全世界,没有一家纯电动车公司是盈利的,也没有任何一家传统车企的电动车业务部门是盈利的。特斯拉在这个季度做到了。
同一场电话会里还有另一个数字。第七百万辆特斯拉在前一天下线。
“昨天我们刚刚生产出第七百万辆车。恭喜让它成为现实的团队。造出七百万辆车,是难以想象的巨大工作量。”(马斯克)
财务负责人 Vaibhav Taneja 给出的数据更直观:本季度运营现金流 63 亿美元,创历史纪录;汽车业务营收同比、环比双增;能源业务毛利率超过 30%,同样是有史以来最好。
资本市场更关心的可能是另一句话。马斯克主动给出了 2025 年的销量预期,20% 到 30% 的增长。
给出 20% 到 30% 之前,他犹豫了一下
“我在这里冒一点风险。”这是马斯克的原话。他说除非出现战争、利率飙升这类不可抗力,否则明年 20% 到 30% 的车辆增长是他目前最好的估计。
行业背景可以对照着看。第三季度全球多数车企的交付量在下滑,特斯拉创了纪录。马斯克特意强调了这个反差,在他眼里,特斯拉跑得快只是表面,别人的电动车业务还在亏钱才是底色。
平均售价在下降,主要原因是金融激励。特斯拉通过第三方银行和金融机构提供低息贷款,把这部分成本记为收入抵减,所以账面 ASP 走低,实际成交价没怎么动。
Vaibhav 提到一个容易被忽略的观察:中国市场的电动车渗透表现,是欧洲和美国的三倍。
“如果要从中学到点什么,这就是其他地区未来的信号,客户对电动车的接受度会沿着同一条曲线走上来。”(Vaibhav Taneja)
他还补了一句关于需求的判断。绝大多数人的购车决策由月供决定,利率下降对汽车需求的拉动效果,比任何营销手段都大。美国市场现在出现的一个现象是,因为负担能力被利率压住,人们把车开得更久了,整个行业的换车周期在拉长。
两万五千美元的普通车,不做了
一位投资人问,什么时候能拿到那台 2.5 万美元的非 robotaxi 普通车型。马斯克的回答直接把这个方向关掉了。
“我们不做那台车。今天在售的所有特斯拉,未来都是自动驾驶车辆。我们很多年前就说过,未来只有自动驾驶电动车。非自动驾驶的燃油车,就像骑马和使用翻盖手机。”(马斯克)
他补了一句更狠的:这件事现在应该已经明显到刺眼,但大部分车企还没消化,这会成为它们未来的损失。
逻辑是这样的。一辆车能自己跑网约车赚钱之后,衡量它的标准就变成每英里的完全成本,方向盘和踏板成了多余重量。一台“既能手动开又能自动开”的混合设计,成本结构比专门为自动驾驶设计的车差得远。
Cybercab 的整车成本大约就在 2.5 万美元这个量级。想私人拥有一台,可以买。只是它没有方向盘、没有刹车、没有加速踏板。
2025 年上半年会有一款更便宜的车型按计划推出。这不是那台两万五的车,方向一致:把购买门槛降下来。
至于未来到底是无人驾驶出租车还是私人车,特斯拉的答案很干脆:今天在路上的七百万辆车,绝大多数已经具备自动驾驶能力。目前每周生产约 3.5 万辆自动驾驶能力的车,对比一下 Waymo 的整个车队还不到一千辆,每周新增 2.5 万辆。
Cybercab 的产线,不在同一个联赛里
马斯克在电话会上花了不少时间讲工厂,而不是产品。
“Cybercab 产线的节拍是真正特别的东西。比别的汽车产线好半个数量级。不在一个联赛里。”(马斯克)
他反复提到一个概念:造机器的机器。产线本身的设计,才是竞争力的核心。
“没有别的车厂在尝试我们正在做的事。我甚至没听说过有人试过。我确信一个都没有。”(马斯克)
一位特斯拉高管接着解释这套思路的实际含义。从一张白纸开始设计某个零件的时候,谁来做、在哪做、怎么运、怎么装到车上,任何一个环节放在孤立的地方,都会变成成本或时间的瓶颈。Robotaxi 的开发是把这些环节合起来重想,然后推翻原有方案。
“这会体现在我们的资本支出效率上,体现在零件数量上,也体现在整车的简洁程度上。”
FSD 现在遇到的瓶颈,是错误太难找
关于全自动驾驶(FSD)的进展,这一季有几个具体节点。
12.5 版本把城市道路和高速公路合并成单一神经网络栈,高速公路表现提升明显。之前高速用的是老一代网络。Cybertruck 的 FSD 发布。智能召唤(Actually Smart Summon)上线,使用次数已经超过一百万次。
Autopilot 总监 Ashok Elluswamy 给出了一组倍数:从今年年初到 12.5,关键干预间隔已经提升 100 倍;即将推送的 v13,相比 1 月会提升 1000 倍。
“v13 相比 12.5,我们预计干预间隔提升 5 到 6 倍。把整年算下来,至少 3 倍以上。这个趋势明年会继续。”(马斯克)
内部目标是 2025 年第二季度,FSD 的关键干预间隔超过人类驾驶员平均水平。马斯克特意说明这不是故意保守:“我们内部估计就是 Q2,没有藏东西。”
瓶颈的位置变了。以前卡训练算力,现在不卡了。得州奥斯汀工厂的 GPU 集群提前投用,本月底会部署 5 万块 GPU。真正的麻烦是模型好到很难找到错误。
“以前开几百英里就能碰上一次失误,现在可能要开一万英里才碰上一次。这时候想比较两个版本谁更好,得跑很久才能看出差别。这是高级烦恼。”(马斯克)
他把验证速度称为当前最大的单一限制因素,训练速度已经排到后面。
还有一层工作量的来源。车队每天回传的数据是 PB 级别的,从海量视频里挑出哪些片段对训练最有价值,这件事本身极难。数据筛选和算力扩容,现在并列成了两条主要的工作线。
Hardware 3 车主可能得换电脑
一位分析师问 Hardware 3 能不能达到 L5。马斯克没有给肯定答复。
“有可能会发现 Hardware 3 达不到无人监督 FSD 所需的安全水平。如果真是这样,我们会为购买了 Hardware 3 FSD 的用户免费升级。系统设计上是可以换计算单元的,摄像头没问题。”(马斯克)
做法上,特斯拉先用 AI4(Hardware 4)把方案跑通,再往回移植到 Hardware 3。Hardware 4 的能力是 Hardware 3 的好几倍,回移要重写内核,因为有些功能在 Hardware 3 上是模拟实现的,跑得慢。
这个回移过程会持续好几个季度。未来一段时间 FSD 的大版本进步,会先在 Hardware 4 上出现。
无人驾驶出租车,明年在得州和加州上路
得州明年肯定能跑。加州要看审批,马斯克的说法是“如果明年拿不到批准我会很吃惊”。年底之前可能扩展到其他州。
湾区特斯拉员工已经能用开发版 App 叫车,在湾区任意地点上下车,目前配安全员。
关于安全员什么时候撤,各州要求不同,涉及需要积累多少英里、多长时间。特斯拉会遵守,安全是优先级,但目标是“从驾驶座挪到副驾,再从副驾挪走”。
马斯克给了一个时间点:2025 年某个时候,会有真正的无人驾驶特斯拉提供付费乘车服务。
David 补充了一个细节,说明这件事不是临时起意。为了做网约车,很多基础件其实早就做出来了,只是先以别的形式给了用户:账号跨车同步(座椅、后视镜、导航、媒体全部一致,换哪台车都像自己的)、手机发送目的地到车、在手机 App 里看车辆在某条路线上的进度。
“我们建这些东西的时候,本来就是打算未来用在网约车上的。只是先释放给用户了。”(特斯拉高管)
每 700 万英里一次碰撞,美国平均是 70 万英里
第三季度车辆安全报告显示,开启 Autopilot 的特斯拉每 700 万英里发生一次碰撞。美国平均水平大约是每 70 万英里一次,十倍差距。
这个数据要带上限定条件看。Autopilot 主要用在高速和封闭道路,路况本身比全美平均简单。十倍这个量级仍然说明问题。
Optimus 的手有 22 个自由度
10 月 10 日的发布会上,特斯拉展示了新一代 Optimus 的前臂和手,22 个自由度,是上一代的两倍,触觉感知大幅提升。马斯克形容它“极其像人”。
“我很有信心地说,我们拥有遥遥领先的最先进人形机器人。而且我们是唯一一家真正具备规模化人形机器人所有要素的公司。”(马斯克)
他点出了别家缺什么:缺 AI 大脑,缺量产能力。有些公司能做出漂亮的演示视频,做不出规模。
还有一句他以前说过、这次重复的话:Optimus 有潜力成为有史以来最有价值的产品。
训练算力在 FSD 和 Optimus 之间共享。奥斯汀 GPU 集群扩得这么快,跟这两条产品线的需求叠加有直接关系。
Megapack 每周 200 台,上海工厂下季度投产
能源业务的数字比汽车业务更干脆。
Megapack 工厂达到每周 200 台,年化 40 GWh。上海第二座工厂 2025 年第一季度投产,起步年化 20 GWh,之后继续扩。
“用不了多久我们每年就会出货 100 GWh 的固定储能。最终会增长到每年几太瓦时。这不是想不想的问题,要有一个可持续能源的未来,就必须到这个量级。不到太瓦时,就谈不上真正推动能源转型。”(马斯克)
关于原材料,他的判断是地球上没有任何一种稀有材料会卡住这件事。全人类文明转成可持续能源,甚至用电量大幅增长,原材料都够。
部署量按季度会有波动,取决于客户准备情况和项目所在地,不代表需求。第三季度部署量下降,第四季度环比会增长,全年比去年翻一倍以上。2025 年的产能已经被订满,团队在尽力跟上。
电动车本身只用掉了地球能源潜力的一小部分。马斯克提到,就算把整个地球表面铺满太阳能板,也只是太阳总输出的十亿分之一。这句话放在储能业务的语境里,是在解释特斯拉为什么把长期重心押在能源上。
用卡尔达舍夫等级看能源
电话会上有一段听起来很飘的话,它是特斯拉能源业务规模的底层逻辑。
“衡量一个文明进步的一个方式,是看它在卡尔达舍夫等级上的完成度。I 型是用尽一颗行星的全部能量,我们现在连 I 型的 1% 都不到。II 型是用尽一颗恒星的能量,III 型是整个星系。我们还有很长的路要走。”(马斯克)
“用卡尔达舍夫尺度来想,最大的能源来源显然是太阳,其他所有东西都在噪声里。”(马斯克)
太阳每秒送到地球的能量,比人类文明现在消耗的总能量高好几个数量级。特斯拉把储能当成长期主赛道,而不是汽车业务的附属品,源头就在这套账上。
Semi 卡车:司机开过就不想换回去
内华达超级工厂的 Semi 产线在快速推进,主要资本开支已经释放。2025 年下半年开始试产,2026 年上半年正式量产,全年爬坡。
“Semi 的需求多到有点荒谬。”(马斯克)
商业逻辑简单得很:每英里每吨的完全成本比柴油卡车低,不换电动重卡的车队会输。这不涉及审美偏好,纯粹是算术。马斯克的原话是,一辆车每英里每吨的运输成本如果能打赢柴油卡车,不用电动重卡的公司就会失去竞争力,这事儿没有主观空间。
百事(Pepsi)那边的反馈是,司机开过电动 Semi 就不想换回去,资深司机抢着开。
马斯克还提到几个附加好处:下坡用动能回收,刹车不会过热,能量回到电池包;防折叠(anti jackknifing)软件;响应快,不会堵在慢车后面,对其他道路使用者也是安全收益。
所有 Semi,从已经交付的几百台到未来生产的车,都带 FSD 所需的全部硬件和摄像头。目前正在用现有车队训练,神经网络就绪之后就推 FSD 上去。长途司机的疲劳问题,是这套系统最大的价值点。
最好的服务,是让车不出问题
一位投资人问服务网点排队时间太长怎么办。回答的第一句是“最好的服务是没有服务”,第二句是“车不坏”。
思路是从源头解决:在工厂里解决,在软件里远程解决。很多问题车主不用到店,车停在家里或者公司,后台推送修好。
北美今年第三、第四季度一共新开近 70 个服务点,单店面积比去年翻倍。大店的好处是可以做分工,专门的车道处理特定类型的维修,重活不堵所有工位。
“传统经销商模式里,经销商大部分利润来自售后,他们显然没有动力降低维修成本。我们的情况相反,维修成本由我们自己承担,所以有动力把它降下来。这是特斯拉相对整个汽车行业的结构性优势。”(马斯克)
他还提到车会“说话”:进店之前车辆已经完成自诊断,零件、举升机、技师全部提前排好。大多数情况下,车到店的时候不需要再诊断。
这套做法还有一层组织上的设计。工厂的人和售后的人互相对调,把现场看到的问题带回产线,把产线上知道的答案带到现场。问题在源头被修掉一次,全国就少修一万次。
一万个零件,每个省五毛钱
被问到 2025 年的规划,马斯克顺势讲了降本的难度,这是他整场电话会里语气最重的一段。
“把一辆车的成本砍掉 20%,比设计这辆车、建整座工厂还要难。这活儿极其折磨人。没有多少电影拍过那些把成本砍掉 20% 的英雄,但说实话,应该有。如果你真看到这件事有多难,你会说,天,真难。”(马斯克)
他给了一个算法。一辆车大概有一万个零件,如果平均每个四美元,就是四万美元。想做到三万五,就得平均每个零件省五十美分。这活儿没有捷径,只能一个零件一个零件抠。
更好的办法是删零件。“我们发现有大量零件可以删掉。”
第三季度单车成本创了历史最低。运营费用环比下降,部分来自此前重组带来的节省。
资本支出全年超过 110 亿美元,第三季度单季超过 35 亿,主要投向 AI 算力。Vaibhav 强调公司在 AI 投入上很克制,先把现有基础设施用尽,再决定新采购。
FSD 及相关功能(包括 Cybertruck 的 FSD 和智能召唤)第三季度贡献 3.26 亿美元收入。监管信贷收入今年累计超过 20 亿美元,仍然维持在高位。汽车业务毛利率环比改善,来自 FSD 功能释放、产销规模提升、大宗商品价格走低,以及更本地化的交付带来的运费和关税节省。第四季度要维持这个水平,Vaibhav 说会很有挑战。
Roadster 是蛋糕上的樱桃
“感谢那些长期忍耐的特斯拉 Roadster 预订用户。”马斯克开场先道歉。
推迟的原因是排序。Roadster 是蛋糕上的糖霜上的那颗樱桃,特斯拉更大的使命是加速可持续能源转型,资源必须先给影响更大的事情。
“设计快要定稿了,会非常惊艳。”
他还提到一件事。朋友彼得·蒂尔(Peter Thiel)抱怨未来没有飞行汽车。
“我们走着瞧。”(马斯克)
xAI 帮了特斯拉什么忙
摩根士丹利的 Adam Jonas 问特斯拉和 xAI 的关系。有投资人担心两家公司在人才和技术上互相竞争,还分走马斯克的时间。
马斯克的回答分两层。
第一层是 xAI 确实帮过特斯拉。训练集群的规模扩展,以及最近一周在训练中断恢复上的改进。大规模训练跑到一半挂掉,能快速恢复继续跑,这件事很值钱。
“xAI 在做通用人工智能,或者说超级智能。特斯拉在做自动驾驶汽车和自主机器人。这是不同的问题。”(马斯克)
第二层是动机。他觉得市面上缺一家真正以追求真相为目标的 AI 公司。
“我创办 xAI 的原因,说到底就是当时没有一家追求真相的数字超级智能公司。哪怕某个结论在政治上不正确,它也应该是真实的。这对 AI 安全很重要。”(马斯克)
技术上他给了一个区分。语言模型的上下文通常是几万到几十万 token,特斯拉的自动驾驶要处理七八个摄像头,最高到九个(含车内摄像头),几百 GB 级别的上下文,然后压缩成少量控制输出,还要在车上几百瓦的算力里跑完。
“我们被迫要在推理效率上做到极致。车里只有几百瓦,特斯拉 AI 芯片设计得相当好,但仍然是个很小的算力。xAI 的模型跑在 10 千瓦的机架上,那个塞不进车里。”(马斯克)
训练算力可以弥补推理算力的不足。人类学开车也是这样,第一次上路占满全部注意力,开够多年之后驾驶变成后台任务,只占一小部分脑力。
两家公司也各有各的稀缺资产。马斯克提了一句,世界上同时拥有世界级整车设计团队和世界级 AI 团队的公司,一只手数得过来。这两样能力都是从头搭起来的。
Grok 进车只是小菜
被问到什么时候把 X 和 Grok 整合进车里,马斯克的评价是“这些是相对小的事情”。
更重要的判断在后面。一旦实现完全自动驾驶,车里的人需要一套能做任何事的系统。上网、问 AI、看电影、打游戏、办公,全都可以,因为不需要开车。
“这就是 Cybercab 有一块大屏幕和好音响的原因。像私人影院。”(马斯克)
他顺手推荐了几款车上的游戏,Castle Doombad、Polytopia。特斯拉会持续看用户最常要求什么功能。
自动驾驶审批,该由联邦管
被问到加州和得州的审批进度,马斯克把话题拉到了联邦层面。
美国对车辆的联邦认证走 FMVSS 法规。特斯拉现在生产的、具备自动驾驶能力的车全部符合,Cybercab 也会符合。车辆能否上路不构成限制,限制在州一级的自动驾驶部署许可。
得州相对容易。加州有路径但要花时间填表、走审批。还有些州根本没设立流程。
“我认为应该有联邦审批流程。如果国会在听,就该搞一个联邦自动驾驶审批。就像 FMVSS 当初那样。2017、2018 年监管机构开始看这件事,之后基本停在那里。全国统一审批很重要。”(马斯克)
他拿保险举例,50 个州各走一遍极其痛苦。顺便开了个玩笑,说如果真有个“政府效率部”,他愿意帮忙推这件事,“为所有人,不只是特斯拉”。
写在最后的行动清单
这场电话会的信息密度很高,真正能落地跟踪的线索其实不多。整理成几条,方便后续验证。
- 盯住 2025 年上半年的“更便宜车型”。特斯拉已经明确不做 2.5 万美元的普通车,这款车大概率是通往自动驾驶产品线的中间台阶,看它的定价和是否有方向盘。
- Hardware 3 车主留意官方关于无人监督 FSD 能力的公告。如果硬件确实不够,特斯拉承诺免费升级计算单元,但时间点取决于 AI4 方案什么时候完全跑通。
- 判断 FSD 进展,别看版本号和演示视频,看“每次关键干预之间的英里数”。内部目标是 2025 年第二季度超过人类平均水平,这个承诺可以被证伪。
- 想体验 Robotaxi,先看监管再看技术。得州确定性高,加州有路径但需要时间,其他州要看是否建立了流程。
- 跟踪储能产业链,上海 Megapack 工厂 2025 年第一季度投产,年化 20 GWh 起步,这是能算出来的产能,不是愿景。
- Semi 的客户名单和 2026 年量产节奏,是观察重卡电动化拐点的窗口。百事这类早期客户的复购意愿是最硬的指标。
- 评估任何自动驾驶公司,把“每英里完全成本”当成核心指标,别被车队规模和融资额带偏。马斯克整场电话会反复回到这个词,值得借用。
- 4680 电池的成本竞争力,明年会有一个可验证的节点:特斯拉说它接近成为北美完全落地成本最低的电芯。看这句话什么时候从“接近”变成确定语气。
- 如果你在关注 AI 算力采购节奏,奥斯汀的 GPU 规模是个公开的追踪点,本月底 5 万块,后续每季度的增量能反推出训练投入的真实意图。
0:00and I'm joined today by muskot TNA and a number of other Executives our Q3 results were announced at about 3M Central Time in the update deck we published at the same link as this webcast during this call we will discuss our business Outlook and make forward-looking statements these comments are based on our predictions and expectations as of today actual events or results could differ materially due to a number of risks and uncertainties including those mentioned in our most recent filings with the CC during the question and answer portion of sp's call please limit yourself to one question and one followup please use the raise hand button to join the
0:39question Q before we jump into Q&A Elon has some opening remarks Elon thank you so to recap as someone seeing something that of Industry I seeing year over your declines in order of volumes in Q3 Tesla at the same time has achieved record deliveries in fact I think if you look at e companies worldwide to the best of my knowledge no EV company is even profitable and and I'm not to the best of my knowledge there's no EV division of any company of any existing order company that is profitable so it is notable that Tesla is profitable despite very challenging for motive
1:23environment and and and this quarter actually is a Rec Q3 for us so we produced our 7th million vehicle actually just yesterday so congratulations to the teams that made it happen at Tesla that's staggeringly immense amount of work to make 7 million cars so let see and we also have energy
1:50storage business is growing like wildfire with strong demand for both mega pack and power wall and as people know on October 10th we laid out our vision for an autonomous analytical future that I think is very compelling you know the Tes team did a phenomenal job there with actually giv people an opportunity to experience the future where where you have humanoid robots walking among the crowd not not you know can but presentation or anything but literally weling among the crowd serving drinks and whatnot and and we had 50 autonomous vehicles there were 20 cyber caps but there were an additional 30 model wise operating fully autonomously the entire night carrying thousands
2:40thousands of people thousands of people doing with with no incidents the entire night so and and for those who went there I it's worth emphasizing that the the Cyber cap had no steering wheel or brake or accelerator pedals meaning there was no there's no POS there was no way for anyone to intervene manually even if they wanted to and the whole night went very smoothly so regarding the vehicle business we are still on track to deliver more affordable U models starting in the first half of 2025 you know this is I think
3:23probably people are wondering what should they assume for vehicle vehicle sales growth next year I and at at the risk of to take a bit of risk here I I I do want to give some some rough estimate which is I think 20 to 30% vehicle growth next year you know not withstanding negative external events like if there's some Force measure events like some big war breaks out or interest rates go Skyhigh or something like that then you know we we can't overcome massive Force mature events but I I think with our vehicles with the Advent of autonomy something like a 20 to 30% growth next year is is my best case
4:09and and then and then c cap reaching volume production in 26 we do feel confident of Cy cap reaching volume production in 26 not just stting production but reaching volume production at 26 and that you know that that should be substantial both we're aiming full you know at least 2 million units a year of of cyoc Camp that that'll be in more than one Factory but I think it's at least 2 million unit year maybe 4 million ultimately so yeah these are just
4:47my best guesses but if you ask me what my best guesses that those are my best guesses the the 46 the the cell 4680 lines the team is actually doing great work there the 4680 is rapidly approaching the point where it is the most competitive cell so when you consider the the fully the fully landed the cost of a of a battery Bank fully
5:16landed in the US net of incentives and duties the the 4680 is tracking to be the most competitive me being lower cost water are fully considered that any other alternative which is not quite there yet but we're we're close to being there which I think is extremely exciting and we've got several a lot of ideas to go well beyond that so if I I think there's if we execute well the 4680 we'll have the the Tesla internally produced cell will be the most cost competitive cell in in cly North
6:03America I tesed to a tremendous amount of hard work there from by the team so that that's say we we'll continue to buy a lot of Sals from our competitors so our tend it's not to make to to provide to to make Sals just interally so I don't want to set off any long bells here we're obviously increasing substantially our vehicle output and our stationary storage output so we need a lot of cells and most of them will still come from suppliers but but I think it is it is a
6:40some good news that the Tesla internal cell is likely this track need to be the most competitive in the US so with respect to autonomy as people have are experiencing in the cars really from week to week there are significant improvements in the miles between interventions so with the the new version 12.5 the release of full celf driving cyber truck the the combining the code
7:17into a single stack so that the city driving and the anti and highway driving are are one stack which is a big Improvement for the highway driving so it's it's just old neural Nets and the releasee of Ashley SM summon we try to have a sense of humor here at Tes U and we're also so that that's 12.5
7:43version 13 of FSD is going out soon sure we'll elaborate more on that later in the call we we expect to see some roughly 5 or six fold Improvement in miles between interventions compared to 12.5 and actually looking at the the year as a whole of the Improvement in miles between interventions we think will BR at least three or mag to so
8:19that's that's a very dramatic Improvement in the course of the year and and we expect that Trend to continue next year so [Music] the current the current internal expectation EMP internal expectation for the Tesla FSD having longer miles between intervention than human is the second quarter of next
8:51year which means it may end up being the third quarter but it's it's next it seems extremely likely to be next year sure you want to anything yeah mentioning miles between critical interventions yeah like you mentioned neon we already made a 100x improvement with 12.5 from starting of this year and then with v13 release we expect to be 1,000x from the beginning from January of this year on my production release software and this came in because technology improvements going to end to end having higher frame rate partly also helped by Hardware force more capabilities
9:33so on and we hope that you know we continue to scale the neural network the data the training compute Etc by Q2 next year we should cross over the average human mins for critical intervention or like Collision in that case I that that is just unboned our internal estimate yes yeah so that that's not sandbagging or anything our interal estimate is Q2 next year to be safe than human and then to continue with rapid improvements
10:12thereafter to a lot vast majority of humanity has no idea that themselves so especially if for something like like a model 3 or model one it looks like a normal car so you don't expect a normal car to be able to be intelligent enough to drive itself Cy cab looks looks different cyber truck looks different but model Y and model 3 are look they're goodlooking cars but look they look fairly normal you don't expect a fairly normal looking car to have the intelligence you enough AI to be able to drive itself itself but it does so we do want to expose that to more people and and so we're doing every time we have as significant Improvement in the software we'll roll we'll roll out
11:12another sort of 30-day trial so to encourage people to try it again and and we're all seeing a significant Improvement in adoption so the the the T rate for FSD is has improved substantially especially after the 10 event yeah so so there there's no need to wait for robot taxi or savab for to experience full autonomy we expect to achieve that next year with the with our existing vehicle Lem I what actually smart summon gives a small taste of what it's going to look like the car able to drive itself to the user within private parking lots currently it's speed Limited but then it's going to quickly be increased and we already had more than a million usage
12:06itms of smart yep so and we actually we have for TLA employees in the Bay Area we already are are offering a right healing capability so so you can actually you with the development app you can request a ride and it'll take you anywhere in the Bay Area we we do have a safety driver for now but the software required to do that we've we've developed I mean David do you want to elaborate on that yeah sure David we showed some screenshots of this in the q1 shareholder deck and you know this is real we've been testing it for the part of the year and the building blocks that we needed in order to build this functionality and deliver it to production we've been thinking about
12:55working on for years it just so happens that we've used those building blocks to deliver great features for our customers in the meantime such as sharing your profile synchronizing it across cars so that every single car that you jump into whether it's you know another car that you own or a car that somebody's loan to you or a rental car that you jump into it looks exactly like yours everything synchronized seat mirror positions you know media navigation everything is the same just what you would expect from one of our Robo taxis but you know we gave that functionality to our customers right now because we built it intending for it to be used in the future
13:32releasing that functionality now all the and then cyber security that we knew we're going to need to deliver that functionality sending a navigation destination from your phone to the vehicle and so you know you're doing that now with the with the ride hailing app but it's something that we've made available to customers for years seen the progress on a route in the mobile app that's something you'll need for the ride healing app but again we released it so it's not like we're just starting to think about this stuff right now while we're building out you know the early stages of our ride healing Network we've been thinking about this for quite a long time and we're excited
14:07to get the function out out there yeah and and we do expect to roll out ride hailing in California Texas next year to the public not the California is somewh there's quite a long regulatory approval process I it shouldn't I think we should get approval next year but but it's continued upon regulatory approval ACC is a lot faster so it's you
14:33I'd say like we'll definitely have available in Texas and probably have it available in in California subject to regulatory approv and then and and maybe some other states actually next year as well but at least California Texas I that'll be very exciting there really a profound change Tesla become comes more than a sort of vehicle and youy Manufacturing
15:05Company at that point so we published a Q3 Vehicle Safety report which shows one crash for every 7 milon miles of a pilot that compares the US average of one crash roughly every 700,000 miles so it's currently showing a 10x safety Improvement relative to the US average and we continue to expand our AI training capacity to combinate the needs of both FSD and Optimus we're currently not not a training compute
15:42constraint it's probably big so many factors that the the empy is actually getting so good that it takes us a while to actually find mistakes when you start getting to where it could take 10,000 miles to find a it's a mistake it's a it takes a while to actually figure out which it is is this is a Sofer B better than is Sofer Bol a better than Sofer b b it actually takes a while to figure it out because neither one of them is making mistakes what takes take a long time to make mistakes so that's actually the single limit big limiting factor is how long does it take us to figure out which which which version is better sort of high class problem
16:30obviously having a dronic fleet is very helpful for B buring the sound and then with Optimus we show we show a massive Improvement in optimus's dexterity movement on October 10th and our next gen handed forarm which has 22 degrees of freedom double which is double the prior and borong
16:56it's it's extremely humanlike says much better tactile sensing it's really I feel confident saying that we have most advanced humanoid robot by Longshot and we're moreover the only company that really has all of the ingredients necessary to scale humanoid robots because the the the things that what other companies missing is they they're they're missing the AI brain and they're missing the ability to really scale to very high volume production so you see some impressive video demos but what but they they like the they like the localized Ai and the Rel to scale volume to very high numbers as I've said on a few occasions before I think Optimus will ultimately
17:54be most I think has a good chance of being the most valuable product ever made for the energy business that's doing extremely well and this the opurity IE is gigantic the L MEAP pack Factory reached 200 megapacks a week which is now 40 G gwatt hour a year run rate and we have a second Factory in Shanghai that will begin with the 20 GW hour a year run rate u in q1 next year so just next quarter and U that that'll also scale up
18:39it won't be long before we're shipping 100 gaw hours a year stationary storage and and it'll that that I mean I that'll ultimately grow I think to to multiple terawatt hours per year it it has to actually in order to have a sustainable energy future if you're not at the ter scale you're you're not really moving the heel so if if you look at our Mally very complicated last master plan which I think actually has too much detail maybe I'll ask gr to analyze it and shorten it up yeah give us the TDR on the last master plan but we we showed in that master plan that it is possible to take all of Earth to a fully sustainable energy situation using sustainable energy power
19:33generation and batteries and electric transport and there are no fundamental material limitations like there's not some very rare material that that we don't have enough of on Earth we actually have enough of the raw materials to yeah take all of human civilization make it fully sustainable and even civilization dramatically increased its electricity usage would still be fully sustainable you know one way to think of the progress of a civilization bace out a little esoteric but is percentage completion of CF
20:16scale so CF scale one would be you're using all the power of of a planet we were we're currently less than 1% on Carter level level one level two would be using all the power of the Sun and level three all the power of the Galaxy so we got a long way to go long way to go when you think in cter ship terms it becomes obvious that by that the by far the biggest source of of energy is the sun everything else is in the noise so
21:01so in conclusion Tesla is focused on building the future of energy transport Robotics and Ai and this is a time when others are just focused on managing around near-term Trends we think what we're doing is the right approach and if we execute on our objectives and I think we will tell that my prediction is T will become the most valuable company in the world probably by a long by a long shot I want to thank the T of Team once again for strong execution in a tough operating environment and we're looking forward to building an incredibly exciting future thank you great thank you very much Elon and that has remarks as well yeah thanks our Q3 results were overall positive and
21:50once again demonstrate the skill in which businesses evolved over the years with the generation of record operating cash flows of 6 3 billion our Automotive revenues grew both quarter on quarter and year on year while we had unit volum growth we did experience reduction in asps primarily due to the impact of financing incentives as a reminder we're providing these incentives primarily using third party Banks and financial institution and recognize the cost of these incentives as an enough reduction to them we released FSD for cyber truck and other features like actually smart someone like talked about in North America which contributed 326 million of
22:34revenues in the quar we continue to see elevated levels of regulatory credit sales with over two billion of revenues so far this year expand on this at an industry level China continues to outperform us and Europe by a factor of three and if there is something to be learned from that this gives a signal of what is to come in other regions as customers acceptance of e grows and we feel that is the right strategy to build affordable and more competive UNS the focus remains on growing unit volume while avoiding a buildup of inventory to support this strategy we're continuing to offer extremely compelling vehicle financing options in every L when you compare any vehicle in our
23:24lineup with other oems believe our vehicles provide much better value particularly when you consider the safety features performance and unparalleled software functionalities like David also talked about include also what had talked about around autonomy music options parental controls and much more while every vehicle in our lineup comes up with these capabilities there is an awareness Gap not just with Buyers but at times even with existing owners we plan on on making these more visible in our interactions with both existing and future customers Automotive margins improved quarter over quarter as a result of FD featur release discussed before increase
24:12in our overall production and delivery volume contined benefit from commodity pricing and more localized deliveries in region which resulted in lower Freight and duties sustaining these margins in Q4 however will be challenging given the current economic environment not that we are focused on the cost per vehicle and there are numerous work streams within the company to exree that cost without compromising on customer experience yeah something that that's a helpful hopefully helpful macro trend is if there's a decline in in in interest rates this has a massive effect on the
24:53on Automotive demand the V vast majority of people is the demand is driven by the monthly payment can they put monthly payment so think most likely we see continue decline interest rates which helps with affordability Vehicles yeah I mean that is one Trend which we observed in the industry that you know because of the affability being impacted because of interest rates people are holding on to the cars longer especially in the US that is actually having an impact on overall industry too
25:30as we discussed earlier as we discussed in P orders energy deployments fluctuate quarter quarter due to customer Readiness location of orders being fulfilled and not necessarily an indicator of demand of production within the quarter while we did see a decline in Q3 we expect to grow deployment sequentially in Q4 to end the year but more than double of last year energy margins in Q3 for record at more than 30% this is a function of mix of projects being deployed in the quarter not that there will be fluctuation in margins as we manage through deployments and ring our Pipeline and backlog continue to grow quarter over quarter as we fill our 2025 production slots and
26:22we're doing our level best to keep up with the demand just coming back on Automotive margins I talked about sorry I talked about what is happening one other thing which I want to also share is that we're that you know we will continue to keep whatever we can to squee like I said before about squeezing out the cost but this is something which we also are very capable of I mean just in Q3 we reach our lowest cost for vle and that is a which we want to keep focus on then going on to service and other we continue to show improvements in Q3 this was a result of better performance both in our service business which includes Collision part sales and merchandise and
27:13continued growth is supercharging these Fleet based revenues will continue to grow as the overall Fleet size increases our operating expenses declined quarter over quarter and a year- on-year basis this is partially due to the restructuring we undertook in KTU cost saving from these initiatives were partially offset by increase in cost related to our efforts we've started using the GPU cluster based out of our Factory in Austin ahead of schedule and now on track to get 50k gpus deployed in Texas by the end of this month one thing which I'd like to elaborate is that we're being very judicious on our AI compute span too and saying how best we can utilize the
27:56existing infrastructure before making qu investment on the capex front we had over 3 and a half billion in the quarter this was a sequential increase largely because of investments in AI compute we now expect the capex for the year to be an excess of 11 billion we shared our vision for the future at the V robot event at the beginning of the month the Tesla team is hyper focus and delivering on that Vision all efforts are underway to make it a reality what we've achieved significant progress this year it will take time to get this as we Pioneer new and Incredibly complex Technologies and navigate a fragmented regulatory landscape future is incredibly bright
28:42and I want to thank the Tesla team once again for all great thank you very much feob now we'll go to investor questions the first one is is Tesla still on track to deliver the more affordable model next year as mentioned by Elon earlier and how does it align with your AI and product redb sure I mean as Elon and VI both said we are in plan to meet that in the first half of next year our mission has always been to lower the cost of of the of of our vehicles to increase the adoption of sustainable energy and transport part of that is is lowering the cost of our current Vehicles which is where all of the personally owned vehicles that we sell today come in but
29:22the next stage in that really is it fits into AI road map is when we bring in robot taxis which lowers the initial cost of of getting into an EV and those that's really where we see the marriage of EV road map and the AI road map it'll be like with with with incent of Sub sub 30k which is kind of a key threshold great thank you very much similar question next when can we expect Tesla to give us the $25,000 nonroot taxi regular car model we're not making it on Rob yeah all our vehicles today are robot I think we've made very clear that we're the future is autonomous I mean it's going to be you and I actually said this many years ago U but that my strong belief
30:16and I believe that is panning out to be true be very obvious in retrospect is that the future is autonomous electric vehicles and non ious gasoline vehicles in the future will be like riding a horse and using a flip bone it's not that there are no horses yeah there are some horses but they're unusual they're Niche and you know
30:43so so everything's going to be electric autonomous I think this this is blind like it should be frankly blindingly obvious at this point that is the future so a lot of Automotive companies most Automotive companies have not not internalized this which is surprising because we've been shouting this from the rooftops for such a long time and it will accre to their detriment in the future all of our vehicles in the future we yes all vehicles that we we really
31:21made all the 7 million Vehicles vast majority are capable of autonomy and you know we're currently making on the order of 35,000 autonomous vehicles a week you compare that to say way more entire fleet it's less than they have less than a thousand cars we make 25k a week yeah and our cars look normal yeah they mostly look normal the Cyber truck looks yes thankfully you know looks abnormal and then the you know cyber cab SL R Taxi you know we wanted to have something futuristic looking I think it does look futuristic it's worth noting with respect to the Cyber cab it's not it's especially not just a revolution in vehicle design but a revolution in vehicle manufacturing
32:13that is also coming with the with the Cyber cab the the cycle time like the the units per per hour of of the Cyber cab line is like this is just really something special I mean this is be yeah half order of Mag better than other car manufacturing lines like like like like like not not in the same league is what I'm saying not in the same league so so it's it's you know
32:51and I I said like you know several years ago that the maybe the most the hotest do look to Copy Factory you just like buy a factory yeah you can't reverse Factory have yeah it's like you know yeah so the and and as we so we're rapidly evolving in manufacturing technology so so anyway there's like basically I think having a regular 25k model is pointless
33:28it would be silly like it would be completely at o with what we believe in in in an autonomous world what matters is the lowest cost per mile of of efficiency of that vehicle and that's what we've done with with the Rope taxi exactly autonomous it's fully considered cost per mile is what matters and if you try to make a car that is you know essentially a hybrid manual automatic cards it's it's not going to be as good as a dedicated autonomous car
34:00so yeah Cy C is is just not going to have steering wheels and pedals it's design optimiz for autonomy but it'll it'll you know cost on the order of cost roughly 25k so it is a 25k car and you can you will be able to buy one Own It exclus exclusively if you want so just what have steer you don't need it great thank you very much the next question is what is Tesla doing to alleviate long weight times at service centers so we aim on solving problems at the source so at the factory before they can even fact our customers we believe the best service is no service and yeah really is don't even have them the if the car doesn't break yeah exact that's
34:52the best thing don't see anyone with the Tes Shir you either do it REM fix the issue upstream or do it remotely do through software they can be at work or at home and you know car can be part and we address fix the issue and we've partnered the field with service to make sure we're looking at the same issues and additionally just in Q3 and Q4 this year alone we we have open and will open in total of nearly 70 locations and in North America we significantly expand the size of each location and have doubled the size last year compared to this year yeah I think I think it's like actually a lot of erit to having large service cers because you can you
35:30can have specialization of Labor you can start you can start to approach yeah it should be more Factory like you know where you can have dedicated lanes for particular types of service and and it's way easier for somebody to come export in a few different types of repairs than in every repair exactly this this helped us with the base that these heavy repairs like clogging up a lane they dedicated lanes for different type of repairs and so it's through matters and really treating it like a factory yeah this is this is where Tesla structure I think Tesla has a structural Advantage relative to the rest of the Auto industry because we make cars and we service the cars whereas I think
36:10there's a bit of a conflict of interest with the the dealer model and the O the de traditional OEM a dealer model where the dealerships make most of their money on service and so they don't they obviously this incentive to reduce the servicing cost whereas in our case we are incented to reduce the servicing cost because we we carry that servicing cost and we've got a good feedback with with our with our cars the master EXA yeah be with the factory with the service leaders together and send fact people from the factories to the field and field to the factory to see it firsthand ex provide suggestions for you know manufacturing as well as for engineering on design
36:51yeah so I view this as a structural fundamental structural advantage of Tesla versus the rest of the order also doing a bunch of work on the software side not only to automate you know Diagnostics so identifying you what needs to be done to a car before it comes into service but also automating all of the preparation work and aligning all the resources that are necessary in order for the car to be very efficiently worked on once it arrives so the parts are there like the lift is scheduled the technician schedule like everything this is what's wrong with me and tells until the service center the car everything ready in advance yeah please fix me and this is what's
37:32wrong instead of a customer trying to translate the car is telling us directly and we're pulling that yeah yeah you don't most the time you don't need diagnose the car when it arrives the car yeah this is like again a fundamental technology advantage and structural Advantage compared to the rest of the order industry it's I think it's underappreciated as to what all we are able to do and and that's why because like you said for most of our cars except for cyber Dru look the same yeah right so people don't realize that it has so much capability yeah yeah they they look better than other cars but but they they're not like obviously like super futuristic
38:15yeah great thank you very much the next question is please provide an update on the semi what will the next stage of growth look like and when will FSD be ready sure so as you we posted in the earnings progressing swiftly on the build of the semi Factory in in our gigafactory in Reno we've released all our major Capo expenditures for that program and we're on track to start pilot builds in the second half of next year with production starting in the first half of 2026 and ramping really throughout the year to to full production sem my you know growth will largely depend on our customer adoption of the product well I don't think we're going to be divine limited
38:52honestly yeah which say which is like aner for for the semi because it's really a commodity of total cost of ownership yes exactly it's it's good we have kind of ridiculous demand for the semi it in that will where it's about how much do I spend to move per mile it's a no-brainer yeah fundamentally if you've got a sui where the fully considered cost per mile per per ton of Transport is better than say a diesel truck any company that doesn't adopt an electric semi will will lose it's not a it's it's subjective thing is like whether do you like comp I mean we like we we want the style we want we want to have a beautiful semi- TR but frankly if
39:34we made an ugly semi- Tru wouldn't matter and this is proving so in our fleets and in Pepsi a partner in fact the Pepsi actually said last week they're having nobody want their drivers don't want to go back yeah yeah as soon as we give anyone the electric semi it's it's it's like the that that's like the choice it's the what they want to drive yeah yeah yeah that's like like so the like the most senior like they top drivers will they get to drive the T the sway it's it's it's it's the thing they want to drive it's it's it's super fun to drive it's also very easy to drive it's it's easy to drive and it RS ass it's like Fast maybe too fast well but I mean like
40:17like you know like like you've seen like the videos of where like think like T Electric someone like you know can go uphill yeah speeding fast like the diesel truck or even cars yeah even cars so like it's responsive it's you you you know you know you floor it and that the truck actually then that's a benefit not only for for the driver and for the goods but also for safety in terms of other drivers on the road you don't get stuck behind the semi you're not like yeah you know in a slow down situation and onr I mean how that plays into you know FSD which is the second part of the question all of the semis have been since the couple hundred we've deployed
40:53already and and the ones that we'll be building next year and throughout the future have all the hardware and the cameras necessary to to deploy FSD and we're currently training with that small F that we have and as soon as the fleet is trained and the neural ns are up you know we we'll get FSD onto that platform yeah I mean it' be a massive Improvement in driver fatigue you know because you at driver safety we've got sort of the anti- Jack knifing software you know you don't have to worry about your brakes overheating if you go down a down a steep hill because because the we use regener like that that energy goes back into the pack actually when we leave R it's just
41:33like it's it's like radically better than than a diesel semi it's what what the drivers love it great guys thank you very much our next question is when will Tesla incorporate X and grock in all the Tesla vehicles well I mean the these are relatively small fry things you know but yeah the I think we'll keep expanding you know what what is available in the car on on the screen and also improving like the browser so like just generally you can access anything you want in the car in fact for the Tesla you know once you get to full autonomy you actually want F A A system that is can do
42:19anything like if you want to browse the internet if you want to you know ask ask AI questions if you want to watch a movie if you want to play a video game if you want to do some productivity thing you can do anything you want in an autonomous vehicle because you don't need to drive and so that's why the Cy cat got a nice big screen and a great sound system so you can watch a watch a great movie it's like being like personal movie theater yeah personal movie theater aome yeah this is why we've been building this functionality adding gaming to the car add you s of different media applications to the car because you know that's what you're yeah the cars that we buil today
43:05really fun games by the way people haven't tried it there's like castle doombad and politopia and a bunch of really fun games in the car y we're we're constantly looking at you know what features to add an ex thing we're paying attention to what's most commonly requested by customers yeah play castle doombad you wonder to Great thank you guys very much the next question is Elon mentioned unsupervised FSD in California and Texas next year does that mean Regulators have agreed to it in the entire state for existing Hardware three and four vehicles no as we as I said earlier you California loves regulation but they have a pathway yeah I mean there's a
43:46pathway obviously weo operates in California so there's just a lot of forms spill out and a lot of approvals that are required I I I mean I'd be shocked if we don't get approval next year but it's it's just not something we we totally control but I think we will get approval next year in California and Texas and and and towards the end of the year I will Branch out Beyond California Texas I mean I think it's important to reiterate this like humig your certifying a vehicle at the federal level in the US is done by meeting fmvss regulations all our vehicles today that are produced that are autonomist capable meet all those regulations the Cyber cat will meet
44:20those regulations and so the deployment of the vehicle to the road is not a limitation but is a limitation is what you said at the state level where they control autonomous FAL deployment some states are relatively easy as you mentioned for Texas yeah and other ones have Pathways like California that may take a little longer other ones haven't set up anything yet and so we will work through those state by state and get I I do think we should have a federal I agree like autonomous vehicles should be approved there should be it should be possible to Congress if you're listening let's get a federal AB there there should be a Fed Federal approval process
44:50for autonomous vehicles I mean that's that's how the fbss is started work Federal Motor Vehicle they def VSS is federal yeah so I mean in 2017 and 18 they we you know it's when Regulators started looking at it it's really kind of stalled since then but we would appreciate and would support helping out with those it really needs to be like at National approval is is important you know if there's a department of government efficiency I'll try to help make that Happ I did it for everyone not just Tesla but you know just we like
45:22like some some things in the US are state-by-state regulated like for example Insurance like a it's incredibly painful to do it state by state for 50 states and and I think we should have there should be a national approval process foron great thanks guys the next question is what is the plan for 2025 I me think we're just talk yeah we I mean basically we talked through this there's a lot going on Elon already mentioned that we're working on cheaper models to come out I mean work which the team is doing to get the factories ready today to try and make that happen yeah by the way the amount of work required to to make a lower cost car is insanely
46:09high but like it is harder to get like 20% of the cost out of a car than it is to design the car and build the entire Factory in the first place yeah it's like excruciating and it's and and there's not a lot of movies made about the the the heroes who got 20% of the cost out of a car but let me tell you there should be got that is incredibly heroic this little changes it's not like I yeah it's like they should be the heroes go 20% cost out of a car is like damn I have a l respect for them talking movie it's like you know I think you only could make a compelling movie but it just no no like if you actually saw how hard if people actually saw how hard
46:53was to do that' be like whoa that's damn hard just yesterday we were talking about potting I mean honestly like literally yeah I mean there's there's a lot of when I do call it sort of like getting cost out of things it's kind of like it's like game of pennies it's like Game of Thrones but pennies you know first approximation youve got if you've got 10,000 items in a car very rough approximation and each of them cost $4 then you have $40,000 car so then if you want to make a you know a $35,000 car if you to get 50 cents on average out of the 10,000 items every every part yeah and it's like you know and then obviously the best is you delete some
47:36parts in fact we found be able to delete a lot of parts I'm very I'm I'm very excited about the the Cyber cap design and and and the both you know how how we're rethinking the design of a car for the Cy cap to desig it alra high volume production and then designing machine that builds the machine Mach that that I think is also revolutionary and it's just there's no other car company that's even trying to do what we're doing like I don't even heard of it actually in fact I'm certain there is no one like like I'm I think this this the on your machine that bus the machine like it's a it's it's inherent like it's it's would it's designed to be like
48:26times better than tral Factory like cycle time cycle time and like Park deletion CH movement I don't think any other car company has the same level of like integration of thought that we have when it comes to like when you design a part from a white sheet of paper who's going to make it where's it going to be made how's it going to be shipped how's it going to be assembled into the vehicle and like at any one point if something is done in a CBO it becomes a bottleneck of either cost or time efficiency but with the with the road taxi you know the development like we've done a good job on the combining all that and then like blowing up how it's
49:07made and saying it should be made this way and rethinking it all so it's the most efficient Factory possible that shows our it will show in our CeX efficiency when we deploy it shows in the number of Parts shows in the Simplicity of vehicle but also how it performs in terms of like end user State yeah just to close out just on the energy front also in 25 we were we will have started manufacturing at the mefactory shankai we'll continue to increase our overall storage deployments with power wall 3 we plan to continue expanding our supercharging Network getting more or on our Network 4680 cell
49:42ramp as Elon talked about that would keep going and then there's we also will have our lithium Refinery starting to produce so there's a lot which is going on great yeah so many things I mean like crazy thing is like Tesla is winning basically on almost every single thing we're doing if we're not winning now we're on try to Wi in arenas where there are entire large companies that that's the only thing they do yeah I mean it's it's a company there multiple companies within the company yeah TLA there's like many companies within one yeah fastic thank you guys just a few more what is going on with the Tesla roadst fun things well I just like to
50:26thank I suffering deposit holders of the Tesla Roadster you know the reason it hasn't come out yet is because it is it the Roadster is not not just the icing on the cake it's the cherry on the icing on the cake and so you know our our our larger mission is is to accelerate the progress towards a sustainable energy future you know try to do things that maximize the probably the future is good for for Humanity and for GR and and so so that necessarily means that like the things like that are kind of like desert we' like we' all love to work on the Tesla next Tes Rose be super it is super fun and we are working on it but but but it has to come behind the more things
51:12the things that have a more serious impact on the good of the world so just thank you to allot longsuffering tesar Royer deposit holders and and and we are actually finally making progress on that and we we're close to finalizing the design on that it's really going to be something
51:36spectacular you know a friend of mine Peter teal you know people think Peter T and our Rivals we're really good friends to be clear you know Peter you know was lamenting how you know the future doesn't have flying cars well we'll see more yeah great thank you very much the next one is quite similar to other questions we've had so I might combine it with with the final question so briefly could you just detail how robot texi will roll out will it start with a tested deployed Fleet and then allow customers to to add theirs on like a subscription model and then we harder 3B capable level
52:23five Asing the hardware three what we saw with 12.5 was it was easier to make rapid progress with starting with Hardware 4 and figuring out the solution then back porting it to Hardware 3 instead of directly working on Hardware 3 given that Hardware 4 has more like fundamental Hardware capabilities I think that Trend will continue into the next few quarters as well but we first figure out the solution rapidly with ai4 and then backport it write the kernels it just takes longer to Lop those things because it's not fundamentally supported in the hardware and it's emulated but yeah it's initially working on Hardware 4 backboarding it to Hardware
53:033 yeah so it's actually the answer is we're not we're not 100% sure but but but as a sh mention because by by some measures H 4 has has really several times the capability of Hardware 3 it's it's easier to to get things to work with hardw 4 and then it takes a lot of effort to sort of squeeze that functionality into Hardware 3 and you know there there is some chance that Hardware 3 is does not achieve the safety level that allows for
53:44unsupervised FSD you know there were some chance of that and and if if that turns out to be the case we will upgrade those who have support Hardware 3 FSD for free and we have designed the system to be operatable you so and it's it's really what you know just switch sort of switch out the computer type of thing the camera the cameras are you know they're they're capable but any we don't we don't actually know the answer to that but if it does turn out we'll take we'll make sure we take care of those who board on three great in the last few minutes that we have left we will try to get in some analyst questions the first question will be coming from Pierre
54:37faru at New Street Pierre please feel free to unmute yourself thanks a lot guys for taking my question I was wondering about like the compute you're you're ramping up so you gave like interesting statistics on how much you have you said you you don't feel your compute constraint and I was wondering you know how you are putting to work this additional compute is that a game for you of creating like larger and larger models like next generation of models that are larger the way open ai go from gpt3 to GPT 4 or is that more like you're set on your model and you need to throw more and more compute to accelerate the base of learning to to improve improve
55:29reliability and then I had a quick followup real quick on your roll out in Texas and in the in California next year the plan as you see today is it to roll out like a fleet or two with a car that will start with like a supervisory like some onboard supervision someone sitting at the wheel just in case and removing the supervisor progressively or are you aiming for going fully fledged without even a human supervisor when you get
56:05started okay well I guess we're going I'll answer yeah the first part of the question the the nature of real world AI is different from say an llm in that you you have a massive amount of context so that like the the you've got case of Tesla seven or eight cameras that that you know nine up to nine if you include the internal camera that that
56:33that so you got gigabytes of context and and that that is then distilled down into a small number of control outputs you know whereas it's it's like you don't really it's very rare to have in fact I'm not sure any a LM out there do gigabytes of context and then you've got to you've got to them process that in the car with a very small amount of compute power
57:03so you know it's it's a it's all doable and it's happening but it is a different problem than than what say a gemini or an openi is doing and now part of the way you can make up for the fact that the inference computer is is quite small is by spending a lot of effort on training and just like just like a human like you the more you train on something the less the less mental workload it takes when you try when when you when you do it like when the first time like a human starts driving it absorbs your your whole mind but then as you train more and more on driving get very good then you the driving camera becomes a background task it doesn't it only
58:00absorbs a small amount of your mental capacity because you have a lot of training so we can make up for the fact that the inference computers it's it's it's tiny compared to you know a 10 kilowatt Bank of gpus because you got you've got a few hundred Watts of inference compute we can make up that with heavy training so
58:26yeah that's and and then there's also vast amounts the actual you know pedabytes of data coming in are tremendous and then sorting out what training is important with you know of the vast amounts of video training video data coming in Fleet what is actually most important for training that's difficult but as I said we're we're not currently training Compu construct want elaborate yeah like you
59:02mentioned the training help both train larger models also that train quicker but in the end you still got to pick which models are performing better so the validation effort to picking the models because the miles SP intervention is pretty large we trve a lot of miles know in close loop we do have simulation and other ways to get those metrics those two help but in the end that's a big b yeah that's why we're not training Compu constraint Alone and there's other actors of scaring as well just like data figuring out which data is more useful that that is an important task and we're focusing on that yeah so as it relates to the second part of your question Pier about Safety
59:47Drivers and rolling it out each state has different requirements that you know in terms of how many miles and how much time you need to have a safety driver and not have a safety driver we're going to follow all those we're not going to violate whatever regulations are out there but safety is a priority but the goal is obviously at when we're ready and safety is there we all D from from from the r chair yeah I mean I guess like we we think that we'll be able to have driverless Teslas of doing paid rides next year
1:00:17sometime next year all right thank you and our next question comes from Adam Jonas at Morgan Stanley Adam please feel free to unmute yourself okay thanks everybody just had a question about the relationship between Tesla and xai many investors are still not clear how the work at xai is is truly beneficial to Tesla some even take the view that the two companies may even be in competition with each other in terms of talent and Tech and even your time Elon so what's your message to investors on that relationship between Tesla X and where do you see it going over time
1:01:01thanks well I should say that you know X has been helpful to Tesla AI you know quite a few times in terms of you know things like scaling up lot like training you know just even like recently in the last week or so improvements in to training where if you you're doing a big training run and and not fails be able to continue training and and easy to recover from on training run X has been pretty helpful so but but there are different problems you know XI is working on artificial general intelligence or artificial super intelligence tells us trying to make autonomous cars and autonomous robots they're different problems so yeah I mean I think we've said
1:02:03this before also like all not all AI is equal right I mean there's AI is a broad spectrum yeah and we have our own swim lengthes yes there are certain things which we can collaborate on if needed but for the most part we're solving different issues yeah t tell us focus on real world and as I saying it is quite a bit different from the llm because you have you have massive context in video and and some amount of audio that's got to be distilled very like with with extremely efficient inference compute I do think Tesla is the most efficient in the world in terms of inference compute like because out of necessity we have to we have to be very good at inut
1:02:56efficient inance we can't put 10 10 kilowatts of gpus in a car we've got a couple hundred wats and you know it's pretty well-designed Tesla AI ch but it's still a c under on but there are different problems I mean it's like like the stuff that X add is like when it's running in bance I mean it's it's it is running inance like like it's answering questions it's answering Tes questions on on a 10 kilowatt rank it's like can't put that in a car it's a different problem please no no exactly so you know X is because because I felt there wasn't there wasn't a truth seeking digital super intelligence company out there like that's what it came down to
1:03:53they needed to be a true like like an AI company that is just very rigorous about being truthful it's I'm not saying X is perfect but that is but that that that is at least the explicit aspiration even if something is you know Politically Incorrect it should still be truthful I think this is very important for AI safety so anyway I think a XI will has been
1:04:23helpful to Tesla and will continue to be helpful to Tesla but they are very different problems great you know and I mean like think like like what is like what what other car company has it has a world class trip design team like zero what other car company has a world class AI team like Tesla does zero those all startups they created from scratch great thank youan and I think that's unfortunately the time that we have for today we appreciate all your questions and we look forward toing quarter thank you very much and goodbye
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版