股东会·中文精读

Tesla 2025 年度股东大会

2025 Annual Shareholder Meeting | Tesla

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

大多数公司的股东大会是用来催眠的。马斯克自己在台上就这么说:

你们看过别的股东大会吗?说实话,如果你需要睡觉,那可以看。股东会就像催眠曲。我们这场是炸场的。(马斯克)

他说这话的时候,身后站着一排真人大小的机器人,没有一根吊线,就在台上晃来晃去地跳舞。会场被布置成赛博朋克夜店的风格。帕洛阿尔托工程总部里,这些机器人二十四小时在办公室溜达,没人看着,电量低了就自己走去充电。

这是特斯拉2025年度股东大会。马斯克开场就说,接下来要讲的东西,已经超出了特斯拉的新篇章,够写一整本新书。

Optimus:他口中"有史以来最大的产品",以及一个"无限金钱漏洞"

马斯克对Optimus的定调毫不含糊:这会是人类历史上最大的产品,比手机大,比任何东西都大。

他的推理方式很马斯克:地球上每个人都想要一个自己的R2-D2或者C-3PO,谁会不想要?而且Optimus比这两个机器人更好用。R2-D2只会对着你哔哔叫,你根本听不懂它在说什么,还得靠C-3PO翻译。Optimus不需要翻译。

数量级上,他给了一个比例:工业场景里,每三到五台机器人对应一台个人机器人。全球最终可能会有数百亿台Optimus。

产能规划听起来像在开玩笑,但他讲得很认真。先在弗里蒙特建一条年产一百万的产线,叫一号线。然后建一条年产一千万的。一亿台一年的产线放哪儿?他说不知道,也许放火星。

再往后,他相信年产量能到十亿台。

这些数字背后是他反复讲的那套逻辑:人类喊了几十年要消除贫困、让每个人获得优质医疗,实际上只有一条路能走通,就是人形机器人。

他举的例子是手术。Optimus最终的精度会超过人类最好的外科医生,超出人类能力的上限。贫困也一样,靠再分配解决不了根本问题,得靠生产成本趋近于零。

然后是那段他明知会被断章取义的设想:

如果一个人犯了罪,也许我们可以给他一种更人道的约束方式。你免费获得一台Optimus,它会一直跟着你,阻止你犯罪。除此之外你想干什么都行。这样你就不需要把人关进监狱之类的地方了。(马斯克)

他承认这听起来很疯狂,但坚持认为未来就是这个样子。他给的阅读推荐是伊恩·班克斯(Iain Banks)的《文化》系列,说如果你想了解他心里的未来长什么样,大概就是那个样子,带一点阿西莫夫,带一点海因莱因。真要贴标签的话,他脑子里的AI与机器人未来是"班克斯式"的。

经济上的推演更直接。AI加机器人可以把全球经济规模扩大十倍,甚至一百倍,看不到明显的上限。所以他给了Optimus一个很互联网的说法:无限金钱漏洞。

他还顺带预言,也许未来根本不存在钱这个东西,货币的计量单位会变成瓦特,看你能调动多少电力。

所以我的意思是,拿好你的特斯拉股票。(马斯克)

公司使命改了一句话,但这句话分量很重

特斯拉成立时的使命是"加速可持续能源的到来"。马斯克说这件事基本做到了:电动车从不存在变成主流,Model Y是地球上销量最高的车,电池包、太阳能,很多公司后来都跟着走。

现在使命更新为实现可持续的富足(sustainable abundance)。

他解释这个词的方式是反问:你能想象的最好的未来是什么?我们想做的是把那个未来造出来。

他给出的答案有两半。一半在物质上,人们想要什么就有什么,商品、服务、医疗全部满足。另一半是,不能因此把自然毁掉,雨林要留着,国家公园要留着,好看的地方都还在。

两半合起来,就是他说的可持续的富足。他还补了一句,如果有人能想出更好的未来版本,他很愿意听。

你的车是穿靴子的猫,只是你不知道

FSD(Full Self-Driving,完全自动驾驶)这段,马斯克用了一个很妙的比喻。

你家里有只猫,正趴在沙发上。你告诉朋友,这只猫其实是穿靴子的猫,它能穿上靴子戴上帽子,能唱歌能跳舞还能耍剑。朋友一定觉得你疯了,那就是只猫。直到猫真的站起来做了这一切,你只能张着嘴说不出话。

特斯拉现在有几百万辆车在路上,它们就是穿靴子的猫。除了Cybertruck,它们看起来都是正常的车,好看但正常。可它们其实是智能的,能自己开。

麻烦在于没人知道。

我们发现有些人买了FSD,从来没打开过。我们的反应是,什么?你至少该试一次。(马斯克)

现在销售和服务团队会坐在客户旁边,手把手教怎么用,有多简单。试过几天的人基本就回不去了。

版本14让事情接近一个临界点:允许驾驶员发短信。因为这才是人真正想干的事,也是他们已经在干的事。目前车对视线离开道路还管得很严,但他说未来一两个月,看完安全数据之后就会放开。

他提到一个很现实的细节:现在有些人为了发短信,会先把FSD关掉,发完再打开。这比开着FSD发短信更不安全。

欧洲是另一个堵点。那边连监督版的FSD都不允许使用。马斯克描述他跟监管机构的对话:

我们手上有几十亿公里的数据,证明FSD提升安全性。他们说,我们得开一堆委员会会议。我说,是,但人命关天。(马斯克)

他把这种对话形容成卡夫卡小说。所以他呼吁欧洲车主去推动监管机构放行。中国这边已经拿到部分批准,对方告诉他,明年二月或三月前后会有全面批准。

Cybercab:没有方向盘,也没有一条像汽车的生产线

Cybercab是第一辆从设计之初就为无人监督自动驾驶准备的车。没有踏板,没有方向盘,没有后视镜。所有优化都指向一个目标:自动驾驶模式下最低的每英里成本。

生产就在股东会所在的这个工厂,明年四月启动。

这辆车特别的地方在生产系统。马斯克说,Cybercab的制造系统更接近高产量消费电子产品,而不是传统汽车产线。结果是节拍时间最终能压到十秒以内,理论上可以做到五秒。

做个换算就明白这意味着什么。Model Y的产线一分钟下线一台车,一年大概五十万辆。如果节拍到五秒,同一条线的理论产能是五百万辆。

他承认这需要几年才能达到,但方向是确定的。这些车未来会满街都是,会改变道路的样子,让人一眼看出这是未来。

AI5:一块只服务自家软件的芯片,成本不到对手的十分之一

马斯克说得很直白:要让机器人真正工作,必须有好的AI芯片,而且这块芯片要便宜、要省电。

他给出的AI5对比数据是这样的:功耗大约是英伟达Blackwell的三分之一,性能大致相当,成本不到对方的十分之一。

为什么能做出这种差距?他归因于两点。

第一是专用。英伟达必须服务过去和未来所有客户的需求,所有客户写过的软件都得能跑,这种兼容性包袱极重。特斯拉只需要让自己的软件跑起来,所以可以把芯片大幅简化。

第二是整数运算。这一点他讲得比较技术,但值得展开。

整数运算在硬件层面比浮点运算高效得多。逻辑门更简单,功耗更低,同样面积能塞进更多算力。麻烦在于,你要用整数做推理,就得在训练阶段就针对整数推理做优化。整个行业都在为浮点推理做训练,特斯拉走了一条很少人走的路。

马斯克自己说这属于小众的技术细节,可它决定了整块芯片的能效。用通用性换能效和成本,取舍就在这里。

AI5会在四个地方生产:台积电台湾、三星韩国、台积电亚利桑那、台积电德州。AI6的改进方案已经想好了,他希望在AI5投产不到一年内,在同一个工厂切换到AI6,所有性能指标翻倍。

他对芯片的投入程度已经到了这个地步:

我现在对芯片超级上头,你们大概能看出来。我脑子里全是芯片,真的会梦见芯片。AI5的物理设计草图,我现在能凭记忆画出来。(马斯克)

但真正的问题不在设计,在产能。就算把台积电和三星最好的情况全部外推,芯片还是不够。

所以他提出了一个叫TeraFab的构想。Giga是十亿级,Tera是万亿级,比超级工厂还要大一个量级。他看不到别的路:要拿到想要的芯片量,只能自己建一座巨型晶圆厂。

规模上,他说一座至少得是每月十万片晶圆起步,也许这样一个厂只是一个综合体里的十座之一,最终合起来是每月一百万片晶圆。他还提到要顺带解决内存和先进封装。

对现有合作伙伴,他的态度很务实。台积电和三星都很值得尊重,跟英特尔也许值得谈一谈,但还没签任何协议。他还开了个玩笑:特斯拉已经同意买下这些厂产出的所有芯片,所以在台积电和三星眼里,这基本是台印钞机,造得越快,钱到得越快。问题是,还是不够快。

太阳能是唯一解,电池的价值被严重低估

马斯克讲能源的时候喜欢从宇宙尺度开始。

太阳占太阳系总质量的99.8%,木星占0.1%,剩下的0.1%里包括地球。所以就算你把木星整个扔进热核反应堆烧掉,太阳输出能量的占比四舍五入还是百分之百。

地球上建聚变堆,在他看来是跟天上那个巨型聚变堆竞争,赢不了。太阳能的未来是必然的,他还预测会有大量太阳能AI卫星,特斯拉会在里面扮演角色。

电池这段是全场信息密度最高、也最容易被忽略的部分。

美国的总发电能力大约是一太瓦,但平均用电量不到半太瓦。差在哪里?昼夜和季节的波动。没有电池,电力没法缓冲,发出来的电用不掉就浪费了。

结论是:就算一座新电厂都不建,光靠电池就能让美国的能源输出翻倍。

他说这才是美国增量能源的主要来源,比表面上看起来重要得多。

产品上,MegaBlock把公用事业级储能变得很容易部署,更多零部件做进电池内部,到场放下就能用。未来MegaPack会集成更多电力电子,直接输出三十五万伏,连变电站都省了。就像Powerwall接进房子一样,一边接电网,一边接家里主线,完事。

充电网络上,特斯拉建了全球最大的超级充电站网络。北美做得太好,其他车企干脆直接来用特斯拉的网络。马斯克的说法是:行啊,我们没意见。

供应链上,得州科珀斯克里斯蒂建了中国以外最大的锂精炼厂,起步产能对应约五十吉瓦时的锂,之后继续扩。旁边还有正极材料工厂。他的理由是地缘政治风险,最坏情况下造电池的原料得握在自己手里。4680电池也在持续改进,会用在Cybertruck、Cybercab和Optimus上。

关于供应商关系,他给了一个很实在的框架:供应商能供多少芯片和电芯,特斯拉就买多少。超过这个量还是不够,那就只能自己造,否则就被卡住。

明年同时量产三个大东西

2026年特斯拉要同时启动三个重量级产品的量产:Optimus、Tesla Semi、Cybercab。

Tesla Semi已经在百事可乐等公司跑了一段时间,明年在内华达北部工厂开始量产。

Cybertruck的描述依旧很有画面感:史上最硬的卡车,真的是防弹的,比保时捷911快,拖曳能力超过福特F-350。

产品线S、3、X、Y都做了更新。他的建议很简单,没试过的去约个试驾,把FSD打开试一次。

问答环节:从产量目标到意识上传

提问环节的信息量不比演讲少。

产量目标。马斯克说,既然相信无人监督的自动驾驶在几个月内就能达到明显优于人类的安全水平,那就该开始大幅提产,因为车的价值主张已经完全变了。

他给了一组目标:明年产量提升大约百分之五十,明年年底年化产能到二百六十万到二百七十万辆,2027年底到四百万辆,2028年底到五百万辆。他强调这是"期望目标"。

他也讲了提产的难点。一辆复杂的车有一万多个零部件,整个系统的速度取决于最慢、最笨的那个环节。

干法正极的教训。有人问到干法正极电极,马斯克的回答很坦诚:比想象中难太多。它看起来能成,相对湿法确实有成本优势,但如果能把时钟拨回去,他大概会选湿法。原因是在高产量、高可靠性下的量产难度被低估了。

Roadster。新一代Roadster的演示定在2026年4月1日。他特意选了这个日子,说自己留了后路,可以声称只是开玩笑。但他又说,这大概会是有史以来最激动人心的产品演示,不管成不成。量产在演示之后大约十二到十八个月。创始人系列的车主会被邀请,Roadster最早的预订者也会优先。

他对Roadster的定位讲得很清楚:它不是可持续富足的必要条件,连蛋糕上的糖霜都算不上,是糖霜上的那颗樱桃。但世界上应该有远远超出以往存在的酷技术,哪怕他自己用不上,知道它存在、看到未来在发生,也让人兴奋。

两个瓶颈。有人问限制因素是什么,他答得干脆:芯片和电力。

月球和火星。火星城市和月球基地,特斯拉的车和Optimus是天然搭配。载人的话需要加压车辆,可以理解成下一代的月球车、火星车。

成本与通缩。Cybercab每英里成本,他说努力之后有路径做到当前美元计价的两毛钱以下。他还顺手讲了一遍通缩逻辑:货币供应量对商品和服务量的比值决定通胀,如果商品和服务的产出增速快过货币供应,结果必然是通缩。他说政府超支大概也追不上生产率提升的速度。

车队总量会降,里程会涨。全球大约有二十亿辆汽车。自动驾驶普及后,车队总规模会下降,但行驶里程会上升。原因很人性:现在要不要穿过拥堵的城市,取决于你要忍受多少痛苦,两个小时堵车大部分人就不去了。如果只是坐在Cybercab里看场电影或者干点活,那就像坐在小客厅里,去的人自然变多。

监管节奏。他预计监管批准的速度大致会跟Cybercab的产量匹配,可能稍微紧一点。他专门感谢了Waymo,说对方铺了路,很有帮助。他的判断是,一旦无人车在城市里变得司空见惯,监管说"不"的理由会越来越少。等到有几十亿英里的事故统计数据摆在那里,证明自动驾驶能救命,拒绝就很困难了。

意识上传。有人问Neuralink能不能把人的意识快照上传到Optimus。他的回答是:

这个技术在某些时点会变得可能,大概不到二十年。但你不会完全是同一个人,你会有点不一样,因为你在一个机器人身体里。心智快照也不会精确,会挺接近,但不会完全一样。反过来说,五年前的你是现在的你吗?也不是,很多东西都变了。(马斯克)

太空太阳能。在星舰之前,把东西送上轨道的成本太高,太空太阳能解决不了任何实际问题。现在他看到了这样一条路径:星舰的每吨入轨成本最终会低于航空货运。比如一架747从本地飞悉尼运货,星舰做同样的运输,每吨成本能更低。

这就打开了很多可能,最明显的是太阳能AI卫星,把AI的算力逐步搬到轨道和深空。深空里能获取的太阳能比地球表面多十亿倍以上。他提到卡尔达舍夫二级文明,也就是能利用恒星相当一部分能量的文明。哪怕只用到太阳能量的百万分之一,也必须去太空搞太阳能。

往地球回传只能传很小一部分,传多了会把地球烤化。他补了一句,从比例上看,地球在太阳面前就是一小粒面包屑。

SpaceX的股票。有人问支持者能不能持有SpaceX。他说自己想了很久,确实希望支持者能拿到SpaceX的股票,但美国有二千五百名股东的门槛,超过就会被视为事实上的上市公司。他不排除某天SpaceX还是得上市,尽管上市公司有一堆坏处。他还提到,作为上市公司的寄生成本这些年一直在涨,各种无谓的诉讼让公司很难高效运转。

FSD的里程估算。有人问他当年怎么估的。他说自己习惯把不确定的估计做到最接近的数量级,当时判断真实值更接近一百亿公里,也就是大约六十亿英里,而不是十亿公里,但也不会到一千亿公里。就取最近的十的倍数。

芯片团队是怎么来的。他强调特斯拉像十几家初创公司捆在一起,成立以来只做过一次像样的收购,就是SolarCity。芯片团队和AI团队都是从零搭起来的,原因简单得很:它变成了限制因素。

Hardware 2时代用的是英伟达。但当时英伟达专注做AI服务器硬件,那是个聪明的赌注,如今他们是全球市值最高的公司。马斯克对黄仁勋和英伟达评价很高,说自己非常钦佩。问题是,那时候英伟达不想做低成本、低功耗的车载AI计算机。所以只能自己组队。他招来了Jim Keller,做出AI3、AI4。他承认AI5阶段犯过一些错误,现在回到了正轨,之后到AI6会保持很快的节奏。

安全指标上,他的说法是:AI4能比人类安全百分之二百到三百,也许四百。AI5能做到百分之一千甚至更好。

车闲下来能干什么。这段是全场最有想象力的商业构想之一。

如果车的智能过剩,那闲置的时候可以把车当成分布式AI推理节点。他对车主的说法是:你的车晚上停在车库里,愿不愿意让它替你赚钱?特斯拉每月付你一百或两百美元,换取在你不用车的时候跑AI推理任务。

他顺带讲了个有点好笑的理由:这样车里的AI也不会无聊。他设想过一辆车被困住,一天的高光时刻就是开车,那剩下的时间它干什么?

特斯拉最终可能会拥有世界上最多的AI推理算力。如果有一亿辆车,每辆跑一千瓦推理,那就是一百吉瓦的分布式推理,还自带散热、电力电子和分布式供电。这个市场现在大概被估值为零。但这件事太顺理成章了。(马斯克)

对比一下:美国全国平均用电大约是四百六十吉瓦。一百吉瓦不是个小数字。

其他零碎。得州工厂可以搞一个Cyber Longhorn。特斯拉保险会继续扩张,但保险监管是州一级的,规则各州不同,非常复杂,他的原话是有点像合法的敲诈。Cybercab运营时特斯拉会自己承保。Optimus的训练算力最终会花掉数百亿美元。对xAI的投资已经获批,未来跟xAI和SpaceX都有很多合作空间。还有个小插曲,特斯拉App还没有覆盖所有语言,他当场说这个必须补上。

这场股东会真正说了什么

把演讲和问答拼起来看,特斯拉的叙事重心已经完成了一次迁移。

过去十年,这家公司的核心命题是"用电动车和电池替代化石能源"。现在这套东西变成了基础设施,成了基本盘。新命题是用AI和机器人把商品与服务的边际成本压到接近于零,然后在这个基础上谈富足。

支撑这个命题的有四根柱子:Optimus提供通用的物理劳动,FSD和Cybercab重构交通,AI5和自建晶圆厂解决算力供给,太阳能加储能解决能源供给。马斯克自己点明了两个终极瓶颈,芯片和电力。整场演讲其实都在回答这两个问题怎么解决。

他给出的时间表密度前所未有:明年四月Cybercab投产,明年Optimus V3量产,明年Semi量产,明年产量提升百分之五十,2026年4月1日Roadster演示,2027年Optimus V4,2028年Optimus V5,二十年内意识上传技术可能成真。

这些话里有大量他自己承认的"不该说的话"。他开场就讲了,我知道有些话会被剪成片段断章取义,但我还是要说。

值得留意的反而是那些最不性感的数字。美国发电能力一太瓦,平均用电不到半太瓦,电池能让现有产能翻倍。这个算术没有任何科幻成分,却可能是整场演讲里最确定、最被低估的一条。

同样确定的是他自己承认的失误。干法正极选错了路线,AI5设计走过弯路。这些细节让那些宏大数字有了可信度,一个从不承认错误的CEO讲出来的十年规划,反而是最不可信的。

如果你想把这场股东会转化成具体行动

  • 开特斯拉且有FSD的用户,这周就打开它跑一趟熟悉的路线。很多人买了从没用过,而版本14的体验跟一年前完全不是一回事。
  • 在欧洲的用户,可以去找当地监管机构的公众意见渠道,反馈你希望监督版FSD获准上路。马斯克在台上明确说这需要用户施压。
  • 判断特斯拉长期价值时,把"一亿辆车变成一百吉瓦分布式推理集群"这条单独算一遍。这个选项目前市场给的估值接近于零。
  • 盯住2026年4月1日的Roadster演示和明年Optimus V3的亮相,这两个是验证马斯克时间表可信度的最近节点。
  • 做硬件或制造的人,把干法正极这个案例记下来。实验室里更优的路线,如果在量产可靠性和产能爬坡上过不了关,代价会超过它带来的成本优势。
  • 关注储能行业而不是只盯电动车销量。电池让存量发电能力翻倍这件事,政策敏感度和资本关注度都远低于它的实际影响。
  • 对"二十年内意识上传"这类说法保持兴趣,但别当规划依据。马斯克自己也说了,上传的那个不完全是你。
第三方整理文稿,仅保留马斯克本人发言。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:The Singju Post (transcription éditée)。

Welcome. So what we’re going to… yeah. And those bots are just dancing. There are no wires. Those are actual robots. Thanks, guys. First of all, I’d like to just give a heartfelt thanks to everyone who supported shareholder votes. I super appreciate it. Thank you, everyone. I’d like to thank the Tesla Board for their immense support. We have a fantastic Board, fantastic group of shareholders. Thank you all. And what we’re about to embark upon is not merely a new chapter of the future of Tesla, but a whole new book. And I’m going to talk about that. So this really is going to be quite the story, and Optimus is a fundamental part of that.

The sheer scale of Optimus… I mean, I’m going to say a bunch of things that probably I shouldn’t say, but that’s what keeps it interesting. I mean, have you watched any other Annual Shareholder Meeting? I mean, honestly, was like… I mean, if you need to go to sleep, sure. I mean, shareholder meetings are like snooze fest. I mean, ours are bangers. I mean, look at this. This is sick. And we got, like, the cyberpunk nightclub here with real robots just standing there and milling around and dancing. And, you know, around our engineering headquarters in Palo Alto, the robots are just walking around the office twenty four seven with no one minding them.

They’re just… and then they go charge themselves. And yeah. So the scale of Optimus, like I said, that’s really going to be something else. I think it’s going to be the biggest product of all time by far. Yeah. So, like, bigger than cell phones, bigger than anything. I guess the way to think about it is that every human on earth is going to want to have their own personal R2-D2 C-3PO. So who wouldn’t? But actually, Optimus will be even better than them. Like R2-D2, it’s kind of would beep at you, and it’s kind of hard to figure out what he’s got talking about. You need C-3PO to translate. But Optimus is going to be like everyone’s going to want one.

I think in terms of industry providing products and services, I think it’s probably, I don’t know, three to five robots in industry for every one that’s a personal robot. I think there could be tens of billions of Optimus robots out there. Now obviously, it’s very important we pay close attention to safety here because we do want the Star Wars movie, not the Jim Cameron movie. I love Jim Cameron’s movies, but you know what I mean. So yes. So we’re going to launch on the fastest production ramp of any product of any large complex manufactured product ever, starting with building a one million unit production line in Fremont. And that’s Line one.

And then a ten million unit per year production line here on the… I don’t know where we’re going to put the one hundred million unit production line. Maybe on Mars, I don’t know. But I think it’s going to literally get to one hundred million a year, maybe even a billion a year. And, you know, people often talk about, like, eliminating poverty, giving everyone amazing medical care. Well, there’s actually only one way to do that, and that’s with the Optimus robot. With humanoid robots, you can actually give everyone amazing medical care. In terms of Optimus will be more precise. Optimus will ultimately be better than the best human surgeon with a level of precision that is beyond human.

So I think that’s a pretty wild concept to say, okay, you… there’s always people always talked about eliminating poverty, but actually, Optimus will actually eliminate poverty. Optimus will actually give people incredible medical care. So I mean and so you start getting, like, sort of some pretty wild sci-fi sort of scenarios where in some of these things I say will obviously be taken out of context and using snippets and, you know, sitting around, but whatever. I’m still going to say them.

You know, like, I think we may… we might may be able to give people a more… if somebody’s committed crime, a more humane form of containment of future crime, which is if you… you say, like, you now get a… you now get a free Optimus, and it’s just going to follow you around and stop you from doing crime. But other than that, you get to do anything. It’s just going to stop you from committing crime. That’s really it. You don’t have to put people in like prisons and stuff, I think. It’s pretty wild to think of the various of all the possibilities, but I think it’s clearly the future.

And, you know, my book recommendation for the maybe the best, mostly utopian sci-fi future are the Ian Banks books, the culture books. So if you’re curious, like, what do I think the future is probably like? I think it’s probably a bit like that. Or, you know, Asimov to some degree, but I think it’s… and Heinlein. But in Banks, if you’re like saying, what does Elon think the future probably will be like for AI and robots? It’s kind of Banksian. So now and things do get kind of wild from an economic standpoint because at a certain point, with AI and robotics, you can actually increase the global economy by a factor of ten or maybe one hundred. There’s not like an obvious limit.

So like Optimus is kind of like an infinite money glitch. And maybe there won’t even be money in the future. Money might be measured in terms of wattage, like how much power can you bring to bear from an electrical standpoint. So I guess what I’m saying is hang on to your Tesla stock. So, yes, man, it’s pretty wild. You’re welcome. Let’s see. I think we’ve got some slides to go through. I’m going to ad lib a lot of this stuff, but so the… you know, when we started Tesla, the goal was to accelerate the advent of sustainable energy, and that is that’s what we’ve done.

I think Tesla has really led the way with electric vehicles, with battery packs, with a lot of solar, and many other companies have then followed our lead and done that. And electric cars, which used to be nonexistent, are now prevalent. And the Model Y, for example, is the number one selling car of any kind on earth. Obviously, now with AI and robotics, we need to update our mission. And our mission, I think it’s a good… it’s a great mission, which is to achieve sustainable abundance, which is… it’s like because I often ask people like, what is the future that you want? What’s the best future you can imagine?

Because we want to try to make that future, like make the best future you can imagine. And I guess probably the best future is if people can have whatever they want from a goods and services standpoint or medical standpoint. And but at the same time, we don’t destroy nature, and we keep the rainforests and the beautiful national parks and all that stuff. And so that’s what I mean by sustainable abundance, is that people can have whatever they want, have all their needs met, but we still keep all of the natural beauty that we want. I mean, if somebody can think of a better future, I’m all ears. But I think that’s probably the best way to go. So, yeah. Let’s see.

So, yes, as you know, every Tesla is designed to be autonomous. So the… it’s sometimes difficult to explain to people if they have not… or in fact, I’m sure you’ve all encountered this, where you try to tell people that the Tesla can drive itself, and they think you’re crazy or something. I mean, especially, like, apart from the Cybertruck, our cars look pretty normal. I mean, they’re good looking cars, but they don’t look super… they look normal. But I guess it’s kind of like having a cat or like… and cat’s just sitting… let’s say you’ve got a cat and it’s like just sitting there on the couch.

And you try to tell people that the cat can actually… it’s actually Puss in Boots and it can actually put on boots and a hat and swashbuckle and sing and dance. And people are like, no way, man, that’s a cat. Until the cat does all those things. And you’re like, damn, what the…? So we’ve got millions of Tesla cars out there that are the kind of like Puss in Boots. They’re intelligent, but people don’t know that they’re intelligent. They look like normal cars, but actually, they’re super smart and can drive themselves.

So I think that’s probably the single biggest thing we need to do is to educate potential customers that you can either have a cat that’s like normal cat or you can have Puss in Boots. And Puss in Boots is very cool. That’s so we’ve… these days, when people come to our stores or even people that have the car haven’t turned it on, we find. And sometimes people have paid for FSD and haven’t turned it on. We’re like, what? You should at least try it once. And so now we’re like the sales team and service team will actually sit with customers and say, look, let us show you how it works and how easy it is. And then once they’ve tried it for even just a few days, they can’t live without it.

And now with version fourteen, we’re actually getting to the point where we almost feel comfortable allowing people to text and drive, which is kind of the killer app because that’s really what people want to do and do do. And actually, now, the version’s… the car is a little strict about keeping your eyes on the road. And but I’m confident that in the next month or two, we should… we’re going to look closely at the safety statistics, but we will allow you to text and drive essentially. So yes. It’s certainly been in the current situation, which often people will actually turn off FSD to text then turn it back on, which is less safe.

So yes, that’s probably the single biggest thing is just get people aware of FSD. And then obviously, we need to get it approved in Europe. So we certainly appreciate the support of our customers in Europe pushing the regulators to approve FSD because you can’t even get a super… even just normal supervised FSD is not allowed in Europe currently, which doesn’t make any sense. And I’ve had these like crazy conversations with the regulators that seem like a Franz Kafka novel, where I’m like, well, look, we have billions of kilometers of data that shows that FSD increases safety. And they’re like, well, we have to have all these committee meetings.

I’m like, yes, but people’s lives are at stake here. So definitely, a pressure from our customers in Europe to push the regulators to approve would be appreciated. And then we have partial approval in China, and we hopefully will have full approval in China around February or March or so. That’s what they’ve told us. Yeah. But yeah. The fact that every Tesla car is capable of full self driving, every car we build and have built for the last several years is capable of full self driving is pretty wild, and most people don’t know that. So… And then we’ve got the first car that is specifically built for unsupervised full self driving to be a robotaxi. It’s called a Cybercab.

It doesn’t even have pedals or steering wheel. Yeah. So there’s no side view mirrors. There’s no… yes. So it’s very much optimized for the lowest cost per mile in an autonomous mode. And that production is happening right here in this factory, and we’ll be starting production in April next year. So the way that Cybercab is designed is it’s designed, obviously, for a purely autonomous world. But also, the manufacturing system is unlike any other car. The manufacturing system of the Cybercab, it’s sort of… it’s closer to a high volume consumer electronics device than it is a car manufacturing line.

So the net result is that I think we should be able to achieve, I think, ultimately, less than a ten second cycle time, basically a unit every ten seconds. Maybe ultimately take a few years to get there, but it’s theoretically possible to get to a five second production time. And so what that would mean is you could get on a line that would normally produce, say, five hundred thousand cars a year at a one minute cycle time, Model Y. This would be maybe as much as two million or three million, maybe ultimately it’s theoretically possible to achieve a five million unit production line if you can get to the five second cycle time. It’s a lot of cars. So these will be everywhere in the future.

And we want to look futuristic, so it changes the look of the roads. The ingredients, when you look at what Optimus is, what’s required to make Optimus and the various ingredients, what do you need to do to make high volume humanoid robot production? I think it’s worth considering that really the cars we make are already robots, but they’re four wheeled robots. So Tesla is already the biggest robot manufacturer in the world because every car we make is a robot. And when you break it down to the fundamental elements, you’ve got batteries, power electronics, motors, gearboxes. You’ve got connectivity. You’ve got a vision based AI. Hi, Optimus.

And you know, all the various pieces that you need for a humanoid robot, you need the AI chip, you need the AI software, you need to be able to manage a large fleet. And so really, Optimus is a robot with arms and legs as opposed to a robot with wheels. So, you know, Tesla’s ideally suited, I think, to succeed in this arena. You will see certainly many companies showing demonstration robots. There’s really three things that are super difficult about robots. One is the engineering of the forearm and hand because the human hand is an incredible thing, actually. It’s super dexterous. So engineering the hand really well, the real world AI, and then volume manufacturing.

Those are generally the things that are missing. One or more of those things are missing from other companies. So Tesla is the only one that has all three of those. So this is the Optimus kind of initial, it’s kind of a prototype production line. The high volume production line will be very automated, obviously, but this is really the production line that we use to make the prototypes. So you can get a sort of rough sense for what it takes to build the robot. Still pull the finger.

And then as I’ve said before, I think once we reach about one million units per year of sustained production or in excess of that, I think probably the cost of production is around twenty thousand dollars in current year dollars. So this will be certainly very affordable. And, yeah, like I said, I think Optimus will ultimately increase the size of the economy probably by a factor of ten or more. You know, next year, we start production with Optimus version three. This is what you’re seeing here is Optimus version two point five. Optimus three is an incredibly good design. The Tesla engineering team is amazing.

When you see Optimus three, it will seem as though that there’s someone like a person in a robot outfit, which is how we started with Optimus. Really, it’s going to be something special. And then Optimus four, that hopefully starts production in twenty seven. And then Optimus five in twenty eight. So it’s kind of like an annual release cycle with significant improvements with each one and gigantic increases in the scale of production. Sustainable abundance via AI and robotics. That’s the future we’re headed for. And as I think most people here know, the safety statistics show that miles driven on FSD are much safer than miles driven without it.

So what this will translate to ultimately is saving the lives of millions of people and preventing hundreds of millions of accidents. So a massive increase in lives saved and tragedies avoided. It’s going to be amazing. How many people here have tried fourteen point one? Okay. All right. Cool. Yeah. You can see that even with the point releases, it’s getting quite a bit better. It should be pretty smooth at this point. But really, fourteen point two, there are major changes to fourteen point two and then fourteen point three. And I think by fourteen point three is when we’re really going to be at the point where you can just pretty much fall asleep and wake up at your destination.

And then I’ve been putting a lot of time into the new Tesla chip design, because in order to have a functional robot, you have to have a great AI chip. And it needs to be an inexpensive chip, and it needs to be very power efficient. So we believe the AI5 chip will be probably about a third of the power of, say, something like a Blackwell, NVIDIA Blackwell, which is a great chip, for roughly comparable performance and much less than ten percent of the cost. So this is a chip that is very much optimized for the Tesla AI software stack. It’s not meant to be a general purpose chip. It’s meant to be an amazing chip for the Tesla AI software.

And I mean, a couple of things that I think make, how is Tesla able to achieve such an improvement? I think it is because we are specialized. We’re not trying, NVIDIA has to serve the superset of all past and future customers. So all of their requirements, all of the software that they’ve written has to work, which is a very difficult problem, whereas we just need to make it work for our software. And so we were able to simplify the chip dramatically. And then we also, I think we’re unique in this, but we have an integer based system. Integer operations are fundamentally more efficient than floating point operations.

So we can do floating point, but the vast majority of our inference is done in integer, which is, if you’re familiar with sort of logic gates, the simplicity of integer, integer is much more power efficient, much more silicon efficient. But you actually have to train for integer inference, which everyone else is training for floating point. That’s kind of a niche technical detail, but it’s actually very important. So, yes, this is going to be a great chip. So this chip will be made in basically in four places: TSMC Taiwan, Samsung Korea, TSMC Arizona and TSMC Texas. And we already know what improvements to make for AI6.

So I’m hopeful that we can within less than a year of AI5 starting production, we can actually transition in the same fab to AI6 and double all of the performance metrics. I’m super hardcore on chips right now, as you may be able to tell. I have chips on the brain. I dream about chips, literally. I can draw the, you know, the at least the broad brush stroke physical design of the AI five chip by heart at this point. It’s a good chip. It’s a good chip, sir. So this is really key. Now one of the things I’m trying to figure out is how do we make enough chips? So I have a lot of respect for the Tesla partners, TSMC and Samsung. Maybe we’ll do something with Intel.

We haven’t signed any deal, but it’s probably worth having discussions with Intel. But even when we extrapolate the best case scenario for chip production from our suppliers, it’s still not enough. So I think we may have to do a Tesla TeraFab. So it’s like Giga, but way bigger. I can’t see any other way to get to the volume of chips that we’re looking for. So I think we’re probably going to have to build a gigantic chip fab. Got to be done. So anyway, some of the stuff I’ve already talked about. Yep. We’ve done a tremendous amount for sustainable energy, and that is only going to grow over time.

The world is moving towards a solar battery economy, which is ultimately where it was going to go anyway, but what Tesla does is accelerate that outcome. Sometimes people don’t understand quite how much energy comes from the sun. So the sun is ninety nine point eight percent of all mass in the solar system. Jupiter being point one percent and then point one percent is miscellaneous, Earth being in the miscellaneous category. So the total amount of energy, the sun, people would say, well, we’ll build fusion reactors on Earth.

It’s like, well, actually, the giant fusion reactor in the sky is basically impossible to beat to such a degree that even if you could burn Jupiter in a thermonuclear reactor, the amount of energy produced by the sun would still round up to a hundred percent. That’s how much energy the sun produces. So solar power is necessarily the future. And I think there’s going to be a lot of solar powered AI satellites. And I think Tesla’s going to play a role in that. We’ve obviously refreshed the product line, so S three X Y. If people haven’t tried the model S, three, X or Y or the Cybertruck, I recommend at least getting a test drive or a test ride as the case may be.

Try out the full self driving, and I think you’ll be blown away. So those who do not, if you might be listening and don’t have a Tesla, you should try one. And, of course, we’ve got the Cybertruck, which is the toughest truck of all time. It’s literally bulletproof, faster than a Porsche nine eleven, and can out tow a Ford F three fifty. So it’s a great car, great truck. And then starting next year, we manufacture the Tesla Semi. So this, we already have a lot of prototype Tesla Semis in operation. PepsiCo and other companies have been using the Tesla Semi for quite some time. But we will start volume production at our Northern Nevada factory in twenty twenty six.

So we got two big products or three, three massive products starting production next year. We got Optimus. We got Tesla Semi, and we got the CyberCab. And then battery packs. So if you look at total US power generation capability, it’s roughly a terawatt. But the average power usage is less than half a terawatt. And that’s because there are big differences in power usage between day and night. So the daily and seasonal variations in power consumption mean that the United States and really every country is only using about half, is only producing about half as much electrical energy as it could. Because without batteries, there’s no effective way to buffer the energy.

So what batteries actually enable is even if you don’t build any incremental power plants, you could double the energy output of the United States just with batteries. This is a super big deal. And in fact, I think that’s really where most of the incremental energy production in the United States is going to come from, is literally batteries. So a bigger deal than it may seem. And then we keep improving the battery design, so the MegaBlock, which makes it really easy to deploy utility scale batteries. So we’ve just simplified and brought more of the components to be internal to the batteries. So you can just show up and drop off a battery and it works.

And then hopefully with, well, not hopefully. Over time, we will actually add more and more of the power electronics so that MegaPack will actually be able to output up to thirty five kilovolts directly. So you won’t need a substation is what I’m saying. You can just literally drop it off, kind of like the way that a Powerwall you just connect it to the house. The utility wires go on one side and the other side goes to the house mains, and that’s it. So we want to get mega pack to the point where you just literally take the utility wires and you plug them in, and it just works. Then we’ve also built the world’s largest supercharger network. So we do a lot of things here at Tesla.

That’s the biggest supercharger network in the world by far. And ultimately, you’ll be able to go anywhere on earth using a Tesla supercharger. And it’s all pretty close to anywhere on earth, but it’s going to be ultimately just anywhere. It will just work anywhere. So the supercharger team has done great work expanding that and improving the efficiency of the supercharger network. And in North America, they did such a good job that the other car companies basically said, well, we’ll just use the Tesla supercharger network. Like, okay, sounds good to us. It’s always important to have a plan on safety in the factory. So we continue to improve safety for our factory workers.

We care a lot about their well-being. And, you know, one way you can just tell if a company is a good company or not, if you just walk through the factory or walk through the office and catch the vibe. And the vibe in the Tesla factory is good. People are happy. That’s how you know it’s a good company. We’ve also put a lot of investment into raw materials. So we’ve built in South Texas and Corpus Christi the biggest lithium refinery outside of China, I believe. So it’s going to, it’s starting off at about fifty gigawatt hours of lithium, and we’ll expand from there. So this is very important to have in a worst case scenario that we have the ingredients necessary to make a battery.

Very important. And then we’ve got here on this site the cathode factory, which is just the sort of giant building about a half mile that way. And we’re just making sure that we, from a supply chain standpoint, are resilient against any potential geopolitical challenges. And then also at this factory, we also make the 4680 cell, which is getting better and better. And that 4680 cell will be used in the CyberCab – it is being used in the CyberTruck and will be used in the CyberCab and also in Optimus. So that’s going well. But we continue, obviously, to get sales from many suppliers. It’s kind of like the chip fab thing. We’ll take as many chips as our suppliers can provide us.

But then beyond that, if they can’t provide us with any incremental cells or chips, we kind of have to make them ourselves or we get stuck. That’s the Tesla Semi Factory. So it’s going to be pretty cool to see these going down the road in scale. So that’s basic plan. Sustainable abundance for all.

All right. So yes, I guess we can go into Q and A now. Maybe at the next Annual Shareholder Meeting, we’ll have Optimus take some of the questions. That would be cool. So I see Alexander, well, thank you for your help, by the way. Please go ahead.

Sure. And we can pay. Yeah. Actually, Tesla will be way, way, way bigger than Berkshire long term. It’s going to be kind of nutty. So, all right. We’ll do the next one. Maybe we’ll do it in the – yeah. Like, yeah. Maybe the downtown arena or the soccer stadium or something like that. And get your security checked. We want to make sure that you’re safe.

All right. Sounds good. Thank you very much for listening to us.

Well, we have done that a few times. I guess we could extend it again. I will extend it for at least another quarter and then play it by ear after that.

Sure. Yeah. Obviously, we need a vehicle that’s big enough to fit a wheelchair accessible. So I think is the Robovan or robust or whatever we call it. I mean, it’s not like we’re slowing down because we want to slow down. It’s like we’re spinning like a zillion plates here. So but I do think it would be very cool because I think aesthetically as well, it just would change the look of the roads and make it feel like the future. So it won’t be long before we make that. But since we do have Optimus, we’ve got CyberCab, Optimus and Semi all next year. It will probably have to be maybe the year – a couple of years from now or something like that.

But we certainly will make a wheelchair accessible vehicle.

Yes. I guess the dry cathode, man, that’s turned out to be a lot harder than we thought. So I mean, it does look like it’s going to be successful, and it will have some cost advantage relative to wet cathode. But if I had to wind the clock back, I would probably have gone with wet cathode instead of dry cathode because it just turned out to be a lot harder to make it high capable of high volume production with super high reliability. So yes. But we will be scaling up battery cell production at Tesla and looking for cell production from our suppliers as well because we’re going to be ramping up production very dramatically at Tesla.

Now that we believe we have full self-driving, that we have autonomy solved or at least within a few months of having it unsupervised autonomy solved at a reliability level significantly better than human – that means it’s time to ramp up production because the value proposition is now much greater than a regular car. The killer app really is for people, can you text and drive? Or can you sleep and drive? Can the car take you to your destination? Or do you need to pay attention and be and have to drive it? And before we allow the car to be driven without paying attention, we need to make sure it’s very safe. Like I said, we’re on the cusp of that.

I know I’ve said that a few times, we really are at this point. And you can feel it for yourselves with the 14.1 release. So we’re going to try to – we’re going to push to expand vehicle production as fast as we possibly can. So aspirationally, we’d aim to increase vehicle production by about 50% by the end of next year. So that’s – it’s very hard to increase production, but that’s roughly – I don’t know, maybe we get to like – I’m just guessing, at exit rate by the end of next year of around 2.6 million, 2.7 million vehicles annualized production and then aim to get to maybe 4 million by the annualized rate by the end of ’27 and then maybe 5 million by the end of ’28.

Those are rough – those are our aspirational goals. So these are – this is a gigantic increase in output, which means that the entire supply chain has to move in unison with that increase in volume. And the nature producing a large complex product is that it moves as fast as the least lucky, dumbest element in the entire system, and there’s 10,000 plus items. But like I said, this really is a new – it’s not just a new chapter for Tesla, it’s a new book. And that new book is massively increasing vehicle production and ramping up Optimus production faster than anything’s ever been ramped up before in history.

Oh, yes. Sure. Absolutely. Definitely. Okay. Yes. So all Founder Series can be invited? Yes, yes. I mean, it’s the least we could do, frankly, for people that have – our long suffering Roadster reservation holders. I feel confident in saying it will be the most exciting product unveil ever. And I hope that – whether it goes well or doesn’t go well.

Well, I guess it’s according to whoever put down their deposit in that sequence. So that’s the – but you’ll get a very early one. And I mean, the new Roadster is very much sort of like – it’s not even the icing on the cake, it’s the cherry on the icing on the cake. So I mean, it’s really kind of like – it’s not essential for sustainable abundance, let me put it that way. But I do think there should be very cool technology in the world that is – you know, it’s way beyond anything that’s ever existed. And I think I’d like those – like, even if I could never have access to those things, I’d like to know that they exist and see the future happening.

So I think it will be inspiring to a lot of people. And just it’s – it’s like the coolest car, if it even is a car, that has ever – that will probably ever exist.

Yes. Chips and electricity are the two limiting factors.

Everyone. It’s like – yeah. Yeah.

Oh, yes, yes. That’s a point. A semiconductor Tera factory? Yes. Well, I mean, thing is that we actually have agreed to buy all the chips that are made from the fab. So it’s basically a money printing machine for TSMC and Samsung. It’s like literally the faster you make the chips, the faster we send them money. But it’s still not going at just fast enough. So that’s why I think, as far as I can see, the only option is to go build some like very big chip fab. And then you go to solve memory and packaging too. But otherwise, you just tap out at whatever the chip production rate is. And so I guess, Tera would be – you’d want to say it’s got to be at least 100,000 wafer starts per month size fab.

And maybe that would be one of ten in a complex. So it would ultimately be one million wafer starts per month. Yeah. Exactly. You can tell when it gets the giggle factor, that’s probably a good sign that we’re onto something special. But I wouldn’t be surprised if long term, it’s like a million wafers a month.

Well, Optimus is going to play a big role. Optimus and, I think, Tesla vehicles will play a big role on the moon and Mars. So for a moon base and a Mars city, Tesla vehicles and Optimus robots are natural fit for building and operating a moon base than a Mars city.

Yeah. Yeah. It’s a Cybertruck. We’ll need to drive around in a pressurized vehicle if there’s a person inside. But, yeah, it’ll be something cool, next level moon buggy or Mars buggy.

Well, in terms of cost per mile, I mean, we do see a path with a lot of work to get below $0.20 a mile in current year dollars. And I somewhat agree that things will probably be deflationary, as productivity increases because you can think of money supply as being the ratio of goods and services to like what’s the growth in goods and services versus the growth in the money supply? That ratio is basically inflation. And so if the output of goods and services grows faster than the money supply, then you necessarily have deflation or vice versa. They try to make it sound complicated, but it’s not.

So I’m not sure that even government overspending can actually—I think government overspending will be lower than the increase in productivity of goods and services, which would imply deflation as you expect.

Economies—sorry, economies of scale? In like auto manufacturing, for instance, it only makes sense if lots of people buy your vehicles. Oh, the total number of vehicles will decrease. The sort of vehicle fleet out there is about two billion cars. Two billion if you add up old cars and trucks that are not at a scrap yard, I think it’s around two billion. But that number would decrease with autonomous vehicles. So the total fleet size would drop. I actually think miles driven will increase because it is now much less painful to travel somewhere.

So if somebody is thinking about traveling across a busy city, then they’ll take into account, well, how much pain do I have to deal with if I have to go through two hours of traffic? I probably won’t do it. But if you’re just, say, sitting in a Cybercab watching a movie or doing some work, then it’s just like sitting in a little lounge. And so I think you’ll see probably a significant increase in total miles driven, but at the same time, a decrease in the total active fleet of vehicles.

Yeah. I think the rate at which we receive regulatory approval will roughly match the rate of Cybercab production. It will be maybe a little tight, but it’s about right. And I’d like to thank Waymo for paving the path here. It’s very helpful. So, yeah, but I think we’ll be able to deploy all the Cybercabs that we produce. And the other thing is, like, once it becomes extremely normal in cities, it’s just going to become like, the regulators will have just fewer and fewer reasons to say no.

And then you’ve got this—you know, you’ve got the accident statistics at scale and you can show that autonomous miles save lives, then and you’ve got unequivocal, you know, billions of miles to prove it, then I think it’s hard for regulators to say no.

Yeah. I mean, it’s not immediate. But if you say that down the road, would you be able to, say, with the Neuralink, have a snapshot of or what is an approximate snapshot of somebody’s mind and then upload that approximate snapshot to an Optimus body, I think that at some point, that technology becomes possible. And it’s probably less than twenty years. But of course, you won’t quite be the same. You’re a little different because you’ll be in a robot body. And the mental snapshot will not be precise. It’ll be probably pretty close, but not exactly the same. On the other hand, are you the same person that you were five years ago? Nope. I mean, a lot of things have changed.

So yeah, but I guess at some point, if you want to be uploaded to a robot body, my guess is that becomes possible.

Well, space solar power—until the advent of something like Starship, where the cost per ton to orbit drops by orders of magnitude, the cost of getting payload to orbit is so high that there was no possibility of space solar power solving anything, really. Now with Starship, like, I see a path to Starship cost per ton to orbit being lower than air flight. Like, if you were to fly, you know, do a long distance air flight with cargo, like, say, took a 747 from here to Sydney, Australia or something like that, I think, ultimately, Starship will be able to do that trip for less cost per ton than an aircraft. So then you—so now you’ve got that opens up a very wide range of possibilities.

Like, the most obvious one, I think, is actually solar powered AI satellites, sort of to move the AI to orbit and essentially deep space over time. Because you can actually access over a billion times more energy from the sun in deep space than you can on Earth, scaling if you—to scale to Kardashev—to make any progress on a Kardashev Type II scale, which is using some nontrivial amount of energy from the sun, you kind of have to do space solar power. Now you could only beam a tiny amount of that back to Earth or you would melt Earth. So Earth actually receives a very, very tiny amount of the sun’s energy. We’re—I mean, Earth is a tiny dust mote.

We see that Earth to scale with the sun, we look like a little crumb. So to scale civilization, like I said, to be at all relevant on a Kardashev Type II scale, like to even use a millionth of a percent of the sun’s energy, you really have to have use solar power in deep space. And you could beam some of that back to Earth too, but you can only beam a tiny bit of it back or you’d melt Earth.

Well, thank you for your support.

Yeah. It is a tough problem that you highlight. I mean, basically, unfortunately, over time, the parasitic load of being a public company has just grown over time. And so you get all these spurious lawsuits, obviously, and they just make it very difficult to operate effectively as public company. So but I do want to try to figure out some way for Tesla shareholders to participate in SpaceX. That would be very cool. I’ve been giving a lot of thought to how to give people access to SpaceX stock because I do want supporters to have SpaceX stock. But there’s a sort of 2,500 shareholder threshold before you become a de facto public company. But I don’t know.

Maybe at some point Tesla—maybe at some point SpaceX should become a public company despite all the downsides of being public.

Well, generally, I mean, in terms of the estimates, like for the self-driving stuff, generally try to get things to within an order of magnitude. So it seemed to me like probably—and this was technically in kilometers, but it would be closer to ten billion kilometers, which is roughly six billion miles, than it would be to one billion. But I thought it would not take one hundred billion kilometers. So it’s really just you’re trying to—when you’re trying to guess something where there’s a lot of uncertainty, just try to get the estimate to the nearest order of magnitude, closest factor of ten. And that’s why I said probably around ten billion kilometers or six billion miles.

And then, yeah, chip—the reason Tesla created a chip team, and it’s important to note, like, Tesla is like a dozen startups in one. And, you know, we’ve only really done one major acquisition, which is SolarCity, and then some very small ones. So all of this is almost all of this is organic. So I built the chip team from scratch and the AI team from scratch. It was just because it became a limiting factor. So for Hardware 2, we used NVIDIA. But NVIDIA was, at that time, focused on making really AI server hardware, which obviously was a smart bet. They’re currently the most valuable company in the world. And Jensen Huang and his team have done an incredible job at NVIDIA. My hats off to them.

I’m a huge admirer of Jensen and NVIDIA. They’ve done amazing work. But they didn’t want to do a low cost, power efficient car computer at the time, AI car computer. So I was like, well, okay, I guess we need to start a chip team to solve that. So then that’s when I hired Jim Keller, and we built the chip team and did AI3, then AI4. And then we—to be honest, we made some mistakes there with AI5, but now AI5 is back on track. And we’ll have a very rapid cadence to AI6 and so forth. And this is really necessary for the car to—like with AI4, I think we can get to 200 or 300 percent better than human safety, maybe 400 percent better.

But with AI5, I think we can do 1,000 percent better or maybe even better than that than human safety. So at a point, actually, it might be too much intelligence for a car. So, like, I was thinking, what if you get stuck in a car and you have too much intelligence? But then one of the things we could do is when the car is idle is use the car as a massive distributed AI inference fleet. So with the concerned customers, we’re like, do you want your car to earn money for you while it’s sitting in your garage at night? I don’t know. We’ll pay $100 a month or $200 a month or whatever the right number is. If you allow Tesla to do AI inference workloads when you’re not using your car.

So that will also help the AI in the car not get bored. Because, like, I sort of imagined, like, what if I got stuck in a car? You know? And then—well and the highlight of your day was driving. It’s like, you know—but they don’t always want to drive. So then what do you do the rest of the time? So I think Tesla could actually end up having the largest—Tesla might end up having the most amount of AI inference compute in the world. Like, if you think, maybe if we had one hundred million car fleet, and at some point, we may have more than one hundred million car fleet, and they’ll have AI six, AI seven, you know.

And if you’re able to run a kilowatt of inference on a one hundred million car fleet, now you’ve got one hundred gigawatts of distributed inference with built in cooling and power electronics and distributed power. Probably the market is valuing that as zero point zero right now is my guess. But it seems like an obvious thing to do. If you’ve got distributed inference AI and you’ve got the power and the cooling, which is very difficult to do the power and the cooling, and one hundred gigawatts is a lot. I mean, the average power consumption in the U.S., I think, is around four sixty gigawatts for the—that’s the entire electrical consumption of the U.S.

So if you do it one hundred gigawatts, that would be a pretty big number. But, yeah, it’s basically something as a limiting factor, and then we take actions to address the limiting factor.

Oh, yeah.

Like a Cyber Longhorn?

All right. We’ll do a cyber longhorn for the factory.

Yep. My permission. Get away. Perfect. You don’t have to do anything. Tesla does energy. We’re going to go through the normal channels. We’ll buy it from a store. All that—sure. If you can give away Tesla, it’s totally cool. Certainly, you don’t need my permission to give away a car. We’ll take, like, maybe a couple more questions and then call it a night. Okay. All right.

No.

I guess, well, so we’re aiming for the—so the product unveiled will be of the Roadster two, which will be very different from what we’ve shown previously. That demo event will be April one of next year. I have some deniability because, like, I could say I was just kidding. But we are actually tentatively aiming for April first for what I think will be the most exciting, whether it works or not, demo ever of any product. And then I guess production is probably about twelve to eighteen months after that. I think production is probably a year or so after that. Well, I can’t give away secrets, but you won’t be disappointed. All right. We’ll take one last question, I guess.

Well, I mean, I think actually long term, the AI is going to be in charge, to be totally frank, not humans. If artificial intelligence vastly exceeds the sum of human intelligence, it is difficult to imagine that any humans will actually be in charge. So we just need to make sure that AI is friendly.

Go ahead. Sure. Yes. Okay.

Need to not have the questions be super long. I mean, Tesla Insurance is trying to expand as quickly as possible, but the regulatory structure for insurance is extremely complex and works on a state by state basis. So it’s really somewhat of a racket. And the rules for insurance are different with every state. It’s a very, very complicated thing. So yes, there’s—but I’m aware that insurance often is too expensive and doesn’t take the right things into account. But so all I can say to that is, yes, we’ll keep expanding Tesla Insurance. When the car is operating as the cybercab, Tesla will simply self insure. So that kind of solves that. But insurance is a real pain in the neck for sure.

But okay, I do need to end this at some point. I’ll take one question and one question there.

Yeah. There’s a lot of training compute needed for Optimus. And because the AI chip in the robot is relatively weak because it’s really limited on power, you can make up with that with a lot of training to have a lot of training result in a very efficient AI that can run on a low power chip in robot. So we will actually have to spend a lot of money on training. Like, ultimately, it will be tens of billions of dollars on training compute. So it’s a big number.

Yeah, I think potentially. Yeah, I think could—yeah, I think there’s potential for accelerating that. So, yeah, did the xAI investment thing get approved? Yes. No? Yes? Okay. Okay. Whatever. You know, whatever, like, it’s like, Tesla and some other company that I have an interest in, and it’s like always quite complicated to do things. It has to go through a lot of hoops to happen. But I do think there’s a lot of potential for collaboration with xAI in the future and with SpaceX. All right. Okay. This is the last one. Yeah.

You mean the app?

Shoot. I thought we had it in all languages. Okay. Well, definitely, the app needs to be in all languages. All right.

从现场发言里挑出的 10 句原话,附中文译文(AI 翻译,以英文原话为准)。点时间戳可跳到视频对应位置。

Optimus 真的能消除贫困。Optimus 真的能给人们提供极好的医疗。

Optimus will actually eliminate poverty. Optimus will actually give people incredible medical care.

其实,长期来看,Tesla 会比 Berkshire 大得多得多得多。

Actually, Tesla will be way, way, way bigger than Berkshire long term.

所以,作为目标,我们希望到明年年底把汽车产量提高约 50%。

So aspirationally, we’d aim to increase vehicle production by about 50% by the end of next year.

我认为,Optimus 和 Tesla 车辆将在月球和火星上发挥重要作用。

Optimus and, I think, Tesla vehicles will play a big role on the moon and Mars.

所以,如果商品和服务的产出增长快于货币供应增长,那就必然会出现通缩,反之亦然。

And so if the output of goods and services grows faster than the money supply, then you necessarily have deflation or vice versa.

我其实认为,行驶里程会增加,因为现在去某个地方出行没那么痛苦了。

I actually think miles driven will increase because it is now much less painful to travel somewhere.

现在有了 Starship,我认为有一条路可以让 Starship 每吨入轨成本低于航空飞行。

Now with Starship, like, I see a path to Starship cost per ton to orbit being lower than air flight.

所以我认为,Tesla 最终可能拥有最大的——Tesla 最终可能拥有世界上最多的 AI 推理算力。

So I think Tesla could actually end up having the largest—Tesla might end up having the most amount of AI inference compute in the world.

如果人工智能远超人类智能的总和,很难想象还会有任何人类真正掌权。

If artificial intelligence vastly exceeds the sum of human intelligence, it is difficult to imagine that any humans will actually be in charge.

所以我们实际上必须在训练上花很多钱。最终,训练算力上的投入会达到数百亿美元。

So we will actually have to spend a lot of money on training. Like, ultimately, it will be tens of billions of dollars on training compute.

资源与出处

相关视频

同频道、同主题或相邻年份