特斯拉第三季度财报电话会,马斯克上台后的第一句话没提交付量,也没提毛利率。
他说的是:特斯拉正处在一个关键的拐点。
接下来一个多小时,他谈机器人、谈芯片、谈变电站、谈外科医生、谈公司治理,还谈了一个被他称作“无限金钱漏洞”的东西。财报会通常把季度数字念一遍就收场,这场更像一次路线图交底。
按信息密度排序,值得停下来的部分在下面。
扩产这件事,他压了很多年才敢按下开关
拐点指什么,他解释得很直白。
路上已经跑着数百万辆特斯拉。这些车不用回厂,一次软件更新就能变成全自动驾驶车。工厂现在每年造两百万辆上下。
过去他一直压着产能不扩。理由听上去很朴素:在没看清无人监督的全自动驾驶到底能不能做成之前,扩产是一场赌博。工厂建起来、车卖不掉,代价太大。
我对实现无人监督的全自动驾驶有了清晰的判断,尽快扩产是合理的。(马斯克)
这句话的分量在于它改变了资本开支的逻辑。以前是等,现在是加速。他把话说到“尽我们所能地快”这个程度,还补了一句,这个决定他自己憋了很久。
有意思的地方在于,他没有宣布任何新工厂选址。扩产的路径是先榨干现有产线的潜力,让供应链跟上,然后再谈增量。这个顺序在后面产能的问题里还会再出现一次。
1 太瓦和 0.5 太瓦之间,藏着一倍的电力
储能这条线,他给了两组数字。
美国全部的连续发电能力大概是 1 太瓦。按 24 小时平均算,实际用电只有 0.5 太瓦。差距来自白天和晚上的用电量差得太远,发电能力必须按峰值配,剩下的时间就在浪费。
用电池把白天的电挪到晚上用,美国的电力输出等于凭空翻一倍,一座新电厂都不用建。
建电厂有多难,他描述得很实在:周期长,许可多,整个行业都不习惯快。电池提供的是另一条路径,交付周期以月计。
Megapack 这条产品线上,这次公布了 Megablock 和 Megapack 3,也提了 Megapack 4 的规划。
Megapack 4 的工程重点是把变电站里的一大堆设备直接做进产品,输出端直接给到 35 千伏。这个改动听起来很技术,实际影响不小:储能项目落地时不再需要先等一座 35 千伏变电站建好。真实项目里,卡住进度的常常是并网这一环,电池反而没那么卡。把这半段吃掉,铺开的速度会完全不一样。
v14 已经能装了,但他劝你先别装
美国任何一位车主,进设置里选择“我想要高级软件”,就能拿到 v14。马斯克说,网上第一批用户的反应是“相当惊讶”。
紧接着他自己泼了盆冷水:大部分人应该等 14.2 再装。
原因是工程顺序。每一次自动驾驶软件大架构更新,第一优先级都是安全,舒适排在后面。初版会安全,但动作可能顿挫。把粗糙的边缘磨平需要时间。
这种坦白在发布会语境里不常见。多数公司会把“安全第一”当成一句正确的话说完就过,他把它变成了具体的用户建议:现在别升。
推理能力会在 14.3 或 14.4 加进来。他举的例子是停车。
一个停车场快满了,车不会傻等在门口找位置。它会先把你放在店门口,然后自己去找车位。它清楚入口附近有空位的概率极低,所以不在这上面耗时间。360 度视觉让它在找空位这件事上强过人类。
真正难的地方在于把推理塞进车里。服务器上慢慢想没问题,车里必须实时决策。这套东西要在 AI4 的算力上跑起来,是团队当下最硬的骨头。
一台“像活物”的车,和 100 吉瓦的分布式算力
他提到一个内部工具:特斯拉现实模拟器。
它生成的视频和真实路测视频看起来完全一样。强化学习因此可以在虚拟世界里大规模地跑,不必等真实里程一点点攒够。训练效率的量级差别就在这里。
下一步是把参数量提升一个数量级。他特意说明这不在 14.1 里,别指望现在就能看到。
这台车会感觉像一个活物。(马斯克)
这句话说的是 AI4,还没算上 AI5。
他甚至顺着这个比喻往下想了一层。如果车里的智能多到会“无聊”,那闲置的车队可以组成一张分布式推理网络。
账是这么算的。假设未来车队规模到 1 亿辆,每辆在空闲时提供 1 千瓦的高性能推理能力,加起来就是 100 吉瓦。这些算力自带电力和冷却,配电、机柜、散热这些数据中心最贵也最麻烦的部分,车本身就解决了。
这个设想现在只是设想,但它解释了特斯拉对“车上到底该装多少算力”的态度为什么和同行不一样。别人在算成本,它在算未来的资产。
robotaxi:年底前,Austin 先摘掉安全员
时间表给得很具体。
年底之前,Austin 至少大部分区域会取消安全员。也就是几个月内。他把这条称为“也许是最重要的一个数据点”。
年底前计划在 8 到 10 个大都市区运营 robotaxi,具体取决于各州审批。内华达、佛罗里达、亚利桑那在名单上。他还提醒了一句,大部分监管申请是公开信息,谁都能查到进度,不用只听公司说。
为什么走得这么慢,他给的理由是怕。
我们对部署这件事保持偏执。哪怕只出一次事故,也会是全球头条。(马斯克)
这话不算客套。他后面补了操作细节:就算监管不要求,进入新城市也会先配安全员,大概三个月。这三个月不是单纯跑里程,是专门找这座城市的怪东西。某个特别难的路口,某个地图上标得含糊的匝道。一次万分之一概率的失误就足以出事,而这类问题只有在真实城市里才会浮出来。
三个月确认没问题,再撤掉安全员。这个节奏比大多数人预期的慢,但它把失败的概率压到了一个可以承受的位置。
Cybercab 没有方向盘,杀手级应用是发短信
明年第二季度投产,这是产能扩张里最大的一块增量。
Cybercab 没有方向盘,没有踏板。整车围绕一个指标优化:每英里运营成本。
马斯克用了个有点刻薄的比喻。现有车型身上还留着“马车的痕迹”:方向盘、踏板,还有为加速和过弯做的性能取向设计。Cybercab 不需要这些,它只求平顺,最高时速 85 到 90 英里就够了。
会不会牺牲毛利?他说不会,因为需求会“相当疯狂”。
支撑这个判断的论证特别朴素。
真正的杀手级应用是这个:你能不能坐在车里发短信。如果你告诉一个人,这车好到你全程都可以低头看手机,那凡是买得起的人都会买。(马斯克)
他顺手给了一个因果解释。全球范围内交通事故在上升,很大一部分原因是开车时发短信。人低头看手机的那几秒,驾驶质量是崩的。自动驾驶在这件事上是净增益,它不需要完美,只要比低头看手机的人强。
产能方面,年化 300 万辆这个数字,他说 24 个月内可能达到,也许更快。前提是整条供应链同步跟上。他把话说得很清楚:能扩多快,取决于供应商能跟多快。Cybercab 之后建新厂,是更后面的事。
AI5:靠删东西换来的 40 倍
这一段的密度最高,值得慢慢读。
先说代工安排,这里他主动做了一个澄清,因为自己之前公开说过的话容易引起误会。AI4 由三星制造,做得很好。AI5 会同时交给台积电和三星,一个在亚利桑那,一个在德州,都在美国境内。他还提到一个细节,三星那座厂的设备其实比台积电略先进一点。
性能上,他给了一个数字,并且特意强调不是笔误:
按某些指标,AI5 会比 AI4 好 40 倍。不是 40%,是 40 倍。(马斯克)
40 倍怎么来的?靠删。
AI5 删掉了 AI4 里的传统 GPU,因为芯片主体本身就在做这件事。图像信号处理器也删了。他用了“一长串删除清单”这个说法,并且强调这些删除非常关键。
删完之后的效果是,整颗芯片能塞进半个光罩,而且从内存到特斯拉自研加速器、ARM 核、PCIe 模块的走线还留有余量。他说这是一颗漂亮的芯片,自己投了很多精力进去。
为什么敢删得这么狠?他给了一个结构性理由。
英伟达要服务很多客户,需求范围极宽。芯片里逻辑块一多,块与块之间的互联就会爆炸式增长。他打的比方是修高速:你不知道两个区域之间到底有多少车流,就只能到处修巨型高速,最后这变成一个几乎无解的设计难题。英伟达把这件几乎无解的事做得很好,这一点他说得很清楚,不含贬义。
特斯拉只有一个客户。需求来源单一,所以能做极简。
预期收益他给了两个:性能功耗比好 2 到 3 倍,性能价格比好 10 倍。紧接着自己补了一句,布丁好不好吃要吃了才知道,芯片得真造出来、真上量。
产能策略有点反直觉:目标是 AI5 供过于求。
多出来的芯片放哪?放数据中心。现在特斯拉的训练集群是 AI4 加英伟达的混合方案,他强调不会替代英伟达。AI5 富余的产能可以直接进自家机房。这个对冲很聪明:车和机器人的需求预测从来不准,多造的芯片不会烂在仓库里。
还有一个概念值得单独拎出来:智能密度。
马斯克说,按每 GB 的智能水平算,特斯拉 AI 可能比其他人好一个数量级。原因跟团队聪明与否无关,约束更狠。这套模型必须塞进 AI4 那个盒子,塞不进去就上不了车,没有第二种选择。
这种在极小空间里榨性能的纪律,到了 AI5 会变成红利。同样的密度,算力大 10 倍,过去所有的积累一次性兑现。
他也顺手对比了 Grok。Grok 是另一个物种,巨大到根本不可能塞进车里,Grok 5 只能在 GB300 这种级别的硬件上有效运行。特斯拉的模型大概是 Grok 的 5% 到 10% 大小。两边从光谱的两端往中间走。互补的地方也有:车里的 Grok 语音交互,Optimus 的语音识别和语音生成,用的就是 Grok。
Optimus:难的不是走路,是那只手
马斯克说过一句听起来像玩笑的话:Optimus 现在确实能在公园里散步。
帕洛阿尔托的工程总部里,有机器人 7×24 小时在办公区走动。访客可以拦下一台,让它带路去某个会议室,它会真的把你带过去。他说这话的时候带着点得意,但马上收住了。
真正的难点在手上。原话是,手和前臂比机器人其余所有部分加起来都更难。
为什么难,他讲了一段解剖学。人手的肌肉大部分不在手上,在前臂。四根手指加一根拇指,每根手指的自由度、每块肌肉的强度、每根手指的长度,研究得越深越会发现都不是随便长的,全部有用处。要在工程上复刻这套结构,难度超出直觉。
第二难点是供应链。这一点决定了 Optimus 为什么不像一家创业公司能做的事。
车有成熟的供应链,电脑有成熟的供应链,人形机器人没有。想造,就得往供应链深处自己走,很多零件必须自产。马斯克的说法大意是,如果把自己放在一家做人形机器人的创业公司的位置上,他也不知道在没有强大制造技术的前提下该怎么完成这件事。
第三是规模。几百台的机器人没有意义,得做到和车同一个量级,年产百万台。
这三个条件叠在一起,才构成“通用机器人”的门槛:一只灵巧的手、一套现实世界 AI、一套能上量的制造体系。他说别的机器人项目读下来,缺的往往就是这三样里的某几样。这也是他判断特斯拉位置独特的原因,这三样恰好都是特斯拉这几年被迫练出来的。
V3 会让你想伸手戳一下
Optimus 第三代预计在第一季度亮相,大概是 2 月或 3 月。
它看起来甚至不像机器人,像一个人穿着机器人服装。真实到你需要戳它一下,才相信这是台机器。(马斯克)
这个说法有前情。Optimus 2.5 在《创:战神》首映式上和杰瑞德·莱托(Jared Leto)一起打功夫,现场没有人在遥控,视频现在网上能搜到。
路人的反应挺有意思。2.5 的腰只有 3 英寸宽,从比例上一眼就能看出不是人。但动作太像人,很多人走过去的时候压根没意识到那是台机器。
工程侧还有一个变化值得记一笔:设计评审和制造评审现在同时开,中间是一个反复的循环。设计出一个零件,制造那边说这东西根本做不出来,就回去改设计。结果是功能在涨,可制造性也在涨,两条曲线一起往上走。
马斯克对 Optimus 2 的评价很直接:几乎造不出来。这也是为什么硬件设计到量产开始都不会冻结,会持续滚动变更。很多制造上的坑,只有走到很后面才会露头。
时间表是这样的:先建一条年产百万台的产线,希望明年年底开始生产。然后是 Optimus 4,一千万台级别。Optimus 5,可能是 5000 万到 1 亿台。
爬坡速度他有个说法,很传神:一条产线的速度,取决于一万个零件里最慢、最笨、最不走运的那一个。
他还提了团队构成。Optimus 工程团队里有很多刚毕业的新人,他认为这批人非常出色。周五晚上的会留给 Optimus,经常开到午夜。周六的会属于 AI 芯片设计团队。他自己说,这两件事决定了公司的未来。
无限金钱漏洞
这个说法是马斯克自己给的。
逻辑很直白。一台 Optimus 的年生产力大概是人的 5 倍,因为它可以 24 小时工作,甚至不需要充电,插着线就能一直干。
他区分了两种 AI 增益。AI 对人类的增强是有上限的,一个人被工具放大得再厉害,一天也只有 24 小时。具身 AI 没有这个上限。
Optimus 规模化就是那个无限金钱漏洞。(马斯克)
往远处推,就是这次财报会上宣布的新使命:可持续富足(sustainable abundance)。
这个提法比“可持续能源”往前迈了一步。能源够用只是前提,目标是没有贫困,以及所有人都能用上最好的医疗。他举的例子是外科手术。Optimus 会是一个出色的外科医生,而全世界绝大多数人一辈子都碰不到一位顶级外科医生。如果这件事能被复制,意义就不在商业层面了。
他当然也补了安全前提。这类话从一位正在造机器人的人嘴里说出来,需要打点折扣听,但方向值得记下来,因为它解释了这家公司为什么把资源往机器人上倾斜。
他真正焦虑的事:造完机器人大军,然后被赶走
这段是全场火药味最重的地方,也是很多人会快进跳过、但其实最该读的地方。
马斯克把投票权这件事的动机说得非常透。他担心的不是钱。他担心的是:如果真造出了一支机器人大军,未来某一天,他会不会被赶出去。
如果我们要造这支机器人大军,我至少要有足够强的影响力。眼下这种控制权不是我要的东西,影响力才是。没有这个,我不放心去挥舞这支军队。(马斯克)
他说需要的比例大概在 25% 上下。多到有实质影响力,少到如果他哪天疯了,股东仍然能把他换掉。这个自我约束他特意讲了一遍,语气里带着自嘲。
为什么不能像其他公司那样搞双层股权?他解释了一句制度现实:超级投票权股必须在上市前设立,谷歌、Meta 都是这么做的,上市之后就再也没有这条路。特斯拉已经上市,所以只能走股东投票这一条。
然后他开火了。他把 ISS 和 Glass Lewis 称为“企业恐怖分子”。
这段话背后的机制值得拆开看,因为它跟特斯拉关系不大,跟所有上市公司都有关系。
大量被动基金、指数基金的投票权,是照着这两家代理顾问的建议走的。如果一家上市公司足够大比例的股份被指数基金持有,那么这家公司的重大决策,事实上就由这两家机构的态度决定。而这些建议未必对股东有利。马斯克的说法是,他们过去给过很多糟糕的建议,如果真的被采纳,会对公司造成极大破坏。
这是结构性的问题,不只是特斯拉一家的抱怨。当所有权和投票权被这样拆开,公司治理的名义主体和实际主体就不是同一个了。他用了“事实上的控制”这个词。
整段话的目的很清楚:股东大会上那份薪酬方案需要投票通过。他把这件事的性质从“给我更多钱”重新定义成了“给我足够的影响力,让我敢把机器人造出来”。
这场财报会真正的信息量
如果把所有细节压成一句话:特斯拉正在从一家造车卖车的公司,变成一家同时在赌四条曲线的公司,自动驾驶、储能、人形机器人、自研芯片。这四条曲线共用同一套底层能力,现实世界 AI、机电工程、大规模制造。
这既解释了它为什么看起来什么都做,也解释了它为什么每一件都做得慢。
顺带说一个容易被忽略的背景。马斯克在这场会上提到,特斯拉起家时没有任何一项核心能力,用他的话说,核心竞争力是零。电池包是自己做的,超充网络是自己从零建的,AI 软件团队是自己从零招的,北美超充网络好到其他厂商集体转用特斯拉标准。他把特斯拉形容成“十几家创业公司装在一家公司里”。
这句话放在今天读,多了一层意思:当年那些从零搭起来的能力,现在被同时押在四条曲线上。风险集中,赌注也集中。
接下来可以怎么用这些信息
- 手里有特斯拉的车,想尝鲜 v14 的,忍到 14.2。初版安全但顿挫,这是官方说法,没必要当测试员。
- 关注 14.3 或 14.4 的推理上线,那是“车会自己找车位”这类体验真正成形的节点。
- 把年底 Austin 取消安全员当成最重要的观察指标。它比任何发布会演示都硬,意味着无人监督的商业模式第一次在真实城市里跑通。
- 查监管申请。他说大部分申请是公开信息,谁都能查到进度,不用只听公司口径。内华达、佛罗里达、亚利桑那这三州是今年内的观察重点。
- 看储能的时候别只看装机量,盯 Megapack 4 的并网方案。能不能绕开变电站,直接决定项目落地的速度。
- 关注明年 Q1 的 Optimus V3 亮相和 2 月到 3 月的生产意向原型。原型和量产之间差着一条百万台产线,这两件事要分开看。
- 看财报的时候,交付数字的权重可以降一降,把注意力放到三件事上:无人监督 FSD 的落地城市数、储能部署量、Optimus 产线的开工时间。
- 如果你持有特斯拉股票,那份关于投票权的议案值得认真读一遍再投票,别默认接受代理顾问的建议。他这段话里最值得琢磨的,恰恰是“被动基金的投票权到底代表谁”这个问题。
Thank you. We are at a critical inflection point for Tesla.
So it's because the cars are all out there. We have millions of cars out there that, with a software update, become full self-driving cars. We are making a couple of million a year. In fact, with the advent of what we see now as clarity on achieving full self-driving, unsupervised full self-driving, I should say, I feel confident in expanding Tesla's production. So that is our intent, to expand as quickly as we can our future production. I was ready to do that until we had clarity on achieving unsupervised full self-driving. But at this point, I feel like we've got clarity, and it makes sense to expand production as fast as we reasonably can.
We are also making a huge impact on the energy sector with battery storage. With both Powerwall and especially with the Megapack, we are dramatically improving the ability to generate more energy from the grid. Let me sort of talk a little bit about that, which is if you look at total US energy capability, for example, there's roughly a terawatt of continuous power available in the US. But the average usage over a twenty-four-hour cycle is only half a terawatt because of the big difference between day and night usage. If you buffer the energy with batteries, you can effectively double the energy output in the United States just with batteries, pulling no incremental power plants.
It's very difficult to build power plants. They take a long time. There's a lot of permitting. It's not an industry that's used to moving fast. We see the potential there for Tesla battery packs to greatly improve the energy output per year for any given grid, US or otherwise. We are also on the cusp of something really tremendous with Optimus, which I think is likely to be, has the potential to be, the biggest product of all time. It's a difficult project. It's worth noting that it's not just automatic. I'm unaware of any robot program by Ford or GM or, you know, in the by USC of car companies. People might think of Tesla as a car company that mostly makes cars and battery packs.
It's not just an obvious fall of a log thing to make Optimus, but we do have the ingredients of real-world AI and exceptional electrical mechanical engineering capabilities and the ability to scale production, which I don't think anyone else has all of those ingredients. With version 14 of self-driving, people can see the reactions of people online. They're quite amazed. Actually, anyone in the US can get version 14 if they just go and select "I want the advanced software" in their car. If you're listening right now and you'd like to try it out, just go into settings and say, "I want the advanced software," and you will get version 14.
On the Megapack front, we unveiled Megablock, Mega Pack three. We also have exciting plans for MegaPack four. MegaPack four will incorporate a lot of what is normally in a substation and be able to output at probably 35 kilovolts directly. This greatly improves our ability to deploy Megapack because it's not dependent on building a substation up through 35 KB for MegaPack four. That's the engineering priority for Megapack. We look forward to unveiling Optimus b three probably in Q1. I think it'll be ready to show off. That, I think, is going to be quite remarkable. It won't even seem like a robot. It'll seem like a person in a robot suit, which is kind of how we started off with Optimus.
It'll seem so real that you'll need to poke it, I think, to believe that it's actually a robot. Obviously, the real-world intelligence we've developed for the car, most of that transfers to Optimus. It's a very good starting point. In conclusion, we're excited about the updated mission of Tesla, which is sustainable abundance. Going beyond sustainable energy to say, sustainable abundance is the mission, where we believe with Optimus and self-driving, we can actually create a world where there is no poverty, where everyone has access to the finest medical care. Optimus will be an incredible surgeon, for example. Imagine if everyone had access to an incredible surgeon.
Of course, we make sure Optimus is safe and everything, but I do think we're headed for a world of sustainable abundance. I'm excited to work with the Tesla team to make that happen.
I'll start off with that, and then Ashok can elaborate. We are expecting to have no safety drivers in at least large parts of Austin by the end of this year. So within a few months, we expect to have no safety drivers at all in at least parts of Austin. We're obviously being very cautious about the deployment. Our goal is to be actually paranoid about deployment because, obviously, even one accident will be front-page headline news worldwide. It's better for us to take a cautious approach here. But we do expect to have no safety drivers in the car in Austin within a few months. I think that's perhaps the most important data point.
We do expect to be operating robotaxi in, I think, about eight to ten metro areas by the end of the year. It depends on various regulatory approvals. You can actually think most of our regulatory applications are online. You can kind of see them because they're public information. We expect to be operating in Nevada, Florida, and Arizona by the end of the year. Ashok?
Thanks.
Yeah. I mean, bringing Optimus to market is an incredibly difficult task, to be clear. It's not like some walk in the park. At some point, I mean, actually, technically, Optimus can walk in the park right now. We do have Optimus robots that walk around our offices at our engineering headquarters in Palo Alto, California, basically twenty-four hours a day, seven days a week. Any visitors that come by can actually stop one of the Optimus robots and ask it to take them somewhere, and it'll literally take them to that meeting room or that location in the building.
I don't want to downplay the difficulty, but it's an incredibly difficult thing, especially to create a hand that is as dexterous and capable as the human hand, which is incredible. The human hand is an incredible thing. The more you study the human hand, the more incredible you realize it is, and why you need four fingers and a thumb, why the fingers have certain degrees of freedom, why the various muscles are of different strengths, and fingers are of different lengths. It turns out that those are all there for a reason. Making the hand and forearm, because most of the actuators, just like the human hand, the muscles that control your hand are actually primarily in your forearm.
The Optimus hand and forearm is an incredibly difficult engineering challenge. I'd say it's more difficult than the rest of the robot from an electromechanical standpoint. The forearm and hand are more difficult than the entire rest of the robot. But really, in order to have a useful generalized robot, you do need an incredible hand. Then you need the real-world AI, and you need to be able to scale up that production to have it be relevant because it's not relevant if it's just a few hundred robots. You need to be able to make Optimus robots at volumes comparable to vehicles, not significantly higher.
So trying to make a million Optimus robots per year, that manufacturing challenge is immense considering that the supply chain doesn't exist. With cars, you've got an existing supply chain. With computers, you've got an existing supply chain. With a humanoid robot, there is no supply chain. In order to manufacture that, Tesla actually has to be very vertically integrated and manufacture very deep into the supply chain, manufacture the parts internally because there just is no supply chain.
This is the kind of thing where I'm like, if I put myself in the position of a startup trying to make a humanoid robot, I'm like, I don't know how to do it without an immense amount of manufacturing technology. That's why I think Tesla is in almost a unique position when you consider manufacturing technology, scaling real-world AI, and a truly dexterous hand. Those are generally the things that are missing when you read about other robots that just don't have those three things. I think we can achieve all those things with an immense amount of work, and that is the game plan.
My fundamental concern with regard to how much voting control I have at Tesla is if I go ahead and build this enormous robot army, can I just be ousted at some point in the future? That's my biggest concern. That is really the only thing I'm trying to address with this. It's called compensation, but it's not like I'm going to go spend the money. It's just, if we build this robot army, do I have at least a strong influence over that robot army, not current control, but a strong influence? That's what it comes down to in a nutshell. I don't feel comfortable wielding that robot army if I don't have at least a strong influence.
Okay. I'm going to give quite a long answer to this question because I have to unpack this question and then answer the unpacked version. First of all, I have nothing but great things to say about Samsung. They're an amazing company. Samsung, it's worth noting, does manufacture our AI four computer and does a great job doing that. Now with the AI five, and here's where I need to make a point of clarification relative to some comments I've made publicly before, which is we're actually going to focus both TSMC and Samsung initially on AI five. The AI five chip designed by Tesla is, I think, an amazing design.
I've spent almost every weekend for the last few months with the chip design team working on AI five. I don't hand out praise easily, but I have to say that I think the TensorFlow team is really an incredible chip here. By some metrics, the AI five chip will be 40 times better than the AI four chip. Not 40%, 40 times. Because we have a detailed understanding of the entire software and hardware stack, we're designing the hardware to address all of the pain points in software. I don't think there's really anyone that's doing this thing the entire stack all the way through real-world.
You know, calibrating against the real world where you've got cars and robots in the real world, we know what the chip needs to do, and we know just as importantly, we know what the chip doesn't need to do. To sort of use some examples here, with the AI five, we deleted the legacy GPU or the traditional GPU, which is in AI four. But AI five does not have, we just deleted the legacy GPU because it basically is a GPU. We also deleted the image signal processor. This looks like a long list of deletions that are very important.
As a result of these deletions, we can actually fit AI five in a half reticle and with good margin for traces from the memory to the Tesla accelerators, the ARM CPU cores, and the PCI blocks. This is a beautiful chip. I've poured so much life energy into this chip personally, and I'm confident this is going to be a winner. Next level. It makes sense to have both Samsung and TSMC focus on AI five. Even though technically, Samsung fab has slightly more advanced equipment than the TSMC fab. These will both be made in the US, one TSMC in Arizona, Samsung in Texas. We're going to make starting off just to be confident. Our explicit goal is to have an oversupply of AI five chips.
If we have too many AI five chips for the cars and robots, we can always put them in the data center. We already use AI four for training in our data center. We use a combination of AI four and NVIDIA hardware. We're not about to replace NVIDIA, to be clear, but we do use both in combination. AI four and NVIDIA hardware, and the AI five excess production can always be put in our data centers. NVIDIA keeps improving. The challenge that they have is that they've got to satisfy a large range of requirements from a lot of customers. Tesla only has to satisfy requirements from one customer, Tesla. That makes the design job radically easier and means we can delete a lot of complexity from the chip.
I can't emphasize how important this is. When you look at the various logic blocks in the chip, you increase the number of logic blocks, you also increase the interconnections between the logic blocks. If you can think of it like there are highways, like how many highways do you need to connect the various parts of the chip? Especially if you're not sure how much data is going to go between each logic block on the chip, then you kind of end up having giant highways going all over the place. It's a very, it becomes almost an impossibly difficult design problem, and NVIDIA has done an amazing job of dealing with almost an impossibly difficult set of requirements.
But in our case, we're going for radical simplicity. The net effect is that I think AI five will be the best performance per watt, maybe by a factor of two or three, and best performance per dollar for AI, maybe by a factor of 10. We'll have to, the proof's in the pudding, so obviously, we need to actually get this chip made and made at scale. But that's what it looks like.
Yeah. So I guess I'll start with that in terms of the semi,
Well, capacity isn't quite 3 million. But it will be 3 million at some point. Aspirationally, it could be 3 million within, we could probably hit an annualized rate of 3 million within twenty-four months, I think. Maybe less than twenty-four months. Bear in mind, there's an entire supply chain, a vast supply chain that's got to also move in tandem with that. I think we're going to expand production as fast as we can and as fast as our suppliers can keep up with it. Then we're going to think about where we build incremental factories beyond that. The single biggest expansion in production will be the Cyber Cap, which starts production in Q2 next year.
That's really a vehicle that's optimized for full autonomy. It, in fact, does not have a steering wheel or pedals and is really an enduring optimization on minimizing cost per mile for fully considered cost per mile of operation. For our other vehicles, they still have a little bit of the horse carriage thing going on where, obviously, if you've got steering wheels and pedals and you're designing a car that people might want to go very direct past acceleration and tight cornering, like high-performance cars, then you're going to design a different car than one that is optimized for a comfortable ride and doesn't expect to go past sort of 85 or 90 miles an hour.
It's just aiming for a gentle ride the whole time. That's what Cyber Cap is. Do I think we'll sacrifice margins? I don't think so. I think the demand will be pretty nutty. Here's the killer app, really. What it comes down to is, can you text while you're in the car? If you tell someone, yes, the car is now so good, you can be on your phone and text the entire time while you're in the car, anyone who can buy the car will buy the car. End of story. That's what everybody wants to do. In fact, not everyone wants to. They do that. That's why, in fact, the reason you've seen an uptick in accidents, pretty much worldwide, is because people are texting and driving.
Autopilot actually dramatically improves the safety here. If someone's looking down at their phone, they're not driving very well. That's really the game changer. At this point, I feel essentially 100% confident, I say not essentially, 100% confident that we can solve unsupervised full self-driving at a safety level much greater than human. We've released 14.1, got a technology roadmap that's, I think, pretty amazing. We'll be adding reasoning to the car. Our world simulator for reinforcement learning is pretty incredible. Our Tesla reality simulator, when you see it, the video that's generated by the Tesla reality simulator and the actual video looks exactly the same.
That allows us to have a very powerful reinforcement learning loop to further improve the Tesla AI. We're going to be increasing the parameter count by an order of magnitude. That's not in 14.1. There are also a number of other improvements to the AI that are quite radical. This car will feel like it is a living creature. That's how good the AI will get with the AI four computer before AI five. AI five, like I said, is by some metrics forty times better. But just to say safely, it's a 10x improvement. It might almost be too much intelligence for a car. I do wonder, like, how much intelligence should you have in a car? It might get bored.
One of the things I thought of, like, well, if we've got all these cars that maybe are bored, well, why they're sort of, if they are bored, we could actually have a giant distributed inference fleet. If they're not actively driving, just have a giant distributed inference fleet. At some point, if you've got tens of millions of cars in the fleet, or maybe at some point 100 million cars in the fleet, and let's say they had, at that point, I don't know, a kilowatt of inference capability of high-performance inference capability, that's 100 gigawatts of inference distributed with power and cooling taken with cooling and power conversion taken care of. That seems like a pretty significant asset.
Actually, I'm not sure what you mean by AI applications outside of Tesla's core competency. We kind of didn't have any of these core competencies when we started, you know. We had zero core competencies, total competency of zero, actually. You can think of Tesla as, like, I don't know, a dozen startups in one company. I've initiated every one of those startups. We didn't used to make battery packs, stationary battery packs, but now we do. We make them for the home, make them for utility scale with Powerwall and Megapack. We created the supercharger network globally. No one else has created a global supercharger network.
In fact, the North American supercharger network is so good that basically, yeah, every other manufacturer in North America has converted to our standard and uses the Tesla Supercharger network. If it was so easy, why didn't they just do it? The Chef Design team started that from scratch. The Tesla AI software team was started from scratch. I literally just said, hey, we're going to start this thing. I posted it on Twitter, now X. Join us if you'd like to build it. In fact, Ashok was, I believe, the first person I interviewed for the Tesla autopilot team, which we now call Tesla AI software team because it is the AI software team. Core competencies created while you wait.
Optimus at scale is the infinite money glitch. It's difficult to express the magnitude of, like, if you've got something that, like, if Optimus, I think, probably achieves five times the productivity of a person per year because it can operate twenty-four seven. It doesn't even need to charge. It can operate tethered. It's plugged in the whole time. That's why I call it, like, if you're true of sustainable abundance, where working will be optional. There's a limit to how much AI can do in enhancing the productivity of humans. There is not really a limit to AI that is embodied. That's why I called the infinite money glitch.
We also have actually added a lot of new engineers as well to the team. A lot of the credit for the Optimus engineering is actually also near new engineers, many of them that are just out of college, actually. The Optimus engineering team is a very talented engineering team. I'd say, like, wow, actually. The Optimus reviews at this point are that there's the engineering review and then there's the manufacturing review. Being done simultaneously. With an iterative loop between engineering design and manufacturing. We design something and we say, like, oh, man, that's really difficult to make. We need to change that design to make it easier to manufacture.
We've made radical improvements to the design of Optimus while increasing the functionality, but making it actually possible to manufacture. I'd say Optimus two is almost impossible to manufacture, frankly. But my two-point, we've gone from a person in a robot outfit to what people have seen with Optimus 2.5 where it's doing kung fu. Optimus was at the Tron premiere doing kung fu, just up in the open, with Jared Leto. Nobody was controlling it. It was just doing kung fu with Jared Leto at the Tron Premier. You can see the videos online. The funny thing is, a lot of people walked past it thinking it was just a person.
Even though with Optimus 2.5, you can see that it has a waist that's three inches wide. It results in not a human. But the movements were so human-like that a lot of people didn't realize they were looking at a robot. What I'm saying is, Optimus three will be a giant improvement on that. Made at scale, like I said, a very difficult thing. The Optimus engineering and manufacturing reviews and there's the Friday night meeting with Optimus, which sometimes goes till midnight. My Saturday meeting is with the AI chip design team. Two things are crucial to the future of the company.
Yeah. There are different forms of AI. The XAI, so Grok is like a giant model. You could not possibly squeeze Grok onto a car. That's for sure. It is a giant piece of a model. With Grok, it's trying to solve for artificial general intelligence with a massive amount of AI training compute and inference compute. For example, Grok five will actually only run effectively on a GV 300. That's how much of a beast Grok five is. Whereas Tesla's models are, I don't know, maybe about less than 10% the size, maybe closer to 5% the size of Grok. They're really at the problem from very different angles. XAI and Grok are competing with Google Gemini and OpenAI ChatGPT and that kind of thing.
Some of it's complementary. For example, for Grok voice, being able to interact with Grok in the car is cool. Grok for Optimus voice recognition and audio voice generation is Grok, so that's helpful there. But they are coming at it from kind of opposite ends of the spectrum.
Well, I think even if the regulators weren't making us do it, we'd still do that as the right sort of cautious approach to a new market. Just to make sure that we're being paranoid about safety, I think it makes sense to have a safety driver or safety occupant in the car when we first go to new markets to confirm that there's not something we're missing. All it takes is one in 10,000 trips to go wrong, and you've got an issue. It just makes sure, like, is there some peculiarity about a city, like a very difficult intersection or something that's an unexpected challenge in a city for that one in 10,000 situation.
We probably could just let it loose in these cities, but we just don't want to take a chance. What we're talking about here is maybe three months of safety driver in a new metro to confirm that it's good, and then we take the safety driver off, that kind of thing.
The first priority when we release a major new software architecture for Autopilot is safety. It starts off with safety, obviously, safety prioritized, and then solve comfort thereafter. That's why I don't recommend people take the initial version. That's why I say, like, yeah, most people should wait until 14.2 before they actually download version 14. By 14.2, we will have addressed many of the comfort issues. The priority is very much safety first and then thereafter, the comfort issues. That's why most people are like, it'll be safe but jerky. We just need time to smooth the rough edges and solve for comfort in addition to safety with a major new autopilot architecture change.
I know what the roadmap is for the Tesla real-world AI in very granular detail. Obviously, Ashok is leading that. I mean, I spend a lot of time with the team going in excruciating detail here on what we're doing to improve the real-world AI. This car is going to feel like it is a living creature. That's with AI four before even AI five.
Yeah. As I mentioned earlier, we'll be adding reasoning to, I don't know, reasoning in 14.3, maybe 14.4, something like that.
Yeah. With reasoning, it's literally going to think about which parking spot to pick. It's going to say, this is the entrance, but actually, probably, there's not a parking spot right at the entrance. If it's a full, you know, if the parking lot is fairly full, the probability of an open parking spot right at the entrance is very low. But actually, what it'll simply do is drop you off at the entrance of the store and then go find a parking spot. It's going to get very smart about figuring out a parking spot. It's going to spot empty spots better than a human. It's got 360-degree vision, and it's going to, yeah.
Yep. Putting that all inside the computer that has AI four is the actual challenge. That's what the team is working on. Obviously, you can do reasoning on the server that takes forever. But then in the car, you need to make real-time decisions. Putting all the, you know, that's in the car, that's the challenge.
Yeah. That's why I say, like, I have a pretty good understanding of AI, you know, the giant model level with Grok and with Tesla. I'm confident in saying that Tesla has the highest intelligence density. When you look at the intelligence per gigabyte, I think Tesla AI is probably an order of magnitude better than anyone else. It doesn't have any choice because that AI has got to fit in the AI four computer. The discipline of having that level of AI intelligence density will pay great dividends when you go to something that has an order of magnitude more capability like AI five. Now you have that same intelligence density, but you've got 10 times more capability in the computer.
Well, the hardware design will not actually be frozen even through the start of production. There'll be continued iteration. A bunch of the things that you discover are very difficult to make. You only find that pretty late in the game. We'll be doing rolling changes for the Optimus design even after the start of production. I do think that the new hand is an incredible piece of engineering. We'll actually have a production intent prototype ready to show off in Q1, probably February or March. We're going to be building a million units Optimus production line, hopefully with the production start towards the end of next year.
That production ramp will take a while to get to an annualized rate of a million because it's going to move as fast as the slowest, dumbest, least lucky thing out of 10,000 unique items. But it will get to a million units. Ultimately, we'll do Optimus four. That'll be 10 million units. Optimus five, maybe 50 to 100 million units. It's really pretty nutty.
Yeah. It's just like the point is that I just need enough voting control to give a strong influence, but not so much that I can't be fired if I go insane. I think that sort of number is in the mid-twenties approximately. As a company that has already gone public, we've investigated every possible way to achieve voting control without, you know, is there some way to have a supervoting stock, but there really isn't. There is no way to have a supervoting stock after you've gone public. For example, Google, Meta, many other companies have this. But they had it before they went public. It sort of gets, I guess, grandfathered in. Tesla does not have that.
Like I said, I just don't feel comfortable building a robot army here and then being ousted because of some asinine recommendations from ISS and Glass Lewis who have no freaking clue. I mean, those guys are corporate terrorists. The problem, yeah. Let me explain, like, the core problem here is that so many of the index funds, passive funds, vote along the lines of whatever Glass Lewis and ISS recommend. They've made many terrible recommendations in the past. If those recommendations had been followed, they would have been extremely destructive to the future of the company.
But if you've got passive funds that essentially defer responsibility for the vote to Glass Lewis and ISS, then you can have extremely disastrous consequences for a publicly traded company if too much of the publicly traded company is controlled by index funds. It's de facto controlled by Glass Lewis and ISS. This is a fundamental problem for corporate governance. They're not voting along the lines that are actually good for shareholders. That's the big issue. That's what it comes down to. ISS, Glass Lewis, corporate terrorism.
如果用电池来缓冲能量,仅靠电池就能让美国的能源输出实际上翻倍,而不需要新增发电厂。
If you buffer the energy with batteries, you can effectively double the energy output in the United States just with batteries, pulling no incremental power plants.
我们确实预计,到今年年底会在约八到十个都市区运营 robotaxi。
We do expect to be operating robotaxi in, I think, about eight to ten metro areas by the end of the year.
对人形机器人来说,根本不存在现成的供应链。为了制造它,Tesla 实际上必须高度垂直整合,深入供应链进行制造,内部生产零件,因为根本没有供应链。
With a humanoid robot, there is no supply chain. In order to manufacture that, Tesla actually has to be very vertically integrated and manufacture very deep into the supply chain, manufacture the parts internally because there just is no supply chain.
从某些指标看,AI five 芯片将比 AI four 芯片好 40 倍。不是 40%,是 40 倍。
By some metrics, the AI five chip will be 40 times better than the AI four chip. Not 40%, 40 times.
此时此刻,我基本上 100% 有信心——不,不是基本上,是 100% 有信心——我们能解决无监督全自动驾驶,并且安全水平远高于人类。
At this point, I feel essentially 100% confident, I say not essentially, 100% confident that we can solve unsupervised full self-driving at a safety level much greater than human.
你可以把 Tesla 看成,怎么说呢,一家公司里的十几家初创公司。每一家都是我发起的。
You can think of Tesla as, like, I don't know, a dozen startups in one company. I've initiated every one of those startups.
只要一万次行程里有一次出问题,你就麻烦了。
All it takes is one in 10,000 trips to go wrong, and you've got an issue.
我们发布 Autopilot 重大新软件架构时,第一优先级是安全。显然,一开始就是安全优先,然后再解决舒适性。
The first priority when we release a major new software architecture for Autopilot is safety. It starts off with safety, obviously, safety prioritized, and then solve comfort thereafter.
我可以自信地说,Tesla 拥有最高的智能密度。如果看每 GB 的智能,我认为 Tesla AI 可能比任何其他人都好一个数量级。
I'm confident in saying that Tesla has the highest intelligence density. When you look at the intelligence per gigabyte, I think Tesla AI is probably an order of magnitude better than anyone else.
我们将建设一条 100 万台 Optimus 的生产线,希望明年年底前后开始生产。
We're going to be building a million units Optimus production line, hopefully with the production start towards the end of next year.
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