SemiAnalysis 分析师解读 Anthropic S-1 招股书背后的 AI 经济账

来源:The Verge:AI(RSS) 2026年10月29日 22:00 AIHOT 评分:83 精选
摘要:The Verge Decoder 节目邀请 SemiAnalysis 分析师 Joey Brookhart 解读 Anthropic 提交的 S-1 招股书,公司拟在纳斯达克上市,估值区间 1.5 万亿至 2 万亿美元。
推荐理由:SemiAnalysis 分析师拆解 Anthropic S-1,给出了推理毛利、每兆瓦收入、客户支出结构等一手财务细节。
# SemiAnalysis 分析师解读 Anthropic S-1 招股书背后的 AI 经济账 > 原标题:Anthropic is asking Wall Street to believe in AGI - 来源:The Verge:AI - 发布时间:2026-10-29T14:00:00.000Z - AIHOT:https://aihot.news/items/r9bmkfh02wkycc8zmhw87h4q9 - 原文:https://www.theverge.com/podcast/1021322/anthropic-ipo-filing-s-1-claude-semianalysis-joey-brookhart-interview ## 摘要 The Verge Decoder 节目邀请 SemiAnalysis 分析师 Joey Brookhart 解读 Anthropic 提交的 S-1 招股书,公司拟在纳斯达克上市,估值区间 1.5 万亿至 2 万亿美元。 ## 推荐理由 SemiAnalysis 分析师拆解 Anthropic S-1,给出了推理毛利、每兆瓦收入、客户支出结构等一手财务细节。 ## 正文 · 中文译文 ![A photo illustration featuring Anthropic CEO Dario Amodei.](https://platform.theverge.com/wp-content/uploads/sites/2/2026/10/VRG_DCD_anthropic_IPO.jpg?quality=90&strip=all&crop=0,0,100,100) 今天的《Decoder》请来了芯片与 AI 研究公司 SemiAnalysis 的分析师 Joey Brookhart,和我一起深挖 Anthropic IPO 文件(也就是 S-1)里那些数字背后的故事。 Anthropic 刚刚[提交了在纳斯达克上市的 S-1 文件](https://www.theverge.com/ai-artificial-intelligence/1018476/anthropic-ipo-details-s1-filed-sec),最快下个月就能开始交易,估值在 1.5 万亿到 2 万亿美元之间。一旦上市,它立刻就会成为全球市值最高的公司之一。 但价格标签并不是故事的全部。本周早些时候公布的这份 S-1 长达数百页,[细节满满](https://www.theverge.com/ai-artificial-intelligence/1018738/anthropic-ipo-s1-filing-weird-details),终于让人看清了一家前沿 AI 实验室的经济账。可以想见,这幅图景既极其复杂,又建立在一系列非常非常[大胆的假设之上——关于未来,以及 AI 将如何改变未来](https://www.theverge.com/ai-artificial-intelligence/1021476/anthropic-s-1-compute-recursive)。 所以我想请 Joey 上节目,帮我们拆解这份文件透露的信息,既关乎 Anthropic 的财务状况,也关乎更广泛的 AI 经济。Joey 和 SemiAnalysis 的同事其实早在 7 月就[从外部搭建了一个 Anthropic 财务模型](https://newsletter.semianalysis.com/p/anthropic-3q26-profit-over-1b-the),所以他一直和大家一样盼着这份文件,我也很期待听听他对我们已知信息的看法,以及更重要的——我们仍然不知道什么。 好,下面是 SemiAnalysis 的 Joey Brookhart,聊 Anthropic 的 S-1,以及打造前沿 AI 的成本与风险。我们开始。 _本次采访经过轻度编辑,以精简篇幅、提升清晰度。_ **Joey Brookhart,你是 SemiAnalysis 的分析师,也是 SemiAnalysis Tokenomics 模型的联合负责人,这个模型在今天的节目里会多次提到。欢迎来到** **_Decoder_****。** 谢谢,Nilay,非常感谢。 **我很期待和你聊。我们是在 Anthropic 那份重磅 S-1 提交几天后录的这期节目。这是我们第一次看到一家领先前沿模型公司的财务状况。你怎么看?你从中读出了什么?** 这份 S-1 非常精彩。投资者日之后的三周挺难熬的,有些人对 9 月的\[年化经常性收入\]数字有点失望。但我们在这里看到的,是一家高效扩张到惊人水平的公司,在极短时间内做到 1 万亿美元以上的估值、700 多亿美元的年化经常性收入。 毛利率、算力利润率和自由现金流,全都是你想看到的样子。所以我觉得这份文件没什么可挑刺的。这也是为什么过去两天,半导体板块、新云厂商和算力股又集体大涨。 **有意思,我觉得 S-1 本身确实挑不出什么毛病。我想深入一些细节,因为里面有很多数字值得琢磨。其实我想先从最大的图景开始。你在做一个叫 Tokenomics 的东西,我先给大家解释一下 Tokenomics 是什么。** **我看 Anthropic,看它可能走向哪里。我看到 Dario Amodei 说他能看到通往 1 万亿美元收入的路。这些钱从哪来?因为这类生意很大程度上就是吃掉其他所有生意,总得有人赚到钱,才能付给 Anthropic 来产生这些收入。我其实还没看到有人把这一点讲得很清楚。** ![](https://platform.theverge.com/wp-content/uploads/sites/2/chorus/uploads/chorus_asset/file/24792604/The_Verge_Decoder_Tileart.jpg?quality=90&strip=all&crop=0,0,100,100) _Verge_ 的订阅用户,别忘了你们可以在任何收听播客的平台独家收听无广告版的 _Decoder_。点[这里](https://www.theverge.com/account/podcasts)。还不是订阅用户?可以[在这里注册](https://www.theverge.com/subscribe)。 这非常有意思。如今,即便两大领先实验室的 ARR 合计已达 1200 亿美元,市场依然极其狭窄。消费最高的那 1% 客户,每个员工要花 5 万到 10 万美元。而更广泛的《财富》500 强或全球 2000 强企业,通常 IT 支出很大,现在却花得不多。 很多支出,尤其是在 Anthropic,都集中在编程这一块。它几乎就是一家单一垂直领域的公司。展望 2027 年,你会想到更广泛的采用和白领工作。眼下正是典型的企业第四季度预算季,但我们其实还没有真正经历过其中一家实验室的完整预算周期,因为它们的 ARR 直到今年年初才真正开始加速。我们也在考虑网络安全等其他领域,它和编程类似,都是与领先企业签下的八位数以上合同,而且由于我们看到的大量网络安全事件,这类合同能非常快地扩张。再往下还有生物等垂直领域。 要达到 1 万亿美元,我们确实必须转向一部分白领工作,从 AI 辅助走向节省人力,甚至在某些领域替代人力。你可以看看[微软向员工提出买断方案这类例子](https://www.theverge.com/tech/1018666/microsoft-expands-buyouts-60000-employees-eligible),但除此之外,并没有大量其他垂直领域能让它们真正撬动增长。 **这是我最大的疑问,不只是关于这份 S-1,而是关于整个行业。如果只有一两家领先实验室,我还能理解。有两家大公司拥有能自动化大量工作的技术。我们会看到这种自动化在各个行业推进得如何,但领先实验室并不只有两家,而是有好几家,还有开放权重模型。** **我听过很多 CEO 说,他们非常不放心把所有数据交给 OpenAI 或 Anthropic,希望有些东西能在本地运行。现在已经有律师事务所在本地跑 AI 了,这大概是整个历史上最有趣的转折。** **你怎么看这件事的走向?因为这份 S-1 承受的压力很大一部分来自数字本身,这个我想聊,但还有外部压力——“你必须增长,而增长并没有那么短暂。”你怎么看这一切会如何发展?** 这很有意思,因为比如看今天的编程领域——我想我们在第三季度看到 OpenAI 有所追赶——编程有些部分的转换成本非常低,我可以随时切换。我们从企业那里听到的,也许还有 OpenAI 那边——显然我认为他们把 Fable 的发布搞砸了,而缺乏零数据保留正是我们看到一些放缓的原因。这是他们可以用 Fable 5.5 解决的问题,也许能在那里推动更多企业采用。 这对企业来说是个大事,我们和企业交流时深有体会。我们确实看到企业,尤其是很多 AI 原生公司,正在开放权重模型上做自定义训练和后训练,然后用它来跑推理,甚至用于内部或外部用例的产品。 显然,谷歌的 DeepMind 和 Meta Superintelligence Labs 都是资金非常充裕的实验室,训练算力也相当接近,而且他们正在大量投入。更不用说 SpaceX 了,埃隆说过明年净增约 10 吉瓦,其中一半是长期的内部用途。 今年上半年,Anthropic 一路领先,占了行业净新增 ARR 的 90%。到了第三季度,OpenAI 可能实际上占了行业净新增 ARR 的大头。开源从第一季度大约 10 亿美元的 token 推理成本、ARR,增长到第三季度的 100 多亿美元。 所以没错,行业竞争激烈多了,但也出现了各种新的用例,其中一些垂直化的用例正是我们未来看到增长的地方。因为很明显,光靠编程不可能推动我们实现哪怕三倍的增长。 **这很有意思。这个概念叫 Tokenomics。你在 SemiAnalysis 和另一位联合负责人一起做这块研究。解释一下它是什么吧,因为在我看来,这是一种非常聪明的思考方式——“我们投一些钱进去,也许能拿一些钱出来。”** 我把它理解为 AI 的单位经济学,或者更宽泛地说,从宏观到微观,经营 AI 业务、乃至观察 AI 需求的方方面面。所以我们覆盖超大规模云厂商,现在也在一个单独的产品里覆盖一些新型云厂商。 但还有 AI 实验室本身。对投资者,或者只是对这门生意感兴趣的人来说,Anthropic 两年前推理业务的毛利率是负 94%,现在则是正 85%,而人们并不理解他们是怎么分配算力的。我们该怎么看待推理的商业模式? 毛利率讨论的一部分是推理的每兆瓦收入。为什么所有做推理的公司都在推理和算力效率上获得巨大提升?所以我们试着对这些业务建模,显然 Anthropic 还没上市时,我们只能依赖从媒体等渠道获得的信息。 但现在有了经过审计的具体数字,解释它的单位经济学就容易了。也能看清这些业务之间的差异。OpenAI 和 Anthropic 是非常不同的生意,无论是在消费还是订阅上,还是在历史权重、消费者与企业客户上。我们试图从需求层面理解这些业务的经济性。 **几年前我和谷歌 CEO 桑达尔·皮查伊聊推理成本时,我说:“你们到底怎么赚钱?”他非常明确。他说:“如果说谷歌有一件事擅长,那就是把这类东西的成本曲线拉下来。”** **这是个很直白的回答。我理解谷歌怎么做得到——他们自己造芯片。而 Anthropic 跑在别人的云上。S-1 的核心部分就是他们承诺在别人的云上花多少钱:SpaceXAI、谷歌、亚马逊。那他们的效率从哪来?** 关键其实在 token 吞吐量。他们的芯片组合非常多样。在 Trainium、TPU 和 GPU 之间,我们在所有这些上面都看到了大量 token 效率提升。比如 Trainium 芯片上的 token 吞吐量,现在同比增长了 5 到 10 倍。 从硬件层面的效率到软件,再到模型与硬件的交互方式,token 吞吐量有大量提升空间:如果你现在每秒处理 X 个 token,仅仅一年后就能做到 5 倍甚至 10 倍。他们也会告诉你,未来算力效率还会继续提升。 Max Kan 和我在 SemiAnalysis 共同负责 Tokenomics 模型,我们甚至就此发表过一些研究。目前,如果 Anthropic 每兆瓦推理算力能带来 6500 万美元收入,而去年这个时候还只有一千多万美元。我们说过,在 \[Nvidia\] Blackwell 上跑 Opus 5.5,每兆瓦收入可以达到 1 亿到 1.2 亿美元。 关键在于,他们现在平均每兆瓦的\[销货成本\],也就是纯算力成本,只略高于 1000 万美元。有了这样的利润率,就可以把钱再投回训练。 很多人没意识到,只有大约 30% 的算力真正用于推理服务创收。其余都用在实验和研究、训练模型以及后训练模型上。所以有大量算力是在打造下一代模型,这更偏向研发侧,也正是这项业务真正获得回报和护城河的地方。 **其实,我原本期待在 S-1 里看到这种动态表现得更加明显——Anthropic 靠卖现有模型赚了很多钱,未来这些模型的客户还会更多。这会释放出大量利润空间。然后,正因为它是 Anthropic,而且他们自认为在打造一个数字之神,他们会尽可能多地去训练下一代模型。** **这是 Anthropic 非常独特的故事。这家公司在追逐 AGI。在那里工作的人会乐于在 S-1 里告诉你,Claude 可能是有意识的,而这一切也可能搞砸。Anthropic 对自身的定位非常独特。但你在 S-1 里其实看不到这种动态充分展开。他们把训练成本与推理利润管理得相当不错。你觉得这个比例会上升或发生什么变化吗?** 我同意。长期来看,我们甚至在公司内部也说过,他们想提高训练所占的比例。如果你相信自己是在打造一个数字之神,或者正走向 AGI,只要推理业务能保持 85% 以上的毛利率,你就能这么做。 他们的增长远超自己预期,也没想到非 GAAP EBITA 利润率能达到 11%。他们巴不得这个数字接近零,甚至更低——只要能拿到那些算力,把它分配给训练,并继续看到算力效率的提升。 他们在路演中也向投资者传递了这一点。你看这份 S-1,看过去几年的盈利曲线,看他们走到今天烧掉了多少现金,就能说:“这是个不错的商业模式,但我们正走向 AGI。别指望我们看到的利润率杠杆会一直保持在这个水平,尽管长期利润率可能高得多。” **这里面有几分亚马逊的影子。杰夫·贝索斯有句名言,他告诉投资者,在能大赚之前他不会赚钱。显然,亚马逊和 Anthropic 之间有很多往来。也许这里面有点家族相似性。** **不过我很好奇,模型能力发展到某个节点,你就能切入编程之外的更多垂直领域,对吧?也许我们现在就到了这个节点。能力过剩的说法是存在的,意思是今天已有的模型可以接手生物研究,它们今天正在接手数学。到了某个节点,最明智的商业决策就是停止训练,把已经能赚钱的模型卖给越来越多的公司。Anthropic 作为一家上市公司会面临这种压力。** **你觉得他们的组织或结构能挡住这种压力吗?** 能,我觉得他们的结构不错。这类领域里有一些投票权股票,他们肯定能挡住。我们在科技行业见过这种情况,甚至公开市场上也见过,马克·扎克伯格以前就做过这种重金投入的决定。 **但并不总是有回报。我只想指出,并不总是有回报。** 对,确实并不总是有回报。 **不过按某种定义,你和我现在其实就在元宇宙里。** _\[笑\]_ 没错。但我觉得这值得一提,S-1 里有个很有意思的数字:过去三个月的收入有 80% 来自过去六个月发布的模型。 所以你会说:“好吧,20% 的模型是 Opus 4.5 和更早的版本。”往前看有一点:如果你是前沿模型公司,你就必须始终站在前沿,竞争极其激烈。 **等等,我能问个特别傻的问题吗?** 可以。 **为什么?我对这个行业的好奇,核心就在于:如果你做出了某种能颠覆全世界的尖端技术,到了某个时候,最赚钱的做法就是把它铺到各个垂直领域去,对吧?而不是去做下一个。真正该做的是在法律服务、医疗以及其他 AI 还没渗透的地方跑一个高效的模型。** **这种“你必须始终站在前沿”、Anthropic 必须始终和谷歌、OpenAI 以及其他公司竞争的说法,我从他们的自我认知角度能理解。但从商业角度看,一家用上 AI 的企业会给我这个 AI 供应商掏更多钱、花更多钱,这未必和站在前沿有关。** 这里面有几层。对于新的垂直领域,这一点在网络安全上尤其明显:如果对手用的是前沿模型,你就得用前沿模型来防御。 **有道理。** 我们和一些企业聊的时候,这不是写个邮件摘要、我随便用开源模型或 Sonnet 就能跑的事。有些垂直领域就是得用……\[比如\]如果你是被告,检方——你愿意让你的律师用非前沿模型吗? **_\[笑\]_** **关于法律市场现在的状况,我有个坏消息要告诉你。** 竞争格局确实决定了这一点。往前看,我们也见过开源在某个时点追上来,但需求会一直在。\[SemiAnalysis CEO\] Dylan \[Patel\] 经常说这一点。对前沿智能的需求永远存在,尤其是在竞争激烈的行业。而且有些任务,当你从辅助人力走向替代人力时,情况也是如此。 与此同时,它们的产品正变得越来越垂直化。如果你去看招聘信息或者他们在招什么岗位,会发现他们推出的这类 SKU 和领域越来越多,而且有大量测试正在进行,比如面向金融服务的 Claude、面向法律行业的 Claude,还有他们的网络安全项目。他们给网络安全项目的参与者每月赠送 25,000 点额度。 到 2027 年,这对他们来说会是一个巨大的 SKU。那些垂直化的领域,以及那些你必须要涉足的领域……我应该说,竞争更激烈、更具对抗性,这是主要原因。但我们都同意,像邮件摘要这类工作,确实不需要在\[前沿模型\]上做。 **我拿 iPhone 做个不太恰当的比较。现在是 iPhone 季。我刚评测了 iPhone Duo。苹果已经好几年让 iPhone 外观保持不变了。他们为新 iPhone 打造了全套工具链,然后直接复用到两三代、甚至四代机型上,把这项投资的全部价值榨干之后,才转向下一个新东西。苹果推出新 iPhone 的同时,还会继续卖上一代 iPhone 一段时间,从那次投资和那套开发计划里获取更多价值。** **AI 行业似乎并不是这么做的。他们的全部重心都放在开发下一个新东西上,旧东西很快就弃用,把你赶到新的贵产品上,与此同时又在打造下一个新的贵产品。** **到某个时候,这种感觉会变得合理起来:总会有一样东西值得他们一边躺着收利润,一边去投资新东西。我只是想不明白 Anthropic 在这个周期里处于什么位置。** 我的意思是,我们确实看到了这种情况。开源模型的 token 用量在上升,而且有很多领域,是的,开源模型已经足够好了。我认为往后看,你会问:“好吧,那是因为蒸馏吗?”然后前沿实验室其实在往前走的过程中是有一定护城河的。我觉得跟客户交流时,重点也在于我们到底怎么看待前沿模型和开源模型。所以像 Sonnet 这类模型,如果你想打个比方,就是 iPhone 18 对 iPhone SE 那种关系。 我们确实看到了这种动态,显然对开源来说,问题在于它能吃掉多少份额。第二,因为这些垂直领域还太早期,我们不知道哪些会完全智能体化。到那时候,就没有人在环了,我应该说,那套执行框架本身才真正有价值、有护城河。 这里我会重点关注的一大领域,尤其是我们今天跟企业交流时看到的,是在很多企业里,只有程序员才被允许使用前沿模型。如果你不是程序员,有些公司就只让你用 Sonnet——如果你是 Anthropic 的客户的话。这反映了管理层如何看待前沿智能在某些领域对他们到底值多少钱,以及想怎么为此做预算。 以我金融分析师的脑子来看,六个月前的 Claude 还建不好一个像样的财务模型。如果你把它接到合适的领域,它正在变好,但我们已经从“六个月前我需要一个初级分析师”,变成“也许我现在没那么需要了”,再到“如果我有无限的 token 预算,我会不会很快就能随时拥有一个甚至多个初级分析师?”我们就是在每个垂直领域和主要职业类别上这样去看的。 **这几类业务的时间跨度会各不相同。这也是我比较大的疑问之一:它们多快能走到自己认为能达到的位置?** **我想细聊一下这里的一些数字。昨天所有论坛的评论区、我们所有文章的评论区里,最扎眼的就是那 800 亿美元。Anthropic 亏损了 800 亿美元\[2026 年上半年\]。这跟一项技术性费用有关。你能解释一下这 800 亿美元吗?因为它看起来跟他们谈的收入和利润关系不大。** 对,这是一笔非现金费用,因为他们的估值涨得太多了,而且他们发行了认股权证或可转换票据。这些必须在资产负债表上体现。之所以是非现金费用,就是因为自发行这些权证和票据以来,他们的估值涨了太多。 这种项目你通常都会剔除掉。股票薪酬这类东西是可以争论的:我们真的该按非 GAAP 的现金口径来考虑它吗?因为它不是现金,如果我只是把股票发给员工,我必须在利润表上体现,但我并没有真的从银行账户里掏现金给他们。所以它不是那种情况。从会计的高层面就能理解。如果你不懂会计,看起来会很糟,但原因只是他们增长太快、估值涨得太多了。 **关于这笔费用,有一种说法是另一方是亚马逊。亚马逊显然很高兴,因为 Anthropic 在大量购买 AWS,估值上涨他们当然开心,但随着时间的推移,亚马逊会因此被稀释。这是其中的一部分动态。** **Anthropic 显然会 IPO。他们会向市场投放大量股票。很明显,随着时间推移,他们会发行更多股票来筹集更多资金。看起来他们短期内不会停止融资的意愿。** **那你怎么看待投资 Anthropic 这类公司这件事——你的估值可能在涨,但你实际持有的股份可能随着时间被稀释?** 这些交易的结构一直非常有创意。数字是,他们实际上只烧掉了 100 亿美元现金。今天资产负债表上有 1200 多亿美元现金和有价证券,IPO 后大概还会多出 1000 亿美元。 这对双方都是好交易。他们实际上只是租用算力。并不是说他们有这些巨额资本开支,甚至也不是租购。你可以就某些会计处理提出一些论点,确实有这些巨额的算力承诺义务,严格来说那是未来债务,或者说基本上是在资产负债表外。但从稀释的角度看,所有科技公司都会稀释。他们将来某个时候会不会需要再融 1000 亿美元的债或股?有可能。他们能不能继续创造性地构建和融资未来的算力建设?大概也能。 我知道过去几周大家对算力融资和数据中心融资的未来有很多焦虑。但如果你是 Anthropic,如果你是大型实验室、超大规模云厂商,或者是为其中很多交易兜底的博通或英伟达,这种情况短期内大概不会改变。所以我不特别担心。 作为股权投资者,我会担心被严重稀释吗?我不这么认为。如果你相信这个长期的 AGI 看涨逻辑,并且相信 Anthropic 会成为其中的领导者,那这大概不该让你担心。你面对的很可能是一家市值 10 万亿美元以上的企业。到那个时候,他们跑赢稀释完全没问题。 **我们来聊聊算力承诺。我看到至少有一份分析说,未来 10 年,他们对各家超大规模云厂商和新型云厂商的承诺总额超过 8000 亿美元。这可不是小数目。你说他们并没有那么多钱,甚至还没花掉那么多钱。你觉得他们能创造出足够的收入来兑现这些承诺吗?** 能。我觉得人们看到那些头条上的承诺数字——7 到 10 年、5000 亿到 6000 亿美元——其实他们今天为算力付的钱,对应的吉瓦数并没有那么多。看他们现在的支出,平均大约是每兆瓦 1100 万美元,也就是每吉瓦 110 亿美元。 把这个换算到年度层面,比如说那 5200 亿美元是分摊在 7 到 10 年里的,那对应的吉瓦数其实并不多,而且他们会远远超出……等他们做到 20 吉瓦左右,他们未来的算力承诺基本上就会轻松达到数万亿美元级别。 所以是的,他们会有足够的收入来支撑。关键更多在于他们能有多少收入,相对于他们能有多少算力。第二是算力效率这个等式。如果他们每兆瓦推理算力能持续带来多得多的收入,那你到底想在训练上投多少,又是否愿意从租赁的角度把利润率压到零?因为你可以走到这样一个状态:即便他们 80% 到 85% 的算力都用在训练和研究上,在这样的利润率下也撑得住。 收入会增长得相当不错。他们在这方面没有任何危急处境。除非我们真的看到这些收入全部消失,或者用户大量流失。并不是说他们必须极快、极快地做到一万亿美元收入,才能实现这一切并为此买单。 **除非你想造的是数字上帝,除非它不会像所有人希望的那样快速扩散到整个经济中。这些都是赌注,对吧?** 是巨大的赌注。你看那些数字,如果你真的看好这波向数字经济的攀升,Anthropic 说在他们的《财富》500 强客户中,他们已经覆盖了 50%,而这仅仅占他们客户总运营支出的 0.1%。按每用户来算,如果你认为运营支出不包含销货成本,那大约是每月 8 美元;如果你认为包含销货成本,那也不到每月 50 美元。 他们也是这么说的。在 Claude Code 用户中排名前 5% 的用户,平均每年花费 1.8 万美元——是前 5% 的用户。而非 Claude Code 用户每月花费不到 75 美元,远低于 75 美元。这里存在巨大的机会,尤其是当我们和企业以及这些传统 IT 软件采购方交流时。 在传统白领工作中,渗透率极低。这也是关键的一点:无论是席位采用还是使用量,这里都有一个很长的爬坡期。你可以争论谁会成为赢家,但很明显,前沿实验室会从中分到一部分,而且这对明年的净新增 ARR 会相当重要。 **其中很有意思的一点是,我们聊到前沿实验室、聊到超大规模云厂商——Anthropic 在这份 S-1 里列了一大堆离谱的风险因素,包括 Claude 可能已经有意识、会杀光我们所有人。** **但其中最直接、也许最容易理解的一条是:AWS、Azure 和 Google Cloud 在某种意义上既是供应商又是竞争对手,Google 尤其如此。他们说,“这些关系可能会变。”他们和 SpaceXAI 关系很深。埃隆·马斯克要去训练 Grok。他们迟早会和 SpaceXAI 正面竞争,那段关系可能会变得岌岌可危。** **就像你说的,他们只是在租算力。到了某个节点,市场可能会逼他们自己拥有更多算力,因为他们的竞争对手不愿意把产能租给这家领先公司。你怎么看这件事的走向?** 我们确实看到 OpenAI 和 Anthropic 都在往更直接的方向走,但不是自己拥有算力,而更像是“先租后买”,或者像一些财务赞助方那样通过 SPV 来做,以更直接的方式产出算力,而不是租用。 这有两点。从某个角度看,Google 也在把自家芯片卖给外部客户,等于在说:“我们不在乎一定要在 Google Cloud 内部变现,甚至不在乎份额分配……就算你觉得我们芯片产能很紧张,我们也不在乎这些芯片是给 DeepMind 还是给内部 GCP 算力去变现。我们就是愿意当芯片卖家。” Anthropic 很喜欢 TPU,订单规模有 3.5 到 4 吉瓦以上,后面可能还有更多在排队——Anthropic 直接买这些芯片,然后再租出去。所以是的,市场是好的。 **等等,不是直接买来出租。那他们是买进一家 neocloud?** 对,他们通过 SPV 买,或者由 SPV 买下,然后他们再从 SPV 租。 **明白了。好。** 即便在 Google Cloud、AWS 或 Foundry 内部,Anthropic 的体量也极大——他们 48% 的业务来自 Bedrock Foundry 和 Vertex,客户通过已有的云关系付费访问这些模型,甚至根本不用联系 Anthropic,不需要和 Anthropic 的销售团队打交道。 **你前面提到过这一点。这个营销渠道,Anthropic 能拿到收入分成。** 没错。如果你是 AWS,你会说:“我在为 Anthropic 跑算力,假设每兆瓦 1000 万美元,但现在我能拿到每兆瓦 650 亿美元中的 16%,外加算力成本。”所以现在各家云的利润率都在上升。一部分原因是编程带来的 CPU 利用率上升,还有强化学习那边的一些因素;但我们所说的“token 即服务”业务——也就是 AWS Bedrock、Microsoft Foundry、Google Gemini Agent Enterprise 平台(原 Vertex)——实际上在大幅抬高这些云厂商 AI 业务的利润率。 股东最近把功劳记在他们头上,他们也因此受益。第二季度,Azure 和 AWS 的业绩都极其、极其强劲,尤其是 AWS,靠的就是 Anthropic 在 Bedrock 上的增长。 **你觉得这会让亚马逊永远不想训练自己的前沿模型吗?他们让 AWS 成为其他所有人的基础设施,这件事做得很好,但他们也是少数几个没有因为缺少自家模型而感到不安的大型科技公司之一。** 是的,这很有意思。亚马逊很乐意服务各种各样的客户,而他们的\[前沿能力\]非常有限。看我们对训练算力的估算,他们在所有超大规模云厂商里是最少的,也许只有 Oracle 比他们更少。看起来那边没什么战略可言。 我知道微软在过去两个季度里谈得稍微多了一些。AWS 在财报电话会上也谈得多了一点,但规模并不大。他们投入的程度比不上 Meta,也没有要在这里争一争的意思,甚至比不上 Google 的 Gemini。尤其是像 Dylan 这样的人会告诉你,如果你不打算在前沿竞争,那真的没有你的位置,特别是你内部只投入几百兆瓦的时候。一来,内部用例是支撑亚马逊零售业务,二来才是训练。想做好这件事,需要做出很多承诺。 **这是另一个话题了,但我看到这些** [**微软裁员**](https://www.theverge.com/tech/1018666/microsoft-expands-buyouts-60000-employees-eligible) **,觉得这算是一丝迹象,说明他们未来会试着参与竞争。他们显然被 OpenAI 坑得挺惨。** 是的。在他们新签的协议里,他们有权将 OpenAI 模型变现,而按协议基本不用付给 OpenAI 任何费用。接下来会有一些有意思的事:他们把 OpenAI 模型变现,或者通过 Microsoft 365 Cowork 和 Copilot 应用变现,而不用为前沿模型向 OpenAI 付钱。 **这里的另一个动态,你之前也提到过,就是过去几周里,人们对 neocloud 有很多担忧——它们会成功还是会失败。我记得 Elon Musk 说过会出现一场** [**类似孔雀鱼的灭绝事件**](https://fortune.com/2026/10/26/elon-musk-ai-boom-debt-neoclouds/)**,用这个比喻来形容一堆数据中心,实在太狠了。** **但这背后是一堆循环融资,英伟达扶持了其中不少,也许超大规模云厂商会直接把这门生意抢走。也许 Anthropic 的需求会这样,因为他们可能想拥有更多自己的算力,毕竟租算力在各种方面都不牢靠。** **你怎么看 neocloud?随着这些公司上市、市场真正开始洗牌,你觉得它们能活下来吗?** Elon 显然有动机说 neocloud 会灭绝。ClusterMAX 的说法不一样,而且[对 neocloud 的评级做得非常出色](https://newsletter.semianalysis.com/p/clustermax-30-the-industry-standard)。有很多 neocloud 是前加密货币矿工转型的,它们不是好生意,也不擅长运营集群。 **我想指出,CoreWeave 就是一家前加密货币挖矿公司,对吧?** 是的。我不是说 CoreWeave,而是名单上更靠后的其他一些公司。CoreWeave、Nebius,以及任何有主要实验室、超大规模云厂商、英伟达或博通支持或兜底的公司,都会没事。 关键在于这条更长的尾部。超大规模厂商占了资本开支和净新增算力的65%,而市场上还有35%的玩家,它们没有超大规模厂商那样的资产负债表。谷歌和Meta最近都进行了融资,或者宣布了融资计划。再看四大超大规模厂商,未来两年它们其实并不需要再融这笔钱,尽管资本开支正从1.2万亿美元快速攀升到2万亿美元。 但对很多其他新型云厂商来说,如果没有这种级别的支持,我认为即便在当下也非常艰难;而如果没有这些玩家的支持或兜底来持续建设算力,情况会变得极其困难。这是一个非常非常难做的融资市场。 **我觉得特别有意思的一点是,埃隆尤其爱说新型云厂商会死掉,同时显然在为自己的数据中心业务造势,而他还想把数据中心搬到太空,以避开美国国内围绕数据中心的强烈反对。** **我们做了大量关于美国数据中心遭遇反对的报道,这正在拖慢很多项目。但照目前的速度发展下去,如果你觉得新型云厂商摇摇欲坠,那最终可能真的会有一批GPU和建到一半的数据中心被弃置,或者被锁在融资协议里,根本无法点亮。** **你怎么看?你觉得我们真的只能上太空吗?还是认为市场会回归理性,真正利用起目前已经分配但尚未使用的算力?** 我认为更多会走向理性。市场很擅长在地面上为芯片找到去处。政治上,下周中期选举会怎样、会带来什么连锁反应,我们拭目以待。显然,还有一些超大型站点仍然面临严重的审批问题。展望明年,整个电力问题也摆在那里。 数据中心会建起来的,总会有办法把这些建起来。最重要的一点——超大规模厂商的CFO们也说过——从过度建设的角度看,相对而言,建数据中心的壳子相当便宜。如果建多了,他们可以这么做:"我们就是不接收芯片,不买芯片,也不把它们装满。"这些数据中心就空着,里面没有芯片。但大部分支出其实在芯片这一侧。 所以,如果我们最终真的走到数据中心供过于求的地步,我知道这和问题问的有点不一样,恰恰相反。他们从投资回报的角度说过这一点。这又回到那个更大的问题:我们是否仍将面临严重的供给约束。现在很多人,尤其是最近三周,都在讨论:"展望2028年,我们是不是实际上正处在一场巨大的算力过剩之中?" 因为我们谈过的同一件事——Anthropic从负94%的利润率转为正85%,以及算力效率的提升——意味着随着你上线更多推理兆瓦,每兆瓦带来的收入大幅增加,在其他条件不变的情况下,你需要的算力反而更少。所以问题在于,到2028年我们能否有足够的收入,来证明2万亿美元以上资本开支的投资回报是合理的。 **这就是为什么我花了那么多时间追问能力过剩的问题,以及今天的模型到底有多强。因为你能看到一些对 AGI 没那么理想主义的竞争对手说:“其实,我们要在这些尚未充分利用 AI 的垂直行业里派驻一大批前置部署工程师,向他们出售高利润率的推理服务,因为我们会把他们的很大一块业务自动化,他们能看到回报。”而现在,一大批财富 500 强企业[**还没有看到 AI 投入的回报**](https://www.apollo.com/wealth/insights-news/insights/daily-spark/no-signs-of-ai-payoff-outside-the-tech-sector)**。**** **所以,如果你在客户服务、部署和模型产品化上确实做得更好,你可能会远远领先于 Anthropic 需要达到的位置,因为他们每分每秒都还在想着 AGI。我看着 Dario 在 Dreamforce 上站在 Marc Benioff 旁边,那哥们儿现在根本不想待在那儿。他想的是机器之神。** **这里有个现象:你能看到竞争对手推出开源模型,然后说:“我们真的要把这个产品化,我们要拿走其他人所依赖的那份收入。”** 是的,这非常、非常有意思。在过渡期,这就是为什么行业越来越垂直化——你需要一些东西来桥接 AGI 完全胜任工作的那一天,你需要一些东西来桥接今天的实际模型。 **_\[笑\]_** **是啊。这个我们就当理所当然吧。** 要达到那一步,你需要在某个节点上搭建桥梁,让模型与真实终端用户对接。一切最终都归结为这个问题:会不会出现某种时机或采用上的错配,导致我们在某个时间点拿不到足够的收入,来支撑“为 AGI 训练和算力投入”的追逐?这是一个风险。 我们关注的另一个风险是,前沿模型——甚至开源模型——在编程能力上越来越有竞争力,前沿的 token 定价需要被压下来,因为正如我之前提到的,切换成本更低了。即便在 SemiAnalysis,我们第二季度末 95% 或 90% 以上的用量还在 Anthropic 的模型上,OpenAI 的 ChatGPT 5.6 和 Astra 发布后降到了 70%。然后随着 Opus 5.5 推出,又回到了 90% 以上。 它确实会回来。即便是今天的收入基础——他们说净收入留存率有 900%,所以他们看到的是巨大、巨大的增长。这不是 ServiceNow 或 Workday 那种传统软件生意:你有一个了不起的客户 logo,毛留存率 95% 到 98%,客户签下漂亮的三年合同,从不流失,增加一些席位,也许再被追加销售采用另一个产品。这纯粹是消费型生意。 另一部分是,由于假期季节,去年最后两周的 token 支出比前一年同期下降了 20%。假期里你不会写工作代码,不会为工作写代码。这是极其、极其纯粹的消费型生意,甚至和我们今天在公开市场上对 Snowflake 或 Datadog 的理解都不一样。 **显然,Anthropic 并不是唯一一家要上市的公司,OpenAI 就在它旁边,只是把上市推迟了。你之前提到 Anthropic 更偏向企业级市场。我觉得 OpenAI 很长时间里都以为谷歌不会反击,自己能直接把搜索业务抢过来,一切都会顺风顺水。但谷歌显然出手了,不管结果好坏。现在搜索已经变成了 AI Overviews。** **看完 Anthropic 的 S-1,你对 OpenAI 有什么期待?毕竟它更偏向消费者业务。** 一季度末时 OpenAI 的营收结构是 60% 消费者、40% 企业,直到今年 1 月或 8 月才首次翻转,变成以企业为主。他们会告诉你,企业业务此前一直没起来,是因为公司太偏消费者和订阅,而这块消费型业务占了 Anthropic 收入的 80%,直到 5.6 和 Astra 发布后才真正起飞。 这些模型确实很出色。OpenAI 拿下了第三季度实际新增 ARR 的大头,但这还需要时间,因为《华尔街日报》报道的是 EBITA 利润率为负 100%,而我们这边看到的是正 11%。 两者的商业模式存在一些差异,而且差异不小:OpenAI 的变现效率就是不如 Anthropic。部分原因在于业务构成——垂直领域不同,消费者与企业的比例不同——而这一点正在改变。 他们会告诉你,第四季度表现会很好,而且他们的企业销售能力比 Anthropic 强得多。我预计他们会追上来,但即便他们在 2027 年初发布 S-1,财务数据恐怕也不会像这份这么好看。 **你觉得这就是他们推迟上市的原因吗?为了让财务数据更好看?** 是的,我认为 OpenAI 的[第三季度非常强劲](https://www.wsj.com/tech/ai/the-ai-price-war-is-heating-upand-openai-is-gaining-ground-on-anthropic-48b8525b)。企业业务环比增长 100%,而整体业务增长 77%。如果再来一个企业业务强劲的季度,而 Opus 5.5 又没有抢走太多份额,那 OpenAI 讲起故事来会从容得多。 **从现在到 Anthropic 真正上市之间还有几个节点。他们会做路演,还会有一些定价环节。你期待从 Anthropic 那里看到什么?** 两三周前他们办了一场投资者日,透露了一些更多细节。我想他们会解释第三季度为什么这么弱。Meta 曾是非常非常大的客户,占比可能超过 10%。公开报道显示,他们的 Claude Code 席位从 6 万个降到了 3 万个。这些 token 用量并没有消失,只是转移到了 Meta 的内部模型上。 Anthropic 开始打击反蒸馏行为,这部分此前占收入的 5%。我记得他们在 S-1 里提到,违反服务条款的相关收入占 5% 到 7%。所以路演上要讲的重点,是 Opus 5.5 发布后 10 月份的这波加速。我们相信他们已经重回正轨。 如果他们在第二季度每月新增 100 亿美元的净 ARR,到第三季度掉到 40 亿美元,那么第四季度他们正回到每月 80 亿到 100 亿美元的节奏。人们想听这些新垂直领域的消息,尤其是网络安全。在完整的网络安全 SKU 推出之前,他们先做了一些补丁式的工作,但今天他们确实已经在通过漏洞代码扫描把网络安全变现了。我得说,这些与现有客户签的合同可不小。 我认为他们会开始对外释放 2027 年 ARR 增长路径的信号。另外还有一点是预期管理,把数字压得足够低,少谈一些 AGI。 **_\[笑\]_** **可当你的风险因素写着“我们可能对全人类构成生存性风险”时,预期管理是很难做的。** _\[笑\]_ 对。所以一边说你们在做这件事——你们终将在某个时点实现 AGI,你们会取代所有人的工作——一边又要把第四季度的数字压得足够低,好让上市第一个季度交出大幅超预期的成绩,并给 2027 年一个像样的指引,这就会是在长期愿景和短期预期管理之间来回摇摆的微妙组合。 **事情确实很多。Joey,这次聊得很愉快。等这些公司一路冲向真正上市、它们的经济状况变得更透明时,我们得再请你回来。非常感谢你做客** **_Decoder_****。** 非常感谢,我很感激。 _有问题或意见?发邮件到 decoder@theverge.com 找我们。我们真的会读每一封邮件!_ ## 正文 · 原文 ![A photo illustration featuring Anthropic CEO Dario Amodei.](https://platform.theverge.com/wp-content/uploads/sites/2/2026/10/VRG_DCD_anthropic_IPO.jpg?quality=90&strip=all&crop=0,0,100,100) Today on _Decoder_, I’m joined by Joey Brookhart, an analyst at the chip and AI research firm SemiAnalysis, to help me dig into the story inside the numbers of the Anthropic IPO filing — a document known as an S-1. Anthropic just [filed its S-1 to go public](https://www.theverge.com/ai-artificial-intelligence/1018476/anthropic-ipo-details-s1-filed-sec) on the Nasdaq, and it could start trading as soon as next month at a valuation of anywhere between $1.5 trillion and $2 trillion. That would immediately make it one of the most valuable companies on the planet. But the price tag really isn’t the whole story here. The S-1 document published earlier this week is hundreds upon hundreds of pages [filled with details](https://www.theverge.com/ai-artificial-intelligence/1018738/anthropic-ipo-s1-filing-weird-details) that finally paint a picture of the economics of a frontier AI lab. As you might imagine, that picture is both really complicated and built on a series of very, very [ambitious assumptions about the future and what AI will do to it](https://www.theverge.com/ai-artificial-intelligence/1021476/anthropic-s-1-compute-recursive). So I wanted to have Joey on the show to help break down what this document tells us, both about Anthropic’s finances and the broader AI economy. Joey and his colleagues at SemiAnalysis actually [built a model of Anthropic’s finances from the outside](https://newsletter.semianalysis.com/p/anthropic-3q26-profit-over-1b-the) back in July, so he’s been waiting for this filing about as eagerly as anyone, and I was excited to hear his take on what we’ve learned and, importantly, what we still don’t know. Okay: SemiAnalysis’ Joey Brookhart on the Anthropic S-1 and the costs and risks of building frontier AI. Here we go. _This interview has been lightly edited for length and clarity._ **Joey Brookhart, you are an analyst at SemiAnalysis. You’re also co-lead on SemiAnalysis’ Tokenomics model, which is going to come up a lot in this episode. Welcome to** **_Decoder_****.** Yeah, thanks so much, Nilay. **I am very excited to talk to you. We’re talking just a couple days after the big Anthropic S-1 was filed. It’s our first look inside of the finances of one of the leading frontier models. What’s your take? What’s your read into it?** It was an awesome S-1. It’s been a rough three weeks since the investor day. Some people were somewhat disappointed in the September \[annualized recurring revenue\] numbers, but what we saw here was a company that has efficiently scaled to an amazing level, with a $1 trillion-plus valuation, $70 billion-plus of ARR in a very short amount of time. In gross margins, compute margins, and free cash flow, everything is what you’d want to see. So I don’t think there was much to nitpick in this filing. It’s the reason you’ve seen a lot of the semiconductor complex, and the neoclouds and compute stocks, really rally again in the last two days. **It’s funny, I don’t think I can nitpick the S-1 itself. I want to get into some of the details because there are a lot of numbers in there that are worth thinking about. I actually just want to start at the really big picture. You run a thing called Tokenomics. I’m going to explain what Tokenomics is to people.** **I look at Anthropic and where it could go. I look at Dario Amodei saying that he can see his way to $1 trillion in revenue. Where does that money come from? Because so much of these businesses are about eating every other business and it feels like someone else has to make some money to pay Anthropic to generate these revenues. I haven’t actually seen that laid out very clearly.** ![](https://platform.theverge.com/wp-content/uploads/sites/2/chorus/uploads/chorus_asset/file/24792604/The_Verge_Decoder_Tileart.jpg?quality=90&strip=all&crop=0,0,100,100) _Verge_ subscribers, don’t forget you get exclusive access to ad-free _Decoder_ wherever you get your podcasts. Head [here](https://www.theverge.com/account/podcasts). Not a subscriber? You can [sign up here](https://www.theverge.com/subscribe). It’s so interesting. Today, because even the scale we’re at is $120 billion of ARR between the two leading labs, it’s still an extremely narrow market. The top 1 percent of customers are the ones that are spending $50,000 to $100,000 per employee. Your broader Fortune 500 or Global 2000, which typically do a lot of IT spend, isn’t spending much. A lot of the spending, especially at Anthropic, is coding-specific. It’s almost a single vertical company. We look in the future to 2027, you think to this broader adoption and white collar work. This typical enterprise Q4 budgeting season, we actually haven’t gone through with one of these labs yet in any meaningful way because their ARR did really start to accelerate until the beginning of the year. We think of other cases like cyber, which can be similar to coding, where these are eight-figure-plus contracts with leading enterprises that can scale very quickly due to all the cybersecurity incidents we’re seeing. You get to other verticals like bio. To get to $1 trillion, we do have to move to some of that white collar work, where it goes from AI augmentation to labor saving or replacement in some areas. You look at things like [Microsoft’s buyout offers to employees as an example of this](https://www.theverge.com/tech/1018666/microsoft-expands-buyouts-60000-employees-eligible), but there’s no wealth of other verticals that they can get into and really start to move the needle. **This is the single biggest question I have, not just about the S-1, but the entire industry. I would understand that if there are one or two leading labs. There are two big companies that have a technology that can automate a whole bunch of things. We will see how well that automation goes sector by sector, but there are not just two leading labs. There are quite a few of them. There are open weight models.** **I’ve heard from a lot of CEOs who are very nervous about giving all their data to an OpenAI or Anthropic and they would like to run some things on-prem. There are law firms running on-prem AI now, which is the funniest turn in maybe all of history.** **How do you see that playing out? Because so much of the pressure on this S-1 was the numbers themselves, which I want to talk about, but then there’s the external pressure of, “You’ve got to grow and growth is not as short.” So how do you see all of that playing out?** It’s very interesting because when we look at coding today, for example — and I think we saw this in Q3 with OpenAI kind of catching up — there are areas of coding where there are very low switching costs, and I can just move between that. What we hear from enterprises and maybe, too, on the OpenAI side — obviously I think they somewhat botched the Fable release, and the lack of zero data retention is why we saw some of the slowdown. That’s something they can fix with Fable 5.5 and maybe drive more enterprise adoption there. That’s a huge thing for enterprises when we talk with them. We definitely see enterprises now, and more so, too, a lot of the AI natives, doing custom training and post-training on open weight models, and then running either their inference off of that or even their products for internal or external use cases. Obviously, DeepMind at Google and \[Meta Superintelligence Labs\] are very well-funded labs that have pretty similar training compute capacity, and they’re bringing on a ton of it. Not even to mention SpaceX as well, with what Elon said on \[about\] 10 net new gigawatts next year and half of that is the long-term internal use case. In the first half of the year, Anthropic ran away with it. They were 90 percent of the industry’s net new ARR. We saw in Q3 that OpenAI was probably actually the majority of industry net new ARR. Open-source has grown from probably $1 billion of token inference costs, ARR, in Q1 to $10 billion-plus here in Q3. So yes, the industry’s gotten much more competitive, but there are all these new use cases, too, and some of these verticalized use cases are really where we’re going to see the growth going forward. Because it’s obvious coding isn’t going to power us to even triple from here. **It’s interesting. The concept is called Tokenomics. You run that practice alongside another co-lead at SemiAnalysis. Explain what it is because to me it’s such a smart way of thinking about, “We’re going to put some money in and maybe we’ll get some money out.”** I think of it as the unit economics \[of AI\], or even just the broader idea — everything from a macro- to microeconomics of running an AI business or even looking at AI demand. So we cover the hyperscalers, now some of the neoclouds as well in a separate product. But then just the AI labs, too. For investors, or just people that are interested in the business, Anthropic had negative 94 percent gross margins on inference two years ago. Now they have a positive 85 percent margin, and people don’t understand how they allocate compute. How do we think about the inference business model? Part of that gross margin conversation is the revenue per megawatt in inference. Why are all companies that are running inference seeing massive gains in inference and compute efficiency? So we try to model out these businesses and obviously when Anthropic was private, we’re dealing with information we can get from the press and things like that. But now that we have concrete audited numbers, it’s easy to explain the unit economics of this. And also how the businesses differ. OpenAI is a much different business than Anthropic just in both their consumption versus subscription and then historical weight versus consumer versus enterprise. \[We are\] trying to understand across the demand level the economics of these businesses. **Several years ago when I talked to Google CEO Sundar Pichai about the cost of inference, I said, “How are you ever going to make money?” He was just very clear. He said, “If there’s one thing Google’s good at, it’s bringing this stuff down the cost curve.”** **It was a very blunt answer. I understand how Google can do it. They make their own chips. Anthropic is running on everyone else’s clouds. The core part of the S-1 is how much they’ve committed to spend on other people’s clouds: SpaceXAI, Google, Amazon. Where does their efficiency come from?** There’s really token throughput. They have a very diverse chip fleet. Between Trainiums, TPUs, and GPUs, across all of those, we see a ton of token efficiencies. Token throughput, for example, on Trainium chips is up 5x to 10x year-over-year right now. Dealing with both the hardware level efficiencies and the software — and how the model interacts with the hardware — there are a lot of token throughput gains where if you’re doing X tokens per second, now you’re doing five or 10 times that just a year later. What they’ll tell you is they see continued compute efficiencies as well going forward. Max Kan, who co-leads Tokenomics model with me at SemiAnalysis, and I have even published some research on this. Right now, if Anthropic’s at $65 million of revenue per inference megawatt, this was mid-teens last year at this time. What we’ve said on maybe Opus 5.5 on a \[Nvidia\] Blackwell, you can get to $100 to $120 million of revenue per megawatt. The big thing here is they’re paying on average now just a little over $10 million per megawatt in \[cost of goods sold\], or just in compute costs. What you can do with those margins is then reinvest in training. People don’t realize this, but only about 30 percent of the compute is actually in inference-serving revenue. The rest is for experimentation and research, training models, and post-training models. So there’s a lot of compute that is building the next model and that’s more of the R&D side and where you actually get your returns and defensibility in this business. **Actually, what I was expecting to see in the S-1 was that dynamic play out much more strongly — that Anthropic makes a bunch of money selling the models they have now and there are more customers in the future for the models they have now. That’s going to throw off a whole bunch of margin. And then, because it is Anthropic and they think they’re building a digital god, they’re going to train the next models as much as they can.** **That is a very particular Anthropic story. The company is chasing AGI. The people who work there will happily tell you that in the S-1 that Claude might be conscious and this could all go to hell. Anthropic is a very unique company in its conception of itself. But you don’t really see that dynamic playing out in the S-1. They’re managing their training costs against their inference margin pretty well. Do you see that going up or changing in any way?** I agree. Long term, and we’ve even said this internally, they want to actually increase the percentage of training. If you believe you’re building a digital god or going towards AGI, and as long as you have the gross margins of 85 percent-plus on inference to do so, you can. They’ve grown a lot faster than they thought, and they didn’t expect to be at 11 percent non-gap EBITA margins. They’d love to be closer to zero, or even below that if they could get that compute and allocate it towards training and continue to see these compute efficiencies. They’ve messaged that, too, to investors in the roadshow. You can look at this S-1 and just the curve of profitability over the last few years and how little cash they burned to get to this point and you can say, “This is a good business model, but we’re going towards AGI. Don’t expect the margin leverage we’re seeing to continue to this level, even though long-term margins can be quite a bit higher.” **There are some real shades of Amazon there. Jeff Bezos famously told investors he wasn’t going to make money until he could make a lot of money. Obviously, Amazon and Anthropic have a lot of back and forth. Maybe there’s some family resemblance there.** **But I’m curious, there’s a point of model capability that would allow you to go into more verticals than just coding, right? We might be at it now. The idea of capability overhang exists, the idea that the models we have today could go and take over bio research. They’re taking over mathematics today. There’s a point at which the smartest business decision is to stop training and to just sell your profitable model into more and more companies. Anthropic as a public company will face that pressure.** **Do you think they’re organized or structured to fend it off?** Yeah, I think they’re structured well. With some of the voting stock in areas like that, they definitely are. We’ve seen this in tech before and even publicly when Mark Zuckerberg has made these decisions before to invest heavily. **It didn’t always pay off. I just want to point out, it didn’t always pay off.** Yeah, it did not always pay off. **Although you and I technically are in the metaverse right now, by some definition.** _\[Laughs\]_ Right. But I think that’s something, and there’s an interesting number in the S-1 that 80 percent of revenue in the last three months comes from models released in the last six months. So then you say, “Okay, 20 percent of the models are like Opus 4.5 in earlier versions.” One thing going forward is that if you’re a frontier model company, you always need to be at the frontier and it’s incredibly competitive. **Wait, can I ask a really dumb question about that?** Yeah. **Why? I think the heart of my curiosity about the industry right now is if you enabled some bleeding-edge technology that could revolutionize the entire world, at some point, the most profitable thing to do is to just go play it out in all the different verticals, right? It’s not to go make the next one. It’s actually to run an efficient model in legal services and medicine and wherever else that there hasn’t been diffusion of AI yet.** **This idea that you always have to be at the frontier, that Anthropic always has to be competing with Google and OpenAI and whoever else, I get it from their self-conception. But from a business point of view, an AI-enabled business is going to throw out more money and spend more money with me, the AI provider, and that isn’t necessarily related to being at the frontier.** There are a few portions of that. For new verticals, and this is very cyber specific, if an adversary is using a frontier model, you need to be defending with a frontier model. **Fair enough.** When we talk to some enterprises, it’s not like an email summary where I can run that on open source or Sonnet. There are just certain verticals where having… \[For example,\] if you were in a legal case as a defendant, would you be okay if your lawyer was using a non-frontier model if the prosecution— **_\[Laughs\]_** **I got bad news for you about what’s going on in the legal market right now.** The competitive dynamics do dictate that. I think from a go-forward perspective, we’ve seen this with open source too catching up in one point, but the demand will be there. \[SemiAnalysis CEO\] Dylan \[Patel\] says this a lot. The demand will always be there for frontier intelligence, especially in competitive industries. And there are certain tasks, as you go towards more labor replacement versus labor augmentation, where that also is the case. At the same time, they are getting more vertical. If you go look at job postings or what they’re hiring for, they’re putting out more of these SKUs and areas, and there’s a lot of beta testing going on like Claude for financial services, Claude for legal, their cyber program. They’re giving away 25,000 of credits per month for people in their cyber program. That’ll be a huge SKU for them in 2027. Those verticalized areas and then ones that are specifically where you need to be… It’s more adversarial, I should say, or competitive, are the main reasons why. But we can both agree there are areas like email summary and work that definitely doesn’t need to happen \[at the frontier\]. **I’ll just make a dumb comparison to the iPhone. It’s iPhone season. I just reviewed the iPhone Duo. Apple kept the iPhone looking the same for several years now. They build all of the tooling for the new iPhone and then they just reuse it for two or three or four models to get all the value out of that investment back before they move on to the next thing. Apple puts out a new iPhone, and it keeps selling the old iPhone for a while to get more value out of that investment and that development program.** **The AI industry does not seem to be doing that. All of their emphasis is on developing the next new thing, deprecating the old thing pretty fast, and moving you onto the new expensive product, while at the same time they’re building the next new expensive product.** **At some point it feels like that will rationalize, that there will be something that is worth just collecting margin from while they go invest in the new thing. I just can’t figure out where Anthropic sits in that cycle.** I mean, we do see this. Open source token volumes are picking up and there’s a lot of areas, yes, where the open source models have become good enough. I think on a go-forward basis, you’re like, “Okay, was that because of distillation?” And then the frontier labs have some defensibility actually going forward. I think when we talk to clients, it’s also about how we actually think about frontier versus open source. So something like a Sonnet type model, if you want to do the iPhone 18 versus the SE type analogy. We do see that dynamic, and obviously with open source, it’s a question of how much that can take. Two, because we’re so early in these verticals, we don’t know what’s going to be fully agentic. There’s no human in the loop where the harness actually has, I should say, some value and defensibility to it. The big area I’d look at here, especially when we talk to enterprises today, is only the coders are allowed to use the frontier models at a lot of these enterprises. If you’re a non-coder, some of these companies just limit you to Sonnet, if you’re an Anthropic customer. That’s reflective of how management views and wants to budget for this, what frontier intelligence is actually worth to them in certain areas. In my financial analyst brain, Claude could not build a good financial model six months ago. It’s getting better if you connect it to the right areas, but we’ve gone from something like, “I needed a junior analyst six months ago,” to, “Maybe I don’t as much anymore,” and then, “Will I have a full junior analyst at my disposal anytime or multiple pretty quickly, if I have unlimited token budget?” We just look at that across every vertical and major job category. **The time horizon in each of those categories is going to be different. That’s just one of the bigger questions I have: How quickly do they get to where they think they can go?** **I want to dig in on some of the numbers here. The one that jumped out in all of the comment sections on all of the forums and all of our articles yesterday was $80 billion. Anthropic lost $80 billion \[in the first half of 2026\]. That is related to a technical charge. Can you just explain the $80 billion? Because it doesn’t seem particularly relevant to the revenue and profits they’re talking about.** Yeah, it’s a non-cash charge because their valuation has gone up so much and they’ve issued warrants or convertible notes. You have to account for that in your balance sheet. It’s just a non-cash charge because their valuation has gone up so much since they issued these warrants and notes. That’s something you would always back out. You can argue on stuff like stock-based compensation. Should we really be thinking about that on a non-GAAP cash basis? Because it’s non-cash, if I just hand employees shares, I have to account for that on my income statement, but I’m not actually handing them pure cash out of my bank balance. So it’s not an area like that. It gets the accounting, the high level. If you don’t know accounting, it looks bad, but it’s because they’ve grown so much and their valuation has gone up so much. **Some of the argument here about that charge is that Amazon was the other party there. Obviously Amazon’s very happy because Anthropic is buying a ton of AWS and sure they’re happy the valuation has gone up, but over time Amazon’s going to get diluted this way. That’s part of the dynamic there.** **Anthropic is going to obviously IPO. They’re going to put a bunch of shares in the market. It’s very clear that over time they will issue more shares to raise more money. It doesn’t seem like they’re going to stop wanting to raise money anytime soon.** **How are you thinking about that in terms of just investment into a company like Anthropic, that your valuation might be going up, but your actual ownership stake might be getting diluted over time?** The deals have always been very creatively structured.  The number is they’ve actually only burned $10 billion of cash. They have $120 billion-plus of cash and marketable securities in the balance sheet today, and post-IPO, probably another $100 billion. It’s a great deal for both parties. They only really rent compute. It’s not like they have these massive capital outlays or even rent-to-own. You can make some arguments about certain accounting and there are these massive compute commitment obligations and that’s technically future debt, or that is sitting somewhat off balance sheet essentially. But from a dilution perspective, all tech companies dilute. Are they going to need to raise another $100 billion of debt or equity at some point? Possibly. Can they continue to creatively structure and finance future compute buildouts? Probably as well. I know there’s been a lot of angst about the future of compute financing and data center financing in the last few weeks. But if you’re Anthropic, if you’re a major lab, a hyperscaler or Broadcom or Nvidia that are backstopping a lot of these deals, that’s probably not going to change anytime soon. So I’m not particularly worried. As an equity investor, would I be worried about getting super diluted? I don’t think so. If you believe in this long-term AGI bull case and that Anthropic is going to be a leader in this, that probably shouldn’t worry you. You’re probably looking at a $10 trillion-plus business. They can outrun dilution at that point. **Let’s talk about the compute commitments. I saw at least one analysis saying over the next 10 years, it’s $800 billion-plus in commitments to the various hyperscalers and neoclouds. That’s a lot. You’re saying that they don’t have that much money. They haven’t even spent that much money yet. Do you see them generating the revenues to pay off those commitments?** Yeah. I think when people look at the headline commitments, the $500 to $600 billion number over 7 to 10 years, let’s call it, what they’re paying for compute today actually isn’t that many gigawatts. If you think about their spending right now, on average they’re close to $11 million per megawatt in spend, or $11 billion per gigawatt. If you take that back to the annual level and say, “Okay, the $520 billion number is over 70 years, for example, that’s really not that many gigawatts and they’re going to be well beyond…” When they get to something like 20 gigawatts, the future compute commitments that they’re running essentially is going to be well into the trillions. So yes, they are going to have the revenue to do so. It’s more about how much revenue they are going to have relative to how much compute. Two, the compute efficiency equation. If they keep getting a lot more revenue out of every inference megawatt, how much do you really want to invest in training and take margins to zero from a rental perspective? Because you can get to something where even 80 percent to 85 percent of their fleet is just on training and research at these kinds of margins. The revenue is going to scale pretty well. They’re not in any dire situations there. Unless we realize all this revenue disappears or people churn. It’s not like they have to hit a trillion dollars of revenue extremely, extremely quickly to achieve all this and pay for this. **Unless you’re trying to build digital God and unless it doesn’t diffuse across the economy as fast as everyone wants it to. Those are bets, right?** They’re huge bets. The numbers you look at, if you’re really bullish on this ramp into the digital economy, is Anthropic saying within their Fortune 500 customers, they have 50 percent of the Fortune 500 and that is just 0.1 percent of their total customers’ operating expenses. When you look at that on a per user basis, it’s like $8 a month if you believe OpEx doesn’t include the cost of goods sold. If you believe they include the cost of goods sold, it’s sub-$50 per month. They said the same thing. Of the top 5 percent of Claude Code users, they’re spending $18,000 a year on average, the top 5 percent of users. Their non-Claude Code users spend under $75 a month, well under $75. There’s this huge opportunity, especially when we talk to enterprises and these traditional buyers of IT software. Penetration is extremely low in traditional white collar work. That’s the big point, too, that there is a large ramp period here of both seats adoption and usage. You can make arguments of who’s going to win in this, but it’s obvious the frontier labs are going to get some portion of this and it’s going to be pretty material to net new ARR next year. **One of the things that’s really interesting about that is we talk about the frontier labs, we talk about the hyperscalers — Anthropic has a lot of wild risk factors in this S-1, including that Claude might be conscious and kill us all.** **But one of the most direct and maybe easiest to understand is that AWS, Azure, and Google Cloud are both suppliers and competitors in some real way, particularly Google. And they say, “These relationships could turn.” They have a big relationship with SpaceXAI. Elon Musk is going to train Grok. They’re going to compete with SpaceXAI at some point. That relationship might become precarious.** **As you say, they’re only renting their compute. At some point, the market might force them to own more of their compute because their own competitors don’t want to lease their capacity to the leading company. How do you see that playing out?** We definitely see both OpenAI and Anthropic going more direct and not owning the compute, but more rent-to-own, or going through an SPV as some financial sponsors do and producing the compute more directly versus renting. There are two points to this. Google, from one perspective, is also selling its external chips to customers and saying, “We don’t care to monetize this within Google Cloud or even allocate the share… If you think we’re very chip-constrained, we don’t even care that these are going to DeepMind or internal GCP compute for us to monetize. We’re just willing to become a chip seller.” Anthropic loves TPUs and there are 3.5 to 4 gigawatts-plus, and probably more in the pipeline, in terms of orders there of Anthropic buying these chips directly to then rent. So yes, the market is good. **Wait, it’s not buying them directly to rent. So they’re buying them into a neocloud?** Yeah, they buy them through an SPV, or the SPV buys them and then they rent it from the SPV. **Got it. Okay.** Even within Google Cloud or AWS or Foundry, Anthropic is massive — 48 percent of their business comes through the Bedrock Foundry and Vertex where people are paying through their existing cloud relationship to access these models. They’re not even talking to Anthropic. You don’t have to interact with the Anthropic sales force. **You mentioned this earlier. This marketing channel where Anthropic gets a rev share.** Right. If you’re AWS and you’re saying, “I’m running Anthropic compute for, let’s say, $10 million per megawatt, but now I’m getting 16 percent of $65 billion-a-megawatt plus the compute costs as well,” so we’re seeing margins at all the clouds go up right now. Part of it is CPU utilization is going up because of all the coding and then some on the \[reinforcement learning\] side, but these token-as-a-service businesses as we call them — so AWS Bedrock, Microsoft Foundry, Google Gemini Agent Enterprise platform, which is formerly Vertex — are actually massively increasing the margins on AI at these clouds. They’re benefiting as well when shareholders are giving them credit for this recently. In Q2, both Azure and AWS had extremely, extremely strong quarters, AWS specifically, on the back of Anthropic’s growth on Bedrock. **Do you think that will keep Amazon from ever wanting to train its own frontier models? They’re doing a good job of letting AWS be the infrastructure for everybody else, but they’re one of the only big tech companies that hasn’t felt the insecurity of not having their own models.** Yeah, it’s interesting. Amazon is very happy to serve a wide variety of customers and they have very little \[frontier capability\]. When we look at our estimates of training compute, they have the least of any hyperscaler, except maybe Oracle. It doesn’t seem like there’s much strategy there. I know Microsoft in the last two quarters has been talking about it a little more. AWS on the call has been talking about it a little more, but it’s not huge. They’re not doing it to the same extent that Meta is or trying to be competitive here or even Google with Gemini. Especially someone like Dylan will tell you, if you’re not trying to compete at the frontier, there’s really not a place for you, especially if you’re only dedicating a few hundred megawatts internally. One, internal use case is like serving Amazon retail, but then two, training. There are a lot of commitments to make to be good at this. **It’s a conversation for a different day, but I look at these** [**Microsoft job cuts**](https://www.theverge.com/tech/1018666/microsoft-expands-buyouts-60000-employees-eligible) **and I see it as the glimmer of evidence that they’re going to try to compete in the future. They obviously got burned pretty badly by OpenAI.** Yeah. They have, in the new agreement that they signed, the right to monetize OpenAI models for essentially not paying OpenAI anything within that agreement. They’re going to have some interesting things ahead where they monetize the OpenAI models or even through Microsoft 365 Cowork and Copilot applications where they’re not paying OpenAI anything for a frontier model. **The other dynamic here, and you brought it up earlier, is over the last few weeks, there’s just been a lot of stress about neoclouds — whether they will succeed, whether they will fail. I think Elon Musk said there will be a** [**guppy-like extinction event**](https://fortune.com/2026/10/26/elon-musk-ai-boom-debt-neoclouds/)**, which is just a brutal metaphor to use to describe a bunch of data centers.** **But this is a bunch of circular financing and Nvidia stood a bunch of them up and maybe the hyperscalers are just going to run away with it. Maybe Anthropic’s demand, because maybe they will want to own more of their own compute because renting it is precarious in all these ways.** **What do you make of the neoclouds? Do you think they’re going to survive the market actually shaking itself out as these companies go public?** Elon obviously has some incentive to say neoclouds are going to be extinct. ClusterMAX said something else and [just does amazing ratings of the neoclouds](https://newsletter.semianalysis.com/p/clustermax-30-the-industry-standard). There are a lot of neoclouds that are ex-crypto miners that are not good businesses and are not good at operating clusters. **I just want to point out CoreWeave is an ex-crypto mining company, right?** Yeah. I don’t mean CoreWeave, but there are others further down the list. CoreWeave, Nebius, and anyone that has backing or backstops from one of the major labs, hyperscalers, Nvidia, or Broadcom is going to be okay. It’s that longer tail. Hyperscalers have been 65 percent of CapEx and net new compute brought online. There’s 35 percent of the market that doesn’t have the balance sheets that the hyperscalers have. Both Google and Meta have recently raised or announced they’re going to raise. When we look at the big four hyperscalers, they don’t actually need to raise that money over the next two years, even though CapEx is growing from $1.2 to $2 trillion pretty quickly. But for a lot of these other neoclouds, if you don’t have that sort of backing, I think it is very hard, even right now, and becoming extremely difficult if you don’t have the backing or backstops from those players to keep building compute. It’s a very, very tough financing market. **One of the things I think is really interesting about Elon in particular talking about neoclouds dying and obviously hyping up his own data center business is he’s trying to put them in space and to get away from the backlash in the United States around data centers in particular.** **We have a lot of coverage about data center backlash in the United States, it is slowing down a lot of projects. But at the rate we’re going, if you think the neoclouds are shaky, you might actually end up abandoning a bunch of GPUs and half-built data centers or have them tied up in financing agreements where they can’t be lit up.** **What do you make of that? Do you think we actually just need to go to space? Or do you think the market is going to rationalize and actually take advantage of all the compute that is currently allocated but not being used?** We’ll rationalize more so. The market is very good at finding places to put chips here on the ground. Politically, we’ll see what happens next week with midterms and the fallout from that. Obviously, there are some very large sites that still have massive permitting problems. There’s the entire power equation as we look towards next year. The data centers will be built. There will be a method to get these built. The biggest thing — and the hyperscale CFOs have said this — from an overbuilt perspective and on a relative basis, it’s pretty cheap to build the data center shell. What they can do if they overbuild is they’re going to say, “We’re just not going to take delivery of the chip. We’re not going to buy chips. We’re not going to fill them up.” They’ll just sit empty and chipless. But the majority of the spend is on the chip side. So if we do end up in a place where we’re actually oversupplied from a data center perspective, I know that’s a little bit different. It’s the exact opposite of what the question was. They have said that from an ROI perspective. This comes back to that larger question of whether we are still going to be super supply constrained. A lot of people now, especially in the last three weeks, are talking about, “Are we in this massive glut of compute as we look towards 2028 actually?” Because the same thing we talked about from Anthropic going from negative 94 percent margins to positive 85 percent and compute efficiencies mean as you bring on more inference megawatts that are making a lot more revenue per megawatt, all else SQL, you need less compute. So it’s a question of whether we we going to have enough revenue in 2028 to justify the ROI on $2 trillion plus of CapEx. **This is why I spent so much time asking about capability overhang and how good the models are today, because you can see some competitor that is less idealistic about AGI saying, “Actually, we’re going to do a bunch of forward-deployed engineers in all of these verticals where they haven’t taken advantage of AI and sell them high-margin inference because we’ll automate big chunks of their business and they’ll see the return.” Right now, a bunch of the Fortune 500** [**has not seen a return on AI spending**](https://www.apollo.com/wealth/insights-news/insights/daily-spark/no-signs-of-ai-payoff-outside-the-tech-sector)**.** **So if you’re actually just better at customer service and deployment and productizing the model, you might get way ahead of where Anthropic needs to be because they’re still thinking about AGI every single minute of every single day. I look at Dario standing next to Marc Benioff at Dreamforce and that dude does not want to be there right now. He wants to be thinking about the machine god.** **There’s something there where you can just see competitors coming up with an open source model and saying, “We’re actually going to productize this and we will actually take the revenue that everyone else is depending on.”** Yeah, it’s extremely, extremely interesting. In the interim, that’s why it is getting more verticalized because you need something to bridge AGI where it just fully does the job, you need something to bridge the actual model today. **_\[Laughs\]_** **Yeah. We’ll just take that for granted.** To get there, you need some bridges of where it sits with the actual end user at some point. Everything is going to come down to this: Is there going to be some timing or adoption mismatch where we just don’t get the revenue at a certain point of time to justify that chase of “train and compute towards AGI”? That’s a risk. The other risk we look at is whether we get more models at the frontier even at the open source that are as competitive in coding, where token pricing across the frontier needs to be taken down, because like I mentioned earlier, there are lower switching costs. Even at SemiAnalysis, we went from 95 percent or 90 percent-plus of our usage at the end of Q2 was on Anthropic’s models down to 70 percent after the release of OpenAI’s ChatGPT 5.6 and Astra. And then with Opus 5.5, it’s gone back to 90 percent-plus roughly. It does come back. Even the base of revenue today — and they’ve said that we have net revenue retention of 900 percent, so they’re seeing this massive, massive growth. This isn’t a traditional software business like ServiceNow or Workday, where you have just an incredible logo or gross retention of 95 percent to 98 percent and customers sign these nice three-year contracts, they never churn, they add some seats, and maybe they adopt another product that is being upsold. This is purely consumption. The other portion was the 20 percent drop in the last two weeks of the year versus the prior two weeks last year in token spend, because of the holiday season. You’re not going to do work code stuff, you’re not going to be coding for work during the holidays. This is extremely, extremely pure consumption business, even different from how we think about Snowflake or Datadog today in public markets. **Obviously, Anthropic is not the only company that’s going to have an IPO. OpenAI is right next to them. They’ve delayed theirs. You mentioned earlier Anthropic is much more enterprise focused. I think OpenAI for a long time thought that Google wouldn’t react to them and they would just take the search business away and everything would be fine. Google obviously reacted for better or worse. Search is now AI Overviews.** **Looking at the Anthropic S-1, what are you looking ahead to with OpenAI? Because they are a more consumer facing company.** OpenAI was 60-40 consumer, 60 percent consumer and 40 percent enterprise at the end of Q1. That didn’t flip for the first time until January or August to be enterprise focused. What they’ll tell you is enterprise business — because it was so consumer and subscription focused, this consumption business that’s 80 percent of Anthropic’s revenue — didn’t take off until a release of 5.6 and then with Astra. Those are really great models. OpenAI took the majority of actual net new ARR in Q3, but it’s going to take some time because what’s been in the _Wall Street Journal_, like negative 100 percent EBITA margins versus we’re seeing positive 11 percent here. There are some differences, some pretty big differences, in the business model where they just don’t monetize at the same rate that Anthropic does. That’s a little bit because the business makeup — the verticals, that consumer-versus-enterprise split — and that’s changing. What they’ll tell you is they see a really good Q4 and they have a much stronger enterprise sales function than Anthropic has. I expect them to come up, but I don’t think the financials are going to look as good as this even when they drop their S-1 maybe in early 2027. **Do you think that’s why they delayed it? To get their financials looking better?** Yeah, I think OpenAI had a [really strong Q3](https://www.wsj.com/tech/ai/the-ai-price-war-is-heating-upand-openai-is-gaining-ground-on-anthropic-48b8525b). Enterprise grew 100 percent quarter-over-quarter versus the total business which was 77 percent. You get another strong quarter of enterprise here. If Opus 5.5 didn’t take a ton of share, that would put OpenAI in a much better position to tell the story. **There are some other turns here between now and when Anthropic actually goes public. They’re going to do a roadshow. There’ll be some pricing events. What are you looking for from Anthropic?** They did an investor day two or three weeks ago where they did give some more color. I think they’ll walk through why Q3 was so weak. Meta was a very, very large customer, probably over 10 percent. What’s been publicly reported is they went from 60,000 Claude Code seats down to 30,000. These token volumes didn’t disappear, they just shifted into Meta’s internal models. Anthropic started cracking down on anti-distillation, and that was 5 percent of revenue prior. I think they said 5 percent to 7 percent of revenue in the S-1 and stuff against their terms of service. So what’s here on the roadshows is a lot of this October acceleration after Opus 5.5. We believe that they’re back on track. If they were doing $10 billion of net new ARR per month in Q2 and that fell down to $4 billion in Q3, that they’re getting back on that $8 billion to $10 billion pace here in Q4. People want to hear about these new verticals, specifically cyber. They’re doing some patching until they have that full cyber SKU rolled out, but they’re definitely monetizing cyber today for vulnerability code scanning. These are not small contracts, I should say, with existing customers. I think they’ll start to message what the 2027 ARR growth path looks like. The other thing too is expectations management, keeping numbers low enough, not talking about AGI as much. **_\[Laughs\]_** **It’s hard to do expectation management when your risk factors are like, “We could pose an existential risk to all of humanity.”** _\[Laughs\]_ Right. And so saying you do that — you’re going to achieve AGI at some point, you’re going to replace everyone’s jobs — but then trying to keep numbers low enough for Q4 and to have this massive beat the first quarter you become public and have a decent guide for 2027, it’ll be a nice mix of seesawing between that long-term vision and short-term expectations management. **There’s a lot going on. Joey, this has been great. We’re going to have to have you back as these companies barrel towards actually being public and their economics become much more visible. Thank you so much for being on** **_Decoder_****.** Thank you so much. I appreciate it. _Questions or comments? Hit us up at decoder@theverge.com. We really do read every email!_