国际新闻 · 中文译文与原文对照

星期六 最新 1,000 篇 · 滚动浏览

财经市场

Anthropic有能力将安全放在首位吗?

来源:The Verge 科技发布 上传时间 (多伦多时间)

中文译文

今天,我在本系列关于人工智能在现实生活中的第三篇也是最后一篇报道中,与《The Verge》的资深人工智能记者海登·菲尔德探讨Anthropic的IPO事宜。

上周四,Claude背后的人工智能公司已在纳斯达克上市。其融资规模巨大,但对于本次讨论而言并非重点,只需知道Anthropic现已成为全球价值最高的公司之一。

真正的关键在于接下来会发生什么——Anthropic能否在兼顾人工智能安全并持续投入的同时,仍作为一家需对股东负责的公司运营。此外,它还必须不断增加盈利,以支付日益高昂的算力费用,未来几年这一费用总计约为9000亿美元。

所有这些因素彼此之间存在着有趣的矛盾,因此我和海登试图分析,在面对公开市场的压力时,Anthropic可能会如何应对。您将听到我们谈论该公司独特的公司治理结构,这一结构在詹森·黄相关事件持续发酵的背景下显得尤为重要。我们还会探讨为什么OpenAI似乎乐于让Anthropic先上市,以及为什么Anthropic向大客户推销的产品核心始终是那套安全理念,而这一理念如今却与其以盈利为目标的业务模式产生了冲突。

在开始之前,请帮忙在YouTube上订阅Decoder频道,我们每周一和周四都会发布新节目。

好了:《The Verge》的资深AI记者海登·菲尔德,将为您解读Anthropic的IPO以及前沿AI领域的经济逻辑。现在开始吧。

为便于阅读和理解,本次访谈内容经过了适度删减。

海登·菲尔德,欢迎回到Decoder频道。

谢谢。很高兴能在这里。

今天能与您交谈让我很兴奋,因为这并非纯粹的闹剧或混乱场面。

我知道。

这就像是一件平常发生的事情。这一次并非围绕人物性格或肥皂剧情节展开,而是关乎冷酷、现实的金钱利益。

[笑声] 嗯,其中还是有一些人物特色的,因为我们会大量讨论Anthropic的IPO事宜,而其中涉及许多有个性的人物。他们构建这家公司的方式,旨在抵御资本主义的贪婪压力,我觉得相当独特,反映了若干特立独行之人的理念。

没错,确实如此。重新研究那份S-1文件,了解谁在哪个委员会任职,哪些创始人拥有最大决策权,以及他们实际的掌控力究竟有多大,这真的很有趣,实际上他们的掌控力比看上去的要强。让我们从开头说起。你和我正在讨论IPO当天的事情。Anthropic是第一家上市的纯AI实验室。他们将IPO定价定为每股83美元,最终股价甚至涨到了111美元。等到大家听到这段消息时,谁也不知道股价会是多少,但对所有投资了Anthropic的人来说,这无疑是一次巨大的胜利。许多律师事务所都修改了首页内容,写着“恭喜大家!”

你对此有何看法?这只是必然会发生的事情,还是表明AI泡沫不会破裂的迹象?人们应该如何看待这个问题?

我认为这并没有为我们提供关于AI泡沫是否会破裂的新信息,因为这正是AI热潮的巅峰时刻。这是第一家上市的AI实验室,几个月来一直处于打破纪录的轨道上。大家都觉得“它的股价很可能会超过马斯克”,而事实也确实如此。我迫不及待想看看他之后的推文。但这件事另一个有趣的地方在于,我听说会有数千名新的百万富翁和数百名新的千万富翁由此诞生,更不用说亿万富翁了。在我看来,这正是人工智能泡沫持续膨胀的体现。

今天早上我甚至与几家不同公司的科技从业者进行了交谈,他们都说在各自的公司里,人工智能泡沫越来越膨胀,他们被要求承担越来越多的与人工智能相关的工作,即便这会降低他们的效率,也让工作质量下降。

所以我认为最终会有一个清算的时刻,但目前我们正处于人工智能热潮的顶峰,Anthropic在将其业务宣传为“安全的人工智能”方面做得相当出色。当然,现在他们在这一宣传理念上面临的压力比以往任何时候都要大。因此,看看这种紧张局势如何成为焦点,以及他们将如何应对,将会很有意思。Verge的订阅用户请不要忘记,无论您在何处收听播客,都能享受无广告的Decoder服务。点击此处即可使用。还不是订阅用户?您可以在此处注册。

本周早些时候我参加了Semafor会议,看到OpenAI的首席执行官Sam Altman以及首席财务官Sarah Friar出现在舞台上。Altman和Greg Brockman则参加了本周的DealBook会议。我感觉他们似乎很高兴Anthropic先一步行动,以便观察市场反应,了解各公司对其主要竞争对手的营收潜力有何评估,从而能更好地讲述自己的IPO故事。

您也是这么认为的,认为Anthropic只是个试验品?还是说OpenAI只是在试图掩饰自己并非第一个采取行动的事实?是的,我认为是这样的。我听说在过去一年里,奥尔特曼一直一心想要抢先Anthropic实现上市。我就是这么听说的,那是他唯一关心的事情。但今年随着OpenAI面临诸多负面新闻,他们接连遭遇挫折。目前很多公众都特别厌恶OpenAI。

显然,存在着强烈的AI民粹主义反弹情绪,很多人讨厌各种AI技术,但看起来今年OpenAI确实承受了大部分公众的厌恶以及各种丑闻。于是他们就想:“你知道吗?我们先退一步,让Anthropic先上市吧。”而且他们似乎也乐意这么做。既然Anthropic率先迈出了这一步,创造了这一具有历史意义的先例,OpenAI就可以借助其势头宣称:“没错,我们也会这么做。”我怀疑他们能否打破Anthropic创下的纪录。也许能,也许不能。但不管怎样,我认为他们明白,如果Anthropic没有率先行动,如果这件事不是在今年发生,他们的处境本可能糟糕得多。

既然Anthropic已成为一家上市公司,我就忍不住从简单的商业基本原理角度去思考它。用户从他们那里购买代币,而他们制造这些代符的成本必须低于售价。同时他们还想要开发通用人工智能,此外还有其他一系列庞大的开支。

仅从目前我们所了解的这家公司的经营状况来看,作为一家上市公司,它表面上有盈利吗?在推理业务能够带来足够利润以支付新模型的训练成本之前,他们还有很长的路要走。他们的业务实际上是如何运作的?他们还有很长的路要走,正因如此,看看他们在增加收入方面会采取哪些具体措施将会非常有趣,因为这正是他们必须做的。我见过史蒂夫·拉特纳的一个估算,称Anthropic在未来10年内需要将其收入提高大约17倍,才能实现与科技行业某些竞争对手相当的回报水平。

但即便没有这个具体数字,也能看出问题所在。我的意思是,当一家公司上市后,投资者希望看到投资回报。他们希望看到利润上升,或者至少有利润。他们还希望看到,即便目前还没有利润,也要有足够的收入增长,以便在短期内实现盈利。在今天之后,Anthropic将比以往任何时候都更需应对这一挑战。我与一些企业治理专家和法学教授交谈过,他们都说Anthropic和OpenAI一样,拥有极为特殊的公司结构,因此并不像某些上市公司那样需要对股东负责。股东永远无法投票罢免高管,他们实际上也没有太多权力来影响任何事情。真正拥有权力的是创始人以及那个我们稍后会谈到的特殊信托机构。

不过不管怎样,他们都说股东是否有权参与投票其实并不重要。关键在于,无论由谁来决策,Anthropic都将面临巨大的压力,必须盈利、购买计算资源并努力赶上OpenAI的步伐,这将会给他们带来巨大的考验。我们很可能会看到很多事情被搁置,甚至一些昂贵的安全措施也可能被放弃。我想从治理结构的角度来谈这个问题。我们正处于英伟达丑闻频发的时代,企业治理如今极为热门,是美国商业界最受关注的话题,不过这一部分我们稍后再谈。

我只想再继续探讨一下“能否以高于制作成本一美元的价格出售代币”这个问题。这种思考方式实在是最愚蠢的。Anthropic从大型云服务提供商那里租用算力,这些提供商拥有高性能GPU,他们需要支付电力费用和GPU费用。有了这些成本之后,他们还必须以高于所有这些成本的价格出售代币。那是思考他们业务时最愚蠢的方式。我向你保证,他们其实会以更复杂的方式来考虑这个问题,但这是我个人的看法。我本周刚参加了Microsoft Ignite大会,见到了萨蒂亚·纳德拉,他明确地告诉我以及其他人,他们打算降低代币的价格,他们的整个系统都是为了让Copilot旗下的所有AI供应商的产品具有可互换性,从而为各项任务选择最合适的模型,并确保用户能以最低的成本完成所需操作。

这似乎与Anthropic需要做的完全背道而驰,后者需要的是让每销售一个代币都能获得更多利润,这样才能继续从数据中心租用算力,或购买比OpenAI更多的芯片,以满足自身的需求。他们打算如何解决这个问题?是通过让更多人使用AI,还是通过提高代币的价格?我觉得两者都有些,这很有趣,因为Anthropic一直在暗中试图这么做。过去一年里我们写过一篇文章,讲述当时出现了强烈的抗议声,尤其是针对Anthropic将成本转嫁给消费者的行为。作为Anthropic客户的个人们非常愤怒,因为他们觉得同样的产品价格却在上涨,质量还更差,而这些成本却都被转嫁到了他们身上。

这种情况在整个行业都出现了,引发了大规模的抗议,但Anthropic承受了最大的压力。Anthropic不得不发表公开声明,表示“我们会解决这个问题”。他们确实稍微降低了价格,但我的意思是,他们仍得以某种方式继续这么做。他们关注的另一个方面是希望自己的模型足够优秀,能让人们对其产生依赖——或许不是情感上的依赖,但至少会对其产生依赖,从而将人们困在自家体系中。如今他们全都致力于构建这种“护城河”:“我们必须打造护城河。”OpenAI也是这么做的。他们希望用户的日常生活和业务尽可能多地与Anthropic的工具结合,这样就很难将其他工具替代过来,用户也不愿这么做,只能继续付费。这就是他们的目标。

当然,他们还希望企业及政府合同能占据很大一部分业务,这样一来就不必太担心那些小额收费的问题,因为他们拥有大量资金。但我认为有趣的是,对于那些非常在意价格的人来说,他们只是试图将这些人困在自家体系中,然后寄希望于最好的结果。有趣的是,消费者在价格方面较为敏感。而Anthropic的传统重点客户——那些大型企业,他们对价格的态度则有所不同。

如果你向Anthropic支付大量费用,你可能会希望能够快速更换服务提供商,还可能要求数据主权以及大型企业所关心的其他种种权益。Anthropic要么必须满足这些要求,那样就会削弱其针对大型企业的竞争优势;要么就得宣称“我们太优秀了,无法满足这些要求”。不过我不确定他们目前是否有能力这么说。这就是为什么他们如此大力推广这款安全、可靠的人工智能产品线,因为这是他们目前唯一的优势。面对企业客户时,大多数情况下你确实必须保障他们的数据主权,不能利用他们的数据谋取利益。在客户愿意支付如此高昂的费用时,Anthropic有很多本可以用来盈利的做法却无法实施。

我认为这正是他们不断强调自己是安全、可靠替代方案的原因,因为只有这样才能真正留住客户,尤其是在当前各种系统频频遭到黑客攻击的背景下,如果他们还能进一步突出其网络安全功能的话。另一件有趣的事情是,即便不是大型企业,就连中小型企业也在这么做。我与很多这类企业的人交谈过,他们都在自行开发内部工具,用来判断哪些任务确实需要使用价格昂贵的Anthropic模型,而哪些任务则可以转而使用开源模型或价格更为低廉的模型。

因此,很多公司,比如律师事务所,以及各行各业中拥有中小型企业的机构,都在自行设定内部标准,比如“好吧,我们要尽可能节省成本,只将10%的任务交给价格昂贵的Anthropic模型,40%的任务可以使用开源模型,50%的任务则可以使用价格较为低廉的中级模型。”

这正是让我觉得有趣的地方。人们都在自己想办法解决这个问题,甚至不公开这些标准,因为这是他们自己的预算管理方式。我本周确实听到纳德拉表示,每家公司都会有自己的一套评估标准,他们会用这些标准来测试各种模型,并始终选择最高效的那个,而这将成为每家公司都会做的事情之一。

不过另一方面,我也听到Anthropic的高管们说:“你总是会想要最先进的模型,因为你的竞争对手也会使用最先进的模型,如果你使用较便宜的模型,就会落后。”在我看来,这种观点有点天真且理想化,因为我认为很多公司会选择低成本方案,只是因为他们知道竞争对手也会这么做。

我完全同意。现在大家都在追求低成本。是的,我认为如果“足够好”就能过关,你最终会弄清楚到底什么才算足够好,而且不会认为竞争对手会通过使用更昂贵的产品来获得5%的优势。所有的成本都应该控制在合理范围内,市场会自行调整。

没错。

只是感觉在企业初创阶段,或许在首次公开募股那天,压力并不大,大家都很开心,收入也很高。但从下一个季度的业绩开始,我们就会看到状况有所变化。在需要不断扩张以发展各项业务的背景下,尤其是当其供应商本身已成为大股东时,它还能保持目前的增长速度吗?

完全正确。我们确实看到过有这样一家企业,一个季度实现了快速增长,但下一个季度的增长幅度就没那么大了。所以,他们确实有很多问题需要解决。让我们来谈谈治理问题。由于英伟达的存在,企业治理已成为美国商界最热门的话题。正如你所说,OpenAI和Anthropic拥有历史上最为复杂的治理结构。

如果你认为Anthropic的未来在于出售最先进的智能技术——即最具能力、最前沿、最接近通用人工智能水平的智能技术——那你很可能不会投入大量精力在安全措施上。你也不会为了那些要求你停止研发、放慢进度或暂不发布最优秀模型的研究人员而牺牲自身的利润空间。

而Anthropic却坚定地致力于这么做。我们之前曾讨论过,这样的做法能否承受资本主义和增长带来的压力。在它们还不是上市公司的时候这就已经很困难了,现在更是如此。让我们来谈谈Anthropic的治理结构,以及他们是如何努力将安全承诺制度化起来的。是的,这确实是个非常好的问题。我刚刚发表了一篇关于这个话题的文章。让我觉得有趣的是,尽管Anthropic目睹了OpenAI在萨姆·奥尔特曼被罢免以及该公司经历的各种重组过程中所出现的种种波折,但Anthropic的架构也同样与众不同。从某种程度上来说,这很有好处,因为正如我们所说,他们不必受股东压力的影响,除非股东开始抛售股票或不再增持。

但在Anthropic中拥有决策权的两类人分别是创始人以及名为“长期利益信托基金”(LTBT)的团体。Anthropic喜欢将这一机制宣传为“不必担心,创始人会为人类利益着想,而且没有任何制约。我们上面还有这个信托基金,可以随时任命或罢免我们”。正如我们在电影《人工智能》中看到的那样,萨姆·奥尔特曼曾说过:“嗯,董事会随时都可以解雇我。”他一遍又一遍地说着这句话。

现在,Anthropic也在暗示同样的意思:“哦,我们有这样一个监督机构在上面。他们可以罢免或任命大多数董事会成员”,尽管实际上并非所有成员都能被如此处理,但我们稍后会谈到这一点。所以这个监督机构目前由四个人组成,而他们中没有人拥有丰富的AI领域经验。这四人分别是[前美联储主席]本·伯南克、一名外交政策分析师理查德·方丹、一名实验物理学家克里斯托弗·斯塔布斯,以及一名前GiveWell高管尼尔·巴迪·沙阿,后者还是LTBT的主席。

我与几位不久前离开Anthropic的前员工交谈过,其中一人表示,这个监督机构是经过精心挑选的,目的是避免在Anthropic内部引发动荡。这些人是创始人任命的,他们不会任命那些很可能会经常与他们意见相左的人。具有讽刺意味的是,该监管机构并无权力干预Anthropic的日常决策,甚至连曾经存在的OpenAI非营利组织董事会也拥有这样的权力。他们可以说:“哦,看吧,不要发布这个模型,它太危险了。”而该监管机构没有任何此类权力,它所能做的只是任命或罢免某些董事会成员。即便如此,他们也必须提前30天通知董事会的一个小型委员会,该委员会的主席是丹妮拉·阿莫代伊,她同时也是Anthropic的创始人之一。简单来说,这七个董事会席位中,信托机构可以挑选三个,创始人可以挑选三个,而作为决定性的第七个席位,虽然名义上由信托机构任命,但实际上必须获得创始人的批准才能这么做。因此,尽管这件事看似复杂,但仔细分析后就会发现,创始人实际上拥有几乎不受限制的控制权。我觉得这一点在“哦,这个信托机构高于我们,不必担心”的说法中被忽视了。实际上,他们并没有太多权力来制衡创始人。

在多年从事相关节目制作及研究商业领域的过程中,我所见过的创始人控制结构最为离奇的就是Anthropic的这种模式。Anthropic有七位创始人,他们共同持有1股股票,而这股股票拥有公司50.1%的投票权。该股票由一家名为Founder LLC的公司持有,而这七位创始人则是Founder LLC的共有人。那到底是怎么运作的?他们是不是聚在一起,然后决定如何投票决定那一股股票的去向?

没错,这真的很有趣,这也是为什么我刚才在解释谁有资格选择这个决胜席位时稍微简化了情况,因为从技术上讲,它属于创始人有限责任公司。整个结构简直一团糟。但如果你看看简化后的版本,就会发现,通过所有这些复杂的结构设计,他们实际上拥有几乎不受约束的权力,而这正是他们所希望的。

我曾与一位最近离职的前员工交谈,他说创始人并不一定是为了赚最多的钱而这么做。在他们看来,他们更看重的是不受限制的掌控权。我认为这是个非常重要的观点。在Anthropic内部,有一些人认为他们的动机并非想要发财,而是渴望获得完全的控制权,这一观点与他们许多公开声明相符。他们说:“无法从次要位置着手确保安全性。”他们都希望加快这场竞赛的节奏,但问题在于,一旦加快人工智能竞赛的速度,就会影响到所有参与者。

对我来说,安全性问题是其中最有趣的部分,因为在没有重大诉讼纠纷或责任风险的情况下,这是人们愿意投入资金解决的领域,而其他公司则不会这么做。

假设法庭最终认定模型公司需对人们通过其模型所做出的行为负责,我认为那时所有的模型公司都会建立安全系统。他们会有强烈的动力来确保自己的模型是安全的。但目前这种情况尚未发生。我们还不清楚。关于安全性以及监管措施将从何处出台、什么因素能形成相应激励,目前还存在诸多争论。与此同时,有七位创始人各持一股股票,目前他们似乎都同意以各种方式在Anthropic内部注重安全性。但如果他们开始出现分歧会怎样?

是的,这是个很好的问题。我认为另一个有趣的问题是,如果有人以“为了更大的利益”为借口为一切安全问题找借口,那会怎样?众所周知,这是每部电影中常见的反派套路。但这也是Anthropic网站上强调的重点之一。他们声称:“我们希望为更大的利益、为全人类的利益而这么做。”如果你说:“嘿,我们必须不惜一切代价赢得这场比赛,因为我们是唯一的好人,而其他任何获胜的人都会比我们表现更差”,那么在那种情况下你真的还是好人吗?难道你就不能把每一个决定,甚至是放弃某些安全措施的行为,都解释为“哦,这会让我们进展太慢。我们会输的。我们必须不惜一切代价取胜”吗?于是你就放弃了这个或那个措施。然后在比赛结束后,你照着镜子说:“哦,实际上,我们和那些我们试图与之区别开来的公司没什么两样。”我觉得这种情况越来越普遍了。

我的意思是,现在他们声称“看,我们真的非常致力于此”。你必须记住,这些人几乎是以近乎宗教般的虔诚态度来追求安全的人工智能理念的。他们中的很多人认为人工智能系统有可能具有意识。这是Anthropic创始人自己的说法。所以我认为,他们并非只是暗中试图赚钱,同时嘴上谈论安全却并无真心。我觉得情况更危险一些,因为他们想要完全掌控一切,因为他们自认为拥有最完善的AI安全方案,而且认为自己能比其他人做得更好。

但正如我们所知,绝对的权力会导致绝对的腐败。因此,我交谈过的许多Anthropic员工,包括前员工,都表示:“没错,这个行业如此缺乏监管实在不合理。没有外部监管机制。即便这七个人自认为拥有最好的AI安全方案,那也该向其他人证明——你们不应是唯一做决定的人,因为你们并非通过选举产生的。”这就是一些人的共识。

他们有一项具有道德约束力的协议,允许外部审计人员进入,海登。一切都会好起来的。

哦,对了,是允许员工级别的访问权限。现在它被称为“超级智能”。一切都会好起来的。他们都去了白宫,彼此深情对视,然后说一切都会没事的。

显然,Anthropic早在Nvidia和黄仁勋的相关事件之前就设计了这种架构,但我还是忍不住在脑海中将它们联系在一起。这些事件似乎是同时发生的。而Nvidia则拥有非常传统的企业治理结构。

黄仁勋持有该公司3%的股份,但他也有董事会。董事会认定他在某些方面违反了公司的行为准则,于是他选择了辞职。很明显,他是被强烈劝告辞职的,但这只是正常的企业治理做法。这是学校里会教的内容——公司有董事会,而董事会的首要职责就是决定CEO的任免。在硅谷情况则大不相同,对吧?在各地,创始人对公司的控制权都是一个重要话题。马克·扎克伯格对Meta拥有控制权,尽管Meta的股东们多次投票要求罢免他,而这其实是美国商业界中常被忽视的现象之一。谢尔盖·布林和拉里·佩奇则控制着谷歌。

这是属于那种“我们需要让创始人免受市场及董事会压力的影响”的典型硅谷做法吗?还是说“我们需要让创始人以及达里奥免受市场压力,以便我们能专注于通用人工智能与安全问题”?因为这似乎是两者的结合。这绝对是两者结合的结果。我的意思是,人们说它是后者,但我觉得在很多方面它更像是前者。完全是两者结合的体现。我认为是各占50%的比例。他们的想法是想要掌控局面,这样就能专注于自己认为最正确的事情,而不让那些他们认为无法做出正确决策的人掌控局面。但具有讽刺意味的是,正如我们所知,不应该由任何单一群体来决定什么才是最好的。至少我的很多消息来源都是这么说的。

所以我觉得,没错,这其实就是两者结合。埃隆·马斯克以不愿管理自己没有投票控制权的公司而闻名。我认为这只是硅谷的一种新现象,由于多种原因正在这里蔓延开来。

一共有七个人。在接下来的5到10年里,这七个人会做些什么,其不可预测性不亚于其他任何事情。如果其中一个人做了违法的事情,或者对董事会撒谎,或者把芯片偷运到中国,那该怎么办?有没有办法将其中一位创始人从这个有限责任公司中除名?是的。我的意思是,据说该信托机构有权罢免董事。他们并未说明是否有权力罢免所有董事还是仅部分董事,但从S-1文件的具体条款来看,如果该信托机构提前30天通知丹妮拉·阿莫代以及该小组委员会中的另外两人,他们就可以提议罢免某位董事。当然,如果丹妮拉认为有必要,他们也可以免除30天的通知期限。

不过,这里面有很多细节规定。基本上,确实可以罢免某人,S-1文件中也提到,如果某人犯了违法行为或做了极其恶劣的事情,就可以将其罢免,但这绝非易事。我认为这也取决于被罢免的是谁。这很有趣。我总是会回到这个话题上,因为它极其复杂、十分罕见,而且与英伟达的情况类似——英伟达的公关人员发表声明称一切正常,从而严重歪曲了他与股东之间的关系。他最终不得不承认情况并不正常,自己确实有婚外情,于是董事会解雇了他。

我为那个不得不在那些最终被证明是谎言的声明上署名的公关人员感到些许同情。

我没什么感觉。正如大家所知,海登也很清楚,我们会要求所有公关人员在任何文件上署名,以防出现类似情况。这只是Verge网站上的一些趣闻而已。

不管怎样,这其实就是正常的公司治理方式。在这里,可能唯一的制衡力量就是市场本身,对吧?

没错。如果他们无法以高于制造成本的价格出售代币,就必须在财务报告中披露这一情况,市场可能会认为这家公司正在倒闭,进而抽回所有资金。这是一种截然不同的压力。

对达里奥来说,唯一可能导致他失去工作的情况就是公司破产。在这种情况下,他必须赶在OpenAI之前购买芯片,还要增加Anthropic在企业领域中的业务范围,同时他还理想地希望训练出下一个模型,再通过RSI技术实现通用人工智能的目标。种种压力似乎都在要求他尽可能快地前进。如果把这一切加在一起,我会说自己绝不会放慢脚步。没错。这就是为什么我说,如果有人一心想要成为第一,那他们几乎可以为任何事情找借口,而他们确实如此。而且他们还公开表明了这一立场。关于资金问题,科罗拉多大学法学院的法律教授安·利普顿本周告诉我,实际上拥有治理权的是信托机构以及创始人们,要么你相信他们会妥善平衡社会利益,要么你不相信。

但不管怎样,制约他们各种冲动的真正因素将是Anthropic需要资金这一事实。问题就在这里,Anthropic确实需要资金。在最近的一次直播中,达里奥告诉摩根大通的首席执行官,计算能力是该公司面临的“最棘手的商业难题”。本质上,如果购买过多,企业可能会陷入财务困境;如果购买过少,则无法为客户提供推理服务。更重要的是,他们无法进行研发,也无法开发新模型。因此他们的计算资源极为有限,甚至可能比OpenAI更为紧张——毕竟OpenAI曾因巨额亏损而长期被嘲笑,但他们一直在大量投资计算资源。Anthropic则没有这么做,现在他们不得不为此付出代价。

这真是一个有趣的转折。希望对他们及其利益而言,如果他们能获得所需的计算资源的话,情况会有所好转。但众所周知,供应远远低于需求。所以,看看他们会采取什么措施来赚钱,进而获得购买计算资源的资金,将会很有趣。

有趣的是,几周前我告诉我们的同事莉兹·洛帕托,Anthropic的IPO将成为今年最重大的新闻之一。今天她却嘲笑我,因为她说由于黄仁勋和Nvidia的缘故,这甚至都不是本周最重大的新闻。我的观点——你可以接受或放弃,再自行判断我是否正确——是这两者其实是同一个问题。如果因为詹森不在而无法迫使大家相信英伟达的融资架构,导致该架构出现问题或动摇,那么数据中心里可能会堆积大量芯片,却缺乏相应的许可或电力供应,届时芯片市场将会发生巨大变化。

也许像CoreWeave这样的公司会变得更加不稳定,或者一些原本就不稳定的新云服务提供商也会出现问题,从而导致计算成本下降,让Anthropic有更多竞争优势。也有可能通过灰色市场分配到中国的那些芯片会大量回流到美国。

显然,这些情况都会发生。

对吧?因此,英伟达的局势会通过多种方式影响Anthropic的财务状况。这家公司对此有过任何表态吗?市场又有何反应吗?还没有。他们刚刚讨论了计算能力的重要性,以及这将会对他们构成一个重大问题。至于Anthropic的IPO,它与Nvidia有着诸多关联,情况相当复杂,因为Nvidia此前已承诺提供100亿美元的资金支持,此外还同意通过私募方式购买价值100亿美元的Anthropic股票。

所以这两家公司关系密切,这一发展也值得关注。Nvidia的新任CEO曾多次宣称要阻止原本可能流向中国的芯片,将其拉回国内,而现在他们不再这么做了。

接下来我们将看看芯片供应是否会增加。此外,詹森亲自参与了其中一些交易,尤其是去年8月与大型金融机构达成的为人工智能基础设施建设提供资金的协议。所以他一直在积极操办各种事务。至于新任首席执行官是否拥有同样的能力来推动事情进展并迫使他人采取行动,还有待观察。这是我的最后一个问题。显然今天大家都在庆祝。我们提及Anthropic的公司结构十分奇特,这无疑给这种氛围泼了冷水。人们应该关注什么?既然Anthropic已经成为一只正在交易的股票,现在应该寻找哪些迹象?在您看来,任何一天Anthropic股票的价值与更广泛的AI行业发展之间存在着怎样的关系?

我认为这类股票有点类似于AI行业的基准指标,因为它有时并不如人们所想的那样具有代表性,而且也并非能反映所有情况。因此看待这类股票时总是要持保留态度。它完全基于氛围和炒作。人们会因为恐惧而卖出股票,也会因为听到一些好消息或传闻而买入。所以我认为重要的是要分析当时究竟是什么导致了恐惧或炒作。但同样重要的是要了解该公司在报告和财报电话会议中表达的观点。我认为最有趣的是这一点。并非总是那些数字,而是他们试图通过这些报告暗中传达的信息,比如“哦,我们让这支团队边缘化了”,或者“我们以争当第一为名终止了这项安全措施”。

我觉得我们会看到很多这样的情况,这也是我想要关注的。因为那些以安全的人工智能为发展基础的公司,越来越显得和OpenAI类似,如果我的消息来源没错的话,他们将会有很多类似被掩盖起来的更新。

这是您首先在这里听到的消息。前CNBC记者海登·菲尔德表示,要透过日常的波动,关注公司的基本面。这很好。另外,股市其实只是情绪的体现。这也是我听到您说过的话。

没错。我的意思是,这很有争议,但作为一名财经记者,确实,股市里有很多情绪因素在起作用。确实是充满了情绪。哦,我同意。我觉得有时候他们让我上CNBC说的只是些感觉而已,然后很快就会让我下台。

海登,这太好了。显然在未来的这一年里,我们会多次和你探讨这个话题。希望我们能再做一期关于真实商业故事的节目,而不只是那种像肥皂剧一样的内容。

两种类型的节目我都会参与。我们两种都得做。

太棒了。非常感谢你,海登。

不客气。

有疑问或意见吗?请发送邮件至[email protected]。我们真的会阅读每一封邮件!

《Decoder》节目,由尼莱·帕特尔主持

这是一档来自The Verge的播客,探讨重大理念及其他问题。

原文

Today, I’m talking with Hayden Field, The Verge’s senior AI reporter, about the Anthropic IPO for the third and final episode in our series about AI in real life.

The AI lab behind Claude went public on the Nasdaq last Thursday. The numbers are gigantic, but not really that important for this conversation, except to say that Anthropic has become one of the most valuable companies on the planet.

The real story is what happens now — and whether Anthropic can continue to thread the needle of caring about and investing in AI safety while also being a public company beholden to shareholders. Oh, and it also continuously growing the amount of money it makes so it can keep paying its ever-increasing compute bills, which total roughly $900 billion over the coming years.

All of these ideas sit in a fascinating tension with one another, and so Hayden and I really tried to break down how we think this might play out as Anthropic navigates the pressures of the public markets.

You’ll hear us talk about the company’s unique corporate governance structure, something that is especially relevant in the wake of the ongoing Jensen Huang saga. We also get into why OpenAI seems happy to have let Anthropic go public first, and why Anthropic’s whole pitch to big customers boils down to the same safety branding that now sits in conflict with the profit-oriented part of its business.

Before we get going, do me a favor and subscribe to Decoder on YouTube, where we put out new episodes every Monday and Thursday.

Okay: Hayden Field, The Verge’s senior AI reporter, on the Anthropic IPO and the economics of frontier AI. Here we go.

This interview has been lightly edited for length and clarity.

Hayden Field, welcome back to Decoder.

Thanks. Great to be here.

I’m excited to talk to you today because it’s not pure soap opera drama chaos.

I know.

It’s like a normal thing that happened.

It’s for once not personality-based and soap opera. It’s about cold, hard would-be cash.

[Laughs] Well, there is some personality in it because we’re going to talk a lot about Anthropic’s IPO, which has a bunch of big personalities in it, and the way they have structured this company to repel the rapacious pressures of capitalism is, I would say, very eccentric. It reflects the beliefs of a number of eccentric people.

Yep, absolutely. It’s really been interesting digging into the S-1 again and just seeing who was on what committee and which founders have the most control here and how much control the founders actually have, which is more than it may seem.

Let’s start at the start. You and I are talking on IPO day. Anthropic is the first pure AI lab to go public. They priced the IPO at $83 a share and they got all the way up to $111. By the time people are listening to this, who knows what the price will actually be, but it’s been a real victory lap for all the people who are invested in Anthropic. A bunch of law firms have changed their homepages to be like, “Congratulations, guys!”

What do you make of this? Is it just the thing that had to happen or is it a sign that the AI bubble is not going to pop? How should people think about it?

I don’t think it really gives us new information about whether the AI bubble is going to pop or not, because I mean, this is peak AI hype right here. It’s the first AI lab to go public. It’s been kind of on track to break records for months. Everyone was like, “It’s probably going to beat Musk,” and then it did. I can’t wait to see his tweets later.

But the other interesting part of this is that I’ve heard thousands of new millionaires and hundreds of new centimillionaires will be born from this, not to mention billionaires, too. It’s one of those things where I think the AI bubble is continuing to swell.

I was even chatting with some tech workers inside a bunch of different companies this morning, and they were all saying within their companies, they’re seeing the bubble swell and swell, and they’re being tasked with doing more and more AI, even if it makes them less productive and makes the work less good.

So I think there’s going to be a reckoning eventually, but right now we’re at peak AI hype and Anthropic does a great job of marketing its business as safe AI. Of course, they’re going to have more pressure around that marketing term now more than ever. So it’s going to be interesting to see the tension point kind of go into the spotlight and see how they handle this.

Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here. Not a subscriber? You can sign up here.

I was at the Semafor conference earlier this week. I saw Sam Altman and Sarah Friar, OpenAI’s CFO, on stage there. Altman and Greg Brockman were at the DealBook conference this week. I get the sense they’re kind of happy that Anthropic went first to see what happens in the markets and to see how companies value the revenue potential of their big competitors so they can tell their IPO story better.

Is that the sense you get, that Anthropic is the trial balloon? Or is OpenAI just retconning the fact that it didn’t go first?

Yes, I think that’s absolutely the case. I heard there was a time in the past year that Altman was really obsessed with beating Anthropic to go public. That’s what I heard. It was all he thought about. But then with all the bad press around OpenAI this year, they’ve had setback after setback. A lot of the public is particularly focused on hating OpenAI right now.

Obviously, there’s a big AI populism backlash and a lot of people hate a lot of AI, but it seems like OpenAI has really borne the brunt of a lot of public hatred and scandals this year. And so they were like, “You know what? Let’s just take a step back and let Anthropic go first.” And they seem happy to do that.

Now that Anthropic has gone first and set this historic precedent, OpenAI can ride on its coattails and say, “Yeah, us too.” I doubt they would beat Anthropic’s record. Maybe they will, maybe they won’t. But either way, I think they know that things could have gone a lot worse for them if Anthropic did not go first, if it’s happening this year.

I can’t stop thinking about Anthropic in terms of just dumb business fundamentals now that it’s a public company. You buy a token from them. They have to make that token for less than you pay for it. They also want to make AGI. Somewhere else they have some other huge set of costs.

Just based on what we know about the business today, now that it’s a public company, are they profitable on the face of it? Do they have a long way to go before inference is so profitable it can pay for training all the new models? How does their business actually work?

They have a long way to go and that is why it’s going to be so interesting to see what exactly they do here with regard to driving more revenue, which is what they’re going to have to do. I saw one estimate from Steve Rattner saying Anthropic is going to have to increase its revenue roughly 17-fold within the next 10 years to deliver a comparable return to some of its competitors in the tech industry.

But even without that specific number, you can tell. I mean, when a company goes public, investors want to see a return on their investment. They want to see profits go up or just profits at all. They also want to see that even if profits aren’t a thing yet, they want to see that you’re driving so much revenue that they could see profit happening in the near term. That is something that Anthropic is going to have to reckon with more than ever before after today.

I was chatting with a bunch of corporate governance experts and law professors and they were all saying Anthropic, like OpenAI, has such an unusual corporate structure that they aren’t actually beholden to shareholders in the way that some public companies are. Shareholders can never vote out executives. They don’t have a lot of power to influence anything really. It’s the founders and this special trust, which we can get to in a second, that have the power.

But either way, they were all saying it doesn’t matter if shareholders have a say here really in voting. It’s about the fact that no matter who has a say, Anthropic is going to be under so much pressure to make money and to buy compute and catch up with OpenAI in that way, that this is going to be a huge reckoning for them. We’re probably going to see a lot of things fall by the wayside, maybe even some expensive safety efforts.

I want to come to the governance structure angle here. We live in a time of Nvidia scandals. Corporate governance is all the rage now. It’s the hottest topic in American business, but let’s come to that in one minute.

I just want to stay on “can you sell a token for a dollar more than it cost you to make it” for one more turn. It’s just the dumbest way of thinking about this. Anthropic rents capacity from big cloud providers, from hyperscalers. They run the GPUs in them. They pay for power, they pay for the GPUs. There’s some input cost and then they have to sell the token for more than all those input costs.

That is the dumbest way of thinking about their business. I promise you that they think about it in a more sophisticated way than that, but that is how I think about it. I was just at the Microsoft Ignite conference this week and I saw Satya Nadella and he was very clear to me and others that they’re going to try to bring the price of tokens down, that their whole system is designed to make all the AI vendors fungible underneath Copilot and they’ll pick the best model for the task and make sure you get the cheapest one for whatever it is you’re trying to do.

That seems totally at odds with what Anthropic needs to do, which is get more and more margin for every single token they sell so they can keep renting capacity from data centers or buy more chips than OpenAI or whatever it is they need to do. How are they structured to solve that problem? Is it just more people using AI or is it more expensive tokens?

I think it’s a little of both, and it’s interesting because Anthropic has been kind of sneakily trying to do this. We wrote a piece in the past year about how there was a big outcry, particularly focused on Anthropic, for passing on costs to consumers. Individuals who were customers of Anthropic were really mad because they felt like the prices were going up for the same stuff or for worse quality and that costs were just getting passed on to them.

This became a thing across the industry, a big outcry, but Anthropic bore the brunt of it. Anthropic had to have a public statement. They’re like, “Look, we’re going to work on this.” And they did roll it back a little, but I mean, they’ve got to keep doing it in some way.

The other thing they’re focused on is hoping that their models are going to be good enough that people just form an attachment — not emotionally, well maybe — but form an attachment to them and they can moat people in. They’re all about the moat these days: “We’ve got to build a moat.” OpenAI is about this, too. They want so much of your everyday life and business to be married to Anthropic’s tools that it’d be really hard to port it all over and you won’t want to and you’ll just keep paying. So that’s what they’re going for.

They’re also, of course, hoping that enterprise and government contracts are going to really make up a bunch of this, in which case you don’t have to worry quite as much about nickel-and-diming because they’ve got a lot of money. But I think that’s the interesting part here. When it comes to people that do care about the price a lot, they’re just trying to moat them in and hope for the best.

What’s interesting there is that consumers are sensitive to price in one way. I think Anthropic’s big business customers, which is where they’ve been historically focused, are sensitive to price in different ways.

If you’re paying Anthropic a lot of money, you might want to be able to switch providers quickly and you might demand data sovereignty and all this stuff that big enterprises are thinking about. Anthropic either has to deliver that, which would kill their moat for their big enterprise customers, or say, “We’re so good, you can’t have that.” I’m not sure that they can say the second one yet.

That’s why they’re marketing this safe, trustworthy AI line so deeply, because it’s the only thing they’ve really got going for them here. When you have an enterprise customer, most of the time you do have to give them data sovereignty. You can’t trade on their data. There’s a lot of stuff that Anthropic could be doing to make more money that they can’t do when someone is paying that much.

I think that’s really why they’re driving home this point of being the safe, trustworthy alternative because that’s the only thing that’s really going to get people to stick around, particularly if they also drive home the cybersecurity tools in an age when everything’s getting hacked all the time.

The other interesting thing is even not with huge enterprises, but with small to medium-sized businesses, I’ve been chatting with a bunch of them and they’ve all been creating their own internal tools to kind of track which tasks they need specifically to use an expensive Anthropic model for and which ones they can port over to maybe an open-source model or a really cheap model.

So a lot of companies like law firms — and people in a lot of different industries that have small to medium-sized businesses — are literally creating their own internal benchmarks about, “Okay, let’s save money as much as we can to just only put 10 percent of our tasks toward the expensive Anthropic model, and then this 40 percent can go to an open-source model and this 50 percent can go to a mid-tier, cheaper model.”

So that’s what’s really interesting to me. People are trying to figure this out on their own and not even sharing those benchmarks publicly because it’s their own way of budgeting.

I actually heard Nadella this week say every company will have its own set of evaluations and they will test the models against them and always use the most efficient one and this will become one of the major things that every company does.

The flip side to that is I’ve heard Anthropic executives say, “Well, you will always want the frontier model because your competitor will always be using the frontier model and if you use a cheaper one, you’ll be left behind.” And that has always seemed somewhere between naive and idealistic to me because I think a lot of companies do cheap things knowing that their competitors will also do cheap things.

I agree completely. Everyone’s being cheap.

Yeah, I think if you can get away with good enough, you’re going to find out exactly where good enough is and you will not assume that your competitor is going to eke out a 5 percent advantage by using a more expensive model. Somewhere on the margin is where all of the costs should be and the market will figure it out.

Exactly.

It just feels like the basics of the business, maybe on IPO day, the pressure isn’t on that, everyone’s just very happy and rich. The next quarter’s earnings are going to be where we start to see this shakeout. Can it keep growing the way it’s growing at the margins it needs to build everything, especially when its own suppliers are now major shareholders?

Totally. And we saw one quarter where they experienced huge growth and then the next quarter wasn’t that much. So yeah, I mean they’re going to have a lot to reckon with.

Let’s talk about the governance issues. Because of Nvidia, corporate governance is the hottest topic in American business. Like you said, OpenAI and Anthropic have some of the most complicated governance structures in history.

If you think that Anthropic’s future is to sell the best intelligence — the most capable, the most cutting-edge frontier, close-to-AGI intelligence possible — you would probably not do a bunch of safety work. You would probably not kill your margins on a bunch of researchers who are telling you to stop or slow down or not release the best model.

Anthropic is deeply committed to doing that. You and I have had conversations before about whether that can withstand the pressures of capitalism and growth. That was hard when they weren’t a public company, but now they are. Talk about Anthropic’s structure and how they’re trying to codify their commitment to safety.

Yeah, that’s a really great question. I just published a piece about this. It’s so interesting to me that Anthropic, despite seeing all the roller-coaster drama that OpenAI went through around the Sam Altman ouster and all of the restructurings that that company has gone through, Anthropic has a similarly unusual structure. And in a way it’s helpful because they’re not beholden, as we said, to shareholder pressure besides if shareholders start selling off their stock or don’t buy more.

But the two kinds of groups of people that have power over what happens at Anthropic are the founders and this group called the Long-Term Benefit Trust (LTBT). Anthropic likes to market this as being like, “Oh, don’t worry about that the founders are doing what’s best for humanity but with no checks. We’ve got this trust above us that can appoint or remove us anytime.” As we’ll see in the movie Artificial, Sam Altman famously says, “Oh, well, the board can fire me whenever.” He says that a million times.

Now, Anthropic is implying the same thing: “Oh, we have this trust above us. They can remove or appoint the majority of board members,” even though it’s actually not all of them, but we’ll get to that in a minute. So this trust is a group of right now four people and none of them have a lot of AI experience. It’s [former Federal Reserve Chairman] Ben Bernanke, a foreign policy analyst named Richard Fontaine, an experimental physicist named Christopher Stubbs, and a former GiveWell executive named Neil Buddy Shah, who is chair of the LTBT.

I spoke to a couple of former Anthropic employees that left not too long ago and one of them said that this trust is kind of handpicked to not rock the boat at Anthropic. The founders appointed them. They’re not going to be appointing someone that’s going to probably disagree with them a lot.

Ironically, the trust doesn’t have the authority to step in on everyday decisions at Anthropic, unlike even the OpenAI nonprofit board that existed. They could say, “Oh look, don’t release this model. It’s too dangerous.” The trust doesn’t have any of those powers. All it can do is appoint or remove certain board members. Even then, they have to give 30 days’ notice to a small committee of the board, which includes Daniela Amodei, the chair of that little committee who is also a founder of Anthropic.

To oversimplify things, out of these seven board seats, the trust gets to pick three, the founders get to pick three, and then the kind of tie-breaking seat, the seventh one, the trust technically appoints, but they have to basically get founder approval to do it. So it’s one of these things that’s so complex, but when you really look at the nitty-gritty of it, the founders have pretty unchecked control here. I think that’s something that’s getting lost in the narrative of, “Oh, this trust is above us. Don’t worry.” Okay, well, they don’t really have much power to check you.

The founder control structure is the wildest thing I have ever read about in years of doing the show and thinking about business. Anthropic has seven founders. They have been issued a single share of stock that holds 50.1 percent of the voting power of the company. That is issued to a company called Founder LLC, and they’re the seven co-owners of Founder LLC.

How on earth does that work? Do they just all get together and then decide how to vote the one share of stock?

Yeah, it’s really interesting and that’s why I was kind of oversimplifying it a minute ago with who gets to choose this tie-breaking seat because it’s technically Founder LLC. It’s just a total mess. But when you really look at the simplified version, it’s that yes, through all of these weird structures and things they’ve done, they basically have pretty unchecked power and that’s the way they want it.

When I was talking to a former employee who recently left, they said it’s not necessarily that the founders have a financial interest in making the most money. In their eyes, it’s more that they have an interest in unchecked control. I think that’s a really important point here.

Some dissenters within the ranks of Anthropic think it’s not about them wanting to get rich, but rather it’s about them wanting that total control, which does track with a lot of their public statements. They say, “You can’t do safety from second place.” They’re all in on speeding up this race, but the problem is when you speed up the AI race, you’re speeding it up for everyone involved.

The safety piece is the really interesting portion of this to me because it’s the thing you would spend money on, absent some huge litigation drama or liability concern, that no other company would spend money on.

So let’s say it’s proven in court that model companies are liable for what people’s agents do. I think then all the model companies would build safety systems. They would have some big incentive to make sure their models are safe. That has not yet happened. We don’t know. There’s a big debate about safety and where the regulation will come from and what will create the incentives.

In the meantime, you’ve got seven founders holding a single share of stock who right now appear to be aligned on investing in safety in various ways inside of Anthropic. What happens if they start to disagree?

Yeah, that’s a great question. I think that the other interesting question here is what happens when you can explain away anything in the name of safety if you’re thinking it’s “for the greater good,” which as we know is a common villain trope in every movie. But it’s also one of Anthropic’s main things on its website. They’re like, “We want to do this for the greater good, for the global good.”

If you’re saying, “Hey, we have to win this race at all costs because we’re the only good guys and anyone else who wins it is going to do worse than us,” are you really the good guy at that point? And can’t you just explain away any single decision, even dropping some safety measures, as saying, “Oh, it’s slowing us down too much. We’re going to lose. We’ve got to win at all costs”? So you drop this or drop that. And then you look in the mirror at the end of the race and you say, “Oh, actually, we’re the exact same as the other company we’re trying to be different from.” I think that’s increasingly happening.

So I mean, right now they’ve said, “Look, we’re really committed to this.” You do have to remember that these people are almost religiously dedicated to this idea of safe AI. A lot of them think that AI systems are potentially conscious. I mean, this is by Anthropic founders’ own statements.

So I think that it’s not like they’re just secretly only trying to make money and just talking about safety and they don’t mean it. I think it’s more something a little more dangerous in that they want complete control because they believe that they have the right plan for AI safety and they will do it better than anyone else.

But as we know, absolute power corrupts absolutely. So a lot of Anthropic employees that I was talking to, former ones, were saying, “Yeah, it does not make sense that this industry is so unregulated. There’s no external regulation. Even if these seven people think they have the best idea for AI safety, okay, then prove it to someone else — that you shouldn’t be the only ones deciding this. Because you’re not elected.” So that’s the consensus among some.

They have a morally binding handshake agreement to allow external auditors, Hayden. It’s going to be fine.

Oh, yeah, with employee-level access.

Also, it’s called “Super Intelligence” now. It’s going to be totally fine. They all went to the White House and they looked deeply into each other’s eyes and said it was going to be fine.

Obviously Anthropic came up with this structure well before the events with Nvidia and Jensen Huang, but I can’t help but put them together in my head. They are happening at the same time. And Nvidia has a very traditional corporate governance structure.

Jensen owned 3 percent of the company, but he had a board of directors. The board of directors determined that he had violated the company’s code of conduct in some way. He resigned. It’s very obvious that he was thoroughly encouraged to resign, but that’s just normal corporate governance. That’s what you learn about in school. A company has a board of directors. Their number one job is to hire and fire the CEO.

That’s very different in Silicon Valley, right? Founder control of companies is a big story everywhere. Mark Zuckerberg famously has control over Meta, even though Meta shareholders repeatedly vote to remove him, which is one of the underemphasized stories in all of American business. Sergey Brin and Larry Page control Google.

Is this more in the, “We need to insulate our founders from the demands of the market and boards of directors” in classic Silicon Valley style? Or is it, “We need to insulate our founders and Dario from the demands of the market so we can stay focused on AGI and safety”? Because it feels like a combination of both.

It’s definitely a combination. I mean, they say it’s the latter. I think in a lot of ways it’s like the former. It’s totally a combination. I’d say it’s about 50/50. I mean, they want control so that they can focus on what they think is best and not lose control to someone that they don’t think knows best. But ironically, as we know, no one group of people should be in control of what is best here. So at least that’s what a lot of my sources say.

So I think that, yeah, I mean, it’s just a combination. Elon Musk famously won’t run a company that he doesn’t have voting control over. So I think it’s just a new Valley thing and it’s pervading here for a couple of different reasons.

There are seven of them. What seven different people do over the course of the next five to 10 years is as unpredictable as anything. What happens if one of them does something illegal or lies to the board or smuggles chips to China or anything? Is there some way to remove one of these founders from this LLC?

Yeah. I mean, allegedly the trust has the power to remove. They haven’t said whether they have the power to remove all of the board or just some, but in some way when you’re looking at the fine print of the S-1, basically if the trust gives 30 days’ notice to essentially Daniela Amodei and her other two people on this subcommittee, then they can suggest removal of a board member. Again, they can waive that 30-day notice if they deem it correct, if Daniela does.

But yeah, there’s a lot of fine print here. Essentially, yes, you can remove someone and there is stuff about that in the S-1 saying if they’ve done something illegal, if they’ve done something really bad, they can be removed, but it’s definitely not easy. It also depends on who it is, I think.

It’s interesting. I keep coming back to this because it is so complicated, it’s so unusual, and it is right next to the Nvidia situation where he materially misrepresented his relationship to the shareholders by having his PR people issue statements saying it was all normal. He had to admit that it wasn’t all normal and he was in fact involved in extramarital affairs and the board fired him.

I felt a little bad for that PR guy who had to put his name on all those statements that ended up being a lie.

I feel nothing. As the audience may know, as Hayden is well aware, we make all the PR people put their names on everything just in case something like this happens. That’s just a little bit of Verge trivia.

Anyhow, right, I mean, that’s just normal governance. Here you have a company where the only check might be the market itself, right?

Yep.

If they can’t sell tokens for more than what they make them for, they’re going to have to report that in the financials and the market might say this company is failing and pull all the money back. That is a very different kind of pressure.

If the only meaningful way that Dario might lose his job is if the company goes bankrupt and then the swirl of things is he’s got to buy chips before OpenAI can buy the chips, he’s got to increase the number of verticals that Anthropic competes in within the enterprise, and he also idealistically has to train the next model and the next model with RSI to get to AGI, it feels like actually the pressure is to go as fast as possible. If I just added that up, I would say I would not slow down ever.

Yeah. That’s why I say you can kind of explain away anything if you’re trying to be first, which they are. And they’ve been very public about that. Regarding the money thing, Ann Lipton, a law professor at the University of Colorado Law School, told me this week that the trust and the founders together are the ones who actually have governance power and that you either trust that they’re going to balance the interest of society appropriately or you don’t.

But either way, the real constraint on whatever impulses they have is going to be the fact that Anthropic needs money. That’s the thing, Anthropic needs money. In a livestream recently, Dario told the CEO of JPMorgan Chase that compute is the “deepest business dilemma” that the company faces. Essentially if you buy too much, you can financially overextend yourself. If you buy too little, then you can’t support inference for customers.

Even more importantly, you can’t do the R&D. You can’t come up with new models. So they’re extremely compute-constrained, even maybe more so than OpenAI because OpenAI was made fun of for a long time for losing a ton of money, but they were investing a lot in compute. Anthropic did not do that as much, and now they’re having to pay the piper.

So it’s an interesting turnaround here. Hopefully for them and their interests, I guess, if they can get the compute they need. But as we know, the supply is a lot lower than the demand. So it’s just going to be interesting to see what they will do to make money and therefore money to buy compute.

It’s funny. A couple of weeks ago I told our colleague, Liz Lopatto, that the Anthropic IPO would be one of the biggest stories of the year. She was laughing at me today, because she says it’s not even the biggest story of the week because of Jensen Huang and Nvidia.

My argument — take it or leave it, and you can judge whether I’m correct — is that they are the same story. That if Nvidia’s financing architecture fails or wobbles because Jensen isn’t there to bully everyone into believing in it, then maybe you do have a glut of chips sitting in data centers without the appropriate permitting or power to light them up and the market for chips will change dramatically.

Maybe a company like CoreWeave does get way more wobbly or some of the already wobbly neoclouds start to get wobbly and the price of compute falls and lets Anthropic compete differently. Maybe a bunch of chips that were allocated to China on the gray market come flooding back to the United States.

That apparently will happen.

Right? So there’s a lot of ways the Nvidia story affects Anthropic’s financials. Has the company said anything about that yet? Is the market reacting to that in any way?

Not yet. They’ve just talked about how important compute is and that it’s going to be a really big issue for them. And Anthropic’s IPO, it has a lot of ties to Nvidia and there’s a lot of complexities here because Nvidia had agreed to buy $10 billion in Anthropic shares via private placement on top of a $10 billion funding promise it made last year.

So they are in bed together, and that’s going to be interesting to see, too. The new CEO of Nvidia has made a lot of claims about diverting chips that were potentially going to China and bringing them back so they’re not doing that anymore.

Now we’re going to see if there’s more supply or not. Also, Jensen personally made some of these deals, especially with large financial institutions to fund the AI infrastructure buildout last August. So he’s wheeling and dealing a lot. And so it’s going to remain to be seen whether the new chief executive has that kind of power to wheel and deal and pressure people into doing stuff like this or not.

This is my last question. Obviously everyone’s just celebrating today. We are probably throwing the most cold water on this by noting that the structure of Anthropic is very odd. What should people be looking for? What are the signs that you’re looking for now that Anthropic is a stock that’s trading? What’s the relationship in your mind between the value of Anthropic stock on any given day and the broader AI story?

I think that stock is a little bit like benchmarks in the AI industry in that sometimes it doesn’t mean what you think it means and also it never means everything. So you always take it with a grain of salt. It’s all vibes based. It’s all hype. So people get scared, they sell, and people hear one good thing or a rumor and they buy. So I think it’s important to look at it and see what’s driving that fear or hype that day or that week. But it’s also really important to see what the company says in these reports and earnings calls.

That’s what I think is the most interesting. Not really the numbers all the time, but instead, what they try to sneak into these reports by saying, “Oh, we sidelined this team,” or, “We discontinued this safety initiative all in the name of being first.”

That’s the type of thing that I think we’re going to see a lot of and that’s what I want to watch for. Because the company that made its bed on safe AI is increasingly looking the same as OpenAI and is going to have a lot of swept-under-the-rug type of updates like that, if my sources are right.

You heard it here first. Former CNBC reporter Hayden Field says, look past the day-to-day swings into the fundamentals of the company. That’s good. Also, the stock market is only vibes. That’s the other thing I heard you say.

Yep. I mean, it’s controversial, but after being a financial reporter, yeah, I mean, there’s a lot of vibes there. It’s a lot of vibes.

Oh, I agree. I think sometimes the only thing they ask me to go on CNBC to say is this is just vibes, and then they usher me off the air very quickly.

Hayden, this is great. We’re obviously going to be talking about this story many times with you in the year to come. Hopefully we can do another straightforward business story and not just a soap opera one.

I’ll be here for both. We’ve got to do them both.

Amazing. Thank you so much, Hayden.

Thanks.

Questions or comments? Hit us up at [email protected]. We really do read every email!

Decoder with Nilay Patel

A podcast from The Verge about big ideas and other problems.