IPO 之后，究竟谁在真正掌管 Anthropic？

Anthropic 成立于 2021 年，旨在确保人类能够掌控 AI 并长期从中受益。然而，随着这家前沿实验室日益临近首次公开募股（IPO），人类中的一类群体却被完全排除在控制权之外：任何在公开市场上购买其股票的人。

在硅谷，允许埃隆·马斯克（Elon Musk）或马克·扎克伯格（Mark Zuckerberg）等创始人在不持有大部分股权的情况下控制上市公司的复杂架构早已屡见不鲜。但 Anthropic 对 AI 安全的执念，催生出了一种真正称得上“巴洛克式”的公司治理风格。这使得普通股股东对这家 AI 公司的绝大多数重大决策（包括谁进入董事会）几乎没有任何发言权。

首先，Anthropic 是一家公共利益公司（Public Benefit Corporation, PBC）——这是一种在法律上被允许将公共使命与利润并重的营利性公司，其目标是促进 AI 安全。这一法律地位保护了董事会有权做出并非完全由股东价值驱动的 AI 安全决策，而免于遭到愤怒投资者的起诉。

目前，董事会由 Anthropic 总裁丹妮拉·阿莫代（Daniela Amodei）担任主席，成员包括 CEO 达里奥·阿莫代（Dario Amodei，丹妮拉的哥哥）、Netflix 联合创始人里德·哈斯廷斯（Reed Hastings）、微软和通用汽车前首席财务官克里斯托弗·利德尔（Christopher Liddell）、诺华制药（Novartis）首席执行官瓦斯·纳拉辛汉（Vas Narasimhan）以及 Spark Capital 合伙人亚斯敏·拉扎维（Yasmin Razavi）。目前董事会尚有一名空缺。

其次，Anthropic 将董事会的重大控制权赋予了一个“长期利益信托”（Long-Term Benefit Trust, LTBT）。这是一个旨在确保公司专注于造福人类使命的独立委员会。该信托持有一种特殊类别的股票，使其能够选举和免职董事会的大多数成员（称为 T 类董事）。LTBT 由全球卫生专家尼尔·沙阿（Neil Shah，曾任慈善机构评级非营利组织 GiveWell 的常务董事）担任主席，成员包括美联储前主席本·伯南克（Ben Bernanke）、国家安全专家理查德·方丹（Richard Fontaine）以及哈佛大学物理学家克里斯托弗·斯塔布斯（Christopher Stubbs）。LTBT 目前也有一名空缺，现有成员将自行推选人选填补。

第三，Anthropic 的七位创始人还共同持有一种特殊类别的股票。只要创始人中的多数达成一致，该股票就能赋予他们 50.1% 的股东投票权，使他们有权选举董事会中的三名成员。如果出于某种原因，创始人群体内部未能形成多数意见，他们的股份才会与普通股持有人的股份按比例一同投票。

最后，董事会也可以制衡 LTBT。当 LTBT 提名其四名董事中的某一“特定”董事，或者在无正当理由（即未指控存在不当行为）的情况下免职该董事时，董事会的提名委员会可以拒绝批准这一举措，并将决策权交由公司员工进行投票表决。

以下是整套治理架构的示意图：

图片来源： Tim Fernholz/TechCrunch

这样一套复杂结构的动态博弈难以预测。我们此前已经目睹过类似结构实际上被推翻的一个著名案例：即 OpenAI 经历的所谓“小插曲”——当时 OpenAI 董事会因担忧 CEO 萨姆·奥特曼（Sam Altman）对其不够坦诚而短暂将其解雇。由于公司员工对奥特曼的高度忠诚，董事会最终被迫让步并恢复其职务，随后建立了一套新的治理结构。Anthropic 的这套架构似乎是将这种员工权力制度化了。

在 Anthropic，若创始人统一战线出现裂痕，可能会让普通股东拥有超出预期的话语权；同时，如果员工选择支持董事会而非那群业界泰斗，员工也可以在董事席位上否决 LTBT。

创始人的控制权并不会永远持续：在 IPO 后的十年，创始人可以选择终止对公司的控制；或者当创始人仅剩不足三人时，其控制权将自动终止。届时，LTBT 将继续存在，直到 85% 的普通股股东投票决定终结它，或者在获得董事会批准的前提下自行决定解散退出。

如果且唯当这两类特殊股东都不复存在之时——但愿这不会是因为一场由 AI 引发的灾难——普通股股东才将最终赢得控制权。

Anthropic was founded in 2021 to ensure humanity can control AI and benefit from it in the long-term. As the frontier lab approaches its initial public offering, however, there’s one class of humanity that is being shut out of control: anyone who buys its stock on public markets.

Complex structures that allow founders like Elon Musk or Mark Zuckerberg to control their public companies without owning most of them are nothing new in Silicon Valley. But Anthropic’s obsession with AI safety has resulted in a truly baroque style of corporate governance that will leave common shareholders with little say over most major decisions at the AI company, including who sits on its board.

First, Anthropic is a public benefit company — a for-profit corporation legally allowed to weigh a public mission alongside profit — with the goal of promoting AI safety. That status protects the board’s ability to make decisions about AI safety that aren’t driven entirely by shareholder value without being sued by angry investors.

Right now, the board is chaired by Anthropic president Daniela Amodei and includes CEO Dario Amodei (her brother), Netflix cofounder Reed Hastings, former Microsoft and GM CFO Christopher Liddell, Novartis CEO Vas Narasimhan, and Spark Capital partner Yasmin Razavi. There is one vacancy on the board.

Second, Anthropic has allocated significant control of its board to a Long-Term Benefit Trust (LTBT), an independent panel meant to keep the company focused on mission to benefit humanity. The trust owns a class of stock that gives it the ability to elect and remove a majority of the company’s board, known as Class T directors. The LTBT is chaired by Neil Shah, a global health expert who was managing director at GiveWell, a nonprofit that ranks charities, and includes former Federal Reserve chair Ben Bernanke, national security expert Richard Fontaine, and Harvard physicist Christopher Stubbs. There is one vacancy on the LTBT right now. The existing members will themselves choose who fills it.

Third, Anthropic’s seven founders also jointly own a special class of stock that gives a majority of them 50.1% of shareholder voting power, allowing them to elect three members of the board of directors. If, for whatever reason, there’s no majority among the founders, only then are their shares voted in proportion with holders of common stock.

Finally, the board can put a check on the LTBT. When the LTBT nominates a “Specified” one of its four directors, or removes that director without cause (that is, without alleging misconduct), the board’s nominating committee can decline to approve the move and throw the decision to the company’s employees for a vote.

Here’s what all that looks like:

Image Credits: Tim Fernholz/TechCrunch

The dynamics of such a complex structure are hard to predict. We’ve already seen one notable instance where a similar structure was essentially overthrown: OpenAI’s so-called blip, when OpenAI’s board briefly fired CEO Sam Altman over concerns that he was not being candid with them. Thanks to the loyalty of the company’s employees to Altman, the board backed down, reinstated him, and a new governance structure was put in place. Anthropic’s structure seems to formalize that kind of employee power.

At Anthropic, cracks in the founders’ unified front could lead to common shareholders having more say than expected, while employees can overrule the LTBT on board seats if they back the board over the assembled luminaries.

The founders’ control won’t last forever: Ten years after the IPO, the founders can choose to end their control of the company, or their control will end automatically when fewer than three of them remain. At that point, the LTBT will continue until 85% of common shareholders vote to end it, or it decides to step down with approval from the board.

If and when neither special class of shareholder exists any more — hopefully not because of an AI-driven catastrophe — common shareholders finally get control.