Anthropic Unveils a Brand-New Governance Architecture, Redefining the Capital Rules for AI Companies
Dario Amodei, who once served as Vice President of Research at OpenAI, co-founded Anthropic. Backed by the Claude large language model, the company has ranked among the world's top AI enterprises.
When news of Anthropic's potential IPO spread, most public attention focused on its revenue growth rate, model capabilities, and the expected trillion-dollar valuation. Few people noticed its governance arrangement that differs from the conventional Silicon Valley model.
This is hardly surprising. For Silicon Valley tech companies going public, the control arrangement has long followed a fixed template, mostly the dual-class AB share structure that keeps voting rights in the hands of founders.
Anthropic did not adopt the traditional model, but instead adopted an architecture combining Public Benefit Corporation (PBC) and a long-term independent trust.
Under this design, the power of Anthropic is neither in the hands of the founders nor the investors, but delegated to a group of external trustees who hold no shares and do not share profits. Former Federal Reserve Chairman Ben Bernanke is on the list of these trustees.
This design breaks the long-standing underlying consensus of the capital market: listed companies should prioritize ensuring returns for shareholders.
But when the priority of AI safety and public mission is raised above capital returns, how should the secondary market price the company?
Has the value of the equity held by investors been quietly discounted? This is not a theoretical legal issue, but a real test for Wall Street in Anthropic's upcoming listing.
Its subsequent development will also provide a reference sample for all large model companies preparing to go public in the future.
Disassembly of the dual-layer structure: The trust holds the real power
In the US capital market, the vast majority of technology enterprises adopt the traditional commercial company form, namely the C Corp.
For such traditional companies, there is a mandatory legal requirement: when the management and the board of directors make major decisions, the primary goal must be to generate more profits for shareholders. If the management gives up short-term high profits for considerations such as safety and social impact, shareholders can fully sue the board of directors and claim for compensation.
The Public Benefit Corporation (PBC) is also a profit-making enterprise, which can conduct normal financing, issue equity, and achieve listing and dividend distribution in the future, competing in the market like ordinary commercial companies.
Its core change is that it rewrites the legal responsibilities of the board of directors.
The entire governance structure of Anthropic is formed by the superposition of two systems. One is the Public Benefit Corporation as the legal carrier, and the other is the long-term interest trust that holds actual control. The combination of the two forms this very special constraint mechanism.
Under the rules of PBC, when making decisions, the management does not have to blindly prioritize shareholder returns, and can legally incorporate public safety and social impact into the scope of trade-offs.
AI companies take vastly different governance paths: Anthropic chooses a compromise route
In fact, this corporate form was not pioneered by Anthropic. Outdoor brand Patagonia, crowdfunding platform Kickstarter, and ice cream brand Ben & Jerry's all adopted the PBC structure very early.
Looking at the world's top AI track, leading enterprises have taken three completely different paths in the choice of corporate structure. Taking OpenAI, Anthropic and xAI as examples, the differences are particularly distinct.
The main body of star company OpenAI for external commercial operation is also registered as a PBC. OpenAI adopts a special structure: a non-profit foundation is at the top to hold control, and a profit-making subsidiary is set up below to carry out commercial operations, so as to restrict the profit-seeking behavior of capital.
Compared with OpenAI, Anthropic's route is more of a compromise. The two seem to be designed to restrict capital, but there are essential differences in their underlying structures.
The top layer of OpenAI is a non-profit foundation; while the main body of Anthropic is still a profit-oriented Public Benefit Corporation (PBC).
Anthropic only separates the corporate control and delegates it to the external trustees of the independent trust. The OpenAI foundation directly holds equity, while Anthropic's trust does not hold shares, and only holds voting control.
In 2023, Elon Musk founded xAI, with the original intention of benchmarking against OpenAI. The company was briefly registered as a PBC, but later chose to abandon this structure and switch back to the ordinary commercial company structure.
xAI removed this additional mission constraint, and all its decisions revolve around business growth. It has flexible capital operations, its expansion speed is not limited by the system, and its safety bottom line relies on corporate self-discipline, without hard constraints from the legal and trust levels.
Anthropic is positioned between the two. It retains the market-oriented financing channel, and will not be constrained by a non-profit entity like OpenAI; at the same time, relying on the independent trust mechanism, it transforms the safety mission into a hard constraint, reducing the possibility that the short-term impulse of capital will override long-term risk control.
Anthropic tries to balance two things that are difficult to be compatible: the freedom of business expansion, and the public responsibility that comes with AI technology.
This differentiation hides the underlying divergence of the industry: are super large models just an ordinary business, or a technology with infrastructure attributes and inherent public risks? Different judgments have given rise to completely different corporate systems.
Why do large model companies attach so much importance to corporate responsibility?
In fact, the technical risks of large models have strong spillover effects. If commercialization is blindly promoted, a series of social risks will be derived.
If the governance framework of an ordinary commercial company is adopted, capital can easily force the management to release model permissions and launch high-risk businesses in order to achieve good short-term financial statement performance. The PBC leaves space for the board of directors at the legal level, allowing it to prioritize AI safety over short-term profits.
But Anthropic judges that PBC alone is not enough. PBC can only protect the board of directors legally, but it is difficult to prevent shareholders or founders from adjusting the development direction under huge commercial temptation. To this end, the company has additionally set up the LTBT long-term independent trust to build a dual constraint mechanism.
PBC is only a legal carrier, and it cannot realize the constraint on capital by itself. The real power holder, which is also the most controversial part, is the LTBT long-term interest trust.
The trustee team is composed of externally selected experts, who do not hold Anthropic's equity, cannot enjoy the appreciation after listing, do not participate in corporate dividends, and their remuneration is only used to cover their time cost.
The trustees have two core powers: first, they have the right to replace most members of the board of directors, and hold the core voice at the personnel level; second, when the company has major transactions such as large-scale unconventional debt, selling more than half of the equity, or transferring core AI models reaching the threshold, it must report to the trust five days in advance.
Many commercial law scholars pointed out after studying the relevant clauses that this five-day report is not a simple notice, and will form a soft veto power after implementation.
Once the trust suspects that the transaction has potential safety hazards, it can prevent the transaction from being implemented through methods such as personnel appointment and removal and continuous pressure.
This is essentially different from the AB share model: the AB share model relies on founders to hold high voting rights, and the power remains inside the company; while Anthropic delegates this core power to a third party independent of shareholders and the founding team.
Long-term barriers and unavoidable valuation discount
Anthropic's architecture is equivalent to giving up part of the shareholders' natural rights in advance.
This architecture has formed two completely opposing views within institutions, and the tolerance between the primary market and the secondary market is particularly different.
The logic of the bulls is very clear. As a third-party gatekeeper, the trust can avoid the control struggle and the internal friction between founders and investors that are common in Silicon Valley companies, and ensure strategic continuity.
In the trend of tightening global AI regulation, the governance structure with built-in safety constraints is itself a layer of compliance moat. Capital can hardly force the team to take technical risks for quarterly reports, and slow variables such as R&D investment and safety assessment can continue to obtain resources.
Institutions in the primary market are more willing to pay for this long-term narrative. The concerns of the secondary market are also realistic, and they are the largest source of uncertainty for future stock prices.
The core contradiction is the structural dilution of shareholder rights. Investors buy stocks, but it is difficult for them to influence transactions that are enough to change the fate of the company.
The switch of major matters is in the hands of a group of trustees who are not responsible to shareholders. Once the trust vetoes a merger and acquisition or licensing plan that can bring huge profits for safety reasons, shareholders can only passively accept the result, and there is no effective means of checks and balances.
Traditional due diligence can calculate revenue, gross profit margin and competitive landscape, but the market cannot build a valuation model based on the subjective judgment of the trustees.
This creates an unavoidable paradox: the company has top-tier technology, rapidly growing business, and all the characteristics of a high-growth target, but the rights carried by its equity are incomplete. The market will eventually give a discount for this part of the right defect.
Betting on long-term space, the primary market can accommodate this discount; most of the funds in the secondary market have a clear assessment cycle, and their attitude towards such uncertain governance risks is naturally more cautious.
This cognitive gap will run through the pricing game for a long time after Anthropic goes public.
A governance experiment for the whole industry, more tests are yet to come
From a longer-term perspective, the significance of this matter has gone beyond the success or failure of Anthropic's own IPO.
In the short term, the AI market is hot enough, and its scarcity will cover up the disputes over governance, so the issuance will most likely be completed smoothly, and there will be no shortage of subscription demand.
In the medium term, with the fluctuation of stock prices after listing and the successive implementation of major transactions, discussions about excessive trust power and opaque rules will appear repeatedly. Once the trustees' decisions come into direct conflict with the interests of investors, shareholder lawsuits, institutional position adjustments and valuation revaluation are all predictable scenarios.
In the long run, its demonstration effect cannot be underestimated. Global regulatory requirements for AI safety will only continue to tighten, and the model that only relies on corporate self-discipline and the personal constraints of founders will gradually lose its supporting power.
If Anthropic's mechanism can run smoothly and maintain a balance between business expansion and risk control, it has the opportunity to become a new model for leading large model enterprises.
The capital market has operated for a hundred years, and its foundation is built on the logic of shareholder supremacy. Round after round of technology waves have reshaped products and business models, but this underlying logic of governance has rarely faced substantive challenges.
AI is different. Its risk spillover and social influence are not possessed by the Internet and semiconductor sectors. The old capital rules may not be suitable for this new technology.
We are currently witnessing a slow process of institutional adaptation. There is no standard answer for the industry for the time being on whether capital should give up part of its power for public safety, and how to weigh commercial returns against long-term risks.
With this IPO, Anthropic is testing the waters for the entire industry. The final direction of this experiment is not only related to the stock price story of this company, but also will define what kind of attitude the giant AI enterprises in the future should take to enter the public capital market.
This article is from the WeChat Official Account "Wall Street Tech Eye", author: Du Jie, editor: Sun Congcong, 36Kr is published with authorization.