Anthropic has filed its S-1 with the SEC, and the governance architecture buried inside it is unlike almost anything the public markets have seen before.
The story begins, as so many do in Silicon Valley, with a founder who wanted to change the world without surrendering control of the company built to do it. When Dario Amodei and six co-founders left OpenAI in 2021 to establish Anthropic as an AI safety company, they were building not just a product but an institutional philosophy — one premised on the idea that artificial intelligence poses risks serious enough to require a different kind of corporate stewardship. That philosophy has now found its way into one of the most elaborate share structures ever submitted to American securities regulators.
The S-1 filing, made public this week, describes five classes of stock — A, E, F, N and T — each engineered to serve a distinct purpose. Two of those classes consist of a single share each. Three carry no economic rights whatsoever. The class available to ordinary investors, Class A, listed on Nasdaq under the ticker “ANT” at one vote per share, can never collectively hold more than 49.9% of the total vote, regardless of how many shares are sold or how many investors buy them. The filing is candid about what this means: Class A holders “generally will not be able to affect the outcome of any matter submitted to a vote of our stockholders or the election of our directors, regardless of how they vote.”
The mechanism that produces this outcome is Class F — a single share of stock, held by a limited liability company owned jointly by the seven co-founders, that carries 50.1% of the vote on almost every matter shareholders decide, including the election of most directors. It has no economic rights, meaning it entitles its holder to no dividends and no share of the company’s value. New share issuances cannot dilute it. Each co-founder retains a seat in the LLC — and therefore a share of that controlling vote — as long as they hold at least 2% of their IPO-era equity stake. The filing notes that each founder has pledged to give away 80% of their Anthropic holdings, but those pledges are described as “voluntary, revocable expressions of philanthropic intent.” If a majority of the co-founders cannot agree on how to cast the Class F vote, it defaults to voting in proportion to the public Class A split — a provision that gives ordinary shareholders a narrow, contingent backdoor into governance.
Class F disappears once fewer than three of the seven founders remain in the LLC for a continuous period of ninety days. That sunset clause distinguishes it from some comparable structures, though it offers little immediate comfort to investors who may hold shares for decades before the clause triggers.
The second pillar of control belongs not to the founders but to Anthropic’s Long-Term Benefit Trust, which holds Class T stock — also a single share, also carrying no economic rights, and carrying no expiration date. The trust’s five trustees, one seat currently vacant, use Class T to elect a majority of the board. The chair is Neil “Buddy” Shah, CEO of the Clinton Health Access Initiative and a former managing director of the charity evaluator GiveWell. Richard Fontaine, who runs the Center for a New American Security and once advised Senator John McCain on foreign policy, also sits on the trust. Ben Bernanke, the former chair of the Federal Reserve, joined in June. Harvard physicist Christopher Stubbs, who joined this month and shared in the discovery that the universe’s expansion is accelerating, rounds out the current membership. Trust policies prohibit trustees from purchasing or holding any Anthropic securities, and their compensation cannot vary with the company’s profits — provisions designed to insulate the board-selection mechanism from financial self-interest.
Taken together, the two structures mean that the trust selects most of the board while the co-founders control every other shareholder vote. Public investors and strategic partners — Amazon and Google among them, holding Class N shares that carry economic rights but no votes — participate in the company’s financial upside while exercising almost no governance influence at all.
None of the individual components are entirely new. The architects of Anthropic’s structure appear to have studied the recent history of founder control carefully, borrowing selectively from three well-known precedents while attempting to address the vulnerabilities each revealed.
The closest parallel in recent memory is Elon Musk’s SpaceX, which went public in June and granted Musk approximately 82.4% of the company’s voting power immediately after its offering, with 81.1% of that figure flowing from his 10-vote Class B shares. SpaceX’s offering documents acknowledged that the structure “concentrates voting control with Mr. Musk” and specified that Class B holders elect a majority of the board independently. That control, however, is tethered to ownership: if Musk were to sell enough Class B shares to lose his majority of that class, another holder could assume board control. Anthropic’s founders face no equivalent constraint. They could sell 98% of their shares and still retain their seats in the LLC that holds the Class F vote.
The Zuckerberg precedent is more complicated, and more instructive. As of April 2026, Mark Zuckerberg held 99.8% of Meta’s Class B stock, which carries ten votes per share, giving him 60.8% of the company’s total voting power while his economic stake sits at roughly 14%. In December 2015, Zuckerberg and his wife Priscilla Chan pledged 99% of their Facebook shares to charity, to be managed through the Chan Zuckerberg Initiative. That created an immediate problem: every share donated would carry ten votes out of the company with it. Facebook’s proposed solution was a new nonvoting Class C stock, issued two-for-one to existing shareholders, which Zuckerberg said would allow him to “keep founder control of Facebook.” Shareholders sued, calling it “a recipe for disaster,” and in September 2017 — days before Zuckerberg was scheduled to testify in Delaware — he abandoned the plan. He cited the rise in the stock price as sufficient to fund his philanthropy while preserving his voting majority. A version of the scheme was revived last week, when Meta announced a new non-voting Class C share structure to allow Zuckerberg to fulfil his Chan Zuckerberg Initiative pledges and, potentially, to address a California wealth tax if that legislation becomes law. Anthropic’s Class F structure is designed to avoid the same dilemma from the outset, separating economic rights from voting rights at the point of founding rather than retrofitting the solution after a public controversy.
The New York Times offers a longer and, in some respects, more reassuring historical comparison. The Ochs-Sulzberger family has controlled the paper since Adolph S. Ochs purchased it in 1896. Public Class A shareholders elect four of the paper’s thirteen directors; the remaining nine are elected by Class B stock, roughly 94.6% of which sits in a family trust directed to keep the paper “entirely fearless, free of ulterior influence and unselfishly devoted to the public welfare,” in the words of Ochs’s will. Family members wishing to sell Class B shares must first offer them to relatives and to the company. The trust’s structure has held for well over a century. Anthropic’s filing cites a similar logic: its co-founders are “distinctly equipped to be stewards of our mission,” and leadership should not face “pressure to prioritize short-term results or purely pecuniary objectives.”
The filing also invokes, by implication, a cautionary tale about what happens when family trust structures fracture. The Bancroft family once controlled 64% of Dow Jones’s vote with less than 25% of its equity, but their Class B shares were distributed across a series of trusts with roughly thirty-five beneficiaries who were under no obligation to vote together. Rupert Murdoch exploited that division with a $60-per-share offer and closed the acquisition in 2007. The special editorial committee he agreed to establish to protect the Wall Street Journal’s independence learned of its managing editor’s resignation the following year through phone calls from Murdoch himself. Anthropic’s structure attempts to pre-empt a comparable outcome by vesting the controlling vote in a single LLC rather than a dispersed family, and by separating that vote entirely from economic value — the asset Murdoch purchased.
The filing does not pretend the arrangement will be universally welcomed. Dual-class structures have been on the rise: in 2025, fifteen of the thirty-one tech companies that went public in the United States had more than one class of stock, and 41% of all IPOs did, compared with roughly 10% across the previous four and a half decades, according to data compiled by University of Florida finance professor Jay Ritter. Ritter’s research also found that dual-class tech IPOs since 1980 outperformed comparable stocks by about nineteen percentage points over their first three years, while single-class tech IPOs trailed by 2.4 points — a finding that complicates the standard shareholder-rights critique without resolving it.
Anthropic’s own filing acknowledges that the structure “may effectively eliminate the ability” of Class A holders to influence major decisions, including a sale of the company, and warns that it “cannot predict whether our structure will result in lower demand for our Class A common stock, increased volatility, adverse publicity, or other adverse consequences.” It also flags the legal uncertainty directly: “the absence of judicial precedent interpreting this structure increases the risk and unpredictability of governance-related litigation.”
One further disclosure in the filing attracted notice for a different reason. Anthropic has agreed to spend up to $120.5 billion on Nvidia-powered computing capacity from SpaceX.AI, a SpaceX subsidiary formerly known as X.AI. Either party may cancel the arrangement on ninety days’ notice once early 2027 has passed — a commercial relationship between two companies whose founders now share, in very different ways, the architecture of control that defines the post-IPO AI industry.
